Alg Solutions of Diff - Eq

Inventiones math.
18, 1 - 118 (1972) 9 by Springer-Verlag 1972
Algebraic Solutions of Differential Equations (p-Curvature and the Hodge Filtration)

Nicholas M. Katz (Bures-sur-Yvette)
Introduction
This paper grew out of an attempt to answer the following question, first raised by Grothendieck. Consider a linear homogeneous n system of first-order differential equations
-
d (Y)=A(z) Y dz
-
in which A(z) is an n x n matrix of rational functions of z. To fix ideas, suppose that the coefficients of the entries of A(z) all lie in an algebraic. number field K. Then for almost all primes p of K, it makes sense to reduce this equation modulo p, obtaining a differential equation over (I) Suvpose that for almost all primes p, the reduced equation has a full set of sdutions (i. e., has n solutions in (Fq(z))" which are linearly independent over Fq(z)). Does the original equation admit a full set of solutions in algebraic functions of z? For example, the equation d 1 dz (Y)= ~-z y with a e Z may be reduced modulo p for all those primes p not dividing a, and the reduced equation admits the solution zb for any integer b such that a b= 1 modulo p. The original equation has for its solution the function z 1/a. Of course, (I) may be reformulated in greater apparent generality. Let R be a subri'ng of C which, as a ring, is finitely generated over Z. Let S be a smooth R-scheme with geometrically connected fibres, and consider a differential equation (M, 17) on S/R, by which we understand a locally free sheaf M on S of finite rank together with an R-linear integrable connection 17: M - , f21S/R| (We considered above the case R=(9[1/n], (P the ring of integers in an algebraic number field, S an open set in P~, M = (9], and V: M ~ f 2 ~ s / ~ | given by
17m=dm-dz|
1 Inventionesmath., VoL 18
A(z) . m.)
N.M. Katz:
Any maximal ideal p of R has a finite residue field Fq of characteristic p > 0. Reducing modulo p, we obtain from (M, V) a differential equation on a scheme which is smooth over a (finite) field of characteristic p > 0. In order for the differential equation (M/p M, V) in characteristic p to have a full set of solutions, in the sense that M / p M is spanned by its subsheaf of horizontal sections; it is necessary and sufficient that a certain p-linear homomorphism, the p-curvature of (M/p M, V), vanish. (Ibis) Suppose that for every maximal ideal p of R, the p-curvature of (M/p M, V) vanishes. Does the complex differential equation (M, V)c on the smooth C-scheme S c have a full set of algebraic solutions, in the sense that it becomes trivial on a finite etale covering of Sc? It is proved in ([24], Theorem 13.0) that the vanishing of the pcurvature for all maximal ideals p of R implies that the complex differential equation (M, V)c on Sc has only regular singular points, and that its local monodromy groups around the "branches at ~ " , in any smooth compactification Sc of Sc such that D = S c - Sc is a divisor with normal crossings in Sc, are all finite groups. In view of the fact that differential equations with regular singular points are determined by their global monodromy groups, (I bis) may be reformulated (lter) Suppose that for every maximal ideal P of R, the p-curvature of (M/p M, 17) vanishes. Does the differential equation (M, V)c on Sc have a finite global monodromy group ? An easy argument, via restricting to curves, projecting and specializing parameters in R shows that if the original question (I) always has an affirmative answer, then (I bis) and (I ter) always have affirmative answers. Unfortunately, (I) is far from being resolved. Consider the special case (M, V)= ((9s, V) of a rank one equation. The connection V is of the form 17(f)=df+fV(1); its integrability implies that the one-form co = V(1) is necessarily closed. This equation admits a solution mod p if and only if 09 is logarithmic mod p. It admits a solution on a finite 6tale covering of Sc if and only if an integral multiple n co, n > 1 of to is logarithmic on Sc [for if co = - d f / f with f an algebraic function of degree n, then putting g=Norm(1/f), we have nto=dg/g]. Thus a special case of (I bis), which is interesting even (or perhaps especially) when SIR is an open subset of a curve of genus g > 1 (c. f. (7.4.4) for g = 0) is (Ilog) If toeF(S, O Xs/R) is closed, and is logarithmic modulo p for every maximal ideal p of R, is an integral multiple n to, n > 1 of to logarithmic on Sc? In applications, it is sometimes more natural to reverse the point of view. Given a smooth connected C-scheme Sc, and a differential equation (M, V)c on Sc, we can always find an affine open set q/c c Sc, a finitely generated subring R c C, a smooth R-scheme ~ with geometrically
Algebraic Solutions of Differential Equations
connected fibres and complex fibre q/c, and a differential equation (M, 17) on ql/R which induces (M, 17)clq/c on q/c. Suppose there exists an affine open set C c q/such that, for any maximal ideal p of R, the p-curvature of (M/o MI V, 17)vanishes. If such a C exists for one set of choices (o//,R,(M, 17)), then such a C exists for any set of choices. It's existence is thus an intrinsic property of the germ of (M, 17)c at the generic point of Sc, which we call "having p-curvature zero for almost all p". Because "p-curvature zero" is a property which is local for the 6tale topology, it follows that if (M, 17)c becomes trivial on a finite 6tale covering of Sc, then it has p-curvature zero for almost all p in the above sense. Grothendieck's question is whether the converse is true: (Iquat)
If (M, 17)c on Sc has p-curvature zero for almost all p, does it become trivial on a finite ~tale covering of Sc?
Our main result is that (I bis) admits an affirmative answer when the differential equation involved is a Picard-Fuchs equation, or a suitable direct factor of one. Recall that if K/C is any function field, and U/K any smooth K-variety, the finite-dimension K-spaces of algebraic de Rham cohomology HER(U/K) are each endowed with a canonical integrable connection V,that of Gauss-Manin ("differentiation ofcohomology classes with respect to parameters"). The resulting differential equations (H~R(U/K), 17)are called the Picard-Fuchs equations. The suitable direct factors are the following. Suppose a finite group G acts as K-automorphisms of U. Then it acts on the de Rham cohomology HgR(U/K) in a horizontal way (i.e., it respects 17). For any irreducible C-representation Z of G and any automorphism a of C, let Z" denote the representation deduced from X by applying tr to its matrix coeff We say that X and X~ are Q-conjugate. Let Z~, ..., X, be the non-isomorphic irreducible representation of G which are Q-conjugate to x. Let P(X,)(HgR(U/K), V) denote the z,-isotypical component of(HgR(U/K), 17), i.e. the part of HER(U/K) which transforms by Xi, with its induced Gaussr
Manin connection. Then 0) P(Xi)(HgR(U/K), 17) is what we mean by a

i=l
suitable direct factor of (HER(U/K), 17). The proof is based upon the somewhat striking fact that in characteristic p, a suitable associated graded form of the p-curvature of the Gauss-Manin connection is a "twisted" form of the mapping "cupproduct with the Kodaira-Spencer class". The proof of this fact is unfortunately computational, and rather long. It depends essentially upon the identity of Hochschild, which asserts that for any derivation D of a commutative Fp-algebra, and for any element X of that algebra, we have DP-I(X p-t DX)+(D(X))P=XP-tDP(X). Indeed, in the case of an H ~, this identity is the proof.
1"
N . M . Katz:
The idea of looking at the associated graded mapping defined by the p-curvature was suggested by Clemens' attempt to generalize the PicardLefschetz formula. The proof of the theorem is concluded by a transcendental argument, based essentially on the Hodge Index Theorem. The argument in the non-proper case is due to Deligne, and depends upon his theory of mixed Hodge structures. Our motivation for extending the theorem to the non-proper case, and to the suitable direct factors defined by the action of a finite group, was Dwork's suggestion that our results should imply an affirmative answer to (Ibis) for the classical hypergeometric differential equation. Indeed, by studying the one-parameter families of curves Y"=xA(x-1)n(X-2) c, Y#:O cover C(2), we show that the hypergeometric equation with parameters a, b, e e C 2(1-2)
d:f -~+(c-(a+b+
1) 2) -~f-abf=O
df
has two algebraic solutions if and only if a, b, c e Q and, for almost all primes p, the reduced equation has two solutions in Fp(2) which are linearly independent over Fp (2). Consider the "general case", which in the theory of the hypergeometric equation means that none of the exponent differences i - c , c-a-b, a-b is an integer. As we learned from Dwork, the existence of two mod p solutions for almost all p is equivalent to the condition that 1, ~a=exp(2nia), ~b=exp(2nib), ~c=exp(2nic) be distinct roots of unity, and that for any automorphism tr of C, the counterclockwise arc on the unit circle leading from 1 to (~ contains either (~ or (~, but not both. The problem of when the hypergeometric equation has two algebraic solutions is of course a classical one, and was solved completely by Schwartz, whose solution (Schwartz's list) is stated with uncommon clarity by Goursat ([10], p. 40-41). It is by no means evident a priori that our solution is equivalent to Schwartz's. The classical theory of Reimann's p-scheme shows that any second order differential equation on an open set of Pd which has at most three singular points, all regular, is isomorphic to (the inverse image, by an automorphism of P~, of) a hypergeometric equation. Thus (I bis) has an affirmative answer for all such equations. We would like to point out that, by projection, the problem (I log) on a hyperelliptic curve of genus g__>1 is equivalent to the problem (I bis) for certain second-order equations on Pc having 3g + 2 singular 1 points. In the final section, we explain how (I log) on an elliptic curve over Q which has a nontrivial rational point of order two, is equivalent to an intriguing diophantine problem on the elliptic curve, which we view
as being (if true) an arithmetic analogue of Manin's" theorem of the Kernel" for abelian varieties over function fields ([29]). For this reason, problem (I bis) for second order differential equations on Pc with five ~ singular points is already of great interest, despite its mundane appearance. Let us briefly indicate the contents of the various sections. In the first section, we explain in intrinsic terms the well-known relation between the Gauss-Manin connection and the cup-product with the Kodaira-Spencer class ((1.3.2), (1.4.1.6), (1.4.1.7)). In the second section, we recall the Cartier operation, which completely analyses the local de Rham cohomology of a smooth morphism in characteristic p > 0 (cf. (2.1.1)). After an interlude of general nonsense on degeneration of spectral sequences after base change (2.2.1.11), we return to reality and study the "conjugate" spectral sequence of de Rhant cohomology in characteristic p > 0 (cf. (2.3.2)). We then explain the relation between this spectral sequence and the classical Hasse-Witt matrix (cf.(2.3.4.1)), introduce the sometimes-defined "higher" Hasse-Witt matrices (2.3.4.22) and extend to them the theorem of Igusa-Manin (2.3.6.3). This divertimento is concluded by a "numerical example" (2.3.7-8), where we calculate the "Hasse invariant" of a hypersurface of geometric genus one. In the third section, we prove the main technical result, relating the p-curvature and the Kodaira-Spencer class (3.2), (3.3). Before giving the proof, we recall some of the basic facts about the modular representation theory of finite groups of order prime to p (3.2.1-3). The fourth section reviews Deligne's theory of mixed Hodge structures. It is independent of the three preceeding sections, and takes place entirely over C. The principal results are (4.1.2), (4.3.3), (4.3.4), (4.3.5), (4.4.2). In the fifth section, we put together the results of the two preceeding sections, to prove our main theorems ((5.1), (5.3), (5.5), (5.7)) on PicardFuchs equations. The sixth section is devoted to the proof of (I bis) for the hypergeometric equation (6.2). The first two parts (6.0-1) give a useful elementary dictionary between "modules with integrable connections" and "n-th order differential equations". The rest of the section is devoted to the proof of (6.2). The essential proposition (6.8.6) is essentially contained in Messing [32]. The entire section can be read independently of the preceeding o n e s - t h e result (5.7) is only applied in a formal way. The final section is devoted to the problem (I log), especially on elliptic curves. The focal point is Conjecture (7.5.11), which emerged in conversations with Tate and Mazur. The appendix gives a simple (not elementary) proof of a useful case of Riemann's Existence Theorem.
N.M. Katz:
1. Generalities on the Kodaira-Spencer Class and the Gauss-Manin Connection 1.0. The Geometric Setting Throughout this Section 1, we will consider the situation
D ~ i ~X,
U=X-D
(1.0.1)
T in which T is an arbitrary base scheme, S is a smooth T-scheme (via g), which will play the role of a parameter space, X is a smooth S-scheme (via f), whose fibres over S are "parameterized" by S, and D is (via /) a union of divisors Di in X, each of which is smooth over S (hence also over T), and which have normal crossings relative to S (hence also relative to T). This situation persists after arbitrary change of base T' -~ T. In practice, X is usually proper over S, and should be thought of as a particularly nice compactification of the smooth "open" S-scheme U = X - D , which is psychologically prior to X. We allow D to be the empty divisor, corresponding to U being proper over S. We do not assume X proper over S except when we explicitly so state. (1.0.2) Let Dero(X/T ) (resp. Dero(X/S)) denote the locally free sheaf on X of germs of T-linear (resp. S-linear) derivations of (9x to (9x which preserve the ideal sheaf of each branch D i of D. We may now define the sheaf of germs of relative (to T, resp. to S) Kahler differentials on X with logarithmic singularities along D, by (1.0.2.1) 12Jc/r(log D) = Hom~x (Dero (X/T), Cx) ~ / s (log D) = Hom~,, (Oero (X/S), (gx).
For every integer p-> 0, we define (1.0.2.2) f2xP/r(log D)= Ae~xf2~/T(log D)
a~/s (log D) = A ~ alx/s(log D).

(1.0.3) In fact, O]/r(Iog D) (resp. 12~/s(log D)) is a subcomplex of j . 0~/r (resp. j, Q~/s), where j: Uc--*X denotes the inclusion. To fix ideas, let us explicate these sheaves in terms of local coordinates. We may cover S by
affine open sets q/i, and cover X by affine open sets V~such that each q / = q/i is 6tale over A t (r depending on i) via local coordinates sl, ..., s, (1.0.3.1) each V= Vi is 6tale over At, (n depending on i) via local coordinates Xl, ..., x, the branches of D which meet V~ are defined by the equation x, = 0, v = 1.... , c~ (a depending on i). Then, over, V the sheaf Dero(X/T) is a free (~v-module with basis (1.0.3.2) x~ - ~x~ (v = 1, ..., ~), ~-~f (j = ct + 1, .... n), ~ a and Dero (X/S) is a free (~v-module with basis (1.0.3.3) x~-ax, (v = 1, ct), (j=0~+ 1. . . . . n). ( # = l , .. . , r)
""'
0xj
Thus, over V, f2~/T(1OgD) is free on (~v with basis (1.0.3.4)
dx~ ( v = l , . . . , ~ ) ,
Xv
dxj ( j = ~ + l, ..., r),
ds, ( # = l , ..., r)
while f21/s(log D) is free on (_9 with basis v (1.0.3.5)

dxv Xv
(v= 1, ..., c0,
dxj
( j = ~ + 1, ..., n).
(1.0.3.6) Clearly the formation of f2]/s (log D) commutes with arbitrary change of base S' -~ S. (1.0.3.7) Remark. (When S is of characteristic zero, the complex I2]/s(log D) is quasi-isomorphic to j. t2~:/s. This is not the case in general, because j . f2g./s has "too much" cohomology, while I2]/s(log D) has only "geometrically meaningful" cohomology, whence our "preference" for t2]/s (log O).)
1.1. The Kodaira-Spencer Class

From the definitions, it follows that we have an exact sequence of locally free sheaves on X (1.1.1) 0 - ~ f * (f2~/r) --~ t2~/r (log D) -~ f2~c/s(log D) -~ 0
N.M. Katz:
which gives rise to an element (1.1.2) p e Ext,,, (f21x/s(log D), f * (t2~/r))
~- H t (X, Dero (X/S) @r f * (I2~/r)).

This element is called the Kodaira-Spencer class of the situation 1.0. It's image in
U ~(S, R ' f , (Dero (X/S) |

(1.1.3)
f * (f2~/r)))
(2~/r)
~--U ~(S, n I f. (Dero (X/S)) |
Home~(Der(S/T)), R~f. (Dero(X/S))

is the Kodaira-Spencer mapping (but still denoted p !), which may be explicated as follows. Let D be a section of Der(S/T) over an affine open set ~ c S. Then p (D) is an element of n I ( f - 1 (q[), Dero ( f - 1(q/)/q/)) which may be given explicitly as follows. Let {Vii} be an affine open cover of f - l ( q / ) . By the exactness of the dual of (1.1.1) over each Vii (or, more "directly", by the explicit description (1.0.3) via local coordinates), we may choose, for each i, a derivation Di~H~ Dero(X/T)) which extends the given derivation D ~H ~(ql, Der(S/T)). Because D i and D~ extend the same derivation D of q/, the difference D i - Dj lies in H ~(Viin Vj, Dero(X/S)). Thus { D i - D j} defines a 1-cocycle on the covering { V~} with coefficients in Dero(f-l(ql)/ql), whose cohomology class is p(D). In particular, p (D)= 0 if and only if there exists a derivation in H ~(f-l(ql), Dero (X/T)) which extends D.
1.2. The Koszui Filtration, Interior Product, and Cup Product (1.2.0) This section is devoted to recording some compatibilities among wellknown constructions from exterior algebra as they occur in homological algebra. For ease of later reference, we record a few more compatibilities than necessary for our application in Section 3.
(1.2.1) Let X be an arbitrary scheme (for even a ringed topos, if that be more to the taste of the reader), and let (1.2.1.1) 0--*fq ~ ,~ff P~Y-~,O
be an exact sequence of locally free Ox-modules of finite rank. (In the application, it will be the exact sequence (1.1.1).) The Koszul filtration of the exterior algebra A~,, ~r which we will henceforth denote simply A~~ is defined by the ideals (1.2.1.2)
Ki(A~
of A i ~ | 1 7 6
gff - , A ~
Algebraic Solutions of Differential Equations The associated graded module is given by (1.2.1.3) grj~ (A~174
Consider the exact sequence (1.2.1.4) 0 -~ gr~ -~ K~
2 -, gr ~ -~ 0 ~ff)-~ W ~-~O.
whose term of degree v is (1.2.1.5)
O-~ f~|
~ - , K~
(1.2.1.6) For each integer v > 1, we define in this way a functor W from the category E X T ( ~ fq) of extensions of ~ by f~ to the category EXT (A~~, f#| ~-a ~ )
of extensions of W o~ by fr | W-a ~ Passing to the (groups of) isomorphism classes of objects of these categories, we obtain a morphism, still denoted A ~, (1.2.1.7)
A~: Ext~x(~, fr
Ext~,,(AV~,c.9|
~).
Because ~ is locally free, the sheaf Ext~x(~, ~9) vanishes, and the local =~ global spectral sequence of Ext furnishes us with an isomorphism (1.2.1.8) Ext~s (~, fq) ~ H a(X, Homr
f9)). ~| ~)).
Similarly, the local freeness of A ~~ furnishes an isomorphism (1.2.1.8bis) Ext~,,(A~, (r174 ~-~ ~),,~HX(H, Homr
Let us make explicit the two identifications (1.2.1.8) and (1.2.1.8 bis), then use them to make explicit the morphism A v (1.2.1.7). (1.2.1.9) Since ~ is locally free of finite rank, the exactness of (1.2.1.1) assures that there is an open covering V~of X and, over each V~, a morphism tpi: ~ l V/-~ Jegl Vi which is a section of/3: ~,ug_~~ (i.e. such that /30 ~oi= id~lv, ). The difference tpi -~pj defines a morphism from ~1 V/n Vj to Yg[ V~n Vj whose image is in fact contained in Ker(/3)l V/r~ V~(because on Vi~ ~,/3 o (tpi - ~0~)=/3 o tp~-/3 o ~0~= id~ - id~ = 0). Thus we may interpret tpi-%, as defining a morphism from ~'IV/c~V~ to fr (~Ker(/3)lV/c~ V~), whence {q~-~o~} is a one-cocycle for the covering {V/} with coefficients in Homex(~,,ff ). The cohomology class in Ha(X, Hom~,,(~,,f~)) of this cocycle is the element corresponding to the extension class of (1.2.11) via the isomorphism 1.2 (1.2.1.10) (1.2.1.10.1) Let us recall that there is another "standard" isomorphism
Ext,: (~, (~)r Ha(X, Horn(,~, ~))
10
N.M. Katz:
which is the negative of the isomorphism (1.2.1.8) explicated in (1.2.1.9), and which we will not use. In the setting of (1.2.1.9), we may choose over each V~ a morphism ~ : ~ 1 V i ~ Ker([3)[ V/~-fg[ Vi which is a section of ~: fq~--~ouf(i.e., such that ~ki. ~ = id,lv, ). (1.2.1.10.2) In fact, we may simply choose ~biso that ~ o ~bi= id~rlv,- q~i~[3, this being possible because, over V~, id~e-Cp~o[3 is a projection onto ker([3). Then the difference ~bi- ~b~defines a morphism from ~r V/~ V~to f~l V~n ~ which vanishes on the image of ct, thus defining by passage to quotients a morphism from ~1 V/c~ V~to fq[ V/c~ V~, still denoted ~ i - ~b~. The cohomology class of {~k~-~b~} in Ha(X, Hom(~, if)) is the element corresponding to the extension class of (1.2.1.1) via the isomorphism (1.2.1.10.1). To see that this isomorphism is the negative of (1.2.1.9), it suffices to recall that (1.2.1.10.3) whence (1.2.1.10.4) 0~o ~i = id~elv, - ~o~o[3 ~ o (~k,- ~s) = - (~o,- ~os)o [3
which gives the equality of the two cocycles {~bi - ~bs} and { - (q~-cp~)}. (1.2.1.11) The "second" isomorphism (1.2.1.10) is the dual of the first (1.2.1.8), in the sense that the diagram Ext a]~, c~)
Ext a (c~, j,~) ~a.2~.~o~, H a(X, nom(~, ~))

in which the left vertical map associates to the class of0-~ fg-, ~ r 0 the class of the dual extension 0 ~ ~ - ~ ~ fq--* 0 and in which the right vertical arrow is deduced by passage to cohomology from the canonical isomorphism of duality Horn (~, f~)~ Horn (f~, ~), is commutative. To see this let us remark simply that if {cp~- cpj} is a cocycle representing the class of 0 ~ f f ~ ~ - - ~ 0 via (1.2.1.8), which is to say, via splitting ~ ~,, then {~bi-c~i } is a cocycle arising from the class of 0 - ~ ~ - ~ 0 by splitting ~ - ~ , which is to say, via (1.2.1.10). Keaping in mind that the isomorphisms (1.2.1.8) and (1.2.1.10) are the negative of each other, we find: (1.2.1.12) Proposition. Let ~eHl(X, Hom(~,ff)) be the class of the extension O--*ff---~--*~--~ 0 via (1.2.1.8), and let ~e HI(X, Horn(if, o~)) be the class of the dual extension O---~-~J/~-~f~--~O, via (1.2.1.8). Then via the canonical isomorphism Hom(~, i f ) ~ Hom((~, ~), ~= -~.
Algebraic Solutionsof DifferentialEquations (1.2.2) Interior Product and Cup Product.
11
Let us now recall the morphism I of interior product for each integer v=>l. (1.2.2.1)
I: Hom~x(~, cg) _, Hom~,,(A~, C~|
~).
In terms of local sections fl, ... ,fv of ~ and q~ of Hom(o~, ~),

v
(1.2.2.2) I((p)(fx A... A f~)= ~ ( - 1)j-1 q~(fj)Qfx ^ " " ^ fj ^ " " ^ f~. j=l (1.2.2.3) Proposition. The diagram Extl(.~,,ff) a v E x f l ( A ~ , f f |
14~(x, Horn (~, ~)) -~-' , ~1 (X, Horn (A" ~, .~ | A v-1 ~))
in which the vertical isomorphisms are (1.2.1.8) and (1.2.1.8 bis), is commutative. Proof. Just as in (1.2.1.9), let us choose a covering V~of X and sections qh: ~1V~-, o~1V~ of ~: ~ - , o ~ Abusing notation, we denote by A v qh the composition
(1.2.2,3.1)
A~IV~-av(~~
A~IV~
Pr~ , K~
(i.e., A v ~oia~" AV(q~i)modulo KE). Clearly the A ~~o give local sections of i A~: K~ and hence 1-cocycle {AVqgi-A~oj} with coefficients in H o m ~ x ( A ~ , f # | has as cohomology class in Hi(X, Hom~x(A~~, C~| ~ ) ) ~ Ext ~(AVo fr174 ~-1 ~ ) the class of ~, A~(0-~fr To conclude the proof, it remains only to notice that (1.2.2.3.2) A ~cpg- A ~cp~= I(~0~- tp~).
To see this, we calculate, for local sections fl ..... f~ of (A ~q9i - A vq~j)(f~ ^ . . . ^ f~)= q0i(fx) ^ . . . ^ q0i(f~) - ~oj(f~) ^ . . - ^ ~o~(f~)
= (q~ ( f l) + (~oi- ~o~)(f,)) ^... ^ (q~i(f~) + (~o,- q~i)(f~))

- ~o~(f~) ^ . . . ^ ~oAL)
= i (-- 1)a-l((Pi--~OJ)(fa)|
a=l
A "'" ^ l---^ ... ^ f , f
+ terms in K z (A ~~ ) =I(~oi-q~)(f t ^ . . . ^f~). Q.E.D.
12
N.M. Katz:
(1.2.2.4) Let us reinterpret the interior product mapping I as a pairing (1.2.2.4.1) I: Homex(O~, ~ ) | 1 7 4
From this pairing we deduce a cup-product pairing for all p, q > 0 (1.2.2,4.2) HP(X, Hom(~, ~))| H~(X, A~ ~ ) - , HP+q(X, ~ ) | A~- ' ~ ) . (1.2.2.4.3) Taking p = l , consider the element ~eHI(X, Hom(~,~)) which corresponds via (1.2.1.8) to the class of the extension (1.2.1.2), (1.2.2.4.4) 0 - , ~r -~ our-+ . ~ - , 0.
(1.2.2.4.5) Proposition. The mapping "cup-product with ~"

~: Hq(X, A ~ ) - , Hq+X(X, ~ @ A ~-1 ~ )
deduced from (1.2.2.4.2) is equal to the coboundary in the long exact cohomology sequence arising from the short exact sequence (1.2.1.5)
(1.2.2.4.6)
0-~ ~ | A ~- 1~ _ ~ KO/K z (A ~~,uf)_~ A ~~ _ ~ O.
Proof Indeed by definition this map is none other than the cupproduct with the image of ~ ~ H 1(X, H o m ( ~ ; if)) under I: Hi(X, Hom(~, ~)) -~ Hi(X, H o m ( a ~ , , ~ | ~-~ ~)).
According to (1.2.2.3), the image 1(0 is none other than the class of the extension (1.2.2.4.6) (via the isomorphism (1.2.1.8 bis)). (1.2.2.4.7) But it is a general fact that the coboundary mapping H~(X, C)-, H q+l (X, A) associated to any short exact sequence 0-~ A - , B-+ C--) 0 of @x-modules is none other than the cup-product with the element of Ext~x(C, A) (this last operation has a sense, thanks to the isomorphism of functors H q(X, - ) ~ E x t ~ , ((gx, -)). Q.E.D.
1.3. Application to Hodge Cohomology

(1.3.1) Proposition. In the geometric situation of 1.0, consider the short exact sequence (1.1.1) (1.3.1.1)
0 - ) f * (O~/T) ---)O~C/T (Iog O) --, O~C/S D) - , 0 (log
which gives rise, via the A p construction (1.2.1.6), to a short exact sequence
(1.3.1.2)
0 - , f * ((2~/r) | (2w (log D) --~ K ~ z (Y2w (log D)) --+ O]/s (log D) - , 0.
Algebraic Solutions of DifferentialEquations
13
The coboundary mapping associated to (1.3.1.2)

(1.3.1.3) ~: Hq(X, f2~/s(log D))-~ Hq+'(X, f*(f2~s/r)|
D))
is the cup-product with the Kodaira-Spencer class (1.1.2)

p ~ Ext,,, (f21XlS(log D), f* (f21Slr))
n ' ( x , Hom(f2'x/s(log D), f*(f2's/r) ) ~ nl(x,f*(f2's/r)| Dero(X/S)). Proof. This is just (1.2.2.4.5) applied to the exact sequence (1.3.1.1), in which case the element ~ of (1.2.2.4.5) is the Kodaira-Spencer class. Localizing on S, we have (1.3.2) Corollary. Hypotheses as in (1.3.1), the coboundary mapping
(3 Rq f , (aVx/s(log O)) - , R q+' f , (f* (fJ~S/T)| O~-fS (log O)) 1
(1.3.2.1)
/
fJ's/T| R" +1f , (f2]-[s,(log D))
associated to (1.3.1.1) is given by cup-product with the Kodaira-Spencer mapping (1.1.3), viewed as an element
p ~ H o m ~ (Der (S/T)), R' f , (Dero (X/S)) (1.3.2.2) t
H ~(S, Y2~S/T) R~ f , (Dero (X/S)). |

1.4. Application to the Gauss-Manin Connection
(1.4.0)
The construction of the Gauss-Manin connection on
HoR (X/S (log O)) = R f , (f2x/s (log O))

is based on the fact that the Koszul filtration of f2~/r(1Og D) arising from the exact sequence (1,1.1) (1.4.0.1)
0 - , f * (f2~S/T)--, fJ~/T (log D) - , f2~/s(1Og D) - , 0
is a filtration by subcomplexes, and that the associated graded complexes are given by (1.4.0.2)
gr~(f2]/T(lOgD))'~f * ~2~/T| 12jc~i(log D).
14
N.M. Katz:
In fact, it is defined ([35]) as the coboundary map 17 in the long exact sequence of the Rif, and the short exact sequence (1.4.0.3) O-~f* (f2s/r) |/ s l ,-l(logD)_~KO/K~-(f2~/~(logD)) f2x
-~ t2]/s (log D) --~ 0.
This "makes sense", because the coboundary is a mapping
17: R~ f , ((2~c/s(log O)) ~ R q+1 L (f* (Q~S/r)| 12]'~S (log D)) 1

(1.4.0.4) /l
~2~/r| Rq f , (t2]/s (log D)).

(The final isomorphism thanks to the local freeness of ~2~/r.) (1.4.1) The Hodge Filtration. Let us recall that for any complex L~ the Hodge filtration of L~ is the filtration by the subcomplexes Fi(L'), where by definition 0 if j--i (1.4.1.1) U(/J)= /j if j>=i. (1.4.1.2) Let us denote by L~[n] the complex L~+" whose term of degree i is/J+", whence (1.4.1.3)
F i (L" In]) = (F i +" (L')) [n].
Applying F ~to the exact sequence (1.4.0.3), we obtain the exact sequence (1.4.1.4)
0 -~ (f* (12~/r) | F i-1 (12],/s(log D))) [ - 1] -~ F~(K~ (Q]/r (log D))) V i (f2~c/s(log D))-~ 0.
Thus the coboundary maps for the W f , and the exact sequence (1.4.0.3) and (1.4.1.4) "fit together" to form a commutative diagram Rqf,(O]ls(log D))--E-~ 12Xs/r| v (1.4.1.5) (2~/s(log D))
|
Rqf, (F *(O]/s (log D))) _ ~ 0 0 ] / r | Rqf, (U- 1(O]/s(log D))).
Thus we find: (1.4.1.6) PrOlmSition (Griffith's Transversality Theorem). The GaussManin connection respects the Hodge filtration up to a shift of one, i.e. (1.4.1.6.1)
17(FiR~f,(g2]/s(log O))) =f2xs/r|
D))).
15
As a corollary of (1.3.2) and the definition of the Gauss-Manin connection, we have (1.4.1.7) Proposition. Suppose that the Hodge=~De Rham spectral sequence (1.4.1.7.1)
EP'~=Rqf,(f2~ls(log D)) =~ RP +qf , (I2~/s(log D))
is degenerate at El, i.e., that gr~ R p + qf , (0~1s (log D)) = Rq f , (O~/s). Then the associated graded mapping induced by the Gauss-Manin connection is the cup-product with the Kodaira-Spencer mapping (1.3.2.2) p ~ H ~(S, a~lT | Ri f , (Oero (X/S))); i.e., the diagram
gr p RP+qf, (Q~c/~(log D)) (1.4.1.7.2) gqf, (f2P/s(log D))" > | +%
v > f2~i T | grF - ' RP +q f , (f2~cis(log D)) p
7
D))
commutes.
(1.4.1.8) Remark. In case X/S is proper, it follows from Deligne's mixed Hodge theory [8] and a slight modification of his argument ([6], Theorem 5.5) that (1.4.1.8.1) If S is any scheme of characteristic zero, the spectral sequence (1.4.1.7.1) is degenerate at El, all of its terms E( 'q, E~;~ are locally free, and its formation commutes with arbitrary change of base S' -~ S. (1.4.1.8.2) If S is any reduced and irreducible scheme whose generic point is of characteristic zero, there exists a non-void Zariski open set q/in S over which the assertions of (1.4.1.8.1) are valid. We conclude this section by stating explicitly a very useful corollary of (1.4.1.7). (1.4.1.9) Corollary. Hypotheses as in (1.4.1.7), fix an integer n>=O, and suppose that M~=R"f,(O~cls(log D)) is an Os-submodule stable under the Gauss-Manin connection, i.e., that (1.4.1.9.1)
V(M) c 0~1T| M.
(We then say that M is horizontal.) Let us define the induced Hodge filtration of M, Fi(M), by
(1.4.1.9.2) F' (M)= M c~ F i R"f, (f2~ls(log D)).
16
N.M. Katz:
Then
gr~ (M) = F p (M)/F p +a(M) ~ gr~ R p +qf, (f2]/s (log D))
(1.4.1.9.3)
l[
Rq f, (~2P/s(log O)).
Because M is horizontal, it follows from (1.4.1.7) that we have a commutative diagram
a f, (C ,,s(log D))-
(O ,/s(log O))
FV(M)
v ~/r|
(1.4.1.9.4) grV(M) v __~f2~/r | grP- 1(M)
R' f , ( ~ / s ( l o g D))
1]
P , ~/r|
Rq+~f, (~xV})~(logD))
We deduce from it that the Hodge filtration Fi(M) of M is horizontal (i. e., each Fi(M) is horizontal) if and only if the restriction to Op gr~(M) of the mapping "cup-product with the Kodaira-Spencer class"
(1.4.1.9.5)
vanishes.
p: @ gr~(M)-, @ f2s~/r|
p p
-I(M)
(1.4.1.10) Remark. In practice, the M in (1.4.1.9) will be either all of R"f,(f2~/s(log D)), or the primitive part of R"f,(f2]/s) in case O is void and X / S is projective and smooth, or the part of R"f, (f2~c/s(logD)) which transforms according to a prechosen irreducible representation of a finite group (of order prime to all residue characteristics of S) which acts as a group of S-automorphisms of X and preserves D. This last case will arise when we discuss Schwartz's list (cf. 6.0).
2. The Cartier Operation and the Conjugate Spectral Sequence

2.0. Throughout this section, we will consider the situation of 1.0, with the additional assumption that S is a scheme of characteristic p (a prime number), i.e. that p. ls = 0 in ~Ys. (2.0.1) Recall that for any S-scheme n: Y---, S, the S-scheme y~v) is by definition the fibre product of 7~: Y---,S and the absolute Frobenius
17
morphism Fab~: S - ~ S (so on the ring level, Fabs is just "raising to the p-th power"). Thus by construction YtP~ sits in a cartesian diagram
(2.0.1.1)
S The pair of morphisms
~(P)= Fa*bs(~)
[~
r~b~
~S
Fabs : y----~ y
7z: Y---~ S
defines a morphism F: Y--~ Y~P~, the relative Frobenius, which fits into a commutative diagram
y f ~ y~p~
~(P)
~ ~y
~ y(p)
in which F . ~ is F,bs: Y(P) ~ Y(P) and a- F is F~bs: Y---~Y. Intuitively, F raises the "vertical coordinates" to the p-th power, and e raises the "S coordinates" to the p-th power. (2.0.2) Consider now the special case Y = X is a smooth S-scheme, say of relative dimension n. Then: (2.0.2.1) F,(Cx) is a locally free (gxc,~ module of rank pn; indeed if xl .... , xn are local coordinates on an open set V c X (i.e., an 6tale morphism V ~ A~), then a base of F, (Cv) as (fv~p~module is given by the p~ monomials x~'~.., x~" having O < w i ~ p - 1. (2.0.2.2) F-l((gxt,,) is precisely the subsheaf of (9x which is'killed by all of Der (X/S). (2.0.2.3) F,(f2~./s(logD)) is an Cx~Minear complex of locally free coherent sheaves on X ~p~.
2.1. The Cartier Operation

The fundamental fact about De Rham cohomology in characteristic p is the following theorem of Cartier whose statement we recall (cf. [3], chpt. 2, and [24], 7.2). (2.1.1) Theorem. There is a unique isomorphism of (PxC,~ modules for each
integer i > 0
2 Inventiones math.,Vol.18
18
N.M. Katz:
(2.1.1.1)
cg-1 : f2jc,~,/s(log Dtp)) ~, ~t~i (F, (Q~/s (log O)))
which satisfies
(2.1.1.2) (2.1.1.3) (2.1.1.4)
.-1(1)=1,
(~-- 1(0) A ~') = ( ~ - a (0)) A ~ - 1 (T),
c~-, (d(a- 1(x))) = the class of x p- 1 dx.
(2.1.2) Putting together all the conditions, we see that, in terms of the local coordinates xl, ..., x. chosen in (1.0.3.1), we have cg- 1(h) = the class of F - 1(h) for h a local sections of Cx~p~ (2.1.2.1)
~e_1 ( d~-l(xv)]] a-X(xv)

\
= t h e class of x~ ,
dx~
for v = 1, ..., a for j = ~ + l .... , n.
c~-l{da-l(xi))=the class o f x ~ - ' d x j

2.2. General Nonsense
We must now give a tautology on functorial spectral sequences, whose precise formulation is somewhat lengthy. (2.2.1) Suppose given f : Y-~ Z an arbitrary morphism of schemes. For every Z-scheme ZI, we form the fibre product Yz,= Yxz Z~, which sits in the cartesian diagram
Yzl ~Y
Z, -
--~ Z.
Suppose we are give, for every Z-scheme Za, a finitely filtered complex (non-zero only in positive degrees). (Kz,, F) of fz~l((.Oz,) modules on Yz,, which is functorial in the variable Z-scheme Z1 in the following sense: (2.2.1.2) For any morphism ~ 0 : Z 2 ~ Z 1 of Z-schemes, denote by ~0r: Yz, ~ Yz, the induced morphism, which sits in the commutative diagram
(2.2.1.3)
Y -
Z2
e , Za
19
We are to be given a morphism K(q~) of filtered complexes of f~l(t~z)modules on Yz~ (2.2.1.4)
q~l(Kz~, F) ~ ~
~" fz~ (r
fz~X((-0z~) ~(~)-~(Kz~, F)
which satisfies the natural transitivity condition for a composition of morphisms of Z-schemes. Consider now the spectral sequence of d~z~-modules (2.2.1.5)
E~' "(Z~)=RP+qfz,.(gr~(Kz,) ) ~ RP+"fz,.(Kz),
on the Z-scheme Z1, whose E~'q term we denote E~'q(ZO. From the given functoriality of (Kz~, F) in ZI, we deduce, for every morphism q~: Z2--, Z1 of Z-schemes, morphisms of (gz~-modules called "change of base morphisms ", (2.2.1.6) ~0"(E~' ~(ZI)) ---~) ~Er (Z2) p'"
* (E p, q (zo) ,,
K(~o)
which render commutative all diagrams
Ep.q~Z 2t r ~,
(2.2.1.7)
q,*~d~'q)l
K(o).~ Ep + r, q + 1 - , ( Z 2 ) .
da, q
These morphisms are compatible with the usual isomorphism of E,+t with the cohomology of (E, d,), in the sense that the diagram ~o*(Ker d~'q in EP'q(ZO)
!
-K3~')
icalprolection)
, Ker d~' q in E~' q(Z2)

canonical projection
(2.2.1.8)
~,*~. . . . .
4,
commutes. Further, the induced mapping on Eoo is the associated graded of the change of base morphism deduced from (2.2.1.4): (2.2.1.9)
(P* RPfzl ,(Kz,)
r~) , Rp fz2,(Kz2).
For each integer ro~ 1, we say that the formation of E,o commutes with base change if for every Z-scheme q~: Z t ~ Z, and all pairs (p, q) of integers, the morphism (2.2.1.6) (2.2.1.10)
~o*(EP,oq(Z1)) -r~,)~ EPgq(Zz)
is an isomorphism. We say that the formation of the spectral sequence from E,o on commutes with base change if for all r > to, the formation of E, commutes with base change.
2*
20
N.M. Katz:
(2.2.1.11) Tautology. Hypotheses as above, suppose that for an integer ro> 1, the formation of Ero commutes with base change, and that the spectral sequence over Z (2.2.1.12) g f 'q (Z) = R p +qf , (grF (Kz)) =~ R" + qf , (Kz) p
is degenerate at Ero. Then after any arbitrary change of base ~o: Z1 -+ Z, the spectral sequence
(2.2.1.13)
_ Elp , q ( Z 1 )_- R P+q
fz,,(gr~(Kz,)) ~RP+qfzl,(Kzx )
P
is degenerate at Ero. Furthermore its formation from Ero on commutes with arbitrary change of base.
(2.2.2) Examples. Let's return to the geometric situation 1.0, so that our morphism f : Y ~ Z in (2.2.1) becomes the morphism f : X ~ S of 1.0. We take for K the complex O]/s (log D)), and for an arbitrary S-scheme S', we take Ks, = (2},/s, (log D'), the symbol ' denoting fibre product with S' over S. There are two filtrations in which we shall be interested. The first is the Hodge filtration (1.4.1), (the "bestial" one in the terminology of [8]). The second is the one, noted Z<o and called "canonical" in [8], which is defined by K" if n<p (2.2.2.1)
Z<p(K)= Ker(d)
= tO
if n=p if n>p.
The spectral sequence defined by this filtration (2.2.2.2)
~<=.Ef'"=RP+qf.(gr~.K) ~RP+qf,(K)
is the d~calage (cf. [8], 1.3.3) of the "second spectral sequence of hypercohomology" (2.2.2.3)
,,E~'q = RPf . (~ffq (K)) =>R p +qf . (K),
which by definition means that we have isomorphisms, compatible with the dr and with the standard isomorphism Er+, ~-H(E, dr), (2.2.2.4)
~ l~q ~< ,E~' q - , lla.~r+ 21, - - p +p
for each integer r > 1 (cf. [8]).
2.3. The Conjugate Spectral Sequence

(2.3.0) Recall that the entire "second spectral sequence" of hypercohomology (2.3.0.1)
E~'q=RPf, (~q(f21,/s(log D))) =~ RP+qf,(t2]/s(log D))
21
is endowed with the Gauss-Manin connection (cf. [24], 3.5). For reasons which will soon become apparent (cf. (2.3.3)below), we call this the conjugate spectral sequence, and the corresponding filtration of R"f,(f2~/s(log D)) the conjugate filtration, noted F~'on. This filtration is i horizontal, and is every bit as interesting as the Hodge filtration. (2.3.1) Let us explicate the Ez term of the conjugate spectral sequence with the aid of Cartier's isomorphism (2.1.1). (2.3.1.1) Lemma. The Cartier isomorphism ~ defines an isomorphism of (gs-modules for each pair a, b of non-negative integers
(2.3.1.1.1)
9 co,Ea,~_ R ~f , (Jr~b(f2x/s(log D))) 2 -
, R"f~,P)(tr*(Qbxls(logD))).
Proof. From the commutativity of the diagram

Xr ~ X~p)
S and the fact that F is a homeomorphism, we have isomorphisms (2.3.1.1.2) ~o,E3'b = Raf, (~'r~b(t2~,./s(log D))) ~- RaJ~p)(F, ~ b (f2.X/s(log D))) and (2.3.1.1.3) F, ,Yfb(f2]/S (log O))~--g/fb(F,(f2"x/s(lOgD))) which combine to give an isomorphism (2.3.1.1.4)
co,E~2'b~R~ft, P)(9~b(F,(I2~ls(logD)))). Jt~ D)))- ~, f2~,p,/s(log D ~p))
Composing (2.3.1.1.4) with Cartier's isomorphism (2.3.1.1.5)
and the inverse of the canonical isomorphism (2.3.1.t.6) (2.3.1.2)
a*(~2bx/s(log ~ , f~,p,/s(log D ~p)) D))

Q.E.D.
gives the asserted isomorphism 2.3.1.1.1.
Corollary. Suppose that either the absolute Frobenius F~b~:S--* S is fiat (which is the case if S is a regular scheme) or that for all pairs a, b of non-negative integers, the Hodge cohomology sheaves ROf . (Ob/s(1Og O)) are fiat Cgs-modules (for instance, if they are locally free (gs-modules). Then via (2.2.1.1) and the inverse of the base-changing isomorphism, we have an isomorphism. (2.3.1.2.1) r
a,b - ~ ' ~
R"f~,P)(a*(f2]./s(log O)))
, F~*s R"f,(f2bx/s(log D)).
22
N.M. Katz:
Proof The only point is that under either of the hypotheses, the canonical morphism of base change
(2.3.1.2.2)
F* s R"f, (bX/S (log D)) --~ R"f~,") (a* (~bx/s(log D)))
which comes from the cartesian diagram (2.0.1.1) X <p)- ~ >X
(2.3.1.2.3)
is an isomorphism.
lf(P)
lf
S Fab~'S
(2.3.1.3) Remark. Under the isomorphism (2.3.1.2.1), the Gauss-Manin connection deduced on F*s R"f, (~2bX/S (log D)) annihilates the image under F*s of R"f,(~x/s(log D)) (compare [24], 5.1.1). (2.3.2) Proposition. In the geometric situation 1.0, suppose that X is proper over S (and that S is a scheme of characteristic p, as it has been throughout Section 2). Suppose further that (2.3.2.1) Each of the Hodge cohomology sheaves R"f,(~x/s(logO)) is a locally free sheaf of finite rank on S and (hence) that its formation commutes with arbitrary change of base S'-* S. (2.3.2.2) The Hodge =~De Rham spectral sequence
E~'b = Rb f , (~x/s (log D)) ~ R a+bf , (I?~c/S(log D)) is degenerate at E 1. Then the conjugate spectral sequence, which, thanks to (2.3.1.2.1) and the hypothesis (2.3.2.1), may be written
(2.3.2.3) r 'b = F~*sR"f, (~x/s (log D)) ~ R" +bf , (g2~/s(log D)),
is degenerate at E 2 . Proof By (2.3.1.4.1), it follows that the conjunction of the hypotheses (2.3.2.1) and (2.3.2.2) remains true after an arbitrary change of base S ' ~ S, and implies that the formation of the Hodge ~ D e Rham spectral sequence commutes with arbitrary change of base S'--, S. From (2.2.1.2) it follows also that the formation of the conE~'b commutes with arbitrary change of base, while by general principles the formation of the entire conjugate spectral sequence commutes with any fiat base change S'-+ S. (2.3.2.4) We may assume that S is affine, because the question is local on S. We wish to reduce to the case in which S is noetherian. So suppose S=Spec(A). Clearly there exists a subring A o c A which is finitely generated over Z, a proper and smooth Ao-scheme Xo, and smooth
23
(over Ao) divisors Di, o in X o which cross normally relative to So = Spec(Ao), such that the geometric situation (1.0) over S (for the purposes of 2.3.2, the base scheme T figuring in (1.0) may be taken to be S itself) comes from the analogous situation over So by the change of base S--, So (cf. E G A IV, 8.9.1, 8.10.5, and 7.7.9). We must show that after replacing Ao by a larger subring At, A = A t ~ Ao, which is still finitely generated over Z, the hypotheses (2.3.2.1.2), that the Hodge ==De Rham spectral sequence degenerate at E1 and have Et locally free of finite rank, are valid over At, which is a noetherian ring. (For then, once the theorem is proved over At, it remains true over A by (2.2.1.11).) In fact, it suffices to find such an A1 over which (2.3.2.1) holds; then (2.3.2.2) follows. For if we suppose that over At the E~ terms noted E'~'b(AO, are locally free of finite rank, the differential (2.3.2.4.1)
d t 9 E~'b(A~) ~ E~ +L b(A~)
must vanish, because after extension of scalars to A ~ Ax~ this differential becomes zero, because of the commutative diagram (cf. (2.2.1.7))
E~,b(A1)QA~A dl _~E~+I,b(A1)QAIA
E~.b(A)
dl , E~+t,b(A)
But clearly if a homomorphism between locally free modules becomes zero after extending scalars by an inclusion A~ c A, the homomorphism is zero. Thus dl = over A1, and hence E 2 - El, whence E2 is locally free of finite rank, and it formation commutes with all change of base. Repeating the above argument, we get d 2 = 0 over At, and inductively d r = 0 for all r ~ 1. So now let's prove that we can find an A~ over which all the E t terms are locally free of finite rank. We'll prove this by descending induction on the integer b of E~' b. Because E~' b(Ao)= H b(Xo, f~Xo/So (log D)) vanishes unless a and b are in the interval [0, N], N being the relative dimension of Xo/Ao it suffices to find, for each integer O < a < N , an A1 which "works" for all the Eal'b with fixed a. By the standard base-changing theorems (cf. [34], pp. 46-55) we know that, for fixed a, (2.3.2.4.3) if, for all b > b o , the module E~'b(Ao) is locally free of finite rank on A o, then the formation of E~' bo commutes with arbitrary change of base. Since E~' b(Ao) vanishes for b > N, our descending induction will work if we can show that, for (a, b) fixed: (2.3.2.4.4) if the formation of E~'b commutes with arbitrary change of base, then the fact that E~' b (A)= E~' b(Ao)| A is a locally free A-module
24
N.M. Katz:
of finite rank implies that for a suitable subring AI of A, A o c A 1 cA, which is finitely generated over Z, the Al-module E~'b(AO is locally free of finite rank, or what is the same, flat, since in any case it of finite type over the noetherian ring A1. The truth of this last assertion is a particularly simple case of (EGA IV, 11.2.6.1), in the notation of which B o = A o , Mo=E~'b(Ao), and the A~ are all the absolutely finitely generated subrings of A with A o c A~ ~ A. (2.3.2.5) Having reduced to the case in which S is the spectrum of a noetherian ring, we may further assume that S is the spectrum of a noetherian local ring (again because the question is local on S). We may next suppose S to be the spectrum of a complete noetherian local ring (by faithful flatness of the completion), and finally that S is the spectrum of an artinian local ring. Let us first explain this last reduction step. Suppose S = Spec(A), A a complete noetherian local ring with maximal ideal m, and suppose that for each integer n_>_0, the conjugate spectral sequence over S,=Spec(A/m ~+1) is degenerate at E2. Let's denote by (c,,.E'~'b(n), d~(n)) the conjugate spectral over S. (including n = o% putting S~ = S). Then as remarked above, we have (2.3.2.6) whence (2.3.2.7) suppose d 2 (oo) . . . . . (2.3.2.8) whence (2.3.2.9) dr+,(oo)= lim d~+l(n)=0.
e~
b(~
~ c~
d2(00) --- ~ d2 (n)-- 0

d, (oo) = O. Then
.,b _ .,b 9 eonEr+l(OO)-conE2 (oo)= ]lnmconEa2'b(n)= lim r
This shows inductively that (conE~'b(oo), d,(oo)) is degenerate at E2, and completes the proof of validity of the reduction to the case in which S is the spectrum of an artinian local ring. (2.3.2.10) Suppose now that S = Spec(A) with A artin local, and denote by lnga (M) the length of an A-module M. Then a necessary and sufficient condition that tile conjugate spectral sequence degenerate at E2 is that (2.3.2.11) ~ lnga(~o.E ~, b)= Z lnga (tonE'b)"
a,b a,b
The second term in (2.3.2.11) is (2.3.2.12) ~ lngA (~o,E~;b)= Z lngA (R"f, (t2}/s (log D)))
a, b II
25
while the first term is, by (2.3.1.2.1), (2.3.2.13)
~ lngA(co,E~'b)= ~ lngA (F~s Raf,(f2bx/s(log D))).

a,b a,b
Now by hypothesis (2.3.2.1), each of the A-modules Raf,(f2bx~s(log S)) is free of finite rank, and hence each of the A-modules F~*sRaf, (~/s(log D)) is free of the same finite rank. In particular, we have, for each a, b (2.3.2.14) lngA (F*s Raf, (f2bx/s(log D))) = lnga (Raf, (f2bX/S D))). (log Putting together (2.3.2.12-14), the criterion (2.3.2.11) for degeneration at E2 may be written (2.3.2.15) ~ lnga(Raf,(f2bx/s(log D)))=~ lngA(R"f,(f2~/s(log D))).
a,b n
This last equality holds, in virtue of the hypothesis (2.3.2.2) that the Hodge => De Rham spectral sequence is degenerate at E~. This concludes the proof of degeneration. Q.E.D. Corollary. Under the hypotheses of proposition (2.3.2), the formation of the conjugate spectral sequence (2.3.0.1) commutes with arbitrary change of base S'--* S. Proof. By (2.3.2), the conjugate spectral sequence is degenerate at E2, (2.3.2.16) and by (2.3.1.2) the formation of its E 2 term commutes with arbitrary change of S ' ~ S. The result follows by (2.2.1.11). (2.3.3) We are now in a position to explain the terminology "conjugate" spectral sequence. With the assumptions of proposition (2.3.2), suppose further that S is the spectrum of a field K of characteristic p, and that the divisor D is void. The Hodge ~ De Rham spectral sequence (2.3.3.1) (2.3.3.2) (2.3.3.3) (2.3.3.4) (2.3.3.5) (2.3.3.6) E~' b = H b (X, ~X/K) =*" Ha +b (X, ~"~'/K) g~ (H n(X, f2~/r)) ~ H"- ~(X, f~a'/~)being degenerate at Et, we have The degeneracy at E2 of the conjugate spectral sequence
co,E~2'b=F~H~(X, f2~,/~)=~ H~+~(X, f2~/r) r =H~(X ~p~,O~c~/K)=~ Ha+b(X, f2~//~),

g~on n~ (X, f2~c/K ~--H a (X ~p~,f~}?~,~/K) ) 9 g ~ o H"(X, f2~/r)"~g~-~ H~(X ~v~,I2],~,/r ) .
which we prefer to rewrite as gives an isomorphism Putting together (2.3.3.5) and (2.3.3.3) (for X tp~)we find an isomorphism
26
N.M. Katz:
In order to explain the transcendental analogue of (2.3.3.6), let Y be a a proper and smooth C-scheme, and denote by Y"" the "underlying" complex manifold. By GAGA ([39, 36]) and Poincar6's lemma ([14]), we have isomorphisms (2.3.3.7) H"(Y,,O~,/c ) ~ , H"(Y~",f2~/~), ~ H"(Y~",C), ~ H"(Y'~",Z)| by means of which any automorphism of the field C operates on H"(Y,O~,/c) (by transporting by (2.2.3.6) its action on H"(Y a", Z ) | through the second factor). In particular, the automorphism "complex conjugation ", denoted (2.3.3.8) F~bs: C ---}C,
operates on H"(Y, I2~,/c), furnishing a canonical (albeit transcendental) isomorphism (2.3.3.9) Fa*s H" (Y, f2~/c)~ H" (Y, f2~./c . )
The complex conjugate of the Hodge filtration, noted ,onFi, is by definition the image under (2.3.3.8) of the filtration F*s(U) of F*sH"(Y,O~,/c ). According to Hodge theory ([42, 5]), we have, for i= 1..... n, a direct sum decomposition (2.3.3.10) H"(Y, f2~,/c)=FiH"(Y, Or/c) G Fdo, H , (Y, Or/c) 9 ,+1-, 9
n
or, what is the same, a bigraduation (Hodge decomposition) (2.3.3.10 bis) H"(Y, t2~,/c)= @ (Fin Fc~ ') (H" ( Y, f2~,/c)). ~=o This bigraduation gives (transcendental) isomorphisms (2.3.3.11) g~oo,(H"(Y, 12~/c) , ~ , F"-"nVgo,, N gr~_~(H,(y,s ) which we regard as the transcendental analogue of (2.3.3.6). (2.3.4) We must however hasten to point out that the analogue over a field K of characteristic p of the Hodge decomposition (2.3.3.10) provided by F and Fr over C is generally false. Indeed, the extent to which the filtrations F and Fr fail to be transversal is an interesting arithmetic invariant. In order to explain this point more fully, it is convenient to first recall the Hass-Witt "operators" and their relation to'the conjugate spectral sequence. (2.3.4.1) We return to the sheltering hypotheses of proposition (2.3.2). Fixing an integer n>0, the degeneration of the conjugate spectral sequence at E2 gives us an inclusion
~-1
(2.3.4.1.1)
F*~ R"f. (d~x)
~ ~~o,E~'0~.._,R"f. (O~/s(log D))
27
and the degeneration of the Hodge =~ De Rham spectral sequence gives us a surjection (2.3.4.1.2) R"f, (f2~/s(log D)) --~ R"f, (d~x)= E ~
Putting these together, we obtain the diagram which defines the HasseWitt operations: r "~~ R"f, (f2;/s (log D)) --~ E ~ (2,3.4.1.3) /l ]l
Fg~ R"f, (ex) ................... __w n:_ ................:,~ R"f, (6x) _
The matrix of this operation in a local base of R"f, ((_gx)is called the HasseWitt matrix of X / S in dimension n. The composite p-linear mapping (2.3.4.1.4) R,f,(Cx) - F*bs ,F~R,f,(Cgx ) n-w ,Rnf,(Cx)
is the one induced by the p-th power endomorphism of d~ x. From the definition (2.2.4.1.3), it follows that we have: (2.3.4.1.5) Proposition. Hypotheses as in (2.3.2), the kernel of the Hasse144tt operation (2.3.4.1.6) H-W: F* s R"f, (d~x)~ R"f, ((gx)
is none other than the intersection (F~"on F 1)(R"f, (f2~/s(log D))). ~
(2.3.4.1.7) Corollary (Assumptions as in (2.3.2)). In order to have a direct sum decomposition (2.3.4.1.8) R"f,(f2~/s(log D)), ~ F ~GFc"on
it is necessary and sufficient that the Hasse-Witt operation
(2.3.4.1.9)
be an isomorphism.
H-W: F*s R"f,((~x)~ R"f, ff)x)
Proof. Since the statements whose equivalence is asserted are both of the form "a certain homomorphism of locally free S-modules of finite rank is an isomorphism", and because the formation of these modules commutes with all changes of base S' --* S, we are immediately reduced to the case in which S is the spectrum of a field K. The isomorphisms
~g
(2.3.4.1.10)
co~E~'b ~,~ F~*~E~'a
28
N.M. Katz:
together with the respective degenerations (2.3.2.2-3) imply, for every 0 < i< n, the formula (2.3.4.1.11) dimK H" (X, f2~./~(log D)) = dimK F i + t + dim r Fr i.
For the homomorphisms (2.3.4.1.8), source and target have the same dimension, just as for (2.3.4.1.9), and both homomorphisms have the same kernel, namely F 1n F~,. Q.E.D. (2.3.4.2) We now discuss the sometimes defined "higher" Hasse-Witt operations (which include the usual one as a special case), still supposing the hypotheses of (2.3.2). As before, we fix an integer n>0. For each integer i, we denote by h (i) the composite mapping
n-9 9 ~oni (R ?1f, (F2~c/s(log D))) ~ R FIf, (f2x/s(log D))
(2.3.4.2.1)
[ R"f. (f2~,./s(log D))/F i+1
(2.3.4.2.2) P r o p o s i t i o n . Hypotheses as in (2.3.2), and n>O fixed as above, suppose that for an integer i, the mapping h(i) (2.2.4.2.1) is an isomorphism. Then there is a unique mapping of locally free Cs-modules, the i + 1-st Hasse- Witt operation (2.3.4.2.3) H-W(i+I): F,~sR"-i-tf,(f2ix-~sl(logD))-~ R"-i-~f,(f2~Ts~(logD)) which renders commutative the following diagram:
H-W(i+I)
Fa~s(R"-'- ~f, (f~ix-~(log D))) ~~ cohEn2 -i-
1,i+1
E i + 1,n - i - 1 ~ R n - i -
If, (~r~-S1 V)) (log
(2.3.4.2.4)
F~'o~i-l(R"f,(f2~c/s(log D))) h(i+l), R,f,(f2~c/s(log D))/F,+2
F~:'(R"f, (f2~c/s(l3g D)))
~ h(0 ) R ~f*(if2" x/s(l~ D))/F i+1
!
0
29
(2.3.4.2.5) Furthermore, h (i + 1) is an isomorphism if and only if H-W (i + 1) is an isomorphism.

Proof Follows formally from the diagram.
(2.3.4.3) (2.3.4.3.1)
Corollary. Assumptions as in (2.3.2), the canonical map F~o~i F' +1 ~ Rnf~
(~-~t'lS(log D))
is an isomorphism if and only if the map h(i) (2.3.4.2.1)
(2.3.4.3.2)
h(i): F~o~i---~R"f,(f2~/s(logD))/F '+1
is an isomorphism. I f this is the case, then the canonical map
(2.3.4.3.3)
F,o~ ,- 19 F' + 2 ~
Rnf,
(~"~t'/s(log D))
is an isomorphism if and only if the i+ l'st Hasse-Witt operation (which is defined, because h(i) is supposed an isomorphism)
(2.3.4.3.4) H-W (i + 1): F~s R " - ' - 'f, (f21~s (log D)) ~ R"-'- - if, (f2~-s (log D)) ~ ~
is an isomorphism. Proof. The second equivalence follows from the first (applied to i + t), in virtue of (2.3.4.2.5). Precisely as in the proof of (2.3.4.1.7), the proof of the first equivalence is immediately reduced to the case in which S is the spectrum Of a field K. Thanks to (2.3.4.1.11), the homomorphism (2.3.4.3.1) has source and target of the same dimension, as does (2.3.4.3.2), and both homomorphisms have the same kernel, namely F i+1(-3 fcnon i. Q.E.D.
(2.3.4.4) Corollary. Assumptions as in (2.3.2), and n>O fixed, suppose that the map h(i) figuring in (2.3.4.2.4) is an isomorphism. Then: (2.3.4.4.1) The Cs-module u+a~F~ox~-t(R"f,(f2}/s(logD))) is locally free, and its formation commutes with arbitrary change of base S'--~ S. (2.3.4.4.2)
The canonical mapping F~oni- 1 RnJ, (f2~c/s(log D)) , - FZo ' @ F i +1 ~ FZo- i- 1 ~ ~ is an isomorphism. Proof. It suffices to prove (2.3.4.4.2), since by hypothesis F~o ~-~/F~o: i ~ is locally free, and its formation commutes with arbitrary change of base S'-~ S. The composite mapping, formed from the bottom half of the diagram (2.3.4.2.4),
F:o~ '-1 (R" f , (t2~:/s(log D))) hi, +x), R"f, (~]:s (log D))/F '+ 2
F~o-~'(R"J, (Q'x/s (log D))) ~hc0 -~ R.f, (f2~,/s(log D))/F '+1
30
N.M. Katz:
is a section of the inclusion F~= i r ~-1, whose kernel is none other then Fi+lnF~o~ ~-1. This shows that the canonical mapping F~+lc~ F~o~i - 1 ~ F~o~i-a/Ffo~iis an isomorphism, which proves (2.3.4.4.2). (2.3.5) In this section we wish to explain the differential equations satisfied by certain of the higher Hasse-Witt matrices introduced in (2.3.4). These differential equations were first noticed by Igusa [22] in the case of elliptic curves, then later explained quite generally by Manin [28]. (2.3.5.1) We place ourselves under the hypotheses of (2.3.2), and assume further that the divisor D is void, and that the geometric fibres of X/S are connected, of dimension N. Under these hypotheses, we have (2.3.5.1.1) R2Nf,(~Q~c/s)--%RNf,(O~/s) is locally free of rank 1 and even canonically isomorphic to the structural sheaf Cs, via the trace morphism (cf. [15] and [42]) (2.3.5.1.2) tr: RaNf , (Y2]/S)~ d)s
which carries the Gauss-Manin connection on RZNf,(f2]/s) into the standard connection on d)s (the one given by exterior differentiation d: (9s'-~ 12~/r). The cup-product pairings (2.3.5.1.3) Rn f , ([2*x/s)@ R 2N-'f, (fd~c/s)~
R2N f (Q~c/s) " ' (gs
tr
are perfect dualities of coherent locally free S modules, for which the filtrations F and F~o~are both self-dual in the sense that
(2.3.5.1.4)
(2.3.5.1.5)
(F' R n f , (O~/s)) = F N +1 -n R a N - nf , (f2~/s)
(F~'onR ' f , (O~c/s)) = F~o, ~-n RaN-n f , (12~;/s). +
The associated graded pairings

N--i 2N-n r iR n 12* f2* gF( f*(X/S)) | gF (R f,(x/s) ~
r gF( R f,( [2 s)))
2N
n-i
N+i-n
N-i
and , grfoo. (R2Sf, (f2~:/s)) (2.3.5.1.7)

F~* i ,bs(Rn - - i f,(~x/s))|
/l
* N+i-n N-f,(f2x/si ))
/l
* N . F.bsR N f,(f2x/s)
"" __~(9 s
31
are the usual perfect pairings of Serre duality, and the lower row of (2.3.5.1.7) is simply the inverse image by Fab s of the lower row of (2.3.5.1.6). (2.3.5.2) Proposition. Hypotheses as in (2.3.5.1), fix integers n>O and i. The morphism (2.3.5.2.1) h(i): Fc"o-~i(R"f,(~'x/s))--~R"f,(f2"x/s)/F i+j
is an isomorphism if and only if the morphism

(2.3.5.2.2) h ( N - i 1): FcNo+ +i-"(RZN-"f,(I?}/s))--, R2u-"f,(O}/s)/FN-i I
is an isomorphism. Proof. Just as in the proof of (2.3.4.1.7), we are immediately reduced to the case in which S is the spectrum of a field K. By the autoduality of the Hodge and conjugate filtrations (2.3.5.1.4-5), the dual of (2.3.5.2.1) is the natural mapping
(2.3.5.2.3)
FU-'(RZU f,(Q'x/s))-~ R2U-"f,(~l,/s)/Fi~o +1+'-".
Thus (2.3.5.2.1) and (2.3.5.2.3) are isomorphisms or not together. But (2.3.5.2.2) and (2.3.5.2.3) have the same kernel, (F u-in.~'u+l-i-"~con ," (RzN- nf , (Q*x/s)). Since the source and target of (2.3.5.2.2) have the same dimension over K, and the same is true of (2.3.5.2.2), it follows that (2.3.5.2.2) and (2.3.5.2.3) are isomorphisms or not together. Q.E.D. (2.3.5.2.4) Corollary. Hypotheses as in (2.3.5.1), suppose in addition that the relative dimension of X / S is 2. If the Hasse-Witt operation (2.3.5.2.5) H-W: F* s R 2f , (Cx) ~ R2 f , ((-gx)
is an isomorphism, then the higher Hasse-Witt operation

(2.3.5.2.6) H-W(1): fabs (R 1f , (f2x/s))--' R xf , ((2x/s) * 1 1
is an isomorphism. Proof. (2.3.5.2) and (2.3.4.2.5). (2.3.6) We are now ready to discuss the differential equations promised in (2.3.5). We recall our situation: T, a scheme of characteristic p, S a smooth T-scheme, f : X--, S a proper and smooth morphism of relative dimension N with geometrically connected fibres, whose Hodge cohomology is locally free, and whose Hodge =~ De Rham spectral sequence is degenerate at El. Fix an integer n > 0, and denote by a the smallest integer such that R " - " f , (t?"x/s) is not zero. (We also assume n so chosen that R " f , (f2"x/s) is not zero, so that such an a exists.) The a'th Hasse-Witt operation
(2.3.6.1) H-W(a): F*s(R"-"f,(f2"x/s))-*R"-"f,(fPx/s)
32
N.M. Katz:
is defined (because, in dimension n, all the "lower" ones H-W(/), i<a, have source and target the zero module, hence are isomorphisms). In fact, it is none other than the composite (analogous to (2.3.4.1.3))
1s,n--a,a cona.~2 ~.~l~n f
Jt~, j , t'~X/S)
[r
~ E1
a,n--a
(2.3.6.2)
F ~ ( R " - " f , (f~x/s)) .................................. .+ R"-"f,(~x/s). n-w(.) ::
(2.3.6.3) Proposition (Igusa, Manin). Assumptions as in (2.3.6), suppose S affine and so small that all the Hodge cohomology sheaves are free (9smodules. Let co1..... me be a base of R ~- af, (~ax/s) ' and denote by (aij) the matrix in Me(F(S, (_gs))of H-W(a) with respect to this base:
(2.3.6.4)
H-W(a) (F~s (coi)) = Z

i
aji o)j.
(2.3.6.5) Consider the dual basis o9",...,o9" of RN+a-nf~(~'~N~sa ). By duality and the definition of the integer a, it is also the least integer with RN+"-"f, (12~7s~) non-zero, so that the degeneration of the Hodge =~ De Rham spectral sequence gives an inclusion (2.3.6.6)
R "+~ "f,
(a;,/s).
(2.3.6.7) Let ~1 ..... ~e be T-linear differential operators on S which are in the algebra generated by Der(S/T), which operate on the De Rham cohomology sheaves Ri f , (Ox/s) via the Gauss-Manin connection 17.Suppose ~ that m* .... ,co*, considered as sections of R2N-n: t~*x/s1, satisfy the J,~ differential equation (2.3.6.8) Z V(~j)(~o*)=0
J
in Rzu-"f,(~2]/s).
Then each column of the matrix (aij) satisfies the differential equation
(2.3.6.9)
Z ~j(aji)=0, J
for i=1 ..... (.
Proof The proof is based on the fact that the composite p-linear mapping
(2.3.6.9bis) R"-"f,(fPx/s)
F'~b. , F , s ( R , - , f , (fPx/s)) n'W(")~ R"-~ f , (fPx/s)
r may be factored through the subsheaf R, J,~q2o ~v of horizontal sections x/s) (el. (2.3.1.3)), as expressed in the commutative diagram

H-W(a)
33
n-a,a
n-a
\
\F~J,
T
(f?~,s))
F~bs
//a / /
/ / / / pr/
a"f,(f2x ~)v //7

/
n--a
a f,(f2x/s) /
Let ( , ) denote the cup-product pairing (2.3.5.1.3) of De Rham cohomology. We have, for each i,
(2.3.6.10)
and, because ~(coi) is
0 = (~ (co,), E V(~j)(co*)) J
horizontal, we have
0=}-" ~j((~ (col), co*.)] J /" J
(2.3.6.11)
Because each co* has Hodge filtration >N-a, the cup-product (~(coi), co*) depends only on the class of ~(coi) modulo F "+1R "c* ~ q2" / S I , d X by (2.3.5.1.4), which is to say on pr. ~ (col)= H-W(a) (F~*s(coi))-Furthermore denoting by ( , ) the cup-product pairing (2.3.6.1.6) of Hodge cohomology, we have
(2.3.6.12)
(~ (coi), co*)= (pr. ~(coi), co*) = (H-W (a)(F,*s(coi)),co*)
= ( ~ akicok,co*)=aji. Q.E.D.
k
(2.3.7) A Numerical Example-Certain Hypersurfaces of Geometric Genus One (2.3.7.0) Recall that for any base scheme S, and any hypersurface X in p~ + 1 which is smooth over S, the Hodge cohomology sheaves. R"f. (f~x/s) (f: X ~ S the structural morphism) are locally free t~s-modules of finite rank (whose formation consequently commutes with all change of base) and the Hodge =~ De Rham spectral sequence degenerates at E1 (cf. [5 a]). In terms of a system of homogeneous coordinates X1,...,X,+2 on p~+l, we can write an equation for X (i.e., an isomorphism between 9 p(-d), d being the degree of X, and the ideal I(X) defining X in P~ 1),
3 lnventionesmath.,Vol. 18
34
N.M. Katz:
at least locally on S. Localizing on S, we may and will assume that X is defined by a homogeneous form of degree d,
H = H(X1 ..... X,+ 2)e F(S, Cs) [X1 ..... X,+2].
The corresponding short exact sequence on P d~,,p~ + 1 (2.3.7.1) 0---~ (_Or(-d) ~ , d~ --* Cx ~ 0 e
gives an isomorphism of cohomology sheaves on S (n: P; 1 ~ S denoting the projection) via coboundary: (2.3.7.2) R"f, (ex) ~ , R "+1 n, ( O r ( - d)).
Using the standard covering of projective space, the ~s-module R"+I n, (Or(-d)) is easily computed:it's the free Cs-module (2.3.7.3) "forms" of degree - d in /the Cgs-span of those monomials /~ which admits as base the monomials (2.3.7.4)
m ( W ) = X w = X W ' . , 9vw,+2 , -~,+z
--1 --1 W Wl Wn+2 " (~s[X1 . . . . . Xn+2 X 1 . . . . . X n + 2]/ X = X 1 ... X n + 2 for w h i c h
W / = - d but W/=>Ofor some i
~ Wi=-d,
Wi<0
for alli.
We denote by the same symbols re(W) the corresponding basis of R"f, (~x) via the inverse of the isomorphism (2.3.7.2). The "residue" exact sequence in highest degree
0 ~ ~1 ~ ~"~+S1(log X) --~ f2]/s --* 0
(2.3.7.5)
+1
II
~/s |
-1
gives us an isomorphism, the "Poincar6 residue" (2.3.7.6)

n, (Y~,~s | I (X)- I ) _=_, f , (~x/s). 1
In terms of the local coordinates x,= Xi/X.+ 2, the global section of +I O~/s (n + 2) given by (2.3.7.7) X1...X.+2-dxl
Xl
^"" ^ - -
dxn +1
Xn+l
defines an isomorphism (2.3.7.8) Or ~ ; f~I/~1(n + 2)
35
from which it follows that n, (f~sl| basis the differentials

X w dxl
-*) is a free tgs-module, with
(2.3.7.9)
xl
A "'" A - -x,+l ,
dx,+t
Y', Wi=d,
Wi>0 for all i.
The image of these differentials under the residue isomorphism (2.3.7.6) are the differentials on X"
X w dx I
o)(W) =
c3H X,+t 8X,+1
xl
A ' " A - -x. ,
dx.
E W/=d,
W~>0 for all i
(2.3.7.10)
m
x w'-I ...xW${'-XdXl A ... A dx, Oh
Ox,+l The bases m(W) of R"f,((_gx) (2.3.7.4) and og(W) of f,(f2~/s ) are dual to each other under Serre Duality: (2.3.7.11) ( m ( - W),og(V))=~I if V= W if V , W .
(2.3.7.12) Suppose once again that S is of characteristic p. Then from the commutative diagram of sheaves on p~+l
0__). ~p (__ g 0 H )(Qp , ~X _._+0
(2.3.7.13)
l~p-,.r.b,
0-,e,(-d)
lr.b.
F.b.
,ex~0
" ,r
it follows that the p-linear endomorphism of R"f, ((gx) induced by the p-th power mapping (gx---~d~ corresponds via the coboundary isox morphism (2.3.7.2) to the p-linear endomorphism of R" + 17z, (d~p( - d)) induced by the composite ( g p ( - d ) --P'thp0we, ~ e ( - p d ) HP-' ,~p(__t0. This permits the calculation of the Hasse-Witt matrix of a hypersurface: (2.3.7.14) Algorithm. Assumptions as in (2.3.7.0) and (2.3.7.12), the HasseWitt matrix in dimension n of a smooth hypersurface X cP~ +1 defined by
an equation H = H (X1 .... , Xn+ E ) ~ F ( S , (.gs)[X 1 . . . . . Xn+2] may be computed with respect to the basis { m ( - W ) [ ~ . W i = d , W/>0 for aU i} of R" f , ((gx) as follows: Raise H to p - 1-st power, and write it explicitly as a sum of monomials
(2.3.7.15)
3*
H p- 1=}-" A v X v.
36
N.M. Katz:
Then the Hass-W~tt matrix is given by

(2.3.7.16) H-W(F*s(m(- W)))= Z Apw_v m ( - V).
(2.3.7.17) Special Case. Hypotheses as in (2.3.7.14), suppose that X has degree d = n+2. Then R"f, ((gx) is free of rank one on S, with base m ( - 1 , - 1 .... , - 1 ) and the Hasse-Witt matrix of X/S in dimension n (or the Hasse invariant, as we shall call it in this case) with respect to the basis m ( - 1, - 1..... - 1) is given by the coefficient of (X2... X,+2) p-2 in H p- 2. We now apply this to a particularly beautiful family of hypersurfaces. (2.3.7.18) Corollary. Let d>2 be relatively prime to p, and put S= Spec(Fp [2] [1/1 - 2d]). Consider the smooth (over S) hypersurface X in Ps of equation d-1
d
(2.3.7.19)
~, X d - d 2 X1... Xd.
i=1
It's Hasse invariant is given by the truncated hypergeometrie series

(2.3.7.20) (d2)P- 2
~,
a~O
(1/d)a(2/d)a...(d-1/d), 2_ad
a ! a !... a !
where for ~4=0, we put (~)o=1, ( ~ ) , = ~ ( ~ + l ) . . . ( ~ + n - 1 ) /f n > l . This may be expressed in terms of the "full" hypergeometrieseries in Zp [[2- 2]] [2] by means of Dwork's congruence (cf. [8 a], pp. 36-37)
(2.3.7.21) (d2) p-2 y'

a=o
(1/d)a(2/d),...(d- 1/d)a 2_ad=_G(2)/G(2P)modulo(p)

a!a!...a!
where G(2)=(d2) - 1 F [ 1/d, 2/d..... d - 1/d. given by \ 1, ..., 1 , 2 -d) is the element of Zp[[2-1]]
(2.3.7.22) G(2)=2-1a~o(-l) a'd-1) ( - 1 / d ) ( - 2 / d ) . . . (-(d-a l)/d)2_ad"
Proof By direct calculation, one finds that (2.3.7.20) is the coefficient of (l--I x,) v-1 in ~ X ( - d 2 I-I X,. To apply the congruences of Dwork,
i i
we need only observe that in the sum (2.3.7.20), we could have let a run from 0 all the way to p - 1 (which is the usual first truncation point for hypergeometric-type series), because, for a ~ p - l but ad>p, we have (2.3.7.23)
(ad)! (1/d)a(2/d)a...(d-1/d)a= dad.a! - 0
modulo p.
Q.E.D.
37
(2.3.8) 7he example, continued. According to (2.3.6), the Hasse invariant of the hypersurface (2.3.7.19) is to satisfy every differential equation satisfied by the differential o~=co(1 ..... 1) (in the notation (2.3.7.10)) considered as a section of R a- 2f. (f2]./s)"By Dwork's congruence (2.3.7.21), this is equivalent to the formal series G(2) (2.3.7.21) satisfying every such differential equation. Because the formal series G(2) is universal, i.e., independent of p, and the formation of the Hodge and De Rham cohomologies of a smooth hypersurface commutes with arbitrary change of base, it follows that we have: (2.3.8.1) Corollary. Consider the hypersurface smooth over the spectrum S of Z[2][1/d(1-2d)] given by the equation (2.3.7.19). For every differential equation d i (2.3.8.2) 2 ai(2) 17 (d-f)(r in Rd-2f,(~/s)
satisfied by c0=09(1, ..., 1) in f, (f2~rTs2),ai(2)~Z[~.] [1/d(l-Aa)], we have

(2.3.8.3) ~ai(2) ~ (G(2))=0
in Z
[[2-1]],
whereG(2)EZ[1][[2-1]]istheseries
(2.3.8.4)
-ad a(R)=(d2)-I ~ (1/d)a'"(d- 1/d)a 1~ . a! ...a!
(2.3.8.5) Remark. In fact, "the" differential equation satisfied by co is (2.3.8.6)
(Z(-'
d,"-'
as may be deduced from 1-23a] and an immediate calculation shows that indeed (2.3.8.7)
[ d ~d-1
~-d2-!
(G(2)) =
(2 d ~d-1
-d2-!
(2G(2)).
(2.3.9) We refer to forthcoming works of B.Mazur for the congruence relations between the higher Hasse-Witt matrices and the zeta function, which generalize the "ordinary" congruence formula [251.
2.4. The Question of Quasieoherence of the Conjugate Spectral Sequence (2.4.0) We return now to the geometric situation of 1.0, and, to fix ideas, we suppose that S is affine. The conjugate spectral sequence
(2,4.0.1)
E~2'b=Raf,(Jfb(f2[/s(logD))) =~ Ra+bf, (~/s(1Og D))
38
N.M. Katz:
is by definition the "second spectral sequence of hypercohomology" (cf. (2.2.2.2)). In general, we do not know whether or not either the E2 terms or the E~ terms of this spectral sequence are quasi-coherent d~s-modules (though of course the R"f, (127,75(logD)) are quasicoherent). (2.4.0.2) Consider a coveting { V/} of X by affine open sets. From it we construct a Cech bicomplex of quasicoherent (gs-modules (2.4.0.3)
c " = c"({ v,}, ; /s(log 09;

C a'l* = C a
({ V/},~X/S(log D)) = [I(flaa) , ((~x/s (log D) Iaa) ffa
where aa runs over the a-simplices V/on . . . n V/oof the nerve of the covering. The homology sheaves of the associated simple complex are the R" f, (g2~;/s(log D)), just because f is quasi-compact and separated, the various f2~x/s(log D) are quasi-coherent (gx-modules, and the d in f2~/s(lOgD) is f - 1(Cs).linear.(cf. [ 12], III). (2.4.0.4) The "first" spectral sequence of this bicomplex, the one associated to the "Hodge" filtration (2.4.0.5)
F j C ~176 ~ C'({V~}, Ob/s(1OgD)), = b>=j
gives the usual Hodge ~ De Rham spectral sequence: (2.4.0.6) E~"b= Rbf, (f~x/s(log D)).
The second spectral sequence of this bicomplex, the one associated to the "conjugate" filtration (2.4.0.7) F k C " = Z Ca({V~}, Q~/s(l~ D))
a>k
gives rise to a spectral sequence (2.4.0.8) E~'b=/~({V~}, ~sh~,(f2]/s(1Og D)))~ Ra+bf,(f2]/s(logD)) of quasi-coherent d~s-modules (~ denoting the quasi-coherent sheaf associated to a F(S, (.0s)-module), which maps canonically to the conjugate spectral sequence (2.4.0.1), (2.4.0.9)
E~2'b=Raf,(~t~
) ~ Ra+bf,(Q~/s(logD)).
Of course, for each affine open coveting of X, the spectral sequence (2.4.0.8) is a spectral sequence of quasi-coherent (gs-modules, but the canonical mapping from (2.4.0.9) need not be an isomorphism, even if we replace (2.4,0.8) by its direct limit over all affine open coverings of X.
39
(2.4.1) In the special case when S is a schema of characteristic p, the conjugate spectral sequence (2.4.0.1) is the second spectral sequence of hypercohomology for the (,0xcp)-linear complex F,(O}/s(log D)) on X ~p), and the functors R"J~ ~. Because f~P): X ~p)--, S is quasi-compact and separated, the various F,(f~x/s(log D)) are quasi-coherent (gx-modules, and F,(d) is Cxc~,-linear, the second spectral sequence (2.4.0.8) of the Cech bicomplex does map isomorphically to the conjugate spectral sequence (2.4.0.9), which is a spectral sequence of quasi-coherent (_9 smodules. The simple interpretation of the conjugate spectral sequence in characteristic p as the second spectral sequence of the Cech bicomplex (2.4.0,3) makes possible effective calculations, as we shall see. On the contrary, the conjugate spectral sequence over C remains in shadow.
3. The Main Technical Result on the p-Curvature of the Gauss-Manin Connection 3.0. We return to the geometric situation 1.0, and assume as before that S, and hence T, is a scheme of characteristic p. Recall that for any (9smodule with integrable T-connection, (M, V), its p-curvature ff is the p-linear homomorphism of Cs-modules (cf. [24], 5.2)
(3.0.1) defined by (3.0.2) ~: Der(S/T)--, Endes(M)
~k(D) = (V (D)) p - V (DP).
Equivalently, we may view ~ as defining by transposition a homomorphism of (~s-modules, also noted ~,, (3.0.3) r M -~ F~(~2~/T)| M
(where F~bSdenotes the absolute Frobenius endomorphism of S). The significance of p-curvature (due to Cartier; cf. [24], 5.1) is the following: (M, V) has p-curvature zero if and only if M is spanned over (9s by the subsheaf M ~ of horizontal sections; more precisely, denoting by S w) the fibre product of g: S--~ T and the absolute Frobenius F~b~: T--~ T, if and only if the canonical mapping (3.0.4) is an isomorphism.
(Me) |
(~s --* M
3.1. Taking for M the De Rham cohomology Rnf,(f2x/s(logD)), and for V the Gauss-Manin connection, we recall once again (cf. [24], 3.5) that the entire conjugate spectral sequence (3.1.1)
E"2"b=R"f,(~t~
~ Ra+bf, (f2~c/s(logD))
40
N.M. Katz:
is endowed with the action of the Gauss-Manin connection, and that on the E2 terms, the p-curvature is zero. This implies that the p-curvature also vanishes on the Eo~ terms, and hence that (3.1.2) ~(F~o,R"f,(gJ~,/s(logD)))cF*~(fJ~/r)| +1 R"f, (gJ~r (log D)).
Passing to the associated graded, there is an induced mapping, again denoted r (3.1.3) ~J: gr~oo,(R"f, (O~/s(log O))) --* F~* (~2s~/7-) gr~ooo(R"f, (~J~?s(logD))). s | Our main technical result 3.2 identifies this mapping, under suitable hypotheses with a twisted form of the Kodaira-Spencer mapping (1.3.2.1). 3,2. Theorem. Under the hypotheses of (2.3.2), the diagram below is commutative g~oo,(Ra+Of,(CJ~./s(logD))) ~ , F*~(fJ~/r)| )
conE~b
17.s [ O 1 ]t'~ a a b ~a'~S/T] ~
con~
g'a+l,b-1
eong~,b
b'*s ~O1 / T ]" ~ t ~ tT'a+l,b-1 ~ eon~L' 2 .t a b I,a a S
~k b F~b (R a f , (Qx/s (log D))) (-- 1)b+'F~*b~(p~ (~/T) | F*~ (R~+~f, (~b/j (log D))) s Fa~
bs
s ~ Fa*bs
Rof,
D))
(- ' ' ....
| g o+ ' / ,
(log D))
in which p is the cup-product with the Kodaira-Spencer class (1.1.3), viewed as a global section over S of Os/r| R 1f , (Dero(X/S)). 1 Before proceeding to the proof, we will recall some basic facts about the modular representation theory of finite groups of order prime to p, and then restate 3.2 "with a group of operators ". (3.2.1) Let G be a finite group of order prime to p, and k a field of characteristic p. Let V be a finite-dimensional k-space on which G acts as a group of k-automorphisms, through a homomorphism Z: G ~ GL(V). Denoting by F,bs: k ~ k the absolute Frobenius endomorphism of k, the representation Z(v) of G on F.'~,s(V)=V| (where k is a module over
k
Algebraic Solutionsof Differential Equations itself by F~bs: k--~ k) is given by
41
(3.2.1.1)
x ~p~ (v | a) d~. x(g) (v) | a. (g)
In matricial terms, the matrix coefficients of Z~p) are the p-th powers of the matrix coefficients of Z. The character of z~P~is the p-th power of the character of Z. (3.2.2) Proposition. Hypotheses as in (3.2.1), /f Z is an absolutely irreducible representation of G in a finite-dimensional k-space, then its degree divides the order of G, and in particular is prime to p.
Proof. This will be a simple consequence of the analogous fact for representation in characteristic zero. In fact, extending scalars if necessary, we may suppose that k is the residue field of a discrete valuation ring ((9, p) whose fraction field K has characteristic zero, and is such that every irreducible representation of G in a finite-dimensional K-space is absolutely irreducible. In fact, we will prove
(3.2.2bis) Proposition. Let ((9, p) be a discrete valuation ring whose fraction field K has characteristic zero, and whose residue field k has characteristic p > O. Let G be a finite group, of order prime to p, such that every irreducible representation of G in a finite-dimensional K-space is absolutely irreducible. Then every irreducible representation of G in a finite-dimensional k-space is the "reduction mod p" of a representation of G in a free (9-module of finite rank, which is irreducible over K, and which is determined over K up to isomorphism.
Proof. Because G has order prime to p, the group-ring k [G] is semisimple, as is K [G], and the isomorphism classes of irreducible representations of G in finite dimensional k-spaces (resp. K-spaces) are in bijective correspondence with the indecomposable central idempotents f~, ..., fr (resp. el ..... e~) in k [G] (resp. K [G]), which give the projections onto the "isotypique" components of any finite dimensional k [G] (resp. K [G]) module. We have
(3.2.2.1)
l=~ei,
f = ~ ai ei,
eiej=6ije i
and for any central function f in K [G], (3.2.2.2)

ai ~ K .
We will first show that the ei give an (9-base for the central functions with values in (9. The ei are given by the following explicit formulaes, in which Z~ denotes the corresponding irreducible representation: (3.2.2.3)
ei = deg(~(i) ~ trace(~(~(g-1))'g" #G
42
N.M. Katz:
Because G is finite, any representation of G on a finite dimensional K-space comes by extension of scalars from a representation of G in a free (9-module of finite rank (take the lattice generated by the G-translates of any lattice), and hence its character takes values in (9. It follows that the functions ei take values in (9; i.e., lie in (gEG]. Because the Z~ are absolutely irreducible, we have (3.2.2.4) deg(xi)l ~ G, hence p4Vdeg(gi).
Hence the values of the e~ at the identity element 1G G are units in (9, because deg(zi) (deg(gi)) 2 (3.2.2.5) ei(1)= 9trace (g~(1))= ~:G :~G Thus iffG (9 [G] is central, f = ~, ai ei and (3.2.2.6) ai
(ai ei)(1) ( f ei)(1) G(9. ei(1) - ei(1)
Thus the e~ give an (9-base of the center of (9 [G]. Now let f~Gk[G] be the indecomposable central idempotent corresponding to an irreducible representation X- We can certainly lift f~ to a central function f with values in (9, which we may write (3.2.2.7)
f = ~ aiei,
aiG(9.
Because fl "f: = f l , we have (3.2.2.8)
~ a { ei = f 2 =_f = ~ ai e, mod(p [G])
and multiplying both sides by e~, we get (3.2.2.9)
a~ ei ---ai ei mod p [ G ] .
Evaluating at the identity element l e G, we have (by (3.2.2.4)) (3.2.2.10)
a~=ai m o d p ,
so that each coefficient ai is congruent to either 0 or 1 mod p. Thus we may lift f~ to a central idempotent (3.2.2.11)
~,eiei,
ei=0
or 1
in (9 [G]. Because the e~ are a basis of the center of (9 [G], the indecomposability of f~ as central idempotent in k I-G] implies that e~ differs from zero for only one value of i, say i = 1. This shows that ei is the unique central idempotent in (9 [G] lifting f~. This shows that Z is the "reduction mod p" of a unique irreducible representation Xl of G in a finitedimensional K-space. Q.E.D.
43
(3.2.3) Corollary. Hypotheses and notations as in (3.2.2), the indecomposable central idempotents in K [G], which all lie in (P [G], have reductions modulo t~ in k [G] which remain indecomposable.
Proof. The number of e i is equal to the number off~, both being the number of conjugacy classes in G. Hence every e i lifts an ft. Q.E.D.
In terms of representations, this gives: (3.2.3 bis) Corollary. Hypotheses as in (3.2.2bis), the reduction mod p of any K-irreducible representation of G in a free C-module of finite rank is irreducible (and every irreducible representation of G in a finite-dimensional k-space arises this way). (3.2.3.1) Corollary. Hypotheses as in (3.2.2his), every irreducible representation of G in a finite-dimensional k-space is absolutely irreducible.
Proof. This follows from (3.3.0bis), applied to arbitrary finite extensions K' of K, and arbitrary extensions of the valuation of K to K', by which we can realize arbitrary finite extensions k' of k as residue field.
(3.2.3.2) Corollary. Hypotheses in (3.2.2 his), the indecomposable central idempotent (noted P(X)) associated to an irreducible representation ;( in a finite-dimensional k-space is given by the formula (3.2.3.3) P(Z) = deg@G(Z_) trace (z(g- 1)). g. g~
(3.2.4) We return to the hypotheses of (2.3.2), and suppose given in addition: (3.2.4.0) a finite group G of order prime to p, which acts as a group of S-automorphisms of X and preserves the divisor D (though not necessarily the individual Di); (3.2.4.1) a subfield k of F(T, (,Or) and an absolutely irreducible repre, sentation X of G in a finite-dimensional k-space. (3.2.4.2) By functoriality, the group G acts on the De Rham cohomology sheaves R" f , (g21/s(log D)) as a group of (~s-linear horizontal (for the Gauss-Manin connection), and respects both the Hodge and the conjugate filtrations. (In fact, G acts on both the Hodge =~ De Rham and the conjugate spectral sequences.) (3.2.4.3) The functoriality of the diagram (3.2.0) shows immediately that it is a diagram of G-morphisms (though the lower vertical arrows F~s, being p-linear, are not k [G]-morphisms). This said, we may now "restate" 3.2.
44
N.M. Katz:
(3.3)(=3.2bis) Theorem. Under the hypotheses of (3.2.4), the following subdiagram of(3.2.0) is commutative.
p(z~v,)(~o,,E~b)_
, (I|
(p)
a, b
))
~(log D)))
P( Z'p)) F ~ (R~f, (fdbx/s)) ~-l)b+' Fa*b'( (1 | P(x~P'))(F~,~(f2~/r)) | F*~(R ~+'f, (s P)~'

bs bs ab~
P(x) R~f,(~x/s(log D))
~_l)b+,p ,(l|174
Proof (Assuming 3.2). Thu upper square is deduced from the upper square of (3.2.0), which is a diagram of k [G]-modules, by applying the projector P(xC~ The lower square is deduced from the lower square of (3.2.0) by applying P(Z) to the lower horizontal line, and noting that
( ~ F*so(P(z))=F*s \ deg (Z) ~, trace(z(g-~)) 9gt oF~s / (3.3.1) = ( deg (Z) '~P \ # G I Z(tracez(g-1))Vg~ Q.E.D.
(3.3.2) Corollary. Hypotheses as in (3.2.4), suppose that the absolute Frobenius endomorphism of Cs is injective (which is the case if T is reduced, for example). If for a fixed integer n, the p-curvature of the Gauss-Manin connection on the "part of R"f,(s D)) which transforms by Z~P. )'', i.e., on the submodule P(ztP~)(R"f, ((2]/s(log D))) is zero, then the induced Hodge filtration on P(z)(R"f,(f2"x/s(log D))) is horizontal, i.e., stable by the Gauss-Manin connection.
3.4. The Proof of 3.2: Reduction Steps
(3.4.0) The question being local on S, we may and will assume S affine. As explained in 2.4, the De Rham cohomology sheaves on S are the total homology of the Cech bicomplex associated to any affine open covering of X, and the Hodge and conjugate spectral sequences are the "first" and "second" spectral sequences of the Cech bicomplex. The proof we will give is not at all intrinsic, but rather depends on explicit calculations on the level of the Cech bicomplex itself. It would be of considerable interest to give an intrinsic proof. (3.4.l) We choose a finite covering of X by affine open sets V/(as in (I.0.3.1)), each 6tale over A~ via "local coordinates" xi(i), ..., x,(i), such that those branches of D meeting V~ are defined by the equations xl(i)=0 ..... x~(i)=0, ~ depending on Vi.
45
We next define a (not necessarily integrable) T-connection 17 on the simple complex deduced by "totalization" from the Cech bicomplex, which yields the Gauss-Manin connection upon passage to homology (cf. [24, 35]). For any DeDer(S/T), we denote (3.4.1.1) D(i)=the unique element of Dero(VjT ) which extends D and which annihilates the chosen local coordinates x~(i) .... , x,(i), and by (3.4.1.2) Lie(D(/)): f2~, s(log D ) ~ f2~,/s(log D)
the "Lie derivative with respect to Di". For each pair of integers i<j, we denote by (3.4.1,3)
9 o- 1 I(D(i)- D(j)): f2v,~vj/s(log D) ~ fJv,-nvj/s(log D)
the operator" interior product with the S-derivation D(i) - D(j)" (cf. (1.2.2)). The connection 17 on C'({Vi},g2]/s(logD)) is given as follows. For a fixed integer b, a section (3.4.l.4) and a simplex io < " ' (3.4.1.5)
<
o)~ Z C"({V~}, ~b/s(log D))

a
iq,
(V (D)(o)))(io, ..., iq)= Lie(D(io))(oo(io ..... iq)) + ( - 1)b I(D (io)- D(ix))(og(il ..... iq)).
Thus 17(D) is the sum of two terms. The first, of bidegree (0, 0), is the cup-product with the 0-chain {Lie(D(/))}. The second, of bidegree (1, -1), is the cup-product with the 1-cocycle I(D (i)-D(j)). This explicit construction makes clear computationally the truth of (1.4.1.6), (1.4.1.7), and (2.3.0.1). In order to prove 3.2, it suffices by linearity to establish the commutativity of the outermost square of (3.2.0), and after "contraction" with any D~Der(S/T). Thus we must prove commutative the diagram
R f,(JY (f2~/s(logD)))
a b 9
~R,+lf,(gfb_l(f2~/s(logD)))
l~e-I ~
(3.4.1.6)
l~e-l~
R"f,(ab/s(log O)) (-1)b+10(D),RO+l f,(f~xSs~(log D)).

Let us explicate the arrows in this diagram. The vertical ones are deduced from the morphisms of sheaves (3.4.1.7)
f~bx/s D) "*, f~x,,,/s (log D ~"))--~21 , o~ b(a}/s (log D)). (log
46
N.M. Katz:
In order to render the vertical arrows still more concrete, we introduce, for each open set Vi, a p-linear endomorphism O5 of ~v,/s (log D), whose image lies in the subsheaf of closed forms, and which "lifts" the mapping 3.4.1.7. Indeed, the formulas (2.1.2.1) give such a lifting: we require that o5(1)=1
(dx~(i) ~ = dxv(i ) O5 \ xv(i) ] x~(i)

(3.4.1.8)
for i = 1..... for i = ~ + 1..... n
os(dx~(i))=(x~(i)) p-~ dx~(i)

os(co ^ ~) = os(co) ^ O5(~) os(co + ~) = os(co) + O5(~)
os(h co) = h p os(co)

We then have our desired lifting (3.4.1.9)
for h~(gv~.
f~v,/s(logD)
~=' , closed forms ~v,/s(logD)
tcanonicalprojection
~ b ' ~ o ~ ( v,/s( 1og D ))~

(3.4.1.10) The lower horizontal arrow in (3.4.1.6) is, as previously (cf. (1.1.3)) noted, ( - 1)b+1 times the cup-product with 1-cocycle
{I(D(i)-O(j))}.
The upper horizontal arrow in (3.4.1.6) is slightly less straightforward to explicate. Let ~ be a section of Raf. (j~ffb(i2~/s (log D))). We may represent by an a-cochain ~ of closed forms (3.4.1.11) whose image in
z~ca({v~},P~x/s(logD)),
dz=O
ca({ Vii},~b(~2]/s(log D)))

is a cocycle representing ~. By the degeneration of the conjugate spectral sequence at Ez, we may choose z to be the component of bidegree (a, b) in a total a + b cocycle a which lies in (3.4.1.12) F~,= 2
i=>O
ca+i( {Vi}' f~x~-s'(logD)).
The section ~, (D) (~) of R a + i f, (ocgb- 1(f2~/s(log D))) is then represented by the component of bidegree (a + 1, b - 1) in the total cocycle (3.4.1.13 ) ( V(D)p - V(D')) (a).
47
Recall that (3.4.1.14)
(V(D) p - V(DP))(Fr176176
'')
(because V(D)(F~,(C~176176 ), and V(D) on gffvr ~176 an inteis grable connection of p-curvature zero, given on Vioc~.-. n Vi, by D--+ Lie (D(i0))). Since a is congruent to z modulo F~o, ( C ,,, It follows from (3.4.1.13) ,+1 ) 9 that (3.4.1.15) (V(D) p - V(Dp))(a)---(V(D) p - V(D'))(z) modV~o+2. Thus the section O(D)(a) is represented by the component of bidegree (a+ 1, b - 1) in ( V ( D F - V(DP))(r). (3.4.2) Lemma. Let zeC"({Vi}, Y~x/s(logD)), and io<ia<...<i,+ 1. For each integer n > 1, the element (3.4.2.0)
V(D)"(z)e ~ C"+'({V~}, ~xTs'(log D))

i>O
has its components of bidegree (a, b) and (a+ 1, b - 1) given by (3.4.2.1) (3.4.2.2)
V(D)"(z)(io .... , i,)=Lie(D(io))"(Z(io, ..., i~)), V(D)'(z)(io,... , i,)

---(-1) b ~
Lie(D~o)k t(D(io)- D(iO) Lie(D~f (z(i I ..... i,)).
k+d~n--1
Proof The proof, by induction, is immediate; the case n = 1 is the definition. Q.E.D.
Putting n = I', we find
(V (D) p - V (DP))(z)(io, ... , i~)

(3.4.2.3) = [(Lie (D (i0)))P - Lie (D (io)P)] (z (io ..... i,)) - 0, since (D(io)) p =- DP(io)
(V(D) p - V(DP))(z)(io ..... ia+O

(3.4.2.4) =(-1) b ~
Lie(D~o)kI(D(io)-D(iO)Lie(D,y(z(q
.... ,i,+~))
k+d=p+l
- ( -- 1 )b I (D (io) p - D (iO p) (z (il ..... i. + 1)).
Thus (3.4.2.4) gives a formula for a cocycle representing ~h(D)(a) in R"+lf. (3r D))), when we take for z a particular representing cochain for ~. (3.4.3) Combining all our explications (3.4.1.9), (3.4.1.10), and (3.4.2.4), we see that the commutativity of (3.4.1.6), and hence the truth of 3.0, is implied by the following assertion:
48
N.M. Katz:
(3.4.3.0) Assertion. For any cochain t e C'({V/}, f~x/s(log D)), the cochain in C a+a({ V~},~x~ ~(log D)) which assigns to (io, ... , i,+ 2) the b - 1 form (-1) b (3.4.3.1) ~
Lie(D(io))kI(D(io)-D(il))Lie(D(il))e(~i,(z(i~ .... ,i,+1)))
k+d=p--i
--(-- 1)bI(O(io)P--O(ia)v)(~,(z(il .... , i,+1)))

-- ( -- 1 )b +1 ~/o
(I (O (i0) -- O (il)) (t (is,
... ,
i, + 1)))
is a cocycle of closed forms, and is cohomologous to zero in
C ~+1 ({ V/}, Jt~b-' (O~/S(log O))). (3.4.3.2) In fact, we will prove that the cochain (3.4.3.1) is in fact a cochain of exact forms, and so vanishes in C ~+ 1({ Vi}, ~ b - 1(f2~/s (log D))). Notice that the occurence of ~o in the last line of (3.4.3.1) may be replaced by ~ , without modifying the class modulo
d ( r (V~o c~ ... c~ ~ . . . . ~x?s 2 (log D)))
of the cochain (3.4.3.1), because both ~,o and -5/, restricted to V/oc~ V/, are liftings of c~-loa* (3.4.1.7). This change made, the truth of (3.4.3.0), and hence of 3.2, results from the following proposition (3.5.0), which may or may not prove to be of any independent interest. In it, V~oc~... c~ V~o is renamed V, D(io) and D(iO are renamed P and Q, t(il, ..., i~+1) is renamed t, ~ , is renamed if, and the coordinates xl(iO ..... x,(i 0 are renamed Xl, ..., x~.
3.5. Conclusion of the Proof of 3.2
(3.5.0) Proposition. Let S be an affine scheme of characteristic p, and V a scheme ~tale over A~, by means of "coordinates" x l , . . . , xn on V. Fix a (possibly empty) subset ~ of {l ..... n}, and let D~ ~ V be the divisor with normal crossings relative to S defined by [-[ xi=O (we put D o =0).
Let D~Der(S/Fp) be any derivation of S, P~Dero~(V/F v) an arbitrary extension of D to a derivation of V which preserves the ideals xi Cv, for i ~ , and Q ~ Der,~ (V/Fv) the unique extension which annihilates xl,..., x,. We denote by ~ : f2~:/s(logD~)--~f2~:/s(logD~) the unique p-linear endomorphism of f2~:/s(log D~) as (ge-module which satisfies
.~(1)= 1
( dx, l _ dx, ,
(3.5.0.1)
\ Xi / Xi i~o~ ~ ( d x i ) = x e-x dxi,
~(~o ^ t ) = ~(~o) ^ ~ ( t ) .
49
Then for any differential form

(3.5.0.2) we have (3.5.0.3)
k+d=p-1
e F (V, f2b/s (log D~)), Lie (P) k I ( P - Q) Lie (Q)e ( ~ (z))- I(P p - Qp)(o~(z)) ~ (I (P - Q) (~)) modulo dr(V, ~?s 1(log D~)).
Proof. Both sides of the asserted congruence (3.5.0.3) are p-linear in z, and exterior differentiation is linear over p-th powers, so it suffices to check the case in which z is a product of the one-forms
dx i
i~
(3.5.0.4)
Xi
dxi,
iq~.
We first introduce some auxilary notation: (3.5.0.5) z(i)= / xi dxi,
dxi,
if i~cz if i~ct, if/ca if i ~ , if i ~ if i~a, if iea if i ~ , if iect if i~e.
(3.5.0.6)
[ dxi , tr(i) = / xi x p-1 dx,

f(i)=/
[ P(x/) x~ '
(3.5.0.7)
(X p-I P(xi) (3.5.0.8)
[ P(xi) g ( i ) = / x~ ' tP(xi), [ PP(xi) , h(i)= / x, [x p-l Pv(xi),
(3.5.0.9)
(3.5.1) Lemma. The following relations hold among the "quantities" z(i), 6(i), f (i), g(i), h(i)
(3.5.1.1)
~(z(i))=tr(i),
Lie (P)(a (i)) = df(i),
(3.5.1.2)
4 Inventiones math., Vol. 18
50
N.M. Katz:
(3.5.1.3)
(3.5.1.4)
Lie(Q)(a(/)) =0,
I(P-Q)(z(i))=g(i), I(P-Q)(a(i))=f(i), I(P p - Qp)(a (i)) = h (i).
(3.5.1.5)
(3.5.1.6)
Proof (3.5.1.1) is by definition (3.5.0.1) of ~ As for (3.5.0.2), we simply calculate

if ie~, (3.5.1.7)
Lie(P)(a(i))=Lie(e) (
\
dx, I = Lie(e)(dxi)
xi ] xi
P(x,) dx~
xi xi
dP(xi)
Xi
P(xi) dxi
Xi Xi
=d \-~-i !=tiT(i)"
(3.5.1.8) if ir we multiply the above calculation (3.5.1.7) by x~.
(3.5.1.3) holds because Q (xi)= 0 by definition (3.5.0) of Q. As for (3.5.1.4), we calculate (3.5.1.9) if ie~, l i P - Q)(z(i))= I ( P - Q) ( dxi ] - ( P - Q) (xi) - P(x,) =g(i),
\ xi ! xi xi
(3.5.1.1o) if ir l(P-Q)(z(i))=I(P-Q)(dxO=(P-Q)(xi)=P(x,)=g(i)
and (3.5.1.5) follows similarly:
(3.5.1.1l)
(3.5.1.12)
if i ~ ,
if ir
\ xl ! xi I(P-Q)(a(i))=I(f-Q)(x~-ldxi)=xe,-1p(xi)=f(i).
I(P-Q)(,~(i))--I(P-Q) ( ,ix, ] = elx,) =f(i),
The proof of(3.5.1.6) is identical to that of(3.5.1.5), depending only on the fact that Qp(xi)=0. Q.E.D. (3.5.2) We now return to the verification of the congruence (3.5.0.3) for a product of the z(i). To begin, we'll check the case z=r(1). In this case (b = 1), the congruence (3.5.0.3) becomes an assertion of equality: Lie (P) k I ( P - Q) Lie(Q)e ( ~ (z (1)))-I(f p - QP) ~ (z (1)) (3.5.2.0) k+:=p-a
= -~(I(P-Q)(~(1))).
Substituting via (3.5.1), (3.5.2.0) becomes (3.5.2.1)
Z
k+g=p-1
Lie(p)RI(p-Q)Lie(Q)e(a(1))-I(PP-QP)(a(1))=-~(g(1))
51
and substituting via (3.5.1) and especially via (3.5.1.3), (3.5.2.1) becomes (3.5.2.2)
Lie(P)P-lI(P-Q)(a(1))-h(1)=
- g ( 1 ) p.
A final substitution via (3.5.1) gives the assertion (3.5.2.3) PP-1 ( f ( 1 ) ) - h(1)= - g ( 1 ) p.
We now "decode" (3.5.2.3) by returning to the definitions (3.5.0.7-9) of f, h, g; (3.5.2.3) becomes the assertion (3.5.2.4) (3.5.2.5) if 1 ~ , if 1r
PP-I(P(xOI\xl !
PP(XO_x,
( ~ 0 ) p,
PP-1(x~-lp(xl))-x(-IPV(xO=-(P(Xl)) p.
Both (3.5.2.4) and (3.5.2.5) are in fact true, and follow ((3.5.2.5) directly, (3.5.2.4) after dividing by x 0 from Hochschild's identity (cf. [20]), according to which, if P is any derivation of any commutative ring A of characteristic p > 0 , then for any element xeA, (3.5.2.6)
Pp-l(xo-~ p(x))-xp-l Pp(x)= -(P(x)) p.
This concludes the proof of (3.5.0.3) in case b = 1. In the following, we will make use of it, through the identity (3.5.2.3). (3.5.3) Because our proof of (3.5.0.3) in the general case is so unenlightening we first present the proof in the case b = 2, which is somewhat more intelligible. The assertion to be verified is Lie (p)k I (P -- Q) Lie (Q)e(~ (r (l) ^ z (2)))
k+g'=p- 1
(3.5.3.0) -
- I (Pp- Qp)(o~(z (1) ^ z (2)))

o~ (I (P - Q) (, (1) ^ z (2))) modulo
dF(V, (gv).
Substituting via (3,5.1.1) and noting that by (3.5.1.3) the terms under the summation sign vanish for ( ~ 0 , (3.5.3.0) becomes (3.5.3.1) Lie (P)P-x (I (P - Q))(~r (1) ^ a (2)) - I(Pp - QO)(a (1) ^ a (2))
~----~(I(PQ)(,(I) ^ ,(2)))modulo dF(V, (9~,).
Expanding the interior products and substituting via (3.5.1), (3.5.3.1) becomes (3.5.3.2)
4*
Lie (P)P-1 (f(1) a (2) - f(2) a (1)) - (h (1) a (2) - h (2) a (1))
--
~(g(1)z(2)-g(2)z(1))modulodr(V, Cv).
52 A final substitution gives (3.5.3.3)
N.M. Katz:
Lie (P)P-1 (f(1) a (2) - f(2) a (1)) - (h (1) a (2) - h (2) a (1))
= g(1)Pa(2)+g(2)Pa(1)modulo dF(V,(_%).
Lie(P) ~- t(f(l) a (2)-f(2) a(t)) = W-t(f(t)} a(2)- p~-i (f(2)) a(l) modulo dF(V, (9~).
Availing ourselves of (3.5.2.3), (3.5.3.3) becomes (3.5.3.4)
Now the right hand member of(3.5.3.4) is the "first term" in the expansion of the left hand member by Leibniz's rule, so that, expanding, (3.5.3.4) becomes (3.5.3.5) Z (Pkl)[pk(f(1))Lie(P)e(a(2))
k+~=p-l,8~:O
_
pk (f(2)) Lie (p)e (or(1))] ~ dr(E r
By (3.5.1.2), we triay substitute (3.5.3.6) Lie (F)e (a (2)) = Lie (p)e- 1 df(2) = dP t-~(f(2)) Lie (p)e (a (1)) = Lie (p)e_ ~ df(1)=dpe_l(f(1)).
Doing so, and remembering that (3.5.3.7) we must show (3.5.3.8) k+e=p-l,e*0 ( P k 1) ---(- 1)k mod p for Ogk<_p-1
~,
( - t)k [p~f(1) dP e-: f(2) -- Pk (f(2)) de e-' f(1)] e dr(V, (gv)
Re-indexing the summation by k and r e = Y - 1 , (3.5.3.8) becomes (remembering that ( - 1)k+l - ( - !) mmod p) (3.5.3.9) ( - 1)k nkf(1) dn"f(2) k+m=p-Z + ~ ( - 1)"1Pk(f(2)) de'f(l)edF(V, (gv).
k+m=p--2
This is the case; in fact, the left member of (3.5.3.9) is (3.5.3.10) d( E

k+m=p-2
( _ 1)kpk(f(1)) P'(f(2))).
This proves (3.5.0.3) in case b = 2, and gives a hint of the combinatorial rearrangements necessary in the general case. (3.5.4) We now turn to the general case. We adopt the convention that a product indexed by a subset of Z is to be taken in increasing order, and
A l g e b r a i c S o l u t i o n s of Differential E q u a t i o n s
53
that, unless otherwise specified, all indexing variables run over the set
{1, ...,hi.
We must verify the congruence (3.5.0.3) in the case (3.5.4.0) ~ =-c (1)/x ~ (2) A... /x ~ (b) = I ] ~ (v).
v
Using the substitutions (3.5.1) and the identity (3.5.2.3)just as in the case b = 2, we see that the congruence in question is equivalent to the congruence Lie (P)e- 1(2 ( - 1)'+~f(i) I-I tr(v)) (3.5.4.1)
i v*i
---~ ( - 1)'+l
i
PP-l(f(i)) H t7(v) modulo dF(V, ~?~/s(log O~)).

v*i
Expanding the left member of (3.5.4.1) be, Leibniz's rule, and remembering (3.5.3.7), (3.5.4.1) becomes
p--1
( - 1)`+' ( (3.5.4.2) '
1)kPP-~-k(f(i))
v*i
k=l 9Lie (p)k( [ I a (v)) ~ dF(V, f~/s(log Dr) ) .
Our next task is to expand each of the terms (3.5.4.3) Lie (p)k ( 1-I a (v))
v~:i
using Leibniz's rule. To facilitate this, we introduce the notations (3.5.4.4) (3.5.4.5) (3.5.4.6) (3.5.4.7) t~= (t'l ..... (b), a b-tuple of non-negative integers Ill = ~ (, ([~1) = (IfD'
i
(~)'
S(f) = {ilf~.O}
(3.5.4.8) for any nonempty subset A c { 1 .... , b } - { i } we denote by sgn (A, i) the sign of the permutation A, {1 ..... b } - {i}-A~
{l, ..., b}- {i}
(We make the convention that subsets of Z are to be enumerated in increasing order.)
54
N.M.
Katz:
Returning to the expansion of (3.5.4.3), we find Lie (P)R( l-I a(v)) =

v*i
(1~}) l-I Lie(P)e~(~r(v))

e; lel = k, S t O = a \ t , /
~'; [ l I = k , i C s ( O x " / v * i
(3.5.4.9)
r * a = {~ ,..-, b } - {i1
"111 Lie(P)e~(tr(v)) H atv).

vezl vC~du{i}
Substituting (3.5.4.9) into (3.5.4.2), we are faced with proving that dr(~ Obs(log D~)) contains E (-- 1)i+1 ~ ( - 1)kPP-l-k(f(i)) ~, sgn(A, i) ~,z~i e; Iel=k.s(e)=~\ f ] (3.5.4.10) i k=I 91--[Lie(P)e~(a(v)) [I a(v).
YeA vCdu{i}
,,_1
UL,
The key point now is to reindex the expression (3.5.4.10) by the subsets A'=A w {i} having at least two elements; then (3.5.4.10) becomes
p-1
~ ( - 1)i+1 sgn(A'-i, i) ~ (d' i~d' k~l
1)kPP-l-~(f(i))
(3.5.4.11)
g; [gl = k, S(g) = a ' {i}
v ~ a ' - {i}
yea'
We now calculate the sign(-1)l+lsgn(A'-i,i) as a function of the position of i in A' and of the position of A' in {1.... , b}.
(_ 1)i+lsgn(A,_i,i)=sg n (i, {1.... ...,b}-{i}]] sgn \(A'-i'{1 {1,, {i}""b}-A') 1, , b .... b } =sgn (3.5.4.12) (i,{1 .... ,b}-{i})sgn(i,A'-i,{1,...,b}-A'] 1,...,b i,{i ..... b}-i /
1. . . . . b ]
=sgn (i, zl'- {i}, {1.... , b} - A ' ]
/i,A'-{i},{1 .... ,b}-d'\

=sgn ~ d',{1 .... ,b}--d'
'
)sgn
(A', {1.... ,b}-A'] 1.... ,b /
{1 . . . . , b } - A " ~
Because each tr(v) is a closed form, in order to show that (3.5.4.11) lies in dF(V, f2~/s(log D~)), it suffices to show that, for each subset zl~C2 {1,..., b}
55
of two or more elements, dF(V, t2~./s(log D~)) contains (3.5.4.13) i~a'
~sgn(i,A'~{i}) p-1 ( - I f PP-l-k(f(i)) ~. [Ifl~ E k=l e;let=k,Ste)=A'-{i}~ E ]

[I
ved'-{i}
Lie (P)ev (a (v)) .
With no loss in generality, we may suppose A'= {1, 2 .... , b}, where b = ~(A'); then (3.5.4.14)
sgn (i' A'A, {i}) = ( - l) '+~.
Let us introduce a final notation: (3.5.4.15)

E m i = p.
m=(mt ..... mb), a b-triple of integers satisfying mi>l, and
A multi-index E occurring in (3.5.4.13) with Il=k and S(~)=A'-{i} gives rise to such an m by putting (3.5.4.16) mv=J~v,
(p-k
if v4=i if v=i.
In terms of m, the multinomial coefficient ( - 1)k (IEI) is easily calculated:
(_l)k I I _ k ! ( p - l - k ) !
(3.5.4.17)
v
(dr)
(p-1),
fl!...fb! = ml!...mb!
k!
(p-1), (p-k)
= 1-I (rod! mi modulo p.

Using (3.5.4.14) and (3.5.4.17), the congruence (3.5.4.13) may be rewritten -1 I-I (m~)! ~ ( - 1)i miPm'-l(f(i)) (3.5.4.18)
9
-1
v*i
I-[ Lie(P)'v(a(v))edr(V,, OLs(l~ D~)).
The proof of (3.5.4.18), and hence of (3.5.0), will be concluded the following calculation. (3.5.5)
Calculation. For any m as in (3.5.4.13),

(-- 1)'+1 m i P"-~(f(i)) I-I Lie(P)m~(a(v))
i v*i
(3.5.5.0)
= d (P"~ -1 (f(1)) ~ ( - 1)i+1 rn, P " ' - ~(f(i)) 1-[ Lie (P)"~ (a (v))).
i~kl v#l,i
56
N.M. Katz:
Proof According to (3.5.1.2), we have

(3.5.5.1) Lie (P)(a (v)) = dr(v) Lie (p)mv(a (v)) = Lie (p)mv-1 (df(v)) = d (pm~-I (f(v))). whence for any integer my > 1, we have (3.5.5.2)
Let us expand the second member of (3.5.5.0), remembering that the a(v), and hence the Lie(P)m~(a(v)), are closed forms.
d(P m'-l(f(1)) ~ ( - 1)i+1 m~em'-~(f(i)) 1-[ Lie(p)mv(a(v)))

i*l
v*l,i
(3.5.5.3)
=Lie(p)m'(a(1))~(-1)i+lmipm'-l(f(i)) ~I Lie(p)mv(a(v))
i#l v*l,i
+ e='-'(f(1)) E ( - 1)'+1 rn, Lie(P)"'(a(i)) 1-I Lie(p)m"(a(v))"

i.1 v*l,i
The first term in (3.5.5.4) is the part of the first member of (3.5.5.0) corresponding to i . 1. Thus it remains to see that (3.5.5.4) ml P"'-x(f(1)) I-I Lie(p)m'(a(v)) v*l = P"'- '(f(1)) Z ( - 1)i+~ m, Lie (P)"'(a(i)) 1-I Lie (P)m~(a(v)).
i*l v*l,i
Rearrainging the final product in the second member, (3.5.5.5) may be rewritten ml Pro'- l(f(1)) 1[I Lie (P)mv(a (v)) (3.5.5.5) ~*l
= --Pm~--l(f(1))Z mi[I Lie(p)m"(a(v)) 9

i.1 v:4:l i
Because ~ mi=p, ~ mi=-O (p), and (3.5.5.6) is true. This concludes the proof of (3.5.5), hence of (3.5.0), and hence of Theorem 3.2 and its Corollaries 3.3 and (3.3.1). Q.E.D.
4. Review of Deligne's Mixed Hodge Structures ([8]) 4.0. The Weight Filtration
(4.0.1) We return to the geometric situation (1.0). The weight filtration of the complex f2~c/s(log D) is by definition the finite increasing filtration defined by (4.0.1.0) W~(~2~/s(logD)) = image of f2~/s(log D) | f 2 ~ f ~ f2~/s(lOg D). Notice that (4.0.1.1) W~A Wjcz Wi+j
57
The "Poincar6 residue" (cf. [7, 8]) furnishes isomorphisms (which are ,, canonical" only after choosing an ordering of the finite set which indexes the smooth divisors Di) O if n<O (4.0.1.2)
grW(f2}/s(logD~= J il<ff<inf2.-, IJ / c~ Dil
. . . . . Oln/S
if n > 1
t t?}/s
if n = 0.
[We are guilty of an abuse of notation in (4.0.1.2) above, because, strictly speaking, gr,, is isomorphic to the direct sum of the extensions by w zero from D i l n . . . ~ D i . to X of the complexes f2g~,~" oi./s]. To avoid .... either further abuses or typographical disaster, we will henceforth adopt the following notation: eat f(i 1.... , i,) denote the composite
Diln...nDi c--~Dc--,X
(4.0.1.3)
s.
IS
S
. . . . Din/S) ,
and let us abbreviate, for each integer k > 0, (4.0.1.4) HkR(DilA...ADi./S) d~ n Rkf(il, ...,in),(~Diln"
which is a quasicoherent sheaf of d?s-modules. The spectral sequence of the filtration W and the functor R f, may be written, via (4.0.1.2), as (4.0.1.5)
wE;"'k+"=
9
il < ' " < i n
Hk-R"(Di,n ' " n D i . / S ) ~
Rkf.(t2;x./s(logD)).
(4.0.2) Proposition. The spectral sequence (4.0.1.5) is (naturally) a spectral sequence with cup-product in the category of quasi-coherent (gs-modules with integrable connection relative to T. This connection is the Gauss-Martin connection on the El terms and on the abutment. In particular, the W filtration on the Rkf,(Q}/s(logD)) is horizontal for the Gauss-Manin connection.
Proof. The product structure in the spectral sequence results from (4.0.1.1). The action of the Gauss-Manin connection on the spectral sequence may be seen directly as follows. The question being local on S, we may suppose S affine. Choose an affine open covering of X by sufficiently small coordinatized open sets Vv as in (3.4.1). The filtration W of f2~/s(logD) induces a filtration W on the Cech bicomplex of quasi-coherent (gs-modules
(4.0.2.0) W~C~
t2}/s(logD)) dN,, C~
W~f21/s(logO) )
58
N.M. Katz:
whose assosciated graded is given by (4.0.2.1) gr,w C~ V~},t2~/s(log D))
~
il<'"<in
C~
.... Di,/S)"
The spectral sequence (4.0.1.3) is the spectral sequence of the filtered complex of quasi-coherent (~s-modules obtained from the filtered bicomplex (4.0.2.0) by totalization. The (not necessarily integrable) T-connection V on the Cech bicomplex C ~({ Vv},I2~/s (log D)) constructed in (3.4.1) preserves the filtration W. This provides a T-connection on the spectral sequence (4.0.1.5), which is that of Gauss-Manin on the abutment. To see that it is that of Gauss-Manin on the E1 term, we simply note that on the assosciated graded for W, (4.0.2.2) . C ~({ V~n Dil n...
n
Di.}, ~-2~)il. . . . Di,/S) .
it induces the same T-connection which the method of (3.4.1) would construct, viewing the affine open sets Vvn Dil n... n Din as a coordinatized open covering of the smooth S-scheme Di~n...nDi, (the coordinates being those on V~ whose restriction to Di~ n . - . n Di, do not vanish). Of course, the truth of (4.0.2) could also be perceived by "pure thought", by contemplating the definition of the Gauss-Manin connection in terms of the Koszul filtration (cf. (1.4)). We leave it to the reader. Q.E.D. (4.0.3) Corollary. Suppose also that T is (the spectrum of) a field of characteristic zero, and that f : X--~ S is proper. Then for every integer r>__l, the E r terms of (4.0.1.6) are locally free (gs-modules of finite rank, and their formation commutes with arbitrary change of base S'--~ S. Furthermore (4.0.1.6) is degenerate at E 2 (and hence by (2.2.1.11) remains so after an arbitrary base change).
Proof. Because f : X ~ S is proper, (4.0.1.6) is a spectral sequence of coherent d~s-modules with integrable connection relative to T. Because T is the spectrum of a field of characteristic zero, any coherent tiTs-module with integrable connection is a locally free (gs-module of finite rank. Thus in particular the Er terms of (4.0.1.6) are locally free for each r ~ 1. To prove that their formation commutes with arbitrary base change, we first remark that it's true for r = 1, by the usual base-changing theorems. Suppose inductively that it's true for all r < ro. We may universally P,q identify E,o+l with the first cohomology of the complex of locally free d~s-modules of finite rank with integrable T-connection
(4.0.3.0) /zp-~o,q+~o-1 dro~E,~q__~ Effo+,O,q+~-~o ~J'O
A l g e b r a i c S o l u t i o n s of D i f f e r e n t i a l E q u a t i o n s
59
Because the d,o are horizontal, all the cohomology groups of this complex are coherent d0s-modules with integrable T-connection, hence are locally free, hence their formation commutes with arbitrary change of base. To prove degeneration at E2, it suffices, by the above, to prove it after an arbitrary change of base S'--. S where S' is the spectrum of a field (because S is reduced). But when S is the spectrum of a field of characteristic zero, the degeneration is proved by Deligne in [8]. Q.E.D.
4.1. Topological Interpretation

Suppose, in addition to the hypotheses of (4.0.3), that T=Spec(C). By the regularity theorem (cf. [7, 11, 24]), the Gauss-Manin connectioo on the terms of the spectral sequence (4.0.1.6) has "regular singular points". By the fundamental comparison theorem of [7], the functor "germs of analytic local horizontal sections" from the category of coherent sheaves on S with integrable connections relative to C, with regular singular points, to the category of local coefficient systems of finite-dimensional complex vector spaces on the "underlying" analytic space Sa" (4.1.0)
(M, V),~(M |
~s
v'o,
is an equivalence of categories. This means that we "know" the spectral sequence (4.0.1.6) once we know the assosciated spectral sequence of local coefficient systems on Sa" to which it gives rise via the functor 4.1.0. (4.1.1) Proposition. The spectral sequence in local systems on S a" corresponding to (4.0.1.6) is the d~calage (cf (2.2.2.4)) of the Leray spectral sequence (relative to the base S an) of the open immersion of San-schemes
U an = X an _
Dan ~ "
s~" , X an n y.n
(4.1.1.0)
in complex cohomology:
(4.1.1.1)
E~'q = RPfg n(R qj , (C)) ~ g p+" zr~ (C) n
(which consequently degenerates at E3). Proof We will first reduce the problem to an analytic one. Notice that the entirity of Section 4.0 may be repeated word for word in the category of analytic spaces, by considering the filtration W on (fl~/s(log O))an.
60
N.M. Katz:
The corresponding spectral sequence in locally free (gs.n-modules of finite rank ty-n,k+n__ nk-nIr~an Flan/~'an)
Wt-'I (an) -(~
DR ~xlil n , . . n L " i ~ / ~ !
(4.1.1.2)
i,<..-<i,
Rkf,"((a~/s(log O)) a")

is of formation compatible with arbitrary change of base S' ~ S an in the category of analytic spaces, thanks to (4.0.3). The canonical morphism of spectral sequences (4.1.1.3) (4.0.1.6)--~ (4.1.1.2)
induces horizontal (for the Gauss-Manin connections (4.0.2)) morphisms (4.1.1.4)
E, | (gs~, ~ Er(an).
d)s
In fact, the morphisms (4.1.1.4) are all isomorphisms, because source and target are locally free (_gsa,-modules of finite rank of formation compatible with arbitrary change of base, and because (by GAGA) the morphism (4.1.1.3) of spectral sequences is an isomorphism when S=Spec(C). Thus the analytic spectral sequence (4.1.1.2) gives rise, via the functor "germs of local horizontal sections" (which is an equivalence of categories between coherent sheaves on (gsa. with integrable connection and local systems of finite-dimensional C-spaces on S an) to the same spectral sequence in local systems on S an as does (4.0.1). It will be convenient to think of this spectral sequence in local systems as a sub-spectral sequence of (4.1.1.2). (4.1.1.5) Consider the "absolute" log complex (12]/c(log D)) an, together with its weight filtration W,, defined as in (4.0.1.0). The spectral sequence of the filtration W and the functors R f , " ,
wE{"'k+"(abs)=
(4.1.1.6)
|
i...... i.
Rk-"f(ix,..., i,),n(C)
Rkf, n((I2x/c (log O)) an)

is a spectral sequence of sheaves of C-vector spaces on S a". [We have used Poincar6's lemma on the Di, n . . . c~ Di, to identify the El term.] The natural mapping of absolute to relative log complexes
(4.1.1.7)
(I2]/c (log O)) an --~ (0~/s (log O)) an
is compatible with the W-filtration, and thus induces a morphism of spectral sequences (4.1.1.8) (4.1.1.6) ---, (4.1.1,2).
61
(4.1.1.9) Lemma. The morphism (4.1.1.8) of spectral sequences induces an isomorphism between (4.1.1.6) and the sub-spectral sequence in local systems of (4.1.1.2) obtained by taking germs of local horizontal sections.
Proof Let's denote by E~(an) the terms of (4.1.1.2) by E,(an) v their sheaves of germs of horizontal sections, and by E,(abs) the terms of (4.1.1.6). It follows from the definition of the Gauss-Manin connection in terms of the Koszul filtration (cf. (1.4)) that the canonical morphisms E,(abs)-* E,(an) factor through E~(an)v. It remains only to prove inductively that the mappings E, (abs) -* E, (an) v are isomorphisms. For r = 1, this is proven in [6]. Below we will indicate another conceptual proof (cf. (4.1.2)). Suppose the result for all r < r o. Because the functor "germs of horizontal sections" is exact, E,o+~(an)v is the first cohomology sheaf of the complex
(4.1.1.10)
E~o(an)v dv~
' E,o(an) v dr~
E~o(an)v.
By induction, this complex receives isomorphically the complex (4.1.1.11) Ero(abs ) ar~ ,E,o(abs) a*~
whose first cohomology sheaf is E,o+l(abs ). This implies that Ero+l(abs) - ,Ero+l(an) v,
and proves the lemma. We may now conclude the proof of (4.1.1) by noting the filtered quasiisomorphisms (cf. (2.2.2.1) and [8], (3.1.7.1) and Prop. (3.1.8)) ((Qi-/c (log D))"n, W), (4.1.1.12) ' ((Q]/c (log D))"n, z < _ _.) l ( j , (t]~/~), z =< .). Q.E.D.
In the course of the above, we made use of the following fact, applied to X a" and the intersections/)anc~ "" c~ D an over S an, in order to prove that -iL i, (4.1.1.9) was an isomorphism for r = 1. (4.1.2) Proposition. Let f: Ys 5r be a proper and smooth morphism of complex manifolds. The canonical morphism of sheaves on 5~ (4.1.2.0) Rqf,(C) ~ , Rqf,(t2~/c) ,Rqf,(a~/~)
62
N.M. Katz:
is an isomorphism between its source and the subsheaf Rq f,(f2~c/~) v of germs of horizontal sections (for the Gauss-Manin connection) of
Rq/, (f2~./~).
Proof The proof is an exercise in the definition of the Gauss-Manin connection (cf. (1.4) and [-35]). We consider the Koszul filtration K (of. (1.4)) of the complex s2]~/c, and the corresponding spectral sequence for the functors R f ,
(4.1.2.1)
Ef'q=RV+q f,(gr~ ~2~c/c)=~.RP+qf,((2~r/c).
By (the analytic version of) (1.4.0.2), we have (4.1.2.2) Fp, ~--- ov~ / C | Rqf, (f2}/y). L'I ~5
By definition of the Gauss-Manin connection (cf. [35]), the differential
(4.1.2.3)
is the mapping (4.1.2.4)
dr, q: E f ' q ~ E f + l ' q
d|
plQV
deduced from the Gauss-Manin connection V= d ~ Because the ROf, (f2]./~) are coherent sheaves on the complex manifold they are locally free of finite rank, and the canonical mapping (4.1.2.5) (_9~| c Rq f , (Q~/~)v__~ Rq f , (~2~:/~)
is an isomorphism. Thus we have an isomorphism (4.1.2.6) in terms of which (4.l.2.7)
El" ~~- f2)/c |
Rq f , (f2~c/~)v
d('q=d|
As Rqf, (fl;c/s~)v is a sheaf of C-spaces, it is automatically fiat over C, and hence we have (4.1.2.8) v.q,,~ 9 9 Ez - ~ v (f2~/c) @cR qf , (f~r/~) v9
By the Poincar6 lemma, (4.1.2.9) Thus we have (4.1.2.10) ~f'v (f2~/c) = if p:60.
EV.q_fRqf,(f2~/~) v if p = 0
2-)0
if p#:0.
Algebraic Solutionsof DifferentialEquations
63
Consequently, the spectral sequence (4.1.2.1) is degenerate at E 2, and the canonical mappings (4.1.2.11) are isomorphisms. Rqf, (O~/c) ~ E~ q --, E ~ Q.E.D.
4.2. Families of Mixed Hodge Structures

(4.2.0) We recall that a pure Hodge structure over Z of weight n, H, is by definition, a Z-module of finite type Hz, together with a (Hodge) filtration F of the complex vector space H c d ~ n H z | which satisfies the condition (4.2.0.0) Hc= ~
p+q=n p, q E Z
FP~ffq(Hc)
(ffq denotes the complex conjugate of F q. Complex conjugation, and indeed all of Aut(C), acts on H c = H z | through the second factor !). For all pairs (p, q) of integers with p + q = n, we put (4.2.0.1) H p' q = F p n Pq (Hc) ,
the "subspace of H c of type (p, q)". The bigraduation (4.2.0.2)

Hc= 0
p+q=n
HV'q;
H q'v=Hv'q
allows us to define an action s of C* as real group on H c by putting (4.2.0.3) s(z)=multiplication by zP~q on H p'q.
Following Well, we put C = s (i). For each z ~C*, the endomorphism s(z) of H c commutes with complex conjugation, and hence comes from an endomorphism still noted s(z) of H n = H z | R denoting the field of real numbers. (4.2.0.4) A morphism q9 between pure Hodge structures H and H' is a group homomorphism ~Oz: Hz---~H'z whose C-linear extension <Pc: Hc--* H~: commutes with the action s of C* (4.2.0.3). (Thus between pure Hodge structure of different weights there is only the zero morphism.) (4.2.0.5) The category of pure Hodge structures has an internal Hom and a tensor product | defined in the expected way (cf. [8]). (4.2.0.6) Tate's Hodge structure Z(n) is the rank one Hodge structure of weight - 2 n which is purely of type ( - n, - n), and whose integral lattice H z is the subgroup (4.2.0.7) (2n i)" Z c C.
64
N.M. Katz:
A polarization of a pure Hodge structure H of weight n is a homomorphism of Hodge structures (4.2.0.8)
(,)H|
such that the real bilinear form on H R (4.2.0.9) (2 g i)" (x, C y)
is symmetric and positive definite. The positive-definiteness means that (4.2.0.10)
if h6H p'q, h+O, then oP-q(2?r i)"(h, h)>0.
(4.2.0.11) The category of pure Hodge structures up to isogeny is the category whose objects are pure Hodge structures, but whose morphisms are defined by (4.2.0.12) Hom(up to isogeny)(H, H ' ) = H o m ( H , H ' ) |
(4.2.0.13) The full subcategory of the category of pure Hodge structures up to isogeny consisting of the polarizable objects (i. e., those which admit at least one polarization) is a semi-simple. It is closed under the formation of internal horn, tensor products, finite direct sums of objects of the same weight, subobjects and quotient objects. Any objects isogenous to a polarizable one is polarizable. (4.2.0.14) Proposition. Let (H, (,)) be a polarized pure Hodge structure. Then its group of automorphisms is finite. Proof Any automorphism +p determines an element Up in the group (4.2.0.15)
Aut(Hz/torsion)~Aut(Hlt,(2gi)"(x, C y))
which, being the intersection o f a discrete and a compact subgroup of Aut (H,), is finite. As ~o-+ Upis clearly a group homomorphism, it remains to prove its kernel is finite. But ~o~ Up is injective, unless H z = torsion. In the latter case, Aut (Hz) is a finite group. Q.E.D. (4.2.1) Let 5~ be a topological space. A family of pure Hodge structures of weight n on 5e is by defintion a local system H z of Z-modules of finite type, together with a continuously warying filtration F+ of (Hc)s, the i complexification of the stalk of Hz at s, which point by point is a pure Hodge structure of weight n. (4.2.1.1) A polarization of a family of pure Hodge structures of weight n is by definition a morphism of local systems on b ~ (4.2.1.2) (,): Hz|
65
which-point by point is a polarization (4.2.0.8). A family of pure Hodge structure is called polarizable if it admits at least one polarization. The considerations of (4.2.0.11-13) apply mutatis mutandis, in this context. (4.2.1.3) Proposition. Let 5P be a connected, locally arcwise connected, locally arcwise simply connected topological space (so that 6~' "has" a universal covering). Let (H z, F) be a polarizable family of pure Hodge structures. Suppose that the filtration F is locally constant, in the sense that it comes from a filtration by sub-local systems of the complexified local system H c. Then there exists a finite ~tale covering n: d,~'-* 6e such that the inverse image n* (H z, F) oj" (Hz, F) on ~ ' is a constant family of pure Hodge structures. Proof. Fix a point s o ~ 5<. The (topological) fundamental group 7q ( ~ so) acts on the stalk (Hz)so, and by hypothesis it preserves the filtration F~ioof (He)so. A polarization ( , ) on the family (H z, F) induces a polarization (,)so on (Hz)so, and n l ( ~ So) acting on (Hz)so preserves this polarization. Thus n l ( ~ So) acts on (Hz)so through the automorphism group of the polarized Hodge structure ((Hz)so, F~o,(,)so), which is a finite group (4.2.0.14). Thus there exists a finite &ale covering rt: 6P'-* 6~ such that n*(Hz) is a constant local system. Hence n*(Hc) is constant, and so necessarily are its sub-local systems re* F i(Hc). Q.E.D. (4.2.2) Let 6r be a topological space. A family of mixed Hodge structures on 6P is by definition a finitely filtered object in the category of local systems of Z-modules of finite type (4.2.2.0) (Hz, W); W~Hzc W~+IHz
together with a continuously varying Hodge filtration F~ on the comi plexifications (He)s, of the stalks of H z, such that, for each integer n, the filtration induced by the F~ makes grw Hz = IV, Hz/W,_ ~Hz into a family i of pure Hodge structures of weight n. (4.2.2.1) A morphism of families of mixed Hodge structures is a morphism of local systems which respects W and whose complexification respects the filtration F~ point by point. ~ (4.2.2.2) An essential fact (cf. [8]) about the category of families of mixed Hodge structures on a topological space 5~ is that it is an abelian category, and in particular that any morphism r is automatically strictly compatible with the filtration W (and also F, for that matter). This means that W~n image(q~)= q~(W0, for every integer i. (4.2.2.3) Proposition. Let 5~ be a topological space as in (4.2.1.3). Let (H z, F, W) be a family of mixed Hodge structures on 6~, such that each of the assosciated graded families (grw Hz, F) of pure Hodge structures is polarizable. Suppose that the Hodge filtration F is locally constant, in the
66
N.M. Katz:
sense that it comes ffom a filtration by sub-local systems of the complexified local system H c. Then there exists a finite Otale covering ~z: Sf'---~ 5t~ such that the inverse image z~*(Hz, W, F) of (Hz, W, F) on s is a constant family of mixed Hodge structures. Proof Applying (4.2.1.3) to each of the finitely many non-zero grw, we find a finite &ale covering 7t: 5P'--* 5~ on which the grw become constant families of Hodge structures. So replacing 5p by 5 '~', it suffices to show that if the grw are all constant, then H z is constant. The constancy of the grw signifies that under the action of zq ( ~ so) on (Hz)~o, for any 7 elq ( ~ So);
(4.2.2.3.0)
(1 - ?)(IV, (Hz)so) c ~ _ 1(Hz)so 9
By hypothesis ? preserves the filtrations W and F, hence T and 1 - 7 are endomorphisms of the mixed Hodge structure ((Hz)~o, W, F). Thus (4.2.2.2), 1 - 7 is strictly compatible with W. Combining (4.2.2.3.0) and strictness, we find (4.2.2.3.1) ( 1 - j ( W , ) c W,_~ c~image ( I - 7 ) = ( 1 - j ( W , _ 0 .
Since IV, = 0 for n,~ 0, and IV, = all for n >>0, this implies 1 - ? = 0, and hence Hz is a constant local system. Q.E.D. (4.2.2.4) Remark. Let us agree to call polarizable a family of mixed Hodge structures whose assosciated graded families grw are all polarizable. Because grw is exact, finite direct sums, sub-objects and quotient objects of polarizable families of mixed Hodge structures are polarizable. Clearly, any object isogenous to a polarizable one is polarizable.
4.3. Geometric Interpretation
(4.3.0) Let T=Spec(C), S a connected smooth C-scheme, f : X ~ S a projective and smooth S-scheme, and D = U D~ a union of divisors in X which are smooth over S and which cross normally relative to S
DC i 9 X~ j ~ U=X-D
(4.3.0.0)
Algebraic Solutions of Differential E q u a t i o n s
67
(4.3.0.1)
For each integer n>0, the sheaf R"n,"(Z) on San is a local system of Z-modules of finite type. As we have seen (4.1.1), the corresponding rational local system R" n," (Q) = R" n." (Z) | Q
(4.3.0.2)
is the abutment of the Leray spectral sequence in local systems on S a". (4.3.0.3)
E~'q =RP f , n(Rqj."(Q)) ~ R p+q 7z,n(Q),
which is degenerate at E3 (because by (4.1.1) its complexification is). (4.3.0.4) As a temporary notational device, let us denote by N i the decreasing filtration of R" rc,"(Q) defined by the spectral sequence (4.3.0.3). We define the weight filtration W of the local system R"rc.n(Z) by: (4.3.0.5) W/R"rc."(Z)=the inverse image of N2"-iR"r~.n(Q) under the an canonical mapping R ~rtan(z)-~ R n r~. (Q) (thus W/= 0 if i < n, and W/= all if i > 2 n). The locally free coherent sheaf on Sa"
g n ~ n ( C ) ~ ) c (gSan ~---Rnf:
(4.3.0.6)
n((f2~c/s(log D))an)
-~ R"f, (t2~:/s(log O)) |162 (~s.n
is the abutment of the Hodge =:- De Rham spectral sequence
E~'q= Rqf,"((t2~/s(log O)) an)
RV+qf, n((f2]/s(log O))a")
Rqf, (f2~/s(log D)) |162 ('0San
II
RP + if* (t2}/s (log D)) |162 (gsa~
which has E1 locally free, and which degenerates at E1 (cf. (1.4.1.8) and [-5]). The corresponding filtration F of R" Tt~"(C) | (gs.~ defines point by point a filtration F~ of the stalk
an ~ 9 (R n ~, (C))~-R nf, an ((t~x/s(logD))~")QCs..(t~s../m~),
where m~ denotes the ideal defining the point s6S"". (4.3.1) Proposition (Deligne-Hodge). The triple (R" rc,"(Z), IV, F) defined above is a polarizable family of mixed Hodge structures on S a". Proof That it is a family of mixed Hodge structures follows from the Deligne's theory, point by point. It remains to see that it is polarizable. Consider the Leray spectral sequence (4.3.1.1)
5*
E~'" =RPf,'(R"j, " Z) ~ R p+~ n," (Z).
68
N.M. Katz:
When tensored with Q, it degenerates at E3 and defines (a renumbering of) the filtration W on R"n,n(Q). Thus the assosciated graded families of Hodge structures grw of our family of mixed Hodge structures are isogenous to various of the E3 terms of the Leray spectral sequence (4.3.1.0) over Z. As these latter are sub-quotients (i.e., quotients of sub-objects) of the E2 terms of (4.3.1.0), it remains to polarize the E2 terms. In the notation of (4.0.1.4), we have ((-q) denoting | cf, (4.2.0)). Ev, q=Si,<.G..<iqRpfa"(i, ) ..... iq).(Z)(-q) if q=~0 (4.3.1.2) lR~'f," (Z) if q = 0. Because finite direct sums of polarizable families are polarizable, and Tate twists H ( - q) of polarizable families H are polarizable, we need only remark the' following proposition, applied to X and to all intersections Dii n... n Diq. (4.3.1.3) Proposition (Hodge's Index Theorem). Let f: X--~S be a projective and smooth morphism of C-schemes. Then for any integer n >-O, R"f, n Z is polarizable. Proof We may assume that X/S has geometrically connected fibres, and is of constant relative dimension N. Let LeHE(Xan, z) be the cohomology class of a hyperplane section, i.e. the inverse image under the composite pr2 o i XC i ' S x P c ~ ' 2 'Pc (4.3.1.4) ~ l p ' ~ S of the class of a hyperplane in H 2 (Pi n, Z). By the "hard Lefschetz theorem" ([1]), the iterated cup-product with L (4.3.1.5) /J: RN-i f,"(Z)(-- i)--~ RN+If,"(Z)
is an isogeny (i. e., becomes an isomorphism when tensored with Q). Thus it suffices to show that R"f.a"(z) is polarizable for n<N. For n < N, the primitive part of R"f," (Z) is defined by (4.3.1.6) Prim"f.~"(Z)= Kernel of LN-"+I: R"f,"(Z) --~ R2N-"+ 2f , n ( Z ) ( N - n + 1).
69
According to the "primitive decomposition" ([43], pp. 77-79), for n < N the mapping (4.3.1.7) t~
0 <- i < [n/21
Prim.-2if, n(z)(_i)
~L, , R " f , " ( Z )
is an isogeny. Thus it remains to polarize Prim"f,"(Z) for n < N . According to the Hodge Index Theorem (cf. [43]), the pairing (4.3.1.8) Prim"f," (Z) x Prim"f," (Z) ---*R 2n f , , (Z) ~ Z ( - N) (x,y)--~xwLN-"y, w = cup-product
is a polarization. This concludes the proof of (4.3.1.6) and of (4.3.1). (4.3.2) Interpretation. When S=Spec(C), the mixed Hodge structure (4.3.1) on H"(U an, Z) depends only on U, and not on the compactification U~---,X chosen to represent U as the complement in a proper smooth variety X/C of a divisor with normal crossings. This mixed Hodge structure is fimctorial in U, in the sense that if h: U -~ V is a morphism of smooth C-schemes, the induced morphisms h*: H"( V an, Z) -~ H"( U a", Z) are morphisms of mixed Hodge structures. (4.3.2.1) By Hironaka [18], any quasi-projective smooth C-scheme U may be compactified as above, and thus its integral cohomology H ' ( U an, Z) carries a functorial mixed Hodge structures. The family (4.3.1) of mixed Hodge structures is just the "interpolation" of these mixed Hodge structures on the fibres of ~: U = X - D - - * S . (4.3.3) Proposition. Hypotheses as in (4.3.0), the following conditions are equivalent, for any integer n > O. (4.3.3.0) The Hodge filtration F on the locally free sheafR"f. (f2}/s (log D)) on S is horizontal for the Gauss-Manin connection V. (4.3.3.1) The Hodge filtration F of the family of mixed Hodge structures on S a", (R" ~t,n(z), I41,F), is locally constant, i.e., it comes from a filtration of R" r~a. (C) by sub-local systems. ~ (4.3.3.2) There exists a finite dtale covering q): S'---~S such that (qoa")*(R"rt,n(z), W,F) is a constant family of mixed-nodge structures o n (S') a". (4.3.3.3) There exists a finite dtale covering q~: S' --* S such that (,o*(R"J. (t2]/s(log D)), V) is isomorphic to (((gs,)b", d) as a coherent (Ps,-module with connection (b, = rank of R"J. (f2~/s (log D))). Proof By (4.3.0.6), we have (4.3.3.0) ~ (4.3.3.1).
70
N.M. Katz:
By (4.2.2.3), (4.3.1), and "Riemann's existence theorem" (cf. [36] and the appendix) we have (4.3.3.1) ~ (4.3.3.2). To see that (4.3.3.2) ~ (4.3.3.3), notice that (4.3.3,2) implies in particular that the local system (~oan)*(R"n."(C)) is constant on (S')an, i.e., isomorphic to (Cb-)(s,),n. Thus, both q~*(R"f.(O~/s(logD)),V) and (((gs,)b",d) are coherent (gs,-modules with integrable connections having regular singular points (cf. 4.1) and they both give rise via (4.1.0) to isomorphic local systems on (S') ~". Hence they are isomorphic. In order to prove that (4.3.3.3)~ (4.3.3.0), we first remark that it suffices to prove that (4.3.3.0) holds after any base change ~o: S'-* S by a finite 6tale morphism, because (4.3.3.0) is of a differential nature, hence local on S for the 6tale topology. The hypothesis (4.3.3.3) implies that the local system R" n, n (C) becomes constant on a finite 6tale covering, hence also R" n.n(Q) and R" n.n(z)/torsion become constant. As the torsion in R"n,'(Z) is a local system of finite groups, it follows that R ' n . ' ( Z ) becomes constant on a finite 6tale covering ~0': S " ~ S, After making the change of base by such a finite 6tale q~': S" ~ S, we are reduced to showing that if R"n,"(Z) is a constant local system, then (4.3.3.0) (or equivalently (4.3.3.1)) holds. This is achieved by the following more general proposition. (4.3.4) Proposition. Hypotheses as in (4.3.0), the largest constant sublocal system of R" n." (Z) is a constant sub-family of mixed Hodge structures.
Proof In down-to-earth terms, this means the following:
(4.3.4.0) For any point seS, any element h~(R"n.n(c))s which is invariant under nl (S ~", s), and any homotopy class of paths ~,~,t from s to a second point t e S , If h~F~i (R , n. (C))s, then 7s, t(h)~Ft ~(R , ~r. (C)),, where " ,~ an 7~,,(h) denotes the element of (R"lr."(C))t deduced by transporting h along the path ~s,t. In order to establish (4.3.4.0), it suffices to do so far all pairs (s, t) of nearby points C nearby" in S~"), and for a single path 7~,, (because S"", being a complex manifold, is locally simply connected). We can always find a connected nonsingular affine curve C c S which contains a given pair (s, t) of nearby points. When we take the inverse image of our situation on C, our chosen element h 6(R" lg~n(C))s remains invariant under lgl(C an, s) because the action of this group factors through ~r1(S an, s). Thus it suffices to prove (4.3.4) when S is a connected smooth affine curve, in which case it follows from the more precise (4.3.5) Proposition. Hypotheses as in (4.3,0), suppose that S is a smooth connected affine curve. Then the canonical mapping of families of
71
mixed Hodge structures on S an (4.3.5.0)
(n n( O an, Z))san ~ R" 7~,n (Z)
has as image a constant sub-family of mixed Hodge structures. Its image is the largest constant sub-local system of R" ft. n (Z). Proof. The first assertion results from the fact that the category of families of mixed Hodge structures is an abelian subcategory of the category of local systems, and that the formation of images (as well as of kernels and cokernels) commutes with the inclusion of the category of families of mixed Hodge structures into that of local systems. The second assertion will result from the equality a.,~itTO'n--,n__ in the usual Leray ~'O 2 x., spectral sequence in integral cohomology of rt~": Ua"--~ San, which itself follows from the following proposition.
(4.3.6)
(4.3.6.0)
Proposition. Hypotheses as in (4.3.5), the Leray spectral sequence

an E~'q=HP(S an, Rq ~z, (Z)) ~ HP+q(U an, Z)
is degenerate at E 2. Proof. It suffices to show that E g'q = 0 unless p = 0 or p = 1. This is true, because the R q n."(Z) are local systems on S"", and because Sa" is an Eilenberg-Maclane space K (n, 1) with n = rq (Sa", s) a free group. This concludes the proof of (4.3.5), hence of (4.3.4) and (4.3.3) as well. Q.E.D.
4.4. We now Wish to Restate 4.3.3 "with a Group of Operators"

(4.4.0) Let G be a finite group. The indecomposable central idempotents in the rational group-ring Q [G] are obtained in the following way. For every irreducible representation X of G in a finite-dimensional C-space, let
(4.4.0.0)
P(Z)= deg(z) ~ trace(g(g-X)).g

~G g
denote the corresponding indecomposable central idempotent in C [G]. (4.4.0.1) As a runs over the group Aut(C), the irreducible representation ~(~ runs over a finite number of isomorphism classes of irreducible C-representations of G, and the projectors P(Xt')) (= tr. P(X), viewed as a function on G with values in C) run over a finite number of distinct indecomposable central idempotents in C [G]. We shall call the XC'~ (resp. the P(Xt'))) the Q-conjugates of x (resp. of P(z)).
72
N.M. Katz:
For each Q-conjugacy class A of irreducible C-representations of G, the element (4.4.0.2)

P(A) der'" ~ P(Z)
Z~A
ofC [G] lies in fact in Q [G], and is an indecomposable central idempotent in Q [G]. Every indecomposable central idempotent in Q [G] is of this form. We have (4.4.0.3)
p(A)P(A,)=~P(A)
if A=A' if not.
(4.4.0.4) Let Hz be a local system of Z-modules of finite type on which G acts. We define the G-sub-local system (4.4.0.5) P(A) H z de_~f'n inverse image in H z of P(A)(Ho) ~ HQ. the The canonical mapping of local systems (4.4.0.6)
(~ P(A) H z --~ H z
A
is a G-morphism and an isogeny, i.e., it becomes an isomorphism when tensored with Q. (4.4.1) If G acts on a family of mixed Hodge structures (Hz, W,F) (meaning that its action on the local system Hz respects W and F), then each P(A) H z is a sub-family of mixed Hodge structures. If (Hz, W, F) is a polarizable (4.2.2.4) family of mixed Hodge structures, so is each subfamily P(A) Hz. (4.4.2) Proposition. Hypotheses as in (4.3.0), suppose G is a finite group of S-automorphisms of X which respects the divisor D. Let n >0 be an integer A a Q-conjugacy class (4.4.0.1) irreducible representation of G in complex vector spaces, and P(A)e Q [G] the corresponding projector (4.4.0.2). The Jollowing conditions are equivalent. (4.4.2.0) The induced Hodge filtration F on the locally free sheaf POO(R"J,(Q'x/s(log D))) on S is horizontal for the Gauss-Manin connection, for every irreducible representation X in the given Q-conjugacy class A. (4.4.2.1) The Hodge filtration F of the family of mixed Hodge structures on S a", (P(A)R"n~,"(Z), W,F), is locally constant, i.e., comes from a filtration of P(A) R" n~n(C) by sub-local systems. (4.4.2.2) There exists a finite Otale covering q~: S ' - * S such that (r (P(A) R" n, ~(Z), W, F) is a constant family of mixed Hodge structures
on
(ST".
A l g e b r a i c Solutions of Differential E q u a t i o n s
73
(4.4.2.3) There exists a finite ~tale covering q~: S'--~ S such that, for every irreducible Z on the Q-conjugary class A, q~*(P(x) R"f,(O'x/s(log D)), V) is isomorphic to (((gs,)b"(x), d) as a coherent (gs,-module with connection (b,(z) = rank of P(z)(R" f , (Y2~:/s(log D)))).
Proof The proof of (4.3.3) applies almost verbatim. The implications (4.4,2.0)~,(4.4.2.1) ~(4.4.2.2) are proved as in (4.3.3), remembering that (P(A) R"~,"(Z), W, F)is polarizable (by (4.4.1) and (4.3.1)). The implication (4.4.2.2)=~(4.4.2.3) is proved as in (4.3.3), remembering that each (P(z) R"f, (Q}/s(log D)), V) has regular singular points, being a sub-object of (R"f, (O}/s (log D)), V). This final implication (4.4.2.3) =~ (4.4.2.1) may be reduced, as in (4.3.3), to the case in which the local system P(A) R" n,"(Z) is constant. But in that case, we have (P(A) being idempotent).
an (4.4.2.4) P(A) R n zc, ( Z )__ P(A). I where I = the largest constant sub-local system f R. . . . '~"
By (4.3.4), I is a constant sub-family of mixed Hodge structures, hence (4.4.1) so is P(A)I=P(A)R"~,"(Z), and in particular (4.4.2.1) holds. Q.E.D. 5. Applications to the Question of Grothendieck 5.0. A Global Situation (5.0,0) Let A be a subring of C which is finitely generated over Z, and put T= Spec(A). Let g: S---~T be a smooth morphism with geometrically connected fibres, and f: X ~ S a projective and smooth morphism. Let D = U Di be a union of divisors in X, each smooth over S, which have normal crossings relative to S. D
\
X
(5.0.0.0)
T (5.0.1) Let q / c S be an affine open neighborhood of the generic point of S over which each of the Hodge cohomology sheaves Re f , (~x/s) is a
74
N.M. Katz:
locally free module of finite rank. (Such a 0//always exists.) Then, over ~, the Hodge ~ De Rham spectral sequence (5.0.1.0)
Ef'~=Rqf, (s
(log D)) ~ R"+qf, (f2~/s (log O))
is degenerate at E 1. Let us recall why this is so. By Deligne's mixed Hodge theory ([8], 3.2.13), this spectral sequence degenerates at E~ over the generic point of q/. Because E1 is locally free over ~, the vanishing of dl at the generic point implies its vanishing on all of q/. Then E2 = E1 is locally free on q/, so d2 = 0 on q/ because dE=0 at the generic point of ~/... this proves inductively that all dr = 0 over Y/. By (2.2.1.11), after any change of base S'--, S which factors through ~ , the spectral sequence (5.0.1.0) will continue to have E1 locally free of finite rank, and to degenerate at E~. (5.1) Theorem. Hypotheses as in (5.0.0), let n>O be an integer. Suppose that there is some non-void affine open set ~ c S satisfying (5.0.1), and an infinite set $ of prime numbers p, such that the following condition (5.1.0) holds (5.1.0) For every point of Twith values in (the spectrum of) a finite field Fq of characteristic peS,, after the base change Spec(Fq)---~ T, D~ | , D~,
(5.1.o.o)
~|
Spec(Fq)
the p-curvature of the free relative to Fq
l 1
, ~-
, T
1 [ 1 1
,X
.,S with integrable connection
~D
(Qall(~AFqmodule
(5A.o.1)
(R"f, ((2~-/s(log D))tq/, V)@Fq
is zero, in other words, that the inverse image of R"f,(f2"x/s(logD)) on | Fq is spanned by its horizontal sections.
75
(5.1.1) Then after the change of base Spec(C)--* T,

Dc .-* D
Xc~ (5.1.1.0) Sc-----~S
X 1
Spec(C)
--, T
the coherent sheaf with integrable connection on S c
(5.1.I.1)
(R"fc,(~ic/sc(log D)), V
becomes trivial on a finite ~tale covering q~: S' --* S c, i. e., becomes isomorphic to (((gs,)~ d), where b,=the rank of R"f,(f~/s(logD) ) over the generic point of S. Proof. By (4.3.3), it suffices to prove that the Hodge filtration F on R"fc,(~}c/sc(log Dc)) is horizontal for the Gauss-Manin connection. For this it suffices to prove that the Hodge filtration is horizontal over ~c, since the obstruction to the horizontality is the (gsclinear KodairaSpencer mapping between locally free Csc-modules (cf. (1.4.1.8))
(5.l,2.0)
@ RqJc,(f2exc/sc(logDc)) o+q=" , _, @ a~Sc/c|

p+q=n
which cannot vanish over the open set q/c without vanishing on Sc. We will in fact prove that the Hodge filtration is horizontal on R"f, (O}/s(log D))lq/, or equivalently that for any section D of Der(ql/T), the (Pc-linear Kodaira-Spencer mapping between free d)e-modules (5.1.2.1) @ RqJ,(I2ets(logD))l~
p+q~n
@ Rq+l(f2~,Tsl(logD))lqJ
p+q=n
vanishes. Let q/=Spec(B). By (3.3.1), the hypothesis (5.1.0) implies that for every point of T with values in (the spectrum of) a finite field F~ of characteristic peX, the inverse image on B | q of the mapping p(D) is
76
N.M. Katz:
zero. This implies that for every maximal ideal m of B such that B/m B has characteristic p~2;, the inverse image on B/m B of the mapping p(D) vanishes, i.e., all the matrix coefficients of p (D) lie in m. Because B is a finitely generated integral domain whose fraction field is of characteristic zero, it follows immediately from Noether's normalization theorem that for any infinite set 2" of prime numbers, the intersection of all maximal ideals of B with residue characteristic in 2" is reduced to zero. Thus the Kodaira-Spencer mapping (5.1.2.1) vanishes. This concludes the proof.
5.2. A Global Situation with a Group of Operators
(5.2.0) Hypotheses as in (5.0.0), let G be a finite group, which acts as a group of S-automorphisms of X which are stable on D. Let g be the least common multiple of the orders of the elements of G, and let Ao= Z [l/g, ~g], where ~g is a primitive g-th root of unity. We suppose A ~ A o. (5.2.1) As is wellknown (cf. [38]), every irreducible representation of G in a finite dimensional complex vector space may be obtained by extending scalars from a representation of G in a locally free Ao-module of finite rank. (5.2.2) By (3.2.2 bis), if )~ is any representation of G in a locally free Ao-module which is irreducible over C, its "reduction" modulo any maximal m of Ao, z(m), is absolutely irreducible over Ao/m. Letting a denote any automorphism of C which extends the Frobenius automorphism corresponding to m of Ao, we have f f ( m ) = x ( m ) (p) (cf. (3.2.1)). In particular, (m) and Z (m) ~p)are obtained from irreducible representations in the same Q-conjugacy class (cf. (4.4.0.1)). Replacing the reference (4.3.3) by (4A.2), the proof of yields: (5.3) Theorem. Hypotheses as in (5.2.0), let n>O be an integer, and let A be a Q-conjugacy class of irreducible C-representations of G. For each z6A, let P(Z) denote the corresponding projector in Ao[G ] (cf (4.4.0), (4.4.0.1)). Suppose that there is a non-void affine open set ~llc S satisfying (5.0.1), and an infinite set Z of prime numbers, such that the jollowing condition holds. (5.3.0) For every point of T with values in (the spectrum of) a finite field Fq of characteristic p~s after the base change Spec(Fq)---~T (cf the diagram (5.1.0.0)), the p-curvature of each of the free (9~| q modules with integrable connection (5.3.0.1)
(P(z) R"f,(f2]/s(log O))[q/, V)|
z~A
is zero.
77
Then after the change of base Spec(C)--~ T (cf. the diagram (5.1.1.0)) each of the coherent sheaves with integrable connection on Sc (P(x) R"fc.(t2~Cc/sc(log Oc)) V),
x~A
becomes trivial on a finite &ale covering q9: S' ~ Sc, i, e., becomes isomorphic to (((gs,)b"(z), d), where b,(z) is the rank of P(X) R'f,(f~/s(log D)) over the generic point of S.
5.4. The Birational Point of View (5.4.0) Let Sc be a connected smooth C-scheme, and (Me, 17c)a locally free sheaf of finite rank with integrable connection on Sc. (5.4.1) Remark. In order that there exist a finite 6tale covering of Sc on which (Me, Vc) becomes trivial (as locally free sheaf with integrable connection), it is necessary and sufficient that there exist a finite extension K of the function field L of Sc such that (Me, Vc)| K is trivial (i. e., so that M c | is spanned over K by horizontal sections). (5.4.2) ProoJl The necessity is obvious. For sufficiency, notice that by hypotheses (M c, Vc) becomes trivial on a finite &ale covering Y/~:of some nonvoid affine open set y/c c Sc. Thus (Me, Vc) on Y/~:has regular singular points, hence on Y/c, hence on S c (cf. [7], 4.1). So it suffices to prove that the associated monodromy representation of nl (Sc, s) on the C-space Ms of germs of holomorphic horizontal sections at s factors through a finite group. Taking s~y/c, we have by hypothesis that the composite nl (y/c, s) -~ nl (Sc, s) -* Aut (Ms) factors through a finite group. As nl(y/c,S)--~nl(Sc, s) is surjective, it follows that the monodromy representation of nl(Sc, s) factors through a finite group. Q.E.D. (5.4.2.1) When either of the equivalent conditions of (5.4.1) holds, we say that (Me, Vc) has a full set of algebraic solutions. (5.4.3) Fix an infinite set 2 of prime numbers. Let A be a finitely generated (over Z) subring of C, T=Spec(A), Y/--~T an affine smooth morphism with geometrically connected fibres whose "complex fibre" y/c "is" an affine open subset of Sc, and (M, V) a free (9~-module of finite rank with integrable connection relative to T, whose inverse image on y/c "is" (Mc, Vc)ly/c. Consider the following property: (5.4.3.1) There is an affine open set 3 v c y / s u c h that for every point of T with values in (the spectrum of) a finite field Fq of characteristic peZ,
78
N.M. Katz:
after the base change Spec(Fq)--~ T VQFq>~cJ#
Spec(Fq)-
. , T
the p-curvature of the free (_9~| Fq-module with integrable connection is zero. ((M, 17)1~/)| F0
(5.4.3.2) A standard "passage to the limit" shows that if the property (5.4.3.1) holds for one set of choices (A, q/, (M, V)) then it holds for every choice. It is thus an intrinsic property of the germ of (Mc, Vc) at the generic point of Sc which we call having p-curvature zero for almost all primes peS. Thus, given a function field K/C (i. e., a finitely generated field extension of C), a differential equation (N, V) over K (i. e., a finite dimensional K-space N together with an integrable connection V: N ~ f 2 ~ / c Q K N ), and an infinite set 27 of prime numbers, it makes sense to say that (N, V) has or has not p-curvature zero for almost all primes pc27. (5.4.3.3) Because p-curvature is a differential invariant, its vanishing is of a local nature for the 6tale topology. It follows that (N, 17) has pcurvature zero for almost all p~27 if and only if there exists a finite algebraic extension L/K such that (N, V)| L has p-curvature zero for almost all peZ. K Putting together (5.4.1) and (5.4.3.3), we have the obvious (5.4.4) Proposition. Let (N, 17) be a differential equation over a function field K/C (5.4.3.2). If (N, 17) has a full set of algebraic solutions, then it has p-curvature zero for almost all primes p. (5.4.5) Let K/C be a function field, and let Ux be a smooth quasi-projective K-variety. By Hironaka ([18]), we can find a projective and smooth K-variety Xx which contains UK as an open set, such that the complement DK = XK UKis a union of smooth divisors D~,~ in XK which have normal crossings. By "general nonsense", we can find a finitely generated subring A of C, a smooth Spec(A)-scheme S with geometrically connected fibres such that K is the function field of its complex fibre So, a projective and smooth S-scheme f : X--* S and divisors D~ in X which are smooth
-
Algebraic Solutionsof DifferentialEquations
79
over S and cross normally relative to S, such that the fibre of X (respectively of Di) over the generic point of Sc is XK (resp. Di, K). (5.4.5.1) The stalk over the generic point of Sc of the coherent (9smodule with integrable connection relative to T (5.4.5.2) (R"J, ((2~/s(log D)), V)
is the differential equation over K (5.4.5.3)
(H~R (UKIK), V).
Combining (5.4.1), (5.4.3.2), (5.4.4) and (5.4.5.1), we may restate Theorem 5.5 birationally: (5.5) Theorem (= 5.1 bis). Let K be a function field over C, U a smooth quasi-projective K-variety, and n > 0 an integer. Then the Jollowing conditions are equivalent. (5.5.1)
(HgR(U/K), V) has a full set of algebraic solutions
(5.5.2) There exists an infinite set Z of prime numbers such that (H~R(U/K), V) has p-curvature zero for almost all p~E. (5.5.3)
(H~R(U/K), F) has p-curvature zero for almost all primes p.
5.6. In the situation of (5.4.5), suppose that a finite group G acts as a group of K-automorphisms of UK. By Hironaka ([18, 19, 41]) we can find a compactification XK of UK as in (5.4.5) on which G acts as a group of K-automorphisms, so that UK~--~XK is a G-morphism. By "general nonsense", we can find A, S, X, D~ as in (5.4.5) such that G acts as S-automorphisms of X preserving D = U D~, in a way which "gives back" the action of G on X K by passage to the generic point of Sc. Thus we may restate 5.3 with a group of operators in a birational way. 5.7. Theorem ( = 5.3 bis). Let K be a function field over C, U a smooth quasi-projective K-variety, G a finite group which acts as a group of K-automorphisms of U, n>O and integer, and A a Q-conjugacy class (4.4.0.1) of irreducible complex representations of G. Then the following conditions are equivalent: (5.7.1) For each irreducible representation z~A, the z-parts of the de Rham cohomology (P (Z) HgR(U/K), F) has a full set of algebraic solutions. (5.7.2) There exists an infinite set ,Y, of prime numbers such that for every z~A, (P(z) H~R(U/K), V) has p-curvature zero for almost all primes p~Z. (5.7.3) For every zeA, (P(z)H~R(U/K), V) has p-curvature zero for almost all primes p.
80
N.M. Katz:
6. Applications to the Hypergeometric Equation 6.0. Relations with Ordinary Differential Equations (6.0.0) Let T a scheme, S a smooth T-scheme, and (M, V) a locally free (gs-module of finite rank with an integrable T-connection. Recall that the dual of (M, V), noted (37/, P), is defined by
(6.0.0.1) ~ / = Homes (M, Cs)
with the connection V defined by requiring that for local section m of M, rh of ~/, and D of Der(S/T), we have (6.0.0.2)
(V(D)(m), th> + (m, V(D)(rh)>=O((m, th>).
Iteration of (6.0.0.2) gives, for every integer n > 1, (6.0.0.3) D'((m, rh))= ~ (~)((V(D))"-i(m),(V(D))i(rh)).
i=0
Taking n =p in characteristic p, we find (6.0.1) Remark. Hypotheses as in (6.0.0), if Tis a scheme of characteristic p, then the p-curvature (cf. [24], 5.2) of (~/, V) is the negative of the transpose of the p-curvature of (M, V): (6.0.1.0)
<Ov(D)(m), rh> + <m, r162 (D)(th)> =0.
(6.0.2) Proposition. Hypotheses as in (6.0.0), suppose T is a reduced and

irreducible scheme of characteristic p. Let k denote its function field. Suppose S is a smooth T-scheme which is irreducible, and denote by K its function field. The following condition on (M, V) are equivalent.
(6.0.2.0) (M, V) has p-curvature zero. (6.0.2.1) (K/I,~z) has p-curvature zero. (6.0.2.2) The dimension over k . K p of (M | K) v is the rank of M. (6.0.2.3) The dimension over k . K p of (A4 | K) r is the rank of M. Proof By (6.0.1), (6.0.2.0)r so it suffices to show that (6.0.2.0) r (6.0.2.2). For this, we form the cartesian diagram S(p) o , S (6.0.2.4) lg~,~ T e~ 1~ ~T
Algebraic Solutionsof DifferentialEquations and "decorate" it with the relative Frobenius F: S--~ Stp)
S F ~, S(p) a ~S F ~, S(p)
81
(6.0.2.5)
T f~bs ~ T
The corresponding diagram of function fields is
K,
, k . KPt . . . .
K+__
\k.K p
o 61
k~ xp~x k
By Cartier's theorem (cf. [24], 5.1), (M, V) has p-curvature zero if and only if the canonical morphism (6.0.2.7)
F* (F, (MY)) --~ M
is an isomorphism. On the other hand, since the p-curvature may be interpreted as an (gs-homomorphism (6.0.2.8)
~: M ~ F*b~((2~S/T)|162 M
between locally free modules, it vanishes if and only if it vanishes over the generic point of S. Using Cartier's theorem over the generic point, the p-curvature vanishes there if and only if the canonical map of K-vector spaces (6.0.2.9) (m | K) v | K -* M
kKP
is an isomorphism. Thus it remains to prove only that (6.0.2.9) is an isomorphism if and only if its source and target have the same dimension. This is indeed the case, because (6.0.2.9) is always injective; in fact we have the apparently more general (6.0.3) Proposition. Let T be a scheme of characteristic p, S a smooth T-scheme, and (M, V) a quasi-coherent (gs-module with integrable Tconnection. The canonical mapping ( cf the diagram (6.0.2.7)) (6.0.3.0)
F, (M v) |
(gs = F* (F, (MY)) --~ M
is always injective. Proof The question is local on S, so we may assume S is 6tale over A~ via X1, ..., X,. Then (9s is a free F-X(~s~,~)-module with basis the monomials X wd~f'n xW'... X w" having 0 < Wi < p - 1. Thus we must show
6 Inventiones math., Vol. 18
82
N.M. Katz:
that whenever horizontal sections m w, one for each exponent-system W as above verify (6.0.3.1) then all mw= O. ; this has a meaning and put DW=[I i W~! because all Wi__<p- 1. For any two exponent V, W as above, we have Let
~ XW mw=O
W
in M
IWl=~ W~eZ,
(6.0.3.2)
Dv(Xw)=o DV (x v) = 1
g
unless W~> V~
~
for all i
~ Dv ( x w) = ( ~ I ( WVi ) ) X w-v i
if W~>V~ for all i.
Suppose that (6.0.3.1) holds, but that not all mw=O. Among exponent systems W having mw 9 O, let V have a maximal weight IV]. Let's apply V(D v) to (6.0.3.1): (6.0.3.3)
~ V(Dv)(x w mw) = O.
W
Because all the mw are horizontal, V(Dv)(x w row)= Dv (Xw) 9row, whence (6.0.3.3) becomes (6.0.3.4)
~ Dv (Xw) 9
W
=0.
By construction, IWI>IV[ implies row=0, while [Wi<[V! implies IVI--IWI,v . w implies DV(xW)=O, as follows from (6.0.3.2). Thus (6.0.3.4) reduces to
Dv(Xw)=o, and
D V(X v) my = 0
whence mv = 0, a contradiction. Q.E.D.
(6.0.4) Let T be any scheme, and S a scheme 6tale over A~- via a section X of (3s. Let (M, V) be a free ~s-module of rank n with integrable Tconnection, such that there exists a basis eo ..... e,_l of M in terms of which the connection takes the form
[7 (d-~-)(el)=ei+ 1
(6.0.4.0)
n l
for O _ < i _ n - 2
(e,,_O = ~aiei;
i=0
aier(S,(gs).
83
Let eo .... , ~,_ 1 denote the dual basis of the dual module AS/. A section (6.0.4.1) ~ fi ei, i=o f/ local sections of (9s
is horizontal for the dual connection V on M / f and only if its coefficients satisfy dfi dX =f/+x for O < i < n - 2 (6.0.4.2) d f n - 1 ,-1 Thus
dX
i=o
Y a,
(6.0.4.3) The projection A7/-*(9s defined by ~Jiki--*fo induces an isomorphism between (A~/) the sheaf of germs of horizontal sections of ~, M, and the sheaf of germs of sections of (9s which are annihilated by
V2- - Fi=0 a, -Z2If T has characteristic p, this isomorphism is F - 1((gs~p,)_linear. Combining (6.0.4.3) and (6.0.2), we find (6.0.5) Proposition. Let T be a reduced and irreducible scheme of characteristic p, S an irreducible scheme which is dtale over A~T by means of a section X of(9 s. Let k (resp. K ) denote the function field of T (resp. S). Let (M, V) be as in (6.0.4). Then (M, V) has p-curvature zero if and only if the field K contains n = rank M solutions of the ordinary differential equation (6.0.5.1) ( f ) = ~ ai
i=0
(f)
which are linearly independent over the subfield k. K p.

(6.0.6) Suppose further that k is a perfect field (then kKP=KP). For any discrete valuation v of K, let vo denote the induced valuation of K p. Clearly the ramification index e(V/Vo) (= the index of the value groups) is p. As K is a p-dimensional vector space over K p, it follows that if t is a uniformizing parameter in K vor v, the elements 1, t .... , t p-I form a basis of K over K p, and thus provide an isomorphism of KP-vector spaces (6.0.6.0) K--- KP(~ ... G K p.
ptimes
In terms of this isomorphism, the v-adic topology on K is just the p-fold product of the vo-adic topology on KP;
6*
84
N . M . Katz:
ordv (i~i fip ti) =mini(ordv(ff ti))

(6.0.6.1) = min i (i + p ordvo (f~P)).
Let Kv (resp. (KV)~o) denote the completion of K (resp. K P) with respect to the valuation v (resp. Vo). Then (6.0.6.0) gives (6.0.6.2)
Kv ~- (KP)vo0 " " 9 (KV)vo~- K | (KV),,o.

9
primes
KP
Since (KP)vo= KW, we have (6.0.6.3)
K~= K | K~.
KP
Let Soln(K) (resp. Soln(Ko)) denote the K v (resp. K p) vector space of solutions of the differential equation (6.0.5.1) which lie in K (resp. in Kv). Then (6.0.6.4) Soln (Ko) ~ Soln (K) | K~.
Kp
This is because the differential operator (6.0.6.5)
-~-
- ~ ai - ~ i=0
: K-* K
is KP-linear, and the differential operator d , ,-1 d i
is deduced from (6.0.6.5) by the (flat !) extension of scalars K p ~ KW. (6.0.6.7) Because we can "clear denominators" by multiplying by p-th powers, the spaces Soln(K) and Soln(K~) are in fact spanned by those solutions which lies in the valuation rings of K and K~ respectively, or indeed by solutions which lie in any subring of K (resp. of K~) whose fraction field is all of K (resp. K~). Putting this together with (6.0.5), we find (6.0.7) Corollary. Hypotheses as in (6.0.5), suppose k perfect, and let v be any discrete valuation of K. Then (M, [7) has p-curvature zero if and only if the complete field Kv, or its valuation ring 0~, contains n = rank M solutions of (6.0.5.1) linearly independent over K~. (6.0.8) Remark. The assumption that k be perfect is used to insure that for any discrete valuation v of K, noting Vo its restriction to k K p, we have ramification e(V/Vo)=p. When k is not perfect, it can happen that e(v/vo)= 1 (cf. Bourbaki, XXX-Algebra Commutative, chpt. 6, excercise 3b) to w8, p. 187).
85
6.1. The Hypergeometric Equation-Definition (6.1.0) For any scheme T, we will denote by 2 the standard coordinate on A~, and by S r the open subset of A~ where the section 2 ( 1 - 2 ) e F(A~, (9A~) is invertible.
For any sections a,b,c~F(T, Or), we define the hypergeometric module E(a, b, c) on Sr to be the free (9sT-module of rank two with base eo, el, and integrable T-connection
(6.1.1) V
( ff_~) (e0 =
( c - ( a + b + l)2) ab 2(1-2) el q A(1-2) e~
As explained in (6.0.4.3), the horizontal sections of the dual of E(a, b, c) over an open set q/of ST "are" the sections f~F(~l, (9~u)which satisfy the hypergeometric equation with parameters a, b, c (6.1.2)
d 2
~-abf
=O.
6.2. Theorem. Let T= Spec(A), with A a finitely generated subring of C, and let a, b, c~A. The following conditions are equivalent: (6.2.1) The hypergeometric module E(a, b, c) on S c becomes trivial on a finite ~tale covering of S c. (6.2.2) There exists a finite extension of the field C(2) over which E(a, b,c) becomes trivial. (6.2.3) The hypergeometric equation (6.1.2) with parameters a, b, c has a full set of algebraic solutions (cf. (5.4.2.1)). (6.2.4) The hypergeometric equation (6.1.2) with parameters a, b, c, has a finite monodromy group. (6.2.5) The inverse image of E (a, b, c) on C0.)/C has p-curvature zero for almost all primes p (cf. (5.4.3.2)). (6.2.6) The parameters a, b, c are all rational numbers, and for almost all primes p, the reduction modulo p of the hypergeometric equation (6.1.2) with parameters a, b, c has two solutions in Fp [2] (resp. Fp(2), resp. Fp[[2]], resp. Fp((2))) which are linearly independent over Fp(2p) (resp. F(2P), resp. Fp(2"), resp. Fp((2P))).
86
N.M. Katz:
Proof By (5.4.1), (6.2.1).r and by (6.0.4.3), (6.2.2)r Because the hypergeometric equation has regular singular points, (6.2.3)~ (6.2.4). By (5.4.4), we have (6.2.2)~ (6.2.5). By ([24], Theorem 1.3.0), (6.2.5) implies that E(a, b, c) has regular singular points on Sc, and that its local monodromy around each singular point 0, 1, oo is of finite order. In particular, (6.2.5) implies that E(a, b, c) on Sc has rational exponents at each of 0, 1, ~ , or, what is the same, that the hypergeometric equation (6.1.2) with parameters a, b, c has rational exponents at 0, 1, ~ . As the exponents are 0 and 1 - c at 0, 0 and c - a - b at 1, and a and b at ~ , (6.2.5) implies that a, b, c~Q. The rest of the implication (6.2.5)~ (6.2.6) results from the fact that E(a, b, c) "comes from" T - a n open subset of Spec(Z), so that we may test for "p-curvature zero for almost all p" by seeing if the reduction modp on E(a, b, c) on Svp has p-curvature zero for almost all p (and perform this latter test prime by prime availing ourselves of (6.0.6.7) and (6.0.7)).
The implication (6.2.6)~ (6.2.5) follows by (6.0.6.7) and (6.0.7). To conclude the proof, we must show that (6.2.6) implies one of the equivalent conditions (6.2.4) or (6.2.2). This will occupy the rest of this chapter.
6.3. Conclusion of the Proof of 6.2 in a Special Case
(6.3.0) Proposition. Suppose that one of the "exponent differences" (6.3.0.0)
1-c,
c-a-b,
a-b
lies in Z. Then (6.2.6) implies (6.2.4). Proof By ([24], Theorem 1.3.0), the hypothesis (6.2.6) implies that the local monodromy of the hypergeometric equation with parameters a, b, c around each singular point 0, 1, ~ is of finite order. The (topological) fundamental group of P~ - {0, 1, ~ } is (the free group) generated by any two of the elements 7o,h,7o~ (where 7s="turning once around s", s=0, 1, ~). Thus the monodromy group of the hypergeometric equation is generated by the local monodromy around any two of the three singular points 0, 1, oo. As each of local monodromy transformation is of finite order, it suffices to show that any one of them commutes with any other of them (then the monodromy group is an abelian group, generated by two elements of finite order, hence is finite !). The integrality of one of the exponent differences means that one of the local monodromy transformations has both eigenvalues equal. As it is of finite order, hence semisimple, this local monodromy transformation is necessarily scalar, and hence commutes with either of the other two local monodromy transformations.
87
6.4. Solutions of the Hypergeometric Equation in Characteristic p (6.4.0) Proposition (compare [23]). Let a, b, c be integers contained in {0, 1, ..., p - 1}. In order that the hypergeometric equation with parameters -a,-b,-c admit two " m o d p" solutions in Fp[2] which are linearly independent over Fp(2), it is necessary and sufficient that either b> c > a or a>c>b.
Proof Let d - - f ( - a , - b , - c ) denote the hypergeometric differential operator with parameters - a, - b, - c;
r=((-a,-b,-c)=2(1-,~) (6.4.0.1)
+ [-c-(1-a-b) 2] J - ~ - a b .
is an Fp [2P]-linear endomorphism of Fp [2], which maps the Fp-module N of polynomials of degree at most p - 1 to itself (in an Fp-linear manner). Because the canonical map (6.4.0.2) Fp 1,2], ~ Fp l'2q @ ~ N
is an isomorphism of Fp I-2p]-modules, we have (Fp [2 p] being flat over Fp l) an isomorphism of Fp 1,2P]-module (6.4.0.3) Ker( in Fpl,2], ~ FpE2P]|
Similarly, we obtain isomorphisms (6.4.0.4) (6.4.0.5) (6.4.0.6) KerY Ker d KerE in in in Fp(2), ~ Fp(2P)| Fp l,Ea]], ~ Fp((2)),Fpl,EZq] | ~ (Ker E in N),
Fp((2P))|
We next calculate the matrix of ( ( - a , - b , - c ) acting on N, in the basis 1, 2, ..., 2 p- 1 of N: for any integer n, we have (6.4.0.7) d ( - a , - b , - c ) ( 2 " ) = n ( n - 1 - c ) 2 "-1 - ( n - a ) ( n - b ) Define polynomials R Q by (6.4.0.8)
P(X) = - (X - a)(X - b)
2".
Q (X) = (X + 1 ) ( X - c).
88
N.M. Katz:
Then the matrix of f ( - a , - b , - c ) on N is /P(O) Q(O) P(1) Q(I!. (6.4.0.9) t @ ( 2 ) ."."."..." ' .
It's proper values are thus its diagonal terms. If a = b, then all but one of these diagonal terms is non-zero, and hence zero is a proper value of multiplicity one, which implies that the kernel of g ( - a, - b, - c) in N is one dimensional. Thus we may assume a +b, and that a < b, since the problem is symmetric in a and b. As a + b, the matrix (6.4.0.9) has precisely two diagonal terms which vanish, namely P(a) and P(v). The subspace N(<b) of N consisting of polynomials of degree <b is stable under (, and f is invertible on N/N(< b). Since the formation of the maximal nilpotent subspace is an left exact functor, it follows that the inclusion of N(< b) in N induces an isomorphism (6.4.0.10) Ker( in N(<b) -,Kerr in N.
Furthermore, the subspace N( < a - 1) of N ( < b) consisting of polynomials of degree < a - 1 is stable under f, and f is invertible on it. It follows that the projection of N(<b) onto N(<b)/N(<a-1) induces an isomorphism (6.4.0.11) K e r r in N(_<_<b)--,Ker( in N(<b)/N(<a-1). The matrix of f ( - a , - b , - c ) on N(<b)/N(<a- 1) is
P(a+ 1) Q(a +,1!.
,,/
P(b)
Algebraic Solutionsof DifferentialEquations which has the shape
89
0 Q(a)
*
(6.4.0.13)
Q(a+ 1)
", Q(b- 1)
the ,'s indicating non-zero diagonal terms. If none of the off-diagonal terms vanishes, this matrix obviously has as image the subspace N ( < b - 1)/N(<a- 1) of N(<=b)/N(<a- 1), and hence has one-dimensional kernel. Conversely, if one of the off-diagonal terms vanishes, then the matrix (6.4.0.13) is in block form
0 *? *? *? 0 ,?
9?
(6.4.0.14)
*? ,? ,?
,?
and clearly has two-dimensional kernel. But the off-diagonal terms are Q(a), Q ( a + l ) . . . . . Q(b-1). As the only possible zero of Q(X) in {0.... , p - 2} is c, it follows that an off-diagonal term of (6.3.0.13) vanishes ifandonlyifc~{a,a+l ..... b - l } . Q.E.D.
6.5. The Calculus of Fractional Parts ([44])

(6.5.0) For any real number x, its fractional part ( x ) is the unique real number satisfying
90
N.M. Katz:
(6.5.0.1)
0<(x)<l
x ~ ( x ) modulo Z.
As real valued function, it satisfies
(x) =0r
(6.5.0.2)
((x)+(x))=(x+y)
(x)=(y)c~x--y
(x)+(-x)=l
mod Z if x e Z .
(6.5.1) For any prime number p, and any ~ Q c ~ Z p (i.e., any rational number with denominator prime to p), we define Rp(a) to be the unique integer such that (6.5.1.0)
O< Rp(~t)<=p- I
= Rp (a) modulo p (Q r~ Zp).
As function from Q n Zp to {0, 1..... p - 1} = Z, it satisfies
(6.5.1.1)
(6.5.2)
Rp(oO=-Rp(fl).r
mod p(Q c~ Zp).
Lemma. Let a~Z, N E Z , and suppose p,~VN. Choose integers B and A such that p A = 1 + NB. Then
P R , = "
(6.5.2.0)
=P
=P
pN
=p (6.5.2.1) (6.5.2.1.0)
Proof As ~ - r
~t
. Q.E.D.
and p,(N, then if p> la], we have
Corollary. I f ~, N e Z , - ~ r ~ Rp
p
pN
aA and A is invertible modulo N, ---~--r Z , and hence

U \ I v / a A \ ~(;1, ,. 2 N . . . . . ~ N - 1 } 9 k-zr-, / e ] ,
(6.5.2.1.1)
91
Asp>l~[, p ~ - < 1-- , a n d h e n c e N
(6.5.2.1.2)
whence (6.5.2.1.3)
o____\ W - / - - ) - N
-< 1
pN
. Q.E.D.
(6.5.3) Corollary. Let ~ ~ Z, N ~ Z with N ~ O. For each invertible element A in Z/N Z, we have the limit formula if (6.5.3.1) limit P
pd=-l(N)
~-~Z
if -~-~Z, ~ < 0 if ~-~Z, ~ - > 0 .
Proof If a/NCZ, this follows immediately from (6.5.2.1). If ~/NeZ, then for all primes p > I~1,
-p-Rp
= ~
~ Np
if N > 0 .
6.6.0. Proposition. Let a, b, c~Q, and suppose that none of the "exponent differences" 1 - c , c - a - b , a - b lies in Z. In order that (6.2.6) be verified, it is necessary and sufficient that both of the following conditions be verified. (6.6.0.1) None of the numbers a, b, c - a , c - b lies in Z. (6.6.0.2) Let N be a common denominator for a,b,c. For every A e Z invertible in Z/N Z, we have
either l > ( a d ) > ( c A ) > ( b A ) > O or l>(bA)>(cA)>(aA)>O.
Proof Let's first prove sufficiency. Fix a A invertible in Z / N Z , and suppose, to fix ideas, that l > ( b A ) > ( c A ) > ( a A ) > O . Our limit formula (6.5.3.1) then implies that for all sufficiently large primes p such that p d = 1(N), we have
(6.6.0.3)
l>l
Rp(-b)>lRv(-c)>l P
Rv(-a)>O.
92
N.M. Katz:
By (6.4.0), this implies that the hypergeometric equation with parameters a, b, c has " t w o " mod p solutions for all sufficiently large primes p satisfying pA - 1 (N). As there are only a finite number of d's to consider ( Z / N Z being finite), (6.2.6) follows. Now for the necessity. Choose a A invertible in Z/NZ. By hypothesis, for every sufficiently large prime p with pA -- 1 (N), we have either (6.6.0.4)
or
1--Rp(-b)>lRp(-C)>= l-- Rp(-a)

P P P
(6.6.0.5)
1 Rp(- a) > 1 Rp(- c) > I Rp(- b). P P P
As there are infinitely many primes with p A--I(N), either (6.6.0.4) or (6.6.0.5) must hold for infinitely many such primes. As the roles of a and b are symmetric, we may suppose (6.6.0.4) holds infinitely often for such primes. We may then apply the limit formula (6.5.3.1) to (6.6.0.4), to obtain one of the following inequalities (6.6.0.6) (6.6.0.7) (6.6.0.8) (6.6.0.9)
l>(bA)>(cA)>(aA)>O I>(bA)>=(cA)>O I>(eA)>(aA)>O

l>(cA)>0 if bCZ, if aCZ,
if aCZ, aeZ
bCZ,
and and a<0. a=<0, b>0,
beZ
and
if a, b ~ Z ,
b>0
We now use the hypothesis that none of the "exponent differences" l - c , c - a - b , a - b lie in Z to eliminate the last three cases. Since a-bCZ, (6.6.0.9) is impossible, i.e., not both a and b are integers. Ifb ~ Z, b > 0, then 1 Rv (b),~ 1 for all p sufficiently large, so for every A P which is invertible modulo N, the limit formula gives (6.6.0.10)
l>(cA)>=(aA)>O.
In particular, replacing A by - A , (6.6.0.11)
l>(-Ac)>(-Aa)>O.
But < - A c ) = l - ( A c > , < - A a ) = l - < A a > , so that (6.6.0.11) is the opposite inequality to (6.6.0.10), and we conclude (cA> = <aA) for all A invertible mod N. Hence (c> = <a), whence c - a lies in Z, whence (as beZ), c - a - b lies in Z, contrary to hypothesis. Thus (6.6.0.8) does not
Occur.
93
If a~Z, a < 0 , then --1 Rx,(_a),,. 1 for all p sufficiently large. Then for P any A invertible mod N, the limit formula gives (6.6.0.12) 1 > (bA) > (cA) > 0 .
Just as above, replacing A by - A shows ( b A ) = (cA) for all invertible A modulo N. Hence c - b lies in Z, hence (as a~Z), c - b - a lies in Z, another contradiction, so (6.6.0.7) does not occur. Thus we have shown that ar bq~Z, and that for every A invertible mod N, we have either (6.6.0.13)
or
l >(b A) > (cA) > (aA) >O

1 > (aA) > (cA) >= A) > O. (b
(6.6.0.14)
We next show that c-aq~Z and c-bq~Z. As by hypothesis a - b r we cannot have both c - a ~ Z and c - b ~ Z. Suppose that c - a ~ Z. Then c-br hence for all A invertible mod N, A c - A br whence we have either (6.6.0.15)
Or"
l > (b A) > ( c A ) = ( a A ) >O

1 > (aA) = (cA) > (bA) > 0 .
(6.6.0.16)
But (6.6.0.15) cannot hold for both A and - A . Hence there exists a A invertible modulo N for which (6.6.0.16) holds. Rewriting (6.6.0.1) for this A via the limit formula (6.5.3.1), we get (6.6.0.17) 1> lim 1 R p ( - a ) = p~ p
pd -~ (N)
lim LRp(--c)> lim 1---Rp(-b)>0. p~ p p~ p

pA =-1 (N) pA =. 1 (N)
Thus for all p sufficiently large with pA = 1 (N), we have (6.6.0.18)
1 Rp(_C)>L Rp(_b)>O"
P P
This is incompatible with the first of the two following inequalities, one or the other of which holds for any sufficiently large prime in virtue of (6.4.0) (6.6.0.19) (6.6.0.20)
l~ R v ( - b ) > l R p ( - c ) > l R p ( - a ) ,
P P P P P P
1 R p ( - a ) > l~ Rp(-C)>__lRp(-b).
94
N.M.
Katz:
Hence (6.6.0.20) holds, and for all sufficiently large p with we have (6.6.0.21)
pA-
1 (N),
1 R p ( - a ) > l R v ( - c ) > l Rp(-b).

P P P
Substituting via (6.5.2.1), (6.6.0.21) may be rewritten (6.6.0.22) Because (6.6.0.23) P
(aA)-~-->(cA)
O
C >(bA) P
--r
b. P
c-aeZ, (aA)= (cA),
and hence >-P
--a
Because (6.6.0.16) holds for A, (6.6.0.15) must hold for - A . Using the limit formula, (6.6.0.15) may be rewritten 1> (6.6.0.24) lim 1 R p ( - b ) > p~ p pn ------- I ( N ) =
p~oo
p d =-- -- I ( N }
lim p~
p A =- - 1 ( N )
1Rp(-c)
p
lim
--Rp(-a)>0.
p
In particular, for all sufficiently large p with p A - - 1 mod (N), we have (6.6.0.25) 1__ Rv ( _ b ) > 1__ R~,(-c). P P
As this is incompatible with (6.6.0.20), (6.6.0.19) holds for such p, and in particular, (6.6.0) gives (6.6.0.26)
1Rp(-c)>__l Rp(-a).
Substituting via (6.5.2.1), this gives (6.6.0.27) As
(_cA)__c_>__(_aA)
P hence
--r --a
a
P
c-aeZ, ( - c A ) = ( - a A ) ,
(6.6.0.28)
--=>_ P
, P
which contradicts (6.6.0.23), and concludes the proof that symmetry we conclude that c-bCZ.
c-aCZ.
By
95
To conclude, we simply observe that as c - a r c-br the inequalities in (6.6.0.15) and (6.6.0.16), already established, are necessarily strict. This concludes the proof of (6.6.0). (6.6.1) Combining 6.3 and 6.6.0, we are "reduced" to proving the implication (6.2.6)~ (6.2.4) under the additional hypothesis that the rational numbers a, b, c satisfy (6.6.1.0) aCZ, b6Z, c~Z, c - a C Z ,
c-bc~Z, a-b6Z, c-a-b~Z.
(6.6.2) Corollary. In order that rational numbers a, b, c satisfy (6.2.6) and (6.6.1.0), it is necessary and sufficient that for any A e Z which is invertible
modulo N for a common denominator N of a, b, c, the rational numbers A a, A b, A c satisfy (6.2.6) and (6.6.1.0). Proof The sufficiency is clear; take A = 1. For the necessity, (6.6.1.0) will still hold because d is invertible mod N, and the condition (6.6.0.2) is obviously invariant under (a, b, c)--* (A a, A b, A c) for such A. By (6.6.0), (6.6.0.2) and (6.6.1.0) imply (6.2.6).
(6.6.3) Corollary. In order that rational numbers a, b, c satisfy (6.2.6) and (6.6.1.0), it is necessary and sufficient that for any integers r, s, t, the rational numbers a + r, b + s, c + t satisfy (6.2.6) and (6.6.1.0).
Proof Sufficiency is clear, necessity follows from (6.6.0).
6.7. In view of (6.6.3), we must show that if a, b, c satisfy (6.6.1.0), then if (6.2.4) holds for a, b, c, it holds for a + r , b+s, c + t whenever r, s, teZ. In fact, we have the more precise (6.7.1) Proposition. Let a, b, c be complex numbers satisfying (6.6.1.0). The monodromy group of the hypergeometric equation with parameters a, b, c depends only on the classes of a, b, and c modulo Z. More precisely, the equivalence class of the two-dimensional complex representation of nl (P~-{0, 1, oo }, 89 on the space of solutions at 89of the hypergeometric equation with parameters a, b, c does not change oJ we replace a, b, c by a + r, b + s, c + t, where r, s, t are arbitrary integers.
Proof The proof is by means of the explicit formulas given in Bateman ([1], pp. 93-94). Given complex numbers a, b, c satisfying (6.6.1.0), a basis of the space of germs of solutions at 89is provided by the functions
ul =F(a, b, c; Z) (6.7.1.0)
u 2 = Z l - C F ( a - c + 1, b - c + 1, 2 - c ; Z)
Let 7o (resp. 71) denote the class in nt(Pcl-{0, 1 , ~ } ; 89 of the loop "turning once counterclockwise aroung 0 (resp. 1)".
96
N.M. Katz:
(6.7.1.1) J
Yo
Let Co (a, b, c) (resp. C l (a, b, c)) be the matrix of "analytic contintiation along 7o (resp. Vl)" of the space of germs of solutions at 89 in terms of the basis (6.7.1.0). Then (cf. [1], p. 93-94), 0 (6.7.1.3)
Cl(a,b,c)=IB1 l(a,b,c) B2 l(a,b,c)~ \B,12(a,b,c) B2'2(a,b,c)] where the coefficients B~.j(a, b, c) are given by
sin(~za) sin(~b) sin (n c) ' sin(n(c-a)) sin (re(c- b)) sin (~ c) '
(6.7.1.4) B1 l(a, b, c)= 1 - 2 i e x p ( ~ i ( c - a - b ) ) ' (6.7.1.5) B2, z (a, b, c) = 1 + 2 i exp (~ i(c - a - b)) (6.7.1.6) B1 2 ( a , b , c ) = - 2 ~ i e x p ( ~ i ( c - a - b ) )
'
r ( c - a) r(c - b) r(b) r(a)'
r(c)r(c-1)
B2,1(a, b, c)
(6.7.1.7) r ( 2 - c) r(1 - c) = 2n i exp(n i(c- a - b)) F(1 - a) F(b - a) F(1 + a - c) F(1 + b - c)"
Clearly, the matrix Co (a, b, c) depends only on the classes of a, b, and c modulo Z. Because the functions exp (i n X), sin (n X), and F(X). F(1 - X) all satisfy the function equation (6.7.1.8)
f(X+n)=(-1)"f(X), Bl,l(a, b, c)
for n ~ Z
it follows immediately that each of
(6.7.1.9)
B2, 2 (a, b, c)
B1,2(a, b, c) B2,1(a, b, c)
depends only on the classes of a, b, and c modulo Z. From this invariance, together with the fact that neither BI,2 (a, b, c) nor B2,1 (a, b, c) vanishes (because a, b, c satisfy (6.6.0.1)!), it follows that for any integers r, s, t, the matrices Cl(a, b, c) and Cl(a+r, b+s, c+t)
97
are conjugate (by a diagonal matrix) in GL(2, C). As already noted, C o ( a , b , c ) = C o ( a + r , b + s , c + t ) , and as nl(P~-{0, 1, oo};89 is the free group on 70 and 71, the representations in question are indeed conjugate. Q.E.D. 6.8.0. In this section, we study the relation between the hypergeometric equation and curves of the form y"= x a ( x - 1 ) b ( x - - 2 ) c. This relation was known to Euler, in the form of his integral representation (cf. [1], p. 115, (6), or [45], p. 293)
9O
F(~, fl,~; 2)=
r(fl) r ( 7 - fl)
F(7)
5x=_~(x_l)~_r
It was reconsidere recently by Messing ([32]), from the conception of whose manuscript we have borrowed heavily. (6.8.1) Let n, a, b, c be strictly positive integers. We denote by X(n;a, b, c) the spectrum of the smooth C(2)-algebra (6.8.1.0) A(n; a, b, c)= C(2) Ix, y, 1/y]/(y" - x"(x - 1)b(x - 2)c) C(2) [x, y, 1/y], which is finite and &ale over (6.8.1.1) with basis (6.8.1.2) 1, 1/7, l/y 2, ..., 1/y "-1 . To any n-th root of unity ~ / ~ , , we assosciate the C(2)-automorphism of X(n; a, b, c) given by (6.8.1.3)
XF--~ X
B = C (2) [x] [ 1/x (x - 1)(x - 2)],
y~--~~y.
This defines an action of/~, on X(n; a, b, c). Let X denote the "identical" character of /~,, i.e., Z(~)=~, and for any integer L let X(f) denote the character of/4, given by (6.8.1.4) ;((~) (~)= ~e. This basis (6.8.1.2) of A(n; a, b, c) over B is just its "isotypique" decomposition according to the characters of/J,: (6.8.1.5) P(X(-E)) A(n; a, b, c ) = y - e B = lim y - e - , , C(2)[x].
F~ 9O
The de Rham cohomology group H~R(X(n; a, b, c)/C(2)) is the cokernel of the exterior differention map (6.8.1.6)
A(n; a, b, c) d , 12~(,;~,b,c)/c~~
because X(n; a, b, c) is affine and smooth of relative dimension one over C(2). The module f2]~,;,,h.~)/C~Z) is free over A(n; a, b, c) with basis dx.
7 Inventionesmath.,VoL 18
98
N.M. Katz:
The isotypique decomposition of the complex (6.8.1.6) with respect to /~, is given by
P()~(-d))(a(n; a, b, c))- d , P(g(-d))(A(n; a, b, c)).dx

(6.8.1.7)
y-e B
Thus
, y- r B.
dx = ~
y - e- r.. C (2) Ix]-
dx.
(6.8.1.8) P()~(- f))H~R(X(n; a, b, c)/C(2)) = cokernel of y-e B ~
f e B . dx.
The gauss-Manin connection V(d/d2) on HIR(X(n;a,b,c)/C(2)) is deduced from the endomorphism of the complex (6.8.1.6) given by Lie(0/O2), where ~?/~2 denotes the unique derivation of A(n;a,b,c) which kills x and extends the derivation d/d2 of C(2). Because 0/~32 commutes with .the action of /1, on A(n; a, b, c), Lie(~?/O2) induces an endomorphism of each of the complexes
y-r.B
a , y-iB.dx
which induces the Gauss-Manin connection on the cokernel. (6.8.2) Proposition. Let n, a, b, c be strictly positive integers. Then for every integer {> 1, there is a horizontal morphism of C(2)-modules with connection (cf (6.1.0))
E ( (c-'n (a+db+(Cn
(6.8.2.0)
1 , ( a ~ E c ) C(2)
--* P(Z(-f))H~R(X(n; a, b, c)/C (2)) eo ~ the class of y--7dx
defined by
(6.8.2.1) el ---, g
the class of .-7- 9 Y
Proof In concrete terms, the assertion is that the operator
(6.8.2.2)
(c (
1)
annihilates the class of dx/y e in H~R(X(n; a, b, c)/C(2)). In fact, direct calculation (cf. [1], p. 60, (11)) shows that in f2~(,:,,b,c)/C(~) we have
99
y ( x - 2) !
(6.8.2.3)
_{2(l_2)(Lie
+ [~a+Cb ( ta+t2+2ec)2
(6.8.3) Proposition. Hypotheses as in (6.8.2), if n does not divide a + b + c, then for every integer ~ > 1 which is invertible modulo n, we have
dx dx xdx t (6.8.3.0) C (2) I-x] 7 - c C (2) 7 - + C (2) - 7 - + d Proof. For every integer m > 2, we readity calculate d [.xm- l(x--1)(X-- 2 ) ~ _ d(xm- l (x - 1)(x--2)) ye } ye dxm-l(x-1)(x-2) ye
(6.8.3.1) 3/
d~c\
dy y
d(x"+~ -(1 +2) x"+2x "-~)

( x " - 2 . x ( x - 1 ) ( x - 2 ) (a/n b/n +P(x) c/n ]
/
= m+l
(a+~b+gc n
xm
7-
ye
where P(x)eC(2)[x] is a polynomial of degree at most m - 1 . Because n does not divide a + b + c, and E is invertible modulo n, n does not divide
fa+fb+Ec,
whencem+l+-(-fa+gb+Ec).
Q.E.D.
(6.8.4) Proposition. Hypotheses as in (6.8.2), if n does not divide a + b + c, then for every integer E > 1 which is invertible modulo n, the image of the mapping (6.8.2.0) contains the classes of dx/y e and x dx/y e.
Proof. Direct calculation shows that in ~ t . ;,. b,c~1Ct~)we have
/
(6.8.4.0)
I
+
n
('a+'c n
1-
j
,c2) n
/
dx ye
7*
100
N.M. Katz:
As n does not divide a + b + c, and f is invertible modulo n, n does not

dividefa+fb+Ec, whence2:t:(fa+~b+fc).
Q.E.D.
(6.8.5) Proposition. Hypotheses as in (6.8.2), suppose that n does not divide a + b + c. For each integer 1 < ~ <-n - 1 which is invertible modulo n, consider the sequence of integers f, f + n, f + 2 n, .... For all r sufficiently large, the mapping (6.8.2.0) for f + r n
E(f+rn)--,(d+rn) n
(6.8.5.0)
. -1,(f+rn)
--, p(z(- +)) H~R(X(n; a, b, c)/C(X))

is surjective. Proof. By (6.8.1.8), any finite number of elements in
P(z (- O) HI,,,(X(n; a, b, c)tC (;,))

may be represented by differentials lying in
C(,b Ix] y++,.

for a suitably large r. By (6.8.3), these elements are linearly dependent upon the classes of dx x dx (6.8.5.1) ye+r, , y++,, " Thus any finite number of elements in P0~(-{))H~oR(X(n; a, b, c)/C(2)) span a subspace having dimension at most two. Hence the space P()~(-{)) H~R(X(n; a, b, c)/C(2)) has dimension at most two, and for all sufficiently large integers r, the two classes (6.8.5.1) span. The result now follows from (6.8.4). (6.8.6) Proposition. Hypotheses as in (6.8.2), suppose that n does not divide a, b, c or a+ b+ c. Then for any integer f > 1 which is invertible modulo n, the mapping (6.8.2.0) is an isomorphism: (6.8.6.0) E(~,
~a+~b+dCn 1, ~a+fc_) C(2)
dx
= , P(Z(-+)) U ~ ( X ( n ; a, b, c)/C(2)).
Proof. By (6.8.5), the target of (6.8.6.0) has dimension at most two, while its source has dimension two. If we can exhibit two linearly in-
10t
dependent elements of the image, it will follows that the target has dimension precisely two, and that (6.8.6.0) is surjective, and hence an isomorphism (source and target having the same dimension). By (6.8.4), the image contains the classes (6.8.6.1)
dx
ye ,
x dx
yt
whence it suffices to prove that they are linearly independent in P()~(-E))H~R(X(n; a, b, c)/C(2)). Suppose the contrary. Then there exist (by (6.8.1.8)) ~, fl ~ C(it), not both zero (6.8.6.2) R (x) e B = C (it) Ix] [ 1/x (x - 1)(x - it)] such that (6.8.6.3)
d ~[R(X)]ye ] = ( ~ + f l x ) -7-dx in ~"~A(n;a,b,c)/C()O
This implies that on the open set where R(x) is invertible, we have (6.8.6.4) or, equivalently (6.8.6.5)
dR(x) R(x) dR(x) R(x) / a/n
dy (o~+ fl x) = dx y R(x) c/n \ (~ + fl x) dx
k-U +
bin
l +
) d =
R (~)
in
~c(z, x)/C(~) 9
We first remark that R(x) is necessarily a polynomial in x. If not, it has a pole at one of the three points x = 0 , 1 or 2 (because R(x)~B), in which case (6.8.6.6)
(a+flx) dx R(x)
would be holomorphic at one of the points x = 0, 1, or 2 and in particular, would have zero residue at one of the points x = 0, 1, 2. But (6.8.6.7)
dR(x) R(x)
(a/n+ bin + c/n ~ dx

~ x x-1 x-2l
necessarily has residues at x = 0, 1, 2 which are congruent modulo Z to -fa/n, -(b/n, -~c/n, which by hypothesis are not integers. Hence (6.8.6.7) has non-zero (even non-integral) residues at each point x = 0, 1, 2. Thus R(x) is a polynomial. We further claim that the polynomial R(x) has zeroes at x = 0 , 1, and 2. For if not, R(x) would be a unit at one of
102 the points x = 0, 1, 2, and hence (6.8.6.8)
N.M. Katz:
(~ + fl x) dx
dR(x)
R(x)
R(x)
would be holomorphic at one of the points x = 0, 1, 2. In view of (6.8.6.5), this is absurd, hence R(x) is a polynomial of degree at least three. Because R(x) is a polynomial of degree at least three, it follows that (6.8.6.9)
(~ + fl x) dx
R(x)
is holomorphic at x = oo, hence has zero residue there. Clearly, the residue at x--- oo of
[ a/n bin e/n \ E ~ - + ~ - ~ _ 1 + ~ - ~ _ 2 ] dx=Ea/n, d +/x X fn b'

(6.8.6.10)
d(X-x_ll)
+~c/n
d(x - 2) x-2 dR(x) R(x)
is just - E a / n - f b / n - g c / n , while the residue at x = oo of (6.8.6.11)
is - d e g r e e (R (x)). By (6.8.6.5), we must have (6.8.6.12) - degree R (x) + Ea/n + ~ b/n + ~ c/n = 0 Q.E.D.
which is absurd' as f ( a+b+c)n i s n o t aninteger.

6.9. Conclusion of the Proof of 6.2
(6.9.0) Let a, b, c, n be integers, n > 1, and suppose that the hypergeometric equation with parameters a/n, b/n, c/n has two "rood p " solutions for almost all p (i. e., suppose a/n, b/n, c/n verify (6.2.6)). We must prove that E(a/n, b/n, c/n) has a full set of algebraic solutions (cf. (5.4.2.1)), in order to conclude the proof of 6.2. (6.9.1) (6.9.1.1) is an integer. (6.9.2) We thus assume that none of the exponent differences is an integer; then 6.6.0 implies that none of the numbers (6.9.2.1) As we saw in 6.3, this is the case if any of the exponent differences 1 - c/n,
c/n - a/n - b/n,
a/n - b/n
a/n,
b/n,
c/n - a/n,
e/n - b/n
A l g e b r a i c S o l u t i o n s of Differential E q u a t i o n s
103
is an integer, and that for every integer C > 1 which is invertible modulo n, the numbers Ca~n, (b/n, Cc/n also satisfy (6.9.0) (cf. (6.6.2)). By (6.6.3) and (6.7.1), the hypothesis and conclusion of the alleged implication (6.2.6)~ (6.2.4) depend only on the classes modulo Z of a/n, b/n, c/n. Thus we may and will suppose in addition that (6.9.2.2) -->0,
n a
----->0,
n n
1+b/n-c/n>O.
We define strictly positive integers A, B, C by (6.9.2.3)
A
n
c-a
n
B
n
-1-~
b-c
n
C
n
a
n
The non-integrality of c - a ,
n
c-b
n
, a ,
n
b
n
(cf. (6.9.1.1), (6.9.2.1)) implies
that n does not divide any of the integers A, B, C, A + B + C. Then we may apply (6.8.6): (6.9.3) For any integer C > 1 which is invertible modulo n, ~ , P (Z ( - C)) H~R(X(n; A, B, C)/C(2)).
(6.9.3.0) E((a/n, (b/n, (c/n)lC(2)
Thus, for every integer 1 < ( < n - 1 which is invertible modulo n, (6.9.3.1) P (Z(- C)) H i , (X (n; A, B, C)/C (2))
with its Gauss-Manin connection has p-curvature zero for almost all prime p (cf.(5.4.3.3)). As C varies over the integers l < C < n - 1 , the characters Z(-C) run over a Q-conjugacy class of irreducible representations of#, (namely, the faithful ones). Thus we may apply Theorem 5.7 to deduce that for every integer C >__1 invertible modulo n, (6.9.3.2) P(X(-C)) H ~ , ( X (n; A, B, C))
with the Gauss-Manin connection has a full set of algebraic solutions. By (6.9.3.0), it follows that for every integer C> 1 which is invertible modulo n, in particular for C = 1, the hypergeometric module (6.9.3.3)
E (C a/n, C b/n, C c/n)[ C (2)

Q.E.D.
has a full set of algebraic solutions.
(6.9.4) Corollary. Let a, b, c, n be integers, n >_1, and suppose none of the exponent d!fference (6.9.4.0) 1 - c/n,
c/n - a/n - b/n,
a/n - b/n
is an integer. Then the hypergeometric equation with parameters a/n, b/n, c/n has two algebraic solutions ![ and only if, for every integer 1 <-C <-n - 1
104
N.M. Katz:
which is invertible modulo n, we have either

(6.9.4.1) or
1> ( f a ) > ( f c ) > ( f b ) > 0 l>(fb)>(fc)>(fa)>0.
Proof This follows from 6.2 and 6.6.0. (6.9.5) Remark. (Interpretation of (6.9.4)). Given three distinct roots of unity 41, 42, ~3 in C all distinct from 1, we say that 41 and 4z separate
~a and 1 if, in marching counterclockwise around the unit circle, starting at i, we encounter either ~1 or ~2 but not both before we encounter 43. Let 4./.=exp (2rti a ) (6.9.5.0)
4b/,=exp (2~ri b )
~c/. = exp (2re i ~-) -
The condition (6.9.4.1) is "simply" that these three roots of unity are distinct from each other and from 1, and that for every automorphism a of C, (6.9.5.1) (4,/,)" and (4b/,)" separate (4c/,)~ and 1.
7. p-Curvature and the Cartier Operation; a Problem on Elliptic Curves

7.0. Let T be a scheme, and S a smooth T-scheme. A connection on Cs relative to T is necessarily of the form
(7.o.11
((gs, Vj|
V(f)=df + fco
where co= V(1)~F(S, O~/r). Denoting this connection by Vo, we have Vo~,)=((gs, V~,+,o,), and the mapping c o ~ Vo, establishes an isomorphism between the group F(S, fJ~/T) and the group of connection on (gs. The subgroup of closed one-forms corresponds to the subgroup of integrable connections: indeed, the curvature K,~ of V,ois the mapping (7.0.2) 7.1.
Ko: Cs--~ f2~/r Ko,(f)= f .dco.

Suppose further that T is a scheme of characteristic p. Then if of the
coeF(S, f2~/r) is a closed one-form, the p-curvature, noted r
Algebraic Solutions of Differential Equations integrable T-connection V,~ on (gs is a p-linear mapping (7.1.0) 0,~: Der(S/T) ~ Endr (9s
105
which is additive in the variable closed one-form o9. (7.1.1) Recall from (6.0.2.5) the commutative diagram (in which g: S --~ T is the structural morphism)
S., F ~ S (p) ~ ~, S v ) S(p)
Fabs ~ T
the middle square of which is cartesian. (7.1.2) (7.1.2.0)

Proposition. The p-curvature 0o, of Vo is given by the formula
0o (D) = F* ((a* (o9)- cg(o9), r (D)))
where ~ is Cartier's isomorphism (2.2.1) jr (F, f2~/r) ~ ' ~s'(,,/r (applied to the class of the closed form o9), ~* (D) is the T-derivation of S (p) defined by
(7.1.2.1)
a*(D)(a*(f))=a*(D(f)),
f a local section of (gs,
and ( , ) is the canonical pairing of Os~prmodules

(7.1.2.1)
f21(p,/r | Der (S~P) T ) --* (gs~, . /
Proof The proof depends "only" on Hochschild's identity (3.5.2.6), and the definitions.
The question is local on S, which we may suppose &ale over A t via coordinates sl, ..., s,. We first calculate 0,o (D). (7.1.2.2) 0,o (D)= ( V,~(D))p - V~,(Dv) =(D + (09, D >)p D p - (o9, Dr).
-
By Jacobson's formula, we have (7.1.2.3) As (7.1.2.4) ((to, D))V=F * a* ((o9, D))=F* ((a* to, tr*D)),
(D+(O9, D))P=DP+DP-a((O9, D))+((O9, D))V.
we have finally the formula (7.1.2.5)
0,o(0)=0"-1((o9, O))+F*((a* o9, a* O ) ) - (o9, Dr).
Comparing (7.1.2.5) with the asserted formula (7.1.2.0), we see that it remains to show that
(7.1.2.6) F* ((cd o9, a* O ) ) = (co, O p) - D p-x ((o9, O ) ) .
106
N.M. Katz: As 09 is closed, it may be written (cf. (2.2.1))
(7.1.2.7)
0 9 = ~ F*(ai) se~-1 d s i + d g ;
al . . . . . a , , g e F ( S , (gs).
As both sides of the asserted equality (7.1.2.6), regarded as function of the closed one-form 09, are F-l(t~s~p~)-linear and annihilate exact oneforms, we are reduced to the case (7.1.2.8) Then (cf. (2.2.1)) (7.1.2.9) and (7.1.2.6) becomes (7.1.2.10) F* ((dtr* (si), tr* (D))) = ( s p- 1 d s i , D p) - D p- 1((s~-1 dsl, O)), or equivalently (7.1.2.11)
(D(si))P=s~-I D P ( s i ) - D P - I ( s ~ - I D(si))
09 -- s~-1 ds i . cg (09)= d (o'* (si)),
which is none other than Hochschild's identity (3.5.2.6). (7.1.3)
Q.E.D.
Corollary. Hypotheses as in 7.1, the following conditions on a closed one-form 09e F(S, f21s/7.) are equivalent.
The connection Vo, on (9s has p-curvature zero
(7.1.3.1) (7.1.3.2) (7.1.3.3)
cg(09)= tr* (09)
09 is locally logarithmic, i.e., locally on S there exists an invertible function g with 09 = dg/g.
P r o o f (7.1.3.1)r by (7.1.2.0). To see that (7.1.3.1)r (7.1.3.3), recall that by Cartier's theorem (cf. (6.0.3) and [-24], Theorem 5.1), ((gs, 17j has p-curvature zero if and only if (9s is spanned as (9s-module by the subsheaf of germs of horizontal functions. Thus (7.1.3.1) is true if and only if there exists locally on S an invertible section of (gs, f , with 0 = V o , ( f ) = d f + f 0 9 , or equivalently (taking g = f - 1 ) , if and only if to is locally logarithmic. Q.E.D.
(7.1.4) Remark. In Cartier's original "operator" (cf. [2]), the absolute Frobenius Fabs: T--, T was an isomorphism; T was, in fact, the spectrum of a perfect field. Then a: S ~ P ~ S was also an isomorphism, and the original Cartier operation rgorigi,a~ was defined as an additive isomorphism (7.1.4.0) which satisfied (7.1.4.1)
(~original: ~
C~original(fP09)=f;
~ ' ~'~/T
~original(09)9
Algebraic Solutionsof DifferentialEquations In our notation, the relation between <gorigi,aland cr is just
(7.1.4.2) <r = (tr- 1), o <g.
107
7.2. Let T be an arbitrary scheme, S a smooth T-scheme. Let ~,e be an invertible sheaf on S, given by transition function f~j with respect to an open covering q/i of S. Then giving a T-connection 17on ~ is equivalent to giving, for each q/i, a one form tDiG_r(dlli, ols/T) subject to the compatibility (7.2.0.0) coi - a~j = dfij/fij.
The connection is integrable if and only if each o9i is closed. If (s 17) and (~', 17') are two invertible sheaves with connection, both given with respect to the same open covering q/i of S by data (col,f0 and (<n'i,f/~) as above, an isomorphism between (s 17) and (~', 17') is just the giving of invertible functions gi~F(qli, C~') subject to the conditions (7.2.0.1) Thus we have: (7.2.1) Proposition. The group of isomorphism classes of invertible sheaves on S with T-connections (Z.g~',17) (under tensor product) is H 1(S, (9*
d log
~, ~'~Is/T)
f~i =f~(gi/gj) CO - CO'= dgi/g i i i
on q/i n q/~ on q/i.
"
The group of isomorphism classes of invertible sheaves on S with integrable T-connection is H I (S, f2*/T) where 12~*/Tdenotes the multiplicative de Rham complex
(7.2.1.0)
(9*
a'o~ , f2~/T
d---~f2Z/T
a , ....
As a corollary of (7.1.2), we have (7.2.2) Proposition. I f T has characteristic p, and (~, 17) is an invertible (gs-module with integrable T-connection, given by the data (Col,fig) on an open covering ~ of S, its p-curvature (7.2.2.0) ~O: Der(S/T) ~ End~s(L~')_~ (gs
is the p-linear mapping given locally by

(7.2.2.1)
~O(D)=F*((tr*(ogi)-cs
a*(D)))
over q/i-
[This formula has a global meaning, because, by (7.1.3.3), (7.2.2.2)
~r*(~o0- ~(o~)-,r* (~o~)+ ~(co~)

= a* (dfij/fij) - ~g(dfiJfij) = 0].
108
N.M. Katz:
(7.2.3) Remark. ~, depends additively upon the invertible sheaf with integrable connection (~, V). 7.3. When T is (the spectrum of) a field k of characteristic zero, the group Hl(S, t2~/r) has a classical interpretation. Recall that a meromorphic closed one-form co on S is said to be a differential of the third kind if there exists an open covering {q/i} of S, and on each q/i a closed holomorphic one form (Diff_F(d~i,~l/T) and a non-zero meromorphic function gi on q/i, such that (7.3.0.1)
co=coiq-dgi/gi
onq/i,
dcoi=O.
Now define f~j = gj/gi, a meromorphic function on q/i n q/j. In fact,
fiijEz F(O~Zi("3q/j, (Q~), because on q/i n q/j

(7.3.0.2)
dfii/fij = d g j / g j - d gi/gi = coi - coj
so that dfij/fij is holomorphic on q/i c3 q/j. Because we are in characteristic zero, this implies that fij and 1/f~j are holomorphic on q/i ~ q/j. Thus we have assosciated to a differential of the third kind co on S a Cech 1-cocycle (coi,flj) for the complex I2~/T. If, on the same covering q/i, we choose different closed holomorphic one forms co'ion q/i and different meromorphic functions gl on q/g, such that (7.3.0.3)
co =coi + dgi/dgi = co'i+ dgl/gl
on q/i,
then hi=g'i/g i lies in F(ql i, (9"), because coi-co'i=dhi/hi is holomorphic on q/i, and hence (col, f/j) and (co'i, fi)) differ by the coboundary (dhi/hi, hi~hi). Thus we have a well-defined mapping from the group of differentials of the third kind on S to the group HI(S, f2~/r). (7.3.1) Proposition. Let S be a smooth connected scheme over a field k of characteristic zero, and let K denote the function field of S. Then the above constructed mapping sits in the exact sequence, in which d. t.k. (S) denotes the group of differentials of the third kind on S: (7.3.1.0)
O---~k*--~ K*
dl~
(S)--~HI(S,O*/k)--q,O.
Proof To prove exactness at Hi(S, O*/k), let (col, f/j) be a Cech 1cocycle for some covering q/i of S, and consider the meromorphic differential co = cot. By the cocycle condition, dco = 0, and
(7.3.1.1)
co=coi+dfli/fli
on
q/i,
so co is a d.t.k., and from the data 7.3.1.0, the procedure of (7.3.0)just reconstructs the cocycle (coi,fij). To prove exactness at d.t.k.(S), notice first that if co--dg/g, in the construction (7.3.0) we may take all coi = 0, all gi = g, whence we construct
109
the zero cocycle (0, 1). If a d.t.k, to dies in H1(S, f2~/k), there exists an open covering q/i of S, and on each a#i there exists a closed (7.3.1.2) such that
coi~F(~[i, ~Qls/T)
a meromorphic function gie K* an invertible function hieF(qli, (9*)
co=coi+dgi/g i on q/i
(7.3.1.3)
coi=dhl/hi gj/gi = hi/h2
on q/i on 0///n q/j.
Then co = d(g i hi)/gi hi, and gi hi = gj hj is a global meromorphic function, so co is logarithmic. Q.E.D.
7.4. The Case of Curves
(7.4.0) Let T be affine, say T = Spec(A), and S a projective and smooth curve over T with geometrically connected fibres. Then f2*/r is just the two-term complex (7.4.0.0)
(9~
dl~
Passing to hypercohomology, we have a long exact sequence (7.4.0.1)
0 -~ F(S, f2~/r)-~ Ha(S, f2*/r)
~ I'P(S, (9*)
dlog , Hi(S, f~s/r)
(the left hand zero because F(S, (9*)=F(T, (9*) is annihilated by d log). The trace morphism (of. [15, 42]) defines a functorial isomorphism (7.4.0.2) H 1(S, Qs~/r) ~ , a = F(T, (9r)-
(7.4.0.3) If P is a section of S/T, its image IPI is a divisor in S which is smooth over T. The inverse of its sheaf of ideals is an invertible sheaf on S, noted classically (gs([P]). If P1, .-., Pr are sections, and n 1, ..., nr integers, we define the invertible sheaf (7.4.0.4)
(9s (Z n i [P~]) = (gs([P1])|174 ... | (9s ([P~])|

trace
The composite mapping

(7.4.0.5) HI(S, (9,) dlog ) HI(S,
~r-~IS/T) ~~,A
allows us to attach to the class in Hi(S, (9*) of Os(Z ni [Pi]) an element of A, which is none other than the image in A of the integer 2; ni. Thus
110
N.M. Katz:
(7.4.0.6) the necessary and sufficient condition for the invertible sheaf ~)s(,F ni [P/]) to admit an integrable T-connection is that ,F nl = 0 in A. (7.4.1.0) Suppose further that A is a principal ideal domain, with fraction field k, and that S admits a section over A. Because S/A is projective, and A is principal, any k-valued point of Sk extends to a unique section of S over A, so it is the same to assume that S k has a rational point. (7.4.1.1) Then the Picard scheme Pies/A exists, is an extension of ZA by the abelian subscheme PiC~/A, and its formation commutes with arbitrary change of base A - * A'. Because PiCs~ is projective, and A is principal, any k-valued point of Pic~k/k= the Jacobian of Sk extends to a unique section of Pic~ over A. (7.4.1.2) (7.4.1.3) Because A is principal and S/A has a section, we have HI(S, C*) ~ ~Pics/A(A ). under
We define Hi(S, t2~/r)0 to be the inverse image of Pic~ the canonical mapping (7.4.1.4) H I (S, f2~/r) --* H 1(S, ~)~)~- Pies~A (A).
The long cohomology sequence (7.4.0.1) gives a short exact sequence (7.4.1.5) 0 ~ F(S, t2~/r) ---)n ~(S, t2*/T)o ---) PicO/A(A) ~ 0
which the specialist will recognize as the exact sequence of T-valued points of the "universal extension of the abelian scheme PiCO/A by a vector group" (cf. [33, 33 a, 40]). We will make no use of this interpretation. (7.4.2) Remark. If the fraction field k of A has characteristic zero, it follows from (7.4.0.5) that (7.4.2.1) H 1(S, f2*/r)o = H 1(S, t2~/r).
If A = k is a field of characteristic zero, we can combine (7.4.2.1) with (7.3.1.0), and (7.4.1.5) becomes the classical exact sequence (K denoting the function field of S) (7.4.2.2) d. t. k. (S) 0 ~ F(S, t2~/k)~ d log (K*) 9 PiC~ (k) --~ 0,
in which a d.t.k, co on S is mapped to the class of its residue-divisor S residue p(co)[P] (compare [33 a]). [Remember that on a curve, co is a d.t.k.r has at worst first order poles, and integer residues.] (7.4.3) The Birational Point of View: Again (7.4.3.0) Let K / k be a function field in one variable, k a field of characteristic zero and let oget2~/k be a (necessarily closed) one-form. We recall the following facts:
111
(7.4.3.1) In order for the rank-one differential equation (K, V~,)on K to become trivial on a finite extension of K, it is necessary and sufficient that there exist an integer n > 1 and a function g~ K* such that n co = dg/g. (If n co = dg/g, then f = g-1/n is a non-zero algebraic solution of the equation V~(f)=0. Conversely, i f f is a non-zero algebraic solution, let n be the degree of K(f)/K. Then n co = - t r a c e ( d f/f)=dg/g, g = Norm(l/f).) (7.4.3.2) If (K, V~,) has p-curvature zero for almost all primes p (in the sense of (5.4.3.2)) (i. e., if co is ~ locally logarithmic mod p" for almost all primes p in the same sense (cf. (7.1.3.3)), then co has at worst first order poles, and rational residues (i. e., there is an integer n > 1 such that n co is a differential of the third kind on the complete non-singular model of K/k). (For by ([24], Theorem 13.0), the differential equation (K, V~,) has regular singular points, and rational exponents.) (7.4.4) Proposition. Assumptions as in (7.4.3.0), suppose K=k(x) is the rational function field in one variable. If co~f2~/R, then the following conditions are equivalent. (7.4.4.0) The differential equation of rank one (K, [7o,) on K becomes trivial on a finite extension L of K. (7.4.4.1) The differential equation (K, VJ has p-curvature zero for almost all primes p (cf. (5.4.3.2)). Proof. (7.4.4.0)~ (7.4.4.1) by (5.4.4). By (7.4.3.2), (7.4.4.1) implies that an integral multiple nco (n> 1) is a differential of the third kind on Pk~, hence (cf. (7.4.2.2)) is logarithmic. By (7.4.3.1), this implies (7.4.4.0). Alternate proof if k = C : again by ([24], 13.0), the differential equation (K, 17o,)has all of its local monodromy transformations of finite order. Because the rank is one, the global monodromy group of the equation is abelian. Because we are working on an open set in Pr the global monodromy group is generated by the (finitely many non-trivial) local ones. Hence the equation has a finite monodromy group. By ([24], 13.0), it has regular singular points, so by the fundamental comparison theorem ([7], Theorem 5.9), it becomes trivial on a finite extension L of K. Q.E.D.
7.5. The Case of Elliptic Curves

(7.5.0) We do not know whether or not the analogue of (7.4.4) is true for elliptic function fields. We would like to explain how this analogue is equivalent, in certain special cases, to a rather striking diophantine statement, which we view as an (unproved !) arithmetic version of Manin's "theorem of the Kernel" (cf. [293). (7.5.1) Let E be an elliptic curve (= abelian scheme of relative dimension one) over T = Spec(Z [I/n]). We suppose that the generic fibre EQ has a
112
N.M. Katz:
non-trivial rational point of order two, for reasons which will become clear. For each prime p which does not divide n, we put Ep=E x r(SpecFp)), the elliptic curve over Fp obtained by reduction modulo p. The fact that Z [i/n] is a principal ideal domain, and the projectivity of E over it, gives (7.5.1.0) E(T) ~ , Eo(Q )
and permits the definition of the homomorphism of '~reduction modulo p" for each p not dividing n
(7.5.1.1)
EQ(Q)
~ , E(T)
, Ep(Fp).
This homomorphism is injective on the subgroup of EQ(Q) consisting of torsion elements of order prime to p. If p 4:2 and p is prime to n, the finite group Ep(Fp) contains a non-trivial element of order two (this being true of EQ (Q) by hypothesis), hence has an even number of elements. (7.5.2) Lemma. Hypotheses as in (7.5.1), for all primes p ~ 7 prime to n, the finite group Ep(Fp) has order prime to p. Proof If not, it has even order divisible by p, hence has > 2p elements. For p > 7, this contradicts the Reimann hypothesis t~ E p ( F p ) - p - 1 [ < 2 ] / p . Q.E.D.
(7.5.3) Lemma. Hypotheses as in (7.5.1), let p>_7, p~n, and suppose ~o is a non-zero differential of the first kind on Ep, i.e., o9~ F(Ep, f2~p/rp), and 09#: O. Then o9 is not locally logarithmic on Ep. Proof If co is locally logarithmic, cg(og)= ~o, and by duality the p-th power mapping induces 1 on HI(Ep, Cep); i.e., the Hass invariant HEFp of Ep is 1. But (cf. [25]), (7.5.3.1) 4~Ep (Fp)--- 1 - H modulo p
which contradicts 7.5.2 if H = 1. Q.E.D. (7.5.4) Let K denote the function field of Eo, and let ~oEf2ax/o be a differential such that (K, 17,o has p-curvature zero for almost all primes p. ) We hope that for some integer n >_ 1, n co will be logarithmic; we will make some "reductions" which are permissible with respect to this hope. (7.5.5) According to (7.4.3.2), we may, replacing co by an integral multiple n co, n > 1, suppose that 09 is a differeiatial of the third on E o. By (7.2.1.0), (7.3.1), and (7.4.2.1), a differential of the third kind co on Eo, taken modulo logarithmic differentials, is an isomorphism class of invertible sheaves on EQ of degree zero with connection (Leo, 17Q).Because Pic~ is projective (it's just E, in fact !) and Z [I/n] is principal, the inver-
113
tible sheaf ~Q on EQ extends uniquely (up to isomorphism) to an invertible sheaf ~ on E which fibre by fibre has degree zero (in fact, ~ Q ~ (gEQ([PQ]--[0]), for a unique PQ~EQ(Q), and, denoting by P its unique prolongation to a section of E over T, we have ~ = ( 9 ~ ( [ P ] - [ 0 ] ) (cf. (7.4.0.3)). (7.5.6) At the expense of enlarging the integer n, i.e. localizing on Spec(Z [1/n]), we may suppose that the conncetion [7Qon s extends to a T-connection 17 on ~e. The hypothesis that the original d. t.k. o~ on EQ was locally logarithmic mod p for almost all p is equivalent to the hypothesis that, for almost all p, the inverse image of ( ~ , V) on Ep has pcurvature zero. (7.5.7) Suppose we view (5r V) as an element in H I(E, sits in the short exact sequence (7.4.1.5) (7.5.7.1)
~/T)O, which
O--~ F(E, g2~/T)--~HI(E,Q~/T)O--~ E(T)--+O.
Reducing modulo p, we find an element (~p, 17p)in H 1(Ep, O~p/vp)o * which sits in the short exact sequence (7.5.7.2)
0 ~ F(Ep, ~p/F~) ~ HI (Ep, ~2*p/v~)o~ Ep (Fp) ~ 0.
For p > 7 prime to n, the sequence (7.5.7.2) is an extension of a group of order prime to p (namely Ep (Fp), cf. (7.5.3)) by a group killed by p (namely the one-dimensional F~-space F(Ep, ~2~,/Fp)), hence has a unique section (7.5.7.3)
0 -+ F(Ep, Q~/vp) ~ H l(Ep, Q~/vp)o --~ Ep (Fp) ~ 0.
(7.5.7.4) Lemma. Let .~q~pbe an invertible sheaf of degree zero on Ep, and let (~v, V(p-can)) be its image under the section (7.5.7.3). The connection V(p-can) on ~L~pis the unique connection on ~q~pwhich has p-curvature zero. Proof. Let m be an integer prime to p such that (~p)| . Then (~p, V(p-can)) | - (t~p, d), the zero element of HI(Er, t2~p/w~), because ~ p ~ ( ~ C p , V(p-can))is a homomorphism. Thus (5r 17(p-can)) | has p-curvature zero, but its p-curvature is just m times the p-curvature of (5r F(p-can))(viewing both as p-linear mappings from Der(Ep/Fp) to (gE~), and, as m is invertible modp, it follows that (s V(p-can)) has p-curvature zero. As the difference between two connections of p-curvature zero on ~ p is a connection of p-curvature zero on (gE~, it follows from (7.5.3) and (7.1.3) that ~ admits at most one connection of pcurvature zero. Q.E.D. (7.5.8) Construction. Given (&o, V)~HI(E, ~/T)0, for each prime p->__7 prime to n, consider the difference (s
Vv)_ (,Lfp, V(p-can)) ~ F(Ep, t2~/vp).
114
N.M. Katz:
Taking this difference simultaneously for all primes > 7, p,~n, we get a group homomorphism (7.5.8.1) H 1(E, 12~/r)o ---, I-I F(Ep, Q~p/r,).
p~7 p~n
(7.5.8.2) Proposition. The kernel of this homomorphism consists precisely of those (SO, V) which have p-curvature zero for all p > 7, pXn. The inverse image of the torsion subgroup of the target (those "tuples" having almost all components zero) consists precisely of those (SO, I7) which have pcurvature zero for almost all primes p.
Proof This follows from (7.5.7.4) and the definition of(7.5.8.1). Q.E.D.
(7.5.9) The restriction of the homomorphism (7.5.8.1) to the subgroup F(E, f2~/r) of T-connections on the structural sheaf (9E is just the diagonal embedding via simultaneous reduction mod p: (7.5.9.0)
F(E, ~/r)--~ 1-] F(Ep, O~,tr,). 1

p>7 p~'n
Passing to the quotient, (7.5.8.1) induces a homomorphism
I-I r(Ep,
(7.5.9.1)
f2E,/r~) 1
EQ(Q)=E(T)___, ,==7.p~,,
F(E, O~/r)
(7.5.9.2) Proposition. A point Po~EQ(Q) lies in the kernel of (7.5.9.1)/f and only if there exists on the invertible sheaf SO = ( 9 ~ ( [ P ] - [0]) on E a T-connection 17which has p-curvature zero for all p> 7, p~'n. I f such a connection exists, it is unique.
Proof This follows formally from (7.5.8.2).

(7.5.9.3) Proposition. A point PQ~EQ(Q) lies in the inverse image of the torsion subgroup if and only if there exists a non-void open set ql = Spec(Z[1/nm]) in T=Spec(Z[1/n]) over which SO~=dTE~([P]--[0]) admits a connection which has p-curvature zero for all primes p> 7 not dividing nm.
Proof Ifm PQ lies in the kernel, Aa| admits a unique connection 17of p-curvature zero for all p >__ not dividing n. Let q / b e the open subset of 7 T where m is invertible. Because m is invertible on q/, F(E~, I2~/~) is uniquely divisible by m, hence there is a unique connection V' on SO~ xuch that (SO~, 17')| ~ (S~174 17) on E~. For all primes p > 7 not dividing n m, (SO~, 17') has p-curvature zero, because its m-th power does, and pXn (compare the proof of (7.5.7.4)). Conversely, if SO~ admits such a connection 17', let's show that for M a high power of m, (SO~, 17,)| extends to all
115
of E. Since ~ extends to ~ , take a finite covering of E by affine open sets V~ which trivialize Lg, and suppose ~ is given by transition functions f u. Then V' is given by 1-forms ~oi holomorphic on V/c~E~ = Vi[1/m]. By the quasicoherence of f2~/r, there is an power M of m such that M co~ is holomorphic on all of V~. As there are only finitely many V~, a common M works for all. Now ( ~ , 17,)| is given by transition functions ( M ~ , f/jM), which do extend. Q.E.D. Combining (7.5.9.3) with the reduction steps (7.5.4)-(7.5.6), we find (7.5.10) Tautology. Let E be an elliptic curve over T=Z[1/n], with a (non-trivial) rational point of order two. Then the truth of the analogue of (7.4.4) for the function field K/Q of EQ is equivalent to the following conjecture: (7.5.11) Conjecture. The kerne; of the homomorphism (7.5.9.1)
FI rcE ,
EQ(Q)=E(T) ~ p>-7,pg, r(E, n~/r) is contained in the torsion subgroup of EQ(Q).
Appendix Riemann's Existence Theorem

An extremely useful form of the theorem is an easy consequence of GAGA, resolution, and the theory of differential equations with regular singular points (cf. [39, 18, 7, 36]). Theorem. Let S be a smooth connected C-scheme, and let S a~ denote the corresponding complex manifold. Denote by Etale(S) (resp. Etale (Sa~)) the category of finite ~tale coverings of S (resp. San). Then the natural functor Etale (S)-~ Etale(S a~) is an equivalence of categories. Proof. We will explicitly construct an inverse functor. Let f : ~---~ S a~ be finite and 6tale, and put ~ = f . d~r. Because f is finite and flat, ~ is a locally free sheaf of algebras. Because f i s 6tale, it is endowed with a canonical integrable connection V, the direct image of the standard connection "exterior differentiation" on tPer, for which the algebra structure is horizontal. By resolution, there is a proper and smooth C-scheme S containing S as an open set, such that complement of S in S is a divisor D with normal crossings. By the theory of differential equations with regular singular points, there exists for any locally free sheaf with integrable connection (fg, 17) on S ~" a unique pair ((~an, ~a~) consisting of a locally free sheaf ~ , on S~n which prolongs fr and a "connection with logarithmic singularities along D", 17r
8*
f#~a, --~ I2~,n(log O)|
116
N.M. Katz:
which extends V, whose exponents along each local irreducible branch D i of D lie in the strip 0__< e (z)< 1. The pair (fgcan, V~an)is called the R quasi-canonical extension of ((~, V). Although functorial in (~, V), the formation of the quasi-canonical extension does not commute with tensor product. However, given (fg, V) and (~, V') there is a unique horizontal morphism
which prolongs the identity. Applying this with ~ - = f~ = ~ , we find a horizontal morphism
By functoriality, the algebra structure on ~ prolongs to a horizontal morphism Composing these last two maps, we obtain a horizontal multiplication on ~an which extends the given algebra structure on The locally free sheaf of algebras ~an on ~a~ corresponds to a finite flat morphism of analytic spaces
f : ~(" ---~ S an
which prolongs f : Y"--~ San. By GAGA, the morphism ~ comes from a unique finite fiat morphism of proper C-schemes X--~ S, and XIS--> S is a finite 6tale covering of S. This construction is the desired inverse functor Etale (San) -~ Etale (S). Q.E.D. References
0. Atiyah, M., Hodge, W.: Integrals of the second kind on an algebraic variety. Annals of Math. 62, 56-91 (1955). 1. Bateman Manuscript Project (Erdelyi Ed.): Higher transcendental functions, vol. 1. New-York: McGraw-Hill 1953. 2. Cartier, P.: Une nouvelle op6ration sur les formes diff6rentielles. C. R. Acad. Sci. Paris 244, 426-428 (1957). 3. Cartier, P.: Questions de rationalit6 des diviseurs en g6om6trie alg6brique. Bull. Soc. Math. France 86, 177-251 (1958). 4. Clemens, H.: Picard-Lefschetz theorem for families of nonsingular algebraic varieties acquiring ordinary singularities. Trans. Amer. Math. Soc. 136, 93-108 (1969). 5. Deligne, P.: Th6or~me de Lefschetz et crit6res de d6g6n6rescence de suites spectrales. Publ. Math. I. H. E. S. 35, 107-126 (1969). 5a. Deligne, P.: Cohomologie des intersections complhtes. Expos6 XI, SGA 7, 1969. Multigraph available from I. H. E. S, 91-Bures-sur-Yvette, France. 6. Deligne, P.: Travaux de Oriffiths. Expos6 376, S6minaire N. Bourbaki, 1969/1970. Lectures Notes in Mathematics 180. Berlin-Heidelberg-New York: Springer 1971.
117
7. Deligne, P.: Equations diff6rentielles ~ points singuliers r6guliers. Lecture Notes in Mathematics 163. Berlin-Heidelberg-New York: Springer 1970. 8, Deligne, P.: Th6orie de Hodge, Publ. Math. I. H. E. S. 40 (1971). 8a. Dwork, B.: P-adic cycles. Publ. Math. I. H. E. S. 37, 27-115 (1970). 9. Goursat, E.: L'Equation d'Euler et de Gauss. Paris: Hermann 1936. 10. Goursat, E.: Int6grales Alg6briques. Paris: Hermann 1938. 11. Griffiths, P.: Periods of integrals on algebraic manifolds; summary of main results and discussion of open problems. Bull. Amer. Math. Soc. 75, 2, 228-296 (1970). 12. Grothendieck, A., Dieudonne, J.: El6ments de g6om6trie alg6brique, Publ. Math. I. H. E. S. 4, 8, 11, 17, 20, 24, 28, 32. 13. Grothendieck, A.: Fondements de la g6om6trie alg6brique. Secr6tariat math6matique, 11, rue Pierre Curie, Paris 5~, 1962. 14. Grothendieck, A.: On the de Rham cohomology of algebraic varieties. Publ. Math. I. H. E. S. 29 (1966). 15. Hartshorne, R.: Residues and duality. Lecture Notes in Mathematics 20. Berlin-Heidelberg-New York: Springer 1966. 16. Hasse, H.: Existenz separabler zyklischer unverzweigter Erweiterungsktirper vom Primzahlgrade p iiber elliptischen Funktionenk/Srpern der Charakteristik p. J. Reine angew. Math. 172, 77-85 (1934). 17. Hasse, H., Witt, E.: Zyklische unverzweigte Erweiterungsktirper vom Primzahlgrade p iiber einem algebraischen Funktionenk6rper der Charakteristik p. Monatsh. f'tir Math. u. Phys. 43, 477-492 (1936). 18. Hironaka, H.: Resolution of singularities of an algebraic variety over a field of characteristic zero I, II. Annals of Math. 79, 109-326 (1964). 19. Hironaka, H.: Bimeromorphic smoothing of a complex-analytic space. Preprint available from Mathematics Institute, University of Warwick (1971). 20. Hochschild, G.: Simple algebras with purely inseparable splitting fields of exponent one. Trans. Amer. Math. Soc. 79, 477-489 (1955). 21. Honda, T.: Differential equations and formal groups. (Preprint.) Presented at the 1971 U.S-Japan Number Theory conference. 22. Igusa, L: Class number of a definite quaternion with prime discriminant. Proc. Nat'l. Acad. Sci 44, 312-314 (1958). 23. Ihara, Y.: Schwarzian equations I. Preprint, 1971. 23a. Katz, N.: On the differential equations satisfied by period matrices. Publ. Math. I. H. E. S. 35 (1968). 24. Katz, N.: Nilpotent connections and the monodromy theorem; application of a result of Turrittin. Publ. Math. I. H. E. S. 39, 355-232 (1970). 25. Katz, N.: Une formule de congruence pour la fonction ~. Expos6 XXII, SGA 7, 1969, multigraph available from I. H. E. S. 91-Bures-sur-Yvette, France. 26. Kodaira, K., Spencer, D.C.: On deformations of complex structures, I, II, Annals of Math. 67, 328-466 (1958). 27. Manin, Ju.: Algebraic curves over fields with differentiation. AMS Translations (2), 37, 59-78 (1964). 28. Manin, Ju.: The Hasse-Witt matrix of an algebraic curve. AMS Translations (2), 45, 245-264 (1965). 29. Manin, Ju.: Rational points of algebraic curves over function fields, AMS Translations (2), 50, 189-234 (1966). 30. Mazur, B.: Frobenius and the Hodge filtration, Preprint available form Mathematics Institute, University of Warwick (1971). 31. Mazur, B.: Frobenius and the Hodge filtration. (To appear.) 32. Messing, W.: On the nilpotence of the hypergeometric equation. (To appear.) 33. Messing, W.: The crystals associated to Barsoth-Tate groups; with applications to abelian schemes. Thesis, Princeton, 1971. (To appear.)
118
N.M. Katz: Algebraic Solutions of Differential Equations
33a. Messing, W., Mazur, B.: Cristalline cohomology and the universal extension of an abelian scheme. (To appear.) 34. Mumford, D.: Abelian varieties. Bombay: Oxford University Press 1971. 35. Oda, T., Katz, N.: On the differentiation of de Rham cohomology classes with respect to parameters. J. Math. Kyoto Univer. 8, 199-213 (1968). 36. Raynaud, M.: Gromrtrie algrbrique et grom&rie analytique. Expos6 XII, SGA 1. Lecture Notes in Mathematics 224. Berlin-Heidelberg-New York: Springer 1971. 37. Serre, J.P.: Reprrsentations linraires des groupes finis. Paris: Hermann 1967. 38. Serre, J.P.: Sur la topologie des varirtrs algrbriques en caractrristique p. Symposio International de Topologia Algebraica. Mexico, 1958. 39. Serre, J.P.: Gromrtrie algrbrique et grom6trie analytique. Ann Inst. Fourier. Grenoble 6, 1-42 (1956). 40. Tate, J.: W-C groups over p-adic fields. Expos6 156. Srminaire Bourbaki 1957/1958. New-York: W.A. Benjamin 1966. 41. Tessier, B., Lejeune, M.: Quelques calculs utiles pour la rrsolution des singularitrs. Multigraph available from Centre de Mathrmatiques, Ecole Polytechnique, 17, rue Descartes, Paris 5~, (1971). 42. Verdier, J.L.: Base change for twisted inverse image of coherent sheaves. Algebraic geometry. Bombay: Oxford University Press 1968. 43. Weil, A.: Varirtrs K/ihlrriennes. Act. Sci. et Ind. 1267, Paris: Hermann 1958. ,14. Weil, A.: Jacobi sums as "GrGBencharaktere". Trans. Amer. Math. Soc. 73, 487-495 (1952). 45. Whittaker, E.T., Watson, G.N.: A course of modem anal~,sis. Cambridge: Cambridge University Press 1962. Nicholas M. Katz Fine Hall Department of Mathematics Princeton University Princeton, N.J. 08540 USA
(Received May 31, 1972)

Alg Solutions of Diff - Eq

Diunggah oleh

Informasi Dokumen

Deskripsi Asli:

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Alg Solutions of Diff - Eq

Diunggah oleh

Hak Cipta:

Format Tersedia

Inventiones math.

18, 1 - 118 (1972) 9 by Springer-Verlag 1972

Algebraic Solutions of Differential Equations (p-Curvature and the Hodge Filtration)

Algebraic Solutions of Differential Equations

Manin connection. Then 0) P(Xi)(HgR(U/K), 17) is what we mean by a

Algebraic Solutions of Differential Equations

For every integer p-> 0, we define (1.0.2.2) f2xP/r(log D)= Ae~xf2~/T(log D)

a~/s (log D) = A ~ alx/s(log D).

Algebraic Solutions of Differential Equations

Thus, over V, f2~/T(1OgD) is free on (~v with basis (1.0.3.4)

dxj ( j = ~ + l, ..., r),

while f21/s(log D) is free on (_9 with basis v (1.0.3.5)

(v= 1, ..., c0,

1.1. The Kodaira-Spencer Class

which gives rise to an element (1.1.2) p e Ext,,, (f21x/s(log D), f * (t2~/r))

~- H t (X, Dero (X/S) @r f * (I2~/r)).

U ~(S, R ' f , (Dero (X/S) |

~--U ~(S, n I f. (Dero (X/S)) |

Home~(Der(S/T)), R~f. (Dero(X/S))

Consider the exact sequence (1.2.1.4) 0 -~ gr~ -~ K~

whose term of degree v is (1.2.1.5)

Ext,: (~, (~)r Ha(X, Horn(,~, ~))

Ext a (c~, j,~) ~a.2~.~o~, H a(X, nom(~, ~))

Algebraic Solutionsof DifferentialEquations (1.2.2) Interior Product and Cup Product.

I: Hom~x(~, cg) _, Hom~,,(A~, C~|

In terms of local sections fl, ... ,fv of ~ and q~ of Hom(o~, ~),

= (q~ ( f l) + (~oi- ~o~)(f,)) ^... ^ (q~i(f~) + (~o,- q~i)(f~))

A "'" ^ l---^ ... ^ f , f

+ terms in K z (A ~~ ) =I(~oi-q~)(f t ^ . . . ^f~). Q.E.D.

(1.2.2.4.5) Proposition. The mapping "cup-product with ~"

0-~ ~ | A ~- 1~ _ ~ KO/K z (A ~~,uf)_~ A ~~ _ ~ O.

1.3. Application to Hodge Cohomology

Algebraic Solutions of DifferentialEquations

The coboundary mapping associated to (1.3.1.2)

is the cup-product with the Kodaira-Spencer class (1.1.2)

H ~(S, Y2~S/T) R~ f , (Dero (X/S)). |

The construction of the Gauss-Manin connection on

HoR (X/S (log O)) = R f , (f2x/s (log O))

0 - , f * (f2~S/T)--, fJ~/T (log D) - , f2~/s(1Og D) - , 0

gr~(f2]/T(lOgD))'~f * ~2~/T| 12jc~i(log D).

-~ t2]/s (log D) --~ 0.

This "makes sense", because the coboundary is a mapping

17: R~ f , ((2~c/s(log O)) ~ R q+1 L (f* (Q~S/r)| 12]'~S (log D)) 1

~2~/r| Rq f , (t2]/s (log D)).

F i (L" In]) = (F i +" (L')) [n].

17(FiR~f,(g2]/s(log O))) =f2xs/r|

Algebraic Solutions of DifferentialEquations

EP'~=Rqf,(f2~ls(log D)) =~ RP +qf , (I2~/s(log D))

v > f2~i T | grF - ' RP +q f , (f2~cis(log D)) p

gr~ (M) = F p (M)/F p +a(M) ~ gr~ R p +qf, (f2]/s (log D))

Because M is horizontal, it follows from (1.4.1.7) that we have a commutative diagram

(1.4.1.9.4) grV(M) v __~f2~/r | grP- 1(M)

2. The Cartier Operation and the Conjugate Spectral Sequence

Algebraic Solutions of Differential Equations

S The pair of morphisms

2.1. The Cartier Operation

cg-1 : f2jc,~,/s(log Dtp)) ~, ~t~i (F, (Q~/s (log O)))

c~-, (d(a- 1(x))) = the class of x p- 1 dx.

~e_1 ( d~-l(xv)]] a-X(xv)

for v = 1, ..., a for j = ~ + l .... , n.

c~-l{da-l(xi))=the class o f x ~ - ' d x j

Algebraic Solutions of Differential Equations

(~ (coi), co)= (pr. ~(coi), co) = (H-W (a)(F,s(coi)),co)