Lectures on D-modules.

Victor Ginzburg

The University of Chicago

Department of Mathematics
Chicago, IL 60637, USA
[email protected]

with collaboration of
Vladimir Baranovsky and Sam Evens

These are lecture notes of a course given at the University of Chicago in Winter 1998. The purpose of
the lectures is to give an introduction to the theory of modules over the (sheaf of) algebras of algebraic
differential operators on a complex manifold. This theory was created about 15-20 years ago in the
works of Beilinson-Bernstein and Kashiwara, and since then had a number of spectacular applications in
Algebraic Geometry, Representation theory and Topology of singular spaces.
We begin with defining some basic functors on D-modules, introduce the notion of characteristic variety
and of a holonomic D-module. We discuss b-functions, and study the Riemann-Hilbert correspondence
between holonomic D-modules and perverse sheaves. We then go on to some deeper results about D-
modules with regular singularities. We discuss D-module aspects of the theory of vanishing cycles and
Verdier specialization, and also the problem of ”gluing” perverse sheaves. We also outline some of the
most important applications to Representation theory and Topology of singular spaces. The contents of
the lectures has effectively no overlapping with Borel’s book ”Algebraic D-modules”.

1
 These lectures can be divided into two parts. The reader who is mostly interested in D-modules is
advised to go directly to Part 2, and to return to results of Part 1 whenever a reference on such a result is
made. There are only a few places where the results of Part 1 are used in Part 2 in an essential way.
The reader who is more tolerant to Abstract Algebra and is interested in some aspects of ”non-
commutative” Algebraic Geometry may find Part 1 interesting in its own right.

Table of Contents

0. Introduction.

Part 1: Differential operators.

I. Algebraic background.
1. Filtered rings and modules.
2. Three theorems of Gabber.
3. Non-commutative localization and microlocalization.
4. Sato-Kashiwara filtration and Duality.

II. Algebraic differential operators on a manifold.

1. Sheaf of algebraic differential operators.
2. Twisted differential operators (TDO).
3. Twisted cotangent bundles and Lagrangian fibrations.
3. Classification of TDO.
4. Sato’s construction of differential operators on a curve.
5. Application: Riemann-Roch Theorem for curves.
6. Leray residue and local cohomology.

Part 2: D-modules.
III. D-modules: basic constructions.
1. D-modules on a line.
2. Direct image from a submanifold.
3. Restriction to a submanifold; Kashiwara’s theorem.
4. Functors f ! and f∗ for a general map.
5. D-modules and local systems.
6. Simple non-holonomic modules.

IV. Holonomic modules.

1. Category of holonomic modules.
2. b-function.
3. Functoriality for holonomic complexes.
4. Diagram Y ,→ X ←- X \ U and the DGM-extension.
5. Vanishing cycles and f s .
6. Verdier specialization.
7. Fourier transform and the second microlocalization.
 1. Algebraic background.
The same way as an Algebraic Geometry course requires certain background in Commutative Algebra, a
D-module course requires some background in non-commutative algebra. Such a background is given below.

1.1. Filtered rings and modules.

Let A be an associative ring with unit. We call A a filtered ring if an increasing filtration . . . Ai ⊂ Ai+1 ⊂
. . . by additive subgroups is given such that
(i) Ai Aj ⊂ Ai+j ,
(ii) 1S∈ A0 ,
(iii) Ai = A , i.e. the filtration is exhausting.
We will usually consider two choices for the range of the index i:
(a) i ∈ N (b) i ∈ Z.
In case (a) we will call A positively filtered. Note that this case may be viewd as a special case of (b) if we
set A−1 = 0. In the latter case we will consider the topology induced by the filtration (in which the {Ai }i∈Z
form a baseT of open subsets) and usually impose two extra conditions on the filtration:
(iv0 ) Ai = {0} , i.e. the topology defined by Ai is separating.
(iv00 ) A is complete with respect to this topology.
L
Denote by gr A the associated graded ring i Ai /Ai−1 .
1.1.1 Notation. For any element a ∈ Ai we denote by σi (a) the image of a in Ai /Ai−1 ⊂ gr A. We
call σi (a) the i-th principal symbol of a. If a ∈ Ai \ Ai−1 then we write σ(a) instead of σi (a), and say that
deg σ(a) = i.
Warning. The assignment a 7→ σ(a) does not give rise to an additive map A → gr A. Nonetheless, if
a1 , a2 ∈ A and
σ(a1 )σ(a2 ) 6= 0 then σ(a1 a2 ) = σ(a1 )σ(a2 ). (1.1.2)

Definition 1.1.3 A ring A is called almost commutative if gr A is commutative.

We will deal mostly with almost-commutative algebras over a field of characteristic zero.
1.1.4 Example Consider a Lie algebra g over k, a ground field, and let A = Ug be its universal
enveloping algebra. By definition, Ug is a quotient of the tensor algebra T g, explicitly, we have Ug :=
T g/{x ⊗ y − y ⊗ x − [x, y]}x,y∈g . The algebra Ug inherits an increasing filtration T g given by
Ai = {span of monomials of degree ≤ i}.
The Poincare-Birkhoff-Witt Theorem states that in this case gr Ug is isomorphic to the symmetric
algebra Sg. In particular, Ug is almost commutative. The reason for this is that for x, y ∈ g ⊂ U1 g we have
x · y − y · x = [x, y] ∈ g ⊂ U1 g. ♦

Here is a partial converse for the previous example.

Proposition 1.1.5 Let A be apositively filtered k-algebra such that
(i) A0 = k,
(ii) A is generated as a ring by A1 ,
(iii) A is almost commutative.
Then A can be represented as a quotient of the universal enveloping algebra of some Lie algebra g:
A ' Ug/I .
3
Proof. Recall that A1 · A1 ⊂ A2 . By almost commutativity for all a, b ∈ A1 , ab = ba modA1 . Therefore
(ab − ba) ∈ A1 and A1 acquires the structure of a Lie algebra. By the universal property of U(A1 ) we have
an algebra map UA1 → A which is necessarily surjective by (ii). 2
Proposition 1.1.6 (i) If A is positively filtered and gr A is Noetherian,
then A is itself Noetherian.
(ii) If A is Z-filtered and complete and gr A is Noetherian, then A is
Noetherian.
(iii) If gr A has no zero-divisors, then neither has A.

1.1.7 Remark If A is Noetherian, gr A does not have to be Noetherian. ♦

Proof. For any ideal J ⊂ A, the L filtration on A induces a filtration Ji := Ai ∩ J on J, and we have an
associated graded ideal gr J = Ji /Ji−1 ⊂ gr A. If we have an increasing sequence J ⊂ I ⊂ . . . of ideals
in A then the sequence gr J ⊂ gr I ⊂ . . . stabilizes, that is starting from some large enough ideal we have
gr J = gr I = . . . , since gr A is Noetherian.
We deduce from this that the sequence J ⊂ I ⊂ . . . stabilizes. To that end consider J0 and I0 , the 0th
terms of the filtration. Since A0 ,→ gr A, we have J0 = I0 . This and J1 /J0 = gr1 (J) = gr1 (I) = I1 /I0 imply
J1 = I1 , which in turn, combined with gr2 (J) = gr2 (I), implies J2 = I2 , etc....
To prove part (ii), take an ideal J ⊂ A. Then gr J is finitely generated over gr A with generators j̄1 , . . . j̄n .
Lift each j̄i toPji ∈ J, and show by going downward that {ji } generate J as follows. P Let b ∈ Js \ Js−1 and
write σ(b) = i σ(ai1 ) · j̄i for some elements ai1 ∈ A, i = 1, . . . , n.PThen b1 = b − i ai1 ji is an element
of Js−1 , and we can choose elements ai2 ∈ A such that σ(b1 ) P = i σ(ai2 ) · j̄i . Continuing
P in this way
we construct br ∈ Js−r and air ∈ A such that br = br−1 − i air ji and σ(br−1 ) = i σ(air )j̄i . By
completeness P of A the elements
T ai = ai1 + ai2 + ai3 + . . . make sense. Using the separation property we
obtain (b − i ai ji ) ∈ r Js−r = {0}.
Finally, (iii) follows from multiplicativity of the symbol map (1.1.2). 2
Corollary 1.1.8 Ug is Noetherian without zero divisors for any finite dimensional Lie algebra g. 2
Modules over filtered rings. Let A be a filtered ring. An A-module M is said to be filtered if an increasing
sequence of subgroups . . . ⊂ Mi ⊂ Mi+1 ⊂ . . . is given, such that:
Ai · Mj ⊂ Mi+j , ∀i, j.
Again we usually consider two types of filtrations:
(a) If A is positively filtered then we require that the filtration on M is bounded below (that is, M−n = 0
for n  0). We do not require M to be positivly filtered since, for any filtration F• M , the shifted filtration,
F•0 M := F•−1 M makes M a filtered module again, and there is no reason why we should start at grade
degree zero. T
(b) If A is Z-filtered then we usually require that the filtration is separating, Mi = 0, and M is
complete in the topology induced by the filtration (completeness will be important for us because of part
(iii) of Lemma 1.1 below). L
The associated graded space gr M = Mi /Mi−1 has an obvious graded gr A-module structure.
M
Definition 1.1.9 One defines the Rees ring of A by A b := Ai .
i
Alternatively, we can embed A into the ring of Laurent polynomials A[t, t−1 ] and define A b as P ti Ai .
The two definitions are clearly equivalent. The element t in the latter definition corresponds in 1.1 to the unit
of A, vieved as an element 1 ∈ A1 (not of A0 ). Since t is invertible in A[t, t−1 ], the imbedding A
b ,→ A[t, t−1 ]
gets identified naturally with the localization (with respect to t) map A ,→ At , The importance of the Rees
b b

4
ring A
b can be best understood geometrically in the case where A is a commutative k-algebra. Then the
imbedding k[t],→ ti Ai = A b → A1 = Affine line over k .
P b induces a flat morphism of schemes: π : Spec A
This morphism may be thought of as an explicit deformation of the ring A to grA since we have, see e.g.
[CG, ch.2]
π −1 (t) ' SpecA ∀t 6= 0 while π −1 (0) ' Spec(grA) .
M X
In exactly the same way, given a filtered A-module M we define Mc= Mi = ti Mi ⊂ M [t, t−1 ].
i
Clearly A
b is a graded ring, and M
c is a graded A-module.
b It follows immediately from the definitions above
that A/tA ' gr A and M /tM ' gr M
b b c c
Let A be a filtered ring, and M an A-module. We will assume that the filtration on A is given and fixed,
but there is no apriori chosen filtration on M , and such a filtration is up to our choice.
Lemma 1.1.10 The following two conditions on the filtration on M are equivalent:
(i) M
c is a finitely generated A-module
b
(ii) The filtration on M has the form
Mi = Ai−r1 m1 + . . . + Ai−rl ml
for some fixed m1 , . . . , ml ∈ M and r1 , . . . , rl ∈ Z.
If the filtration on A is complete, and M is finitely generated over A, then the above conditions are also
equivalent to the following one:
(iii) gr M is a finitely generated gr A-module.
Proof. (iii) follows from (i) if we choose a set of generators for M c and project them to M c/tMc ' gr M . The
implication (ii) ⇒ (i) is trivial. To show that (iii) implies (ii), let {ui } be the finite set of generators of gr M
over gr A. We cam assume that each ui is homogeneous. Let mi ∈ M be lifts of ui to M . If A is positively
filtered then (iii) implies that Mi = 0Pfor i  0 and we P can proceed by induction: the statement is true for
i  0 and if m ∈ Mi+1 then σ(m) = σ(ai )ui , so m = ai mi (mod Mi ). If A is Z-filtered then we proceed
as in the proof of (1.1) using the completeness property. 2
Definition 1.1.11 A filtration on M is called good if the equivalent conditions (i)-(ii) of Lemma 1.1
hold. If, in particular, A is positively filtered then the filtration on M is good iff gr M is a finitely generated
gr A-module.
From now on we will assume unless otherwise stated that
(1) A is an almost commutative algebra over a field k (⊆ A0 ) of characteristic zero and, moreover,
(2) gr A is a finitely generated k-algebra.
Let M be a finitely generated A-module. Choose a good filtration on M . Then gr M is a module over
gr A, a commutative ring, so we can consider
p the support Supp(gr M ) ⊂ Spec(gr A) with its reduced structure
so that Supp(M ) is given by the ideal Ann(M ) ⊂ gr A.
Definition 1.1.12 (Characteristic variety) The support Supp(gr M ) with its reduced scheme structure
is called the characteristic variety (or singular support) of M and is denoted by SS M . If S is an irreducible
component of SS M , and A0 is its coordinate ring then the rank of the A0 -module gr M ⊗gr A A0 is called the
multiplicity of M at S and is denoted by mult(M, S).
Theorem 1.1.13 (J. Bernstein)
(i) Supp(gr M ) does not depend on the choice of a good filtration on M .
(ii) For any irreducible component S of SS M , the multiplicity mult(M, S) does not depend on the choice
of a good filtration.
Moreover, the multiplicity function mult(•, S) is additive on short exact sequences, that is given
00
0 → M0 → M → M → 0
5
 00
one has: mult(M, S) = mult(M 0 , S) + mult(M , S) whenever mult(M, S) is defined (if S is not an
00
irredicible component of SS M 0 or SS M we set the corresponding multiplicity to be zero).

1.1.14 Remark The theorem fails if we would not take the reduced structure on Supp(gr M ). ♦

Lattices. Let A be a (not necessarily commutative) Noetherian ring, and t ∈ A a central non-zero
divisor. Then one defines, in a standard way, see §1.3 below, the localization, At , of A with respect to the
multiplicative subset {tn }n∈Z .
Let M be a finitely generated At -module.

[ 1.1.15 A subgroup L ⊂ M is called a lattice if L is a finitely generated A-submodule of M

Definition
such that t−k L = M .
k
The following properties of lattices are immediate:
(1) For an exact sequence of modules
0 → M 0 → M → M 00 → 0
and a lattice L ⊂ M , the subgroup L0 = M 0 ∩ L (resp. L00 = { image of L in M 00 }) is a lattice in M 0 (resp.
M 00 ) and one has an exact sequence
0 → L0 → L → L00 → 0
(2) For any two lattices L1 , L2 in M there exist big enough integers, k, l  0, such that
tk L2 ⊂ L1 ⊂ t−k L2
(of course, one can switch the roles of L1 and L2 ).

1.1.16 Notation Given a ring B, write K + (B) for Grothendieck semigroup of finitely generated B-
modules, that is the abelian semigroup freely generated by symbols [N ], for all B-modules N , modulo
relations [N ] = [N 0 ] + [N 00 ] for any short exact sequence 0 → N 0 → N → N 00 → 0.
Clearly, for any lattice L, the quotient L/tL is a module over Ā = A/tA.
Theorem 1.1.17 For any two lattices L, L0 , the classes [L/tL] and [L0 /tL0 ] are equal in K + (Ā).

Corollary 1.1.18 The assignment [M ] 7→ [L] gives rise to a well-defined semigroup homomorphism
K + (At ) → K + (Ā). 2

1.1.19 Remark This is similar to the situation one encounters in representation theory of p-adic groups:
one has natural maps
Qp ←- Zp → Fp .
For any G(Qp )-module M we can choose a lattice L (i.e. a G(Zp )-submodule such that L ⊗Zp Qp = M . Then
L/pL is a module over G(Fp ) which depends on the choice of the lattice L, but whose class in K + (G(Fq ))
does not depend on this choice. ♦

Proof of Theorem 1.1. First consider a special case when L is adjacent to L0 , i.e. tL0 ⊂ tL ⊂ L0 ⊂ L. The
natural short exact sequence
0 → L0 /tL → L/tL → L/L0 → 0
induces the equality [L/tL] = [L0 /tL] + [L/L0 ] of classes in K + (Ā). Another exact sequence
0 → tL/tL0 → L0 /tL0 → L0 /tL → 0
implies that [L0 /tL0 ] = [tL/tL0 ] + [L0 /tL]. But tL/tL ' L/L0 as Ā-modules (since t is invertible, multiplica-
tion by t is an isomorphism). Hence the two equalities imply [L/tL] = [L0 /tL0 ].
6
 In the general case of an arbitrary pair of lattices, L, L0 , we introduce a sequence of lattices Lj = L+tj L0 ,
j ∈ Z. One can easily prove that Lj is adjacent to Lj+1 for all j ∈ Z. Moreover, Lj = L for j  0, while
Lj = tj L0 for j  0. Since
. . . = [Lj−1 /tLj−1 ] = [Lj /tLj ] = [Lj+1 /tLj+1 ] = . . .
by the first part of the proof, and multiplication by tj is an isomorphism, the theorem follows. 2

Lemma 1.1.20 Let A Z-filtered complete ring. If grA is Noetherian then A b is Noetherian.
Proof. We will find a filtration on A
b such that associated graded will be Noetherian. Put
X X X
Fj A
b= ti Ai + ti Aj (= ti Ai + Aj tj+1 [t]).
i≤j i>j i≤j
F b ' (gr A)[t, t−1 ]. Since gr A is Noetherian, gr A[t, t−1 ] is also Noetherian. Hence, grF (A) and A
Then gr (A) b
are both Noetherian, due to Lemma 1.1. 2
Now let A be a filtered ring and M a filtered A-module. From the previous lemma and Lemma 1.1 we
obtain the following relationship between filtrations and lattices.
Lemma 1.1.21 A filtration on M is good iff M c is an A-lattice
b in the A[t, t−1 ]-module M [t, t−1 ]. 2
Therefore the choice of a good filtration corresponds precisely to a choice of a lattice.
Proof of theorem 1.1. For any two good filtrations on M , their respective Rees modules are two lattices in
M [t, t−1 ] which have the same class in K + (gr A) by (1.1). Hence not only their supports are equal but also
they have equal multiplicities along each irreducible component of the support which has maximal dimension.
The additivity of the multiplicity with respect to short exact sequances folows from its definition. 2
1.1.22 Remark It is important for the proof to use K + , not the K-group, since K(gr A) is less friendly
with supports and multiplicities. ♦

1.1.23 Elementary properties of characteristic varieties.

Keep the setup of Theorem 1.1. The general properties of lattices mentioned above imply the following
(1) Suppose M 0 ⊂ M is an A-submodule. If a good filtration {Mi } on M is chosen then the induced filtration
Mi0 = Mi ∩ M 0 on M 0 is also good.
(2) The induced filtration on the quotient module M/M 0 is good and we have a short exact sequence
0 → gr M 0 → gr M → gr(M/M 0 ) → 0
(3) The exact sequence above and standard properties of supports of modules over a commutative algebra
yield
SS(M ) = SS(M 0 ) ∪ SS(M/M 0 )
(4) If M ' A/J then SS(M ) is the zero variety of gr J.
Given a filtered ring A such that grA is a commutative Noetherian ring, we have the scheme Spec(grA).
The grading on grA corresponds geometrically to an algebraic Gm -action on Spec(grA). If A is positively
filtered, then the projection
M
grA  A0 = grA/ gri A gives an inculsion Spec(A0 ) ⊂ Spec(grA) .
i>0

The subscheme Spec(A0 ) is precisely the fixed point scheme of the Gm -action on Spec(grA). Moreover,
Spec(grA) is a cone-scheme over Spec(A0 ), i.e. there is a projection Spec(grA) → Spec(A0 ) induced by the
imbedding A0 ,→ A. Furthermore, the Gm -action contracts Spec(grA) to the fixed point variety, Spec(A0 ),
along the fibers of this projection.
The following simple criterion is quite useful.
7
 Lemma 1.1.24 Let A be a positively filtered algebra such that grA is a finitely generated A0 -algebra,
and M a finitely generated A-module. Then
SSM = SpecA0 ⊂ Spec(grA)
iff M is finitely generated over A0 . 2

1.2. Three theorems of Gabber.

Recall the definition of a Poisson structure.
Definition 1.2.1 A Poisson algebra consists of the following data:
(1) a commutattive (associative) algebra (B, ·) with unit
(2) a Lie bracket (B, { , }).
(3) These two structures are related by the Leibniz identity:
{a1 · a2 , b} = a1 · {a2 , b} + a2 {a1 , b} .

Proposition 1.2.2 If A is a filtered almost commutative algebra then grA has a canonical structure of
a Poisson algebra.
Proof. Let āi ∈ Ai /Ai−1 and āj ∈ Aj /Aj−1 . We will construct an element {āi , āj } ∈ Ai+j−1 /Ai+j−2 such
that the operation {, } together with multiplication in grA will satisfy the definition of a Poisson structure.
To that end, choose a lift ai ∈ Ai (resp. aj ∈ Aj ) of āi (resp. āj ). Form an element ai aj − aj ai . Apriori,
this is an element of Ai+j . But since grA is commutative, (ai aj − aj ai ) ∈ Ai+j−1 . We define {āi , āj } to be
the image of (ai aj − aj ai ) in Ai+j−1 /Ai+j−2 . One can show the class of (ai aj − aj ai in Ai+j−1 /Ai+j−2 does
not depend on the choice of lifts ai and aj (while, of course, (ai aj − aj ai ) itself depends on this lift). 2
1.2.3 Example Let g be a Lie algebra over a field k, and g∗ the dual space. Consider the enveloping
algebra A = Ug equipped with the standard filtration. Then grUg = Sg = k[g∗ ], see Example 1.1. We will
give three equivalent formulas for the Poisson structure on grUg arising from Proposition 1.2:
(1) For x, y ∈ g one has {xn , y m } = (mn)xn−1 y m−1 [x, y], where [x, y]
is the Lie bracket in g.
(2) Choose a base x1 , . . . , xr of g. Each xi gives a linear function on g∗ ,
so {xi }i=1,... ,n is a coordinate system on g∗ . In these coordinates we have
X ∂P ∂Q
{P, Q} = ckij · xk · , P, Q ∈ k[g∗ ] ,
∂xi ∂xj
i,j,k
X
where ckij ∈ k are the structure constants of the Lie algebra, defined by [xi , xj ] = ckij xk .
k
(3) For any λ ∈ g∗ one has
P, Q ∈ k[g∗ ] ,
 
{P, Q}(λ) = hλ, dP (λ), dQ(λ) i ,
where dP (λ), dQ(λ) ∈ g and h , i : g∗ × g → k is the natural pairing. ♦

Definition 1.2.4 Let B be a Poisson algebra. We say that a subvariety V ⊂ Spec(B) is coisotropic if
the ideal IV of functions vanishing on V satisfies
{IV , IV } ⊂ IV .

Theorem 1.2.5 (Involutivity of characteristic variety) Let A be an almost commutative algebra such
that grA is Noetherian, and M be a finitely generated A-module. Then the characteristic variety SS(M ) ⊂
Spec(grA) is coisotropic with respect to the Poisson structure on grA.
8
 √
1.2.6 Comment. In a special case when M = A/J for some left ideal J ⊂ A we have ISS(M ) = grJ. This
special case actually implies the Theorem. Note further that is easy to prove that {gr(J), gr(J)} ⊂ gr(J).
However, in general, given an ideal I in a Poisson algebra, we have
√ √ √
{I, I} ⊂ I ; { I, I} ⊂ I .♦

1.2.7 Remark One recent application of the Involutivity of Characteristic Variety is the proof by
Beilinson-Drinfeld (cf. [BeFM]) of the fact that conformal blocks for the Virasoro Lie algebra are finite-
dimensional, based on Lemma (1.1). ♦

The original proof of the Involutivity of characteristics theorem by Gabber [Ga] was quite tricky. We
present here a simplified version of the argument due to F. Knop.
Consider the dual numbers D := C[]/2 . For each D-module M put M := M/M . For m ∈ M let m be
its image in M . Multiplication by  induces a map M → M . Then M is D-free if and only if this map is
an isomorphism.
Let A be a finitely generated D-algebra and M a finitely generated A-module.
q Assume that both A and
M are D-free. Assume moreover that A is commutative. Consider I := kAnn|A M . Then {I, I} ⊆ I.
Let A be a D-algebra. Assume that A is commutative and that A is D-free. Then for each ā, b̄ ∈ A one
can define a Poisson product {ā, b̄} ∈ A by the formula [a, b] = {ā, b̄}. The theorem of Gabber follows from:
Theorem 1.2.8 Let A be a finitely generated D-algebra and M a finitely generated q
A-module. Assume
that both A and M are D-free. Assume moreover that A is commutative, and put I := kAnn|A M . Then
{I, I} ⊆ I.
Proof of the Involutivity of characteristics theorem. The ideal I can also be defined as the intersection of all
minimal associated prime ideals p of M . Thus it suffices to show {p, p} ⊆ p for each of these p.
Choose x1 , . . . , xl ∈ A such that the xi + p form a maximal algebraically independent set in A/p. Let
R ⊆ A be the subalgebra generated by the xi . Then there is an 0 6= f ∈ R such that B := Af /pf is a free
Rf -module (of finite rank). Since p is a minimal associated prime of M one can find f and s > 0 such that
ps M f = 0 and each of pi M f /pi+1 M f is a free B-module. Thus, one can find m1 , . . . , ms ∈ M such that
the mi form an Rf -basis of M f with the property pmi ⊆ j≺i Rmj . Here j ≺ i means mj ∈ pt M f and
P

mi 6∈ pt M f for some t.
Lemma 1.2.9 For all a, b ∈ A with ā,P b̄ ∈ p one can find
P integers n1 , n2 , n3 ≥ 0 and elements eij ∈ A with
eij ∈ R such that f n1 [f n2 a, f n2 b]mi =  j eij mj and i eii = 0.
We show first that this Lemma implies the Theorem. We have f n1 [f n2 a, f n3 b] = c where c ∈ f t {ā, b̄} + p
with t = n1 + n2 + n3 . The action of c on the free Rf -module M f is given by the matrix (eij ) whose trace is
zero. Since p acts nilpotently, we conclude that the trace of {ā, b̄} on M f is zero. Apply this to ā replaced
by xā for any x ∈ Af . From {xā, b̄} ∈ x{ā, b̄} + pf we conclude that ktr|Rf (x{ā, b̄} : M f ) = 0 for all x ∈ A.
On the other hand, the trace of y := x{ā, b̄} can be calculated as the trace on ⊕i pi M f /pi+1 M f ∼ = B`
for some ` > 0. Thus ktr|Rf (y : M f ) = `ktr|Rf (y : B). The extension B|Rf is (generically) separable, hence
its trace form is non-degenerate. Thus the image of {ā, b̄} in B is zero, i.e., {ā, b̄} ∈ p. 
Proof of the Lemma: Let A0 ⊆ A be the set of a with ā ∈ R. By construction, one can find n2 ≥ 0 and
(0) (1)
uij , uij ∈ A0 such that
X (0) X (1)
f n2 ami = uij mj +  uij mj .
j≺i j

9
Similarily, we obtain
(0) (1)
X X
f n3 bmi = vij mj +  vij mj .
j≺i j
(0) (0) (1) (1)
Define the matrices U = (uij ), . . .
,V := Then (vij ).
X (0) (0)
X (1)
f n2 af n3 bmi = (vij f n2 a + {f n2 a, vij })mj +  vij f n2 amj =
i≺j j
(0)
X X
((V (0) U (0) )ij + {f n2 a, vij })mj (V (0) U (1) )ij + (V (1) U (0) )ij mj .


= +
i≺j j
Thus
(0) (1)
X X
[f n2 a, f n3 b]mi = cij mj +  cij mj .
j≺i j
where
(0) (0) (0)
cij = [V (0) , U (0) ]ij + ({f n2 a, vij } − {f n3 b, uij })
(1)
cij = [V (0) , U (1) ]ij + [V (1) , U (0) ]ij .
(0) (0)
Since ā and b̄ commute we have cij ∈ A. Thus one can find n1 ≥ 0 and dij ∈ R such that f n1 j≺i cij mj =
P
(1)
 j≺i dij mj . Now we define eij = dij + f n1 cij . The trace of the matrix eij is zero since it it the sum of a
P
strictly upper triangular matrix and two commutators. 
The most important application is as follows: let D be a filtered C-algebra whose associated graded
algebra D̃ is commutative. Then the commutator induces a Poisson product on D̃. Let X be a D-module
with compatible filtration. Then the associated graded object X̃ is a D̃-module.
Corollary 1.2.10 qAssume à is a finitely generated commutative C-algebra and X̃ a finitely generated
D̃-module. Let I := kAnn|D̃ X̃. Then {I, I} ⊆ I.
Proof: Let (Dn )n∈Z and (Xn )n∈Z denote the filtrations of D and X repectively. Then apply [?] to
A := ⊕n∈Z Dn /Dn−2 , M := ⊕n∈Z Xn /Xn−2 , and  := 1 + D−1 ∈ D1 /D−1 . 

Let A be as before and assume in addition that grA is the coordinate ring of a smooth connected affine
variety, char(k) = 0. For any finitely generated (hence Noetherian) A-module M define a finite filtration of
M by A-submodules, called the Gabber filtration, by setting
Gi (M ) = { largest A-submodule N ⊂ M such that dim(SSN ) ≤ i}
(“largest” makes sense by the Noetherian property). Alternatively we may define Gi (M ) as follows
Gi (M ) = {m ∈ M | dim SS(A · m) ≤ i}.

Theorem 1.2.11 (Equi-dimensionality) Assume grA = k[X] is the regular ring of a smooth affine
algebraic variety over k, char(k) = 0. Then, for any i = 0, 1, . . . , the characteristic variety of Gi M/Gi−1 M
is of pure dimension i, provided Gi M/Gi−1 M is non-zero. 2

Corollary 1.2.12 If Spec(grA) is smooth and equidimensional, then the characteristic variety of any
irreducible A-module is of pure dimension. 2

1.2.13 Remark The characteristic variety of an irreducible module need not be irreducible. For example,
d
let A = C[z, dz ] be the ring of polynomial differential operators in one variable with the standard filtration
by the order of differential operators, see ch.2 below. Then we have grA = C[z, ξ], so that Spec(grA) = C2 .
d d
Fix λ ∈ C and put Mλ := C[z, dz ]·z λ = A/A·(z dz − λ). Then, SS(Mλ ) = {z · ξ = 0} ⊂ C2 is the ”coordinate
cross” consisting of the two coordinate lines. Now, it is easy to verify, that if λ is not an integer, the A-module
10
Mλ is simple. We see that the characteristic variety of this simple module has two irreducible components
of the same dimension. ♦

We now state the third theorem due to O.Gabber. Let r be a solvable finite dimensional Lie algebra
over a field k of characteristic zero. Suppose further that r can be represented as a sum n ⊕ k · δ such that
(1) n is a nilpotent Lie ideal and
(2) The adjoint action of δ on n is semisimple with strictly positive rational eigenvalues.
Let M be a finitely generated Ur-module. Assume that
there is filtration {Mi }i∈N on M compatible
with Ur-action such that grM is finitely generated over gr Un = Sn (hence, over Sr, in particular).
Theorem 1.2.14 (Separation theorem)
\ If M and n are as above, one has n · M 6= M , or equivalently,
ni M = 0.
i

1.2.15 Remarks. (1) If U+ := n · Un denotes the augmentation ideal in Un, then the theorem above
can be restated as U+ M 6= M and by the Artin-Rees Lemma one has
\
i
U+ M = 0,
i.e. the augmentation filtration on M is separating. This explains why two claims of the Separation Theorem
are equivalent.
(2) Nilpotency of n is essential: if n = sl2 , then any non-trivial finite dimensional simple n-module M
satisfies nM = M . T i
(3) If n is abelian, then U+ M = 0 is a standard fact in Commutative Algebra. Assume first that the
point 0 ∈ Spec(Sn) does not belong to Supp(M ). Then there is a polynomial P ∈ Sn that vanishes on
Supp M and such that P (0) 6= 0. Replacing P by its high enough power we may achieve that P annihilates
M , i.e. P ∈ Ann(M ) ⊂ Sn. But the space Ann(M ) is clearly stable under the adjoint δ-action on Sn.
Moreover, since P = P (0) + P1 , where P1 ∈ n · Sn, and all weights of adδ on n · Sn are strictly positive,
we deduce from P ∈ Ann(M ) that P (0), P1 ∈ Ann(M ). Since P (0) 6= 0 this yields 1 ∈ Ann(M ), a
contradiction. Thus we have proved 0 ∈ Supp M .
Now we can localize M at 0 ∈ Spec(Sn) to get a non-zero module M(0) . But then Nakayama’s lemma
yields, M(0) = nM(0) , hence M(0) = 0, a contradiction. 2
1.2.16 Casselman Theorem in Representation Theory.
The Separation Theorem was discovered as an attempt to find a purely algebraic proof of a theorem of
Casselmann. The latter is a rather deep result in Representation theory originally proved by Casselmann
using analytic methods. To state the Casselman Theorem we need some notation.
Let G be a real semisimple Lie group with Lie algebra g (over R). Let K ⊂ G be a maximal compact
subgroup. We have an Iwasawa decomposition
G=N ·A·K,
where N is a unipotent subgroup, and A is isomorphic to a product of several copies of R>0 . For example
G = SLn (R), K = SUn , A = { diag(α1 , . . . αn ), αi > 0 } ,

N = {upper-triangular matrices with 1 on the diagonal}

Writing n = Lie N , a = Lie A, k = Lie K, we have the corresponding Lie algebra direct sum decompo-
sition (as vector spaces, not as Lie algebras)
g=n⊕a⊕k (1.2.17)

11
 Now let M be an “admissible”, e.g. irreducible unitary, representation of G in a complex (infinite-
dimensional) topological vector space. Write
M fin = {m ∈ M | m belongs to a finite-dimensional K-stable subspace}
By a deep theorem of Harish-Chandra one knows, [Wa], that
(i) M fin is dense in M , and M 7→ M fin is an exact functor.
(ii) The Lie algebra action of any x ∈ g on any m ∈ M fin is well-defined,
i.e. ∀m ∈ M fin , the map g 7→ g · m, G → M is differentiable,
hence M fin is a Ug-module.
(iii) For any “admissible” representation M , see e.g. [Wa], the Ug-module M fin is finitely generated,
and if M is irreducible then M fin is a simple Ug-module.
Assume M fin is simple. Then a version of Schur Lemma yields:
(iv) Z(g) = (center of Ug) acts on M by scalars.
Further, from (iii) we deduce
(v) M fin is generated as Ug-module by a finite dimensional Uk-stable subspace M0 , i.e. M fin = Ug · M0 .

Define a filtration {Mi , i ∈ N} on M fin by Mi = Ui g · M0 where {Ui g} is the standard increasing filtration
on Ug.
Lemma 1.2.18 gr(M fin ) is a finitely generated gr(Un)-module.
Proof. By (1.2.17) we have
Ug ' Un ⊗ Ua ⊗ Uk
Hence by (v) we get M = Ug · M0 = Un · Ua · M0 . Further we have a Harish-Chandra algebra homomorphism
Z(g) → Ua. It is compatible with filtrations and gr(Ua) is a finite module over the image of grZ(g). The
claim can be derived from this and (iv). 2
Next we find δ ∈ a, a generic Q-rational linear combination of simple coroots in a, such that adg δ
is a diagonalizable diagonalizable, its eigenvalues on n are positive integers, and such that Ker adg δ, the
centralizer of δ has minimal possible dimension. Then l = Ker adg δ is a Levi subalgebra in g, and one has a
triangular decomposition
g = n− ⊕ l ⊕ n
where n− is the Lie subalgebra spanned by the negative weight spaces for adg δ. Thus, p := l + n− and l + n−
are the opposite parabolic subalgebras of g with Levi subalgebra l.
Write U+ = n · Un for the augmentation ideal. The Gabber theorem yields
\
Theorem 1.2.19 (Casselman Theorem) i
U+ M fin = 0 . 2

Let O be the abelian category of finitely generated Ug-modules V such that Up-action (recall p := l+n− )
on V is locally finite, i.e.:
dim Up · v < ∞ ∀v ∈ V.
We define an exact faithful functor (called Jacquet functor)
J : Admissible G-representations → O
as follows. First we introduce a naive functor
i
Jb : M 7→ lim
←−
M fin /U+ M fin
i

This functor is faithful by 1.2 and exact (since completion is exact) but the Ug-module J(M
b ) is too large to
be an object of O, it is not even finitely generated as an Ug-module.
12
 We may do better. Notice that M fin /U+ M fin is finite dimensional. It follows that ∀i, the space
i
fin
M /U+ M fin is finite dimensional (since grM fin is finitely generated over gr Un). The action of δ ∈ a
fin
on M induces a δ-action on each finite dimensional space in the following inverse system

2 3
M fin /U+ M fin ← M fin /U+ M fin ← M fin /U+ M fin ← . . . (1.2.20)

One deduces from the positivity of adδ-eigenvalues on n that, for each λ ∈ C, the generalized λ-
j
eigenspaces, (M fin /U+ M fin )λ , (= all Jordan blocks with eigenvalue λ) of δ in (1.2.20) stabilize, i.e., the
j+1 sim j
projection give isomorphisms (M fin /U+ M fin )λ −→ (M fin /U+ M fin )λ , for all j sufficiently large. Let J(M )
be the direct sum of all such “stable” generalized eigenspaces of δ. It is clear that
(i) J(M ) is an U(g)-submodule in J(M
b ) and J(M ) is dense in J(M b ) in the n-adic topology, hence

M 6= 0 ⇒ J(M
b ) 6= 0 (by Casselman Theorem) ⇒ J(M ) 6= 0.
(ii) J is exact, since taking (generalized) eigenvalues is an exact functor;
(iii) The δ-action on J(M ) is locally finite and each generalized eigenspace is finite-dimensional.
(iv) All the eigenvalues are bounded from below by some λ = λ(M ) ∈ C.
Property (iii) follows from the stabilization of eigen-spaces in the inverse system (1.2.20). To prove (iv)
observe that, since the action of l ⊂ g commutes with that of δ, it takes each generalized eigenspace of δ into
itself. Hence Ul-action is locally finite by (iii). The n− -action strictly decreases the eigenvalue of δ, hence,
Un− -action on J(M ) is locally finite by (iv). It follows that J(M ) ∈ O, as promised. 2

1.2.21 Remark An argument similar to the one used above will be used again, in chapter 4, in the
construction of the second micro-localization functor Φ. ♦

1.3. Non-commutative localization and microlocalization.

Definition 1.3.1 Let A be an associative ring with unit 1 ∈ A. A subset S ⊂ A is called multiplicative
if
(1) 1 ∈ S; (2) 0 ∈
/ S; (3) s1 , s2 ∈ S =⇒ s1 s2 ∈ S.
In the commutative situation these conditions are enough to localize A at S. In non-commutative
situation we encounter the following obstacles:
(a) One can form both s−1 a and as−1 and it is not clear which to choose.
(b) It is hard to say when s−1 −1
1 a1 = s2 a2 .
(c) It is not clear how to multiply s1 a1 by s−1
−1
2 a2 .
(d) We don’t have a common denominator for s−1 −1
1 a1 + s2 a2 .
To remove these obstacles one has to impose Ore conditions on S. There are two left Ore conditions
and two right Ore conditions:
(1left ) Every left fraction can be written as a right fraction: ∀s ∈ S, a ∈
A , ∃t ∈ S, b ∈ A such that at = sb (informally, this means: s−1 a = bt−1 )
(1right ) Every right fraction can be written as a left fraction.
(2left ) If s ∈ S, a ∈ A and sa = 0 then ∃t ∈ S such that at = 0.
(2right ) If t ∈ S, a ∈ A and at = 0 then ∃s ∈ S such that sa = 0
Below, we will usually try to escape from having to verify the second condition by requiring that: no
element of S is a zero divisor.
Consider the category whose objects are ring homomorphisms f : B → A such that
(i) B is a ring with unit and f (1) = 1.
13
 (ii) All the elements of f (S) are invertible in B,
and whose morphisms are obvious commutative triangles.
Theorem 1.3.2 (Ore, see [?, ch.3.6]) If Ore’s conditions are satisfied then there exists the universal object
A → S −1 A in this category, i.e. for any morphism f : A → B satisfying (i)-(ii) there is a canonical
commutative diagram
S −1 A
yy< 
yy 
yyy 
y
A / B.
Idea of Proof. One can easily see that the Ore conditions remove obstacles (a) - (d) mentioned above.
Obstacle (a) is handled by the conditions (1lef t ) − (1right ). Obstacle (b) is removed by declaring two left
fractions to be equal if they are equal to the same right fraction. Further, one defines multiplication of left
fractions as follows:
s−1 −1 −1 −1 −1 −1
1 a1 · s2 a2 = s1 (a1 s2 ) a2 = s1 s b a2 = (s s1 )−1 (b a2 ) ,
where we have used the Ore conditions to write : a1 s−1 2 = s
−1
b. Similarly, the Ore conditions ensure that,
given s1 , s2 ∈ S, one can find t1 , t2 such that t1 s1 = t2 s2 = t, and such that t1 ∈ S. It follows that t ∈ S is
a common denominator for s−1 −1
1 a1 + s2 a2 . 2

1.3.3 Digression: Localization of categories. A similar localization techniques applies for additive
categories instead of rings (note that giving an additive category C with one object, X, amounts to giving a
ring A = HomC (X, X); thus rings are just categories with one object).
Let C be an additive category and Φ a family of morphisms in C. Assume that Φ is closed under
composition and contains the identity maps for all objects. Motivated by the ring case, we say that the
category CΦ is a localization of C with respect to Φ if a functor αΦ : C → CΦ is given, satisfying the following
universal property:
For any category D and a functor F : C → D such that F (φ) is an isomorphism
in D for every φ ∈ Φ, there exists a unique functor FΦ : CΦ → D such that F
is naturally isomorphic to FΦ ◦ αΦ .
We claim that if the multiplicative family Φ satisfies obvious analogues of Ore conditions, then the localized
category, CΦ , exists. Specifically, given (C, Φ) we construct CΦ as follows. Put Ob(CΦ ) = Ob(C). Define an
α φ
element of HomCΦ (X, Z) to be a diagram of morphisms X −→ Y ←− Z, where φ ∈ Φ. It is easy to see as in
the proof of Theorem 1.3 that the Ore conditions for Φ ensure the possibility of composing morphisms thus
defined.
The most important example of such a situation is the construction of the derived category of an abelian
category A. Given such an A, let C = C(A) be the category of complexes of objects of A. Recall that a
morphism of complexes is called a quasi-isomorphism provided it induces isomorphisms on the cohomology.
We would like to declare all quasi-isomorphisms to be invertible, i.e. we would like to localize the category
of complexes with respect to the family, Φ, of all quasi-isomorphisms. This family does not satisfy the Ore
conditions, however. To fix the situation, one has first to pass from C(A) to the homotopy category.
In more detail. Let X • ∈ C(A) be a complex. Then Cone(X • ) is a complex such that Cone(X • )i =
X ⊕X i+1 with the differential given by the differential of X. One has the short exact sequence of complexes:
i

0 → X • → Cone(X • ) → X • [1] → 0.

Definition 1.3.4 A morphism of complexes φ : X • → Y • is said to be homotopic to 0 if it factors as

X • → Cone(X • ) → Y • . In other words, one should be able to define morphisms hi : X i+1 → Y i such that
hot
φi = dY hi ± hi dX . We write that φ ∼ 0.
14
Definition 1.3.5 The homotopy category K(A) is defined by
Ob K(A) = Ob C(A),
HomK(A) (X , Y ) = HomC(A) (X • , Y • )/{morphisms homotopic to 0}.
• •

Proposition 1.3.6 The class of quasi-isomoprhisms staisfies the Ore conditions in K(A)
Proof. First note that any map of complexes C → D is homotopic to the embedding C → C ⊕ Cone(C).
Now suppose we have a diagram
a /
A B
qis
φ

C
where a is an embedding and φ is a quasi-isomorphism, then we can apply the pushout construction to
obtain a commutative diagram
0 /A a /B /L /0

φ ψ
 
0 /C b /D /L /0
where ψ is also a quasi-isomorphism (by 5-lemma), and this is exactly what the Ore conditions require us
to prove. 2
Definition 1.3.7 The derived category D(A) is defined to be the localization of the homotopy category
K(A) by the class Φ of all quasi-isomorphisms.

∗ ∗ ∗

We will now give two criteria to verify Ore conditions for a multiplicative subset S of a ring A. First,
given x ∈ A write adx(y) = xy − yx.
Proposition 1.3.8 If for any s ∈ S the operator ad s on A is (locally) nilpotent then the Ore conditions
hold for S.
Proof. Let a, s be as in the first Ore condition. Then, there exists n ∈ N such that
Xn r
0 = (ad s)n (a) = (−1)r sr asn−r .
r=0
n
(The last equality can be easily proved by induction on n.) Hence if take t = sn , we get 0 = asn + sb , where
Xn r
b=− (−1)r sr−1 asn−r . 2
r=1
n
To state the second (less trivial) criterion, assume that we have an almost commutative algebra A such
that grA is finitely generated over an algebraically closed field. Start with a multiplicative subset S̄ ⊂ grA
which contains no zero divisors (this assumption is not imprtant but that is what we will have in applications
anyway). Define
S = {s ∈ A | σ(s) ∈ S̄} .
Proposition 1.3.9 The subset S ⊂ A satisfies the Ore conditions.
Proof. We need to show that ∀a ∈ A, s ∈ S, there is an element t ∈ S such that ta ∈ As. Define
I = {x ∈ A | xa ∈ As}
15
(apriori I could be zero). Hence, if we denote by r(a) the right multiplication map r(a) : A → A, x 7→ xa,
r(a)
then I is the kernel of the composition φ : A −→ A → A/As. Therefore A/I ,→ A/As is an injective map
of A-modules and hence SS(A/I) ⊂ SS(A/As). The point here is that the inclusion is not compatible with
filtrations (since r(a) is not comatible with them) but we use the fact that the characteristic variety does
not depend on the choice of filtration.
With respect to the natural filtrations we have gr(A/I) = grA/gr(I) and gr(A/As) = grA/(grA) · σ(s)
(the latter equality uses that σ(s) in not a zero-divisor). It follows now from the inclusion SS(A/I) ⊂
SS(A/As) established above, that σ(s) vanishes at SS(A/I) = {the zero-variety of grI}. Hence (σ(s)n ∈
gr(I), by Hilbert Nullstellensatz. But then σ(sn ) ∈ gr(I) and therefore ∃t ∈ I such that σ(t) = σ(sn ). Since
S̄ is multiplicatively closed σ(sn ) = (σ(s)n ∈ S̄, hence t ∈ S by definition of S 2
1.3.10 Remark In contrast with the first criterion the element t provided by (1.3) is not a power of s
but one rather has t ≡ sn (modulo lower terms of the filtration). ♦

Suppose that grA has no zero divisors. Then by (1.3) for any multiplicatively closed subset S̄ ⊂ grA we
can construct a subset S ⊂ A satisfying Ore conditions. Hence the localization S −1 A exists.
We introduce the filtration on S −1 A by requiring that s−1 a ∈ Aj−i if deg σ(s) = i and deg σ(a) = j.
This is well-defined since we assume that two left fractions are equal iff they are equal to the same right
fraction and s−1 a = bt−1 implies at = sb hence σ(a)σ(t) = σ(s)σ(b). However this filtration on S −1 A is in
fact a Z-filtration with non-trivial terms in all negative degrees.
Definition 1.3.11 The completion AS̄ of S −1 A in the topology defined by the filtration above is called
the formal microlocalization of A at S. We define the formal microlocalization of an A-module M to be
MS̄ = AS̄ ⊗A M viewed as a left AS̄ -module.

Definition 1.3.12 Let f : M → N be a morphism of filtered A-modules. We say that f is compatible

with filtrations if f (Mi ) ⊂ Ni . We say that f is strictly compatible if f (Mi ) = f (M ) ∩ Ni (i.e. the two
possible filtrations on f (M ) arising from viewing f (M ) either as the surjective image of M or as a subobject
of N , in fact coincide).
1.3.13 Properties of the formal microlocalization.
(1) The natural embedding A ,→ AS̄ is sitrictly compatible with filtrations.
(2) AS̄ is flat over A (since the localization and completion are both exact functors).
(3) gr(AS̄ ) = (S̄)−1 grA (this is true since taking associated graded “does not feel” the completion).
To formulate the fourth property, recall first that the scheme Spec(grA) has a natural Gm -action, and
is cone-scheme over Spec(A0 ). Consider Zariski cone-topology on Spec(grA), i.e. the topology generated by
open cone-subsets U ⊂ Spec(grA). For each open cone-subset U we denote by S̄(U ) the set of all elements
of grA that are invertible on U . This is clearly a multiplicative subset of grA, and abusing the notation we
write AU (instead of AS̄(U ) ) for the microlocalization of A with respect to S̄(U ).
Given a finitely generated A-module M , and an open cone-subset U as above, put MU := AS̄(U ) ⊗A M ,
an AU -module. Then we have
Proposition 1.3.14 (i) The assignment U 7→ AU defines a sheaf of algebras on the cone-scheme
Spec(grA).
(ii) For any finitely generated A-module M , the assignment U 7→ MU gives a sheaf of modules. 2

1.3.15 Remark Notice that in the case when Spec(A0 ) is a point (= the vertex of the cone Spec(grA)),
we have AU = A if U contains Spec(A0 ). One can invert much more elements on A if U does not contain
Spec(A0 ). ♦

16
 Digression: non-commutative determinants.
Recall that to any (n × n)-matrix kaij k ∈ Mn (A) with entries in a commutative ring A one can associate
its determinant, det kaij k ∈ A, given by the standard alternating sum over all permutations s ∈ Sn :
X
det kaij k = (−1)s · a1 ,s(1) · a2 ,s(2) · . . . · an ,s(n) . (1.3.16)
s∈Sn

This formula does not make sense, however, for a noncommutative ring since, in the non-commutative case,
the order of factors in (1.3.16) becomes essential.
Now, let A be an almost commutative filtered ring. We will see that, under certain mild conditions, to
any A-valued (n × n)-matrix P ∈ Mn (A), one can associate in a canonical way an element Det(P ) ∈ grA
satisfying most of the expected properties of a determinant.
Theorem 1.3.17 Assume that grA is a unique factorization domain. Then, there is a natural map
Det : Mn (A) → grA which satisfies the following properties:
(i) Det(P ) · Det(Q)
 = Det(P · Q) , ∀P, Q ∈ Mn (A),
P Q
(ii) Det = Det(P ) · Det(R),
0 R
(iii) P is invertible iff Det(P ) is invertible,
(iv) Let P = kaij k. Assume that ∃l ∈ Z such that, for any permutation s ∈ Sn , a1 ,s(1) ·a2 ,s(2) ·. . .·an ,s(n)
∈ Al , and the following holds:
σ(a1 ,s(1) ) · σ(a2 ,s(2) ) · . . . · σ(an ,s(n) ) 6= 0 and det kσ(aij )k =
6 0.
Then we have Det(P ) = detkσ(aij )k .
Proof. Put S̄ := grA \ {0} and let S ⊂ A be as in (1.3). Then the localization K = S −1 A (without
completion) is a skew-field. The filtration on A induces an increasing Z-filtration on K, moreover, gr(K) is
just the field of fractions of grA.
Given a matrix P over any skew-field, K, one can associate to it its Dieudonne determinant, det(P ) ∈
K × /[K × , K × ], where K × = K \ {0} denotes the multiplicative group of K, [K × , K × ] denotes the derived
group, the multiplicative subgroup of K × generated by the elements ghg −1 h−1 . Recall the construction of
det(P ). First, one puts P into uppertriangular form by elementary transformations in Mn (K). If one of the
elements on the diagonal is zero, we set det(P ) = 0. Otherwise we multiply the elements on the diagonal and
denote the product by det(P ). However, since the diagonal entries do not commute in general, and there is
no natural order in which we should multiply these elements, det(P ) is only well-defined as an element of
K × /[K × , K × ].
^) of det(P ) to K × . Consider the
Next, fixing some order of the diagonal elements defines a lift det(P
^)) ∈ gr(K)× . Any other lift of det(P ) to K × will differ by a product of elements of the type
symbol σ(det(P
−1 −1
ghg h . Since gr(K) is almost commutative, this other lift will have the same symbol in gr(K). Thus,
we have associated to P ∈ Mn (A) a well-defined element Det(P ) := σ(det(P^)) ∈ grK.
All the properties of the theorem hold for Det(•): they follow readily from the corresponding properties
of commutative determinants and multiplicativity of the symbol map. The only thing that remains is the
following claim:
Lemma 1.3.18 ^)) ∈ grA.
For any P ∈ Mn (A) we have σ(det(P
Sketch of proof. Step 1. Assume that grA is a discrete valuation ring. In this case we can easily say when
^)) is integral.
an element is integral, therefore to verify directly that σ(det(P
^)) ∈ grA(p) and
Step 2. Take any prime element p ∈ grA and localize at p. One can show that σ(det(P
\
the rest of the proof follows by the fact that grA = grA(p) . 2
p

17
 1.4. Sato-Kashiwara filtration and Duality.

Let A be a positively filtered almost commutative algebra over a field k of characteristic zero. Put B :=
grA. Then B = ⊕ Bi is a graded k-algebra with augmentation ideal B+ = ⊕i>0 Bi and A0 = B0 ' B/B+ .
Assume further that B = k[X] is the coordinate ring of a smooth affine algebraic variety. The grading
on B gives an algebraic Gm -action on X. The projection B → B/B+ = B0 induces an imbedding i : X0 =
Spec(B0 ) → X = SpecB, and X0 is just the fixed point set of the Gm -action. By assumption, X is smooth.
Hence, X0 is also smooth.
Consider the normal bundle TX0 X of X0 in X. Since X0 is the fixed point set, the group Gm acts along
the fibers of TX0 X. It follows that TX0 X, viewed as the locally free sheaf on X0 , splits into a finite direct
sum of weight components, each of them a locally free sheaf on X0 .
To interpret the normal bundle algebraically, recall that, for any graded B-module N the quotient
N/B+ N is naturally a B0 -module and if N is finitely generated over B then N/B+ N is finitely generated
over B0 . We see that the weight decomposition of TX0 X corresponds to the decomposition of the B0 -module
2
B+ /B+ into its graded components.
2
Lemma 1.4.1 Assume that B+ /B+ is free over B0 and that N/B+ N is free and finitely generated over
B0 . Then if N is projective graded (i.e. a graded direct summand of a free graded B-module) then N is free
over B.
Proof. Note that, for any graded B-module N , any set of B0 -generators in N/B+ N lifts (non-uniquely) to
a set of B-generators in N . By our assumption N ⊕ N 0 ' B ⊕k . Then N/B+ N ⊕ N 0 /B+ N 0 ' B0⊕k . Choose
a base of N/B+ N over B0 and lift it to a set of generators of N over B. There can be no relations between
these generators as can be seen by passing to the field of fractions of B0 . 2
Lemma 1.4.2 Let M be a filtered A − module such that grM has a finite presentation
0 ← grM ← B ⊕k ← B ⊕l .
If grM/B+ grM is free over B0 then one can lift the presentation above to the non-filtered level
0 ← M ← A⊕k ← A⊕l
so that all maps are strictly compatible with filtrations.
Sketch of proof. The first arrow grM ← B ⊕k of the presentation gives k generators m1 , . . . , mk of grM (of
degrees d1 , . . . , dk respectively) which we lift to elements m01 , . . . , m0k of M . Consider the exact sequence
0 ← M ← A⊕k ← Ker ← 0
Notice that A⊕k has a shifted filtration:
k
M
A⊕k


j
= Aj−di .
i=1

First of all we need to show that A⊕k j maps surjectively onto Mj (that will imply that the first arrow is

surjective and strictly compatible with filtrations). To that end, we induct on j. Take an element m of Mj
and write σj (m) as a linear combination of mi ’s. If we try to lift this linear combination to M , it will not in
general be equal to m, but will differ from m by
an element of Mj−1 . By inductive assumption any element
of Mj−1 is an image of some element in A⊕k j−1 . Hence, any element of Mj is an image of some element
of A⊕k j .

The arrow B ⊕k ← B ⊕l gives us l relations between m1 , . . . , mk . We lift these relations to non-graded

level as follows. Let b1 m1 + . . . bk mk = 0 be one such relation (and deg (b1 ) + d1 = deg (b2 ) + d2 = . . . =
deg (bk )dk =: d). Suppose that bi = σ(ai ). Then a1 m01 + . . . ak mk is not necessarily zero but rather an
element of Md−1 . By the strong compatibility of the first arrow we have a1 m01 + . . . ak mk = a001 m01 + . . . a00k mk
18
where a00i ∈ Md−di −1 . But then σ(ai − a00i ) = σ(ai ) = bi hence (a1 − a001 )m01 + . . . + (ak − a00k )m0k = 0 is the
lift of the relation form grM to M .
The l lifts of relations to M define a map A⊕l → Ker . We need to show that this map is surjective
and strictly compatible with filtrations. The proof of this repeats the proof of correspoding facts for the lift
of the first arrow of the resolution. 2

Let A be a positively filtered almost commutative ring. Notice that for any a0 ∈ A0 the operator
ad a0 : A → A is nilpotent, since it sends Ai to Ai−1 for all i = 0, 1, . . . . Therefore, any multiplicative subset
S of A0 satifies the Ore conditions (cf. 1.3) and we can localize A with respect to S. The localization is
positively filtered (hence, in particular, we do not have to take the completion).
Corollary 1.4.3 Given a graded projective resolution of grM
. . . → N2 → N1 → N0 → grM → 0
one can (localizing with respect to A0 if necessary) lift this resolution to a free resolution of M itself
. . . → Ak2 → Ak1 → Ak0 → M → 0 ,
where all morphisms are strictly compatible with filtrations.
Proof. We can shrink Spec(A0 ) to make the resolution free (since both
Spec(A0 ) and Spec(grA) are smooth). Then we apply the lemma above step by step to lift the resolution.
2
Corollary 1.4.4 If X = Spec(grA) is smooth, then the algebra A has the same homological dimension
as grA, that is dA = dim X. 2

Proposition 1.4.5 Let M be finitely generated A-module. Then

(i) codim SS(ExtjA (M, A)) ≥ j, ∀j.
(ii) ExtjA (M, A) = 0 unless codim (SSM ) ≤ j ≤ dim X.
Proof. If we replace A by grA , M by grM , and ”SS” by ”Supp”, the corresponding statement in commutative
algebra is well-known, see e.g. [Eis].
Now, given M , we first choose a free graded resolution of grM , and then use Lemma 1.4 to lift it to a
free filtered resolution . . . → F1 → F0 → M , with Fi finitely generated. Then ExtA (M, A) can be computed
using the complex
0 → HomA (F0 , A) → HomA (F1 , A) → . . . (1.4.6)
of free right A-modules. Observe that the Hom-group between any two filtered objects, X, Y , has a
natural filtration: a morphism φ : X → Y is contained in the lth term of the filtration, if (Xi ) ⊂ Yi+l , ∀i.
Thus, all the groups HomA (Fj , A) are filtered. Furthermore, it is easy to verify that grHomA (Fj , A) =
HomgrA (grFj , grA). Hence the spectral sequence of the filtered complex (1.4.6) implies that grExtjA (M, A)
is a subquotient of Extjgr A (grM, grA). Since passing to a subquotient does not increase the support, the
proposition follows. 2

Duality. Given a left A-module M define its naive dual to be HomA (M, A). The multiplication on A on
the right induces a right A-module structure on HomA (M, A). Similarly, given a right A-module M define
its naive dual to be HomA (M, A), a left A-module. The main defect of this naive construction is that the
module HomA (M, A) is often zero (e.g. when M = A/Aa for some element a ∈ A).
The standard way to go around this difficulty is to use the language of derived categories. Let Db (A-mod)
be the derived category of bounded complexes of projective left A-modules, and write Db (mod-A) for a similar
19
derived category of right A-modules. Then, the naive duality above, applied term by term to a bounded
complex . . . → Pi → Pi+1 → . . . of projective left A-modules gives a bounded complex
. . . → HomA (Pi+1 , A) → HomA (Pi , A) → . . .
of projective right A-modules. This way one gets duality functors
D : Db (mod-A)  Db (mod-A) . (1.4.7)

We say that the ring A has finite homological dimension dA , provided any finitely-generated left A-module
M has a projective resolution of length ≤ dA . This is equivalent to the requirement that ExtiA (M, N ) = 0
for any A-modules M , N , and all i > dA . In such a case any finitely-generated left A-module M gives rise
to a well-defined object of Db (A-mod) represented by its projective resolution. Thus, applying the above
duality one gets a well-defined object of DM ∈ Db (mod-A).
1.4.8 Example
(i) We have DA = A.
(ii) If M = A/Aa then we have a free resolution of M :
·a
0 → A −→ A → M → 0
·a a·
which we can use to compute RHom. Since the dual of the complex 0 → A −→ A → 0 is 0 → A −→ A → 0
we have D(A/Aa) = A/(aA)[1] where [1] stands for the shift of complexes in the derived category.

1.5. Sato-Kashiwara filtration.

Throughout this subsection A stands for a filtered algebra over a field of characteristic zero such that
SpecgrA is a smooth algebraic variety. In particular, by Corollary 1.4, A has finite homological dimension
d = dim SpecgrA.
qis
Lemma 1.5.1 There is a functorial quasi-isomorphism in Db (A − -mod): DD(M ) ' M .
Proof. The assertion is true for a free module, hence for a projective module M . The proof is now completed
by induction on the length of finite projective resolution. 2
Definition 1.5.2 Consider a complex of abelian groups
di−1
i d
K• = {. . . → Ki−1 −→ Ki −→ Ki+1 → . . . }.
The i-th truncation, τi≤ K• , is defined as
τi≤ K• = {. . . → 0 → Im(di−1 ) ,→ Ki → Ki+1 → . . . } .
We have a natural morphism of complexes K• → τi≤ K• , given by
di−1
... / Ki−2 / Ki−1 / Ki di
/ Ki+1 / ...

di−1
 
... /0 / Im(di−1 ) - Ki
di
/ Ki+1 / ...

Also, it follows from the definition that


0 for j < i,
H j (τi≤ K• ) =
H j (K• ) for j ≤ i.
This cohomological property implies that the truncation functor gives rise to a well defined functor on the
derived category.
20
 For a finitely generated A-module M , and any i, we have the object τi≤ (DM ) ∈ Db (A-mod), and
the canonical morphism DM → τi≤ (DM ). Applying the functor D to this morphism we get a canonical
morphism D(τi≤ (DM )) → DD(M ) ' M
Definition 1.5.3 Given an A-module M , and any i ∈ Z, write Sj (M ) for the image of the induced map
on zero cohomology
H 0 (D(τi≤ (DM ))) → H 0 (DM ) = M .
The {Sj M }j=0,1,... form an increasing filtration of M by A-submodules (it is clear that Sj M = 0 if
j < 0). This filtration is called Sato-Kashiwara filtration.
1.5.4 Remark Since M has a finite projective resolution, the Sato-Kashiwara filtration is finite too. In
fact, Sj (M ) = M if j > dA since the corresponding truncation complex vanishes. ♦
Recall that earlier we have defined the Gabber filtration G• M of M by
Gi M = {largest A-submodule M 0 ⊂ M such that dim SS(M 0 ) ≤ i}.
Gabber filtration is relatively easy to define but hard to handle while Sato-Kashiwara filtration, due
to its abstract definition, has good functorial properties. Fortunately, the following theorem says that they
coincide anyway.
Theorem 1.5.5 The Gabber filtration is equal to the Sato-Kashiwara filtration, i.e.
Gi (M ) = Si (M ) , ∀i = 0, 1, . . . .

Before we prove (1.5) we need two lemmas.

Lemma 1.5.6 Sato-Kashiwara filtration is functorial, i.e. for any
morphism f : M1 → M2 of A-modules one has
f (Sj M1 ) = Sj M2 .

Proof. Everything is functorial. 2

For an A-module M denote by d(M ) the dimension of SS(M ).
Lemma 1.5.7 (i) d(Sk (M )) ≤ k,
(ii) Sd(M ) (M ) = M .
Proof. To prove the lemma we reformulate (i) and (ii) above in the following way:
(i0 ) d(Si (M )/Si−1 (M )) ≤ k, ∀i ≤ k,
(ii0 ) Si (M )/Si−1 (M ) = 0 ∀i > d(M ).
We now prove (i0 ) and (ii0 ) by using the spectral sequence for the composition IdDb (A-mod) = D ◦ D of the
duality functors, applied twice. The spectral sequence reads:

d(M )−i Si (M )/Si−1 (M ) if i = j
E2i,j = ExtjA (ExtA i,j
(M, A), A) ⇒ E∞ =
0 otherwise
This spectral sequence arises as follows. Choose a finite projective resolution of M
0 → M−d → . . . → M−2 → M−1 → M0 → M → 0 ,
so that DM can be computed using the complex {HomA (Mi , A)}−d≤i≤0 . To apply the duality functor again,
we have to replace this complex by a quasi-isomorphic complex of projectives. It is more convenient however
to use a different approach and to choose instead a (possibly infinite) resolution of the (A-A)-bimodule A
by graded injective bimodules:
0 → A → I0 → I1 → . . . .
21
Then DDM can be computed using the simple complex associated with the double complex K i,j =
HomA (DMi , Ij ). Note that since Mi is projective, DMi is projective. But for a free module, hence for
the projectives Mi , one has canonical isomorphisms:
K i,j = HomA (DMi , Ij ) ' HomA (HomA (Mi , A), Ij ) ' Ij ⊗A Mi .
Thus, the spectral sequence above becomes the standard spectral sequence of the following double complex:
0 / I0 ⊗A M−d / I0 ⊗A M−d+1 / ... / I0 ⊗A M0 /0

  
0 / I1 ⊗A M−d / I1 ⊗A M−d+1 / ... / I1 ⊗A M0 /0

. . . . . . . . .

  
0 / Id ⊗A M−d / Id ⊗A M−d+1 / ... / Id ⊗A M0 /0

  
Now we can apply (1.4) for the Ext-groups in this spectral sequence (twice). The assertion follows. 2
Proof of (1.5). By definition of Gabber filtration and (1.5(i)) we have Sk (M ) ⊂ Gk (M ). To show the
opposite inclusion, consider the map Gk (M ) ,→ M . By functoriality of the Sato-Kashiwara filtration (1.5)
we get
Si (Gk M ) ⊂ Si (M ).
Take i = k. Since d(Gk (M )) = k we can appply (1.5(ii)) to get Gk (M ) = Sk (Gk (M )) ,→ Sk (M ). 2

Lemma 1.5.8 Taking Gabber filtration commutes with microlocalization

Proof. Formal microlocalization is exact and hence commutes with truncation functors. 2
Proof of Gabber’s Equi-dimensionality Theorem. Recall that we need to prove that Gi M/Gi−1 M is of
pure dimension i. Suppose that SS(Gi M/Gi−1 M ) is not of pure dimension i. Then it has a component Y
of dimension dim Y ≤ i − 1. Choose an open affine subset U ⊂ X such that
U ∩ SS(Gi M ) ⊂ Y \ Y sing ,
Taking microlocalization at U we clearly have
(M/Gi−1 (M ))U 6= 0 , (1.5.9)
since SS((Gi (M )/Gi−1 (M ))U ) is a subset of SS(M/Gi−1 (M ))U containing Y \ Y sing .
On the other hand d(MU ) ≤ i − 1 since by localizing at U we excluded the components of characteristic
variety of dimension ≥ i. Therefore MU = Gi−1 (MU ) = (Gi−1 (M ))U . Hence localizing the short exact
sequence
0 → Gi−1 (M ) → M → M/Gi−1 (M ) → 0
at U we obtain (M/Gi−1 (M ))U = 0 which contradicts (1.5.9). 2

22
 2. Algebraic differential operators on a manifold.
2.1. Sheaf of algebraic differential operators.
2.1.1 Quasi-coherent sheaves.
We will assume that all varieties are defined over the field of complex numbers C.
Let X be an algebraic variety and U be its affine open subset. For any regular function f ∈ O(U ) we
denote by Uf the open subset U \ f −1 (0). Given an O(U )-module M , set Mf = M [f −1 ] = O(Uf ) ⊗O(U ) M .
Lemma 2.1.2 The following conditions on a sheaf F of OX -modules are equivalent:
(i) F is a direct limits of its coherent subsheaves.
(ii) For any Zariski open affine subset U ⊂ X and any f ∈ O(U ) one has Γ(Uf , F) = Γ(U, F)f . 2
Definition 2.1.3 A sheaf of OX -modules satisfying either of the two equivalent conditions above is
called quasi-coherent.
Now we present two different approaches to differential operators on X.

2.1.4 Coordinate approach. We define a notion of “local coordinates” on X as follows. Assume that
X is a smooth affine variety (all the constructions will be local in Zariski topology). Fix a point x ∈ X.
By Noether Normalization Lemma we can find an open neighbourhood U ⊂ X of x and a finite unramified
morphism φ : U → V of U to an open affine neghbourhood V of {0} ∈ Cn (where n = dim X) such that
φ(x) = 0. In fact, and n regular functions φ1 , . . . , φn with linearly independent differentials dφ1 , . . . , dφn
in U will define such a map φ. While it is not true that φ gives local coordinates on U , it still induces an
isomorphism of tangent spaces. Hence any vector field on V can be lifted in a unique way to U . We can see
that algebraically it we recall that B = Γ(U, OX ) is a finite extension of A = Γ(V, OCn ). Hence any element
b ∈ B satisfies some monic equation bk + a1 bk−1 + . . . + ak = 0 with a1 , . . . , ak ∈ A. If ∂ is a derivation of
A then, differentiating the equation above, we obtain
k bk−1 ∂(b) + (k − 1)a1 bk−2 ∂(b) + bk−1 ∂(a1 ) + . . . + ∂(ak ) = 0

and this defines ∂(b) uniquely.

Denote by x1 , . . . , xn the standard coordinates on Cn and by ∂i , i = 1, . . . , n, the lift of the vector field
∂/∂xi from V to U . Then the vector fields ∂i on U commute with each other and Γ(U, TX ) = O(U )·∂1 +. . .+
O(U ) · ∂n since ∂1 , . . . , ∂n generate the tangent space at each point of U . We call the pullbacks of x1 , . . . , xn
to U an algebraic “local coordinate” system. It must be kept in mind that these “local coordinates” do not
separate points in the same fiber of φ : U → V .
Definition 2.1.5 We define the sheaf DX of differential operators on X as the subsheaf End C (OX ) of
all complex-linear endomorphisms of OX generated locally at each point by OX and TX .
Lemma 2.1.6 In algebraic “local coordinate” system we have
Γ(U, DX ) = O(U )[∂1 , . . . , ∂n ]
where the RHS stands for the polynomial algebra over O(U ) generated by commuting vector fields ∂1 , . . . , ∂n .
Proof. An exercise left to reader (one checks that any section u of DX can be uniquely expressed in terms
of ∂1 , . . . , ∂n by applying u to pullbacks of functions from V ). 2
In particular DX is a quasi-coherent sheaf.
Corollary 2.1.7
i
(i) The filtration {DX }i∈Z+ of DX by the order of the operator is preserved under coordinate change,
thus making DX into a filtered sheaf of rings.
i
(ii) For any i ∈ Z+ the subsheaf DX of differential operators of order ≤ i is coherent.
(iii) The associated graded sheaf gr DX is isomorphic to the symmetric algebra STX of the tangent
bundle TX . In particular, DX is almost commutattive. 2
23
 Corollary 2.1.8 (Poincaré-Birkhoff-Witt for differential operators) Let U be any affine open subset of
X such that the module of global sections Γ(U, TX ) is free over Γ(U, OX ) with generators v1 , . . . , vn . Then
any differential operator on U of order ≤ k can be written uniquely in the form
X
fk1 ...kn v1k1 . . . vnkn .
k1 +...kn ≤k

Proof. The vector fields v1 , . . . , vn may no longer commute. From commutation relations one can deduce
the existence of the representation above, while uniqueness can be proved by covering U with open subsets
on which “local coordinates” exist. In fact, it suffices to prove that
X
fk1 ...kn v1k1 . . . vnkn 6= 0 (2.1.9)
k1 +...kn ≤k

if at least one of the functions fk1 ...kn is not zero. Without loss of generality we can assume that such
function satisfies k1 + . . . kn = k. By shrinking U if necessary, we can find functions φ1 , . . . , φn with linearly
independent differentials that we deonte by ∂1 , . . . , ∂n . Now rewrite (2.1.9) in terms of ∂i instead of vi and
apply (2.1). 2
Since STX is an associated graded of DX , it carries a natural Poisson bracket {·, ·}gr DX . One also has a
Poisson structure on STX on it arising from symplecticP topology as follows. Consider the cotangent bundle
∗ ∗
π : TX → X with its natural symplectic form ω = dpi ∧ dqi . For any two functions f and g on TX we
∗
have corresponding Hamiltonian vector fields ξf and ξg . Hence ω(ξf , ξg ) is a function on TX that we denote
by {f, g}symplect . Since STX ' π∗ (OTX∗ ) one also has an induced Poisson bracket {·, ·}symplect on STX .
Proposition 2.1.10 The two brackets {·, ·}gr DX and {·, ·}symplect on
STX coincide. 2
2.1.11 Coordinate-free interpretation of DX (Grothendieck).
Definition 2.1.12 Let A be a commutative ring. For any pair of A-modules M , N we define modules
DiffkA (M, N ) inductively by putting
(1) Diff0A (M, N ) = HomA (M, N ),
(2) Diffk+1 k
A (M, N ) = {additiive maps M → N, s.t. ∀a ∈ A, (au − ua) ∈ DiffA (M, N )}.
k k+1
It follows from the definition that DiffA (M, N ) ⊂ DiffA (M, N ). We put
DiffA (M, N ) = ∪k DiffkA (M, N ).

2.1.13 Exercise Show that if M = N then Diff(M ) is a filtered almost commutative ring.
2.1.14 Remarks.
(1) Note that for any a ∈ A the operator [·, a] of commuting with a maps Diffk+1 (M, N ) to Diffk (M, N )
hence it is locally nilpotent. Recall that this is precisely one of the properties that guarantees that the
Ore conditions are satisfied (in the case M = N ). This property is necessary is we want to show that Diff
can be glued into a quasi-coherent sheaf of algebras (since we want to be able to localize Diff. In a sense,
Diff ⊂ Hom C (M, N ) is the lagrest subspace on which the adjoint action of A is locally nilpotent.
(2) The abelian group Hom C (M, N ) has a natural structure of an (A − A)-bimodule. Since A is
commutative, we can view Hom C (M, N ) as a sheaf on Spec A × Spec A. Since the ideal of the diagonal
∆ ⊂ Spec A × Spec A is generated by functions of the type (a ⊗ 1 − 1 ⊗ a), we can say that Diff(M, N ) is
the largest submodule of Hom C (M, N ) that is supported on the diagonal.
(3) If X is not smooth it is hard to say anything about Diff(M, N ) (for example, identify the associated
graded of Diff(M, M )). Moreover, for a singular X the sheaf of rings Diff(M, M ) may not even be Noetherian.
Hence from now on we will assume that X is a smooth algebraic variety and M, N are two coherent
sheaves on X.
24
 Definition 2.1.15
(i) We define the sheaf D(M, N ) by gluing Diff(M, N ) on open affine subsets, i.e. by requaring that for any
affine open subset U ⊂ X
Γ(U, D(M, N )) = DiffO(U ) (M(U ), N (U )).

(ii) The sheaf DX of differential operators is define by DX := DX (OX , OX ).

2.1.16 Exercise Let M, N be locally free OX -sheaves. Show that

DX (M, N ) = N ⊗OX DX ⊗OX M∗ .
Show also that this isomorphism is compatible with filtrations and
gr DX (M, N ) ' Hom(M, N ) ⊗ STX . 2
2.2. Twisted differential operators (TDO).

Definition 2.2.1 A sheaf D of twisted differential operators (TDO for short) on X is a positively filtered
sheaf of almost commutative algebras together with an ismorphism of Poisson algebras ψD : gr D ' STX .

2.2.2 Example DX is a TDO on X.

It follows from the definition that for any TDO D on X one has D0 ' OX . Also one has a bracket
[·, ·] : D1 × D0 → D0 coming from commutator of elements in D. This bracket by its definition satifies the
Jacobi identity. Hence every ∂ ∈ D1 induces a derivation of D0 = OX which we denote by ∂. ¯ In particular,
1
one has a map of sheaves D → TX . One easily checks that this gives rise to a short exact sequence
0 → OX → D1 → TX → 0.
In fact, a TDO structure on a given filtered sheaf of algebras can be reconstructed from the following
data:
(a) An isomorphism D0 ' OX and almost commutativity on D.
(b) The fact that the commutator of element induces a short exact sequence 0 → OX → D1 → TX → 0
wher OX is identified with D0 and hence D1 /D0 is isomorphic to TX .
(c) The induced bijective map STX → gr D is bijective (then its automatically a Poisson algebra isomor-
phism).
In fact, the three properties above may be taken as a definition of a TDO.

2.2.3 Exercise Prove that two TDO D1 and D2 are isomorphic if and only if
(a) D1 ' D2 as sheaves and
(b) the two embeddings OX ,→ D1 , OX ,→ D2 agree with the isomorphism above.
(HINT: define Di1 as {u | [u, OX ] ⊂ OX }, etc.)
The previous exercise essentially says that the TDO structure is uniquely defined by the embedding
OX → D.

2.2.4 Example If L is a line bundle on X then DX (L, L is a TDO.

0
2.2.5 Remark For a vector bundle E of rank greater than 1, DX (E, E) is not a TDO since DX (E, E) '
HomOX (E, E).

Definition 2.2.6 An Atiyah algebra on a smooth algebraic variety X is a sheaf (A, [·, ·]) of OX -modules
with a Lie bracket [·, ·] such that
(a) There exists a short exact sequance of OX -modules
0 → OX → A → TX → 0
25
(we denote the map A → TX by a 7→ ā),
(b) [a1 , f · a2 ] = (a¯1 f )a2 + f [a1 , a2 ] and
(c) 1A ∈ OX ⊂ A is a central element of A.
2.2.7 Remark It is immediate from (b) and (c) that OX is an abelian ideal in A.
2.2.8 Example For any TDO D the first term D1 of the filtration on D is an Atiyah algebra.
Lemma 2.2.9 The set of isomorphism classes of Atiyah algebras has a natural structure of a vector space.
Moreover, there is a linear map from this vector space to Ext1 (TX , OX ) given by forgetting the Lie bracket.
Proof. First of all, we let any λ ∈ C act on the set of classes of Atiyah algebras by multiplying the map
A → TX (and leaving the embedding OX ,→ A unchanged. The sum of two Atiyah algebras is defined via
the Bauer ??? sum of extensions (one checks that this standard explicit constructions actually gives a Lie
bracket on the sum of two extensions). 2
2.2.10 Construction of the Atiyah class. Any Atiyah algebra A by its definition gives a class in
Ext1 (TX , OX ). Since TX and OX are locally free, one has Ext1 (TX , OX ) = H 1 (X, Hom(TX , OX )) =
H 1 (X, Ω1 ). The class in H 1 (X, Ω1 ) given by A is called Atiyah class of A.
1
2.2.11 Example. If L is a line bundle then DX (L, L) is a TDO hence DX (L, L) is an Atiyah algebra
1 1
defining a class in H (X, Ω ). This class is nothing but the first Chern class of L. If L is a vector bundle of
1
rank ¿ 1, then DX (L, L) is not a TDO.
Definition 2.2.12 Given an Atiyah algebra A, we define UX (A) to be the quotient of the sheaf UC (A)
of the universal enveloping algebras by the relations
1UC (A) = 1A ; f ⊗ a = f · a, ∀f ∈ OX ⊂ A, ∀a ∈ A.

2.2.13 Remark This definition imitates an attempt to consider the universal enveloping algebra of A
over OX . This is of course not possible, since OX is not central in A. However, the effect of the relations
above is that UX (A) has the “size” of DX . More precisely, one has the following
Proposition 2.2.14 For any Atiyah algebra A, UX (A) is a TDO. 2
We can construct another object from A which is rather similar to UX (A) and can be easily confused
with it.
Definition 2.2.15 Define SX (A) to the the quotient of the sheaf SC• (A) of symmetric algebras on X,
by the following relations:
1SC• (A) = 1A ; f ⊗ a = f · a, ∀f ∈ OX ⊂ A, ∀a ∈ A.

Proposition 2.2.16 For any Atiyah algebra A, SX (A) is a sheaf of filtered commutative algebras on X
such that
0
(i) SX (A) = OX ,
i j i+j−1
(ii) SX (A) has a natural Poisson bracket {·, ·} : SX (A) × SX (A) → SX (A),
(iii) SX (A)/SX (A) ' TX and gr SX (A) ' STX . 2.
1 0 i

Definition 2.2.17 A sheaf of filtered algebras satisfying the properties of (i) - (iii) of the proposition
above is called a twised symmetric algebra of the tangent bundle TX .
Theorem 2.2.17.1. There exist natural bijections between the following classes of objects
{TDO} ←→ {Atiyah algebras} ←→ {Twisted symmetric algebras}. 2

26
 2.2.18 Exercise Given an Atiyah algebra A, construct a flat family D≈ of sheaves of filtered associative
algebras on X, parametrized by t ∈ CP1 such that
(1) For t 6= ∞ one has Dt = UX (t · A) (in particular, D1 = UX (A) and D0 = DX ),
(2) D∞ = SX (A),
(3) gri (D) ' S i TX (i) as sheaves on X × CP1 .
(HINT. Use two filtrations on UX (A): the natural filtration of the enveloping algebra and the filtration
induced
X by a two-step filtration on A in which A0 = OX and A1 = A. Then form an object of the type
i j
x y Mij where x and y are homogeneous coordinates on CP1 .)
i,j≥0

2.2.19 Remark The sheaf Dt is an example of a mixed twistor structure of Simpson.

2.3. Twisted cotangent bundles and Lagrangian fibrations.

Definition 2.3.1 A twisted cotangent bundle on X is an affine bundle π : T # X → X modeled on vector
∗
bundle TX → X such that
(i) The total space of T # X is endowed with a symplectice form,
(ii) The fibers of π are Lagrangian with respect to this symplectic form,
(iii) The direct image π∗ (OT # X ) is a sheaf of commutative algebras with the Poisson bracket {·, ·}
sending π∗ (OT # X )≤1 × OX to OX , where π∗ (OT # X )≤1 stands for polynomials of degree ≤ 1 along the fibers
on π.
2.3.2 Remarks.
(1) Notice that on an affine space one only has a well-defined notion of a polynomial finction of degree
≤ i. This is because we don’t have a preferred zero point, hence the notion of a homogeneous polynomial of
degree i is not well-define since it is not preserved by affine transformations.
(2) The property (iii) above is equivalent to
{π∗ (OT # X )≤i , π∗ (OT # X )≤j } ⊂ π∗ (OT # X )≤i+j−1 .

2.3.3 Example Any Atiyah algebra A gives rise to a twisted cotangent bundle T A X defined by
T A X = {φ ∈ A∗ | hφ, 1A i = 1}.

Proposition 2.3.4 There exists a canonical isomorphism of Poisson algebras

SX A ' π∗ (OT A X )
which gives a bijection between twisted cotangent bundles and twisted symmetric algebras. 2

2.3.5 Exercise Show that in C ∞ -category the all isomorphism classes of twisted cotangent bundles are
of the form  
T ∗ X, ω + π ∗ (β) /modulo exact forms β,
where ω is the standard symplectic 2-form on T ∗ X and β is a closed 2-from on X which we pull back on
T ∗ X. Hence in the C ∞ -situation T # X is always a vector bundle and symplectic forms on it giving rise to
a twisted cotangent bundle structure, are parametrized by H 2 (X, R).
2.3.6 Twisted cotangent bundle associated with a linear bundle.
Let us construct explicitly the twisted cotangent bundle T L X corresponding to the Atiyah algebra
1
DX (L, L) of differential oprators of order ≤ 1 on a line bundle L. To that end, take the total space L of
the principal C∗ -bundle corresponding to L. The natural C∗ -action on L lifts canonically to a Hamiltonian
C∗ -action on T ∗ L. Moreover, there is a canonical choice of a moment map

∗
µ : T ∗ L → Lie C∗ = C
27
(this is because any element x ∈ Lie C∗ gives rise to a vector field on X which can be viewed as a function
on T ∗ L).
Claim 2.3.7 T L X is naturally isomorphic to the symplectice reduction µ−1 (1)/C∗ . Moreover, the quotient
µ−1 (0)/C∗ is isomorphic to the usual cotangent bundle TX
∗
. 2
2.3.8 Remark Denote by φ the projection from the C∗ bundle L to X. Then the Claim above is nothing
but a commutative version of the followig statement: the sheaf φ∗ (DL) is naturally isomorphic to the sheaf
∗
(DL )C of C∗ -invariant differential operators on L.
Proposition 2.3.9 There exists a canonical connection on the pullback π ∗ L of L to T L X with curvature
equal to the standard symplectic form on T L X. Moreover, any connection on L is obtain from this canonical
connection on T L X via a uniquely defined section X → T L X of π. 2
2.3.10 Lagrangian fibrations and holomorphic coordinates.
Proposition 2.3.11 Let M be a symplectic manifold and assume that a map π : M → X to a manifold
X is a Lagrangian fibration (i.e. its fibers are Lagrangian subvarieties). Then the fibers of π have canonical
affine structure.
Proof. Of course, since π is not assumed to be proper, the affine structure means a local action of Rn ,
i.e. n commuting vector fields. Let α be a local section of T ∗ X. Then π ∗ α is a 1-form on M and we can
find a unique vector field ξ on M such that ω(ξ, ·) = π ∗ ω. It follows that ξ is tangent to the fiber. Hence we
proved that locally M is a principal homogeneous space over the vector bundle T ∗ X. In particular, M has
a local affine structure. 2
2.3.12 Remarks.
(1) There is another way to define the affine structure: define O≤0 to be the normalizer on π ∗ O with
respect to the Poisson bracket on M . Then the Hamiltonian vector fields obtained from the sections of π ∗ O
are tangent to the fibers and define an affine structure on them.
Proposition 2.3.13 If fibers of the Lagrangian fibration π above are compact then each fiber is a torus,
i.e. a quotient of a vector space by a maximal rank lattice.
Proof. By compactness the local action of Rn on each fiber can be integrated to a global action, hence
the fiber becomes a compact homogeneous space of Rn which is necessarily a torus. 2
We now give a very important application of the proposition above: a construction of canonical holo-
morphic coordinates in a neighbourhood of a point on a Kähler manifold. This construction is due to
Bershadsky-Cecoti-Ooguri-Vafa.
2.3.14 Construction.
Let M be a Kähler manifold with a Kähler form ω. Denote by M the same manifold with conjugate
complex structure. The underlying real manifold MR can be embedded diagonally into M × M . Denote by
OR the sheaf of complex-valued R-analytic functions on MR . Any section f of this sheaf can be expanded
locally in a series in zi and z̄j , local holomorphic and antiholomorphic coordinates on M . If we view zi as
functions on M and z̄j as functions on M , then we can extend f locally to a function in the neighbourhood
of MR in M × M . We assume that ω is a real-analytic form (this assumption is not essential but it simplifies
the construction). Consider the corresponding section ωC of the complexified tangent bundle TC MR . Note
that TC MR ' TM ×M |MR . Since ω is a closed (1, 1)-form, it follows that the extension of ω to a small
neightbourhood of MR in M × M is holomorphic. Moreover, each fiber of the projection pr2 : M × M → M
is Lagrangian. By proposition above each fiber acquires an affine structure. Such an affine structure on
the fiber over x ∈ M define a holomorphic exponential map exphol x : Tx M → U , where U is an open
neighbourhood of x in M . This exponential map is defined only in some neighbourhood since ωC can be
extended only to a neighbourhood of MR in M × M . Moreover, this exponential map exphol x depends on x
in an antiholomorphic way (since we take the conjugate struture on M .
28
 2.3.15 Exercise
(1) Prove that for the Fubini-Studi metric on the Riemann sphere CP1 , the exponential map coincides
with the standard stereographic projection form the point opposite to x ∈ CP1 .
(2) Generalize the result of (1) to Grassmanians.
2.3.16 Remark The existence of canonical holomorphic coordinates on the moduli space of complex
structures on a Calabi-Yau manifold has been recently used in Mirror Symmetry (some equations look
especially nice in these local coordiates).

2.4. Classification of TDO.

Definition 2.4.1 A TDO D is said to be locally trivial if the embedding of algebras OX ,→ D is locally
isomorphic to the standard embedding of algebras OX ,→ DX .
2.4.2 Example For any line bundle L, the TDO D(L) is locally trivial.
Theorem 2.4.2.2. Locally trivial TDO’s on a smooth variety X are classified by the first cohomology
1
group HZar (X, Ω1cl ), where Ω1cl stands for the sheaf of closed algebraic differential forms and the cohomology
is taken in the zariski toplogy.
First we will prove the following “meta-lemma”.
Lemma 2.4.3 The objects P of “sheaf nature” on X are classified by the first cohomology group
H 1 (X, Aut P) of the sheaf Aut P of automorphisms of P if the latter is independent of P.
Proof. A standard argument invloving choosing a Cech covering {Ui } of X and considering transition
functions φij ∈ Γ(Ui ∩ Uj , Aut P). 2

Proof of (2.4.2.2) We have seen bofore that the embedding OX ,→ D defines the TDO structure on D
uniquely. Hence, by the meta-lemma above we have to establish the isomorphism of sheaves
Ω1cl ' Aut(OX ,→ DX )
since locally OX ,→ D is isomorphic to OX ,→ DX . In other words we have to find a group of automorphisms
1
of DX as a TDO. Any such automorphism is uniquely defined by its restriction to DX ' OX ⊕ TX . Since
1
the embedding OX ,→ DX is fixed, any automorphism of DX is given by sending a section ξ of TX to a
section with the same symbol, which is necessarily of the form ξ + φ(ξ) for some function φ(ξ). Since the
correspondence ξ 7→ ξ + φ(ξ) is to be OX -linear, it is given be a global 1-form φ. By Cartan’s formula, the
operation ξ 7→ ξ + φ(ξ) preserves brackets if and only if φ is closed. Hence Ω1cl ' Aut(OX ,→ DX ). 2

2.4.4 Example Let us give an example of a TDO which is not locally trivial. We will describe only the
correspondong Atiyah algebra A. We take A to be a direct sum OX ⊕ TX as a sheaf of OX -modules and
define the bracket [·, ·] by the rule
[f1 ⊕ ξ1 , f2 ⊕ ξ2 ] = (ξ1 · f2 − ξ2 · f1 + β(ξ1 , ξ2 )) ⊕ [ξ1 , ξ2 ],
where β is a 2-form on X. The Jacobi identity for this bracket is satisfied if and only if β is closed. If we
now suppose that β is not locally exact, then the corresponding TDO will not be locally trivial. Of course,
this never happens in the holomorphic setup but it is quite possible in the algebraic situation.

To classify the TDO’s in the general case we consider the truncated De Rham complex
d
Ω≥1 := Ω1 −→ Ω2cl .

29
 Theorem 2.4.4.3. The TDO’s on a smooth variety X are classified by the first hypercohomology group
H 1 (X, Ω≥1 ) of the truncated De Rham complex Ω≥1 . Moreover, the short exact sequence of complexes
Ω1 Ω1
     
0
0→ ↓ → ↓ → ↓ →0
Ω2cl Ω2cl 0
induces the long exact sequence of (hyper)cohomology
. . . → H 0 (Ω1 ) → H 0 (Ω2cl ) → H 1 (Ω≥1 ) → H 1 (Ω1 ) → . . .
and the last arrow is given by the Atiyah class of the corresponding Atiyah algebra.
Proof. Choose an affine covering X = ∪Ui so that the restriction of any Atiyah algebra A to Ui is
given as an OUi -module by the direct sum OUi ⊕ TUi . Since the symbol of the bracket on A is fixed, any
A is necessarily given by some closed 2-form βi on Ui as in the example above. We reflect it by writing
A|Ui = Aβi .
Moreover, on double intersection Ui ∩ Uj any map Aβi → Aβi is necessarily given by
f 7→ f, ξ 7→ ξ + αij (ξ),
where f is a function, ξ is a vector field and αij is a 1-form on Ui ∩ Uj .
Hence, with respect to the covering X = ∪Ui , any Atiyah algebra A is given by the data
{αij , βi }, αij ∈ Ω1 (Ui ∩ Uj ), βi ∈ Ω2cl (Ui ).
These data a required to satisfy the following conditions
βi − βj = d αij on Ui ∩ Uj , αij + αjk + αki = 0
which say exactly that {αij , βi } represents a class in H 1 (Ω≥1 ). 2

2.4.5 Remark To compare locally trivial TDO’s with all TDO’s note that there is an embedding of
complexes
Ω1cl Ω1
   
 ↓ → ↓ 
0 Ω2cl
In holomorphic situation this embedding is a quasi-isomorphism since Ω1 → Ω2cl is surjective and Ω1cl is the
kernel of it. In the lagebraic situation this map may not me a quasi-isomorphism and in general it only
induces a map of the corresponding hypercohomology groups.
2.4.6 Vector bundle case. If E → X is a vector bundle of rank r > 1 then the ring of differential
operators D(E, E) is not a TDO since D0 (E, E) ' End(E). Hence the extension
0 → End(E) → D1 (E, E) → TX ⊗ End(E) → 0
is not an Atiyah algebra because of its size.
We consider a diagram
0 / End(E) / D1 (E, E) / TX ⊗ End(E) /0
O O

0 / End(E) / A1 / TX /0

 
0 / OX / A2 / TX /0

30
the second row of which is obtained from the first as a pullback with respect to TX = T ⊗OX ,→ TX ⊗End(E)
Tr
and the third is obtained from the second as a pushout with respect to End(E) −→ OX . Notice that the
standard construction of pullback and pushout endow A1 and A2 with a bracket. This barcket defines the
structure of an Atiyah algebra on A2 .
Theorem 2.4.6.4. Let E be a rank r vector bundle on X and A2 = A2 (E) be the Atiyah algebra obtained
from the diagram above. Then there exists a natural isomorphism of Atiyah algebras r · A2 (E) ' A(Λr E). 2
2.5. Sato’s construction of differential operators on a curve.
To give an illustration, suppose we consider the ring of smooth differential operators on the real line R.
The basic example of such an operator is ∂ = ∂/∂ x. If f is a smooth function on R then one can write
Z
∂ · f = − δ 0 (x − y)f (y)dy, where δ 0 is the Dirac delta function.
R
This formula suggests that differential operators on X should have something to do with “functions” on the
diagonal in X × X.
Let now X be a smooth algebraic curve. Denote by ∆ the diagonal in the Cartesian product X × X
which we sometimes identify with X itself. Let OX×X (∞ · ∆) be the sheaf
of functions with pole of any
finite order along ∆. We will also write ΩX  OX (∞ · ∆) for ΩX  OX ⊗ OX×X OX×X (∞ · ∆).
Theorem 2.5.0.5. (Sato) One has a natural isomorphism
ΩX  OX (∞ · ∆)
' DX
ΩX  OX
of sheaves on X × X supported on ∆ ' X.
Sketch of proof. For any section φ of ΩX  OX (∞ · ∆) we construct a differential operator uφ on X
which acts on functions on X as follows. Choose such a function f and denote the local coordinates on the
factors of X × X by z and w respectively. We put
Z
(uφ f )(w) = f (z)φ(z, w), for ε small enough .
|z−w|=ε

In other words, we pull f back to X × X via the first projection, multiply it by φ and take the fiberwise
residues along the points on diagonal with respect to the second projection.
dz
If φ is written locally as (z−w) k+1 then integrating by parts we obtain

Z
f (z) d z
(uφ f )(w) = =
|z−w|=ε (z − w)k+1

∂ k f (z) d z  ∂ k f (z) 
Z
=± k (z − w)
= ± (w)
|z−w|=ε ∂ z ∂ zk
and the rest of the proof follows by this local computation. 2

More generally, let E and F be two vector bundles on X. We put E ∨ = E ∗ ⊗OX ΩX .

Theorem 2.5.0.6.
(i) There exists a canonical isomorphism of sheaves
(E ∨  F )(∞ · ∆)
' DX (E, F ).
E∨  F
Moreover, the two OX module structures on DX (E, F ) (left and right) correspond under this isomorphism
to the two OX -actions on E ∨  F (via pullback under the first and the second projections).
31
 (ii) One the level of filtrations one has
(E ∨  F )((k + 1)∆) k
' DX (E, F ).
E∨  F
Proof. Take a section φ of E ∨  F (∞ · ∆). For any section e of E, the product e · φ is a section of
ΩX  F (∞ · ∆) hence when we take the residue over the first factor of X × X we will be left with a section
of F . The fact that (E ∨  F )((k + 1)∆)/E ∨  F indeed maps to DX k
(E, F ) is shown by a local computation
as in the proof of the previous theorem. We will show that this map
is an isomorphism by showing that it
induces an isomorphism on graded objects. Indeed, since O(∆)/O |∆ is isomorphic to the tangent sheaf
TX on ∆ ' X, one has
(E ∨  F )((k + 1)∆)
' S k+1 TX ⊗ (E ∗ ⊗ Ω) ' E ∗ ⊗ F ⊗ S k TX . 2
(E ∨  F )(k∆)
2.5.1 Exercise Understand how the ring structure on DX (E, E) apperas from Sato construction (i.e.
how to find the product in terms of sections with poles and residues).
2.5.2 Left and right D-modules.
To establish the connection between left and right D-modules, we first notice that any rigth D-module
may be viewed as a left Dop -module. Since D is a TDO, Dop is also a TDO. We will compute Dop using
Atiyah algebras. In fact, given any Atiyah algebra
0 → OX → A → TX → 0
we can define the second OX -module structure on A (which we identify by writing the fuctions of the right
while for the old OX -module structure we write functions on the left):
a · f := f · a + ā(f ).
where ā stands for the image of a in TX . Of course, this is not compatible with the old OX -module structure
on A, i.e. in general (g · a) · f 6= g · (a · f ). We also replace the old bracket [·, ·] on A by −[·, ·]. This new
OX -module structure together with the new bracket define a structure of an Atiyah algebra on the same
sheaf A that we denote by Aop .
Proposition 2.5.3
(i) For any A, one has the following identity between isomorphism classes of Atiyah algebras:
[A] + [Aop ] = [AΩX ].
op
(ii) DX ' DX (Ω) ' Ω ⊗OX DX ⊗OX Ω−1 . Hence if a sheaf E has a structure of a left (resp. right)
DX -modules then the tensor product Ω ⊗OX E has a structure of a left (resp. right) Dop -module.
(iii) For any left DX -module M, the tensor product ΩOX M has a natural right DX -module structure.
2
2.5.4 Example Let us consider the case M ' OX , i.e. describe the canonical right DX -module structure
on ΩX . If ξ is a vector field and ω is a top-degree differential form, the we can form a Lie derivative Lξ ω
but it does not define a left DX -module structure since
Lf ·ξ ω = f · Lξ ω + ξ(f ) · ω 6= f · Lξ ω.
However since we have an equality of differential operators
ξ · f = f · ξ + ξ(f )
hence ξ : ω 7→ −Lξ ω defines a right DX -module structure on Ω.
In general, for any left DX -module M the formula
ξ : (ω ⊗ m) 7→ −ω ⊗ ξ · m − (Lξ ω) ⊗ m
gives a right DX -module structure on Ω ⊗OX M.
32
 If E and F is a pair of vector bundles on X then for any differential operator u ∈ DX (E, F ) one can
consider ut ∈ DX (F ∨ , E v ee).
Proposition 2.5.5
(i) Given three vector bundles E, F and V on X and opertors u ∈ DX (E, F ), w ∈ DX (F, V ), one has
(u ◦ w)t = ut ◦ w.
(ii) For any vector bundle E one has a natural isomorphism of sheaves of rings
op
DX (E, E) ' DX (E ∨ , E ∨ ). 2

2.5.6 Remark One can interpret the proposition above in terms of Sato’s construction as follows. the
isomorphism
(E ∨  F )(∞ · ∆)
DX (E, F ) '
E∨  F
allows us to write
(F ∨∨  E ∨ )(∞ · ∆) (F  E ∨ )(∞ · ∆)
DX (F ∨ , E ∨ ) ' ∨∨ ∨
=
F E F  E∨
and the last term now coincides with DX (E, F ) with the order of the factors flipped.

2.6. Application: Riemann-Roch Theorem for curves.

Let X → S be a smooth mosphism of smooth algebraic varieties and E be a vector bundle on X. One
can define a line bundle Rdet(π∗ E) on S by a formula
 O  O 
Rdet(π∗ E) = det(Ri π∗ E) ⊗ det−1 (Ri π∗ E) .
i even i odd

One can show that even though the individual sheaves Ri π∗ E are not locally free, the expression above
defines a locally free sheaf of rank one on S.
Then the ordinary Grothendieck-Riemann-Roch theorem gives an answer to the following question:
Compute the Chern class of Rdet(π∗ E) in terms of Chern classes of E.
The Riemann-Roch theorem in Arakelov geometry deals with the following situation:
Given a hermitian metric h on E, compute the induced Quillen metric on Rdet(π∗ E).
The Atiyah algebra of Riemann answers the question which is somewhat half the way betweem ordinary
Riemann-Roch and Arakelov Riemann-Roch
1
Compute the Atiah algebra of Rdet(π∗ E) in terms of DX (E, E).

2.6.1 Remark The problem of this type arises when one cosideres natural determinant bundles on
various moduli spaces, i.e. moduli Mg,n of genus g curves with n points or the moduli space of bundles on
a fixed curve, etc.
the logical connection is weak here

We assume that X is a smooth projective curve. Let De denote the quotient (ΩX  OX )(∞ · ∆)/(ΩX 
OX )(−∆). By Sato’s construction one has a natural extension
Ω X  OX
0→ →D
e → DX → 0
(ΩX  OX )(−∆)
33
Notice that ΩX  OX /(ΩX  OX )(−∆) is nothing but ΩX . Similarly, for any vector bundle E on X we can
define the sheaf D(E)
e and consider the diagram of sheaves on X

0 / Ω ⊗O End(E) / D(E)
e / DX (E, E) /0
X

 
0 /Ω / D(E)
b / DX (E, E) /0

where the second row is obtained as a pushout of the first under the trace map Ω ⊗OX End(E) → Ω. Now
the second row gives a long exact sequance of cohomology and in particualr the connecting homomorphism
∂
H 0 (X, DX (E, E)) −→ H 1 (X, ΩX ).
Since we assumed that X is smooth and projective H 1 (X, ΩX ) is canonically identified with C and for any
u ∈ H 0 (X, DX (E, E)), ∂(u) is just a complex number. Notice also that any such differential operator u
induces a C-linear transformation of each cohomology group H i (X, E).
Proposition 2.6.2 For any operator u ∈ H 0 (X, DX (E, E)) we have
∂(u) = T r(u, H 0 (X, E)) − T r(u, H 1 (X, E)).

Corollary 2.6.3 In the situation above, ∂(1) is equal to the Euler characteristic χ(E) of E. 2
Sketch of proof of (2.6).
Step 1. One has the following cahin of isomorphisms:
H •+1 (X × X, E  E ∨ ) ' ⊕k H k (X, E) ⊗C H •+1−k (X, E ∨ ) '

∗
⊕k H k (X, E) ⊗C H k−• (X, E) ' End•C H ∗ (X, E).
In particular, one has maps
∆∗
End0C (H ∗ (X, E)) ' H 1 (X × X, E  E ∨ ) −→
∆∗ Tr
−→ H 1 (X, E ⊗ E ∗ ⊗ Ω) −→ H 1 (X, Ω) = C.
It follows from the definitions that the composition φ : End0C (H ∗ (X, E)) → C of maps above takes u to
T r(u, H 0 (X, E)) − T r(u, H 1 (X, E)).
Step 2. The exact sequence
o → E  E ∨ → E  E ∨ (∞ · ∆) → DX (E, E) → 0
gives a connecting homomorphism H 0 (X, DX (E, E)) → H 1 (X × X, E  E ∨ ) = End0C H ∗ (X, E)
Now the proof of the theorem follows from the following claim
Claim 2.6.4 The diagram
∂ /
H 0 (DE ) H 1 (E  E ∨ ) H 0 (E) ⊗ H 1 (E ∨ ) ⊕ H 1 (E) ⊗ H 0 (E ∨ )
NNN
NNN
NNN
NN&
End0 H • (E) End H 0 (E) ⊕ End H 1 (E)
commutes (where DE stands for D(E, E), ∂ is the connecting homomorphism and the diagonal arrow is
given by the action on cohomology).
Proof. Since we are essentially proving an equality of two operators, it suffices to check it on a particular
vector. We will only consider the End H 0 (E) part (the other part is similar).
34
 Therefore want to prove that the diagram
H 0 (DE ) ⊗ H 0 (E) / H 1 (E  E ∨ ) ⊗ H 0 (E)
ii
ii iiiii
iii a
 iiii 
End0 H • (E) ⊗ H 0 (E)
b / H 0 (E)

commutes (where a is given by projecting H 1 (E E ∨ )⊗H 0 (E) to H 0 (E)⊗H 1 (E ∨ )⊗H 0 (E) ' End H 0 (E)⊗
H 0 (E) and applying the action map End H 0 (E) ⊗ H 0 (E) → H 0 (E), similarly for b).
Pick a point x ∈ X and let Ex be the fiber of E over x. We need to show that the following diagram
commutes
H 0 (DE ) ⊗ H 0 (E) / Ex ⊗ H 1 (E ∨ ) ⊗ H 0 (E)

action
 
H 0 (E)
restriction to x / Ex

In fact, take the element ψ ⊗ f ∈ H 0 (DE ) ⊗ H 0 (E). Mapping it to H 0 (E) and then the result to Ex we
obtain Resx2 (ψ(x, x2 )f (x2 )) (this follows from definitions). To compute the other composition, take a unit
disk in some local coordinates centered at x and call it Xin . Also denote X \{x} by Xout . There is a theorem
saying that both Xin and Xout are Stein manifiold, hence cohomologically trivial. Then by Meyer-Vietoris,
Γ(Xin ∩ Xout , E ∨ )
H 1 (E ∨ ) =
Γ(Xin , E ∨ ) + Γ(Xout , E ∨ )
In thhis represetation we see that an element ψ(x, x2 ) in fact gives an element of H 1 (E ∨ ) and that the image of
ψ⊗f is also given by Resx2 (ψ(x, x2 )f (x2 )) since the residue is equal to zero for ψ ∈ Γ(Xin , E ∨ )+Γ(Xout , E ∨ ).
(This should be said better) 2

2.7. Leray residue and cohomology with support.

Let us assume first that we deal with Hausdorf topology on C-analytic sets. Choose an open disk U in
Cn = {x1 , . . . , xn } centered at 0 (so that 0 ∈ U is given by equations {x1 = 0, . . . , xn = 0}). Denote by U ∗
the punctured disk U \ {0}. Cover U ∗ by open subsets Ui = U \ {xi = 0}, i = 1, . . . , n.
Consider a holomorphic differential form ω ∈ Ωn (U1 ∩ . . . Un ). The residue Res x=0 (ω) of ω at x = 0 is
defined by restricting ω to (S 1 )n = {x ∈ U ∗ such that |xi | = 1, i = 1, . . . , n} and computing the integral
Z Z Z
Res x=0 (ω) = ... ω
| {z }
n
1 n
over (S ) .
We can also consider U = Cn with Zariski topology and put Ui = Cn \ {xi = 0}, i = 1, . . . , n. For any
algebraic differential n-form ω ∈ Ωnalg (U1 ∩ . . . ∩ Un ) the residue Res x=0 (ω) is defined via the same formula.
2.7.1 Cohomological interpretation.
Let X is any topological space and A is a sheaf of abelian groups on X. Choose an open covering {Ui }i∈I
of X such that for any finite subset J ⊂ I we have H j (∩i∈J Ui , A) = 0 for all j ≥ 1 (such a covering is called
A-acyclic). It is well known that, for such a covering, the cohomology groups H j (X, A) can be computed
via the Cech complex C • ({Ui }, A).
In our situation above (holomorphic or algebraic) the open sets Ui form a covering of U ∗ = U \ {0}.
One can prove that this covering is Ωn -acyclic (in the algebraic situation this is true since all Ui and their
intersection are affine, in the holomorphic setting one uses (??)). Therefore the Cech complex C • ({Ui }, Ωn )
computes H • (U ∗ , Ωn ).
35
 Any holomorphic (or algebraic) n-form ω represents represents a Cech n − 1 cocycle (since there are no
(n+1)-multiple intersections). Coboundaries in the Cech complex of Ωn are formed by all linear combinations
of forms that extend to at least one of the Ui ’s. It is easy to see that the residue at x = 0 of any form that
represents a coboundary, is zero. Hence we have a well-defined residue map:
Resx=0 : H n−1 (U ∗ , Ωn ) → C.
Alternatively, in the holomorphic setup we could notice that one has
H n−1 (U ∗ , Ω) ' H∂n,n−1
¯ (U ∗ ),
since on (n, q)-forms ∂¯ is equal to d. One has a natural map
H∂n,n−1
¯ (U ∗ ) → H 2n−1 (U ∗ ) = C
that coincides with the residue map.
2.7.2 Cohomology with support.
i
Consider a closed subvariety Y of an algebraic variety X. We denote by Y ,→ X the closed embedding
j
of Y in X and by U = X \ Y ,→ X the open embedding of the complement U of Y .
For any sheaf of abelian groups A on X consider the subsheaf Γ[Y ] A of A formed by all section of A
supproted on Y . One has an exact sequence
0 → Γ[Y ] A → A → j∗ j ∗ A.
i
Moreover, Γ[Y ] is exact on injective (or flabby) sheaves. Hence there exist derived functors H[Y ] A which
i
allow us to continue the exact sequence above. Namely, since R (id) = 0 for i > 0, we have an exact sequence
0 ∗ 1
0 → H[Y ] A → A → j∗ j A → H[Y ] A → 0

and isomorphisms
0 → Ri j∗ j ∗ A → H[Y
i+1
] A → 0, i ≥ 1. (2.7.3)

2.7.4 Remark The functor j∗ may not be exact. For example, if one removes a point from C2 the
resulting variety is no longer affine, so the functor R1 j∗ will in fact be non-trivial.
2.7.5 Example In the algebraic situation described above one has a vector space isomorphism
h1 1 i dx1 dxn
n
H[0] (U, Ωnalg ) ' C ,... , ∧ ... ∧ .
x1 xn x1 xn
h i
In fact, Ωn (U1 ∩ . . . ∩ Un ) = C x1 , x11 , . . . , xn , x1n dx1 ∧ . . . ∧ dxn . However, any form that does not have
a pole along all the divisors {x1 = 0}, . . . , {xn = 0}, represents a Cech coboundary. The quotient of
Ωn (U1 ∩ . . . ∩ Un ) by the linear subspace generated by all such for is exactly the RHS of the expression above.

Denote the defining ideal of Y by IY ⊂ OX . First, notice that

HomOX (OX /IY , M) ⊂ Γ[Y ] M,
(since any section killed by IY is supported on Y .) To go backwards, we notice that if a section of M is
supported on Y then it is killed by some power of IY . In other words,
Γ[Y ] M = lim
−→
HomOX (OX /IYn , M). (2.7.6)
n

By uniqueness of derived functors, we deduce from (2.7.6) that

i
H[Y ] M = lim
−→
ExtiOX (OX /IYn , M). (2.7.7)
n

36
 2.7.8 Special case: X is affine and Y is a smooth divisor in X given by the equation {f = 0}. In this
case U = X \ Y is affine hence j∗ is exact and we just have an exact sequence
0 → ΓY [M ] ,→ M → M [f −1 ] → H[Y
1
] M → 0.
−1
We see now that ΓY M = {m ∈ M | f i(m) · m = 0 for some i(m)} and H[Y
1
] M = M [f ]/Image(M ).
2.7.9 Example If X = C1 , Y = {0} and M = C[t] = OX . Then C[t, t−1 ]/C[t] = D · δ in the notations
of (???).
2.7.10 Local duality.
Now we explain the relationship between differential operators and local cohomology. Let us assume
first that Y ,→ X is a point given locally by vanishing of n regular functions. Namely, we choose an affine
open neighbourhood U of x := Y and functions fi ∈ Γ(U, OX ), i = 1, . . . , fn such that Ix is generated by
f1 , . . . , fn .
Then we can construct a pairing
n
RES : Ox × H[x] (ΩX ) → C (2.7.11)
n ∗
as follows. Let f ∈ Ox and ω ∈ H[x] Consider the open covering of U = U \ {x} by affine open subsets
(ΩX ).
Ui = U \ {fi = 0}. Using (2.7.3) we can represent ω by an algebraic differential n-form on U1 ∩ . . . ∩ Un .
Put RES(f, ω) = Res x (f · ω).
h i
2.7.12 Example Let X = C and x = {0}. Then H[x] 1
(Ω1 ) = dx 1
x x and

dx h 1 i
RES : Ox × →C
x x
is given by the usual one-dimensional residue of f · ω (i.e. the coefficient of dx
x in the Laurent expansion of
f · ω).

Assume for simplicity that we are in the situation of the example above, i.e. X = C and x = {0}. Take
1 dx
ω= xk x
. Then integrating by parts we obtain
dk f (x) dx
Z Z
f (x) dx
k x
= ± .
S1 x S1 dxk x
Similarly, when X = Cn and x = {0} we can conclude that
(a) RES( · , dx 1
∧ . . . ∧ dxn
) acts on Ox as delta function supported at x, i.e. f ∈ Ox maps to f (0).
x1 xn
∂ k1 ∂ kn
(b) RES( · , k1 . . . xkn x1 ∧ . . . ∧ dx
1 1 dx1
xn ) maps f (x) to ( k1 . . . ∂xkn f )(0).
n
x1 n ∂x1 n

2.7.13 Cohomological residues. Using the direct limit presentation (2.7.7) of local cohomology we can
view the residue map (2.7.11) as follows. By Serre duality, for any m = 1, 2, . . . there exists a natural pairing
OX /Ixm × ExtnX (OX /Ixm , ΩnX ) → C.
Notice that since X is smooth we have
ExtnX (OX /Ixm , ΩnX ) = Γ(X, Extn (OX /Ixm , ΩnX )).
For any ω ∈ ExtnX (OX /Ixm , ΩnX ) ⊂ H[x]
n
(ΩX ) the image of the residue map RES( · , ω) is equal to the
m
composition of Ox → OX /Ix with the Serre duality map.
n
Proposition 2.7.14 The pairing (2.7.11) is (topologically) perfect, i.e. it identifies H[x] (ΩnX ) with a
subspace of Hom C (Ox , C) formed by all linear functions that depend only on some finite jet of an element
in Ox . Moreover, under this identification, the subspace ExtnX (OX /Ixm , ΩnX ) ⊂ H[x]
n
(ΩnX ) corresponds to
Hom C (OX /Ix , C) ⊂ Hom C (Ox , C). 2
m

37
 2.7.15 The general case. Now suppose that Y ,→ X is a smooth subvariety of codimension d. By
passing to an open subset we can assume that Y is given by vanishing of d functions t1 , . . . , td .
Lemma 2.7.16 If Y , X and t1 , . . . , td are as above, then

i 0 if i 6= d
H[Y =
] C[∂/∂t1 , . . . , ∂/∂td ] ⊗ C OY if i = d. 2

To construct the analogue of (2.7.11), notice that

OY ' HomOX (OY , OX /IY ). Hence one has a Yoneda pairing
OY × Extd (OX /IY , ΩX ) → Extd (OY , ΩX ) ' ΩY
(I don’t know what to to with higher powers of IY !!)
Proposition 2.7.17 There exists a (topologically) perfect pairing of
sheaves on X
d
RES : OX,Y × H[Y ] (ΩX ) → ΩY (2.7.18)
where OX,Y denotes the completion of OX along Y .
2.7.19 Higher-dimensional Sato construction.
Denote by ∆ the diagonal in X × X and by i∆ its embedding in X × X. We want to represent DX as
some sheaf on X × X supported at ∆. Let dim X = n.
Theorem 2.7.19.1. There exists an isomorphism
n
(i∆ )∗ (DX ) = H[∆] (ΩX  OX ) (2.7.20)
of sheaves on X × X. 2
n
Sketch of proof. Suppose one has an section ω of H[∆] (ΩX  OX ) and a function f (x) on X. Then
ω · f is a function on X which can be described as follows. Given a point x2 ∈ X, the restriction of ω to
pr2−1 (x2 ) ' X2 is an element of H[x
n
2]
(ΩX  OX ). We put the value of the function ω · f at x2 to be equal
to RES(f, ω). 2

We can describe the isomorphism of the theorem above from the point of view of local duality. For any
function f on X the pullback pr1∗ (f ) is a section of OX×X which we can project to a section of OX×X,∆ .
The proposition (??) above provides a pairing
n
RES : OX×X,∆ × H[∆] (ΩX×X ) → Ω∆ .
n n
However since ω is a section of H[∆] (ΩX OX ), not of H[∆] (ΩX×X ), the result of the pairing will be a section
of O∆ , which is a function on ∆ ' X (i.e. we “untwist” ΩX on the second factor of X × X).
2.7.21 Local cohomology modules.
i
Let M be a DX -module and let Y ,→ X is a submanifold.
i
Claim 2.7.22 For any i ≥ 0, H[Y ] M has a natural structure of a DX -module.
Proof. Since this claim is very important we will give two equivalent constructions of the DX -module
structure.
i+1 i
2.7.23 First construction. We use the isomorphisms H[Y ] (X, A) ' H (X \ Y, A). Choose a Cech
•
covering for X \Y . All the groups C (Ui1 ∩. . .∩Uin , M) have a natural DX -module structure which descends
to the cohomology due to combinatorial nature of the Cech differential.
38
 2.7.24 Second construction. For the second construcion of the DX -module structure we extend
isomorphism (2.7.6):
Γ[Y ] M ' lim
−→
HomOX (OX /IYn , M) ' lim
−→
HomDX (DX /(DX · IYn ), M)
n n

(the second isomorphism follows from DX /(DX ·IYn ) ' DX ⊗OX OX /IYn . Notice also that the first term in the
line above makes sense in the general sheaf-theoretic setup, the second term in the category of OX -modules
while the third only in the category of DX -modules.
n ·u
Take u ∈ DX . The right multiplication map DX −→ DX does not descend to an endomorphism of
DX /(DX ·IY ) since (·u) does not preserve IY . However, using the Leibniz rule we can show that (DX ·IYk )·u ⊂
k k
·u
DX · IYk−n hence there is an induced map DX /(DX · IYk ) −→ DX /(DX · IYk−n ). Therefore the right action of
u on lim
←−
DX /(DX · IYk ) is well-defined.
k

Hence lim
←−
DX /(DX · IYk ) has a DX -bimodule structure. Since
k

Γ[Y ] M ' lim

−→
HomDX (DX /(DX · IYn ), M) = HomDX (lim
←−
DX /(DX · IYn ), M),
n n

Γ[Y ] M inherits a left DX -module structure.

2.7.25 Warning. There is no DX -module structure on any of the individual terms HomOX (OX /IYn , M),
only on the direct limit above.
To extend this construction of higher cohomology with support (i.e. to
lim Exti (. . . )) we can do two a priori different things:
−→
consider RHomDX (DX /(DX · IYn ), M) or consider RHomOX (OX /IYn , M).
However these two functors coincide since we can compute the latter using locally free OX -resolutions
of OX /IYn . Given such resolution, we can tensor it with DX over OX and obtain a DX -resolution computing
the former functor.

39
 3. D-modules: basic constructions.

3.1. D-modules on a line.

Here we will consider the case of D-modules on a one-dimensional affine space over the field of complex
numbers C (any other algebraically closed field of characteristic zero will do). In other words, we consider
dn
P
the ring of all differential operators of the form ai (x) dx n where ai (x) ∈ C[x] are polynomials in x.

d
3.1.1 Notation Everywhere below we will write ∂ instead of dx .
The ring D of differential operators on the line can be also represented as a quotient of the ring Chx, ∂i
of polynomials in non-commuting variables x and ∂, by the relation x∂ − ∂x = 1.
We define an anti-involution (·)t : u 7→ ut by
xt = x, ∂ t = −∂
(then, for example, (x∂)t = −∂x). The meaning of this involution becomes clear if we think of differential
operators on C as acting on the space of smooth functions on C = R2 (say, the Schwartz space S(R2 )). Then
for any differential operator P and any pair of functions f, g, one has
Z Z
(P f )g = ± f (P t g).
R2 R2
Let us give several examples of D-modules.
3.1.2 Examples.
(i) Consider O := C[x]. Then O is a simple D-module generated by 1 ∈ O and
O = D · 1 = D/D∂
(ii) We can act by a differential operator on the ring C[x, x−1 ] of Laurent polynomials in x. This ring
is not finitely generated as a O-module (and in general, many D modules that we will consider will not be
coherent). However, over D, C[x, x−1 ] is generated by the element x1 which is annihilated by (∂x):
1
C[x, x−1 ] = D ·= D/D(∂x)
x
(iii) For any λ ∈ C \ Z introduce the formal symbol xλ and put ∂ · xλ = λxλ−1 . Generalizing the previous
example, we can write
C[x, x−1 ]xλ = D · xλ = D/D(x∂ − λ).
(iv) Note that O is a submodule of C[x, x−1 ]. Consider the short exact sequence
C[x, x−1 ]
0 → O → C[x, x−1 ] → → 0,
C[x]
C[x,x−1 ]
and denote the image of x−1 in C[x] by δ. The reader can easily prove that
C[x, x−1 ]
= D · δ = D/Dx
C[x]

3.1.3 Remark What are the characteristic varieties of the modules above? If we notice that gr(D) is
C[x, ξ] (where ξ is the symbol of ∂) then Spec(gr(D)) = C2 and
(i) SS(O) = {ξ = 0}.
(ii) SS(C[x, x−1 ]) = {xξ = 0}.
(iii) SS(C[x, x−1 ]xλ ) = {xξ = 0}. Notice that in this case the characteristic variety is not irreducible
even though C[x, x−1 ]xλ is a simple D-module.
(iv) SS(D · δ) = {x = 0}.
40
 −1
]
Note that in the fourth example the element δ ∈ C[x,x
C[x] satisfies x · δ = 0: a property that is true for
R
the delta function in calculus. What will happen when we try to find “ δ”? In calculus this integral is given
by the Heaviside function θ(x). To see what happens for D-modules we consider the following
3.1.4 Example
Introduce a formal symbol log(x) with the property ∂·log(x) = x1 and consider the D-module C[x]log(x)+
C[x, x−1 ]. This module is generated over D by log(x) and one has
C[x]log(x) + C[x, x−1 ] = D · log(x) = D/D(∂x∂).

3.1.5 Remark In general, let P = an (x)∂ n + an−1 (x)∂ n−1 + . . . + a0 (x) ∈ D. The idea due to Sato is
that one should study the properties of M = D/D · P instead of looking for solutions of P · a(x) = 0.
3.1.6 Notation When a module M has a filtration M1 ⊂ M2 ⊂ M3 = M we write
 
M3 /M2
M ;  M2 /M1 
M1
and
 say  that M3 /M2 is on the top of the filtration while M1 is on the bottom. For example, C[x, x−1 ] ;
D·δ
.
O
 
O
Lemma 3.1.7 D · log(x) ;  D · δ 
O
Proof. One has an exact sequence
0 → C[x, x−1 ] → D · log(x) → C[x] → 0. 2
R
Now mod out the bottom of D · log(x) and denote by δ the image of log(x)in the quotient
 module
O
(this image generates the quotient). It follows from the definition that D · δ ;
R
. Denote by θ
D·δ
−1
R
the generator of the quotient O of D · δ. Thus, θ isR “log(x) modulo C[x, x ]”. Geometrically θ gives the
monodromy around zero. If we compare that with δ in calculus, we will see that the situation is similar
to the relationship between real Morse theory (which tells how the topology changes when we go through a
critical level) and complex Morse theory (which describes Picard-Lefschetz transformation, i.e., the result of
going around a critical level).
3.1.8 Exercise The Verdier dual of a module was defined in (??). By its definition, dual of a left
D-module is a right D-module. However, the anti-involution u 7→ (u)t allows us to view any right D-module
as a left D-module, so we will assume that the dual is a left D-module.
(i) Prove that O is a self-dual module
(ii) Prove that (C[x, x−1 ])∨ = D · δ and (D · δ)∨ = C[x, x−1 ] (i.e. taking the dual flips the order of
R R

subquotients).
(iii) Prove that (D · xλ )∨ = D · x−(λ+1) .

3.1.9 Remark Finally, we give an example which illustrates that in general one should be careful when
considering the charactersitic varieties. Let P1 = x∂ and P2 = ∂x. Consider the ideal J = {D · P1 + D · P2 }.
Then J = D, so SS D/J = ∅. However, if we just look at the zero-set of the principal symbols of P1 and
P2 , we get {xξ = 0}. The point is that the statement {Pi generate J} does not imply in general that {σ(Pi )
generate gr J}.
41
 3.2. Direct image from a submanifold.
When M = OX we can give a third construction of the DX -module structure on the local cohomology.
To that end, introduce the following
Y
Definition 3.2.1 DX = {u ∈ DX |u(IYi ) ⊂ IYi , ∀i}.
Y
Proposition 3.2.2 DX has the following properties:
Y
(i) DX is a subalgebra of DX .
Y
(ii) OX ⊂ DX .
Y
(iii) Denote by TX ⊂ TX the subsheaf of all vector fields on X that are tangent to Y at any point of Y .
Y Y
Then TX ⊂ DX .
Y Y
(iv) DX is generated by OX and TX
Proof. First notice that (i) and (ii) are clear, and we can prove (iii) and (iv) locally.
Since Y is smooth, it is locally a complete intersection in X. Therefore, for any point y ∈ Y ⊂ X, we
can choose an affine neighborhood U of y in X and regular functions y1 , . . . , yk , t1 , . . . , tl such that
(i) Y ∩ U is given by vanishing of t1 , . . . tl ,
(ii) y1 , . . . , yk form a system of “local coordinates” on Y (in the sense explained in ??).
Then one can check that
(1) DX = OX [∂/∂ti , ∂/∂yl ]
Y
(2) TX is generated by ∂/∂yi , tj ∂/∂tk .
Y
(3) DX = OX [ti , tj ∂/∂tk , ∂/∂yl ].
Now the assertions of the proposition follow from (1)-(3). 2
Y Y
By definition of DX , there is a well-defined action of DX on OX /IY = OY . Moreover, one has the
following
Claim 3.2.3 There exists a diagram

Y Y Y
DX ←- DX  DY ' D X /(IY · DX ) ∩ DX (3.2.4)
Y Y
Proof. The intersection (IY · DX ) ∩ DX is in the kernel of the map: DX → DY since it sends OX to
Y
IY . By choosing locally a suitable coordinate system we can show that DX → DY is indeed surjective and
Y
(IY · DX ) ∩ DX coincides with the whole kernel of it. 2
Considering DX · IY instead of IY · DX leads to a different situation:
Y Y
Proposition 3.2.5 The quotient DX /(DX · IY ) ∩ DX is isomorphic to the ring of twisted differential
operators on the top exterior power det(TY X) of the normal bundle to Y.
Y
Proof. First we observe that by Proposition 3.2, TX acts on IY /IY2 , the sheaf of sections of the conormal
∗
bundle TY X. The sections of the normal bundle TY X are given by HomOY (IY /IY2 , OY ). If v and λ are
sections of TY X and TY∗ X respectively, and a vector field ξ ∈ TXY
acts on v by the formula
< ξ · v, λ >= ξ· < v, λ > − < v, ξ · λ > .
Y
Thus, ξ also acts on det(TY X) and it is easy to check that [ξ, f ] = ξ(f ) for f ∈ OX . Since DX is generated
Y Y
by TX and OX , it follows that DX acts on det(TY X). It is routine to check that IY acts trivially, so that
Y Y
DX · IY ∩ DX is in the kernel of this action. Checking that DX gives the full algebra of differential operators
on det(TY X) is easy to do locally. 2
We can now define direct image for DY modules for the closed embedding f : Y → X by
f∗ M = f· (f · DX ⊗f · DX
Y (M ⊗ det(TY X)), (3.2.6)

42
where f· and f · are sheaf theoretic direct and inverse images. Here DXY
acts on M through DY , and vector
Y
fields in DX act on the tensor product by derivations. On affine open subsets U , then the global sections of
f∗ M (U ) are given by
DX (U ) ⊗DX
Y (U ) (M ⊗ det(TY X)(U )).

When we are verifying local properties of f∗ , we will often omit the sheaf functors f· and f · from the notation.
3.2.7 Remark The twist by det(TY X) in the definition of direct image can be explained if we think in
terms of distributions. We would like to think of f∗ (OY ) as the sheaf of distributions supported on Y. In the
case where Y = {0} ⊂ C = X, the affine line, this means f∗ (OY ) should be spanned by the δ function at the
origin and its derivatives. It is routine to check that the Euler vector field x∂ · δ = −δ, so that x∂ + 1 = ∂x
Y
should act as zero. This is the case for the action of DX on det(TY X) but is not the case for the action of
Y
DX on OY . Ignoring left vs. right difficulties that will be explained later, the factor det(TY X) can also be
explained as follows. Usually, we think of D-modules as being like functions, so that it is natural to pull
them back. If we want to push forward a DY -module M to get a DX module, we should think of M as a
distribution by tensoring it with ΩY . Then it can be pushed forward naturally, and then we should tensor
the push-forward with ΩX −1 in order to think of the push-forward as functions again. This produces the
factor det(TY X).
3.2.8 Remark This construction of direct image is reminiscent of parabolic induction in representation
theory. Let g be a reductive Lie algebra and p be its parabolic subalgebra. Then p can be written as a direct
sum l + n where n is the nil-radical of p (which is defined canonically) while l is intrinsically defined only as
a quotient p/n (and the direct sum above is a non-canonical splitting). The diagram of maps of Lie algebras
g ←- p  l
gives rise to a diagram of maps of associative algebras
Ug ←- Up  Ul.
hence for any Ul-module N we can consider a Ug-module Ug ⊗U p N .
However, it is often more natural first to shift the action of p. For x ∈ p, let ρ(x) = 21 T r(x|g/p ). ρ
defines a one dimensional representation Cρ of p. Then if N is a Ul-module, its twisted induced module is
Ug ⊗U p (N ⊗ Cρ ). This twisted induction is much more natural in representation theory. The twist by ρ in
representation theory plays the same role as the twist by det(TY X) in D-modules. 2
Any morphism f : Y → X of algebraic varieties by definition induces a map of sheaves f · OX →
OY . However, there is no covariant (or contravariant) map between differential operators, since they are
constructed from functions (contravariant objects) and vector fields (covariant objects).
To illustrate the way in which this difficulty is resolved replace all the D-modules with the corresponding
graded objects which are modules over OT ∗ X or OT ∗ Y . A map f : Y → X does not induce a map
between the tangent spaces. So, instead of taking the graph of a map T ∗ Y → T ∗ X we consider the graph
∗ ∗ ∗
Graph(f ) ⊂ X × Y , and take its conormal bundle TGraph(f ) (Y × X) ⊂ T Y × T X which is the only natural
object that arises in this situation. This conormal bundle is only a correspondence (i.e. a multivalued map)
between T ∗ Y and T ∗ X. The complexity of D-module behaviour is depicted by deviation of TGraph(f
∗
) (Y ×X)
from being a graph of a map.
To see how a correspondence can define a functor on some geometric objects, we think of sections of
a D-module as “functions” or “distributions”. For any such “function” F (y) on Y we can define its direct
image under f by a formula used in physics:
Z
(f∗ F )(x) = δ(x − f (y))F (y)dy.
Y
When we try to adapt this formula to our situation we view δ(x − f (y)) as a D-module supported on
Graph(f ) ⊂ Y × X.
43
 f∗ (M) = f· (BGraph(f )|Y ×X ⊗prY∗ DY M)
where f· stands for sheaf-theoretic direct image which mimics the integral in the formula above. Here, if Z
is a closed smooth subvariety of codimension d in a smooth variety W,
d
BZ|W := H[Z] (OW ).

Proposition 3.2.9 If Y is a smooth subvariety of X of codimension d, then


i 0 if i 6= d
H[Y ] (OX ) =
f∗ (OY ) if i = d.

Remarks.
(1) The vanishing part of the proposition can be proved by a local computation.
(2) We will give the proof in 3.3.
3.2.10 Notation Denote by DY →X the quotient DX /IY · DX (this is a particular case of the object to
be introduced later).
With this notation, we can write the following equalities (the first one uses that IY · ΩY = 0):
ΩY ⊗DX
Y DX ) ⊗O
X
Ω−1
X = ΩY ⊗DX
Y DX /IY · DX ) ⊗O
X
Ω−1
X = (3.2.11)

ΩY ⊗DX
Y DY →X ) ⊗O
X
Ω−1 −1
X = ΩY ⊗DY DY →X ) ⊗OX ΩX .
Y
The point is that, while DX is not locally free over DX , we have the following
Claim 3.2.12 DY →X is locally free over DY
Sketch of proof. Locally we can choose d regular functions t1 , . . . , td on X so that Y is given by vanishing of
these functions. We can also choose (again locally) (n − d) regular functions y1 , . . . , yn−d that have linearly
independent differentials on Y (i.e. form an “etale local coordinate system” on Y ). Then we can write a
non-canonical splitting (depending on the choices above):
DX = IY · DX ⊕ h∂t , y, ∂y i.
which implies that DY →X = DX /IY · DX is a free module over DY with generators ∂t . 2

3.2.13 Other definitions of f∗ for closed embeddings.

(1) For any right DY module N one can form a right DX -module N ⊗DX Y DX .

(2) For any left DY -module M one can consider the left DX -module DX ⊗DX
Y M.

We will see later that the latter of the two functors defined above is “wrong” (it does not agree with
other operations on D-modules to be defined later). So we have to redefine the functor on left DY -modules
by sending
 
⊗OX Ω−1


M 7→ M ⊗OY ΩY ⊗DX Y DX
X (3.2.14)

(i.e. we use tensoring with top degree differential forms on Y to make right modules from left modules, and
then we tensor with Ω−1X to go back from right modules to left modules.) This last expression coincides with
 
M 7→ M ⊗OY ΩY ⊗DY DY →X ⊗OX Ω−1


X (3.2.15)

using (3.2.11) above.

3.2.16 Exercise Verify that f∗ (M) coincides with the expression defined in (3.2.15).
44
 d
The exercise allows us to construct a good filtration on f∗ (OY ) = H[Y ] OX . It follows from the proof of
claim 3.2 and the above exercise that locally f∗ (OY ) ∼
= C[∂/∂tk ]ω, polynomials with constant coefficients in
d
the normal derivatives times a section of det(TY X). We filter by requiring locally that ∂/∂tk ∈ (H[Y ] OX )1 .
d ∗ ∗
One deduces immediately that SS(H[Y ] OX ) = TY X ⊂ T X.
Corollary 3.2.17 f∗ is exact.
Proof. This follows from Claim 3.2 and Exercise 3.2. 2

3.3. Restriction to a submanifold; Kashiwara’s theorem.

3.3.1 Pull-back of D-modules (definition of f + ).
Let f : Y → X be a morphism. In our definition of a pullback, we want the following property to be
satisfied:
Let φ be a function on X and Ann φ = {u ∈ DX | uφ = 0} be the sheaf of all differential
operators that annihilate φ. Then the pullback of a DX -module DX /Ann φ is isomorphic
to DY /Ann f ∗ (φ).
This means that we pullback a DX module by pulling back its solutions.
Recall that we have the sheaf-theoretic pullback functor f · and the O-module pullback functor f ∗ defined
by
f ∗ (M) = OY ⊗f · OX f · M.

Claim 3.3.2 For any left DX -module M, the pullback f ∗ (M) in the sense of O-modules, has a natural
DY -module structure.
Proof. Recall that we have a natural morphism df : TY → f ∗ TX = OY ⊗f · OX TX (the image of a tangent
vector at a particular point of Y is a tangent vector at its image, but when we take a vector field of Y its
image is well-defined only as a section of f ∗ TX since a point in X can have many points in Y mapping to
it).
Now define a DY -module structure on OY ⊗f · OX P f · M as follows. The sheaf of functions OY acts as
usual, while for a vector field ξ on Y we write df (ξ) = gi ⊗ ηi and define
X
ξ(g ⊗ m) = (ξg) ⊗ m + g · ( gi ⊗ ηi m).
One can write the same formula in a slightly different form by choosing local coordinates {xi } on X:
X
ξ(g ⊗ m) = (ξg) ⊗ m + g(ξf ∗ (xi )) ⊗ ∂i m.
i

One checks that either of this two formulas above defines a DY -module structure. 2
Definition 3.3.3 We denote by f + the restriction of the pullback functor f ∗ from the category of
O-modules to the category of D-modules.
3.3.4 Remark One can give another definition of f + : consider the sheaf f ∗ DX on Y . By the claim
above it has a left DY -module structure and it also has a right f · DX -module structure. Then the equality
OY ⊗f · OX f · M ' f ∗ DX ⊗f · DX f · M
allows us to define f + (M) as f ∗ DX ⊗f · DX f · M.
3.3.5 Notation The sheaf f ∗ DX above will be denoted by DY →X .
3.3.6 Another pullback for closed embeddings (definition of f ! ).
Let f : Y ,→ X be a closed embedding. We define a functor
f ! : mod − DX → mod − DY
45
by
f ! (M) = det(TY∗ X) ⊗OY IY M
for a (left) DX -module M. To see that f ! (M) is a DY -module, note that IY M has a natural action by
DXY Y
/(DX · IY ) ∩ DX , the differential operators on det(TY X) by Proposition 3.2. After we twist by det(TY∗ X),
we have action by DY .
3.3.7 Properties of f + and f ! for closed embeddings.
Proposition 3.3.8 Let f : Y ,→ X be a closed embedding of codimension d. Then
(i) f ! is left exact (hence we can speak of right derived functors Ri f ! ),
(ii) f + is right exact (and we can consider the left derived functors Li f + ), and
(iii) For any left DX -module M there exist natural isomorphisms
Rd f ! (M) ' f + (M) and L−d f + (M) ' f ! (M)
(iv) The functor f∗ is left adjoint to f ! :
HomDX (f∗ N , M) = HomDY (N , f ! M)

Proof. The first asserion follows from the definition of f ! and left exactness of HomOX (OY , ·). The second
assertion is a consequence of right exactness of (OY ⊗f · OX · ). The third claim follows from the equalities
Ri f ! (M) ' det(TY∗ X) ⊗OY ExtiOX (OY , M), L−i f ∗ (M) ' T oriOX (OY , M)
(which can be proved using uniqueness of derived functors) and the isomorphism of sheaves
OX
Extd−i


OX OY , M ' T ori (OY , M) ⊗OY det(TY X).
The latter isomorphism can be proved in local coordinates by considering the Koszul complex and using its
self-duality property.
By easy sheaf arguments, it suffices to prove (iv) on an affine open cover and to check that the iden-
tifications agree on intersections. On affine open sets, the adjunction property (iv) follows by Frobenius
reciprocity. We have
HomDX (DX ⊗DX
Y (M ⊗ det(TY X), N )) = HomD Y (M ⊗ det(TY X), N )
X

= HomDdet(TY X) (M ⊗ det(TY X), IY

N ) = HomDY (M, det(TY∗ X) ⊗ IY
N)
where the second equality follows since IY annihilates M ⊗ det(TY X). Identifications agree on intersections
of affine open sets since they are canonical. 2
3.3.9 Remark The proposition above essentially says that for a closed embedding the functor f + is
“unnecessary” since it can be expressed via f ! .
Proposition 3.3.10
(i) If g : Z → Y and f : Y → X are arbitrary morphisms then for any DX -module M (f g)+ (M) '
g + (f + (M)).
(ii) If g and f are closed embeddings then (f g)! (M) = g ! (f ! (M)) and (f g)∗ (N ) ' f∗ (g∗ (N )).
Proof. The first claim follows directly fron the definition of f + and the corresponding statement from
algebraic geometry. Recall that one can also write f + (M) = DY →X ⊗f · DX M. Hence (i) applied to
M = DX gives
DZ→X ' DZ→Y ⊗g· DY g · DY →X . (3.3.11)

3.3.12 Exercise Prove (3.3.11) form the point of view of Sato construction.
The first claim of (ii) is routine to check, and the second claim follows from uniqueness of the adjoint
functor from the first.
46
 3.3.13 Kashiwara’s Theorem.
Theorem 3.3.13.1. Let f : Y ,→ X be a closed embedding and M be a DY -module. Then
(i) f ! f∗ (M) ' M (i.e. f ! is left inverse of f∗ ).
(ii) If a DX -module N is supported on Y (in the sense on O-modules) then N ' f∗ f ! N .
(iii) In this way one has an equivalence of categories:



f∗ : DY -modules → DX -modules supported on Y .

3.3.14 Remark
The analogue of Kashiwara’s theorem for the category of O-modules fails for the obvious reason: OX -
modules supported on Y have a non-trivial invariant: multiplicity along Y . However, for D-modules, differ-
entiating in directions transversal to Y destroys this invariant. For example, on the affine line, C[t]/(tn ) is
not a D-module.

Proof of Kashiwara’s Theorem. We will proceed in several steps.

Step 1. There exists a canonical adjunction map
M → f ! f∗ (M) = det(TY∗ X) ⊗ IY
(DX ⊗DX
Y (M ⊗ det(TY X)))

that in local coordinates maps m to φ ⊗ 1 ⊗ m ⊗ φ , where φ is a local nonvanishing section of det(TY∗ X)

−1

and φ−1 is the inverse section. Using local coordinates as in the proof of Claim 3.2 one can see that this
map is an isomorphism. This proves (i).
Step 2. There exists a canonical adjunction map f∗ f ! (N ) → N . To prove that it is an isomorphism we
can represent Y locally as intersection of hypersurfaces and by functoriality of f∗ and f ! (cf. Proposition
3.3) we may assume that codim(Y, X) = 1. Hence locally Y is given by one equation Y = {t = 0}. Put
∂ := ∂/∂t and θ := t · ∂. For N as in (ii), let
N (i) := {n ∈ N |θ · m = i m}.
Since ∂ · t − t · ∂ = 1, we have maps:
t : N (i) → N (i + 1), , ∂ : N (i + 1) → N (i).
For i < −1, ∂ · t and t · ∂ are invertible operators on N (i), N (i + 1), hence the maps t, ∂ above are
isomorphisms.
Step 3. Consider N I = {n ∈ N | tn = 0}. We want to show that locally N = C[∂] ⊗C N I . In fact, for
any n ∈ N I , θ · n = −n. Therefore NI ⊂ N (−1) and ∂ i · N I ⊂ N (−i − 1). Now, by the Nullstellensatz, since
N is supported on Y , any element n of N is annihilated by some power of t, say tN . Hence tN −1 n ∈ N I
and n ∈ ∂ N −1 · N I . In particular N I = N (−1). The rest of the proof follows easily. 2
3.3.15 D-modules on singular varieties.
We assumed everywhere that Y is a smooth subvariety of X. For a singular variety Y , the category
of DY -modules can (and should) be defined by choosing a closed embedding Y ,→ X and considering the
subcategory of all DX -modules supported on Y . By Kashiwara’s theorem, the definition of DY -modules is
independent of X. Using the equivalence between left and right D modules, we can define right DY -modules
in the same way.
3.3.16 D-crystals. Let µ : X → Y is a finite morphism of schemes. Then there is the Grothendieck
functor µ! : OY − mod → OX − mod satisfying the adjunction formula
HomOY (M, µ! N ) = HomOX (µ· M, N ).
When X = SpecA and Y = SpecB, by letting M = B, we see that µ! (N ) = HomB−mod (A, N ). In other
words, if A = B/J then µ! (N ) is the maximal submodule of N annihilated by J. It follows immediately
from the adjunction property that for any pair of morphisms π, µ as above one has (µ ◦ π)! = π ! ◦ µ! .
47
 Let Y be an affine variety (possibly singular). We say that a morphism Y ,→ X is a nilpotent thickening
of Y , if Y is a closed subscheme of X defined by a nilpotent ideal. Morphisms between two thickenings are
defined in an obvious way.
We will consider in particular nilpotent thickenings of Y given by the diagonal embeddings Y → Y m , y 7→
(y, y, . . . , y). Let I be the ideal defining the diagonal and let Ynm be the nth infinitesimal neighborhood of the
diagonal, the scheme defined by I n+1 , i.e., Ynm = Spec(OY m )/I n+1 . We have the projections to the ith factor,
pi,n : Ynm → Y, and the obvious closed embeddings in : Y ,→ Ynm . Consider also the formal completion Y <m>
corresponding to the diagonal embedding and the category of discrete O-modules on Y <m> , i.e., modules
whose sections are supported on some Ynm . Note that from the description p!i,n F = HomOY (OYnm , F), we
have inclusions p!i,n F ⊂ p!i,n+1 F. There are projections onto the ith factor pi : Y <m> → Y. We define the
discrete module p!i F as the union of all p!i,n F. The functors p!i satisfy similar adjunction properties.
In the remainder of this section, we will focus on the completions Y <2> and Y <3> , and denote the
projection on the ith factor Y <3> → Y by qi . We will denote the projections to the ith and jth factor
Y <3> → Y <2> by pij . Functors p!ij are defined as above, and are right adjoint to the corresponding direct
image functors.
Proposition 3.3.17 [Grothendieck] Let F be a OY module and assume that Y is smooth.
∼ F ⊗O D Y
p! F =
1 Y

where the tensor product is defined using the left OY module structure on DY . Moreover,
q1! F ∼
= F ⊗OY DY ⊗OY DY .

Proof. First let F = OY . A differential operator ∂ of order n defines a section of p!1,n F = HomOY (OY n , OY )
by the formula, φ∂ : f ⊗ g 7→ f ∂(g). One checks by induction that if ∂ is a nth order operator, then
φ∂ annihilates the ideal defining Yn , so φ∂ is well-defined. The inverse map is given by φ → ∂φ , where
∂φ (h) = φ(1 ⊗ h). Hence, DY ∼ = p!1 OY . The general case follows from the isomorphism p!1 F ∼ = F ⊗ p!1 OY ,
which can be proved by using the freeness of OY <2> over OY . The second claim follows as above by defining
a map φ : DY ⊗OY DY → q1! (OY ) by φ(∂1 ⊗ ∂2 )(f1 ⊗ f2 ⊗ f3 ) = f1 ∂1 (f2 ∂2 (f3 )). 2
Definition 3.3.18 A D-crystal on Y is a collection of OX -modules FX on all nilpotent thickenings
Y ,→ X, such that for any morphism
>X
}}}
}
}}
}}
Y A µ
AA
AA
AA
A 
X0
an isomorphism µ! (FX 0 ) ' FX compatible with compositions is given. The category of all D-crystals on Y
is denoted by Mcrys (Y ).
We recall the notion of a formally smooth algebra due to Grothendieck ([EGA IV] or [Mt]). A noether-
ian topological ring A is said to be formally smooth if for any discrete ring R with a nilpotent ideal I, any
homomorphism A → R/I lifts to a homomorphism A → R. It suffices to requires the above lifting property
for any ideal I with I 2 = 0. If A is the ring of functions on a smooth variety (with the discrete topology on
A), then A is formally smooth (see [Ha1], Exer. II.8.6, pp. 188-189). Moreover, if A is formally smooth and
I is an ideal of A, then the completion  of A along I (with I-adic topology) is formally smooth.
We give an equivalent formulation of the notion of a D-crystal. We choose a closed embedding of Y
into some smooth variety Z. By taking the formal completion V of Z along Y, we may find an embedding
48
k : Y → V , a formally smooth variety. As for Y <2> , a D-crystal determines a discrete O-module on V <2> .
Note that when Y = Z, we may take V = Z.

V< <2>
k×k y yy
y
yy
yy
Y FF pi
FF k
FF
FF
F" 
V
we obtain an OV <2> module isomorphism FV <2> ∼
= p!i FV . By composing two of these we obtain an isomor-
phism
τ ∈ HomOV <2> (p!1 FV , p!2 FV ).
It is easy to check that
p!23 (τ )p!12 (τ ) = p!13 (τ ). (3.3.19)
Conversely, given τ ∈ HomOV <2> (p!1 FV , p!2 FV ) satisfying (3.3.19), we can recover the D-crystal F. Since
both notions of D-crystal are defined locally, we may assume Y is affine. Consider a nilpotent thickening
i : Y → X. Since V is formally smooth, we may choose a finite morphism j : X → V such that j ◦ i = k.
Then j ! FV is a OX module, which we can take to be FX . If l : X → V is another finite morphism such
that l ◦ i = k, then we claim there is a natural isomorphism j ! FV ∼ = l! FV . Indeed, we obtain a morphism
(j, l) : X → V <2>
and j FV = (j, l) p1 FV , l FV = (j, l) p2 FV . Then (j, l)! (τ ) : j ! FV ∼
! ! ! ! ! !
= l! FV is our natural
isomorphism. We can show that FX defines a D-crystal. Indeed, the structure isomorphisms can be given
using the above argument, and compatibility follows from (3.3.19).
Thus, we have a description of Mcrys (Y ) in terms of V <i> . Using this description, we see that if
i : Y → Z is a closed embedding into a smooth variety, there are functors i∗ : Mcrys (Y ) → Mcrys (Z) and
i! : Mcrys (Z) → Mcrys (Y ). Indeed, consider V as above, and the induced map ĩ : V → Z. For i∗ , consider
the sheaf ĩ· FV with morphisms (ĩ × ĩ)· p!1 FV → (ĩ × ĩ)· p!2 FV . The reader can check that (ĩ × ĩ)· p!j = p!j ĩ· , and
that (3.3.19) is satisfied. This defines i∗ . The functors i! are left to the reader to define.
Let Mcrys (Z)Y be the D-crystals FZ such that j ∗ FZ = 0, where j : Z − Y → Z is the open embedding.
These are D-crystals supported on Y.
Lemma 3.3.20 The functors i∗ and i! induce an equivalence of categories between Mcrys (Y ) and
Mcrys (Z)Y .
The proof is easy, and left to the reader.
Proposition 3.3.21 There exists a natural equivalence between the category of right DY -modules (in
the sense of the last subsection) and Mcrys (Y ).
Proof. Suppose first that Y is non-singular and F is a D-crystal on Y . Then we are given an isomorphism
τ ∈ HomOY <2> (p!1 FY , p!2 FY ) satisfying (3.3.19). Using the isomorphism p!1 FY ∼
= FY ⊗OY DY and adjunction,
we may regard τ as a(τ ) ∈ Homp−1 OY (FY ⊗OY DY , FY ). We claim that a(τ ) makes FY into a right DY -
2
module. The property (3.3.19) implies that a(τ ) is a ring action. Indeed,
p!13 (τ ) ∈ HomOY <3> (p!13 p!1 FY , p!13 p!2 FY ) = HomOY <3> (q1! FY , q3! FY ).
Using Proposition 3.3 and adjunction, we may regard
p!13 (τ ) ∈ HomOY <3> (FY ⊗OY DY ⊗OY DY , q3! FY ) =

Homq−1 OY (FY ⊗OY DY ⊗OY DY , FY ).

3

49
By chasing through the identifications, one can check that p!13 (τ ) corresponds to the map v1 ⊗ ∂2 ⊗ ∂3 →
a(τ )(v1 ⊗ ∂2 ∂3 ). Similarly, p!23 (τ )p!12 (τ ) corresponds to the map v1 ⊗ ∂2 ⊗ ∂3 → a(τ )(a(τ )(v1 ⊗ ∂2 ) ⊗ ∂3 ).
Thus, a(τ ) gives a right action. By pulling back to Y, one sees that a(τ )(v ⊗ f ) = f v, for v ∈ FY , f ∈ OY .
Hence, FY is a right DY -module.
Conversely, if FY is a right DY -module, by adjunction we get an OY <2> morphism τ : p!1 FY → p!2 FY
which satisfies (3.3.19) from the action hypothesis. Moreover, we can prove by induction that τ is an
isomorphism. Recall that p!1,k FY consists of functions vanishing on I k+1 . It is clear that τ is an isomorphism
on functions living on the 0th infinitesimal neighborhood and that τ : p!1,k FY → p!2,k FY . Assume that τ is
an isomorphism through the kth infinitesimal neighborhood and let φ ∈ p!1,k+1 FY be such that τ (φ) = 0.
Let dx = 1 ⊗ x − x ⊗ 1 ∈ I. Then τ (dx · φ) = dx · τ (φ) = 0. The morphism dx · φ ∈ p!1,k FY , so dx · φ = 0 for
all x ∈ OY . Hence, φ vanishes on I · I k , so φ = 0. Surjectivity follows from an induction argument, which
we leave to the reader.
Now suppose Y is arbitrary, and choose a closed embedding Y → Z. By Lemma 3.3, we may regard a
D-crystal on Y as a D-crystal on Z supported on Y. It is routine to check that the corresponding right DZ
module FZ is supported on Y, and hence defines a D-module on Y. 2
3.3.22 Remark A more direct way to associate a D-crystal to a right DY -module can be given as
follows. Let Y ,→ X be a nilpotent thickening. Choose a closed embedding X ,→ Z of X into a smooth
variety Z. Then any object M of DY − mod is represented by some DZ -module MZ supported on Y . Let
FX = HomOZ (OX , MZ ).
See [BD] for more on D-crystals.
Now we give an interpretation of local cohomology in terms of the functors introduced above.
Proposition 3.3.23 If i : Y ,→ X is a closed embedding then
Γ[Y ] M ∼
= i∗ i! M.
Moreover, the same property is true for higher derived functors.
Proof. First note one has a diagram

i∗ i! Γ[Y ] M / i∗ i! M

 
Γ[Y ] M /M

where the vertical arrows are the augmentation morphisms. It follows from the definitions that the top
arrow is in fact an equality. Moreover, by Kashiwara’s theorem the left arrow is an isomorphism. Hence
Γ[Y ] M = i∗ i! M as submodules of M. To prove the property for higher derived functors, note that since i! is
left exact and i∗ is exact, i∗ i! is left exact. Since Γ[Y ] is also left exact, the property follows from uniqueness
of higher derived functors. 2
Proof of Proposition 3.2. Use the above Proposition and Proposition 3.3 and Kashiwara’s theorem. 2.
3.3.24 Remark The functor i! takes just those sections of M that are annihilated by the ideal sheaf
IY itself. By “differentiating in transversal directions” the functor i∗ restores the sections of M that are
annihilated by higher powers of IY .

3.3.25 Applications of Kashiwara’s Theorem.

Let f : Y ,→ X be a closed submanifold. Recall that one has an exact sequence of vector bundles on Y :
π
0 → TY∗ X → T ∗ X|Y → T ∗ Y → 0.
50
 Corollary 3.3.26 (Bernstein) If N is a coherent DY -module then f∗ N is a coherent DX -module and,
moreover, SS(f∗ N ) = π −1 (SS N ).
Proof. Locally Y is given by vanishing of some set of funtions t1 , . . . , td . Then locally one has f∗ N =
DY [∂t ] ⊗DY N = N [∂t ] and both claims of the corollary follow. 2
Corollary 3.3.27 (Weak Gabber theorem) If M is a DX -module and dim SS M < dim X then M = 0.
Proof. The characteristic variety SS M ⊂ T ∗ X always projects surjectively onto Supp M ⊂ X. Hence if
SuppM = X we are done. Hence we can assume that Y = Supp M is a proper subvariety of X. We can
choose an affine open subset U such that the intersection Y 0 = U ∩ Y coincides with the smooth locus of a
certain irreducible component of Y of maximal dimension (i.e. dim Y ). Replace X by U and Y by Y 0 and
apply Kashiwara’s theorem to M|U . It follows that M = i∗ N where N = i! M. By the previous corollary,
SS M = π −1 (SS N ). Arguing by induction on dimension we can assume that dim SS N ≥ dim Y hence
dim SS M ≥ dim X. 2
3.3.28 Remark Comparing this weak form of Gabber theorem with the original statement one notices
two differences: first, “strong” Gabber’s theorem implies that every irreducible component of SS M has
dimension at least dim X; and second, the weak version says nothing about the coisotropicity property.
3.3.29 f∗ for open embeddings.
If j : U ,→ X is an open embedding, then the higher direct image sheaves Ri j∗ (M ) have DX -module
structure (constructed exactly as in the case of cohomology with support, i.e. either via action on the Cech
j+1
complex or via the local cohomology construction using Ri j∗ (M ) ' H[X\U ] (M )).

3.4. Beilinson-Bernstein theorem for projective space. Let X be a d-dimensional projective

space CPd .
Theorem 3.4.0.1. (Beilinson-Bernstein)
(i) The functor of global sections
Γ : M → Γ(X, M)
is exact on the categry of D-modules on X.
(ii) Every DX -module is generated by its global sections.
3.4.1 Remarks.
(1) Both claims of the theorem can be restated as a claim that DC Pn is a projective generator in the
category of D-modules on C Pn .
(2) A well known analogue of this theorem for the category of OX -modules (due to Serre) claims that the
functor of global sections is exact on the category of O-modules on an algebraic variety X if and only if X is
affine. For this reason Beilinson and Bernstein call the varieties satisfying the property above D-affine. The
only known examples of such varieties are products of affine varieties and projective homogeneous spaces of
reductive Lie groups.
(3) From this theorem one deduces immediately that there exists an equivalence of categories:
{coherent DCPd − modules} ←→ {f.g. modules over Γ(CPd , DCPd )}
Moreover, if CPd = P(V ) and x0 , . . . , xd are coordinates on V , then the algebra Γ(CPd , DCPd ) is generated
over Γ(CPd , OCPd ) = C by the global vector fields vij = xj ∂j satisfying the following relations:

(xi ∂j )(xk ∂l ) = (xi ∂l )(xk ∂j ) + δjk · (xi ∂l ) − δkl (xi ∂j ).

d
Thus Γ(CP , DCPd ) is a quotient of the universal enveloping algebra U gl(V ). The reason for this is that
gl(V ) acts on P(V ) by vector fields, hence there is a morphism gl(V ) → Γ(CPd , TCPd ) which extends to a
morphism U gl(V ) → Γ(CPd , DCPd ) by the universal property of U gl(V ).
51
 Lemma 3.4.2 The morphism of sheaves OCPd ⊗ U gl(V ) → DCPd is surjective. In particular, DCPd is
generated by its global sections.
Proof. This follows from the surjectivity of the associated graded map OCPd ⊗ S gl(V ) → S TCPd . 2
One can show in a similar way that the relations above form a complete set.
Proof of the theorem. As in the remark above, we assume that CPd = P(V ). We also denote by V∗ the open
subset V \ {0} of V and consider the maps:
p j
P(V ) ←− V∗ ,→ V.
We write j· for the sheaf-theoretic direct image and compare Γ(V, j· p∗ M) with Γ(P(V ), M). First of all,
∗
note that there is a natural embedding Γ(P(V ), M) ,→ Γ(V,
P j· p M). To identify Γ(P(V ), M) as a subspace
∗
of Γ(V, j· p M), recall that the Euler vector field θ = i xi ∂i measures the homogeneous degree of any
polynomial (or tensor) on V (or V∗ ). One can easily see that the image of Γ(P(V ), M) in Γ(V, j· p∗ M)
coincides with the subspace annihilated by θ.
Hence the theorem reduces to showing that the functor
M → Γ(V, j· p∗ M)θ
is exact. To do that we consider the exactness properties of the various functors involved:
(1) p∗ is exact since p is a smooth morphism.
(2) j· is left exact, but it is not right exact since j· is not affine.
(3) Γ is exact on V since V is affine.

θ
(4) . . . is exact on modules of the type j· p∗ M since the action of θ on them is semisimple.
00
Therefore, a short exact sequence 0 → M0 → M → M → 0 of DP(V ) -modules induces a short exact
00
sequence 0 → p∗ M0 → p∗ M → p∗ M → 0. Applying j· we obtain the long exact sequence
00
0 → j· p∗ M0 → j· p∗ M → j· p∗ M → R1 j· p∗ M0 → . . . . (3.4.3)
Now notice that the eigenvalues of θ on the global sections of R≥1 j· p∗ M are negative. In fact, since j is an
isomorphism away from 0 ∈ V , we have Supp(R≥1 j· p∗ M) = {0}. We also know that each sheaf R≥1 j· p∗ M
has a structure of a DV -module by 3.3. By Kashiwara’s theorem any such module has to be of the type i∗ (W )
where i is the embedding {0} ,→ V and W is a coherent D-module on {0}, i.e. just a finite-dimensional
vector space. We can assume without loss of generality that W = C. Then i∗ (C) = C[∂0 , . . . , ∂d ] · δ. But
θ · δ = −(dim V )δ (since xi ∂i · δ = −δ + ∂i xi · δ = δ), and applying ∂i to δ can only decrease the eigenvalue
of θ. Hence the eigenvalues of θ on Γ(V, R1 j· p∗ M0 ) are negative and the long exact sequence (3.4.3) induces
the short exact sequence
00
0 → Γ(V, j· p∗ M0 )θ → Γ(V, j· p∗ M)θ → Γ(V, j· p∗ M )θ → 0
hence the short exact sequence
00
0 → Γ(P(V ), M 0 ) → Γ(P(V ), M ) → Γ(P(V ), M ) → 0.
This proves (i).
To prove (ii) let M0 ⊂ M be the submodule generated by the global sections of M. One has an exact
sequence
0 → M0 → M → M00 → 0
and by exactness of Γ(P(V ), ·) this gives Γ(P(V ), M00 ) = 0. We will show that this implies M00 = 0. In fact,
consider the decomposition of Γ(j· p∗ M00 ) into a sum of eigenspaces with respect to θ:
X
Γ(j· p∗ M00 ) = Γ(n).
n∈Z

52
We have shown before that Γ(P(V ), M00 ) = Γ(0). Similarly one can show that Γ(P(V ), M00 (n)) = Γ(n).
Hence if M00 6= 0 then Γ(n) 6= 0 for n  0. Take a non-zero section m ∈ Γ(n) for such n. If all ∂i annihilate
m, then θ ·m = 0, which is a contradiction. Hence, by appplying some nth order constant coefficient operator
to m, we will get a non-zero element of Γ(0), giving a contradiction. 2

3.5. Functors f ! and f∗ for a general map.

Let f : Y → X be a morphism of varieties. The functor f + from DX modules to DY modules was
defined in 3.3. We want to generalize the functor f ! , but we will do this only on the derived category. Recall
that we denote by Db (DX ) the derived category of bounded complexes of DX modules. Denote also by
f + : Db (DX ) → Db (DY ) the derived extension of f + . It is computed by replacing a complex M with a
locally free resolution P → M and then applying f + .
Definition 3.5.1 f ! : Db (DX ) → Db (DY ) is the functor given by f ! (M) := f + (M[d]), where
d = dim(Y ) − dim(X).
Note that this definition is consistent with the earlier definition in 3.2, by Proposition 3.3 (iii).
We also want to define a direct image functor f∗ : Db (DY ) → Db (DX ). This functor is defined as the
composition of a left exact functor with a right exact functor, and as such does not have a good definition at
the level of modules, but is only defined in the derived category. We will partially circumvent this problem
by first defining direct image for a projection, where one can give an explicit resolution.
Let p : Y = Z × X → X be projection on the second factor and let q : Y → Z be projection on the first
factor. There is an isomorphism DY ∼ = p−1 DX ⊗CY q −1 DZ .
Definition 3.5.2 For M ∈ Db (DY ),
p∗ (M) := Rp. (q ∗ ΩZ ⊗L
q ∗ DZ M) = Rp. (q
−1
ΩZ ⊗L
q −1 DZ M).

Here the symbol ⊗L means we take the derived functor defined by ⊗, i.e., we resolve M and/or q −1 ΩZ
by projectives and work with the complex. Rp. is the usual direct image of sheaves. Since p−1 DX commutes
with q −1 DZ , it follows that DX acts on p∗ (M).
In this definition, q −1 ΩZ ⊗Lq −1 DZ M is not obviously a complex of OY -modules, so it is not clear that
the direct image can be computed in the O category. Thus, it is not clear that p∗ (M) is a quasi-coherent
OX module, and moreover, it is not clear that when Z is affine, Ri p. = 0 for i > 0. We give an equivalent
definition for q −1 ΩZ ⊗Lq −1 DZ M which exhibits the O module structure.
Let
DX←Y = DY →X ⊗OY ΩY /X = p∗ (DX ⊗OX ΩX −1 ) ⊗OY ΩY .
Since DX has left and right DX module structures, it follows that DX ⊗ ΩX −1 has two commuting DX
module structures. In particular, the pullback p∗ (DX ⊗ ΩX −1 ) has a left DY module structure given by
pulling back the shift of the right module structure, so p∗ (DX ⊗ ΩX −1 ) ⊗OY ΩY is a right DY module, as
well as a left p−1 DX module. Thus, for M ∈ Db (DY ), DX←Y ⊗L DY M has a p−1 DX module structure. We
claim that
DX←Y ⊗L DY M ∼ = q −1 ΩZ ⊗L q−1 DZ M.
The claim follows from the definition and the identification DY = q −1 DZ ⊗ p−1 DX .
Now we can show that q −1 ΩZ ⊗L q−1 DZ M can be treated as an OY module. Since we are working locally
on the base, we can assume that X is affine, and we assume also that Z is affine. Then we can replace M
by a free resolution F → M be a complex of free DY modules. Then DX←Y ⊗L DY M ∼ = DX←Y ⊗L DY F
is a direct sum of copies of DX→Y and hence is a quasi-coherent OY module. It follows that p∗ (M) is a
quasi-coherent OX module and Rp. can be computed in the category of OY modules. In general, we use a
covering of Z by affine open subsets, and the Cech resolution.
53
 To compute p∗ (M), we give an explicit resolution of ΩZ as a right DZ module. Consider the complex
of regular differential forms
A0Z → A1Z → . . . → AnZ
with de Rham differential d, and n = dim(Z). Choose local coordinates z1 , z2 , . . . on Z and regard them also
as coordinates on Y by pullback. If M is a DY module, we can consider the relative de Rham complex
A·Y /X (M) := q −1 A·Z (M) =

q −1 A0Z ⊗q−1 OZ M → q −1 A1Z ⊗q−1 OZ M → . . . → q −1 AnZ ⊗q−1 OZ M

with differential d given by
X
d(ω ⊗ m) = dw ⊗ m + dzi ∧ ω ⊗ ∂/∂zi · m.
i

It is easy to check that d2 = 0.

Let Li (M) = H i (A·Y /X (M)[n]). Then
L0 (M) = q −1 ΩZ ⊗q−1 (OZ ) M/TZ M.
If the constant coefficient operators C[∂/∂z ] act freely on M, then L−i (M) = 0 for i > 0. This follows by
realizing ΩZ · as a Koszul complex.
We apply this in particular to the case M = q ∗ (DZ ), which is both a left DY module and right q ∗ DZ
module. Then A·Y /X (q ∗ DZ )[n] is a complex of free right q ∗ DZ modules. By the above comments, it follows
that Lj (q ∗ DZ ) = 0 if j 6= 0, and L0 (q ∗ DZ ) ∼
= ΩZ since DZ /TZ DZ ∼
= OZ . Thus, A·Y /X (q ∗ DZ )[n] → q ∗ ΩZ is
∗ ∗
a resolution of q ΩZ by free right q DZ modules. Hence, we can identify
q ∗ ΩZ ⊗L∗
q (DZ ) M∼= A· (q ∗ DZ )[n]⊗L∗ M.
Y /X q DZ

But this last module is isomorphic to A·Y /X (M)[n] with the differential d. We conclude that
p∗ (M) = Rp. A·Y /X (M)[n].

3.5.3 Remark If M is a DY module, then Li (M) can have cohomology in degrees −n ≤ i ≤ 0. If p is

affine, p∗ (M) can have cohomology only in degrees −n ≤ i ≤ 0, but in general p∗ (M) can have cohomology
in degrees −n ≤ i ≤ z.
3.5.4 Examples. Consider the projection p : Cn → Cn−1 . Then

0 if i 6= −1
Hi (p∗ OCn ) =
OCn−1 if i = −1.

i 0 if i 6= 0
H (p∗ B0|Cn ) =
B0|Cn−1 if i = 0.

i 0 if i 6= 0
H (p∗ DCn =
DCn /∂/∂zn DCn ⊗ OCn−1 dzn if i = 0.
Note that in this last case, direct image of the coherent module DCn is not coherent.
We can factor an arbitrary map f : Y → X as Y → Y × X → X, where the first map is i : y 7→ (y, f (y))
and the second map is p, projection on the second factor.
Definition 3.5.5 f∗ for a morphism
f∗ (M) := p∗ i∗ (M), M ∈ Db (DY ).

54
 3.5.6 Remark if f : Y → X is a closed embedding, then f∗ coincides with the definition given before,
as follows from Lemma 3.5 below. This would not be the case if we had not inserted the shift by n in the
definition of direct image for a projection.
We want to prove that direct image is a functor. First we need some easy lemmas which are special
cases of the claim.
Lemma 3.5.7 Let p = p1 ◦ p2 : Z × Y × X → Y × X → X be the sequence of projections. Then
p∗ = p1 ∗ ◦ p2 ∗ .
Proof. When Z and Y are affine, this is a routine exercise. The general case follows using Cech coverings.2
Lemma 3.5.8 Let i : Y → W be a closed embedding, and consider the diagram
X ×Y /Y
pY

id×i i
 
X ×W /W
pW

where the horizontal arrows are the projections. Then i∗ ◦ pY ∗ = pW ∗ ◦ (id × i)∗ .
Lemma 3.5.9 Let p : U × V → V and q : W × V → V be the projections, and let i : U → W be a closed
embedding, and i × idV : U × V → W × V be the induced embedding. Then p∗ = q∗ (i × idV )∗ .
We leave the proofs of these last two lemmas to the reader, noting the idea that direct image for a closed
embedding adjoins normal derivatives, while direct image for a projection applied to a module free in the
projected variables kills the normal derivatives.
Proposition 3.5.10 Let f : Z → Y and g : Y → X be morphisms of smooth varieties. Then
(g ◦ f )∗ = g∗ ◦ f∗ .
Proof. g∗ f∗ = pY ×X,X ∗ ig ∗ pZ×Y,Y ∗ if ∗ , where pU ×W,W : U × W → W is the projection, and for a map
h : U → V, ih : U → U × V is the closed embedding to its graph. By Lemma 3.5 applied to the diagram of
embeddings and projections
Z ×Y /Y ,

 
Z ×Y ×X / Y ×X

g∗ f∗ = pY ×X,X ∗ pZ×Y ×X,Y ×X ∗ (iid×g )∗ if ∗ .

By Lemma 3.5 and Proposition 3.3, it follows that g∗ f∗ = pZ×Y ×X,X ∗ (if ×g◦f )∗ . On the other hand,
(g ◦ f )∗ = pZ×X ∗ ◦ (if ◦g )∗ = pZ×Y ×X,X ∗ (if ×idX )∗ (ig◦f )∗ = pZ×Y ×X,X ∗ (if ×g◦f )∗ ,
using successfully, Lemma 3.5, Lemma 3.5, and Proposition 3.3.2

3.6. D-modules and local systems.

Note that any D-module M can be viewed as a quasi coherent O-module with flat connection. In fact,
the D-module structure on an O-module M is completely determined by the action map T ⊗ M → M, or
equivalently, as a connection map M → M ⊗ Ω1 which is flat by definition of a D-module. In particular,
any local system, i.e. a O-coherent locally free sheaf with flat connection, has a structure of a D-module.
Proposition 3.6.1 Let M be a D-coherent module. Then the following properties are equivalent:
(i) SS M is a subset of the zero section subvariety in T ∗ X (by coisotropicity property it then has to be
equal to the zero zection),
(ii) M is O-coherent,
(iii) M is a local system.
55
Proof. First of all, (ii) follows from (iii) by definition of a local system. To show that (i) and (ii) are
equivalent, note that SS M = Supp grM is a subvariety of the zero section iff the images of ∂/∂xi in grDX
act on M by nilpotent operators. This in turn implies that gr M is coherent over OX ⊂ grDX , hence M is
coherent over OX . If M is O-coherent, we can consider a one-term filtration on M consisting of M itself
and reverse the argument.
To show that (ii) implies (iii), suppose that M is OX -coherent and let mx ⊂ OX be the maximal ideal
of the point. Then the geometric fiber M/mx M over x is finite-dimensional. It suffices to show that the
function dimC M/mx M is constant on X. Indeed, if we choose a set of n local sections of M generating
the geometric fiber over x, then Nakayama’s lemma implies that the sections generate the local ring Mx ,
and hence they generate the geometric fibers in a neighborhood of x. Since the dimension of the fibers is
constant, it follows that the sections are linearly independent in each geometric fiber and hence functionally
independent in a neighborhood. [ASK ABOUT A REFERENCE FOR THIS] To show that the dimension
of the geometric fiber is constant, choose two point x and y in X and assume that they can be connected
by a smooth irreducible curve Y ⊂ X (since in general we can always connect x and y with a chain of such
curves).
Denote by ix , iY and j the inclusions {x} ,→ X, {Y } ,→ X and {x} ,→ Y , respectively. Then one
+ +
can write M/mx M = i+ x M = j (iY M). The objects obtained at each step are O-coherent (recall that
+
the pullback iY M etc., in the sense of O-modules has a natural D-module structure). So, it suffices to
prove the statement assuming that X = Y . In this one dimensional case a sheaf is locally free iff it has
no torsion and this is what we need to show for our module M on Y . In fact, suppose this module has a
torsion submodule at some point z ∈ Y . Then i!z M = 6 0 and by right exactness of (iz )∗ we have a non-zero
submodule (iz )∗ i!z M = Γ[z] M ,→ M . But by a local computation we know that the direct image of any
non-zero module from {z} is not an O-coherent module on Y . 2

3.7. Simple non-holonomic modules. In was believed for some time that the characteristic variety
of every simple D-module has the minimal possible dimension, i.e., that every simple D-module is holonomic,
see §4.1 of the next chapter. We will show below (following Bernstein and Lunts) that in some sense “most”
simple modules are not holonomic. In what follows we will assume that X = Cn .

Conjecture 3.7.1 Let u ∈ D(Cn ). If σ(u) is generic among homogeneous polynomials of deg k ≥ 4
then M = D/D · u is simple. Since SS M is equal to the zero-variety of σ(u), M is not holonomic for n ≥ 2.

Theorem 3.7.1.1. (Bernstein-Lunts[BeLu]) The conjecture above is true for n = 2.

The proof of this theorem will occupy the rest of this section.
As in the previous section, we will use the Bernstein filtration defined by deg x = deg ∂ = 1. Of course,
this does not exist on a general smooth algebraic variety, and also the notion of holonomicity changes
completely. However, we will work with this filtration (and hence a different notion of holonomicity) since it
makes the basic idea of the proof simpler and in any case it can be repeated on the original situation with
slight modifications.?????
Observe that grD(Cn ) can still be identified with functions on the cotangent bundle, so that the char-
acteristic variety still lives in T ∗ (Cn ). By the theorem of Involutivity of the Characteristic Variety Theorem
(see section 1.2 REFERENCE), SS M is a coisotropic subvariety of T ∗ Cn = C2n . We also note that SS M
is a cone-subvariety of C 2n with respect to the natural C∗ -action (this is the main effect of changing the
filtration since for the usual filtration the group C∗ acts along the fibers of T ∗ Cn → Cn ). We will call any
coisotropic cone-subvariety a cc-(sub)variety.

Definition 3.7.2 A cc-subvariety Σ ⊂ C4 is minimal if is does not contain any proper cc-subvariety
Σ0 ( Σ.
56
 In particular, any minimal cc-subvariety is irreducible. Also, any irreducible Lagrangian cone-subvariety
(equivalently, the closure of the conormal bundle to a smooth subvariety) is minimal. The main observation
of Bernstein and Lunts is that there are plenty of minimal cc-subvarieties that are not Lagrangian.
More precisely, we will show that
Theorem 3.7.2.2. If P is a generic homogeneous polynomial of deg k ≥ 4 on C4 then the zero variety
{P = 0} is a minimal cc-subvariety
We will show below that any hypersurface in a symplectic variety is automatically coisotropic, and
thus {P = 0} is a cc-subvariety. Once the theorem is proved, we can take any differential operator u with
σ(u) = P (of course, the symbol is taken here in the sense of the Bernstein filtration), to obtain an example
of a simple non-holonomic module:
Lemma 3.7.3 If u is as above then D/D · u is simple, i.e. D · u is a maximal left ideal.
Proof. If the claim is not true then there is a non-trivial inclusion of left ideals D · u ( J, which induces
a nontrivial inclusion of associated graded ideals, and hence a nontrivial quotient map D/D · u  D/J. It is
easy to check that SS D/J is the zero set of σ(J), the ideal generated by symbols of J. Since P is generic, it
may be assumed irreducible,
so the characteristic variety SS D/J is a proper cc-subvariety of the minimal
cc-subvariety SS D/D · u , which cannot happen. 2
3.7.4 Digresssion on coisotropic subarieties.
To prove that the zero variety of a generic polynomial is minimal we need to study the geometry of
coisotorpic subvarieties in C4 endowed with the natural syplectic form ω = dx1 ∧ dx2 + dx3 ∧ dx4 . We will
denote C4 with this symplectic structure by M . Recall that a subvariety Σ ⊂ M is called coisotropic is its
ideal sheaf IΣ satisfies {IΣ , IΣ } ⊂ IΣ where {· , ·} is the Poisson bracket (algebraic definiton). It follows that
a hypersurface is coisotropic. Alternatively, we can say that for any smooth point x ∈ Σreg the tangent space
Tx Σ ⊂ Tx M is a coisotropic subspace, i.e. Tx Σ⊥ ⊂ Tx Σ (geometric definition) . Therefore in the tangent
bundle to the smooth part Σreg we have a subbundle formed by Tx Σ⊥ ⊂ Tx Σ. This subbundle is called the
nul-foliation and we summarize its properties in the following
Lemma 3.7.5
(i) T Σreg ⊥ coincides with the radical of the form ω|Σreg .
(ii) T Σreg ⊥ is integrable, i.e. a bracket of two vector fields in T Σreg ⊥ is again a vector field in T Σreg ⊥ .
2
By Frobenius integrability theorem T Σreg ⊥ is tangent to a familiy of curves in Σreg . In a particular case
when Σ is a hypersurface in M given by an equation P = 0, the nullfoliation T Σreg ⊥ is generated by the
Hamiltonian vector field ξP corresponding to P .
Idea of proof of (3.7.2.2). The main idea behind the proof of Bernstein-Lunts is that a solution of an
algebraic differential equation is usually just analytic. Suppose now that Σ0 ( Σ is a proper cc-subvariety.
Then at a point x which is smooth for both Σ0 and Σ one has (Tx Σ0 )⊥ ⊃ (Tx Σ)⊥ . Therefore Σ0 is formed
by the leaves of the foliation, i.e. ξP is tangent to Σ0 at its smooth points. When we apply Frobenius
integrability in general we will not get an algebraic subvariety Σ0 .
Suppose that Σ0 is algebraic, and denote the projective variety corresponding to Σ0 by C. The condition
of genericity we will impose on P basically says that if C exists then it has only ordinary double points as
singularities. We know that C is tangent to the direction field ξ¯ on the projectivization Σ obtained from
ξP . Suppose for a moment that ξ¯ is actually a vector field. Then, at each singular point, ξ¯ defines a linear
operator on its tangent space to Σ. Suppose this operator is diagonalizable. Then the ratio of its eigenvalues
makes sense even if xi¯ is just a direction field (not a vector field). Now we can state the genericity conditions
imposed on P :
(a) dP does not vanish away from the origin of C4
(b) For any singular point of ξ,¯ the ratio of the corresponding eigenvalues is irrational.

57
 To show that these two conditions imply that C has only ordinary double points one uses the following
theorem from the theory of differential equations:
Theorem 3.7.5.3. ([Ar]) Suppose that a vector field ξ on a two-dimen sional manifold defines a diago-
nalizable operator on the tangent space to its singular point. Suppose futher that the ratio of the eigenvalues
λ/µ is irrational. Then there exists a system of (formal) coordinates which linearizes ξ:
∂ ∂
ξ = λy1 + µy2 . 2
∂y1 ∂y2

Proposition 3.7.6 Suppose ξ is as in the theorem and ξ is tangent to a curve C at a point m. Then at
m, either C is smooth or has an ordinary double point.
Now suppose ξ is tangent to a curve C at a point m and let f generate the ideal J of C. If m is a regular
point for ξ, then we can find local coordinates (y1 , y2 ) on a formal neighborhood such that ξ = ∂/∂y1 , and
it follows that the ideal of C is generated by y2 in the formal neighborhood.
P If m isβ1a singular point, choose coordinates in a formal neighborhod as in the theorem. Write f =
aβ1 ,β2 y1 y2 β2 . We show that each monomial y β := y1 β1 y2 β2 is in the ideal of C. Since the ratio of
eigenvalues is irrational, it follows that ξ has distinct eigenvalues on its eigenvectors, the monomials y β .
Hence, for any integer N > 0, we can find a polynomial Qβ such that Qβ (ξ)(y β ) = y β and Qβ (ξ)(y ν ) = 0 for
deg(ν) < N, ν 6= β. Since ξ preserves the ideal J, it follows that y β ∈ J + mN , where m is the maximal ideal
of m. Since J is closed in the m-adic topology by teh Artin-Rees lemma, it follows that y β ∈ J. But now
recall that f generates J and let y β be a monomial of minimal degree of f. Then y β = hf for some power
series h and h(m) 6= 0. Hence, h is invertible in the ring of formal power series, so y β generates J. Since J is
radical, it follows that y β = y1 , y2 , or y1 y2 . 2
Proof of (3.7.2.2). Recall that in our case M = C4 and Σ is a three-dimensional subvariety given by the
equation {P = 0} where P is a homogeneous polynomial of degree ≥ 4. Since be have a natural C∗ -action,
we can pass to the projective picture, i.e. consider a projective algebraic surface Σ ⊂ CP3 . Then any proper
cc-subvariety Σ0 will give us a curve C on Σ, so we need to study the curves on Σ. A classical theorem of
Noether asserts that since k ≥ 4, any such curve is given by vanishing of a homogeneous polynomial Q of
degree l [De]. Therefore, the ideal sheaf of PΣ0 is generated by two functions P and Q.
Recall that the Euler vector field Eu = i xi ∂i allows us to define a 1-form α = iEu ω such that dα = ω.
One easily sees that a cone-subvariety of M is isotropic iff the restriction of α to it vanishes.
Since Σ0 in our case is 2-dimensional and ω is non-degenerate, we have (Tx Σ0 )⊥ = Tx Σ0 , so Σ is also
isotropic and α|Σ0 = 0. Since P and Q generate the ideal of Σ0 , dP and dQ generate the conormal bundle,
and hence there exist regular functions f, g on C4 such that the equality
α|Σ0 = f dP + g dQ.
holds on the regular part (Σ0 )reg of Σ0 (where f and g are regular functions on (Σ0 )reg . Since α, P and Q
are homogeneous, we deduce that f and g are also homogeneous of degrees (2 − k) and (2 − l) respectively.
We denote by OC (i) the restriction of the linear bundle O(i) from CP3 to C. Hence we can think of f (resp.
g) as a section of OC (2 − k) (resp. OC (2 − l)) over the regular part of C.
Case 1. Assume C is smooth. Then by positivity of OC (1) the bundles OC (2 − k) and OC (2 − l) have
no global sections if 2 − k < 0 and 2 − l < 0. Since we assumed that k ≥ 4, we have f = 0 and l = 1 or 2. [[
DIDN’T UNDERSTAND THIS. REWROTE AS IN BERNSTEIN-LUNTS, BUT MAYBE I AM MISSING
SOMETHING OBVIOUS We will consider here the case l = 2 (the other case is easier since for l = 1 the
curve C is just a hyperplane section of Σ).
Let SQ = {x ∈ C4 | ξg = c · Eu for some non-zero constant c}. We claim that SQ is a finite union of
linear subpaces. In fact, define an operator A : C4 → C4 by viewing Q as a symmmetric bilinear form and
putting Q(·, ·) = ω(A·, ·) then the equation defining Sg translates into condition that x is an iegenvector of
58
A. Since Σ0 is coisotropic, it has to be a two-dimensional linear subpace. Hence C has to be a projective
line which is impossible since deg C = kl > 1. ]]
Thus, α is proportional to dQ at each point of Σ0 . We claim that Σ0 lies in a hyperplane. If l = 1, this
is obvious. If l = 2, let SQ = {x ∈ C4 | ξQ = c · Eu for some non-zero constant c}. We claim that SQ is
a finite union of linear subpaces. In fact, define an operator A : C4 → C4 by viewing Q as a symmmetric
bilinear form and putting Q(·, ·) = ω(A·, ·) then the equation defining SQ translates into condition that x is
an eigenvector of A. Since Σ0 is irreducible, it follows that Σ0 lies on a hyperplane.
Let φ be a linear function defining the hyperplane with constant Hamiltonian vector field ξφ . Choose
some vector v ∈ Σ0 not proportional to ξφ . Since Σ0 is a cone, Cv ∈ Σ0 . Since ξφ is a constant vector field
tangent to Σ0 , Cv + Cξφ ⊂ Σ0 . Since Σ0 is two-dimensional, it is a linear plane. Hence C has to be a degree
one projective line which is impossible since deg C = kl > 1.
Case 2. C is singular. Since the vector field ξP is homogeneous of degree (k − 1), it does not descend to a
vector field on C P3 . However, is still gives a one-dimensional subbundle in the tangent bundle (or a direction
field) over an open part of C P3 and our curve C is tangent to it. The preimage of singular points of C in
C4 \ {0} belongs to the zero set of the field ξP (since this preimage is defined by conditions d P = d Q = 0
and the first of the describes the zeros of ξP ).
Recall that C has only ordinary double points (that is in some formal neighborhood with coordinates
y1 , . . . , y4 in C4 of any singular point, Σ = {y3 = 0} and Σ0 = {y3 = 0 = y1 y2 }. Let C e → C be the
0 reg reg
normalization. We claim that regular functions f and g on (Σ ) defined by α|Σ0 = f dP + g dQ, viewed as
reg
sections of some line bundles on C , extend to C. For this we can use the formal local coordinates y1 , . . . , y4
e
above. One has to show that f and g do not have any poles on the two branches y1 = 0 and y2 = 0 of C in
the neighbourhood of a singular point. In fact, P = y3 (since Σ = {y3 = 0}) and Q = ε P + δ y1 y2 where ε
and δ are formal power series in y and δ(0) 6= 0. Then we can write an equality of differential forms on Σ0
X
α= αj dyj = f dP + g dQ = (f + εg)dP + gd(δ · y1 y2 ) + gP dε,

and we can ignore the last term because P vanishes on Σ0 . Since d(δ · y1 y2 ) = δd(y1 y2 ) + y1 y2 dδ and y1 y2
vanishes on Σ0 , we can absorb δ into g, i.e. assume that Q = y1 y2 .
Recall that we want to extend f and g to Σ0i = {yi = y3 = 0}, i = 1, 2. We see that α4 |Σ0 = 0 and
α3 |Σ0 = f + εg. Then, g|Σ02 = αy12 and since α2 vanishes at the origin of Σ02 , the ratio does not have a pole
at the origin of Σ02 which means that g extends to Σ02 . It follows that f = α3 − εg extends to Σ01 . Similarly,
f, g extend to Σ01 .
Now our goal is to show that for under the genericity condition for P described above we will in fact
have only ordinary double points. Recall that Σ is given by the equation {P = 0}. First of all we assume
that df does not vanish on C4 \ {0} (in particular P is irreducible and Σ is smooth away from the origin).
Let SP = {x ∈ C4 | ξP = a · Eu for some non-zero constant a}.

Lemma 3.7.7 Under the conditions described above one has dim Sf = 1 Proof. Define SP = {x ∈
C4 | ξP = Eu}. Then SP0 ⊂ SP = C∗ · SP0 so it is enough to prove that SP0 is a finite set. To that end,
Pi ∂∂xi and note that by our assumption Pi have no common zeros outside the origin. Since
P
write ξP =

k−1
|Pi |2 ≥ C · |xi |2 for all x ∈ C4 . The
P P
deg Pi = k − 1, there exists a constant C > 0 such that
set SP is given by the equations {P1 = x2 , P2 = x1 , P3 = x4 , . . . } hence for any point x ∈ SP0 one has
0
1
|xi |2 ≤ C k−2 . If SP0 is not finite we obtain a contradiction since any algebraic set of positive dimension is
P
necessarily unbounded. 2

Now the theorem on a canonical form of a vector field quoted above finishes the proof. 2

3.7.8 A generalization of Hol(DX )

59
 Recall that by Hol(DX ) we denoted the abelian category of holonomic DX -modules. Now we can also
define Min(DX ) to be the category of all DX -modules M such that every component of SS M is a minimal
coisotropic subvariety. Generalizing from the lagrangian case one easily proves the following
Proposition 3.7.9 Every object in Min(DX ) has finite length. 2
In other respects the category Min(DX ) is not studied at all.

60
 4. Holonomic D-modules.
4.1. Category of holonomic modules.
Definition 4.1.1 A coherent D-module M is called holonomic if
dim SS M = dim X.

Proposition 4.1.2
(i) Holonomic D-modules form an abelian subcategory HolX of the category of all coherent D-modules.
In particular, every subquotient of a holonomic module is itself holonomic.
(ii) Every holonomic module has finite length as a D-module.
Proof. To prove the first claim, note that for any subquotient M0 of a holonomic module M we have
SS M0 ⊂ SS M. Hence dim SS M0 ≤ dim X. But by the Involutivity of Characteristic Variety theorem for
any D-module M0 one has dim SS M 0 ≥ dimX. Hence in our case we have dim SSM0 = dim X.
To prove the second claim we need to show that any decreasing chain of submodules of a holonomic
module M necessarily terminates. In fact, let S1 , . . . , Sn be the irreducible components of SS M which have
multiplicities m1 , . . . , mn respectively. Since the characteristic variety of any subquotient of M is a union of
several components Si and the multiplicities are additive with respect to short exact sequences, any strictly
decreasing chain of submodules of M has at most m1 + . . . + mn terms. 2
Now we prove a small lemma which gives a partial description of the characteristic variety of a holonomic
D-module. Recall that T ∗ X is a symplectic manifold (REFERENCE TO CHAPTER TWO) with a C∗ action
along the fibers and corresponding Euler vector field Eu. A subvariety of T ∗ X is called a cone-subvariety if
it is C∗ -invariant. Let λ = iEu ω, the “tautological” 1-form on T ∗ X, i.e., if ξ is tangent to T ∗ X at a covector
α, then by definition λ(ξ) = α(π∗ ξ), where π is the projection T ∗ X → X.
Lemma 4.1.3 Let X be a manifold and Λ an irreducible Lagrangian cone-subvariety in T ∗ X. Then there
exists a locally closed smooth subvariety Y of X such that Λ = TY∗ X and Y = π(Λ)reg .
Proof. Since Λ is a cone-subvariety, Eu is tangent to Λreg . Since Λ is Lagrangian, ω(Eu, ξ) = 0 if ξ is tangent
to Λ, so λ vanishes on Λ. Let Y = π(Λ)reg , and choose y ∈ Y and generic α ∈ Λreg such that π(α) = y and
the tangent map Tα (Λ) → Ty Y is onto using the Bertini-Sard theorem. Then it follows from the definition
of λ that α ∈ TY∗ X. Since a generic vector α ∈ Λ lies in TY∗ X, and Λ and TY∗ X are both irreducible and of
the same dimension, it follows that they coincide. 2
Corollary 4.1.4 Let M be a holonomic module and Supp M = Y. Then SS M = TY∗reg X∪{other irreducible components}.
2
Proof. Let Z ⊂ T ∗ X be the union of the irreducible components of Λ ⊂ SS M such that the image of the
generic point of Λ is not the generic point of Y . Thus, π(Z) ∩ Y is a proper closed subset in Y . Write
Y reg for the smooth locus of Y , and set Y ◦ := Y reg r π(Z). Choose a Zariski open subset U ⊂ X such
that Y ∩ U = Y ◦ . Then, i : Y ◦ ,→U is a closed imbedding of a smooth submanifold. Hence, by Kashiwara
theorem, we have M|U = i∗ L, where L is a finitely generated D-module L on Y reg . Moreover, Kashiwara’s
theorem says that SS L is contained in the zero section of T ∗ Y ◦ The result follows. 2

4.2. b-function.
j
Let f be a (global) regular function on X and U ,→ X be the open subset defined by X \ U = {f = 0}.
Let M be a DU -module. We introduce an auxilary variable s and consider X and U as varieties defined
over the field of rational functions K := C(s) (even though the equations defining X and U do not contain
s). Denote DU ⊗ K by DU (K) and similarly for DX(K) . The subrings with C[s] coefficients will be denoted
DU [s] and DX [s]. Let Mf s = {a(s) · m · f s }, be the space formed by all formal expressions a(s) · m · f s , where
a(s) is a rational function in s, m is a section of M and f s is a formal symbol. The O-module structure on
61
Mf s is defined by the action of OX on m ∈ M. We define the DX -module structure by letting a vector
field ξ act via the formula
(ξf )  s
ξ : mf s → ξ(mf s ) = (ξm) + s

m f .
f
This way, Mf s becomes a DU (K) -module.

4.2.1 Remark One can define an automorphism τ of DU [s] by

ξf
ξ 7→ ξ + s , φ 7→ φ,
f
where ξ is a vector field and φ is a function. Then it follows immediately from the definitions that Mf s = Mτ ,
i.e. Mτ is obtained from M by twisting the DU [s]-module structure with τ (this means that P ∈ DU [s] now
acts by τ (P )).
Notice that if M is a holonomic DU module, then Mf s is a holonomic DU (K) module. This follows
by extending filtrations by scalars. Notice also that if κ is an automorphism of C(s), then it induces an
automorphism of DX(K) modules, which does not affect holonomicity. We will apply this in particular to
the automorphism induced by s → s − 1.
Lemma 4.2.2 (Lemma on the b-function.) Suppose that M is a holonomic DU module. Then for any
section m ∈ M there exist u ∈ DX [s] and a polynomial b(s) ∈ C[s] satisfying
m
u(mf s ) = b(s − 1) f s . (4.2.3)
f

4.2.4 Remark
The meaning of the lemma is that if we view s as a complex number (not a variable) then m s
f f can be
s−1
written as mf and hence the Lemma says that we can achieve one extra f in the denominator applying
u and dividing by b(s − 1) (which is possible for all but finitely many values of s, the roots of the polynomial
b(s − 1)). Obviously, if a polynomial b(s) satisfies the property (4.2.3) of the theorem, then any multiple of
it satisfies the theorem too (since u can be also multiplied by a polynomial). This justifies the following
Definition 4.2.5 The generator of the (principal) ideal of all polynomials b(s) with the property (4.2.3)
is called the b-function.
Proof of the lemma on b-function.
We will prove a slightly more general statement. To that end, consider M as a DX -module and let
M0 ⊂ M be an OX -coherent subsheaf (rather than just submodule generated by one section). We will show
that the module N := DX(K) · M0 f s is a holonomic DX(K) module over the ground field K.
The strategy here is to notice that N |U is a holonomic DU (K) -module since it is a submodule of the
holonomic module Mf s , and then extend N |U to some holonomic DX(K) -module and finally compare this
extension with N itself.
To that end, denote by N 0 ⊂ N the lowest term of Gabber’s filtration, see REFERENCE. Of course,
this filtration was defined earlier only for affine varieties and to make sense of it in the general situation one
needs to prove that this filtration is compatible with localizations, i.e. Gj (M)|U = Gj (M|U ). The latter
property follows from the fact that Gabber’s filtration coincides with Sato-Kashiwara filtration. Since the
Sato-Kashiwara filtration is defined via truncations and other homological operations, it is compatible with
localization. One deduces further that N 0 6= 0 since N 0 |U = N |U .
Thus, we’ve obtained a holonomic submodule N 0 ⊂ N such that N 0 |U = N |U . Then N /N 0 is a DX(K) -
module supported on X \U . Therefore any particular section of N /N 0 is annihilated by a certain power of f .
Hence, by the Nullstelenzats, the image of M0 in N /N 0 is annihilated by a certain power of f , in particular
for some k f k · M0 f s ⊂ N 0 . Hence, DX(K) M0 f s+k ⊂ N 0 , and therefore DX(K) M0 f s+k is holonomic, as
62
a submodule of a holonomic D-module. But DX(K) M0 f s ' DX(K) M0 f s+k as DX(K) -modules after an
automorphism of K. Indeed, the map is P (s)mf s 7→ P (s + k)mf s+k and it is obvious that it it one-to-one
and onto, and the translation s → s + k makes the map a DX(K) morphism. Hence, DX(K) M0 f s is also
holonomic.
To deduce the original statement from the holonomicity note that any holonomic module has finite
length. Hence the chain of submodules
DX(K) (M0 f 1 ) ⊃ DX(K) (f · M0 f 1 ) ⊃ DX(K) (f 2 · M0 f 1 ) ⊃ . . .
necessarily stabilizes. Hence DX(K) (f r · M0 f 1 ) = DX(K) (f r+1 · M0 f 1 ) for some r. If M0 is now generated
by one section m, we have DX(K) (f r · mf s ) = DX(K) (f r+1 mf s ). But the LHS of this equality is isomorphic
to DX(K) (mf s ) after a field automorphism. Hence DX(K) (mf s ) = DX(K) (f · mf s ) and therefore there exists
a section u0 of DX(K) such that u0 (mf s+1 ) = mf s . But this u0 may have denominators involving s, i.e.
u0 = b(s)
u
where u and b depend on s polynomially. Then u · mf s+1 = b · mf s . 2

In the previous chapter we have constructed a functor

j∗ : DU − mod → DX − mod.
In general this functor does not map finitely generated modules to finitely generated modules: for example
j∗ (DU ) is not finitely generated over DX . However we have seen that sometimes the finite generation property
is preserved: for example we can take X = C, U = {z 6= 0} and consider j∗ (OU ) = C[z, z −1 ]. The theorem
that we state below says that it is the size of the characteristic variety that makes the difference.
Theorem 4.2.5.1. Let X and U be as above. Then j∗ maps holonomic DU -modules to holonomic DX -
modules. In particular, the image under j∗ of any holonomic DU -module is finitely generated.
Proof. M is a holonomic DU -module generated by an OX coherent module M0 . Then
DX(K) M0 f s = ∪k∈Z DX(K) M0 f s−k = DU (K) M0 f s = j∗ Mf s ,
where the first equality follows from the lemma on the b-function. Thus, j∗ Mf s is a holonomic DX(K)
module, again by the lemma on the b-function. We can assume that M0 is generated by a single section
m (othewise induct on the number of generators). Then DX(K) · mf s = DX(K) /I where I stands for the
annihilator ideal of (mf s ). Then the zero variety of gr I is a Lagrangian variety in T ∗ X (again viewed
over K). The ideal gr I is generated by finitely many (K-valued) functions on T ∗ X. If we specialize s to
an integer n, for all but finitely many values of n the complex dimension of the characteristic variety of
DX · mf n is the same as the K dimension of the characteristic variety of DX(K) · mf s = dim X, and in
addition the multiplicities of components are preserved. It follows immediately that if n  0 then DX · mf n
is holonomic. In addition, if n  0 then DX · mf n = DX · mf n−1 = DX · mf n−2 = . . . . Indeed, if not
DX · mf n is a proper holonomic submodule of DX · f m for some m < n, and the sum of the multiplicities of
DX · f m exceeds the sum of the multiplicities of DX(K) · mf s . It follows that j∗ M = DX · mf n for n  0,
so j∗ M is holonomic. 2

4.2.6 Another proof for X = Cn .

Now we will run through a proof of Theorem ?? in the case where our variety X is Cn . Let A = D(Cn ).
This ring is generated by elements x1 , . . . , xn , ∂1 , . . . , ∂n . Filter A by giving all the generators degree 1
(earlier we used a filtration where xi had degree 0 and ∂i degree 1). We still have
gr A = C[x1 , . . . , xn , ξ1 , . . . , ξn ]
but now all generators have degree 1. The advantage of this choice of degrees is that each homogeneous
piece of gr A is a finite dimensional vector space. This new filtration is called the Bernstein filtration.
Given an A-module M , consider M with a filtration such that
Ai · Mj ⊂ Mi+j .
63
As before (REFER TO CHAPTER 1.1) we have a notion of a good filtration. However, with such a filtration,
each piece of gr M is a finite dimensional vector space. Let h(gr M ) be the Hilbert polynomial of gr M ; so
that
h(gr M )(i) = dim(Mi /Mi−1 ) for i  0.
This is indepentdent of particular choice of a good filtration; so we just write h(M ). Now
c · id
h(M )(i) = + terms of degree ≤ d − 1
d!
where d = dim SS M .
Note that this characteristic variety SS M is defined using our particular filtration on A. It does not
coincide with the characteristic variety defined using the other filtration on A. Nevertheless, the dimensions
of these characteristic varieties do coincide. This follows from the result in Chapter 1 relating dim SS M to
vanishing of Ext groups, which is independent of filtration. This is done in 1.4. If so, we should probably
EXPLAIN this a little bit here ??). We use our filtration of A where deg xi = 1, and all notions like SS M
will be with regard to this filtration.
Proposition 4.2.7 For any A-module M we have dim SS M ≥ n.
Proof. The Hilbert poynomial h(A) of A looks like
i2n
h(A)(i) = + ...
(2n)!
Let 0 6= M0 ⊂ M1 ⊂ . . . , ∪Mi = M , be a good filtration of M .
Lemma 4.2.8 The natural map Ai ,→ Hom(Mi , M2i ) is injective for all i ≥ 0.
Proof. Use induction on i. The case i = 0 just means M 6= 0. Suppose we proved the statement for values
les then i. Proceed by contradiction. Suppose P ∈ Ai and P · Mi = 0. Note that it follows that each [P, xj ]
and [P, ∂j ] act by zero on Mi−1 . We may assume that P is not a scalar. Then for some j either [P, xj ] 6= 0
or [P, ∂j ] 6= 0. If [P, xj ] 6= 0 then [P, xj ] is a non-trivial element of Ai−1 that acts by zero on Mi−1 , which
cannot happen by the inductive assumption. Similarly, [P, ∂j ] 6= 0 also leads to a contradiction. 2
End of proof of the Proposition. The lemma above implies that h(A)(i) ≤ H(M )(i)h(M )(2i). Since h(A)
has degree 2n, we conclude that h(M ) has degree ≥ n. 2
Corollary 4.2.9 Any holonomic module has finite length
n
Proof. Since h(M ) = c · in! + . . . where c is a positive integer, and since the Hilbert polynomial is additive
on short exact sequences, we immediately get length M ≤ c. 2

Theorem 4.2.9.2. If M is a holonomic A-module and f ∈ C[x1 , . . . , xn ] then M [f −1 ] is a holonomic

A-module.
Proof. Let M0 ⊂ M1 ⊂ . . . be a good filtration on M . Write N = M [f −1 ]. Let d = deg f . Define
Nj = {mf −j | m ∈ M(d+1)j }.
We will not try to prove that this filtration is good. However, one can check by direct computation that
h(gr N ) has degree n.
To prove that N is holonomic, take a finitely generated A-submodule N 0 of N and choose a good filtration
{Nj } on N . There exists m0 large enough such that Nj0 ⊂ Nj+m0 for all j. Hence deg h(N 0 ) = n, so N 0 is
0

holonomic. Moreover, the leading coefficient is bounded by that of h(gr N ). Hence any increasing sequence
of submodules finitely generated over A terminates. Since we proved that any such submodule is holonomic,
N itself is holonomic. 2
64
 4.3. Functoriality for holonomic complexes.
Definition 4.3.1
For any smooth algebraic variety X of dimension n we define the Verdier duality functor by
D : M 7→ RHomDX (M, DX )(n)
(i.e. we extend the algebraic definiton we used before to the setup of coherend DX -modules).

4.3.2 Remark By definition, for any left DX module M the Verdier dual DM is a complex of right
DX -modules. However, by tensoring with Ω−1 X as usual we can think of DM as a complex of left DX -modules.
Apriori the complex DM can be infinite. This would make using the duality functor very difficult. We
will show now that this is not the case.
b
Recall (??) that we denote by Dcoh (DX ) the derived category of all bounded complexes of DX -modules
with DX -coherent cohomology.
b
Proposition 4.3.3 The Verdier duality functor D maps Dcoh (DX ) to itself. In particular, for any
DX -module M the Verideir dual D M is a bounded complex.
Proof. It suffices to prove the statement for a signle DX -module M (the proof will extend to the general
case easily). If M = DX we have D(M) = Ω−1 ⊗OX DX and the statement is true. For a general DX -module
M we consider a resolution (possibly infinite) with free DX -modules:
. . . → M2 → M1 → M0 → M → 0.
As in the commutative case we prove that the kernel of the map M2n−1 → M2n−2 is a projective DX -module
(this follows from the vanishing of non-commutative Ext-groups which can be deduced from the vanishing of
commutative Ext-groups for the corresponding graded objects). Note that the statement of the proposition
is local in nature. Hence, once we know that it’s true for free modules it has to be true for projective modules.
If we know it is true for projective modules, we can deduce it for any module using the resolution above and
long exact sequences of RHom’s. 2
Note that by our computations (reference !) of Ext-groups we know that ExtiDX (M, DX ) vanishes if i > n
or i < codim SS M. Therefore if M is holonomic, only one Ext group can survive, namely ExtnDX (M, DX ).
This can be expressed in the following
Corollary 4.3.4 If M is holonomic then
(i) the complex RHom(M, DX ) is quasi-isomorphic to the single object ExtnDX (M, DX ),
(ii) ExtnDX (M, DX ) is a holonomic DX -module,
(iii) D defines a contravariant exact functor D : Hol(DX ) → Hol(DX ) satisfyingD ◦ D = Id. Proof.

The fact that ExtnDX (M, DX ) is holonomic can be proved using the estimates on dim SS ExtnDX (M, DX )
proved in Chapter 1 (REFERENCE !). 2
4.3.5 Example: duality and local systmes. For any pair of DX -modules M nad N one can defined
the O-tensor product M ⊗DX N with the DX -module structure defined by the Leibniz rule ξ(m ⊗ n) =
ξ(m) ⊗ n + m ⊗ ξ(n). In general, this product is not considered, since it may not even be coherent even for
N = DX . However, it is immediate that if N = L is a local system then the O-tensor product is indeed a
coherent DX -module.
Proposition 4.3.6 Let L is a local system and M is a holonomic DX -module then
(i) The tensor product L ⊗OX M is a holonomic DX -module,
(ii) One has D(L ⊗OX M) ' L∗ ⊗OX D(M) where L∗ is the dual in the sense of local systems. In
particular, D(L) ' L∗ . 2
4.3.7 Digression: on derived categories of holonomic modules.
65
 b
Recall that Dcoh (DX ) denotes the derived category of bounded complexes of DX -modules with DX -
coherent cohomology. Inside it one has an abelian subcategory Coh(DX ) formed by all DX -coherent modules.
b
One might ask the following question: is it true that Dcoh (DX ) is equivalent to the bounded derived category
of Coh(DX )? The answer to this question is not known (and it might be “no”).
However, if we consider the holonomic version of this problem, then the answer is known to be “yes”.
b
More precisely, consider the bounded derived category Dhol (DX ) of complexes of DX -modules with holonomic
cohomology and its abelian subcategory Hol(DX ) formed by all holonomic DX -modules. Then one has the
following
Theorem 4.3.7.3. There exists a natural equivalence of categories
Db (Hol(DX )) → Dhol
b
(DX ).
This theorem will be proved later and now we will explain how its proof can be reduced to a statement
about objects of Hol(DX ).
4.3.8 Devissage.
Recall that for any complex A (of sheaves, DX -modules, etc.) and any integer k we defined the truncated
complex τk≥ A (GIVE ITS CONSTRUCTION) together with the morphism A → τk≥ A. The crucial feature
of τk≤ is that it defines an operation of the derived category of compexes.
Lemma 4.3.9 Suppose that H i A = 0 for i > k. Then the natural morphism A → τk≥ A induces a
qis
quasi-isomorphism H k A −→ τk≥ A. 2
Now the devissage procedure says that if we want to prove some property for the derived category of
complexes, it suffice to show that
(a) The property holds for indvidual objects(complexes concentrated in degree zero)
(b) The property respects the distingiushed triangles, i.e. if any two complexes in a distinguished triangle

A _@ /B
@@ [1] ~
@@ ~~
@@ ~~~
~
C
satisfy the property then the third complex also satisfies it.
In fact, in (b) we can take B = τk≥ A, C = Cone(A → τk≥ A) and this allows us to prove the statement
by induction on the degree of the top non-zero cohomology of A.
One useful application of devissage is the following theorem
b b
Theorem 4.3.9.4. Let U be an open subset of X. If M ∈ Dhol (DU ) then j∗ M ∈ Dhol (DX ).
Proof. By the first part of devissage it suffices to prove the statement for DU -modules. We will prove
later that if j : U ,→ X is affine, then the statement follows from the theorem on b-function. If j is not affine
then we consider an open affine cover U = ∪Ui where each open subset Ui is given by non-vanishing of a
regular function fi . We assume, for simplicity, that U = U1 ∪ U2 (the general case follows from this easily).
Then one has a long exact sequence
0 → j∗ M → (j1 )∗ M|U1 ⊕ (j2 )∗ M|U2 → (j12 )∗ M|U1 ∩U2 → . . .
the second part of the devissage is obvious. 2

4.3.10 A criterion of holonomicity.

66
 i
Suppose that Y = X \U ,→ X is a smooth subvariety. Then for any DX -module M one has the following
distinguished triangle:
i∗ i! MdI /M
II xx
II x
II xx
[1] II xxx
{x
j∗ M|U

Since by the theorem above we know that j∗ M|U is holonomic, the module i∗ i! , M is also holonomic. But
holonimicity of i∗ i! M is equivalent to the holonomicity of i! M . In fact, one has exact sequence 0 → TY∗ X →
π
T ∗ X|Y −→ T ∗ Y → 0 and equality SS i∗ i! M = π −1 (SS i! M ) and the assertion now follows by counting
dimensions. This proves the following
Corollary 4.3.11 For any closed embedding i : Y ,→ X and any holonomic module M the DY -module
i! M is also holonomic.
These results allow us to prove the following criterion of holonomicity:
b
Theorem 4.3.11.5. Let M ∈ Dcoh (DX ). Then the following are equivalent:
b
(i) M ∈ Dhol (DX ),
(ii) For any point ix : x ,→ X, the complex i!x M has finite dimensional cohomology.

Proof. The fact that (i) implies (ii) is an easy consequence of the previous corollary. We prove that (ii)
implies (i) by induction on dim SuppM. Let SuppM = Y ,→ X. Let Y reg ⊂ Y be the subset of regular
points of Y . We can choose an open subset U ⊂ X such that U ∩ Y = Y reg . We can replace X by U to
get to the situation in which Kashiwara’s theorem applies. By this theorem one has M|U = i∗ N for some
D-module N on Y reg . Since for a closed embedding the functors i+ and i! coincide up to shift in derived
category, we deducde that i!x N is finite-dimensional for any x ∈ Y reg .
Choose a goood filtration on N . The sheaf gr N is a coherent sheaf on T ∗ Y reg . By a basic result from
algebraic geometry (reference) we can choose a smaller open set U 0 ⊂ U such that gr N becomes flat when
restricted to Y 0 = Y ∩ U 0 . In fact, first by shrinking U we acheive that gr N is flat over T ∗ Y 0 . This will
imply that each graded component of it is finitely generated and flat (hence projective) over Y 0 . Hence the
first term N 0 of the filtration on N is projective. Since any extension of two projective modules is again
projective, we deduce that N 1 is projective. Continuing in this manner we obtain that all N i are flat. Hence
N , being a direct limit of projective modules, is flat over Y 0 . Then for any x ∈ Y 0 the complex i!x (N ) is
in fact one object i+ ! +
x N concentrated in some degree (since ix is equal up to shift to ix and higher derived
+
functors of ix all vanish by flatness). Hence N |Y 0 is a flat sheave with finite-dimensional geometric fibers.
This implies that N |Y 0 is a local system, hence holonomic.
We have obtained a locally closed embedding i0 : Y 0 ,→ X such that (i0 )! M is holonomic. Now we deduce
that (i0 )∗ (i0 )! M is holonomic. This follows from factoring i0 into a composition of an open embedding and
j0
a closed embedding and applying the holonomicity theorems above. Let V 0 = X \ Y 0 ,→ X. Then the cone
of the morphism (i0 )∗ (i0 )! M → M is isomorphic to j∗0 M|V 0 which has support of strictly smaller dimension.
So, by the inductive assumption j∗0 M|V 0 is holonomic, hence M itself is holonomic. 2

4.3.12 Remark In terms of Corollary (4.1), one can describe the locally closed subvariety Y 0 as follows.
Decompose SS M as TY∗reg X ∪ {other components}. Then Y 0 = Y reg \ {images of the other components}.
4.3.13 The functors f ∗ and f! .
Recall that earlier we have defined for any map f : Y → X of smooth algebraic varieties the functors f∗
and f ! on D-modules.
67
 In the special case of holonomic modules we can define another pair of functors, f ∗ , f! , using Verdier
duality functor:
f ! := D ◦ f ! ◦ D; f! = D ◦ f∗ ◦ D.
We want to emphasize that we assume that the domain of f ! is Dholb b
(DX ) and the domain of f! is Dhol (DY ),
since in general the Verdier duality functor is ill-behaved.
There seems to be no nicer formula for f ∗ and f! and all the properties of these functors are established
using Verdier duality.
4.3.14 Example Let j : U ,→ X be an affine open embedding (i.e. Y = X \ U is of codimension
1. Then j∗ is exact. We will prove in the next section that for any DU -module M, the module j∗ M has
no submodules supported on Y . By duality one deduces that j! M has no quotient modules supported on
Y . Similarly, if i : Y ,→ X is the closed embedding and M is a DX -module, by Kashiwara’s theorem we
deduce that i∗ i∗ M is the maximal quotient module of M which is supported on Y . Notice that the notion
of “maximal quotient module” is well defined only for holonomic modules due to their artinian properties
(while, of course, the notion of “maximal submodule” makes sense for any Noetherian module).

4.4. Diagram Y ,→ X ←- U and the DGM-extension.

4.4.1 Direct image from U .
i j
Let Y ,→ X be a smooth closed submanifold and U = X \ Y ,→ X be its complement.
Proposition 4.4.2 For any DU module N one has i! j∗ N = 0. Proof. This is a consequence or (??). In
fact, since i∗ is fully faithful, it suffices to prove that i∗ i! j∗ N = 0. By (??) this amounts to showing that
Γ[Y ] j∗ N = 0 which follows from definitions. 2
For any holonomic DX -module M one has a canonical map M → j∗ (M|U ). Hence by duality and
definition of j! one obtains a map j! (M|U ) → M. If we take M of the form j∗ N for some holonomic
DU -module N , then one gets a canonical map j! N → j∗ N . It follows immediately that the kernel and
cokernel of it are supported on Y .

j
Definition 4.4.3 For any holonomic module N on a locally closed subvariety Y ,→ X we define the
Deligne-Goresky-MacPerson extension (or minimal extension) j!∗ N of N by
j!∗ N = Image (H0 j! N → H0 j∗ N ).
The basic properties of the DGM-extension are summarized in the following
Proposition 4.4.4 Let N be a holonomic module on Y and j!∗ N its minimal extension. Then
(i) j ! (j!∗ N ) = N ,
(ii) j!∗ N has neither submodules nor quotient modules supported on
Y \Y,
(iii) Taking the minimal extension commutes with Verdier duality:
j!∗ (DN ) = D(j!∗ N ).

Proof. To show (i) take an open set U ⊂ Y such that U ∩ Y = Y and apply Kashiwara’s theorem to U
and Y . Part (iii) follows by self-duality of the map j! N → j∗ N . To prove (ii) note that j!∗ N ⊂ j∗ N has
no submodules supported on Y \ Y . The statement about quotient modules follows by duality. 2
4.4.5 Warning. It might well be the case that j!∗ N has a filtration with submodules such that one of
the intermediate terms is supported on the boundary.

68
 j
Theorem 4.4.5.6. Let Y ,→ X be a locally closed embedding of smooth varieties and L be an irreducible
local system on Y . Then j!∗ L is a simple holonomic DX -module. Moreover, any simple holonomic DX -
module M arises from an appropriate pair (Y, L).

Proof. Let N ⊂ M = j!∗ L be a submodule. By the proposition above N cannot be supported on Y \ Y .

Hence 0 6= j ! N ⊂ j ! M = L and this implies j ! N = j ! M. Therefore N = M.
To prove the second part notice that in fact the argument for the first works backwards too: as in the
j
previous section we can find a locally closed subvariety Y ,→ X such that j ! M is a local system. By a similar
argument one showd that the minimal extensionof this system coincides with M. Finally, the irreducibility
of this local system follows from simplicity on M. 2

4.4.6 Remark Suppose that Y is a locally closed submanifold of X and Y 0 is an open subset of Y .
Since all the varieties are complex, the fundamental group of Y 0 maps surjectively. Hence the restriction
any irreducible local system L on Y to Y 0 is still irreducible. Moreover, one can show that the minimal
extension of L|Y 0 coincides with the minimal extension of L. Hence we can always assume that Y \ Y is of
codimension 1 in Y (if not, we can just shrink Y ).

From now on we will often consider the diagram

j i
U ,→ X ,→ Y = f −1 (0)

where Y is a divisor given (locally) by an equation {f = 0}. Let M be a holonomic DU -module. Choose an
OX -coherent subsheaf M0 ⊂ j· M such that DU · M0 = M.

Proposition 4.4.7 For any k  0 one has j!∗ M = DX · (f k M0 ).

Proof. The module j∗ M/j!∗ M is supported on the divisor Y hence all sections of this quotient are
annihilated by f k if k  0. This implies that DX (f k M0 ) ⊂ j!∗ M. Recall that j!∗ M is holonomic, in
particular, artinian. We deduce that the descending chain of submodules DX (f k M0 ) ⊇ DX (f k+1 M0 ) ⊇ . . . ,
stabilizes as k is increasing. The result follows. 2
b b
Recall that for a singular Y we defined Dhol (DY ) to be the full subcategory of Dhol (DX ) generated by
holonomic modules supported on Y .
b
Lemma 4.4.8 For any M ∈ Hol(DU ) the following objects are quasi-isomorphic in Dhol (DY ):
qis qis
Cone(j! M → j∗ M) ∼ i! j! M ∼ i∗ j∗ M[−1??]

Moreover, H−1 i! j! M = H0 i∗ j∗ M and H0 i! j! M = H1 i∗ j∗ M. Proof. One has the following self-dual

diagram
0 → i! j! M → j! M → j∗ M → i∗ j∗ M → 0
(notice that since Y is a divisor, j! M, etc. is in fact just one object rather than a complex). Here i∗ j∗ M is
the maximal quotient module supported on the boundary, and i! j! M is obtained by duality. Furthermore,
for any DX -module N one has the following exact sequence:

0 → H0 i! N → N → j∗ N → H1 i! N → 0

If we take now N = j! M and look at the cokernel, we will get H0 i∗ j∗ M ' H1 i! j! M. The other equality is
proved similarly. 2
If Y is a divisor we can summarize all the information about the objects involved in the following diagram
(I’ll straighten it later)
69
 0


H1 i∗ MX 90 H−1 i! MX
11 KK
K r rr ss9 F
11 KK K r s
K r rr sss
11 KK rr ss
11 KK
KK rr sss
11 KK s s
ss
M
t9 LLL
j
KK !∗
11 KK
K ttt LLL sss
11 KK t LLL ss
KK tt ss
11 % tt L& ss
11 j! M j∗ M
11 tt: JJ
J r rr9 KKK
KKK
11 t tt JJ
J rr r KKK
1 tttt JJ r
J$ rrr K%
H0 i! j! M i∗ j∗ M
u: KK
MX
r9 JJ K t 9
r J u KK tt
r rr JJ
J uuu K K tt
rr JJ uu KK tt
rr
rr J$ uu K% ttt
0 ∗
i! MX
: JJ r9
0 H i MX
ttt JJ rr r
t JJJ r
tt JJ rr
tt $ rrr
tt 0 !∗
0 H i M

where i!∗ M = {Image of i! M in i∗ M}, MX stands for a DX -module and M stands for its restriction to U .
In general (i.e. when j is not affine) we just have a three-dimensional octahedral diagram in the derived
category

MX
 ? 00 IIII
 00 I
 00 III
 00 II
 II
 00 II
 00 II
 II
 0 II
 00 II
  0 II
00 II
  00 II
  II
  0 $
 ∗ 0 0 /
 i? ! MX
i i M X 00 i

∗ X
r
∗9 ~
 00 rr ~
 00 r rrrr~~~
y  rr ~~
j! j ∗ MX / j∗ j ∗ MX ~~~
dHH ~
HH
HH  ~~~
HH  ~~
HH
HH  ~~~
HH  ~
HH  ~~~
HH  ~
HH
HH  ~~~
HH  ~~
HH
HH  ~~~
HH ~
~~
i∗ i∗ j∗ j ∗ MX

(To check the directions of the arrows) The equivalence of the two diagrams in the case when j is affine
is a non-trivial fact and we express it as a theorem
70
 Theorem 4.4.8.7. If Y is a divisor (i.e. if j is affine) then the octahedral diagram is equivalent to the
plane diagram above.

4.4.9 Remark The octahedral diagram above has a close relation to the octahedral diagram involved
in the definition of a derived category. The latter is usually represented as two “hats” (upper and lower):

A0 Ao C A0 `Bo C
AA [1] ? O BB [1] ~ O
AA  BB ⊕ ~~~
AA dist  BB ~
A  B ~~~~
⊕ B _? ⊕ dist B 0 dist
~~ ??? [1] |> AA
~~ ? ||| ⊕ AAA[1]
~~ dist ??? || AA
A
 ~~~  ||
C0 /A C0 /A

The condition imposed on this diagram is that all triangles marked by dist are distinguished triangles, all
triangles marked by ⊕ commute and the two maps from B to B 0 (via C and C 0 ) coincide (these three
conditions imply ?? that the two possible maps from B 0 to B also coincide). The octahedron axiom of
the derived category says that any upper hat can be completed by a lower hat, and any lower hat can
be completed by an upper hat. This axiom originates from the Ore conditions which have to be satisfied
by quasi-isomorphisms when the derived category of complexes (of abelian groups, say) is obtained as a
localization of the homotopy category of complexes.

4.6. Vanishing cycles and f s .

Let C ∗ be the category of holonomic D-modules on C∗ such that all simple subquotients are isomorphic
to Γ(OC∗ ) = C[t, t−1 ] = R.
Lemma 4.6.1 The category C ∗ is naturally equivalent to the category of all finite dimensional vectors
spaces V endowed with a nilpotent operator s : V → V .
Proof. Any object M of C ∗ is necessarily a local system on C∗ . Given such M, put V = M|{1} and let
u be the monodromy operator. Since all the subquotients have trivial monodromy, u has to be unipotent.
Hence the operator s = log u is nilpotent. 2

The lemma above allows us to identify the indecomposable objects of C ∗ . In fact, by Jordan normal
form any indecomposable object of the category of vector spaces with a nilpotent operator, is given by a
nilpotent Jordan block of the size n. One can easily check that this block corresponds to the D-module
Mn = R · logn−1 t + R · log n−2 t + . . . + R (with the monodromy operator u given by exp(t∂t )).
We can also see from the lemma above that C ∗ has no projective modules. We can consider a pro-object
 
O
 O 
E proj = lim Mn ;  
 . 
←
.
and an ind-object
 
.
 . 
E ind = lim Mn ;  
 O 
→
O
71
 Now let C be the category of all D-modules M on C such that M|C∗ ∈ C ∗ . In particular, one has a simple
object δ = i∗ C where i : {0} ,→ C. Another simple object of C can be obtained as a minimal extension of a
DC∗ -module R to C. In fact, one can see that
       
O δ O δ
j! R ; j∗ R ; , & ,
δ O δ O
and the map j! R → j∗ R mods out the bottom of j! R and maps the quotient O isomprhically onto the bottom
of j∗ R. Hence j!∗ R = O.
From now on we will denote Mn−1 by Log n−1 . Then one has j! Log n−1 = DC /DC · (t∂)n , j∗ Log n−1 =
DC /DC · (∂t)n hence
   
δ O
 O   δ 
   
.   . 
n−1
; n−1
;
 
j∗ Log 
, j! Log  
 .   . 

 δ   O 
O δ
where in both cases on has n pairs in the columns. Notice that Dj! Log n−1 = j∗ Log n−1 . We introduce the
following pro- and ind-objects:

∇proj = j! E proj , ∆proj = j∗ E proj , ∆ind = j∗ E ind , ∇ind = j! E ind .

Lemma 4.6.2 ∇proj and ∆proj are indecomposable and projective. 2

Note that there exist two surjective maps ∇proj → O → 0 and δ proj → δ → 0.
Lemma 4.6.3 The category C is equivalent to the category of pairs of vector spaces V0 , V1 together with
two linear maps u : V0 → V1 , v : V1 → V0 satisfying (uv)n = 0 for some n.
Proof. We define the equivalence by putting
M 7→ (V0 = HomC (∆proj , M ), V1 = HomC (∇proj , M )),
where the maps u and v are induced by the natural maps δ proj  ∇proj . Note that V1 is nothing but the
fiber of M at 1. 2

4.6.4 Notation For any object M of C we denote HomC (∆proj , M ) by Φ(M) (the functor of vanishing
cycles) and HomC (∇proj , M )) by Ψ(M) (the functor of nearby cycles).
We have a canonical residue pairing C[[s]] × C((s))/C[[s]] −→ C given by taking the residue f × g 7−→
Res s=0 (f ·g). It is a perfect pairing making C((s))/C[[s]] the continuous dual of C[[s]], viewed as a topological
C-vector space equipped with s-adic topology.
Lemma 4.6.5
R ts [s]
Log n−1 = .
sn R ts [s]
Xn
Proof. Define a map R ts [s] → Log n−1 as follows. Let ai si ts be an element of R ts [s]. Put formally
i=0
∞
X sk · log(kt)
ts = es log t = . So, we send
k!
k=0
n n ∞
X 1 X
X sk log(kt)

ai si ts 7→ Ress=0 ai si .
i=0
s i=0
k!
k=0

72
This induces a bijection
R ts [s]
Log n−1 = . 2
sn R ts [s]
Using this lemma we can make the following key observation:
Rts [s]
E proj = Rts [[s]] = lim .
← sn Rts [s]

Moreover, under this isomorphism the logarithm of monodromy corresponds to s. When we take the direct
image with respect to j we obtain the following diagram of objects on C
Rts (ss)
∆proj = j∗ E proj = Rts [[s]] ,→ Rts ((s))  = ∆ind ,
Rts [[s]]
where the middle term is a pro-ind-object (self-dual with respect to s 7→ s−1 ) and the last equality is proved
by using the correspondence
∞
X 1 X
X sk log(kt)

aj si ts 7→ Ress=0 ai si
s k!
i∈Z k=0
Rts ((s))
and noting that Rts [[s]] ' R[log t] = ∆ind .
Corollary 4.6.6 The following sequences are exact:
0 → j! E proj → j∗ E proj → δ → 0,
0 → δ → j! E ind → j∗ E ind → 0 2

Let f : X → C be a regular function such that df never vanishes outside Y = f −1 (0):

j
U =X \Y ,→ X
↓ ↓f
C∗ ,→ C
Define Efproj
= f + E proj . (So, formally one has ts 7→ f s , log t 7→ log f .)
For any holonomic DU -module M, we have
D(Efproj ⊗OU M) = Efind ⊗OU DM.

Proposition 4.6.7 In our pro-ind category

j! Mf s ((s)) ' j∗ Mf s ((s)) .

Proof. Recall that

j!∗ Mf s ((s)) = Im j! Mf s ((s)) → j∗ Mf s ((s))

 

and this module can be computed as

j!∗ N = DX (f k N0 ) ⊂ j∗ N ,
where N0 is an O-coherent submodule of N and k  0. Assume for simplicity that M = DX · m where m
is a section of M (the general case follows from this by induction on the number of generators). Then
j!∗ Mf s ((s)) = DX mf k f s ((s)) = DX mf k+s ((s)) = Mf s ((s))

by the lemma on b-function. Hence our map is surjective. By duality it is also injective, therefore it is an
isomorphism. 2
73
 Corollary 4.6.8 The canonical map
j! Mf s [[s]] → j∗ Mf s [[s]]
is injective. 2
Corollary 4.6.9 If M is generated by an OX -coherent subsheaf M0 , then

(i) j! Mf s [[s]] = DX [[s]] · (f k+s M0 ), k0

(ii) j∗ Mf s [[s]] = DX [[s]] · (f ks M0 ), k  0.
Proof. j! Mf s [[s]] = j!∗ Mf s [[s]] = DX [[s]] · (f k+s M0 ). The second assertion is proved similarly. 2
Corollary 4.6.10 For k  0 one has the follwoing ismorphisms
DX [s](f s+k M0 ) DX [s](f s−k M0 )
j! M = , j∗ M = .
s · DX [s](f s+k M0 ) s · DX [s](f s−k M0 )
Proof. We have the following short exact sequance on U
s·
0 → Mf s [[s]] −→ f s M[[s]] → M → 0.
By exactness of j! this gives
s·
0 → j! Mf s [[s]] −→ j! f s M[[s]] → j! M → 0,
and the result for j! follows. The statment for j∗ is proved similarly. 2

Definition 4.6.11 For any DU -module M we define a submodule

j∗ Mf s [[s]]
Ψ(M) =
j! Mf s [[s]]
supported on the divisor Y = f −1 (0).
Proposition 4.6.12 Ψ(M) is a holonomic module.
Proof. It suffices to show that, for some N  0,
sN j∗ Mf s [[s]] ⊂ j! Mf s [[s]]
(by the previous corollary). Note that:
sN DX [[s]](f s−k M0 ) ⊂ DX [[s]](f s+k M0 ).

Assume for simplicity that M0 is generated by one section m. By the lemma on b-function, there exists
an operator u(∂, s) such that u(∂, s)(m · f s+1 ) = b(s) · (mf s ). Then u(∂, s + k − i)(m · f s+1+k−i ) =
b(s + k − i)m · f s+k−i . Hence we can find an operator Q such that
Q(∂, s)(m · f s+k ) = b(s + k − 1) · b(s + k − 2) . . . b(s − k)(m · f s−k ).
Now take N to be equal to the number of integral roots of b-finction. (Since s − α is invertible in C[[s]].)
(That’s was someone’s notes and I don’t understand this last part). 2

Corollary 4.6.13 The functor Ψ : Hol(DU ) → Hol(DY ) is exact.

Proof. The functor Dhol (DU ) → Dhol (DY ) preserves the abelian core of these derived categories. (same
comment as before). 2

Corollary 4.6.14 Ψ(DM) = D(ΨM).

Proof. One has the following short exact sequence
j! Mf s ((s))
0 → j! Mf s [[s]] → j! Mf s ((s)) → →0
j! Mf s [[s]]
74
 j! Mf s ((s))
where the quotient j! Mf s [[s]] is isomorphic to j! (M ⊗ Efind ). Now

D Ψ D(M) = D i∗ j∗ (DM ⊗ Efproj ) =

= D i∗ j∗ D (M ⊗ Efind ) = i! j! (M ⊗ Efind ) = Ψ(M). 2

4.6.15 Remark From now on all the arguments will makes sense not only in the category of D-modules,
but also in the category of D–,odules with regular singularities, mixed Hodge modules, perverse sheaves, etc.
Beilinson in his (reference) introduced another functor that we are about to define.
Definition 4.6.16 For any DU -module M define
j∗ Mf s [[s]]
Ξ(M) =
j! Mf s s[[s]]
The functor Ξ is called the maximal extension (and sometimes the “double tail”)
Mf s [[s]]
It follows from the definition that Ξ(M) has a submodule jj!!Mf s s[[s]] with the quotient module isomorphic

to Ψ(M):
α− β−
0 → j! M −→ Ξ(M) −→ Ψ(M) → 0. (4.6.17)
s
Mf s[[s]]
We can also consider a submodule jj∗! Mf s s[[s]] which is itself isomorphic to Ψ(M) since mulitplication

by s is an isomorphism. This leads to a short exact sequence

β+ α+
0 → Ψ(M) −→ Ξ(M) −→ j∗ M → 0. (4.6.18)

Lemma 4.6.19 The functor Ξ has the following properties:

(i) Ξ is exact,
(ii) Ξ commutes with Verdier duality,
(iii) The two exact sequences (4.6.17) and (4.6.18) get interchanged by the Verdier duality,
(iv) The composition α+ ◦ α− : j! M → j∗ M coincides with the canonical map form j! M to j∗ M,
(v) The composition β− ◦ β+ : Ψ(M) → Ψ(M) is multiplication by s.
Proof. The proof of exactness is the same as for Ψ. The rest was omitted. 2
4.6.20 Example Condider a D-module on the punctured affine line C \ {0} given by the ring of functons
R = C[x, x−1 ]. One has  
δ
Ξ(R) ;  O  so called ‘two-sided δ-tail’.
δ
In this case we can see that Ξ(R) is indeed self-dual and also that this extension is maximal (since we know
all the indecomposable objects, it follows that the only way the make Ξ(R) bigger is to add at least one
subquotient isomorphic to O which would change the restriction to the open part C \ {0}).

Let MX ∈ DX and MU = MX |U . Consider the following complex

(α− ⊕γ− ) (α+ ⊕−γ+ )
j! MU / Ξ(MU ) ⊕ MX / j∗ MU

where the γ’s are the canonical maps. It follows immediately that the first map is injective and the second
map is surjective.
Definition 4.6.21

Ker α+ ⊕ −γ+
Φ(MX ) =
.
Im α− ⊕ γ−
75
Lemma 4.6.22 The module Φ(MX ) is supported on Y = f −1 (0).
Proof. On the open part U the complex above restricts to
0 → MU → MU ⊕ MU → MU → 0. 2

Proposition 4.6.23 The functor Φ has the following properties

(i) Φ takes D-modules to D-modules (i.e. not just to complexes of D-modules),
(ii) Φ commutes with the Verdier Duality functor D,
(iii) Φ is exact.
Proof. The first assertion follows from the definition, the second from self-duality of the complex defining
Φ and the third is proved by diagram chase. 2
4.6.24 Remark One can also give alternative definitions of the functors Ψ and Φ:
Ψ(MX ) = i∗ (∆ind ⊗OX MX ), Φ(MX ) = i∗ (∇ind ⊗OX MX ).
These definitions will not be used on these notes.

Note that the maps (β+ , 0) : Ψ(MU ) → Ξ(MU ) ⊕ MX and (−β− ) : Ξ(MU ) ⊕ MX → Ψ(MU ) descend
to a pair of maps u : Ψ(MU ) → Φ(MX ) and v : Φ(MX ) → Ψ(MU ), respectively. Moreover, the composition
v ◦ u = β− ◦ β+ is equal to the monodromy map s. We will also use the following exact sequence:
0 → i! MX → Φ(MX ) → Ψ(MU ) → H1 i! MX → 0,
and the one obtained from it by Verdier duality:
0 → i∗ MX → Ψ(MU ) → Φ(MX ) → H1 i∗ MX → 0.

Now we apply this formalism to the following situation:

i j
4.6.25 Problem. Let Y = f −1 (0) ,→ X ←- U be our usual diagram and suppose that Y is a smooth
divisor. One can ask the following question: how can we glue the category Hol(DX ) from the two categories
Hol(DU ) and Hol(DY )?
The answer to this question in the case of sheaves is that the gluing conditin is given by the specialization
map. This data, however, is not “finite” (i.e. cannot be expressed in the language of finite-dimensional vector
spaces) since the category of all sheaves on X is not artinian (IS THERE A BETTER EXPLANATION ?).
The problem for D-modules was solved independently by Deligne, Verdier and Beilinson (and all the
constructions can be also applied to perverse sheaves, mixed Hodge modules, etc.) Here we will give the
construction due to Beilinson. One advantage of it is that it works even when Y is singular.

4.6.26 Gluing data. Consider the quadruples (MU , MY , u, v) where u : Ψ(MU ) → MY and v :
MY → Ψ(MU ) are two morphisms of DY -modules satisfying (u ◦ v)n1 = 0, (v ◦ u)n2 = 0, for n1 , n2  0.

4.6.27 Remark If M is an object of Hol(DX ) then MU = M |U, MY = Φ(M) and the morphisms u, v
described above form a quadriple F (M) = (MU , Φ(M), u, v) satisfying the conditons imposed on gluing
data.

Definition 4.6.28 Define the gluing category to be the category with objects being the gluing data
quadruples and obvious morphisms.
Theorem 4.6.28.1. The category Hol(DX ) is equivalent to the gluing category.
76
 Proof. The remark above provides a functor F in one direction. Suppose a quadruple (MU , MY , u, v) is
given. Consider
(β+ ⊕u) (β− ⊕−v)
Ψ(MU ) / Ξ(MU ) ⊕ MY / Ψ(MU )
One can see that the first arrow is injective (since β+ is injective) and the second arrow is surjective (since
β− is surjective). Define

Ker β− ⊕ −v
G(MU , MY , u, v) =
.
Im β+ ⊕ u
By direct diagram chase one shows that G defines an exact functor which is inverse to the functor F . 2

4.6.29 Remark The functor Ξ(M) was introduced exactly because it is necessary to prove the theorem
above kosnoyazychno poluchilos’.
4.6.30 Proof of (4.3.7.3).
(the second proof uses functor Ξ. Maybe one could give the first proof right after the statement of the
theorem and then give the second one much later).
To formulate a general statement that implies the theorem, recall the definition of a t-structure on a
triangulated category D. Let h : D → Ab be a functor from D to some abelian category Ab (one should
think of h as a zero homology functor). For any i ∈ Z, define the functors hi by hi (M ) = h(M [i]), where M
is an object of D and [i] stands for shift in a triangulated category. We will say that h is homological if it
M → {hi (M )}i∈Z maps distinguished triangles to long exact sequences. Given such h, we can define
D+ = {M | hi (M ) = 0, ∀i < 0}, and
− i
D = {M | h (M ) = 0, ∀i > 0}.
Then we have an ”exact sequence of categories”: ‘0 → D− → D → D+ → 0.’ Such a triple (D, D+ , D− ) is
called a t-structure (one can show that any t-structure in the sense of the usual definition (cf. ???) comes
from some homological functor h, so this definiton is equivalent to the one of (cf. ???)).
One defines a heart of t-structure to be C = D+ ∩ D− . It follows that C is an abelian category and
there exists a functor Db (C) → D (one should says smth abou iterated cones, and also that the axioms of
the derived category are “bad”). We will answer the following general question: is it true that the functor
Db (C) → D is an equivalence of categories?
Of course, for any abelian heart C the categories Db (C) and D have the same objects. However, in
general there is no reason why the Hom-groups in D should coincide with the Yoneda Ext’s in the category
of complexes. We will outline two approaches to detecting in which case one in fact has such coincidence.
The first approach is due to Beilinson:
Lemma 4.6.31 The natural functor Db (C) → D is an equivalence of categories if and only if one of the
following equivalent conditions holds:
(i) Ext∗C (M, N ) ' Ext∗D (M, N ), ∀M, N ∈ Ob(C).
(ii) For all M, N ∈ Ob(C), i > 0 and e ∈ ExtiC (M, N ) there exists an embedding N ,→ N 0 (depending
on e) such that the image of e under the map ExtiC (M, N ) → ExtiC (M, N 0 ) is equal to zero.
(Maybe one has to put finiteness conditions that are satisfied in examples) 2
4.6.32 Example Let us consider the situation when D is the derived category of complexes of sheaves
with locally constant cohomology on some topological space X and C is the category Loc of local systems
on X. Then we have to establish whether or not the groups ExtiD (L1 , L2 ) = H i (X, L∗1 ⊗ L2 ) are isomorphic
to ExtiLoc (L1 , L2 ) = ExtiLoc (1, L∗1 ⊗ L2 ) (wheref 1 denotes the trivial one-dimensional local system. Denote
the tensor product L∗1 ⊗ L2 by L. Recall that the category Loc is equivalent to the category Rep(π1 (X)) of
finite-dimensional representations of the fundamental group π1 (X).
Now we want to know under what conditions the groups H i (X, L) are isomorphic to Yoneda groups
ExtRep(π1 (X)) (1, L). In general this is a dificult question but one can give a sufficient condition. First note
77
that we have to use the Yoneda Ext’s since the category Rep(π1 (X)) of finite-dimensional representations
does not have enough projective objects. To use projective resolutions one has to consider a larger category
Rep∞ (π1 (X)) if infinite-dimensional representations of π1 (X). With this understood, we can formulate the
sufficient conditions on X:
(1) X is a K(π1 (X), 1)-space and
(2) Ext-groups in the category Rep∞ (π1 (X)) coincide with the Ext-groups in the category Rep(π1 (X)).
Note that the second condition depends on the group π1 (X) only. It is known to hold in the following
cases:
(a) π1 (X) is a finite group.
(b) π1 (X) is a finitely generated free abelian group.
(c) π1 (X) is a finitely generated free (non-abelian) group.
The condition (c) is of particular importance to us since it is satisfied when X is, for example, an affine
algebraic curve. Hence we have proved the following proposition
Proposition 4.6.33 If X is a K(π, 1)-space and π satisfies one of the conditions above then the derived
b
category DLoc is equivalent to the derived category Db (Loc). 2
b
Now we are ready`to prove that the category Dhol (DX ) is equivalent to the derived category of Hol(DX ).
In fact, let X = Xα be an algebraic stratification of X such that:
(1) Each stratum Xα is a K(π, 1)-space and π1 (Xα ) is a free finitely generated group, and
(2) For any α the morphism jα : Xα ,→ X is affine (this condition ensures that all the higher derived
functors of (jα )∗ and (jα )! vanish).
Consider the category C ⊂ Hol(DX , {Xα }) of holonomic DX -modules on X smooth along the strata.
`
Theorem 4.6.33.2. If X = Xα is a smooth algebraic stratification as above and and M, N ∈ C then,
for any i,
ExtiC (M, N ) = ExtiDb (M, N ).
hol

Proof. For i = 1 the statement follows from the fact that he category C is stable under extensions. For
the general case we need the following
Proposition 4.6.34 The following properties of a D-module N ∈ Ob(C) are equivalent:
(a) For any α and any local system L on Xα one has
Ext1Db ((jα )! L, N ) = 0,
hol

i
(b) The same property holds for Ext with any positive i.
(c) N has a finite filtration {Ni } such that Ni /Ni−1 = (jβ )∗ I, where Iβ is an injective object of the
category of local systems on Xβ .
Comments on proof. (c) implies (b) since
Exti ((jα )! L, (jβ )∗ I) = Exti (jβ∗ (jα )! L, I) = Exti (0, I) = 0
if β 6= α and if β = α we use injectivity of I. The implication (b) ⇒ (a) is obvious. Finally, (a) ⇒ (c) is
proved by categorical diagram chasing using induction on strata and one trick due to Jantzen. 2
If all the fundamental groups of the strata are finite (e.g. trivial) then the category C has enough injective
and projective objects. Hence we can proceed as follows:
Step 1. Using long exact sequences and the 5-lemma we can reduce to the case when M is simple (since
any holonomic module is of finite length).
Step 2. Choose an injective resolution 0 → N → I0 → I1 → . . . in C. This resolution gives two spectral
b
sequences for computing Ext-groups in C and Dhol (X), respectively. Since the differentials in the spectral
sequences commute with the maps between them it is enough to prove the theorem in the case when N is
injective.
78
 Step 3. Argue by induction on dim SuppM. Suppose Supp M = Xα . Then
u
0 → Keru → (jα )! (M|Xα ) −→ M → 0,
where the map u is surjective by simplicity of M and Ker(u) is supported on the strata of smaller dimension.
It follows from the long exact sequence of Ext’s that it suffices to show that the groups Exti (jα )! (M|Xα , N )
b
are the same in C and Dhol (X). Since we assumed N to be injective both these groups are trivial by the
proposition above.

In the general case (when the fundamental groups are not finite) we cannot assum that N is injective
since C doesnot have enough injectives. Instead, given any particular e ∈ ExtiDb (M, N ), we want ot find
hol
an inclusion N ⊂ N 0 such that the image of e in ExtiDb (M, N ) is equal to zero. Of course, if N embeds
hol
into injective module N 0 then N 0 will work for all e. If this is not the case we can still find a module N 0
which annihilates each particular e. (details still to be worked out) 2
This result suffices to prove (4.3.7.3) since, given any particular pair of DX -modules M and N , we can
first choose a stratification of X such that M and N are smooth along the strata of it. By removing finitely
many divisors from each stratum we can guarantee that the inculsion of this strarum in X is affine and the
stratum itself is an iterated locally trivial fibration with fibers isomorphic to affine curves (hence a K(π, 1)-
space. Hence, for any particular pair M and N we can choose a stratification such that the conditions (1)
and (2) are satisfied. 2

We now outline a second approach to (4.3.7.3) due to Beilinson. Recall that it suffices to find an
embedding N ,→ N 0 annihilating some particular element e ∈ ExtiDb (M, N ).
hol
Step 1. Choose a Zariski open subset U ⊂ X that has a smooth morphism to some variety Z with fibers
isomorphic to affine curves. We will find a module N 0 which solves the problem fo ExtiDb (U ) . Once we know
hol
0 0
how to doLthat, we0 can choose an open cover X = ∪Uα , find a module Nα on each Uα and take N to be
equal to (jα )∗ Nα . To that end, note that the Ext-groups on U can be computed via some Ext-groups on Z
using Leray spectral sequence. Shrinking U if necessary we can assume that all the direct image sheaves on
Z involved in the spectral sequence are locally free. We will deduce the statement for U from the statement
for Z, inducting on dimension.
Step 2. Induction on min(dim Supp M, dim Supp N ). Suppose that both dimensions are less then
dim X. Then there exists a divisor Y containing both Supp M and Supp N (such Y may not be smooth
and irreducible). We represent Y (locally) as a zero set of some function f and consider our usual diagram
Y ,→ X ←- V . We obviously have ExtiDb (X) (M, N ) = ExtiDb (Y ) (M, N ) since both sides of the equality
hol hol
have topological meaning. Hence it suffices to prove ExtiC(X) (M, N ) = ExtiC(Y ) (M, N ), where the category
calHol(Y ) of holonomic modules on Y is taken to be the full subcategory of Hol(X) generated by modules
supportedon Y . We always have a map from the LHS of the equality to the RHS of it. Since the functor
Φ : Hol(X) → Hol(Y ) restricts to the identity on Hol(Y ), the map ExtiC(X) (M, N ) → ExtiC(Y ) (M, N ) is
injective.
To show the surjectivity recall that the Yoneda Ext groups classify all the extensions
0 → M → ... → N → 0
modulo the equivalence relation generated by all commutative diagrams
0 /M / ... /N /0

0 /M / . . . /N / 0.

79
Hence our goal is to connect a representative of the Yoneda Ext-group on X to a representative of a Yoneda
Ext-group on Y by a chain of such commutative diagrams. To that end, assume that we have an exact
sequence
0 → M → A• → N → 0.
Then one also has an exact sequence
0 → M → Φ(A• ) → N → 0.
Now the chain of diagrams in question is generated by the following diagram of maps:
qis qis A• ⊕ Ξ(A• |U )
A• −→ A• ⊕ Ξ(A• |U ) −→ ← Φ(A• )
j! A•
(Here we use the fact that Ξ(A• |U ) is acyclic.)
Step 3. As in the first proof we may assume that M is simple. Suppose also that dim Supp N < dim X
j
and let U = X \ Supp N ,→ X. Consider the diagram
u
0 → Ker(u) → j! M|U −→ M → 0,
where the morphism u is again surjective by simplicity of M. Since
Exti (j! M|U, N ) = Exti (M|U , N |U ) = 0,
the long exact sequence associated with the short exact sequence above reduces us to the case where the
support of M is strictly smaller than X.
Step 4. Hence we can assume that M and N are simple an that both Supp M and Supp N are strictly
smaller than X. In this case the sequence
u
0 → Ker(u) → j! M|U −→ M → 0,
and a similar sequence for N
v
0 → N −→ j∗ NU → Coker(v) → 0
allow us to induct on the dimensions of the supports. 2
4.6.35 Malgrange construction.
ε
Let Y = f −1 (0) ,→ X ←- U be as before and M be a DU -module. Consider an embedding X ,→ X × C
given by ε(x) = (x, t = f (x)). Note that the (co)-normal bundle to ε(X) is trivial, hence we do not have
to worry about left/right modules when we take direct images under ε. Consider ε∗ M = DX×C ⊗Dε(X) M.
X×C
The ring DX×C is generated by the operators in DX , multiplication by the coordinate t on C and the partial
derivative ∂t .
Proposition 4.6.36 (Malgrange) The assignment mf s 7→ 1 ⊗ m, s 7→ −t ∂t , defines an isomorphism
DX [s] (Mf s ) → ε∗ M 2
4.6.37 Remarks
t ∈ DX×C corresponds under the isomorphism above to the operator t̃ : ui (s)f s mi 7→
P
P (1) The the operator
ui (s + 1)f · f s mi .
(2) One can easily check that (f − t) · (1 ⊗ m) = 0 and, more generally (f − t)n · (u ⊗ m) = 0 for n  0
depending on u.
(3) The proposition above allows us to think of DX [s] (Mf s ) as a DX×C -module. Alternatively, we can
view it as a DX [s, t̃, t̃−1 ]-module.

Denote DX [s] (Mf s ) by N and choose a DX [t]-lattice L ⊂ N . Then the quotient L/tL is naturally a
DX -module.
Theorem 4.6.37.3. For any lattice L, L/tL is a holonomic DX -module.
80
 We will prove this theorem later and now we will state one corollary of it. Note that s still acts on L/tL
and this action commutes with that of DX . Since L/tL is holonomic, the space of its endomorphisms as a
DX -module is finite dimensional. Hence s has a minimal polynomial on L/tL.
Corollary 4.6.38 There exists a polynomial bL ∈ C[s] such that bL (s) acts on L/tL by zero. 2
Definition 4.6.39 We denote by Spec L (s) the set of roots of bL .
Proposition 4.6.40 There exists a unique lattice L0 such that
−1 < Re Spec L0 (s) ≤ 0.

Definition 4.6.41 If L0 is the lattice given by the proposition above we denote the quotient L0 /tL0 by
Ψtot (M). The reason fo this notation is that Ψ(M) is isomorphic to the submodule in Ψtot (M) on which
the action of s is nilpotent.

Proof of (4.6.37.3). First of all, proving that the class of L/tL in K + does not depend on L, we
have shown that any lwo lattices have isomorphic subquotients. Since the holonomicity is preserved under
extensions, it suffices to prove the result for one particular lattice.
Choose an O-coherent subsheaf M0 of M such that M = DX · M0 . Then the lattice L = DX [s](M0 f s )
[s]M0 f s
is stable under s (and of course t). Then the quotient L/tL is isomorphic to DDXX[s]M 0f
s+1 . Assume for
simplicity that M0 is generated by one section m (and induct in general on number of generators). By the
lemma on b-function we can find an operator u(s) such that u(s)(mf s+1 = b(s)mf s . Hence b(s)·DX [s]mf s ⊂
DX [s]mf s DX [s]mf s
DX [s]mf s+1 . This implies that L/tL is a quotient of b(s)D X [s]mf
s and it suffices to show that b(s)D [s]mf s
X

is holonomic. If b(s) = (s − βi )ki then

Q

DX [s]mf s M DX [s]f s
= .
b(s)DX [s]mf s i
(s − βi )ki · DX [s]nf s
Since holonomicity is preserved under extensions, we can assume that all ki are equal to 1. Then for β = 0
the statement was proved before and for β 6= 0 the proof follows the same pattern. 2

4.6.42 Construction of the lattice L0 .

First choose some lattice L and let [a, b] be the interval of minimal length containing Re Spec L (s). Let
l be the biggest integer satisfying b + l ≤ 0, respectively and k be any integer satisfying a + k < 0, l > k.
Then the module tl L/tk L is holonomic. Decompose it with respect to the action of s:
M
tl L/tk L = Vλ ,
λ
where Vλ is the λ-generalized eigenspace.
Since multiplication by tl shiftsM
eigenvalues by tl , all λ in the decomposition above are non-positive.
Define L̄0 to be the subspace L̄0 = Vλ and let L0 be the inverse image of L̄0 under the projection map
−1<λ
tl L → tl L/tl L. Then L0 satisfies all the properties required from it. 2
4.6.43 Remarks.
(1) The construction of L0 above is independent of the original choice of L.
(2) The operation of taking eigenvalues of s = t∂t on the DX [s, t]-module N = Mf s [s] itself would not
make sense since the action of s is not locally finite. Hence we mod out a “small” lattice, which cannot
affect th eigenvalues in the (−1, 0]-range. Once we do that, choosing the eignespaces corresponding to all
λ ∈ (−1, 0] becomes an exact functor.
(3) Instead of the interval (−1, 0] we could choose any bounded set G representing the points of R/Z.
81
Proposition 4.6.44 The zero weight component in L0 /tL0 is isomorphic to Ψ(M). 2
λ
4.6.45 Remark One can also take the λ-component of L0 /tL0 and obtain an exact functor Ψ (M)
corresponding to Mf s [[s − λ]] instead of Mf s [[s]].
4.7. Verdier specialization.
Let Y be a submanifold of X (not necessarily a divisor). We will construct the Verdier specialization
map
SpX/Y : Hol(DX ) → Hol(DTY X ).
This fonctor is connected (unproved?) with our previous constructions in the following way. Suppose
Y = f −1 (0) for some function f on X. Such a function is not defined uniquely an choosing a particular
f amounts to trivializing the normal bundle TY X induced by the trivialization of TY∗ X via the section df .
Once this bundle is trivialized we can write
ε : Y ,→ Y × C = TY X, y 7→ y × 1,
(in invariant terms one can say that the dual map ε : Y ,→ TY∗ X
is given by df ). In this particular case one
has
ε! (SpX/Y M) = Ψtot
f M.
So, if g is another function vanishing along Y then Ψtot tot
f M and Ψg M are isomorphic but non-canonically.
Thus the functor Ψ has an advantage of being defined even for singular divisors Y (but it depends on the
choice of f ), while Sp is canonical and works in any codimension.
Consrtuction of SpX/Y uses the standard deformation to normal bundle diagram:
 /X o ? _ X × C∗ pr1 / X
TY X
g pr2
   
{0}  /Co ? _ C∗
M

If X is affine then X is defined to be Spec IYi and in general X is obtained by gluing the spectra
i≥0
of this type. Note that the projection g is a flat map. For any holonomic DX -module M we define
+
SpX/Y (M) := Ψtot
g pr1 M.

It follows that in fact the image of Verdier specialization functor belongs to the subcategory of mon-
odromic (i.e. C∗ -invariant) holonomic D-modules on TY X.

4.7.1 Verdier gluing.

If M is a holonomic module on X \ Y then, applying the same procedure as above, we obtain a module
SpX/Y (M) on TY X \ Y . Given a holonomic module M on X \ Y and an monodromic module MY we define
the glueing data to the pair (M, MY ) together with an isomorphism u : SpX/Y (M) ' MY |TY X\Y . This
gluing data is equivalent to Beilinson’s glueing data when Y is a divisor since then bundle TY X has rank
one and giving a monodromic module on TY X “almost” amounts to giving a module on Y itself. As before,
from any glueing data we can obtain a D-module on X itself (was it explained how?).

To provide another view on Verdier specialization, let Y ⊂ X be a submanifold and denote the conormal
bundle TY∗ X by Λ. We introduce a non-separating Z-filtration FiΛ DX on DX . We will in fact give three
equivalent definitions. In all of them we take F0Λ DX to be DX
Y
(the subring generated by OX and the sheaf
Y
TX of vector fields tangent to Y at all points of Y ).
4.7.2 First definition.
FiΛ DX = {u | u(IYk ) ⊂ IYk−i , ∀k ≥ i}.
82
 4.7.3 Second definition. X j
FiΛ DX = IYk DX Y
DX .
j−k≤i

4.7.4 Third definition. Recall that BY |X denotes the sheaf

codim Y
H[Y ] (OX ) = i∗ OY ,
where i : Y ,→ is the embedding. The sheaf BY |X has a natural OX -module filtration (by the number of
derivatives in directions transversal to Y ). We now define
FiΛ DX = {u |u(BYk |X ) ⊂ BYk+i
|X ∀k}.

Let p be the projection TY X → Y . The geometrical significance of the filtration introduced above is
that, for any k ∈ Z, there exists a canonical isomorphism
 
FkΛ DX /Fk−1
Λ
DX ' p· DTY X (k) ,
where DTY X (k) stands for the sheaf of differential operators of homogeneous degree k along the fibers, i.e.
if Eu is the Euler vector field generating the C∗ action on the fibers of p then u ∈ DTY X (k) if and only if
[Eu, u] = k u.

4.7.5 Example If Y is adivisor in X and t denotes a local coordinate along the fibers of p then
t ∈ DTY X (−1), ∂t ∈ DTY X (1) and t ∂t ∈ DTY X (0).
Recall that in general TY X = Spec(⊕IYi /IYi+1 ), so any section of IY gives rise to a section of p· DTY X (−1).
Y
If ξ ∈ TX then ξ·IYi ⊂ IYi hence ξ acts on ⊕IYi /IYi+1 and therefore gives a vector field on TY X of homogeneous
degree -1. This establishes the isomorphism
Λ
grF ' p· DTY X .
One has a notion of a good filtration with respect to F Λ on a DX -module M. Any such filtration is
obtained by choosing an OX -coherent submodule M0 generating M and defining MΛ Λ
i = Fi · M0 . Then
Λ FΛ
gr M is a gr DX -module or, equivalently, a D-module on TY X. However, since p· DTY X(k) is not OY -
coherent, the action of the vector filed Eu on grΛ M may not be locally finite. We will say that a filtration
on M is very good if the action Eu is locally finite.
Theorem 4.7.5.4. (Kashiwara) If M is holonomic then there exists a unique very good filtration on M
such that
i − 1 < Re Spec MΛ Λ
i /Mi−1 ≤ i,
for any k ∈ Z. 2

4.7.6 Remark The main difficulty of the proof is in finding at least one very good filtration. Once this
is done, the restriction on the spectrum will be easy to achieve koryavo.
Definition 4.7.7 The unique filtration provided by the theorem above is called Kashiwara filtration.
From now on we will only consider the KAshiwara filtration on M.
Corollary 4.7.8 The fuctor M → grΛ M is exact. 2
Theorem 4.7.8.5. There exists a natural isomorphism of DTY X -modules
grΛ M ' SpX/Y M.

4.7.9 Remark This theorem was first proved by several people in 1984-85 (unpublished). The assertion
of this theorem is a generalization of the coincidence of two definitions of Ψ.
83
 4.8. Fourier transform and second microlocalization. Note that one can also consider the dual
situation: let p∨ : TY∗ X → Y be the projection of the conormal bundle to Y . Then we can formulate the
dual counterpart of Proposition ???:
Proposition 4.8.1 There exists a natural isomorphism of algebras
Λ
grF DX ' p∨
· DTY∗ X .

The isomorphism in this case has to be constructed differently: any section ξ of TX defines a function
on T ∗ X. If ξ is in fact a section of TX
Y
then ξ vanishes on Λ = TY∗ X hence the hamiltonian vectore field vξ
associated to ξ is tangent to Λ (this is a property of any coisotropic variety).
Hence, given any DX -module M, we can construct a D-module on TY X and another D-module on TY∗ X.
To understand the relationship between these two modules we will need a notion of Fourier transform.
Let us suppose first that E is a (complex) vector space and E ∨ is the dual vector space. One has a
natural isomorphism of algebras
DE ' D E ∨ ,
which interchanges multiplication by the coordinate functions with partials. Therefore, any DE -module can
be also viewed as a DE ∨ .
If p : E → X is a vector bundle over an algebraic variety X, and p∨ : E ∨ → X is the dual bundle, then
the isomorphism above becomes an isomorphism of sheaves
p∨ ?
· DE ∨ = det E ⊗ p· DE ⊗ det
?−1
E.
Therefore, given any DE -module N , we can consider p· N as a p· DE -module and hence det? E ⊗p· N becomes
a p∨
· DE ∨ -module. Finally, we obtain a DE ∨ which is called the Fourier transform of N .

Definition 4.8.2 We denote by Φ∧ (M) the D-module on TY∗ X obtained from a DX -module |M via the
procedure described above.
Proposition 4.8.3 The DTY∗ X -module Φ∧ (M) is isomorphic to the Fourier transform of SpX/Y (M).
Now suppose that Y is defined by a single equation {f = 0}. This gives a trivialization on both normal
and conormal bundles to Y and we define the maps ε : Y → TY X and ε∨ : Y → TY∗ X by y 7→ (y, 1). In this
notations we have the followng important formulas:
ε! SpX/Y (M) = Ψf (M), (ε∨ )! Φ∧ (M) = Φf (M).
4.8.4 Comments.
(1) Last time we’ve had a fixed p ∈ C[x1 , . . . , xn ] and we used the D-module M = C[x1 , . . . , xn ]p · pλ .
Now V = {v ∈ Cn | p(v) 6= 0} is affine Zariski open in Cn and M = C[V ] · pλ . In case n = 1 this is the
exact analogue of M = C[x, x−1 ] · xλ for p(x) = x.
(2) For any fixed ϕ ∈ C.∞ (Rn ) and Re λ > 0, the assignment
Z
λ 7−→ ϕ(x)p(x)λ dx
U
is holomorphic w.r.t. λ, in fact,
Z Z
d λ
ϕ(x)p(x) dx = ϕ(x)[log p(x)]p(x)λ dx
dλ U U

k n
4.8.5 Example On Rn , consider a quadratic form p(x) = x2i − x2j , and let U = {p(x) > 0}.
P P
i=1 j=k+1
We are looking for an algebraic differential operator u ∈ D(Cn )[λ], such that
u(pλ+1 ) = b(λ)pλ , (∗)
84
where b is a b function.
d λ+1
[last time: p(x) = x, dx x = (λ + 1)xλ , b(λ) = λ + 1.]
Note that for equation (∗), the signs do not matter because we can make a complex-linear change of
coordinates.
n n
∂2
x2i . Then u = ∆ =
P P
So we assume p(x) = ∂x2
.
i
i=1 i=1
4.8.6 Claim.
n
∆(pλ+1 ) = 4(λ + 1)(λ + )pλ
2
∂
P
4.8.7 Hint. The operators p, ∆, Eu = xi ∂x i
form an ξl 2 (C).

Proposition 4.8.8 For X = Cn , the original definition of D(Cn ) ⇐⇒ Grothendieck’s definition.

4.9. Additional Comments on Grothendieck’s Definition. A is a commutative ring, S ⊂ A a

multiplicatively closed subset which does not contain any zero divisors =⇒ get S −1 A.
Moreover, for any A-module M , we get S −1 M .
Proposition 4.9.1 If M , N are A-modules and u ∈ DiffA (M, N ), then u extends canonically to a
differential operator u ∈ DiffS −1 A (S −1 M, S −1 N ).
4.9.2 Example If ξ : A −→ N is a derivation, we extend it by
a s · ξ(a) − a · ξ(s)
ξ( ) = .
s s2

Corollary 4.9.3 If X is a smooth affine variety with A = C[X] and U ⊆ X is a Zariscki open affine, then
any differential operator u on X “restricts” canonically to U .
4.9.4 Notation ∀ A-bimodule D, have (ad a)(d) = a · d − d · a, ∀a ∈ A, d ∈ D.
Lemma 4.9.5 [Ore] Let D be an A-bimodule. Assume that every ad s, s ∈ S, is locally nilpotent on D.
Then
(1) Any left fraction s−1 d can be written as a right fraction ut−1 , i.e., (∀s ∈ S, d ∈ D)(∃t ∈ S, u ∈
D)(su = dt).
(2) Similarly, any right fraction can be written as a left fraction.
In fact we can choose t = sk , some k ∈ N.
Proof. In general, we have an identity
n  
X n r n−r
(ad s)n (d) (−1)n−r s ds (ind. on n).
r=0
r
Choose n so that (ad s)n d = 0. Hence
n  
n−1 n
X
n−r n r n−r
(−1) ds = (−1) s ds .
r=1
r

Corollary 4.9.6 There exists an unique A-bimodule isomorphism ϕ making the following diagram com-
mute:
DF
xxx FF
FF
xx FF
xx F"
|x
−1
S D
' / DS −1
ϕ

In particular, via this identification, S −1 D = DS −1 is an S −1 A-bimodule.

85
 Last time: If X is any connected smooth algebraic variety, we define a sheaf DX by U DGroth (U ).
By the above, DX is a presheaf on the affine opens of X. To proof that it is a sheaf, we must show that
for any smooth affine X, we have DGroth (X) = Γ(X, DX ) = global sections of the associated sheaf.
k
Proposition 4.9.7 If X is smooth affine, then DX is a sheaf for each k = 0, 1, 2, . . . .

4.10. Comments on the definition of DX .

1) DX ⊆ EndC C[X] s.t. on each open U ⊂ X, u(C[U ]) ⊂ C[U ] and u|C[U ] ∈ DGroth (U ).
2) More elegantly: DX ⊆ EndC OX .
3) If X = Cn , we have defined D(Cn ) as the subalgebra generated by C[Cn ] and vector fields. For general
X, let us assign to any open affine U ⊂ X the subalgebra of EndC (C[U ]) generated by C[U ] and vector fields
on U . However, we can not even prove that this is a presheaf.
4) In holomorphicP geometry, we have holomorphic local coordinates, therefore can define differential
operator locally by uk ∂ k .

4.11. Principal symbol map.

Special case X = Cn :
X
Dk (Cn ) 3 u = uk1 ...kn ∂ k1 · · · ∂ kn 7−→ σk (u)
k1 +...+kn 6k
X
= uk1 ...kn (x)ξ1k1 · · · ξnkn
k1 +...+kn =k
∈ C[x1 , . . . , xn ][ξ1 , . . . , ξn ] homogeneous in ξ’s.
In general, let X be a smooth affine variety, A = C[X]. Then
X 3 x ←→ mx = {f ∈ A|f (x) = 0} ⊂ A maximal ideal.
k k
Have D (X) = D (A), the Grothendieck’s differential operator of order 6 k on X.
Lemma 4.11.1 If u ∈ Dk (A) and f ∈ mk+1 x ⊂ A, then (uf )|x = (uf )(x) = 0.
Proof. Induction on k. We have f ∈ mxk+1 =⇒ f = sum of terms of the form f0 · f1 · . . . · fk , where
fi ∈ mx . By definition, [f0 , u] ∈ Dk−1 (A). By induction, [f0 , u](f1 f2 . . . fk )|x = 0.
Hence (uf )|x = [u, f0 ](f1 f2 . . . fk )|x + f0 (x) · u(f1 f2 . . . fk )|x = 0 because f0 ∈ mx .

First construction of the principal symbol.

We want to define
Dk (X) 3 u 7→ σk (u)|x = a degree k homogeneous polynomial on Tx∗ X
4.11.2 Reminder. C[x]/mx ∼ = C via f 7→ f (x).
mx /m2x ∼= Tx∗ X via f 7→ df |x .
x is smooth ⇔ the canonical map.
S k (mx /m2x ) −→ mkx /mk+1
x is an isomorphism ∀ k (S k = k-th. symm. power).
The definition of σk (u)|x is as follows. Define a C-linear map mkx −→ C by f 7→ (uf )|x . By the lemma,
this descends to a C-linear map mkx /mk+1 x −→ C. By the above, this is the same as a map σk (u)|x ∈
HomC (S k (mx /m2x ), C) = HomC (S k Tx∗ X, C).
But to give a degree k homogeneous polynomial on a vector space V is the same as to give a linear map
S k V −→ C.

4.11.3 Reformulation. If ϕ ∈ mx , then f = ϕk ∈ mkx , and (uf )|x depends only on dϕ|x and the
dependence is polynomial of degree k.
86
 Varying x ∈ X, we get σk (u) ∈ C[T ∗ X].
∞ fk
Lemma 4.11.4 [Useful formula] u ∈ Dk (X), f ∈ C[X] = A ⇒ etf = tk
P
∈ A[[t]].
k=0 k!
Then
e−tf · u(etf ) = tk · (σk (u) · df ) + lower powers of t.
df σk (u)
[here, X −→ T ∗ X −→ C].
Proof. (DX).
This lemma is the second construction of the principal symbol.

Third construction of the principal symbol.

If u ∈ Dk (A), then by definition, for all f0 , f1 , . . . , fk ∈ A = C[X], we have [f0 , [f1 , . . . , [fk , u] . . . ]] = 0.
We define a map
σ : A ⊗C A ⊗C . . . ⊗C A −→ A
by
(f1 , . . . , fk ) 7→ [f1 , [f2 , . . . , [fk , u] . . . ]] ∈ A.
In fact, σ is symmetric. If suffices to note that [fi , [fi+1 , v]] = [fi+1 , [fi , v]] for any operator v, because
[fi , fi+1 ] = 0. Thus σ induces
σ : SCk A −→ A.
Moreover, σ(f1 ⊗ . . . ⊗ fk ) depends only on the differentials dfi , and so it gives a map S k T ∗ X −→ A.
Indeed, it suffices to show that if dfi |x = 0 for at least one i, then σ(f1 ⊗ . . . ⊗ fk )|x = 0. By symmetry,
we may assume that i = 1. We may also assume that fi (x) = 0, and therefore f1 ∈ m2x . By linearity, we
may then assume that f = ϕ · ψ, where ϕ, ψ ∈ mx . But now
σ(f1 ⊗ . . . ⊗ fk )|x = [ϕψ, ?]|x = ϕ|x · [ψ, ?]|x + ψ|x · [ϕ, ?]|x = 0.
This way we have defined σk as a sheaf map
k
DX −→ π.OT ∗ X (k)
where π : T ∗ X → X is the projection and π. is pushforward, and “(k)” means “homogeneous of degree k
along the fibers”.
Theorem 4.11.4.6. We have a short exact sequence of sheaves
k−1 k k σ
0 −→ DX −→ DX −→ π. OT ∗ X (k) −→ 0.
In fact, for any smooth affine X, we have a short exact sequence
σ
0 −→ Dk−1 (X) −→ Dk (X) −→
k
π. Ckhomog [T ∗ X] −→ 0.
Proof. By Grothendieck’s definition, we have Dk−1 (X) = Ker(σk ). So we must prove that σk is surjective.
Recall that we’ve had a naive definition of Dk (X) as the subspace Dorig k
(X) ⊆ EndC (C[X]) generated
k
by all elements of the form f0 ξ1 . . . ξl , where l 6 k, f0 ∈ C[X], ξi are vector fields. It is clear that Dorig (X) ⊆
k
DGroth (X).
We will prove surjectivity of σk together with the following
Proposition 4.11.5 Dorig = DGroth .
k−1 k−1
Indeed, by induction, w.m.a. Dorig = DGroth k
. Then it suffices to prove that σk : Dorig −→ Ckhomog [T ∗ X]
is surjective.
Lemma 4.11.6 Let A = C[X] and M a locally free f.g. A-module. Then
k k
HomA (SA M, A) = SA HomA (M, A).
87
 Proof. This is true for M = A =⇒ true for any M = A⊕n =⇒ true for any direct summand of A⊕n .
But locally free f.g. =⇒ projective f.g.
Now we have Ckhomog [T ∗ X] = HomA (SA
k
Γ(X, T ∗ X), A) = S k HomA (M, A) because M is locally free
| {z }
=:M
(since X is smooth).

Also, if M is locally free, then it is projective, and hence

HomA (Hom( M, A), A) ∼ = M canonically.
Now Γ(X, T ∗ X) = HomA (Γ(X, T X), A),whence
Ckhomog [T ∗ X] = S k Γ(X, T X).
Hence every element of this can be written as a sum of products of the form ξ1 , . . . , ξk where ξi ∈
Γ(X, T X).
Let X be an algebraic variety and U ⊆ X open affine. For any f ∈ C[U ], put Uf = U \ {f = 0}. For
any C[U ]-module M , we have Mf .
Lemma 4.11.7 For an OX -sheaf F, the following are equivalent:
(1) F = lim
−→ Fi (union) of coherent subsheaves Fi ⊆ F.
k
[E.g. DX = −lim
→ XD ]
(2) For any open affine U ⊆ X and any f ∈ C[U ], we have
Γ(U, F)f = Γ(Uf , F).
If (1) and (2) hold, we say F is quasi-coherent.
4.11.8 Homework.
1) p ∈ C[x1 , . . . , xn ] =⇒ show that D(Cn ) · ep is holonomic.
2) (a) For which λ, µ ∈ C do we have D(C) · eλx ∼ = D(C) · eµx as D-modules.
(b) Is D(C) · e1/x ∼
2
= D(C) · e1/x ?
∗
3) Let M be a holonomic D(Cn )-module.
(a) Prove that dimC EndD(Cn ) (M ) < ∞
c n
(b) Suppose we have a filtration on M such that hM (t) = t + lower order terms, c ∈ N. Let
n!
a : M −→ M be a D(C )-module map. Show that the elements IdM , a, a2 , . . . , ac are linearly depended
n

over C.
4.12. Prehomogeneous vector spaces.
V = finite dimensional vector space over C, G ⊆ GL(V ) is a closed connected reductive algebraic
subgroup.
Definition 4.12.1 The pair (V, G) is called a regular pre-homogeneous vector spae if G has an open
dense orbit V reg such that V \ V reg is an irreducible hypersurface, i.e., there exists irreducible f ∈ C[V ]
such that V \ V reg = {f = 0}.
4.12.2 Examples.
(1) Cn with f (x) = x21 + x22 + . . . + x2n (assume n > 2).
Take G = SO(n) · C∗ .
(2) Counterexample: G = GL(V ) for dim V > 1.
(Prehomogeneous but not regular)
(3) G = GL(V ) acting on End(V ) by left multiplication.

Elementary properties:
88
 (1) f is a G-semiinvariant, i.e., there exists an algebraic group homomorphism χ : G −→ C∗ such that
f (gx) = χ(g)f (x) ∀x ∈ V, g ∈ G.
(2) f is homogeneous of degree d := deg f . [Indeed, the G-action commutes with the C∗ -action =⇒ the
∗
C -action must preserve V reg .]
Theorem 4.12.2.7. There exists b(λ) ∈ C[λ] such that deg b = deg f and
f (∂)f λ+1 = b(λ)f λ
(f (∂) = differential operator with constant coefficients obtained by replacing the xi ’s by ∂i ’s in f ).

Corollary 4.12.3 det(∂) · det(x)λ+1 = b(λ) · det(x)λ ←− related to the “Capelli identity”.
In fact, it is known that Y
b(λ) = (λ + k).
16k6n

4.12.4 Claim. T is a D-endomorphism of M .

Proof. Straightforward.
Hence T is well defined as an endomorphism of
M |1 = M/(t − 1)M.
Also, since
T : tn 7→ exp(2πin) · tn = tn ,
it is clear that T is unipotent.
Conversely, to define the functor in the other direction, it suffices (by the Jordan decomposition) to
consider the case  
1 1 0

 1 1 

T =
 1 


 . .. 1 
0 1
For this T , the corresponding D-module is
(log t)n−1 · R ⊕ (log t)n−2 · R ⊕ . . . ⊕ (log t) · R ⊕ R.

To prove that the functor C ∗ → (V, T ) constructed above is an equivalence, it is enough to show that


it is fully faithful and essentially surjective. Last time we have shown essential surjectivity.
Let C2 be the category of pairs (V, T ) as above, and let
M on : C ∗ → C2
be the monodromy functor. Let
  
1 1



 1 1 

 n
J ordn = V = C , T = 
 1  ∈ C2 ,

  .. 
  . 1 
1
and let
Jn = R · (log t)n−1 ⊕ R · (log t)n−2 ⊕ . . . ⊕ R · (log t) ⊕ R ∈ C ∗ .
4.12.5 Claim. M on(Jn ) ∼
= J ordn .
89
 
Proof. Jn has a C-basis tk · (log t)l | k ∈ Z, 0 6 l 6 n − 1 , and the action of t ∂t is given by

t ∂t : tk · (log t)l 7→ ktk · (log t)l + ltk · (log t)l−1

For a fixed k, the matrix of t ∂t is thus
 
k n−1
 k n−2 
,
 
 ..
 . 1 
k
and we see that  
1 1 ∗
 1 
T = e2πit ∂t =  ,
 
..
 . 1 
0 1
which is what we want.

Next we have to prove that

HomDC∗ (Jn , Jm ) ∼
= HomC2 (J ordn , J ordm )
Geometrically,
e0 e1 e2 en−1
J ordn : · −→ · −→ · −→ . . . −→ · −→ 0
e0 e1 e2 em−1
J ordm : · −→ · −→ · −→ . . . −→ · −→ 0
Using this, we can compute the dimensions of both homomorphisms and thus prove that they are isomorphic.

Alternative interpretation.
Let M ∈ C ∗ , than we can write M =
L
Mi , dim Mi < ∞, where each Mi is t ∂t -stable. [Mi = the i-th
i∈Z
eigenspace of t ∂t ]. Picture:

...
t / t / t / ...
M−1
. o M. 0
Wo
M. 1
Wo
M. 2
∂t ∂t ∂t

t∂t t∂t
∗
Since we are working in C , t is clearly an isomorphism; in particular, all the Mi have the same dimension.
The category C ∗ has a single simple object, namely R. It does not have projective objects, but it does
have a pro-object E proj .
Note that Jn has a filtration will all quotients ∼=R
   
R R
 R   R 
 ←−
   
Jn ∼  .. ..
 ∼ Jn+1
 
.  
 . 
R  R 
0 ←− R
90
 We take  
R
 R 
E proj := lim J ∼
 
←− n 
 R 

..
.

Lemma 4.12.6 The functor M HomD (E proj , M ) is given by

HomD (E proj , M ) = M0 ;
in particular, E proj is projective.
4.12.7 Remark We can also define E ind = lim −→ Jn with respect to the natural inclusions Jn ,→ Jn+1 ,
and then E ind is the injective hull of R.
To prove the lemma above, we have the following
4.12.8 Claim. Jn ∼ = D/D · (t ∂t )n .
Proof. We have a natural map D/D · (t ∂t )n → Jn given by 1 7→ (log t)n−1 , and it is clearly surjective.
Since t is invertible in R, we have
D = C[t, t−1 ][∂t ]
= C[t, t−1 ][t ∂t ].
Thus D/D · (t ∂t )n is a free R-module with basis 1, (t ∂t ), (t ∂t )2 , . . . , (t ∂t )n−1 , completing the proof.
Now we immediately get the tautological projections
Jn+1 −→ Jn .
Moreover, we see that
E proj = ← n
− D/D · (t ∂t ) .
lim
n
This implies that
Hom(E proj , M ) = Hom lim D/D · (t ∂t )n , M = M0


← −
n

Now we have
J ordproj = ← − J ordn = e·0 −→ e·1 −→ e·2 −→ . . .
lim
It remains to prove that our functor takes
'
End(E proj ) −→ End(J ordproj )
We claim that both are isomorphic to C[[u]].

E proj = lim n
←− D/D · (t ∂t ) = ← − Jn
lim
E ind = R[log t] = lim
−→ nJ

4.12.9 Remark We have duality D on holomorphic modules. Note that D(R) = R. Hence D : C ∗ −→ C ∗
observe that    
A1 DA1
D :  A2  7−→  DA2 
A3 DA3
proj ∼ ind


We have D E =E .
91
 Recall the diagram
j
{0} ,→ C ←- C∗
Let δ = D/Dt and O = C[t] = D/D ∂.
Definition 4.12.10 Let C be the category of D-modules of finite length on C with all subquotients
isomorphic to either δ or O.
We have j∗ R. As a space, j∗ R = R = C[t, t−1 ].
We have a short exact sequence of DC -modules
0 → O → R → δ → 0.
Thus  
δ
R = C[t, t−1 ] ∼
O
Now we have  
O
j! R = Dj∗ DR = D(j∗ R) ∼
δ
There is a canonical morphism j! R → j∗ R; it looks like

OW W W W /
   
δ
W W+
δ W W W O
W W W
+
0

[because this map has to restrict to the identity on C∗ ]

Explicitly,
j∗ R = D/D(∂t) and j! R = D/D(t∂),
and the map j! R → j∗ R is multiplication on the right by t.

Now we have D-modules j∗ Jn , j! Jn . Pictorially,

   
δ O
 O   δ 
   
 δ   O 
   
j∗ Jn ∼  O  and j! Jn ∼  δ  .
   
 ..   .. 
 .   . 
   
 δ   O 
O δ
We can do four limit constructions. We set

∇proj := j! E proj , ∆proj := j∗ E proj ,

∇ind := j! E ind , ∆ind := j∗ E ind .
Pictorially,
.. ..
       
O δ
 .   . 
∇proj =  δ , ∆proj =  O , ∆ind =  δ , ∇ind =  O .
   
.. ..
. . O δ
For future reference:
∇proj = D(∆ind ), ∆proj = D(∇ind ), etc.
92
 An important thing is to calculate
j! E proj −→ j∗ E proj .
We have a short exact sequence
0 → j! E proj → j∗ E proj → δ → 0,
and
0 → δ → j∗ E ind → j! E ind → 0.
Proposition 4.12.11 The category C is equivalent to the category of diagrams of finite dimensional vector
spaces
u
·j *· with (uv)N = 0 for some N > 1
v
Proof. The functor is given by

,

M 7−→ Hom ∆proj , M Hom ∇proj , M
l
where the maps are induced by

δ YYYYYYYYY eeeeeeee O
eeYY
eeeeeee YYYYYYY
O YreYYYYYYYY eeeeeeee, δ
eeYeYYYYYYYY
eeeeee
δ YreYYYYYYYY eeeeeeee, O
eeYY
eeeeeee YYYYYYY,
O re δ
.. ..
. .
Another proof. We still have the picture

M−1
t / M. 0
t / M. 1
t / M. 2
M
. o o o M= Mi ,
−1
∂t
0
R ∂t
1
∂t
2
i∈Z
t∂t

but t is no longer invertible. However, since ∂t = t∂ + 1, both ∂t and t∂ are invertible on Mi for i 6= 0, −1,
which implies that t : Mi → Mi+1 is an isomorphism whenever i 6= −2, −1, 0.
For M = O, we have M0 = C and M−1 = (0).
For M = δ, we have M0 = (0) and M−1 = C.
In general we get
Hom(∇proj , M ) = M0 and Hom(∆proj , M ) = M−1
We have a category C, and we have shown that
v
C ' M−1  M0 | (uv)N = 0 for some N > 1

u
v
A quasi-inverse can be constructed as follows. Given F  E, we set M = C[t] ⊗C E ⊕ C[∂] ⊗C F , and
u
define
t(1 ⊗ f ) := 1 ⊗ v(f ),
∂(1 ⊗ e) := 1 ⊗ u(e).
Relation to duality:
D:C→C (since D(O) = O, D(δ) = δ)
93
 By definition,
D = R HomD (−, D).
We have a free resolution
p a
0 → N → N → M → 0,
where
N = D ⊗C E ⊕ D ⊗C F,
a = the obvious action map, and P is given by
P : (a ⊗ e, b ⊗ f ) 7→ (a∂ ⊗ e, −au(e)) + (−bv(f ), bt ⊗ f ).
Obviously,
HomD (D ⊗ E, D) = E ∗ ⊗ D (E ∗ = HomC (E, C))
Now applying HomD (−, D) to the resolution above, we get
p∗
0 −→ N ∗ −→ N ∗ −→ 0
where
N ∗ = E ∗ ⊗C D ⊕ F ∗ ⊗C D,
and
P ∗ : (e∗ ⊗ a, f ∗ ⊗ b) 7→ (e∗ ⊗ ∂a, −v ∗ (e∗ ) ⊗ a) + (−u∗ (e∗ ) ⊗ b, f ∗ ⊗ tb).
We see from this that D(M ) corresponds to the diagram

u∗ /
∗
E o F∗
∗
v
[We have to use the involution to switch from right to left modules.]

Now on the category C, we have two functors:

Ψ /
C / Vect
Φ
where
Ψ(M ) = Hom(∇proj , M ) = M0 ,
Φ(M ) = Hom(∆proj , M ) = M−1 .
4.12.12 Remark Ψ(M ) is the geometric fiber of M at the point t = 1.
We want to produce maps
can /
Ψ(M ) o Φ(M )
var
with
can var
T − Id : Ψ −→ Φ −→ Ψ, T = exp(−2πit∂)|M0
We will define
e2πi(vu)−1
can = u, var = ·v
vu
4.12.13 Remark The universal cover of C∗ is
C −→ C∗
z 7−→ t = e2πiz
94
Then we have tn = e2πinz , log t = z. So, geometrically,
R[log] = C[t, t−1 , log t] = “quasi-polynomials” on the universal cover.
We have
R[log] = ∆ind .
We can define, equivalently,
Ψ(M ) = TorD (M, ∆ind ).
If M = C[t] ⊗C E ⊕ C[∂] ⊗C F as above, then we get

Ψ(M ) = (e e, fe) ∈ E ⊗ R[log] ⊕ F ⊗ R[log] | ∂e
e = −v(fe), tfe = u(e
e) .
But t is invertible on ∆ind , so
)
e = −v(fe)
∂e
⇐⇒ t∂e
e = −vue
e,
tf = u(e
e e)
and a solution of the last equation is
ee = exp(−vu log t) · e, ∀e ∈ E.
We have
1 1
fe = u(e
e) = exp(−uv log t)u(e).
t t

Recall that on C∗ , we’ve had Jn . Let us introduce a new variable s. We have a short exact sequence
C((s))
0 → C[[s]] → C((s)) → → 0.
C[[s]]
We also have the residue map
Res : C((s)) −→ C
given by Z
Res
a(s) 7−→ a−1 = “ a(s) ds00
C
Note that
• Res(C[[s]]) = (0)
• If a ∈ C((s)) is such that Res(a · C[[s]]) = (0), then a ∈ C[[s]].
This can also be done for any ring:
R((s))
0 → R[[s]] → R((s)) → → 0.
R[[s]]
We take R = C[t, t−1 ]. We also consider
Rt2 ((s))
0 → Rt2 [[s]] → Rt2 ((s)) → → 0.
Rt2 [[s]]

Lemma 4.12.14 There is a DC -module isomorphism (where coordinate on C is t):

Rt2 ((s)) '
−→ R[log t].
Rt2 [[s]]

Proof. Formally, we have

∞
X si
ts = e(log t)s = (log t)i .
i=0
i!
95
 We define our map as the composition
∞
X X sj (log t)j Res
ai si ts 7−→ ai si 7−→? ∈ R[log t], ai ∈ Rts ((S)).
i,j
j!
i=−N

Clearly, the kernel of this composition is precisely Rts [[s]], and the map is surjective.
4.12.15 Remark The universal cover of C∗ is
C −→ C∗
z 7−→ t = e2πiz

Then we have tn = e2πinz , log t = z. So, geometrically,

R[log] = C[t, t−1 , log t] = “quasi-polynomials” on the universal cover.
We have
R[log] = ∆ind .
We can define, equivalently,
Ψ(M ) = TorD (M, ∆ind ).
If M = C[t] ⊗C E ⊕ C[∂] ⊗C F as above, then we get

Ψ(M ) = (e e, fe) ∈ E ⊗ R[log] ⊕ F ⊗ R[log] | ∂e
e = −v(fe), tfe = u(e
e) .

But t is invertible on ∆ind , so

)
e = −v(fe)
∂e
⇐⇒ t∂e
e = −vue
e,
tf = u(e
e e)

and a solution of the last equation is

ee = exp(−vu log t) · e, ∀e ∈ E.
We have
1 1
fe = u(e
e) = exp(−uv log t)u(e).
t t

Recall that on C∗ , we’ve had Jn . Let us introduce a new variable s. We have a short exact sequence
C((s))
0 → C[[s]] → C((s)) → → 0.
C[[s]]
We also have the residue map
Res : C((s)) −→ C
given by
Z
Res
a(s) 7−→ a−1 = “ a(s) ds00
C
Note that
• Res(C[[s]]) = (0)
• If a ∈ C((s)) is such that Res(a · C[[s]]) = (0), then a ∈ C[[s]].
96
 This can also be done for any ring:
R((s))
0 → R[[s]] → R((s)) → → 0.
R[[s]]
We take R = C[t, t−1 ]. We also consider
Rt2 ((s))
0 → Rt2 [[s]] → Rt2 ((s)) → → 0.
Rt2 [[s]]

Lemma 4.12.16 There is a DC -module isomorphism (where coordinate on C is t):

Rt2 ((s)) '
−→ R[log t].
Rt2 [[s]]

Proof. Formally, we have

∞
X si
ts = e(log t)s = (log t)i .
i=0
i!
We define our map as the composition
∞
X X sj (log t)j Res
ai si ts 7−→ ai si 7−→? ∈ R[log t], ai ∈ Rts ((S)).
i,j
j!
i=−N

Clearly, the kernel of this composition is precisely Rts [[s]], and the map is surjective.
Corollary 4.12.17
Rts ((s))
j∗ R[log t] = ∆ind =
Rts [[s]]

||

−→ JneKK)
j∗ (lim
KK
KK
KK
KK S
3
Jn

Question. How can we see Jn inside RHS?

Answer. We have

0 / Jn / R[log t] / R[log t]/Jn /0

o| o|
s−n ·Rts ((s))   /
s
Rt ((s))
Rts [[s]] Rts [[s]]

Corollary 4.12.18
Rts [[s]] Rts [s]
Jn ∼
= n s = n s (∗)
s Rt [[s]] s Rt [s]

We have, on C∗ ,
E proj = ← s
− Jn = Rt [[s]]
lim
(this is obvious from (∗)). Hence
∆proj = j∗ E proj = Rts [[s]].
97
Now we get
0 −→ ∆proj −→ j∗ Rts ((s)) −→ ∆ind −→ 0.
Pictures.

..
.
O
∆ind = j∗ E ind =
δ
O

δ
proj = j∗ E proj = O
∆
δ
..
.

..
.
O
δ
O
j∗ Rts ((s)) =
δ
O
δ
..
.

..
.
δ
∇ind = j! E ind =
O
δ

O
proj = j! E proj = δ
∇
O
..
.

98
 ..
.
δ
O
δ
j! Rts ((s)) =
O
δ
O
..
.

'
Last time we’ve have j! E proj ,→ j∗ E proj . It induces an isomorphism j! Rts ((s)) → j∗ Rts ((s)):

O ZZZZZZZZZZ δ
ZZZZZZZZ
ZZZZ-
δ XXXXXXX O
XXXXX
XXXXX
XXXXX
X+
O ZZZZZZZZZZ δ
ZZZZZZZZ
ZZZZ-
δ O

General situation
We have X = smooth variety, f : X → C a regular function, Y = f −1 (0) = divisor:
 /Xo j _
? U =X \Y
Y = f −1 (0)
f
   
{0}  /Co ? _ C∗
j

Assume X is affine. Let NU be a holonomic DU -module. We have a map j! (NU ) → j∗ (NU ), and


Im j! (NU ) → j∗ (NU ) = j!∗ (NU ).

Proposition 4.12.19 Let N0 ⊆ NU be a finite dimensional generating subspace, so that DU · N0 = NU .

Then
j!∗ (NU ) = DX (f k N0 )
for sufficiently large k.
Proof. Note that for k  0,
DX (f k N0 ) does not depend on k
(because j∗ NU is holonomic =⇒ has finite length). Let M := j!∗ (NU ). We have
• M |U = NU
• NU /M is supported at f −1 (0)
Now there is exists k  0 s.t. f k N0 = (0) in NU /M , whence f k N0 ⊆ M , so that D(X)f k N0 ⊆ M . If this is
not an equality, then
M
M |U NU
k
= k
= = (0),
D(X)f N0 U D(X)f N0 |U NU
which contradicts the definition of M .
4.12.20 Homework. Jn = R logn−1 +R logn−2 + . . . + R, R = C[t, t−1 ]. This is a D-module on C∗
which is free of rank n as an R-module.
99
 (1) Describe the corresponding connection on the trivial bundle Rn .
(2) Check that the dual connection on HomR (Jn , R) is isomorphic to the original connection.
' 
We’ve proved that C −→ M−1  M0 . Find the images of j∗ Jn and j∗ Jn under this equivalence.
Set-up.

 /Xo ? _ U = X \ f −1 (0)
Y = f −1 (0)
f
   
 /Co j
? _ C∗
{0}
Let us replace t f , ts f s , log t log f . On U , we have
∗
Jn (f ) = f Jn = OU · (log f )n−1 + OU · (log f )n−2 + . . . + OU · log f + OU .
We can consider
∼ OU f s ((s))
lim
−→ Jn (f ) = O U [log f ] =
n OU f s [[s]]
as before (write f s = e(log f )s ). We have
f ∗ ∆ind = −→ n Jn (f ),
lim
f ∗ ∆proj = OU f s [[s]],

0 / OU f s [[s]] / OU f s ((s)) / OU f s ((s)) /0

OU f s [[s]]

f ∗ ∆proj f ∗ ∆ind

Remainder:
Lemma 4.12.21 [on the b-function] If MU is holonomic on U and MU = DU · M0 , where M0 is a
finite-dimensional generating subspace, then there exists b ∈ C[s] with
b(s)f s M0 ⊆ DX [s] f s+1 M0 .

Proposition 4.12.22 For any holonomic MU on U , we have a canonical isomorphism

'
j! MU f s ((s)) −→ j∗ MU f s ((s))

in the appropriate category.

Proof. Surjectivity. Image = j!∗ MU f s ((s)) ⊂ j∗ MU f s ((s)) . By the proposition proved last time, we
have
j!∗ (−) = DX f k · generating subspace for k  0.

Choose M0 ⊂ MU , a finite-dimensional generating subspace of MU over DX [we can do it because M0

is holonomic]. We get
j!∗ MU f s ((s)) = DX M0 f s f k ((s)) =

= DX M0 f s+k ((s)) ⊇

⊇ DX b(s + k − 1)M0 f s+k−1 ((s)) ⊇

100
 ⊇ DX b(s + k − 1)b(s + k − 2) · . . . · b(s) M0 f s ((s)) = j∗ MU f s ((s)) .


| {z }
because this is a unit in C((s))

We write Efproj = ←
lim ∗ proj
− Jn (f ) = f E (on U ),
ind
Ef = − → Jn (f )
lim

Lemma 4.12.23 On U , we have

s

DMU f s ((s))
D MU f [[s]] =
,
DMU f s [[s]]
i.e.
D Efproj ⊗OU MU = Efind ⊗OU DMU .

Note that f ∗ Jn = free OU -module of rank n.

Hence f ∗ Jn ⊗OU MU ∼ = MU⊕n as OU -modules. Also,
 
MU
 MU 
f ∗ Jn ⊗OU MU ∼  . 
 
 .. 
MU
Step 1. DEfproj ∼
= Ef . It is enough to check that D f Jn ∼
∗
= f ∗ Jn .
ind

4.12.24 Exercise Define a canonical pairing

f ∗ Jn ⊗ f ∗ Jn −→ OU .
Step 2. General case. Write a free resolution
0 → pn → pn−1 → . . . → p1 → p0 → MU → 0
(the pj are free DU -modules).
We get
0 ← DMU ← HomDU (pn , DU ) ← . . . ← HomDU (p0 , DU ) ← 0.
But now
0 → Efproj ⊗OU pn → . . . → Efproj ⊗OU p0 → Efproj ⊗OU MU → 0
is a free resolution of Efproj ⊗OU MU . But
proj
v
HomDU Efproj ⊗OU pj , DU ∼ ⊗OU HomDU pj , DU ,



= Ef
which completes the proof.

In the proposition above, it remains to prove injectivity.

But we have
j! (DMU )f s ((s))  j∗ (DMU )f s ((s))
   

Applying D gives

Dj! [−] o ? _ Dj∗ [−]

using the lemma

j∗ MU f s ((s)) o ? _ j! MU f s ((s))

101
 Corollary 4.12.25 The map
j! MU f s [[s]] −→ j∗ MU f s ((s))

is injective.
Definition 4.12.26 The nearby cycle functor ψ(−) : HolDU -mod −→ HolDX -mod supported on X \ U
is defined by
j∗ Mf s [[s]]
ψ(M) :=
j! Mf s [[s]]
This comes with a canonical DX -module endomorphism given by multiplication by s.
Let M be a holonomic DU -module; M0 finite dim’l such that DU M0 = M.
Lemma 4.12.27
We have the following
isomorphisms of DX [[s]]-modules:
(a) j! Mf s [[s]]
∼
= DX [[s]] f s+k M0

(b) j∗ Mf s [[s]] ∼ = DX [[s]] f s−k M0
for k  0
Proof. (a) follows from the fact that j! M f s [[s]] ,→ j∗ (−), whence j! (−) ∼


= j!∗ (−).
(b) follows by duality.
Proposition 4.12.28 We have

DX [s] f s−k M0
j∗ M =
,
s · DX [s] f s−k M0
and

DX [s] f s+k M0
j! M =
,
s · DX [s] f s+k M0
for k  0.
Proof. Look at the exact sequence on U :
s
0 → Mf s [[s]] → Mf s [[s]] → M → 0.
Applying j∗ and j! , we get the result (because these functors are exact).

We go back to the nearby cycle functor ψ.

Theorem 4.12.28.8. ψ(M) is a holonomic DX -module supported on f −1 (0), and the functor ψ(−) is
exact.
Proof. Let us first prove exactness.
We have the following general nonsense result.
Lemma 4.12.29 Let F , G be exact functors C → C 0 . Suppose we have in addition a morphism (natural
transformation) F → G, which is injective on all objects. Then the functor M G(M )/F (M ) is also
exact.
Proof. Obvious.

We have
 /Xo j
? _U
Y = f −1 (0)

   
{0}  /Co ? _ C∗
If M is a DU -module, we get j! M → j∗ M. Hence
   
M[[s]] M[[s]]
lim
← j
− ! sn M[[s]] = j! M[[s]] −→ j∗ M[[s]] = lim
← j
− ∗ sn M[[s]]
n n

102
 Generalization of the general nonsense lemma from last time:
Let F 0 , F, F 00 : C −→ D be three exact functors. Assume that we have morphisms of functors F 0 →
F → F 00 , whose composition is zero, such that F 0 (M ) ,→ F (M ) and F (M )  F 00 (M ) for all M ∈ C. Then
the functor M Ker(F (M ) → F 00 (M ))/Im(F 0 (M ) ,→ F (M )) is exact.
Now we want to prove that
j∗ (f s M[[s]])
ψ(M ) =
j! (f s M[[s]])
is a holonomic DX -module supported on Y . We claim that it is enough to show holonomicity. Indeed, if
ψ(M ) is holonomic, then it has finite length, whence
M[[s]]f s M[[s]]f s
    
ψ(M ) = Coker lim←− j! sn f s M[[s]] −→ lim ←− j∗ sn f s M[[s]]
M[[s]]f s M[[s]]f s
    
= Coker j! −→ j∗
sn f s M[[s]] sn f s M[[s]]
for some n  0, and the latter is clearly supported on Y .
Main theorem of last time. If K = C((s)), then
'
K ⊗ j! (Mf s ) −→ K ⊗ j∗ (Mf s ).
Proof. Choose a finite dimensional subspace M0 ⊂ M such that M = DU · M0 . Then last time we have
proved that
DX · M0 f s+k [[s]] ⊇ B(s)DX · M0 f s−l [[s]],

B(s) = b(s + k − 1)b(s + k − 2) . . . b(s − l).

Now B(s) is invertible in K. Also note that
DX · M0 f s−l [[s]] ⊆ DX · M0 f s−l−1 [[s]],

and this is an increasing chain of DX [[s]]-submodules of j∗ Mf s [[s]]. When we tensor them with K, they
become the same. But [
K ⊗ j∗ (Mf s ) = K ⊗ DX · (M0 f s−l ),
l
which completes the proof.

Key lemma. There is exists n  0 such that

sn j∗ Mf s [[s]] ⊆ j! Mf s [[s]]
This implies that
j∗ (f s M[[s]]) j∗ (f s M[[s]])
s
is a quotient of ,
j! (f M[[s]]) sn j∗ Mf s [[s]]
and the latter has a finite filtration with all quotients isomorphic to j∗ M, hence is holomorphic. This implies
our theorem.
Proof of the key lemma. We know that
j! Mf s [[s]] = DX · M0 f s+k [[s]],
j∗ Mf s [[s]] = DX · M0 f s−k [[s]],
for k  0. By the same argument as above, we have
B(s) · DX · M0 f s−k [[s]] ⊆ DX · M0 f s+k [[s]].


We can write B(s) = sn · unit of C[[s]] , and this gives the desired result.

4.12.30 Note. The proof above shows that we can choose k so that [−k, k] contains all the integral
roots of the b-functions, and n = the number of integral roots of the b-function.
103
 4.13. Digression. Given N , a holonomic DX -module, we have N[Y ] = the maximal submodule of N
supported on Y . We can define N [Y ] = the maximal quotient N supported on Y .
We also have the duality functor D : HolX −→ HolX which preserves supports (because D commutes
with restrictions to open subsets). This implies that
N [Y ] = D (DN )[Y ] .

Now we suppose M is a holonomic DU -module. By definition,

(j∗ M)[Y ] = j∗ M/j! M.
Therefore
j∗ (f s M[[s]])   proj
[Y ]
ψ(M ) = = j∗ M ⊗OU
Ef
j! (f s M[[s]])
Lemma 4.13.1 There is a canonical isomorphism
ψ(M ) ∼
= j! M ⊗OU Efind [Y ]

Proof. Write
OU f s ((s))
0 → OU f s [[s]] ,→ OU f s ((s)) → → 0.
OU f s [[s]]
=⇒
0 → M ⊗ Efproj → Mf s ((s)) → M ⊗ Efind → 0.
Let us apply j! and j∗ to these short exact sequences.
j! M ⊗ Efproj M ⊗ Efind
| {z }
z }| {
j∗ M ⊗ Efproj M ⊗ Efind

=⇒ ψ(M) ∼
= Ker j! M ⊗ Efind → j∗ M ⊗ Efind = j! M ⊗ Efind [Y ]




Corollary 4.13.2 ψ commutes with D.

Proof. We compute
  [Y ]    
Dψ D(M) = D j∗ (DM)Efproj = Dj∗ (DM)Efproj
[Y ]
     
proj proj
= j! D (DM)Ef = j! MEf = ψ(M).
[Y ] [Y ]

104
 Bibliography

[AM] Atiyah, M. F.; Macdonald, I. G.: Introduction to commutative algebra. Addison-Wesley, Mass.-London-Don Mills,
Ont. 1969.
[Ar] Arnold, Singularities of differentiable maps.
[BFM1] Baum, P.; Fulton, W.; MacPherson, R. Riemann-Roch for singular varieties. Inst. Hautes tudes Sci. Publ. Math.
No. 45 (1975), 101–145.
[BFM2] Baum, P.; Fulton, W.; MacPherson, R. Riemann-Roch and topological K theory for singular varieties. Acta Math.
143 (1979), no. 3-4, 155–192.
[Beil] Beilinson, A.: On the derived category of perverse sheaves. K-theory, arithmetic and geometry (Moscow, 1984–
1986), 27–41,Lecture Notes in Math., 1289, Springer, Berlin-New York, 1987.
[BeBe] Beilinson, A.; Bernstein, J. : Localisation de g-modules. C. R. Acad. Sci. Paris Sér. I Math. 292 (1981), no. 1,
15–18.
[BBD] Beilinson, A.; Bernstein, J.; Deligne, P.: Faisceaux pervers. (French) [Perverse sheaves] Analysis and topology on
singular spaces, I (Luminy, 1981), 5–171, Astérisque, 100, Soc. Math. France, Paris, 1982.
[BD] Beilinson, A.; Drinfeld, V.: “Hitchin’s integrable system”, preprint.
[BeFM] Beilinson-Feigin-Mazur: “Algebraic Field theory”, preprint.
[Be1] Bernstein, I.: Modules over a ring of differential operators. Funct. Anal. Appl., 5 1971 no. 2 1–16.
[Be2] Bernstein, I.: Algebraic theory of D-modules, Harvard course, 1982.
[Be3] Bernstein, I.: Algebraic theory of D-modules, IAS course 1997.
[BeLu] Bernstein, I., Lunts, V.: On non-holonomic irreducible D-modules. Invent. Math., 94 1988 223–243.
[Bj1] Björk, J.-E.: Rings of differential operators. 18th Scandinavian Congress of Mathematicians (Aarhus, 1980), pp.
202–222, Progr. Math., 11, Birkhuser, Boston, Mass., 1981.
[Bj2] Björk, J.-E.: Rings of differential operators. North-Holland Mathematical Library, 21. North-Holland Publishing
Co., Amsterdam-New York, 1979.
[Bo1] Borel, A.; Grivel, P.-P.; Kaup, B.; Haefliger, A.; Malgrange, B.; Ehlers, F.: Algebraic D-modules. Perspectives in
Mathematics, 2. Academic Press, Inc., Boston, MA, 1987.
[Br1] Brylinski, J.-L.: (Co)-homologie d’intersection et faisceaux pervers. Bourbaki Seminar, Vol. 1981/1982, pp. 129–
157, Astérisque, 92–93, Soc. Math. France, Paris, 1982.
[Br2] Brylinski, J.-L.: Transformations canoniques, dualité projective, théorie de Lefschetz, transformations de Fourier
et sommes trigonométriques. Géométrie et analyse microlocales. Astérisque No. bf 140-141 (1986), 3–134.
[BDK] Brylinski, J.-L.; Dubson, A.; Kashiwara, M.: Formule de l’indice pour modules holonomes et obstruction d’Euler
locale. C. R. Acad. Sci. Paris Sér. I Math., 293 (1981), no. 12, 573–576.
[BK] Brylinski, J.-L.; Kashiwara, M.: Kazhdan-Lusztig conjecture and holonomic systems. Invent. Math., 64 (1981),
no. 3, 387–410.
[BMV] Brylinski, J.-L.; Malgrange, B.; Verdier, J.-L.: Transformation de Fourier géométrique. I, II. C. R. Acad. Sci. Paris
Sér. I Math. 297 (1983), no. 1, 55–58, 303 (1986), no. 5, 193–198.
[CG] Chriss, N.; Ginzburg, V.: Representation theory and complex geometry. Birkhäuser Boston, Inc., Boston, MA,
1997.
[De] Deligne, P.: Le theoreme de Noether. Groupes de monodromie en geometrie algebraique, SGA 7 II, 328–340,
Lecture Notes in Math., 340, Springer, Berlin-New York, 1973.
[EGA IV] Dieudonné, J.; Grothendieck. A.: Eléments de Géométrie Algébrique IV, Étude locale des schémas et des mor-
phishmes de schémas, Publ. Math. IHES, 20, 1964.
[Di] Dixmier, J.: Enveloping algebras. Revised reprint of the 1977 translation. Graduate Studies in Mathematics, 11.
American Mathematical Society, Providence, RI, 1996.
[DV] Douady, A.; Verdier, J.-L.: Séminaire de Géométrie Analytique. Tenu á l’École Normale Supérieure, Paris, 1974–75.
Astérisque, No. 36-37. Société Mathématique de France, Paris, 1976.
[Eis] Eisenbud, . Commutative Algebra.

105
[Fu] Fulton, W.: Intersection theory. Ergebnisse der Mathematik und ihrer Grenzgebiete (3), 2. Springer-Verlag, Berlin-
New York, 1984.
[FM] Fulton, W.; MacPherson, R.: Categorical framework for the study of singular spaces.Mem. Amer. Math. Soc. 31
(1981), no. 243,
[Ga1] Gabber, O.: The integrability of the characteristic variety. Amer. J. Math. 103 (1981), no. 3, 445–468.
[GS] Gelfand, I.; Shilov, G.: Generalized functions, Vol. 1. Academic Press, 1964.
[Gi1] Ginzburg, V.: G-modules, Springer’s representations and bivariant Chern classes. Adv. in Math. 61 (1986), no. 1,
1–48.
[Gon] Gonzalez-Sprinberg, G.: L’obstruction locale d’Euler et le théorme de MacPherson. , pp. 7–32, Astérisque, 83–83,
Soc. Math. France, Paris, 1981.
[GM1] Goresky, M.; MacPherson, R.: Stratified Morse theory. Ergebnisse der Mathematik und ihrer Grenzgebiete (3), 14.
Springer-Verlag, Berlin-New York, 1988.
[GM2] Goresky, M.; MacPherson, R.: Intersection homology. II. Invent. Math. 72 (1983), no. 1, 77–129.
[GH] Griffiths, P.; Harris, J.: Principles of algebraic geometry. Reprint of the 1978 original. Wiley Classics Library. John
Wiley & Sons, Inc., New York, 1994.
[GS] Guillemin, V.; Sternberg, S.: Geometric asymptotics. Mathematical Surveys, No. 14. American Mathematical
Society, Providence, R.I., 1977.
[GQS] Guillemin, V.; Quillen, D.; Sternberg, S.: The integrability of characteristics. Comm. Pure Appl. Math. 23 no. 1
(1970), 39–77.
[Ha1] Hartshorne, R.: Algebraic geometry. Graduate Texts in Mathematics, No. 52. Springer-Verlag, New York-
Heidelberg , 1977.
[Ha2] Hartshorne, R.: Residues and duality. Lecture notes of a seminar on the work of A. Grothendieck, given at Harvard
1963/64. With an appendix by P. Deligne. Lecture Notes in Mathematics, No. 20 Springer-Verlag, Berlin-New York
1966.
[Hi] Hironaka, H.: Resolution of singularities of an algebraic variety over a field of characteristic zero. I, II.Ann. of
Math. (2) 79 (1964), 109–203; ibid. 79 1964, 205–326.
[Ka1] Kashiwara, M.: Index theorem for a maximally overdetermined system of linear differential equations. Proc. Japan
Acad., 49 (1973), 803–804.
[Ka2] Kashiwara, M.: B-functions and holonomic systems. Rationality of roots of b-functions. Invent. Math., 38
(1976/77), no. 1, 33–53.
[Ka3] Kashiwara, M.: On the holonomic systems of linear differential equations. II. Invent. Math. 49 (1978), no. 2,
121–135.
[Ka4] Kashiwara, M.: Systems of microdifferential equations. Based on lecture notes by Teresa Monteiro Fernandes
translated from the French. Progress in Mathematics, 34. Birkhäuser Boston, Inc., Boston, MA, 1983.
[Ka5] Kashiwara, M.: Vanishing cycle sheaves and holonomic systems of differential equations. Algebraic geometry
(Tokyo/Kyoto, 1982), 134–142, Lecture Notes in Math., 1016, Springer, Berlin-New York, 1983.
[Ka6] Kashiwara, M.: The Riemann-Hilbert problem for holonomic systems. Publ. Res. Inst. Math. Sci.20 (1984), no. 2,
319–365.
[Ka7] Kashiwara, M.: The characteristic cycle and the index formula for constructible sheaves, Preprint Res. Inst. Math.
Sci., 1983.
[KaKa1] Kashiwara, M.; Kawai, T.: On the characteristic variety of a holonomic system with regular singularities. Adv. in
Math., 34 (1979), no. 2, 163–184.
[KaKa2] Kashiwara, M.; Kawai, T.: On holonomic systems of microdifferential equations. III. Systems with regular singu-
larities. Publ. Res. Inst. Math. Sci., 17 (1981), no. 3, 813–979.
[KaKa3] Kashiwara, M.; Kawai, T.: Second-microlocalization and asymptotic expansions. Complex analysis, microlocal
calculus and relativistic quantum theory pp. 21–76, Lecture Notes in Phys., 126, Springer, Berlin-New York, 1980.
[KS] Kashiwara, M.; Schapira, P.: Sheaves on manifolds. Grundlehren der Mathematischen Wissenschaften , 292.
Springer-Verlag, Berlin, 1990.
[KT] Kashiwara, M.; Tanisaki, T.: The characteristic cycles of holonomic systems on a flag manifold related to the Weyl
group algebra. Invent. Math. 77 (1984), no. 1, 185–198.
[KL1] Kazhdan, D.; Lusztig, G.: Representations of Coxeter groups and Hecke algebras. Invent. Math., 53 (1979), no. 2,
165–184.
[KL2] Kazhdan, D.; Lusztig, G.: Schubert varieties and Poincaré duality. Geometry of the Laplace operator, Proc.
Sympos. Pure Math., XXXVI, Amer. Math. Soc., Providence, R.I., 1980, pp. 185–203.
[La1] Laumon, G.: Sur la catégorie dérivée des D-modules filtrés. Algebraic geometry (Tokyo/Kyoto, 1982), 151–237,
Lecture Notes in Math., 1016, Springer, Berlin-New York, 1983.
[La2] Laumon, G.: Transformations canoniques et spécialisation pour les D-modules filtrés. Differential systems and
singularities (Luminy, 1983). Astérisque, 130 (1985), 56–129.

106
[Le] Levasseur, T.: Exposé of a lecture of Gabber, Paris 1982.
[Lu] Lunts, V.: Algebraic varieties preserved by generic flows. Duke Math. J. , 58 (1989), 531–554.
[Mcd] Macdonald, I. G.: Symmetric functions and Hall polynomials. Second edition. With contributions by A. Zelevinsky.
Oxford Mathematical Monographs. Oxford Science Publications. The Clarendon Press, Oxford University Press,
New York, 1995.
[McPh1] MacPherson, R.: Chern classes for singular algebraic varieties. Ann. of Math. (2), 100 (1974), 423–432.
[McPh2] MacPherson, R.: Global questions in the topology of singular spaces. Proceedings of the International Congress of
Mathematicians, Vol. 1, 2 (Warsaw, 1983), 213–235, PWN, Warsaw, 1984.
[MacV] MacPherson, R.; Vilonen, K.: Construction élémentaire des faisceaux pervers. C. R. Acad. Sci. Paris Sér. I Math.,
299 (1984), no. 10, 443–446.
[Ma1] Malgrange, B. L’involutivité des caractéristiques des systèmes différentiels et microdifférentiels. Séminaire Bour-
baki, (1977/78), Exp. No. 522, pp. 277–289, Lecture Notes in Math., 710, Springer, Berlin, 1979.
[Ma2] Malgrange, B.: Polynômes de Bernstein–Sato et cohomologie vanescente., 243–267, Astérisque, 101-102, Soc. Math.
France, Paris, 1983.
[Ma3] Malgrange, B.: Rapport sur les théorèmes d’indice de Boutet de Monvel et Kashiwara. 230–242, Astérisque, 101-
102, Soc. Math. France, Paris, 1983.
[Ma4] Malgrange, B.: Sur les images directes de D-modules. Manuscripta Math., 50 (1985), 49–71.
[MW] Marsden, J.; Weinstein, A.: Reduction of symplectic manifolds with symmetry. Rep. Mathematical Phys., 5 (1974),
no. 1, 121–130.
[Mt] Matsumura, H.: Commutative algebra. W.A. Benjamin, Inc. 1970.
[Me1] Mebkhout, Z.: Théorèmes de bidualité locale pour les DX -modules holonomes. Ark. Mat., 20 (1982), no. 1, 111–124.
[Me2] Mebkhout, Z. Une équivalence de catégories. Compositio Math., 51 (1984), no. 1, 51–62. Une autre équivalence de
catégories, ibid. 63-88.
[Qu] Quillen, D.: On the endormorphism ring of a simple module over an enveloping algebra. Proc. Amer. Math. Soc.,
21 1969 171–172.
[SKK1] Sato, M.; Kawai, T.; Kashiwara, M.: Microfunctions and pseudo-differential equations, in Hyperfunctions and
pseudo-differential equations pp. 265–529. Lecture Notes in Math., 287, Springer, Berlin, 1973.
[SKKO] Sato, M.; Kashiwara, M.; Kimura, T.; Oshima, T.: Microlocal analysis of prehomogeneous vector spaces. Invent.
Math., 62 (1980/81), no. 1, 117–179.
[Se] Serre, J.-P.: Faisceaux algébriques cohérents. Ann. of Math. (2), 61, (1955). 197–278.
[SGA6] Théorie des intersections et théorème de Riemann-Roch. (SGA 6). Dirig par P. Berthelot, A. Grothendieck et L.
Illusie. Lecture Notes in Mathematics, 225, 1971.
[Ve1] Verdier, J.-L.: Stratifications de Whitney et théorème de Bertini-Sard. Invent. Math., 36 (1976), 295–312.
[Ve2] Verdier, J.-L.: Categories derivées, État 00. SGA 4 12 , Lecture Notes in Mathematics, 569, Berlin–Heidelberg–
New-York: Springer 1977.
[Ve3] Verdier, J.-L. Spécialisation de faisceaux et monodromie modérée. Analysis and topology on singular spaces, II, III
(Luminy, 1981), 332–364, Astérisque, 101-102, Soc. Math. France, Paris, 1983.
[Ve4] Verdier, J.-L. Géométrie microlocale. Algebraic geometry (Tokyo/Kyoto, 1982), 127–133, Lecture Notes in Math.,
1016, Springer, Berlin-New York, 1983.
[Ve5] Verdier, J.-L.: Prolongement des faisceaux pervers monodromiques. Differential systems and singularities (Luminy,
1983). Astérisque, 130 (1985), 218–236.
[Wa] Wallach, N.: Real Reductive groups I.

107