Holonomic D-Modules and Positive Characteristic
Holonomic D-Modules and Positive Characteristic
Holonomic D-Modules and Positive Characteristic
4, 1–25 (2009)
DOI: 10.1007/s11537-009-0852-x
1.1. Quantization
In the algebraic setting there is no obvious analog of the Hilbert space, even
in the case of the cotangent bundle. A possible replacement for the notion of a
function is the one of a holonomic -module. Here we recall the definition and
several basic and well-known facts (the standard reference is [4]).
Let X be a smooth affine algebraic variety over field k of zero characteristic,
dim X = n. The ring (X) of differential operators is k-algebra of operators
acting on (X), generated by functions and derivations:
f → g f , f → ξ ( f ), g ∈ (X), ξ ∈ Γ(X, TX/Spec k ).
Algebra (X) carries the filtration (X) = k≥0 ≤k (X) by the degree of op-
erators, the associated graded algebra is canonically isomorphic to the algebra
of functions on T ∗ X. In geometric terms, the grading comes from the dilation
by Gm along the fibers of the cotangent bundle.
Let M be a finitely generated (X)-module, and choose a finite-dimensional
subspace V ⊂ M generating M. Then consider the filtration
M≤k := ≤k (X) ·V ⊂ M, k ≥ 0.
The associated graded module gr(M) is a finitely generated (T ∗ X)-module.
By theorem of Gabber, its support (which is a reduced conical subscheme in
T ∗ X)
supp (gr(M)) ⊂ T ∗ X
is coisotropic. In particular, the dimension of any irreducible component is at
least n = dim X. The support does not depend on the choice of the generating
space V , and is denoted by supp (M).
A finitely generated module M is called holonomic if and only if the dimen-
sion of its support is exactly n. Let (Li )i∈I be the set of irreducible components
Holonomic -modules and positive characteristic 3
1 Same formulas give the definition of algebra An,R for arbitrary commutative ring R.
4 M. Kontsevich
Here the inductive limit is taken over the filtered system consisting of all
finitely generated subrings R ⊂ R, and the index p runs over primes 2, 3, 5, . . ..
It is easy to see that the ring R∞ is defined over Q (all primes are invertible in
R∞ ), and the obvious map R → R∞ gives an inclusion R ⊗ Q → R∞ . Also, there
is a universal Frobenius endomorphism given by
where the limit is taken over pairs consisting of a finitely generated subring
R ⊂ k and an affine scheme of finite type S/ R endowed with a constructible
over k bijection between S ×Spec R Spec k and S. The scheme SR,p
is defined
as the pullback under Fr of the scheme S × Spec Z/pZ. Finally,
Spec (R⊗Z/pZ),p
the abbreviation C.S. means the set of constructible sections of a morphism of
schemes of finite type.
Notice that we automatically identify collections of constructible sections
which differ only at a finite set of primes. The reason is that R = k contains Q
and we can always add to R inverses of any finite set of primes.
In the special case S = A1k we have an embedding
k∞ → (A1k )tw
∞.
For an ind-constructible set over k represented as a countable limit of con-
structible sets S = lim Si we define S∞
tw as the inductive limit of sets (S )tw .
i ∞
−→
Conjecture 1. The degree of supp p,v (MR ) is bounded above by const · pr where
r = dim supp (M) − dim X.
We have a good evidence for this conjecture in the case of cyclic (X)-modules
of the form M = (X)/ (X)·P where P = 0 is a non-zero differential operator,
see section 3.2 for the special case X = A1k and section 5.1 for X = Ank , n > 1.
In the case of holonomic -modules the conjecture implies that the degree
is uniformly bounded. Also, one can expect an analog of Gabber theorem:
Conjecture 2. In the above notation, for holonomic M the support supp p,v MR is
Lagrangian for sufficiently large p and any v.
Let us assume the above conjecture, and let Li be the generic point of an irre-
ducible component of supp p,v (MR ). Then the length of
Mv := MR ⊗R kv
at Li is divisible by pdim X . Hence, we have an effective algebraic cycle on T ∗ Xv
where Xv
:= XR
p ×Spec R p Spec kv given by
lengthLi Mv
p,v (MR ) := ∑
suppnum [Li ].
i pdim X
Let us use the notation form section 2.1, and assume that module M is holo-
nomic. Let us assume3 conjectures 1 and 2. Then there exists prime p0 such
that for any p ≥ p0 and any v ∈ Spec R p we have an effective Lagrangian cycle
∗
p,v (MR ) in T Xv . Let us replace R by its localization obtained by inverting
suppnum
all primes < p0 .
It is easy to see that the definition is consistent, i.e., that for any model the
collection of cycles comes from a collection of constructible maps for all suffi-
ciently large primes, and that models form a filtered system.
The main advantage of the arithmetic support is that it gives a more elaborate
signature of a holonomic module, and the Lagrangian cycle is no longer conical
3 In fact, conjectures 1, 2 are very plausible statements which should not be hard to prove. On
the contrary, all the further conjectures made in the paper seem to really deserve their name, and
need some new ideas.
8 M. Kontsevich
There are two cases when one can easily calculate the arithmetic support.
First, for any F ∈ (X) we have an associated holonomic X -module given by
the trivial line bundle X endowed with the flat connection
∇ = d + (dF ∧ ·).
One can think about this X -module as exp(F) · X . We claim that the corre-
sponding arithmetic support is the pullback by the universal Frobenius Frk∞ of
the graph of differential form −dF. This follows easily from the identity
d dG p d p dG p
+ = +
dx dx dx dx
which is held in (R[x]) for any ring R over Z/pZ and any element G ∈ R[x]
(see proposition 3 in [3]). In particular, for the case X = Ank and F polynomial of
degree ≤ 2, the support is the pullback by the Frobenius of the affine Lagrangian
∗
subspace in A2n k = T X corresponding to F.
One can also calculate the arithmetic support for the connection on the trivial
bundle corresponding to a closed but not exact 1-form. For example, for X =
Spec k[x, x−1 ] and for the connection on := X given by 1-form λ dx/x for
some λ ∈ k, the arithmetic support is the curve in T ∗ X given by the equation
Here x(p) , y(p) are coordinates on T ∗ X p
⊂ T ∗ A1k with symplectic form dx(p) ∧
dy(p) . In the case λ ∈ Q the expression (λ p − λ ) (mod p) vanishes for all suf-
ficiently large p, hence the arithmetic support of this X -module is just the zero
section of T ∗ X p
. If λ ∈
/ Q then the arithmetic support is not equal to the zero
section, as follows from Chebotarev density theorem in the case when λ is al-
gebraic, and by elementary reasons when λ is transcendental.
Finally, let M be a holonomic X -module corresponding to the vector bundle
on X endowed with a flat connection ∇ of Gauss–Manin type (for variations
of pure motives). This means that ( , ∇) is a subquotient of the natural con-
nection of the bundle of de Rham cohomology of fibers of a smooth projective
morphism Y → X. Then by a classical result of N. Katz (see [8]) the p-curvature
is nilpotent. Hence the support is the zero section of the cotangent bundle, taken
with the multiplicity equal to rank .
3. One-dimensional case
Let us consider the case X = A1k , char(k) = 0. The algebra (X) is the first Weyl
algebra A1,k , we denote its generators by x̂ = x and ŷ = d/dx. We consider holo-
nomic (A1k )-module which is a non-trivial cyclic module (A1k )/ (A1k ) · P,
where
P = ∑ ai j xi (d/dx) j
i+ j≤N
is a non-zero differential operator on X. Let us fix a finitely generated subring
R ⊂ k containing all coefficients ai j . The center of A1,R p (recall R p := R⊗Z/pZ)
is the polynomial algebra R p [x̂ p , ŷ p ]. We extend it by adding central variables x̃, ỹ
satisfying
x̃ p = x̂ p , ỹ p = ŷ p .
The resulting extension of A1,R p is isomorphic to the matrix algebra
Mat(p × p, R p [x̃, ỹ]).
Indeed, this extension is the algebra over R p [x̃, ỹ] generated by two elements x̂, ŷ
satisfying the relations
[ŷ, x̂] = 1, x̂ p = x̃ p , ŷ p = ỹ p .
Shifted generators (x̂ − x̃, ŷ − ỹ) satisfy the same relations as the operators x and
d/dx in the truncated polynomial ring Z/pZ [x]/(x p ):
x p = 0, (d/dx) p = 0, [d/dx, x] = 1.
For example, for p = 5 the corresponding matrices are
⎛ ⎞ ⎛ ⎞
00000 01000
⎜1 0 0 0 0⎟ ⎜0 0 2 0 0⎟
⎜ ⎟ ⎜ ⎟
X p = ⎜ 0 1 0 0 0 ⎟ , Yp = ⎜
⎜ ⎟ ⎟
⎜0 0 0 3 0 ⎟.
⎝0 0 1 0 0⎠ ⎝0 0 0 0 4⎠
00010 00000
12 M. Kontsevich
Mp := ∑ ai j X pi ·Ypj
i+ j≤N
contains non-zero terms only at distance at most N from the main diagonal. We
are interested in its determinant for p N.
Consider the general situation: let M be a square matrix of size L × L (with
coefficients in a commutative ring) such that Mi j = 0 if |i − j| > N for some
N < L/2. For every integer i ∈ [N + 1, L] denote A(i) the square matrix of size
2N × 2N given by
⎧
⎨Mi−N,i if j1 = j2 − 1
(i)
A j1 , j2 = −Mi−N,i−2N+ j2 −1 if j1 = 2N
⎩
0 otherwise
The idea of the proof. Suppose that the matrix M is degenerate and all elements
Mi−N,i are non-zero for i ∈ [N + 1, L]. Hence the left hand side of the above
identity vanishes, and we want to prove that the right hand side vanishes, too.
Let us consider the sequence
(v j ) j=1,N+L := (0, . . . , 0, x1 , . . . , xL )
N times
14 M. Kontsevich
U (1) ∈ Im(B
)
U (2) = M1,N+1
−1
· A(N+1) U (1)
U (3) = M2,N+2
−1
· A(N+2)U (2)
···
U (L−N+1) = ML−N,L
−1
· A(L) U (L−N)
0 = BU (L−N) .
The above proposition allows to calculate p-determinants up to a simple
factor which has the form
p−N
prime p → ∏ f ( j) (mod p),
j=1
g.c.d.({d} ∪ {i | ai = 0}) = 1.
We identify such data factorizing by the free action of the group μd of roots 1
of order d
ξ ∈ k, ξ d = 1 acts as ai → ξ i ai .
The interpretation of the pair d, (ai )−d<i<d is as the truncated germ of a curve
d−1
x = x(t) = t d , y = y(t) = ∑ ait i + O(t −d ), t→∞
1−d
+ · · · + a1−d x−
d−1 d−1
y = ad−1 x d d + O(1/x).
y∼ ∑ cλ xλ
λ ∈Q, λ <1
Acting by the group GL(2, k) we obtain the description of the whole set Jk∞ .
There is a Z≥1 -valued function deg on Jk∞ , with the value equal to d in the above
notation.
Any planar curve C ∈ CurvesA2 gives a function νC : Jk∞ → Z≥0 with finite
k
support. Namely, we count with multiplicities all the germs of C intersecting P1k
at infinity. The degree of C coincides with the sum over Jk∞ of the product of νC
with deg.
We define a logarithmic family to be the set of all curves C with a given
function νC .
Here are examples in small degrees. First, we have the logarithmic family
consisting of the empty curve C with νC = 0. Next, we have one-point loga-
rithmic families each of which consists of a line in A2k . The simplest non-trivial
example is the family of hyperbolas xy = t where t ∈ k, including the degenerate
case t = 0.
It is easy to see that there is a natural action of Aut A2k on Jk∞ , and the decom-
position by logarithmic families is Aut A2k -equivariant. An intrinsic definition of
Jk∞ is as the inductive limit of the set of divisors where the volume form dx ∧ dy
has pole of order one, over the partially ordered set of smooth compactifications
of A2 on which the volume form does not vanish at infinity (i.e., it is a Poisson
compactification, compare with [9]). For any curve C ∈ CurvesA2 there exists
k
a Poisson compactification of A2k such that C intersects only those divisors at
infinity where the form has logarithmic pole. Also, If C1 and C2 are two curves
such that νC1 · νC2 = 0 (i.e., functions νC1 and νC2 have disjoint support), then
the intersection C1 ∩ C2 is finite and the intersection number [C1 ] ∩ [C2 ] can be
determined entirely in terms of functions νC1 and νC2 .
For a non-algebraically closed field k of zero characteristic we define log-
arithmic families using the embedding k → k to the algebraic closure. Also,
logarithmic families for curves of a given bounded degree can be defined for
positive characteristic if it is large enough.
There is an alternative meaning of Jk∞ in terms of singularities of holonomic
(Ak )-modules. Let M be such a module. We will associate with M a Z≥0 -
valued function νM on Jk∞ with finite support.
First, there exists a finite set S ⊂ k such that M|A1 \S is a vector bundle with
k
connection ∇. It can have irregular singularities at S and at ∞.
It is well-known (see e.g. [11]) that the category of bundles with connections
over the field of Laurent series k((z)) is decomposed into the direct sum of
blocks corresponding to Puiseux polynomials in negative powers of z:
F(z) = ∑ bλ zλ , bλ ∈ k, bλ = 0 for almost all λ ,
λ ∈Q<0
Theorem 1. For any non-zero differential operator P ∈ A1,k = (A1k ) with coef-
ficients in a finitely generated ring R ⊂ k, char(k) = 0, for all sufficiently large
p and for any point v ∈ Spec R over p, the multiplicities νM at point v where
M = A1,k /A1,k · P coincides with the (pullback by Frobenius of) multiplicities of
(p)
the curve given by the equation DP = 0 at point v.
The idea of the proof. First of all, it is easy to identify contributions of germs
of the line y = 0. Namely, in the case of a curve given by equation H(x, y) =
0, H = ∑i, j Hi j xi y j = 0 ∈ k[x, y], this multiplicity can be read from the Newton
polygon of P. Namely, the multiplicity is equal to
max{ j | Hi j = 0, ∀(i
, j
) Hi
j
= 0 =⇒ (i − j) ≥ (i
− j
)}
A similar description works for cyclic A1 -modules. For the multiplicities at
k
other points of Jk∞ one can apply automorphisms of the Weyl algebra, and also
take the tensor product with -modules corresponding to exponents of Puiseux
polynomials.
Finally, one can show that for two holonomic A1 -modules M1 , M2 such that
k
supports of νM1 and νM2 are disjoint, there is no non-trivial homomorphisms
from M1 to M2 , and the dimension of Ext1 (M1 , M2 ) coincides with the intersec-
tion number of the corresponding curves.
18 M. Kontsevich
We expect that in the case dim X > 1 also there exists a notion of a logarithmic
family of effective Lagrangian cycles in T ∗ X, and the arithmetic support should
always belong to such a family. In the special case when a Lagrangian cycle is
a smooth closed Lagrangian variety L ⊂ T ∗ X (taken with multiplicity one) we
expect a more clearer picture of what is the logarithmic family:
Definition 3. A smooth logarithmic family of smooth Lagrangian subvarieties
in T ∗ X is a pair (S, ) where S is a smooth variety over k and ⊂ T ∗ X × S
is a smooth closed submanifold such that its projection to S is smooth, all fibers
s, s ∈ S are Lagrangian, and the following property holds. For any s ∈ S the
natural map
Ts S → Γ(s , (TX )|s /Ts ) = Γ(s , T∗ s )
identifies Ts S with the space of 1-forms on s with logarithmic singularities4 .
supp νM ⊂ supp νM .
We see that the dimension of an isosingular family is greater than the dimension
of the corresponding logarithmic family. The difference between two dimen-
sions is the dimension of the Picard variety of any smooth compactification L of
L.
Informally speaking, holonomic X -modules correspond to Lagrangian sub-
manifolds in T ∗ X together with a line bundle on L (or something like that5 ).
More precisely, there should be the decomposition of the space of holonomic
modules into isosingular families, and the decomposition of Lagrangian sub-
manifolds with line bundles into logarithmic families, such that there is a canon-
ical one-to-one correspondence between families (equivalence classes) of both
kinds, and dimensions of the corresponding families coincide.
The importance of the ring Pk is that the arithmetic supports should be effective
Lagrangian cycles parametrized by Spec Pk . For countable k (e.g. for k = Q or
k = Q) the ring Pk is a countable subring in uncountable ring k∞ , similar to the
subalgebra of periods in C for algebraic varieties over Q.
The result from section 3.3 indicates that Pk should be related with the dif-
ferential Galois group of difference equations with coefficients in k.
Also, there are strong indications that Pk should be related somehow to an-
other group, the motivic Galois group of k. Indeed, for holonomic X -module
corresponding to the flat connection on the trivial bundle given by a closed but
not exact 1-form, the support is controlled by the Cartier operator, which is re-
lated to the the comparison of de Rham and crystalline cohomology of X (in
degree 1). Roughly speaking, one can expect that the motivic Galois group of k
in de Rham realization acts on the category of holonomic X -modules. The cor-
respondence between holonomic modules and their arithmetic support could be
related to torsors comparing de Rham and crystalline cohomology (and maybe
Hodge realization which is the associated graded with respect to the Hodge
filtration on de Rham cohomology). Fixed points of the motivic Galois group
should correspond to motivic (or, more generally, motivic-exponential) holo-
nomic X -modules. This fits well with the fact that supports of motivic and
exponential X -modules are defined over Fr∗k∞ (k) ⊂ k∞ .
For a given field k, char(k) = 0 we define extended motivic-exponential X -
modules on smooth algebraic varieties over k as the minimal class which is
closed under extensions, subquotients, pushforwards and pullbacks, and contain
all X -modules of type exp(F) · X for F ∈ (X).
One can ask what are the “most canonical” holonomic modules and correspond-
ing Lagrangian varieties6 .
The reason for the condition H 1 (L(C), Z) = 0 is that it guarantees that there is
no non-trivial local system of rank 1 over L. This condition can be reformulated
6 In a recent preprint [2] a related but different question was studied for X = A1k .
Holonomic -modules and positive characteristic 21
which after the transcendental change of variables xi = log yi has a more familiar
form
∂ 2
∑ ∂ xi + exp(xi+1 − xi ) .
i∈Z/nZ
where W (x) is a fixed polynomial of degree (n + 1) and (ci )i≤n−1 are parameters
of the curve. The full solution of the matrix model is based on the associated
family of holonomic A1 -modules. Also, Seiberg–Witten curves for N = 2 pure
k
SU(2) gauge theory, given by the equation
x + 1/x = y2 + u
where u is a parameter, form a logarithmic family of curves in T ∗ Gm endowed
with coordinates x, y (x = 0) and the symplectic form dx ∧ dy/x.
6. Trigonometric version
and it is an Azumaya algebra over its center, a twisted form of the algebra of
matrices of size N n × N n .
In the case n = 1 any element
P= ∑ ai j x̂i1 x̂2j
i, j∈Z, |i|+| j|≤d
References