a

r
X
i
v
:
m
a
t
h
/
0
7
0
2
2
0
6
v
2


[
m
a
t
h
.
A
G
]


3
0

M
a
y

2
0
0
7
Notes on motives in nite characteristic
Maxim Kontsevich
February 2, 2008
To Yuri Manin on the occasion of 70-th birthday, with admiration.
Introduction and an example
These notes grewfrom an attempt to interpret a formula of Drinfeld (see [3]) enu-
merating the absolutely irreducible local systems of rank 2 on algebraic curves
over nite elds, obtained as a corollary of the Langlands correspondence for
GL(2) in the functional eld case, and of the trace formula.
Let C be a smooth projective geometrically connected curve dened over a
nite eld F
q
, with a base point v C(F
q
). The geometric fundamental group

geom
1
(C, v) :=
1
(C
Spec Fq
SpecF
q
, v) is a pronite group on which the Galois
group

Z = Gal(F
q
/F
q
) (with the canonical generator Fr := Fr
q
) acts. In what
follows we will omit the base point from the notation.
Theorem 1. (Drinfeld) Under the above assumptions, for any integer n 1
and any prime l = char(F
q
) the set of xed points
X
(l)
n
:=
_
IrrRep(
1
(C
Spec Fq
Spec F
q
) GL(2, Q
l
))/conjugation
_
Fr
n
is nite. Here IrrRep(. . . ) denotes the set of conjugacy classes of irreducible
continuous 2-dimensional representations of
geom
1
(C) dened over nite exten-
sions of Q
l
. Moreover, there exists a nite collection (
i
) Q

of algebraic
integers, and signs (
i
) {1, +1} depending only on C, such that for any n, l
one has an equality
#X
(l)
n
=

n
i
.
From the explicit formula which one can extract from [3] one can see that
numbers
i
are q-Weil algebraic integers whose norm for any embedding Q C
belongs to q
1
2
Z
0
. Therefore, the number of elements of X
(l)
n
, n = 1, 2, . . . looks
like the number of F
q
n-points on some variety over F
q
. The largest exponent is
q
4g3
, which indicates that this variety has dimension 4g 3. A natural guess is
that it is closely related to the moduli space of stable bundles of rank 2 over C.
At least the dimensions coincide, and Weil numbers which appear are essentially
1
the same, they are products of the eigenvalues of Frobenius acting on the motive
dened by the rst cohomology of C.
The Langlands correspondence identies X
(l)
n
with the set of Q
l
-valued un-
ramied cuspidal automorphic forms for the adelic group GL(2, A
Fq(C)
). These
forms are eigenvectors of a collection of commuting matrices (Hecke operators)
with integer coecients. Therefore, for a given n 1 one can identify
1
all sets
X
(l)
n
for various primes l with one set X
n
endowed with an action of the absolute
Galois group Gal(Q/Q), extending the obvious actions of Gal(Q
l
/Q
l
) on X
(l)
n
.
These days the Langlands correspondence in the functional eld case is es-
tablished for all the groups GL(N) by L. Laorgue. To my knowledge, almost
no attempts were made to extend Drinfelds calculation to the case of higher
rank, or even to the GL(2) case with non-trivial ramication.
It is convenient to take the inductive limit X

:= lim

X
n
, X
n1
X
n1n2
which is an innite countable set endowed with an action of the product
2
Gal(Q/Q) Gal(F
q
/F
q
) .
The individual set X
n
can be reconstructed from this datum as the set of xed
points of Fr
n
Gal(F
q
/F
q
).
In spite of the numerical evidence, it would be too naive to expect a natural
identication of X

with the set of F

q
-points of an algebraic variety dened over
F
q
, as there is no obvious mechanism producing a non-trivial Gal(Q/Q)-action
on the latter.
The main question addressed here is
Question 1. Does there exist some alternative way to construct the set X

with the commuting action of two Galois groups?

In the present notes I will oer three dierent hypothetical constructions.
The rst construction comes from the analogy between the Frobenius acting
on
geom
1
(C) and an element of the mapping class group acting on the funda-
mental group of a closed oriented surface, the second one is almost tautological
and arises from the contemplation on the shape of explicit formulas for Hecke
operators (see an example in Section 0.1), the third one is based on an analogy
with lattice models in statistical physics.
I propose several conjectures, which should be better considered as guesses
in the rst and in the third part, as there is almost no experimental evidence
in their favor. In a sense, the rst and the third part should be regarded as
science ction, but even if the appropriate conjectures are wrong (as I strongly
suspect), there should be some grains of truth in them.
On the contrary, I feel quite condent that the conjectures made in the sec-
ond part are essentially true, the output is a higher-dimensional generalization
1
It is expected that all representations from X
(l)
n
are motivic, i.e. they arise from projectors
with coecients in Q acting on l-adic cohomology of certain projective varieties dened over
the eld of rational functions F
q
n(C).
2
One can replace Gal(Q/Q) by its quotient Gal(Q
qWeil
/Q) where Q
qWeil
Q is CM-
eld generated by all q-Weil numbers.
2
of the Langlands correspondence in the functional eld case. At the end of the
second part I will show how to make a step in the arithmetic direction, extending
the formulas to the case of an arbitrary local eld.
In the fourth part I will describe briey a similarity between a modication
of the category of motives based on non-commutative geometry, and two other
categories introduced in the second and the third part. Also I will make a
link between the proposal based on polynomial dynamics and the one based on
lattice models.
Finally, I apologize to the reader that the formulas in Sections 0.1 and 1.3
are given without explanations, this is the result of my poor knowledge of the
representation theory. The formulas were polished with the help of computer.
Acknowledgements: I am grateful to many people for useful discussions, es-
pecially to Vladimir Drinfeld (on the second part of this paper), to Laurent
Laorgue who told me about [3] and explained some basic stu about automor-
phic forms, to Vincent Laorgue who proposed an argument for the Conjecture
5 based on lifting, to James Milne for consultations about the category of mo-
tives over a nite eld, to Misha Gromov who suggested the idea to imitate
Dworks methods in the third part, also to Mitya Lebedev, Sasha Goncharov,
Dima Grigoriev, Dima Kazhdan, Yan Soibelman and Don Zagier. Also I am
grateful to the referee for useful remarks and corrections.
0.1 An explicit example
Here we will show explicit formulas for the tower (X
n
)
n1
in the simplest truly
non-trivial case. Consider the ane curve C = P
1
Fq
\ {0, 1, t, } for a given
element t F
q
\ {0, 1}. We are interested in motivic local SL(2)-systems on C
with tame non-trivial unipotent monodromies around all punctures {0, 1, t, }.
A lengthy calculation lead to the following explicit formulas
3
for the Hecke
operators for cuspidal representations. In what follows we assume char F
q
= 2.
The Hecke operators act on the spaces of functions on certain double coset space
for the adelic group, which can be identied with the set of equivalence classes of
vector bundles of rank 2 over P
1
Fq
together with a choice of one-dimensional sub-
spaces of bers at {0, 1, t, }. This double coset space is innite, but the eigen-
functions of Hecke operators corresponding to cuspidal representations have
nite support which one can bound a priori.
For any x F
q
the Hecke operator T
x
(on cuspidal forms) can be written
as an integral q q matrix whose rows and columns are labelled by elements of
F
q
, (i.e. T
x
Mat(F
q
F
q
; Z)). Coecients of T
x
are given by the formula
(T
x
)
yz
:= 2 #{w F
q
| w
2
= f
t
(x, y, z)} + (correction term)
where f
t
(x, y, z) is the following universal polynomial with integral coecients:
f
t
(x, y, z) := (xy +yz +zx t)
2
+ 4xyz(1 +t (x +y +z)) .
3
I was informed by V. Drinfeld that a similar calculation for the case of SL(2) local
systems on P
1
Fq
\ {4 points}, with tame non-trivial semisimple monodromy around punctures
was performed few years ago by Teruji Thomas.
3
The correction term is equal to

_
q + 1 x = y {0, 1, t}
1 x = y / {0, 1, t}
0 x = y
+
_

_
q if x / {0, 1, t} and
_

_
y =
t
x
, z = 0
y =
t x
1 x
, z = 1
y =
t(1 x)
t x
, z = t
0 otherwise
Operators (T
x
)
xFq
satisfy the following properties:
1. [T
x1
, T
x2
] = 0,
2.

xFq
T
x
= 1 = id
Z
Fq ,
3. T
2
x
= 1 for x {0, 1, t}, moreover {1, T
0
, T
1
, T
t
} form a group under the
multiplication, isomorphic to Z/2Z Z/2Z,
4. for any x / {0, 1, t} the spectrum of T
x
is real and belongs to [2

q, +2

q],
any element of Spec(T
x
) can be written as + where || =

q is a q-Weil
number,
5. for any = + Spec(T
x
) and any integer n 1 the spectrum of the
matrix T
(n)
x
corresponding to x F
q
F
q
n (if we pass to the extension
F
q
n F
q
) contains the element
(n)
:=
n
+
n
,
6. the vector space generated by {T
x
}
xFq
is closed under the product, the
multiplication table is
T
x
T
y
=

zFq
c
xyz
T
z
where c
xyz
= (T
x
)
yz
.
Typically (for generic t, x) the characteristic polynomial of T
x
splits into
the product of 4 irreducible polynomials of almost the same degree. The splitting
is not surprising, as we have a group
4
of order 4 commuting with all operators
T
x
(see property 3). Computer experiments indicate that the Galois groups of
these polynomials (considered as permutation groups) tend to be rather large,
typically the full symmetric groups if q is prime, and the corresponding number
elds have huge factors in the prime decomposition of the discriminant.
Notice that in the theory of automorphic forms one usually deals with in-
nitely many commuting Hecke operators corresponding to all places of the
global eld, i.e. to closed points of C (in other words, to orbits of Gal(F
q
/F
q
)
acting on C(F
q
)). Here we are writing formulas only for the points dened over
F
q
. The advantage of our example is that the number these operators coincides
with the size of Hecke matrices, hence one can try to write formulas for struc-
ture constants, which by luck turn out to coincide with the matrix coecients
of matrices T
x
(property 6).
4
This is the group of automorphisms of P
1
\ {4 points} for the generic cross-ratio.
4
1 First proposal: algebraic dynamics
As was mentioned before, it is hard to imagine a mechanism for a non-trivial
action of the absolute Galois group of Q on the set of points of a variety over
a nite eld. One can try to exchange the roles of elds Q and F
q
. The rst
proposal is the following one:
Conjecture 1. For a tower (X
n
)
n1
arising from automorphic forms (or from
motivic local systems on curves), as dened in the Introduction, there exists a
variety X dened over Q and a map F : X X such that there is a family of
bijections
X
n
(X(Q))
F
n
, n 1
covariant with respect to Gal(Q/Q) Z/nZ actions, and with respect to inclu-
sions X
n1
X
n1n2
for integers n
1
, n
2
1.
1.1 The case of GL(1)
Geometric class eld theory gives a description of the sets (X
n
)
n1
in terms of
the Jacobian of C:
X
n
= (Jac
C
(F
q
n))

(Q) = Hom(Jac
C
(F
q
n), Q

) .
The number of elements of this set is equal to
#Jac
C
(F
q
n) = det(Fr
n
H
1
(C)
1)
where Fr
H
1
(C)
is the Frobenius operator acting on, say, l-adic rst cohomology
group of C.
One can propose a blatantly non-canonical candidate for the corresponding
dynamical system (X, F). Namely, let us choose a semisimple (2g 2g) matrix
A = (A
i,j
)
1i,j2g
(where g is the genus of C) with coecients in Z, whose
characteristic polynomial is equal to the characteristic polynomial of Fr
H
1
(C)
.
Dene X/Q to be the standard 2g-dimensional torus G
2g
m
= Hom(Z
2g
, G
m
), and
the map F to be the dual to the map A : Z
2g
Z
2g
:
F(z
1
, . . . , z
2g
) = (

i
z
Ai,1
i
, . . . ,

i
z
Ai,2g
i
) .
Moreover, one can choose A in such a way that
F

= q where =
g

i=1
dz
i
z
i

dz
g+i
z
g+i
.
On the set of xed points of F
n
act simultaneously Gal(Q/Q) (via the cy-
clotomic quotient) and Z/nZ (by powers of F). Nothing contradicts to the
existence of an equivariant isomorphism between two towers of nite sets.
5
1.2 Moduli of local systems on surfaces
One can interpret the scheme G
2g
m
as the moduli space of rank 1 local systems
on a oriented closed topological surface S of genus g, the form is the natural
symplectic form on this moduli space.
In general, for any N 1, one can make an analogy between the action of
Frobenius Fr on the the set of l-adic irreducible representations of
geom
1
(C) of
rank N, and the action of the isotopy class of a homeomorphism : S S on
the set of irreducible complex representations of
1
(S) of the same rank. Sets
of representations are similar to each other, as it is known that the maximal
quotient of
geom
1
(C) coprime to q is isomorphic to the analogous quotient of
the pronite completion
1
(S) of
1
(S). Also, if we assume that there are only
nitely many xed points of acting on
IrrRep(
1
(S) GL(2, C))/conjugation
then the sets
X
(l)
:=
_
IrrRep(
1
(C
Spec Fq
SpecF
q
) GL(2, Q
l
))/conjugation
_

do not depend on the prime l for l large enough.

All this leads to the following conjecture (which is formulated a bit sloppily),
a strengthening of Conjecture 1:
Conjecture 2. For any smooth compact geometrically connected curve C/F
q
of genus g 2 there exists an endomorphism
C
of the tensor category of
nite-dimensional complex local systems on S such that

C
is algebraic and dened over Q, in the sense that it acts on the moduli
stack of irreducible local systems of any given rank N 1 by a rational
map dened over Q,

C
multiplies the natural symplectic form on the moduli space of irre-
ducible local systems of rank N by the constant q,
for every n, N 1 there exists an identication of the set of isomorphism
classes of irreducible motivic local systems of rank N on C
SpecFq
Spec F
q
invariant under Fr
n
, with the set of isomorphism classes of Q-local systems
of rank N on S invariant under
n
C
, compatible with the relevant Galois
symmetries and tensor constructions.
One can not expect that
C
comes from an actual endomorphism of the
fundamental group
1
(S), as it is known that for g 2 any such is necessarily
an automorphism. That is a rationale for replacing a putative endomorphism
of
1
(S) by a more esoteric endomorphism of the tensor category of its nite-
dimensional representations.
6
1.2.1 Example: SL(2)-local systems on sphere with 3 punctures
A generic SL(2, C)-local system on CP
1
\ {0, 1, } is uniquely determined by 3
traces of monodromies around punctures. A similar statement holds for l-adic
local systems with tame monodromy in the case of nite characteristic. Motivic
local systems correspond to the case when all the eigenvalues of the monodromies
around punctures are roots of unity, i.e. when the traces of monodormies are
twice cosines of rational angles. This leads to the following prediction:
X = A
3
, F(x
1
, x
2
, x
3
) = (T
q
(x
1
), T
q
(x
2
), T
q
(x
3
))
where T
q
Z[x] is the q-th Chebyshev polynomial,
T
q
( +
1
) =
q
+
q
.
In this case the identications between the xed points of F and motivic local
systems on P
1
Fq
\ {0, 1, } exist, and can be extracted form the construction
of these local systems (called hypergeometric) as summands in certain direct
images of abelian local systems (analogous the classical integral formulas for
hypergeometric functions). The identication is ambiguous, it depends on a
choice of a group embedding F

q
Q/Z.
1.3 Equivariant bundles and Ruelle-type zeta-functions
The analogy with an element of the mapping class group acting on surface S
suggest the following addition to the Conjecture 1. Let us x the curve C/F
q
and the rank N 1 of local systems under the consideration. For a given point
x C(F
q
) we have a sequence of Hecke operators T
(n)
x
associated with curves
C
Spec Fq
SpecF
q
n. The spectrum of T
(n)
x
is a Q-valued function on X
n
, i.e.
according to Conjecture 1, a function on the set of xed points of F
n
. We expect
that the collection of these functions for n = 1, 2 . . . comes from a F-equivariant
vector bundle on X.
Conjecture 3. Using the notations of Conjecture 1, for given x C(F
q
) there
exists a pair (E, g) where E is a vector bundle on X of rank N together with an
isomorphism g : F

E E (dened over Q), such that the eigenvalue of T

(n)
x
at
the point of its spectrum corresponding to z X
n
coincides with
Trace(E
z
= E
F
n
(z)
E
F(z)
E
z
)
where arrows are isomorphisms of bers of E coming from g.
In particular, one can ask for an explicit formula for the F-equivariant bundle
E in the case of SL(2)-local systems on the sphere with 3 punctures where we
have an explicit candidate for (X, F).
In the limiting most simple non-abelian case when the monodromy is unipo-
tent around 2 punctures, and arbitrary semisimple around the third puncture,
one can make the above question completely explicit:
7
Question 2. For a given x F
q
\ {0, 1}, does there exist a rational function
R = R
x
on CP
1
with values in q
1/2
SL(2, C) Mat(2 2, C) which has no
singularities on the set
_

n1
{z C|z
q
n
1
= 1}
_
\ {1}
such that for any n 1 two sets of complex numbers (with multiplicities):
X
n
:=
_
_
_

yF
q
n\{0,1,x}

_
y(1 xy)
1 y
_

: F

q
n C

, = 1
_
_
_
where runs through all non-trivial multiplicative characters of F
q
n, and
X

n
:=
_
Trace
_
R(z)R(z
q
) . . . R(z
q
n1
)
_
| z
q
n
1
= 1, z = 1
_
coincide?
Elements of the set X
n
are real numbers of the form + where Q is
a q-Weil number with || = q
1/2
. Therefore it is natural to expect that R(z)
belongs to q
1/2
SU(2) if |z| = 1.
The Galois symmetry does not forbid for the function R (as a rational func-
tion with values in (2 2)-matrices) to be dened over Q, after the conjuga-
tion by a constant matrix. Moreover, the existence of such a function over
Q leads to certain choice of generators of the multiplicative groups
_
F

q
n
_
for
all n 1, well-dened modulo the action of Frobenius Fr
q
(the Galois group
Gal(F
q
n/F
q
) = Z/nZ), as in a sense we identify roots of unity in C and multi-
plicative characters of F
q
n. In particular, there will be a canonical irreducible
polynomial over F
q
of degree n for every n 1. This is something almost too
good to be true.
1.3.1 Reminder: Trace formula and Ruelle-type zeta-function
Let X be now a smooth proper variety (say, over C), endowed with a map
F : X X, and E be a vector bundle on X together with a morphism (not
necessarily invertible) g : F

E E. Let us assume that for any n 1 all xed

points z of F
n
: X X are isolated and non-degenerate, i.e. the tangent map
(F
n
)

z
: T
z
X T
z
X
has no nonzero invariant vectors (in other words, all eigenvalues of (F
n
)

z
are
not equal to 1). Then one has the following identity (Atiyah-Bott xed point
formula):

vX:F
n
(z)=z
Trace(E
z
= E
F
n
(z)
E
z
)
det(1 (F
n
)

z
)
=
= Trace((g

)
n
: H

(X, E) H

(X, E))
8
The trace in the r.h.s. is understood in the super sense, as the alternating sum
of the ordinary traces in individual cohomology spaces.
If one wants to eliminate the determinant factor in the denominator in the
l.h.s., one should replace E by the superbundle E

(T

X
).
The trace formula implies that the series in t
exp
_
_

n1
t
n
n

zX:F
n
(z)=z
Trace(E
z
= E
F
n
(z)
E
z
)
det(1 (F
n
)

z
)
_
_
is the Taylor expansion of a rational function in t. It seems that in many cases
for non-compact varieties X a weaker form of rationality holds as well, when
no equivariant compactication can be found. Namely, the above series (called
the Ruelle zeta-function in general, not necessarily algebraic case) admits a
meromorphic continuation to C; also the zeta-function in the version without
the denominator is often rational in the non-compact case.
1.3.2 Rationality conjecture for motivic local systems
In the case hypothetically corresponding to motivic local systems on curves (in
the setting of Conjecture 3), one can make a natural a priori guess about the
denominator in the l.h.s. of the trace formula. Namely, for a xed point z of
the map F
n
corresponding to a xed point [] in the space of representations
of
1
(C
Spec Fq
Spec F
q
) in GL(N, Q
l
), we expect that the vector space T
z
X
together with the automorphism (F
n
)

z
should be isomorphic (after the change
of scalars) to
H
1
(C
Spec Fq
SpecF
q
, End()) = Ext
1
(, )
endowed with the Frobenius operator.
Eigenvalues of Fr
n
in this case have norm q
n/2
by the Weil conjecture, hence
not equal to 1, and the denominator in the Ruelle zeta-function does not vanish
(meaning that the xed points are non-degenerate).
In our basic example from Section 0.1 one can propose an explicit formula
for the denominator term. Dene (in notation from Section 0.1) for given t
F
q
\ {0, 1} a matrix T
tan
Mat(F
q
F
q
, Q) by the formula
T
tan
:=
1
q

xFq
(T
x
)
2
+ (q 3 1/q) id
Q
Fq .
This matrix satises the same properties as Hecke operators
5
. Namely, all
eigenvalues of T
tan
belong to [2

q, +2

q], any element of Spec(T

tan
) can be
written as + where || =

q is a q-Weil number, and for any = +
Spec(T
x
) and any integer n 1 the spectrum of the matrix T
(n)
tan
corresponding
to x F
q
F
q
n (if we pass to the extension F
q
n F
q
) contains the element

(n)
:=
n
+
n
.
5
The only dierence is that eigenvalues of operators Tx are algebraic integers while eigen-
values of Ttan are algebraic integers divided by q.
9
We expect that the eigenvalue of T
tan
at the point of the spectrum cor-
responding to motivic local system is equal to the trace of Frobenius in a
two-dimensional submotive of the motive H
1
(C, End()), corresponding to the
deformations of preserving the unipotency of the monodromy around punc-
tures.
Notice that any motivic local system on C can be endowed with a non-
degenerate skew-symmetric pairing with values in the Tate motive. This ex-
plains the main term of the formula:
the sum of squares of Hecke operators means that we are using the trace
formula for Frobenius in the cohomology of C with coecients in the
tensor square of ,
the factor 1/q comes from the Tate twist,
the minus sign comes from the odd (rst) cohomology.
The candidate for the denominator term in the putative Ruelle zeta-function
is the following operator commuting with the Hecke operators (we write the
formula only for the rst iteration, n = 1), considered as a function on the
spectrum:
D := (q + 1 T
tan
)
1
.
The reason is that the eigenvalue of D at the eigenvector corresponding to
motivic local system is equal to
1
(1 )(1 )
=
1
1 +q
where , are Weil numbers, eigenvalues of Frobenius in H
1
(C, End()) satis-
fying equations
+ = , + = q .
The l.h.s. of the putative trace formula for the equivariant vector bundle
E
x1
E
x
k
(here E
x
is the F-equivariant vector bundle corresponding to point
x C(F
q
), see Conjecture 3), is given (for the n-th iteration) by the formula
Trace
_
T
(n)
x1
. . . T
(n)
x
k
D
(n)
_
.
It looks that in order to achieve the rationality of the putative Ruelle zeta-
function one has to add by hand certain contributions corresponding to missing
xed points. For example, for any x F
q
\ {0, 1, t} one has
Trace(T
x
D) =
q
(q 1)
2
and the corresponding zeta-function
exp
_
_

n1
t
n
n
q
n
(q
n
1)
2
_
_
=

m1
(1 q
m
t)
m
Q[[t]]
10
is meromorphic but not rational. The above zeta-function looks like the contri-
bution of just one
6
xed point z
0
on an algebraic dynamical system z F(z)
on a two-dimensional variety, with the spectrum of (F

)
z0
equal to (q, q), and
the spectrum of the map on the ber E
z0
E
z0
equal to (q, 0). Here is the
precise conjecture coming from computer experiments:
Conjecture 4. For any x
1
, . . . , x
k
F
q
\ {0, 1, t}, k 1 the series
exp
_
_

n1
t
n
n
_
Trace
_
T
(n)
x1
. . . T
(n)
x
k
D
(n)
_
+ Corr(n, k)
_
_
_
where
Corr(n, k) :=
(1 q
n
)
k
(1 q
n
) (1 q
2n
)
,
is a rational function.
The rational function in the above conjecture should be an L-function of a
motive over F
q
, all its zeroes and poles should be q-Weil numbers.
Finally, if one considers Ruelle zeta-functions without the denominator term,
then rationality is elementary, as will become clear in the next section.
2 Second proposal: formalism of motivic func-
tion spaces and higher-dimensional Langlands
correspondence
The origin of this section is property 6 (the multiplication table) of Hecke op-
erators in our example from Section 0.1.
2.1 Motivic functions and the tensor category C
k
Let S be a noetherian scheme.
Denition 1. The commutative ring Fun
poor
(S) of poor mans motivic func-
tions
7
on S is the quotient of the free abelian group generated by equivalence
classes of schemes of nite type over S, modulo relations
[X S] = [Z S] + [(X \ Z) S]
where Z is a closed subscheme of X over S. The multiplication on Fun
poor
(S)
is given by the bered product over S.
6
Maybe the complete interpretation should be a bit more complicated as one can check
numerically that Trace(D) =
q
2
(q2)
(q1)
2
(q+1)
.
7
The name was suggested by V. Drinfeld.
11
In the case when S is the spectrum of a eld k, we obtain the standard
denition
8
of the Grothendieck ring of varieties over k. Any motivic function
on S gives a function on the set of points of S with values in the Grothendieck
rings corresponding to the residue elds.
For a given eld k let us consider the following additive category C
k
. Its
objects are schemes of nite type over k, the abelian groups of homomorphisms
are dened by
Hom
C
k
(X, Y ) := Fun
poor
(X Y ) .
The composition of two morphisms represented by schemes is given by the
bered product as below:
[B Y Z] [A X Y ] := [A
Y
B X Z]
and extended by additivity to all motivic functions. The identity morphism id
X
is given by the diagonal embedding X X X.
One can start from the beginning from constructible sets over k instead of
schemes. The category of constructible sets over k is a full subcategory of C
k
,
the morphism in C
k
corresponding to a constructible map f : X Y is given
by [X
(idX,f)
X Y ], the graph of f.
Finite sums (and products) in C
k
are given by the disjoint union.
We endow category C
k
with the following tensor structure on objects:
X Y := X Y
and by a similar formula on morphisms. The unit object 1
C
k
is the point
Spec(k). The category C
k
is rigid, i.e. for every object X there exists another
object X

together with morphisms

X
: X X

1,
X
: 1 X

X such
that both compositions:
X
idXX
X X

X
XidX
X, X

Xid
X

X X

id
X
X
X

are identity morphisms. The dual object X

in C
k
coincides with X, the duality
morphisms
X
,
X
are given by the diagonal embedding X X
2
.
As in any tensor category, the ring End
C
k
(1
C
k
) is commutative, and the
whole category is linear over this ring, which is nothing but the Grothendieck
ring of varieties over k.
2.1.1 Fiber functors for nite elds
If k = F
q
is a nite eld then there is an innite chain (
n
)
n1
of tensor functors
from C
k
to the category of nite-dimensional vector spaces over Q. It is dened
on objects by the formula

n
(X) := Q
X(F
q
n)
.
8
Usually one extends the Grothendieck ring of varieties by inverting the class [A
1
k
] of
the ane line, which is the geometric counterpart of the inversion of the Lefschetz motive
L = H
2
(P
1
k
) in the construction of Grothendieck pure motives. Here also we can do the same
thing.
12
The operator corresponding by
n
to a morphism [A XY ] has the following
matrix coecient with indices (x, y) X(F
q
n) Y (F
q
n):
#{a A(F
q
n) | a (x, y)} Z
0
Q
The functor
n
is not canonically dened for n 2, the ambiguity is the
cyclic group Z/nZ = Gal(F
q
n/F
q
) Aut(
n
).
2.1.2 Extensions and variants
The abelian group Fun
poor
(S) of poor mans motivic functions can (and prob-
ably should) be replaced by the K
0
group of the triangulated category Mot
S,Q
of constructible motivic sheaves (with coecients in Q) on S. Although the
latter category is not yet rigorously dened, one can envision a reasonable can-
didate for the elementary description of K
0
(Mot
S,Q
). This group should be
generated by equivalence classes of families of Grothendieck motives (with co-
ecients in Q) over closed subschemes of S, modulo a suitable equivalence
relation. Moreover, group K
0
(Mot
S,Q
) should be ltered by the dimension of
support, the associated graded group should be canonically isomorphic to the
direct sum over all points x S of K
0
groups of categories of pure motives (with
coecients in Q) over the residue elds
9
.
Similarly, one can extend the coecients of the motives from Q to any eld
of zero characteristic. This change will aect the group K
0
and give a dierent
algebra of motivic functions.
Finally, one can add formally images of projectors to the category C
k
.
Question 3. Are there interesting non-trivial projectors in C
k
?
I do not know at the moment any example of an object in the Karoubi closure
of C
k
which is not isomorphic to a scheme. Still, there are interesting non-trivial
isomorphisms between objects of C
k
, for example the following version of the
Radon transform.
2.1.3 Example: motivic Radon transform
Let X = P(V ) and Y = P(V

) be two dual projective spaces over k. We assume

that n := dimV is at least 3.
The incidence relation gives a subvariety Z X Y , which can be inter-
preted as a morphism in C
k
in two ways:
f
1
:= [Z X Y ] Hom
C
k
(X, Y )
f
2
:= [Z Y X] Hom
C
k
(Y, X)
The composition f
2
f
1
is equal to
[A
n2
] id
X
+ [P
n3
] [X pt X] .
9
I do not know how to ll all the details in the above sketch.
13
The reason is that the scheme of hyperplanes passing through points x
1
, x
2
X
is either P
n3
if x
1
= x
2
, or P
n2
if x
1
= x
2
. On the level on constructible sets
one has P
n2
= P
n3
A
n2
.
The rst term is the identity morphism multiplied by the (n 2)-nd power
of the Tate motive, while the second term is in a sense a rank 1 operator. It
can be killed after passing to the quotient of X by pt which is in fact a direct
summand in C
k
:
X pt (X \ pt) .
Here we have to choose a point pt X. Similar arguments work for Y , and as
the result we obtain an isomorphism (inverting the Tate motive)
X \ pt Y \ pt
in the category C
k
which is not a geometric isomorphism of constructible sets.
2.2 Commutative algebras in C
k
By denition, a unital commutative associative algebra A in the tensor category
C
k
is given by a scheme of nite type X/k, and two elements
1
A
Fun
poor
(X), m
A
Fun
poor
(X
3
)
(the unit and the product in A) satisfying the usual constraints of unitality,
commutativity and associativity.
The formula for the structure constants c
xyz
= (T
x
)
yz
of the algebra of Hecke
operators in our basic example (see 0.1) is given explicitly by counting points
on constructible sets depending constructibly on a point (x, y, z) X
3
where
X = A
1
, for any t k \ {0, 1} (one should replace factors q by bundles with
ber A
1
). Hence we obtain a motivic function on X
3
which gives the structure
of a commutative algebra on X for any t k \ {0, 1}, for arbitrary eld k.
A straightforward check (see Proposition 1 in Section 2.4 below for a closely
related statement) shows that this algebra is associative.
2.2.1 Elementary examples of algebras
The rst example of a commutative algebra is given by an arbitrary scheme
X (or a constructible set) of nite type over k. The multiplication tensor is
given by the diagonal embedding X X
3
, the unit is given by the identity
map X X. If k = F
q
is nite then for any n 1 the algebra
n
(X) is
the algebra of Q-valued functions on the nite set X(F
q
n), with the pointwise
multiplication.
The next example corresponds to the case when X is an abelian group scheme
(e.g. G
a
, G
m
, or an abelian variety). We dene the multiplication tensor
m
A
Fun
poor
(X
3
) as the graph of the multiplication morphism X X X.
Again, if k is nite then the algebra
n
(X) is the group algebra with coecients
in Q of the nite abelian group X(F
q
n). Its points in Q are additive (resp.
multiplicative) characters of k if X = G
a
(resp. X = G
m
).
14
Also, one can see that the algebra in C
k
corresponding to the group scheme
G
a
is isomorphic to the direct sum of 1
C
k
(corresponding to the trivial additive
character of k) and another algebra A

which can be thought as parameterizing

non-trivial additive characters of the eld, with the underlying scheme A
1
\{0}.
Finally, one can make quotients of abelian group schemes by nite groups
of automorphisms. For example, for G
a
endowed with the action of the antipodal
involution x x, the formula for the product for the corresponding algebra
is the sum of the following main term
[Z (A
1
)
3
], Z = {(x, y, z)| x
2
+y
2
+z
2
2(xy +yz +zx) = 0}
(the latter equation means that

x+

y =

z), and of certain correction terms.
Similarly, for the antipodal involution (x, w) (x, w) on the elliptic curve
E P
1
P
1
given by w
2
= x(x1)(xt) (with (, ) serving as zero for the
group law), the quotient is P
1
endowed with the multiplication law similar to one
from the example 0.1. The main term is given by the hypersurface f
t
(x, y, z) = 0
in the notation from Section 0.1. The spectrum of the corresponding algebra is
rather trivial, in comparison to our example. The dierence is that in Section 0.1
we consider the two-fold cover of (A
1
)
3
ramied at the hypersurface f
t
(x, y, z) =
0.
2.2.2 Categorication
One may wonder whether a commutative associative algebra A in C
k
(for general
eld k, not necessarily nite) is in fact a materialization of the structure of a
symmetric (or only braided) monoidal category on a triangulated category, i.e.
whether the multiplication morphism is the class in K
0
of a bifunctor dening
the monoidal structure. The category under consideration should be either the
category of constructible mixed motivic sheaves on the underlying scheme of A,
or some small modication of it not aecting the group K
0
(e.g. both categories
could have semi-orthogonal decompositions with the same factors).
2.3 Algebras parameterizing motivic local systems
As we noticed already, the Example 0.1 can be interpreted as a commutative
associative algebra in C
k
parameterizing in a certain sense (via the chain of
functors (
n
)
n1
) motivic local systems on a curve over k = F
q
. Here we will
formulate a general conjecture, which goes beyond the case of curves.
2.3.1 Preparations on ramication and motivic local systems
Let Y be a smooth geometrically connected projective variety over a nitely
generated eld k. Let us denote by K the eld of rational functions on X and
by K

the eld of rational functions on Y

:= Y
Spec k
Spec k. We have an
exact sequence
1 Gal(K/K

) Gal(K/K) Gal(k/k) 1
15
For a continuous homomorphism
: Gal(K/K

) GL(N, Q
l
)
where l = char(k), which factorizes through the quotient
geom
1
(U) for some
open subscheme U Y

one can envision some notion of ramication divisor

(similar to the notion of the conductor in one-dimensional case) which should
be an eective divisor on Y

.
One expects that for a pure motive of rank N over K with coecients in
Q, the ramication divisor of the corresponding l-adic local system does not
depend on prime l = char(k), at least for large l.
Denote by IrrRep
Y

,N,l
the set of conjugacy classes of irreducible representa-
tions : Gal(K/K) GL(N, Q
l
) factorizing through
geom
1
(U) for some open
subscheme U Y

as above. The Galois group Gal(k/k) acts on this set.

Denote by IrrRep
mot,geom
Y,N
the set of equivalence classes of pure motives in
the sense of Grothendieck (dened using the numerical equivalence) of rank N
over K, with coecients Q, which are absolutely simple (i.e. remain simple
after the pullback to K

), modulo the action of the Picard group of rank 1

motives over k with coecients in Q. This set is endowed with a natural action
of Gal(Q/Q). The superscript geom indicates that we are interested only in
representations of the geometric fundamental group.
One expects that the natural map from IrrRep
mot,geom
Y,N
to the set of xed
points (IrrRep
Y

,N,l
)
Gal(k/k)
is a bijection. In particular, it implies that one can
dene the ramication divisor for an element of IrrRep
mot,geom
Y,N
. Presumably,
one can give a purely geometric denition of it, without referring to l-adic
representations.
2.3.2 Conjecture on algebras parameterizing motivic local systems
Conjecture 5. For a smooth projective geometrically connected variety Y over
a nite eld k = F
q
, an eective divisor D on Y , and a positive integer N, there
exists a commutative associative unital algebra A = A
Y,D,N
in the category C
k
satisfying the following property:
For any n 1 the algebra
n
(A) over Q is semisimple (i.e. it is a nite di-
rect sum of number elds) and for any prime l, (l, q) = 1 there exists a bijection
between Hom
Qalg
(
n
(A), Q) and the set of elements of IrrRep
mot,geom
Y
Spec Fq
Spec F
q
n,N
for which the ramication divisor is D. Moreover, the above bijection is equiv-
ariant with the respect to the natural Gal(Q/Q) Z/nZ-action.
One can also try to formulate a generalization of the above conjecture, allow-
ing not an individual variety Y but a family, i.e. a smooth projective morphism
Y B to a scheme of nite type over k, with geometrically connected bers,
together with a at family of ramication divisors. The corresponding algebra
should parameterize choices of a point b B(F
q
n) and a irreducible motivic sys-
tem of given rank and a given ramication on the ber Y
b
. This algebra should
map to the algebra of functions with the pointwise product (see 2.2) associated
with the base B.
16
In the above conjecture we did not describe how to associate a tower of
nite sets to the algebra A, as a priori we have just a sequence of nite sets
X
n
:= Hom
Qalg
(
n
(A), Q) without no obvious maps between them. This
leads to the following
Question 4. Which property of an associative commutative algebra A in C
Fq
gives naturally a chain of embeddings
Hom
Qalg
(
n1
(A), Q) Hom
Qalg
(
n1n2
(A), Q)
for all integers n
1
, n
2
1 ?
It looks that this holds automatically, by a kind of trace morphism.
2.3.3 Arguments in favor, and extensions
First of all, there is a good reason to believe that Conjecture 5 holds for curves.
Also, it would be reasonable to consider local systems with an arbitrary structure
group G instead of GL(N). The algebra parameterizing motivic local system
on curve Y = C with structure group G should be (roughly) equal to some
nite open part of the moduli stack Bun
G
L of G
L
-bundles on C, where G
L
is
the Langlands dual group. The multiplication should be given by the class of a
motivic constructible sheaf on
(Bun
G
L)
3
= Bun
G
L Bun
G
L
G
L
which should be a geometric counterpart to the lifting of automorphic forms
corresponding to the diagonal embedding
G
L
G
L
G
L
.
Presumably, the multiplication law from Example 0.1 corresponds to the lifting.
If we believe in the Conjecture 5 in the case of curves, then it is very natural
to believe in it in general. The reason is that for a higher-dimensional variety
Y (not necessarily compact) there exists a curve C Y such that
geom
1
(Y )
is a quotient of
geom
1
(C). Such a curve can be e.g. a complete intersection
of ample divisors, the surjectivity is a particular case of the Lefschetz theorem
on hyperplane sections. Therefore, the set of equivalence classes of absolutely
irreducible motivic local systems on Y
Spec Fq
SpecF
q
n should be a subset of the
corresponding set for C for any n 1, and invariant under Gal(Q/Q)-action as
well. It looks very plausible that such a collection of subsets should arise from
a quotient algebra in C
Fq
.
From the previous discussion it looks that the motivic local systems in
higher-dimensional case are less interesting, the 1-dimensional case is the
richest one. Nevertheless, there is denitely a non-trivial higher-dimensional
information about local systems which can not be reduced to 1-dimensional
data. Namely, for any motivic local system
arith
on Y and an integer i 0
the cohomology space
H
i
(Y

, )
17
where is the pullback of
arith
to Y

, is a motive over the nite eld k = F

q
.
We can calculate the trace of N-th power of Frobenius on it for a given N 1,
and get a Q-valued function
10
on the set
X
n
:= Hom
Qalg
(
n
(A), Q)
for each n 1. This leads to a natural addition to Conjecture 5. Namely, we
expect that systems of Q-valued functions on X
n
associated with higher coho-
mology spaces arise from elements in Hom
C
Fq
(1, A) (i.e. from motivic functions
on the constructible set underlying algebra A).
More generally, one can expect that the motivic constructible sheaves with
some kind of boundedness will be parametrized by commutative algebras.
Formulas from the example from Section 0.1 make sense and give an algebra
in C
k
for arbitrary eld k. This leads to
Question 5. Can one construct algebras A
Y,D,N
for arbitrary ground eld k,
not necessarily nite? In what sense will these algebras parameterize motivic
local system?
In general, it seems that the natural source of commutative algebras in C
k
is not the representation theory, but (quantum) algebraic integrable systems.
2.4 Towards integrable systems over local elds
Here we will describe briey an analog of commutative algebras of integral
operators as above for arbitrary local elds, i.e. R, C, or nite extensions of Q
p
or F
p
((x)). Let us return to our basic example. The check of the associativity
of the multiplication law given by formula from Section 0.1 in the case of nite
elds is reduced to an identication of certain varieties. The most non-trivial
part is the following
Proposition 1. For generic parameters t, x
1
, x
2
, x
3
, x
4
the two elliptic curves
E : f
t
(x
1
, x
2
, y) = w
2
12
, f
t
(y, x
3
, x
4
) = w
2
34

E : f
t
(x
1
, x
3
, y) = w
2
13
, f
t
( y, x
2
, x
4
) = w
2
24
given by equations in variables (y, w
12
, w
34
) and ( y, w
13
, w
24
) respectively, are
canonically isomorphic over the ground eld. Moreover, one can choose such an
isomorphism which identies the abelian dierentials
dy
w
12
w
34
and
d y
w
13
, w
24
.
In fact, it is enough to check the proposition over an algebraically closed
eld and observe that the curves E,

E have points over the ground eld
11
.
10
Here there is a small ambiguity which should be resolved somehow, as one can multiply

arith
by a one-dimensional motive over k with coecients in Q.
11
Curve E has 16 rational points with coordinate y {0, 1, t, }, same for

E.
18
Let now k be a local eld. For a given t k\{0, 1} we dene a (non-negative)
half-density c
t
on k
3
by the formula
c
t
:=

_
|dx
1
|
1/2
|dx
2
|
1/2
|dx
3
|
1/2
|w|
_
where
: Z(k) A
3
(k), (x
1
, x
2
, x
3
, w) = (x
1
, x
2
, x
3
)
is the projection of the hypersurface
Z A
4
k
: f
t
(x
1
, x
2
, x
3
) = w
2
.
We will interpret c
t
as a half-density on (P
1
(k))
3
as well.
One can deduce from the above Proposition the following
Theorem 2. The operators T
x
, x k \ {0, 1, t} on the Hilbert space of half-
densities on P
1
(k), given by
T
x
()(y) =
_
zP
1
(k)
c
t
(x, y, z) (z)
are commuting compact self-adjoint operators.
Moreover, in the non-archimedean case one can show that the joint spectrum
of commuting operators as above is discrete and consists of densities locally
constant on P
1
(k) \ {0, 1, t, }. In particular, all eigenvalues of operators T
x
are algebraic complex numbers. Passing to the limit over nite extensions of k
we obtain a countable set upon which acts
Gal(Q/Q) Gal(k/k) .
Also notice that in the case of local elds the formula is much simpler then
the motivic one, there is no correction terms. On the other hand, one has a
new ingredient, the local density of an integral operator. In general, one can
imagine a new formalism
12
where the structure of an algebra is given by data
(X, Z, , ) where X is a (birational type of) variety over a given eld k, Z is
another variety, : Z X
3
is a map (dened only at the generic point of Z),
and is a rational section of line bundle K
2
Z

(K
1
X
3
). If k is a local eld
then the pushdown by of ||
1/2
is a half-density on X
3
. The condition of the
associativity would follow from a property of certain data formulated purely in
terms of birational algebraic geometry.
Presumably, the spectrum for the case of the nite eld is just a low fre-
quency part of much larger spectrum for p-adic elds, corresponding to some
mysterious objects
13
.
The commuting integral operators in the archimedean case k = R, C are
similar to ones found recently in the usual quantum algebraic integrable systems,
see [5].
12
A somewhat similar formalism was proposed by Braverman and Kazhdan (see [1], who
had in mind orbital integrals in the usual local Langlands correspondence.
13
It looks that all this goes beyond motives, and on the automorphic side is related to some
kind of Langlands correspondence for two (or more)-dimensional mixed local-global elds.
19
3 Third proposal: lattice models
3.1 Traces depending on two indices
Let X be a constructible set over F
q
and M be an endomorphism of X in the
category C
Fq
(like e.g. a Hecke operator). What kind of object can be called
the spectrum of M?
Applying the functors
n
for n 1 we obtain an innite sequence of nite
matrices, of exponentially growing size. We would like to understand the be-
havior of spectra of operators
n
(M) as n +. A similar question arises
in some models in quantum physics where one is interested in the spectrum
of a system with nitely many states, with the dimension of the Hilbert space
depending exponentially on the number of particles.
Spectrum of an operator acting on a nite-dimensional space can be recon-
structed from traces of all positive powers. This leads us to the consideration
of the following collection of numbers
Z
M
(n, m) := Trace((
n
(M))
m
)
where n 1 and m 0 are integers. It will be important later to restrict atten-
tion only to strictly positive values of m, which mean that we are interested only
in non-zero eigenvalues of matrices
n
(M), and want to ignore the multiplicity
of the zero eigenvalue.
Observation 1. For a given n 1 there exists a nite collection of non-zero
complex numbers (
i
) such that for any m 1 one has
Z
M
(n, m) =

m
i
.
Observation 2. For a given m 1 there exists a nite collection of non-
zero complex numbers (
j
) and signs (
j
{1, +1}), such that for any n 1
one has
Z
M
(n, m) =

n
j
.
The symmetry between parameters n and m (modulo a minor dierence
with signs) is quite striking.
The rst observation is completely trivial. For a given n the numbers (
i
)
are all non-zero eigenvalues of the matrix
n
(M).
Let us explain the second observation. By functoriality we have
Z
M
(n, m) = Trace(
n
(M
m
)) .
Let us assume rst that M is given by a constructible set Y which maps to
X X:
Y X X, y (
1
(y),
2
(y)) .
Then M
m
is given by the consecutive bered product
Y
(m)
= Y
X
Y
X

X
Y Y Y
20
of m copies of Y :
Y
(m)
(F
q
) = {(y
1
, . . . , y
m
) (Y (F
q
))
m
|
2
(y
1
) =
1
(y
2
), . . . ,
2
(y
m1
) =
1
(y
m
)}
The projection to X X is given by (y
1
, . . . , y
m
) (
1
(y
1
),
2
(y
m
)). To take
the trace we should intersect Y
(m)
with the diagonal. The conclusion is that
Z
M
(n, m) is equal to the number of F
q
n-points of the constructible set

Y
(m)
:= Y
(m)

XX
X ,

Y
(m)
(F
q
) = {(y
1
, . . . , y
m
) Y
(m)
(F
q
)|
1
(y
1
) =
2
(y
m
)} .
The second observation is now an immediate corollary of the Weil conjecture
on numbers of points of varieties over nite elds
14
. The general case when M
is given by a formal integral linear combination

[Y

X X] can be
treated in a similar way.
3.2 Two-dimensional translation invariant lattice models
There is another source of numbers depending on two indices with a similar
behavior with respect to each of indices when another one is xed. It comes
from the so-called lattice models in statistical physics. A typical example is
the Ising model. There is a convenient way to encode Boltzmann weights of a
general lattice model on Z
2
in terms of linear algebra.
Denition 2. Boltzmann weights of a 2-dimensional translation invariant lat-
tice model are given by a pair V
1
, V
2
of nite-dimensional vector spaces over C
and a linear operator
R : V
1
V
2
V
1
V
2
.
Such data give a function (called the partition function) on a certain set of
graphs. Namely, let be a nite oriented graph whose edges are colored by
{1, 2} in such a way that for every vertex v there are exactly two edges colored
by 1 and 2 with head v, and also there are exactly two edges colored by 1 and
2 with tail v. Consider the tensor product of copies of R labelled by the set
V ert() of vertices of . It is an element v
R,
of the vector space
(V

1
V

2
V
1
V
2
)
V ert()
.
The structure of an oriented colored graph gives an identication of the above
space with
(V
1
V

1
)
Edge1()
(V
2
V

2
)
Edge2()
where Edge
1
(), Edge
2
() are the sets of edges of colored by 1 and by 2. The
tensor product of copies of the standard pairing gives a linear functional u

on
the above space. We dene the partition function of the lattice model on as
Z
R
() = u

(v
R,
) C .
14
Here we mean only the fact that the zeta-function of a variety over is rational in q
s
, and
not the more deep statement about the norms of Weil numbers.
21
An oriented colored graph as above is the same as a nite set with two
permutations
1
,
2
. The set here is V ert(), and permutations
1
,
2
correspond
to edges colored by 1 and 2 respectively.
In the setting of translation invariant 2-dimensional lattice models we are
interested in the values of the partition function only on graphs corresponding
to pairs of commuting permutations. Such a graph (if it is non-empty and
connected) corresponds to a subgroup Z
2
of nite index. We will denote
the partition function
15
of the graph corresponding to by Z
lat
R
().
Finally, Boltzmann data make sense in an arbitrary rigid tensor category
C. The partition function of a graph takes values in the commutative ring
End
C
(1). In particular, one can speak about super Boltzmann data for the
category Super
C
of nite-dimensional complex super vector spaces.
3.2.1 Transfer matrices
Let us consider a special class of lattices Z
2
depending on two parameters.
Namely, we set

n,m
:= Z (n, 0) Z (0, m) Z
2
.
Proposition 2. For any Boltzmann data (V
1
, V
2
, R) and a given n 1 there
exists a nite collection of non-zero complex numbers (
i
) such that for any
m 1 one has
Z
lat
R
(
n,m
) =

m
i
.
The proof is the following. Let us introduce a linear operator (called the
transfer matrix) by formula:
T
(2),n
:= Trace
V
n
1
((
n
id
V
n
2
) R
n
) End(V
n
2
)
where
n
End(V
n
1
) is the cyclic permutation. Here we interpret R
n
as an
element of
(V

1
)
n
(V

2
)
n
V
n
1
V
n
2
= End(V
n
1
) End(V
n
2
) .
It follows from the denition of the partition function that
Z
lat
R
(
n,m
) = Trace
_
T
(2),n
_
m
for all m 1. The collection (
i
) is just the collection of all non-zero eigenvalues
of T
(2),n
taken with multiplicities.
Similarly, one can dene transfer matrices T
(1),m
such that Z
lat
R
(
n,m
) =
Trace
_
T
(1),m
_
n
for all n, m 1. We see that the function (n, m) Z
lat
R
(
n,m
)
has the same two properties as the function (n, m) Z
M
(n, m) from Section
3.1. For super Boltzmann data one obtains sums of exponents with signs.
15
In physical literature it is called the partition function with periodic boundary conditions.
22
3.3 Two-dimensional Weil conjecture
Let us return to the case of an endomorphism M End
C
Fq
(X). In Section 3.1
we have dened numbers Z
M
(n, m) for n, m 1. Results of 3.2 indicate that one
should interpret pairs (n, m) as parameters for a special class of rectangular
lattices in Z
2
. A general lattice Z
2
depends on 3 integer parameters
=
n,m,k
= Z (n, 0) Z (k, m), n, m 1, 0 k < n .
Here we propose an extension of function Z
M
to all lattices in Z
2
:
Z
M
(
n,m,k
) := Trace((
n
(M))
m
(
n
(Fr
X
))
k
)
where Fr
X
End
C
Fq
(X) is the graph of the Frobenius endomorphism of the
scheme X. Notice that
n
(Fr
X
) is periodic with period n for any n 1.
Proposition 3. Function Z
M
on lattices in Z
2
dened as above, satisfy the
following property: for any two vectors
1
,
2
Z
2
such that
1

2
= 0 there
exists a nite collection of non-zero complex numbers (
i
) and signs (
i
) such
that for any n 1 one has
Z
M
(Z
1
Z n
2
) =

n
i
.
In other words, the series in formal variable t
exp
_
_

n1
Z
M
(Z
1
Z n
2
) t
n
/n
_
_
is rational.
The proof is omitted here, well just indicate that it follows from the con-
sideration of the action of the Frobenius element and of cyclic permutations on
the (etale) cohomology of spaces

Y
(m)
introduced in Section 3.1.
Also, it is easy to see that the same property holds for the partition function
Z
lat
R
(
m,n,k
) for arbitrary (super) lattice models.
16
The analogy leads to a two-
dimensional analogue of the Weil conjecture (the name will be explained in the
next section):
Conjecture 6. For any endomorphism M End
C
Fq
(X) there exists super
Boltzmann data (V
1
, V
2
, R) such that for any Z
2
of nite index one has
Z
M
() = Z
lat
R
() .
16
In general, one can show that for any lattice model given by operator R, and for any
matrix A GL(2, Z) there exists another lattice model with operator R

such that for any

lattice Z
2
one has Z
lat
R
() = Z
lat
R

(A()).
23
Up to now there is no hard evidence for this conjecture, there are just a
few cases where one can construct a corresponding lattice model in an ad hoc
manner. For example, it is possible (and not totally trivial) to do that for the
case when X = A
1
Fq
and M is the graph of the map x x
c
where c 1 is an
integer.
The above conjecture means that one can see matrices
n
(M) as analogs of
transfer matrices
17
. In the theory of integrable models people are interested in
systems where the Boltzmann weights R depends non-trivially on a parameter
(spaces V
1
, V
2
do not vary), and the horizontal transfer matrices commute with
each other
[T
(2),n
(
1
), T
(2),n
(
2
)] = 0
because of Yang-Baxter equation. Theory of automorphic forms seems to pro-
duce families of commuting endomorphisms in category C
Fq
, which is quite anal-
ogous to the integrability in lattice models. There are still serious dierences.
First of all, commuting operators in the automorphic forms case depend on dis-
crete parameters whereas in the integrable model case they depend algebraically
on continuous parameters. Secondly, the spectrum of a Hecke operator in its
n-th incarnation (like T
(n)
x
in Section 0.1) has typically n-fold degeneracy, which
does not happen in the case of the usual integrable models with period n.
3.4 Higher-dimensional lattice models and a higher-dimensional
Weil conjecture
Let d 0 be an integer.
Denition 3. Boltzmann data of a d-dimensional translation invariant lattice
model are given by a collection V
1
, . . . , V
d
of nite-dimensional vector spaces
over C and a linear operator
R : V
1
V
d
V
1
V
d
.
Similarly, one can dene d-dimensional lattice model in an arbitrary rigid
tensor category. The partition function is a function on nite sets endowed with
the action of the free group with d generators. In particular, for abelian actions,
it gives a function Z
lat
R
() C on the set of subgroups of nite index in
Z
d
. Also, for any lattice
d1
Z
d
of rank (d 1) and a vector Z
d
such
that / Q
d1
, the function
n 1 Z
lat
R
(
d1
Z n)
is a nite sum of exponents. Analogously, for any d-dimensional lattice model
R and any integer n 1 there exists its dimensional reduction, periodic with
17
At least if one is interested in the non-zero part of spectra. In general, the size of the trans-
fer matrix depends on n as an exact exponent, while the size of n(M) is a nite alternating
sum of exponents.
24
period n in d-th coordinate, which is a (d 1)-dimensional lattice model R
(n)
satisfying the property
Z
R
(n)
(
d1
) = Z
R
(
d1
Z ne
d
) ,
d1
Z
d1
where e
d
= (0, . . . , 0, 1) Z
d
= Z
d1
Z is the last standard basis vector.
Conjecture 7. For any (d1)-dimensional lattice model (X
1
, . . . , X
d1
, M), d
1 in the category C
Fq
, there exists a d-dimensional super lattice model (V
1
, . . . , V
d
, R)
in Super
C
such that for any n 1 the numerical (d 1)-dimensional model

n
(M) gives the same partition function on the set of subgroups of nite index
in Z
d1
as the dimensional reduction R
(n)
.
In the case d = 1 this conjecture follows from the usual Weil conjecture.
Namely, a 0-dimensional Boltzmann data in C
F
k
is just an element
M End
C
Fq
(1) = End
C
Fq
(
i
X
i
)
of the Grothendieck group of varieties over F
k
(or of K
0
of the category of pure
motives over F
k
with rational coecients). The corresponding numerical lattice
models
n
(M) are just numbers, counting F
q
n-points in M. By the usual Weil
conjecture these numbers are traces of powers of an operator in a super vector
space, i.e. values of the partition function for 1-dimensional super lattice model.
Similarly, for d = 2 one gets the 2-dimensional Weil conjecture from the
previous section.
3.4.1 Evidence: p-adic Banach lattice models
Let K be a complete non-archimedean eld (e.g. a nite extension of Q
p
).
We dene a d-dimensional contracting Banach lattice model as follows. The
Boltzmann data consists of
2d countable generated K-Banach spaces V
in
1
, . . . , V
in
d
, V
out
1
, . . . , V
out
d
,
a bounded operator R
vertices
: V
in
1

. . .

V
in
d
V
out
1

. . .

V
out
d
,
a collection of compact operators R
edges
i
: V
out
i
V
in
i
, i = 1, . . . , d.
Such data again give a function on oriented graphs with colored edges, in the
denition one should insert operator R
edges
i
for each edge colored by index i, i =
1, . . . , d. In the case of nite-dimensional spaces (V
in
i
, V
out
i
)
i=1,...,d
we obtain the
same partition function as for a usual nite-dimensional lattice model. Namely,
one can set
R :=
_

d
i=1
R
edges
i
_
R
vertices
, V
i
= V
in
i
, i = 1, . . . , d
or, alternatively,

R := R
vertices

d
i=1
R
edges
i
_
,

V
i
:= V
out
i
, i = 1, . . . , d .
25
In particular, for any contracting Banach model one get a function
Z
lat
R
() K on the set of sublattices of Z
d
. This function satises the property
that for any lattice
d1
Z
d
of rank (d 1) and a vector Z
d
such that
/ Q
d1
, one has
Z
lat
R
(
n1
Z n) =

, n 1
where (

) is a (possibly) countable Gal(K/K)-invariant collection of non-zero

numbers in K (eigenvalues of transfer operators) whose norms tend to zero.
Similarly, one can dene super Banach lattice models.
Here we announce a result supporting Conjecture 7, the proof is a straight-
forward extension of the Dwork method for the proving of the rationality of
zeta-function of a variety over a nite eld.
Theorem 3. The Conjecture 7 holds if one allows contracting Banach super
lattice models over a nite extension of Q
p
where p is the characteristic of the
nite eld F
q
.
3.5 Tensor category A and the Master Conjecture
Let us consider the following rigid tensor category A. Objects of A are nite-
dimensional vector spaces over C. The set of morphisms Hom
A
(V
1
, V
2
) is dened
as the group K
0
of the category of nite-dimensional representations of the free
(tensor) algebra
T(V
1
V

2
) :=

n0
(V
1
V

2
)
n
.
A representation of the free algebra by operators in a vector space U is the same
as an action of its generators on U, i.e. a linear map
V
1
V

2
U U .
Using duality we interpret it as a map
V
1
U V
2
U .
The composition of morphisms is dened by the following formula on generators:
[V
1
U V
2
U] [V
2
U

V
3
U

]
is equal to
[V
1
(U U

) V
3
(U U

)]
where the expression in the bracket is the obvious composition of linear maps
V
1
U U

V
2
U U

V
3
U U

.
The tensor product in A coincides on objects with the tensor product in
V ect
C
, the same for the duality. The formula for the tensor product on mor-
phisms is an obvious one, we leave details to the reader.
Like in Section 2.1 (Question 3), we can ask the following
26
Question 6. Are there interesting non-trivial projectors in A?
18
We denote by A
kar
the Karoubi closure of A.
There exists an innite chain of tensor functors (
A
n
)
n1
from A to the
category of nite-dimensional vector spaces over C given by

A
n
(V ) := V
n
on objects, and by
[f : V
1
U V
2
U]
n
Trace
U
n((
n
id
V
n
2
f
n
) Hom
V ect
C
(V
n
1
, V
n
2
)
on morphisms, where
n
: U
n
U
n
is the cyclic permutation. The cyclic
group Z/nZ acts by automorphisms of
A
n
. Moreover, the generator of the cyclic
group acting on V
n
=
A
n
(V ) is the image under
A
n
of a certain central element
Fr
V
in the algebra of endomorphisms End
A
(V ). This Frobenius element is
represented by the linear map : V U V U where U := V and =
2
is the permutation. As in the case of C
Fq
, for any V the operator
A
n
(Fr
V
) is
periodic with period n.
Let us introduce a small modication A

of the tensor category A. Namely,

it will have the same objects (nite-dimensional vector spaces over C), the mor-
phism groups will be the quotients
Hom
A
(V
1
, V
2
) := K
0
(T(V
1
V

2
) mod)/Z [triv]
where triv is the trivial one-dimensional representation of T(V
1
V

2
) given by
zero map
V
1
1
0
V
2
1
All the previous considerations extend to the case of A

.
Amazing similarities between categories C
Fq
and A

suggests the following

Conjecture 8. For any prime p there exists a tensor functor

p
: C
Fp
A

kar
and a sequence of isomorphisms of tensor functors from C
Fp
to V ect
C
for all
n 1
iso
n,p
:
A

n

p
i
V ect
Q
V ect
C

n
where i
V ect
Q
V ect
C
is the obvious embedding functor from the category of vector
spaces over Q to the one over C. Moreover, for any X C
Fp
the functor
p
maps the Frobenius element Fr
X
End
C
Fp
(X) to Fr
p(V )
.
This conjecture we call the Master Conjecture because it implies simultane-
ously all higher-dimensional versions of the Weil conjecture at once, as one has
the bijection (essentially by denition)
{(d 1)-dimensional super lattice models in A

}
{d-dimensional super lattice models in V ect
C
} .
18
A similar question about commuting endomorphisms in A is almost equivalent to the
study of nite-dimensional solutions of the Yang-Baxter equation.
27
Remark 1. One can consider a larger category A
super
adding to objects of A
super vector spaces as well. The group K
0
in the super case should be dened
as the naive K
0
modulo the relation
[V
1
U V
2
U] = [V
1
(U) V
2
(U)]
where is the parity changing functor.
It suces to verify the Master Conjecture only on the full symmetric monoidal
subcategory of C
Fp
consisting of powers
_
A
n
Fp
_
n0
of the ane line. The reason
is that any scheme of nite type can be embedded (by a constructible map)
in an ane space A
n
Fp
, and the characteristic function of the image of such an
embedding as an idempotent in End
C
Fp
(A
n
Fp
).
3.5.1 Machine modelling nite elds
Let us x a prime p. The object A := A
1
Fp
of C
Fp
is a commutative algebra
(as well as any scheme of nite type, see 2.2.1), with the product given by
the diagonal in its cube. The category A(C
Fp
) of ane schemes in C
Fp
(i.e. the category opposite to the category of commutative associative unital
algebras in C
Fp
) is closed under nite products. In particular, it makes sense to
speak about group-like etc. objects in A(C
Fp
). Ane line A is a commutative
ring-like object in A(C
Fp
), with the operations of addition and multiplication
corresponding to the graphs of the usual addition and multiplication on A
1
Fp
. In
plain terms, this means that besides the commutative algebra structure on A
m : A A A
we have two coproducts (for the addition and for the multiplication)
co a : A A A, co m : A A A
which are homomorphisms of algebras, and satisfy the usual bunch of rules for
commutative associative rings, including the distributivity law.
If the Master Conjecture 8 is true then it gives an object V
p
:=
p
(A)
A

kar
, with one product and two coproducts. One can expect that it is just C
p
as a vector space. For any n 1 the A

-product on V
p
denes a commutative
algebra structure on V
n
p
. Its spectrum should be a nite set consisting of p
n
elements. Two coproducts give operations of addition and multiplication on
this set, and we will obtain a canonical construction
19
of the nite eld F
p
n
uniformly for all n 1.
Even in the case p = 2 the construction of such V
p
is a formidable task: one
should nd 3 nite-dimensional super representations of the free algebra in 8
generators, satisfying 9 identities in various K
0
groups.
19
Compare with question 2 in Section 1.3, and remarks afterwards.
28
3.6 Corollaries of the Master Conjecture
3.6.1 Good sign: Bombieri-Dwork bound
One can deduce easily from the Master Conjecture that for any given p and
any system of equations in arbitrary number of variables (x
i
) where each of
equations is of an elementary form like x
i1
+ x
i2
= x
i3
, or x
i1
x
i2
= x
i3
or
x
i
= 1, the number of solutions of this system over F
p
n is an alternating sum of
exponents in n, with the total number of terms bounded by C
N
where C = C
p
is a constant depending on p, and N is the number of equations. In fact, it is a
well-known Bombieri-Dwork bound (and C is an absolute constant
20
), see [2].
3.6.2 Bad sign: cohomology theories for motives over nite elds
Any machine modelling nite eld should be dened over a nitely generated
commutative ring. In particular, there should be a machine dened over a
number eld K
p
depending only on the characteristic p. A little thinking shows
that the enumeration of the number of solutions of any given system of equations
in the elementary form as above, will be expressed as a super trace of an operator
in a nite-dimensional super vector space dened over K
p
. On the other hand,
it looks very plausible that the category of motives over any nite eld F
q
does
not have any ber functor dened over a number eld, see [9] for a discussion.
I think that this is a strong sign indicating that the Master Conjecture is just
wrong!
4 Categorical afterthoughts
4.1 Decategorications of 2-categories
Two categories, C
k
and A introduced in this paper have a common feature which
is also shared (almost) by the category of Grothendieck motives. The general
framework is the following.
Let B be a 2-category such that for any two objects X, Y B the category
of 1-morphisms Hom
B
(X, Y ) is a small additive category, and the composition
of 1-morphisms is a bi-additive functor. In practice we may ask for categories
Hom
B
(X, Y ) to be triangulated categories (enriched in their turn by spectra,
or by complexes of vector spaces). Moreover, the composition could be only a
weak functor (e.g. A

-functor), and the associativity of the composition could

hold only up to (xed) homotopies and higher homotopies. The rough idea is
that objects of B are spaces (non-linear in general), whereas objects of the
category Hom
B
(X, Y ) are linear things on the product X Y interpreted as
kernels of some additive functors transforming some kind of sheaves from X to
Y , by taking the pullback from X, the tensor product with the kernel on XY ,
and then the direct image with compact supports to Y .
20
A straightforward application of [2] gives the upper bound C 17
4
which is presumably
very far from the optimal one.
29
In such a situation one can dene a new (1-)category K
tr
(B) which is in fact
a triangulated category. This category will be called the decategorication of B.
The rst step is to dene a new 1-category K(B) enriched by spectra. It
has the same objects as B, the morphism spectrum Hom
K(B)
(X, Y ) is dened
as the spectrum of K-theory of the triangulated category Hom
B
(X, Y )
21
.
The second step is to make a formal triangulated envelope of this category.
This step needs nothing, it can be performed for arbitrary category enriched by
spectra. Objects of the new category are nite extensions of formal shifts of the
objects of K(B), like e.g. twisted complexes by Bondal and Kapranov.
At the third step one adds formally direct summands for projectors. The
resulting category K
tr
(B) is the same as the full category of compact objects
in the category of exact functors from K(B)
opp
to the triangulated category of
spectra (enriched by itself).
Finally, one can dene a more elementary pre-additive
22
category K
0
(B) by
dening Hom
K0(B)
(X, Y ) to be K
0
group of triangulated category Hom
B
(X, Y ).
Then we add formally to it nite sums and images of projectors. The result-
ing additive Karoubi-closed category will be denoted by K
kar
0
(B) and called
K
0
-decategorication of B. In what follows we list several examples of decate-
gorications.
4.1.1 Non-commutative stable homotopy theory
R. Meyer and R. Nest introduced in [8] a non-commutative analog of the tri-
angulated category of spectra. Objects of their category are not necessarily
unital C

-algebras, the morphism group from A to B is dened as the bivariant

Kasparov theory KK(A, B). One of main observations in [8] is that this gives a
structure of a triangulated category on C

-algebras. Obviously this construction

has a avor of the K
0
-decategorication.
4.1.2 Elementary algebraic model of bivariant K-theory
One can dene a toy algebraic model of the construction by Meyer and Nest. For
a given base eld k consider the pre-additive category whose objects are unital
associative k-algebras, and the group of morphisms from A to B is dened
as K
0
of the exact category consisting of bimodules (A
op
B-modules) which
are projective and nitely generated as B-modules. This is obviously a K
0
-
decategorication of a 2-category.
4.1.3 Non-commutative pure and mixed motives
Let us consider the quotient of the category of Grothendieck Chow motives
Mot
k,Q
over given eld k with rational coecients, by an autoequivalence given
21
It is well-known that in order to dene a correct K-theory one needs either an appropriate
enrichment on Hom
B
(X, Y ), or a model structure in the sense of Quillen, see e.g. [10].
22
Enriched by abelian groups in the plain sense (without higher homotopies).
30
by the invertible functor Q(1). The set of morphisms in this category between
motives of two smooth projective schemes X, Y is given by
Hom
Mot
k,Q
/Z
Q(1) (X, Y ) =

nZ
Hom
Mot
k,Q
(X, Q(n) Y ) =
=
_
Q
Z

nZ
Cycles
n
(X Y )
_
/( rational equivalence ) =
= Q
Z

nZ
CH
n
(X Y ) = Q
Z
K
0
(X Y )
because the Chern character gives an isomorphism modulo torsion between the
sum of all Chow groups and K
0
(X) = K
0
(D
b
(CohX)), the K
0
group of the
bounded derived category D(X) := D
b
(CohX) of coherent sheaves on X. Fi-
nally, the category D(X Y ) can be interpreted as the category of functors
D(Y ) D(X).
Triangulated categories of type D(X) where X is a smooth projective va-
riety over k belong to a larger class of smooth proper triangulated k-linear
dg-categories (another name is saturated categories), see e.g. [7],[12]. We
see that the above quotient category of pure motives is a full subcategory of
K
0
-decategorication (with Q coecients) of the 2-category of smooth proper
k-linear dg-categories. This construction was described recently (without men-
tioning the relation to motives) in [11].
Analogously, if one takes the quotient of the Voevodsky triangulated category
of mixed motives by the endofunctor Q(1)[2] , the resulting triangulated
category seems to be similar to a full subcategory of the full decategorication
of the 2-category of smooth proper k-linear dg-categories.
4.1.4 Motivic integral operators
We mentioned already in Section 2.1 that the category C
k
should be considered
as a K
0
-decategorication of a 2-category of motivic sheaves. A similar 2-
category was considered in [6] in the relation to questions in integral geometry
and calculus of integral operators with holonomic kernels.
4.1.5 Correspondences for free algebras
The category A is a K
0
-decategorication by denition.
4.2 Trace of an exchange morphism
Let G
1
, G
2
be two endofunctors of a triangulated category C, and an exchange
morphism (a natural transformation)
: G
1
G
2
G
2
G
1
31
is given
23
. Under the appropriate niteness condition (e.g. when C is smooth
and proper) one can dene the trace of , which can be calculated in two ways,
as the trace of endomorphism of Tor(G
1
, id
C
) associated with G
2
and , and as
a similar trace with exchanged G
1
and G
2
(see [4] for a related stu). Passing
to powers and natural exchange morphisms constructed from nm copies of :

(n,m)
: G
n
1
G
m
2
G
m
2
G
n
1
we obtain a collection of numbers Z

(n, m) := Trace(
(n,m)
) for n, m 1. It is
easy to see that these numbers come from a 2-dimensional super lattice model.
Let C = D(X) for smooth projective X, and functors are given by F

and
by E where F : X X is a map, and E is a vector bundle endowed with a
morphism g : F

E E (as in Section 1.3). In this case Z

(n, m) is the trace

(without the denominator) associated with the map F
n
and the bundle E
m
.
For example, one can construct a 2-dimensional super lattice model with the
partition function
Z
lat
R
(
n,m
) =

xC:F
n
(x)=x
x
m
where F : C C is a polynomial map
24
, e.g. F(x) = x
2
+c.
The conclusion is that two dierent proposals concerning motivic local sys-
tems in positive characteristic: the rst (algebraic dynamics) and the third one
(lattice models) are ultimately related. It is enough to nd the dynamical real-
ization, and then the lattice model will pop out. As it was mentioned already,
most probably these two proposals would fail, but they still can serve as sources
of analogies.
References
[1] A. Braverman, D. Kazhdan, Gamma-functions of representations and lift-
ing, with appendix by V. Vologodsky, GAFA 200 (Tel Aviv, 1999), Geom.
Funct. Anal. 2000, Special Volume, Part I, 237278, math.AG/9912208.
[2] E. Bombieri, On exponential sums in nite elds. II. Invent. Math. 47
(1978), no.1, 2939.
[3] V. Drinfeld, The number of two-dimensional irreducible representations of
the fundamental group of a curve over a nite eld, Functional Anal. Appl.
15 (1981), no. 4, 294295 (1982).
[4] N. Ganter, M. Kapranov, Representation and character theory in 2-
categories, math.KT/0602510.
23
We do not assume that is an isomorphism.
24
This seems to be a new type of integrability in lattice models, dierent from the usual
Yang-Baxter ansatz.
32
[5] A. Gerasimov, S. Kharchev, D. Lebedev, S. Oblezin, On a Gauss-Givental
Representation of Quantum Toda Chain Wave Function, Int. Math. Res.
Not. 2006, Art. ID 96489, 23 pp., math.RT/0505310.
[6] A. Goncharov, Dierential equations and integral geometry, Adv. Math.
131 (1997), no.2, 279343.
[7] M. Kontsevich, Y. Soibelman, Notes on A

-algebras, A

-categories and
non-commutative geometry. I, math.RA/0606241.
[8] R. Meyer, R. Nest, The Baum-Connes Conjecture via Localization of Cat-
egories, Topology, vol. 45 (2006), no. 2, pp. 209259, math.KT/0312292.
[9] J. S. Milne, Motives over F
p
, math.AG/0607569.
[10] M. Schlichting, A note on K-theory and triangulated categories, Invent.
Math. 150 (2002), no. 1, 111116.
[11] G. Tabuada, Invariants additifs de dg-categories, Int. Math. Res. Not.
2005, no. 53, 3309-3339, math.KT/0507227.
[12] B. Toen, M. Vaquie, Moduli of objects in dg-categories, math.AG/0503269.
IHES, 35 route de Chartres, Bures-sur-Yvette 91440, France
maxim@ihes.fr
33