Automorphisms of the Weyl algebra

Maxim Kontsevich
I.H.E.S.
35 route de Chartres, Bures-sur-Yvette, 91440
France

Mathematische Arbeitstagung June 10-16, 2005

Max-Planck-Institut für Mathematik, Bonn, Germany

I report on a joint work with Alexei Belov-Kanel, which was started about
1,5 years ago, and is not finished yet.

1 Main Conjecture
For integer n ≥ 1 denote by An,C the Weyl algebra of rank n over C

Chx̂1 , . . . , x̂2n i/ ( relations [x̂i , x̂j ] = ωij , ∀i, j, 1 ≤ i, j ≤ 2n ) ,

where (ωij )1≤i,j≤2n is the standard skew-symmetric matrix,

ωij = δi,n+j − δi+n,j .

Algebra An,C is isomorphic to the algebra D(AnC ) of polynomial differential

operators in n variables x1 , . . . , xn :

x̂i 7→ xi , xn+i 7→ ∂/∂xi , i = 1, . . . , n .

Denote by Pn,C the Poisson algebra over C which is the usual polynomial algebra
C[x1 , . . . , x2n ] ' O(A2n
C ) endowed with the Poisson bracket

{xi , xj } = ωij , 1 ≤ i, j ≤ 2n .

Conjecture 1 There exists a canonical isomorphism Φ = Φ n

Aut(An,C ) ' Aut(Pn,C )

between the automorphism group of the Weyl algebra and the group of polyno-
mial symplectomorphisms of A2n C . Isomorphism Φ is covariant with respect to
the natural action of the Galois group Aut(C).
1.1 First positive evidence: case n = 1
The structure of group Aut(P1,C ) is known after H.W.E.Young (1942) and W.
van der Kulk (1953). This group contains the group G 1 = SL(2, C) n C2 of
special affine transformations, and solvable group G 2 of polynomial transfor-
mations of the form

(x1 , x2 ) 7→ (λx1 + F (x2 ), λ−1 x2 ), λ ∈ C× , F ∈ C[x] .

Group Aut(P1,C ) is equal to the amalgamated product of G 1 and G2 over their

intersection. L.Makar-Limanov (1984) proved that if one replaces commut-
ing variables x1 , x2 by noncommuting variables x̂1 , x̂2 in above formulas, one
obtains the description of group Aut(A 1,C ). Hence, in the case n = 1 two
automorphism groups are isomorphic.

1.2 Negative evidence: Lie algebras are not isomorphic

Groups Aut(An,C ) and Aut(Pn,C ) are in fact groups of complex points of certain
group-like ind-affine schemes Aut(A n,C ) and Aut(Pn,C ). Namely, an automor-
phism φ of An,C such that both φ, φ−1 map all generators x̂i to expressions
of degree bounded by some integer N ≥ 1, is a solution of certain system of
polynomial equations in variables c i,I , c0i,I (coefficients of φ and φ−1 ):
X X
φ(x̂i ) = ci,I x̂I , φ−1 (x̂i ) = c0i,I x̂I ,
I: |I|≤N I: |I|≤N

where for multi-index I = (i1 , . . . , i2n ) ∈ Z2n

≥0 we define

x̂I := x̂i11 . . . x̂i2n

2n
, |I| := i1 + · · · + i2n .

Similarly, one has an ind-scheme Aut(P n,C ), an inductive limit of a chain of

affine schemes of finite type over C. Using dual numbers C[t]/(t 2 ) one defines
as usual Lie algebras of ind-schemes Aut(A n,C ) and Aut(Pn,C ). These algebras
are derivations Der(An,C ) and Der(Pn,C ) respectively. It is well-known that all
derivations of An,C are inner,

Der(An,C ) ' An,C / C · 1An,C .

Similarly, derivations of Pn,C are hamiltonian vector fields,

Der(Pn,C ) ' Pn,C / C · 1Pn,C .

These two Lie algebras are not isomorphic: Der(P n,C ) contains many non-trivial
Lie subalgebras of finite codimension (e.g. vector fields vanishing at some point
of A2n
C ), while Der(An,C ) does not contain such subalgebras (in a sense it is
similar to sl∞ ).
Thus, we conclude that the conjectural isomorphism Φ can not be an iso-
morphism of ind-schemes. In fact, we expect that Φ preserves the filtration of
automorphism groups by degree, and it is a constructible continuous map, both
in Zariski and usual topology.
1.3 Another positive evidence: tame automorphisms
Symplectic group Sp(2n, C) acts by automorphisms of A n,C and of Pn,C by linear
transformations of generators. Also, for any polynomial F ∈ C[x 1 , . . . , xn ] we
define non-linear transvections

TFP ∈ Aut(An,C ), TFA ∈ Aut(Pn,C )

by formulas

TFP (xi ) = xi , TFP (xn+i ) = xn+i + ∂i F (x1 , . . . , xn ), 1 ≤ i ≤ n ,

TFA (x̂i ) = x̂i , TFP (x̂n+i ) = x̂n+i + ∂i F (x̂1 , . . . , x̂n ), 1 ≤ i ≤ n .

The last formual makes sense, as variables x̂ 1 , . . . , x̂n commute with each other.
Correspondence F 7→ TFP (resp. F 7→ TFA ) gives a group homomorphism
C[x1 , . . . , xn ]/C · 1 → Aut(Pn,C ) (resp. to Aut(An,C )). Automorphisms of An,C
and of Pn,C generated by Sp(2n, C) and transvections are called tame.
We prove in the next section that groups of tame automorphisms of A n,C and
of Pn,C are canonically isomorphic. Symplectic group Sp(2n, C) is generated by
Fourier transform

xi 7→ xn+i , xn+i 7→ −xi ; x̂i 7→ x̂n+i , x̂n+i 7→ −x̂i , i = 1, . . . , n

and by transvections corresponding to quadratic polynomials F . Our assertion

follows from the following theorem:
Theorem 1 For any sequence of polynomials F 1 , . . . , Fk ∈ C[x1 , . . . , xn ]/C · 1
the composition
F ourier ◦ TFA1 ◦ F ourier ◦ · · · ◦ TFAk
is identity in Aut(An,C ) if and only if the composition

F ourier ◦ TFP1 ◦ F ourier ◦ · · · ◦ TFPk

is identity in Aut(Pn,C ).

2 Proof of theorem 1
Notice that in the definition of algebras A n,C and Pn,C one can replace C by
arbitrary commutative ring.
The only known proof of Theorem 1 is based on considerations in finite
characteristic. It is a challenge to find a purely complex proof. In fact, in the
case n = 1, J.Dixmier in 1968 described automorphisms of A 1 in finite charac-
teristic, and L.Makar-Limanov used the result of Dixmier in zero characteristic.
Here we follow the same line.
Let us assume that the composition in Aut(A n,C ) is equal to identity. We
want to prove that the corresponding composition in Aut(P n,C ) is also equal to
identity.
Let us denote by R ⊂ C the subring of C generated by all coefficients of
polynomials F1 , . . . , Fk . It is a finitely generated integral domain. Hence for all
sufficiently large primes p  1 the reduction R/p := R ⊗ Z/pZ is a non-zero
ring. For any prime p Weyl algebra An,R/p has a large center,
Center(An,R/p ) = R/p [x̂p1 , . . . , x̂p2n ] .
Any automorphism φp of An,R/p induces an automorphism of its center. If we
replace generators x̂pi , i = 1, . . . , 2n of Center(An,R/p ) by letters xi , we obtain a
polynomial automorphism φcentr p in characteristic p. One sees immediately that
the Fourier transfrom maps to Fourier transform. Almost the same happens
for transvections:
(TFA (mod p) )centr = TFPrp (F ) ,
P p I P I
where F rp (F ) := I cI x for F = I cI x . It follows immediately from a
simple identity in zero characteristic for differential operators in one variable:
(∂/∂x + g 0 (x))p = (∂/∂x)p + (g 0 (x))p (mod p) .
Thus, we see that the composition F ourier ◦ T FP1 ◦ F ourier ◦ · · · ◦ TFPk (after
application of the Frobenius map F r p to its coefficients) coinsides with the
identity morphism modulo p for all sufficiently large primes p. Hence it is equal
to identity in Aut(An,R ) ⊂ Aut(An,C ). Implication in one direction is proven.
Conversely, let us assume that the composition φ A is not equal to identity
in Aut(An,R ). Then it is either a non-trivial affine map on generators, or it
maps some generator x̂i to an expression of degree d > 1. In the first case one
can show that the composition φP is not equal to identity, as after application
of Frobenius one obtains affine map φcentr
p 6= id. In the second case the image
p
of x̂i has degree strictly equal to pd as its symbol is p-th power of the degree d
symbol of x̂i . Hence, we have again φcentr
p 6= id. This finishes the proof of the
inverse implication.

3 Conjectural description of Φ
The above prooof give an indication how the homomorphism Φ : Aut(A n,C ) →
Aut(Pn,C ) should be defined in general. Namely, as before, we can assume that
φ ∈ Aut(An,C ) is defined over a finitely generated integral domain R ⊂ C.
For any prime p we obtain a polynomial map φ centrp ∈ Aut(R/p [x1 , . . . , x2n ] by
restriction of φ mod p to the center of A n,R/p .
The next step is to prove that φcentr
p is a symplectomorphism. For suffi-
ciently large p ring R is flat over p ∈ Spec(Z), hence A n,R /p2 is flat over Z/p2 .
We can consider An,R/p2 as an infinitesimal one-step deformation of associative
algebra An,R/p . In the usual way, one associates to this deformation a canonical
Poisson bracket on the center of An,R/p :

[ã, b̃]
{a, b} := (mod p) ,
p
where ã, b̃ ∈ An,R /p2 are arbitrary lifts of central elements a, b ∈ Center(A n,R/p ).
A straightforward calculation:
[(∂/∂x)p , xp ] = −p (mod p2 )
shows that one get the usual Poisson bracket on R/p [x 1 , . . . , x2n ] ' Center(An,R/p ).
Hence, we conclude that φcentr
p is a symplectomorphism for p  1. For any p
centr
the degree of φp is bounded from above by the degree of φ.
Let us denote by R∞ the following ring:
Y M
(R/p) / (R/p)
primes p primes p

It contains R ⊗ Q, and it has a “universal” Frobenius endomorphism F r R :

R∞ → R∞ , (ap ) 7→ (app ). Our construction of symplectomorphisms φ centr
p for
all p  1 can be interpreted as a map
Φcentr : Aut(An,R ) → Aut(Pn,R∞ ), φ 7→ (φcentr
p ).
The example of tame automorphisms shows that it is not exactly the conjec-
tural map Φ which we want to construct, one should first untwist by Frobenius
coefficents of symplectomorphisms φ centr
p . This can be done in general, by the
following result:
Theorem 2 For any finitely generated ring R which is smooth and dominant
over Z, the image of the map Φcentr belongs to Aut(Pn,F rR (R∞ ) ) ⊂ Aut(Pn,R ).
This theorem is applicable to automorphisms over C because any finitely gen-
erated integral domain R ⊂ C can be made smooth over Z by adding inverses
of finitely many non-zero elements.

3.1 Proof of Theorem 2

Let us choose (under the assumption that R is smooth and dominant over Z)
a finite collection of derivations δ 1 , . . . , δk ∈ Der(R) which span the tangent
space at every closed point of Spec(R). An element a ∈ R/p belongs to (R/p) p
if and only if it is killed by all derivations δ j . Therefore, we have to prove that
for all p  1 and any i, j, one has
δj (φcentr
p (xi )) = 0 ∈ R/p [x1 , . . . , x2n ] .
Equivalently, we have to show that
δj (φp (x̂pi )) = 0 ∈ An,R/p .
The l.h.s of the above expression is equal to
δj ((φp x̂i )p ) = ap−1 b + ap−2 ba + · · · + bap−1 (mod p)
where
a := φ(x̂i ), b := δj (a), a, b ∈ An,R .
Notice that for any i element x̂i is locally ad-nilpotent in An,R , that is for any
element f ∈ An,R the iterated commutator (ad(x̂i ))m (f ) vanishes for sufficiently
large m (in fact m = deg(f )+1 suffices). Hence, element a = φ(x̂ i ) is also locally
ad-nilpotent, and there exists positive integer N such that (ad(a)) N (b) = 0.
Finally, for any prime p ≥ N + 1 one has
 
p−1 p−1 p − 1 p−2
0 = (ad(a)) (b) = a b − a ba + · · · + bap−1 =
1
= ap−1 b + ap−2 ba + · · · + bap−1 (mod p) .
3.2 Lifting to characteristic zero
By Theorem 2 we know that the Φcentr (φ) is a polynomial symplectomorphism
with coefficients in F rR (R∞ ) ⊂ R∞ . We define Φuntwisted (φ) to be an element
of Aut(Pn,R∞ ) obtained from Φcentr (φ) by the inverse Frobenius map (it is
well-defined because F rR is injective for any integral domain R).

Conjecture 2 For any finitely generated integral domain R smooth and domi-
nant over Z and any φ ∈ Aut(An,R ), element Φuntwisted (φ) belongs to subgroup
Aut(Pn,R⊗Q ) ⊂ Aut(Pn,R∞ ).

If we assume Conjecture 2, then the map Φ from Conjecture 1 is defined by

Φ(φ) := Φuntwisted (φ) ⊂ Aut(Pn,R⊗Q ) ⊂ Aut(An,C )

We know that conjecture 2 holds for tame automorphisms. In general, I can

prove a good approximation:

Theorem 3 Under the same assumptions as in Conjecture 2, element Φ untwisted (φ)

belongs to Aut(Pn,R̃ ), where R̃ ⊂ R∞ is a finitely generated ring (not necessarily
integral) containing R and finite over the generic point of Spec(R).

The above theorem implies that there exists a homomorphism

Aut(An,Q ) → Aut(Pn,Q )

which may be not covariant with respect to the natural Aut(Q) action.
I will not give here the proof of theorem 3. It is based on some geometric
considerations relating general holonomic modules over the Weyl algebra, and
singular lagrangian subvarieties of A 2n .

4 Conjectural description of Φ−1

Here we propose a candidate for the inverse map

Φ−1 : Aut(Pn,C ) → Aut(An,C ) .

As before, we may assume that ψ ∈ Aut(P n,C ) is defined over a finitely gen-
erated integral domain R ⊂ C. We know already the automorphism Φ −1 (ψ)
(mod p) of algebra An,R/p should induce symplectomorphism F r p (ψp ) on its
center for any prime p  1. We will show here that it lifts to at most unique
automorphism of An,R/p .
Algebra An,R/p is an Azumaya algebra over its center, a twisted form of
matrix algebra M at(pn × pn , R/p [x̂p1 , . . . , x̂p2n ]). It has an invariant

[An,R/p ] ∈ Br(R/p [x1 , . . . , x2n ])

in the Brauer group of its center Center(A n,R/p ) ' R/p [x1 , . . . , x2n ], the group
of Morita equivalence classes of Azumaya algebras.
 For any commutative algebra S in finite characteristic p > 0 there is a
canonical map
Ω1abs (S)/dΩ0 (S) → Br(S), Ω1abs (S) := Ω1 (S/Z) ,
given by the formula f dg 7→ [Af,g ]. Here for any two elements f, g ∈ S one
defines Azumaya algebra Af,g over S as
Af,g := Shξ, ηi/ relations ξ p = f, η p = g, [ξ, η] = 1 .
It follows from definitions that the class [A n,R/p ] is given by the class of 1-form
n
X
α := xi dxn+i ∈ Ω1abs (R[x1 , . . . , x2n ]) .
i=1

Let us denote by α0 the same 1-form considered as a relative form over R.

Symplectomorphism ψ preserves 2-form dα 0 , hence (by Poincaré lemma) there
exists a function W ∈ R ⊗ Q[x1 , . . . , x2n ] such that ψ ∗ (α0 ) = α0 + dW . We
enlarge R by adding inverses to prime numbers which appear in denominators
of coefficients of W . Symplectomorphism F r p (ψ) for p  1 preserves the class
of the absolute form α modulo differentials, because all coefficients of F r p (ψ) are
p-th powers, and therefore can be treated as “constants” (they have vanishing
derivatives).
The conclusion is that for all p  1 there exists a Morita self-equivalence of
algebra An,R/p inducing automorphism F rp (ψ) on its center. This Morita self-
equivalence is given by a bimodule M ψ,p over An,R/p , which is finitely generated
projective both as left and as right module. Moreover, it is easy to see that
such bimodule is unique up to an isomorphism, and its automorphism group is
the group of multiplications by scalars (R/p) × .
Conjecture 1 follows from Conjecture 2 and the following
Conjecture 3 Bimodule Mψ,p is free rank one as left and as right A n,R/p -
module.
Namely, if Mψ,p is free rank one module on both sides, then choosing the
generator of it we obtain an automorphism of A n,R/p . Conjecture 3 implies
that ind-schemes Aut(An ) and Aut(Pn ) over Z have the same points over fi-
nite fields of sufficiently large characteristic (when we keep the degree of auto-
morphisms bounded), therefore (assuming Conjefcture 2) there will be a con-
structible homeomorphism of schemes over Q.
There is no good reasons to beleive in Conjecture 3, as an analog of Serre
conjecture fails for Weyl algebras. Namely, there are non-trivial projective mod-
ules of “rank one” over A1,C which are not free (J.T.Stafford (1987), Yu.Berest
and J.Wilson (2000)). Maybe, one should replace in Conjecture 1 the group
Aut(An,C ) by the group of Morita self-equivalences of A n,C .

5 Generalizations
There are many generalizations of Conjecture 1. Here I’ll describe few of them.
First of all, the idea of the use of large centers in finite characteristic can be
applied to holonomic D-modules. For example, with any holonomic D X -module
M where X is a smooth algebraic variety over C one can associate the collections
of its supports (mod p) in F rp (T ∗ X). One can prove that these supports are
Lagrangian varieties. In general, these supports behave in a complicated way
as functions of prime p, the dependence on p is (presumably) related to the
motivic Galois group in the crystalline realization. Nevertheless, I expect that
in some situations (including the one considered in Conjecture 1), there exists
a clean correspondence:

Conjecture 4 For any smooth variety X/C and a smooth closed Lagrangian
subvariety L ⊂ T ∗ X such that H1 (L(C); Z) = 0 there exists a canonical holo-
nomic DX -module ML characterized uniquely by the property that after reduc-
tion to finite characteristic p  1 module M L considered as a module over the
center of algebra of differential operators, is supported over F r p (L) and has
rank pdimX .

For example, any polynomial symplectomorphim ψ of A 2n C gives a contractible

Lagrangian subvariety in A4nC , its graph. The corresponding holonomic module
can be interpreted as a bimodule over A n,C .
One of motivations for me to propose Conjecture 1 came from the defor-
mation quantization. Namely, I proved several years ago the following result,
which is in fact not a theorem but a construction:

Theorem 4 Let X/C be a smooth affine variety endowed with an algebraic

Poisson bracket α ∈ Γ(X, ∧2 TX ). Assume that X is rational, and there exists
a smooth compactification X by divisor D = X \ X such that D is ample, α
extends to X, and the ideal of D is a Poisson ideal. Then there exists a canonical
associative algebra A/C[[~]] which is free as C[[~]]-module and quantizes O(X)
in direction α.

In fact, I expect that the quantized algebra does not depend on the choice
of compactification if it exists. In the case when there exists a symmetry of
X identified α and rescaled bivector field λα where λ ∈ C × is arbitrary, one
can remove parameter ~. In the case X = A 2n C with the flat bracket we get
Conjecture 1. Also we get

Conjecture 5 For any finite-dimensional Lie algebra g/C the group of auto-
morphisms of the universal enveloping algebra U g is canonically isomorphic to
the group of polynomial automorphisms of g ∗ preserving the standard Kirillov
bracket.

Finally, using the same circle of ideas, together with A.Belov-Kanel we

proved that the stable Jacobian Conjecture:
étale maps AnC → AnC , n ≥ 1 are invertible
is equivalent to the stable Dixmier Conjecture:
Endomorphisms of An,C , n ≥ 1 are invertible .