1306.6754v2

DIFFERENTIAL OPERATORS ON G/U AND THE AFFINE GRASSMANNIAN
VICTOR GINZBURG AND SIMON RICHE
A BSTRACT. We describe the equivariant cohomology of cofibers of spherical perverse sheaves on the
affine Grassmannian of a reductive algebraic group in terms of the geometry of the Langlands dual
arXiv:1306.6754v2 [math.RT] 24 Mar 2014
group. In fact we give two equivalent descriptions: one in terms of D-modules of the basic affine
space, and one in terms of intertwining operators for universal Verma modules. We also construct
natural collections of isomorphisms parametrized by the Weyl group in these three contexts, and
prove that they are compatible with our isomorphisms. As applications we reprove some results of
the first author and of Braverman–Finkelberg.
1. I NTRODUCTION
1.1. The geometric Satake equivalence relates perverse sheaves (with complex coefficients in our
case) on the affine Grassmannian Gr of a complex connected reductive algebraic group Ǧ and rep-
resentations of the Langlands dual (complex) reductive group G. The underlying vector space of
the representation S(F) attached to a perverse sheaf F is given by its total cohomology H (Gr, F).
q
It turns out that various equivariant cohomology groups attached to F also carry information on
the representation S(F), see e.g. [G1, YZ, BF]. In this paper, if Ť is a maximal torus of Ǧ, we de-
scribe, in terms of G, the equivariant cohomology of cofibers of F at Ť -fixed point, with respect
to the action of Ť or of Ť × C× , where C× acts on Gr by loop rotation. In fact these groups can
be described in two equivalent ways, either in terms of D-modules on the basic affine space or in
terms of intertwining operators for universal Verma modules. We also describe the Weyl group
action on this collection of spaces induced by the action of NǦ (Ť ) on Gr.
1.2. To state our results more precisely, choose some Borel subgroup B̌ ⊂ Ǧ containing Ť , and
let T, B be the maximal torus and the Borel subgroup of G provided by the geometric Satake
equivalence. Note that the Tannakian construction of G also provides no zero vectors in each
simple root subspace of g := Lie(G). In this paper we study three families of graded modules over
a polynomial algebra, attached to G or Ǧ, and endowed with symmetries parametrized by their
common Weyl group W .
Let t := Lie(T ) and S~ := S(t)[~], considered as a graded algebra where ~ and the vectors in t
are in degree 2. (Here, S(t) is the symmetric algebra of the vector space t.) Let also X := X ∗ (T ) be
the character lattice. Let Rep(G) be the category of finite dimensional algebraic G-modules.
Our first family of graded modules over S~ is of “geometric” nature. Let U be the unipotent
radical of B, and let X := G/U be the basic affine space. Consider the algebra D~ (X) of (global)
asymptotic differential operators on X, i.e. the Rees algebra of the algebra Γ(X, DX ) of differential
operators on X, endowed with the order filtration (see §2.4 for details). This algebra is naturally
graded, and endowed with an action of T induced by right multiplication on X. We denote by
D~ (X)λ the weight space associated with λ ∈ X. Then we set
G
Mgeom
V,λ := V ⊗
(λ)
D~ (X)λ .
The work of V.G. was supported in part by the NSF grant DMS-1001677. The work of S.R. was supported by ANR
Grants No. ANR-09-JCJC-0102-01 and No. ANR-2010-BLAN-110-02.
1
(Here the twist functor (λ)(·) will be defined in §2.4.)
Our second family of graded S~ -modules is of “algebraic” nature. Let U~ (g) be the asymptotic
enveloping algebra of g (i.e. the Rees algebra of the algebra U (g) endowed with the Poincaré-
Birkhoff-Witt filtration, see §2.1 for details). For λ ∈ X we let M(λ) be the asymptotic universal
Verma module associated with λ, a graded (S~ , U~ (g))-bimodule whose precise definition is re-
called in §2.1. Then we set

Malg
V,λ := Hom(S~ ,U~ (g)) M(0), V ⊗ M(λ)
where we consider morphisms in the category of (S~ , U~ (g))-bimodules.

We will also construct a third family of graded modules, of “topological” nature, which is as-
sociated with the “Langlands dual data.” Let ť := Lie(Ť ). We have canonical identifications
X∼= X∗ (Ť ) and S~ ∼
= S(ť∗ )[~]. Consider the category PervǦ(O) (Gr) of Ǧ(O)-equivariant perverse
sheaves on the affine Grassmannian Gr of Ǧ. Then for any λ ∈ X and F in PervǦ(O) (Gr) we set
q+λ(2ρ̌)
Mtop !
F ,λ := HŤ ×C× (iλ F).
Here iλ is the inclusion of the point of Gr naturally associated with λ, C× acts on Gr by loop
rotation, and ρ̌ is the half sum of positive coroots of G. Then Mtop
F ,λ is in a natural way a graded
S~ -module.
1.3. Each of these families is endowed with a kind of “symmetry” governed by the Weyl group
W of (G, T ) or (Ǧ, Ť ). (Note that these Weyl groups can be canonically identified.) Namely, we
have isomorphisms of graded S~ -modules
∼ ∼
→ wMV,wλ
AV,λ,w : MV,λ − or → wMF ,wλ
AF ,λ,w : MF ,λ −
for all w ∈ W . (Here the twist functor w(·) will be defined in §2.4.)
In the “geometric” case, the isomorphisms Ageom V,λ,w are constructed using a W -action on D~ (X)
given by partial Fourier transforms due to Gelfand–Graev and studied in particular by Bezrukav-
nikov–Braverman–Positselskii in [BBP]. These operators depend on a choice of (non-zero) simple
root vectors in g, which we choose to be those provided by the geometric Satake equivalence.
In the “algebraic” setting, the isomorphisms Aalg V,λ,w are constructed using properties of inter-
twining operators between a Verma module and a tensor product of a G-module and a Verma
module. Our constructions are “renormalized” variants of classical constructions appearing in
the definition of the dynamical Weyl group (see [TV, EV]) but, as opposed to those considered in
loc. cit., our isomorphisms do not have poles. Again, the operators Aalg V,λ,w depend on a choice of
simple root vectors in g, which we choose as above.
In the “topological” setting, the isomorphisms Atop F ,λ,w are induced by the action of NǦ (Ť ) on Gr
by left multiplication.
In each setting, the collection of operators is compatible with the product in W in the sense that
y
y

AV,yλ,x ◦ AV,λ,y = AV,λ,xy or AF ,yλ,x ◦ AF ,λ,y = AF ,λ,xy
for any λ, V, F as above and x, y ∈ W .
1.4. In addition, these families of graded modules are endowed with morphisms
ConvV,V ′ ,λ,µ : MV,λ ⊗S~ (λ)MV ′ ,µ → MV ⊗V ′ ,λ+µ , ConvF ,F ′ ,λ,µ : MF ,λ ⊗S~ (λ)MF ′ ,µ → MF ⋆F ′ ,λ+µ
related to the monoidal structure on the category Rep(G) (denoted ⊗) or Perv Ǧ(O) (Gr) (denoted ⋆).
2
In the “geometric” setting, morphisms Convgeom
V,V ′ ,λ,µ are induced by the product in the algebra
D~ (X). In the “algebraic” setting, morphisms Convalg
V,V ′ ,λ,µ are induced by composition of mor-
phisms of bimodules. In the “topological” setting, morphisms Convtop
F ,F ′ ,λ,µ are defined using a
standard construction considered in particular in [ABG].
1.5. Our main result might be stated as follows (see Corollary 2.4.2, Theorem 2.5.5, and Proposi-
tion 8.1.5).
Theorem. For F in PervǦ(O) (Gr) and λ ∈ X there exist canonical isomorphisms
Mtop ∼ geom ∼ alg
F ,λ = MS(F ),λ = MS(F ),λ ,
∼
where S : PervǦ(O) (Gr) −→ Rep(G) is the geometric Satake equivalence. These families of isomorphisms
are compatible with operators A and with morphisms Conv.
The proof of this theorem is based on another crucial property of the modules Mgeom alg
V,λ , MV,λ and
Mtop
F ,λ : they are all compatible with restriction to a Levi subgroup in the appropriate sense. This
property is used to reduce the proof of our claims to the case G and Ǧ have semisimple rank one,
in which case they can be checked by explicit computation. This strategy is rather classical in this
context, see e.g. [BFM, BF, BrF, AHR].
1.6. The present paper is closely related to, and motivated by, results of [ABG] and [BF]. In fact, in
a follow-up paper the results of the present article will be used to obtain a common generalization
of the equivalences of categories established in these papers. A similar generalization can also be
obtained using recent results of Dodd [Do], but our approach is different and, we believe, more
explicit. We will follow the strategy of [ABG] and a key technical step in our approach is the
following algebra isomorphism, which is a ”quantum” analogue of [ABG, Theorem 8.5.2] and
which follows from the theorem stated in §1.5:
 
M M
ExtŤ ×C× (RG , W λ ⋆ RG ) ∼
= U~ (g) ⋉  D~ (X)λ  .
q
λ∈X+ λ∈X+
Here, W λ is the Wakimoto sheaf associated with λ, RG is an ind-perverse sheaf on Gr correspond-

ing to the regular representation of G, and we refer to [ABG, §8] for this and other unexplained
notation.
1.7. We will also consider “classical analogues” of the above constructions, by which we mean
specializing ~ to 0, hence replacing S~ by S(t) or S(ť∗ ). The classical analogues of Mtop are easy to
define: we simply set
top q+λ(2ρ̌) !
MF ,λ := HŤ (iλ F).
top top
We also have morphisms A and Conv given by the same constructions as for Atop and Convtop .
There is no interesting classical analogue of Malg . The classical analogues of Mgeom are defined
using the geometry of the Grothendieck–Springer resolution e g. More precisely we set
geom G
MV,λ := V ⊗ Γ(e g, Oeg (λ))
geom
where Oeg (λ) is the G-equivariant line bundle on e
g associated with λ. The operators Conv are
induced by the natural morphisms
Γ(e g, Oeg (µ)) → Γ(e
g, Oeg (λ)) ⊗ Γ(e g, Oeg (λ + µ)).
3
geom
Finally, the operators A are defined using the W -action on the regular part of e
g. Again, these
operators depend on a choice of simple root vectors in g. This construction seems to be new, and
has interesting consequences (see §5.5).
Then we prove the following (see Corollary 2.4.5, Theorem 2.5.7 and Remark 8.1.6(2)).
Theorem. For F in PervǦ(O) (Gr) and λ ∈ X, there exist canonical isomorphisms

top geom
MF ,λ ∼
= MS(F ),λ .
This family of isomorphisms is compatible with operators A and with morphisms Conv.
The modules appearing in the theorem (and the corresponding morphisms) are related to those
appearing in the theorem of §1.5 by the functor C ⊗C[~] (−) (where ~ acts by zero on C). For Mtop
top
and M , this easily follows from the parity vanishing of H (i!λ F), see Lemma 6.2.4. For Mgeom
q
geom
and M , this requires a more subtle argument, see §3.5. In particular, our results establish
a relation between the automorphisms of D~ (X) induced by partial Fourier transforms and the
W -action on the regular part of e
g, which seems to be new.
1.8. As applications of our constructions we give new proofs of two results: a geometric descrip-
tion of the Brylinski–Kostant filtration due to the first author (see [G1]), and a geometric construc-
tion of the dynamical Weyl group due to Braverman–Finkelberg (see [BrF]). We also observe that
some of our technical preliminary results have interesting applications: they allow to give simpler
proofs of results on the structure of the algebra D(X) of differential operators on X (see §3.6) and
to construct an action of W on the regular part of T ∗ X which “lifts” the action on the regular part
of eg, see §5.5.
One important tool in the first proof of the geometric Satake equivalence in [G1] was specialized
equivariant cohomology of cofibers (see in particular [loc. cit., §3.5]), while in [MV] the authors
replaced this tool by cohomology of corestrictions to semi-infinite orbits Tλ . Our descriptions of
HŤ (i!λ F), HŤ (t!λ F) (where tλ denotes the inclusion of Tλ ) and the natural morphism between them
q q
(see Theorem 2.3.1) shed some light on the precise relation between these points of view.
1.9. Description of the paper. In Section 2 we define our main players, and state our main results.
In Section 3 we study the modules Mgeom
V,λ and define their symmetries. In Section 4 we study the
modules MalgV,λ , define their symmetries, and relate this algebraic family with the geometric one. In
geom
Section 5 we study the modules MV,λ , define their symmetries, and relate them with the modules
geom
MV,λ . In Section 6 we recall the construction of the geometric Satake equivalence and its main
properties. In Section 7 we prove our main results. In Section 8 we give some complements
and applications of these results. Finally, the paper finishes with two appendices: Appendix A
collects computations in semi-simple rank one that are needed in our proofs, and Appendix B is a
reminder on partial Fourier transforms for (asymptotic) D-modules.
1.10. Conventions. Throughout,

L we will work over the ground field C of complex numbers and
write ⊗ = ⊗C . If M = n∈Z Mn is a graded vector space and m ∈ Z, we define the graded vector
space M hmi by the following rule: (M hmi)n = Mn−m . Note that h1i is a “homological” shift, i.e. it
shifts graded vector spaces to the right. We will always consider C[~] as a graded algebra where
~ has degree 2. If A and B are C[~]-algebras, by an (A, B)-bimodule we mean an (A ⊗C[~] B op )-
module. If A is an algebra, we write Hom−A (−, −) for HomAop (−, −).
4
1.11. Acknowledgements. We thank Pramod Achar for useful discussions at early stages of this
work, and Pierre Baumann. Part of this work was completed while the second author visited
University of Chicago.
2. S TATEMENT OF THE MAIN RESULTS

S
2.1. Asymptotic Verma modules. Given a filtered C-algebra A = i∈Z≥0 Fi A, we let A~ be the
Rees algebra (sometimes referred to as “graded” or “asymptotic” version) of the filtered algebra
A. It can be defined as the following subalgebra of A[~]:
(
M 0 if i is odd;
A~ := Ai~ with Ai~ = i
(2.1.1)
i∈Z
~ · Fi/2 A if i is even.
Thus, A~ is a graded C[~]-algebra, where the indeterminate ~ has grade degree 2. (The reason for
our convention will become clear later.) Moreover, one has a natural isomorphism
A~ /~ · A~ ∼
= grF A
(where degrees are doubled on the left-hand side).
If k is a Lie algebra, the enveloping algebra U (k) comes equipped with a natural ascending filtra-
tion, the Poincaré-Birkhoff-Witt filtration, such that gr U (k) = S(k). The corresponding asymptotic
enveloping algebra U~ (k) := U (k)~ has an alternative (equivalent) definition as the C[~]-algebra
generated by k, with relations xy − yx = ~[x, y] for x, y ∈ k. Here elements of k have degree 2. We
will use this description of U~ (k), and still denote by x the image of an element x ∈ k. (If we were
using the description (2.1.1), this element should rather be denoted ~x.)
Let G be a connected reductive group over C with Lie algebra g. We fix a triangular decompo-
sition g = u ⊕ t ⊕ u− , so b = t ⊕ u is a Borel subalgebra. Let T be the maximal torus and B = T · U
the Borel subgroup corresponding to the Lie algebras t and b, respectively. We will denote by R
the set of roots of G (relative to T ), by R+ the positive roots (i.e. the roots of u), and by W the
Weyl group of (G, T ). Let ρ ∈ t∗ , resp. ρ̌ ∈ t, be the half sum of positive roots, resp. coroots. We
also let X be the lattice of characters of T , and X+ , resp. X− , be the sub-semigroup of dominant,
resp. antidominant, weights. We will frequently consider elements of t∗ (resp. of X) as linear forms
on b (resp. characters of B) which are trivial on u (resp. on U ). Also, as usual, when convenient
we identify X with a subset of t∗ via the differential.
We consider asymptotic C[~]-algebras U~ (g) and U~ (t). The latter algebra is a commutative
graded algebra which is clearly isomorphic to S~ := S(t)[~] where S(t), the symmetric algebra of
t, is equipped with its natural grading. Let Z(g) be the center of the algebra U (g). The Poincaré-
Birkhoff-Witt filtration on U (g) induces, by restriction, a filtration on Z(g). The corresponding
asymptotic algebra Z~ (g) is the center of the algebra U~ (g). One has a Harish-Chandra isomor-
phism Z~ (g) ∼ = SW W
~ = S(t) [~], a graded C[~]-algebra isomorphism induced by the composition
∼ ∼
Z~ (g) ֒→ U~ (g) ։ U~ (g)/(u · U~ (g) + U~ (g) · u− ) ←
− S~ −
→ S~
where the isomorphism on the right-hand side sends t ∈ t to t + ~ρ(t). Using this isomorphism,
we may (and will) identify the algebra Z~ (g) with a subalgebra of S~ .
For any λ ∈ t∗ , let S~ hhλii be the U~ (b)-bimodule defined as follows. As a C[~]-module, it is
isomorphic to S~ . The left U~ (b)-module structure is given by the natural isomorphism S~ ∼ =
U~ (b)/U~ (b) · n. Then in the right U~ (b)-module structure, the Lie ideal u ⊂ b acts by 0, and t ∈ t
acts by multiplication by t + ~λ(t).
We define a graded (U~ (b), U~ (g))-bimodule, a certain asymptotic version of the universal
Verma module, as follows:
M(λ) = S~ hhλ + ρii ⊗U~ (b) U~ (g).
5
(We will mainly only consider M(λ) as an (S~ , U~ (g))-bimodule.) The action of U~ (g) on the vector
vλ := 1 ⊗ 1 ∈ M(λ) induces an isomorphism of right U~ (g)-modules
M(λ) ∼ = U~ (g)/(u · U~ (g)). (2.1.2)
Under this isomorphism, the right S~ -module structure is such that the action of t ∈ t is induced
by right multiplication by t − ~ · (λ + ρ)(t) on U~ (g).
It is immediate from definitions that there is well defined ‘adjoint’ action b : m 7→ ad b(m) of
the Lie algebra b on M(λ), which is related to the bimodule structure by the equation
~ · ad b(m) = (b + ~ρ(b)) · m − m · b ∀b ∈ b.
L of the subalgebra t ⊂ b is semisimple. Therefore, one has a weight decompo-

The adjoint action
sition M(λ) = µ∈t∗ M(λ)µ . In particular, we have M(λ)−λ = C[~] · vλ , and M(λ)µ = 0 unless
µ ∈ −λ − Z≥0 R+ . For λ ∈ X, the adjoint action on M(λ) can be exponentiated to an algebraic
B-action.
Let Rep(G) be the tensor category of finite dimensional rational G-modules. For V ∈ Rep(G)
and λ ∈ X, let Vλ denote the T -weight space of V of weight λ.
The assignment ~ 7→ 1 ⊗ ~, x 7→ −x ⊗ ~ + 1 ⊗ x has a unique extension to an algebra homo-
morphism U~ (g) → U (g)op ⊗ U~ (g). Via this homomorphism, for any right U~ (g)-module M and
V ∈ Rep(G), the vector space V ⊗ M acquires the structure of a right U~ (g)-module. This gives
an (U~ (b), U~ (g))-bimodule structure on V ⊗ M(λ), where the left action of the algebra U~ (b) on
V ⊗ M(λ) comes from its action on M(λ) on the left. If λ ∈ X, the differential of the diagonal
B-action on V ⊗ M(λ) and the bimodule structure are related as follows: if b ∈ b ⊂ U (b), ~ times
the action of b is given by the assignment n 7→ (b + ~ρ(b)) · n − n · b.
One has a natural morphism of left S~ -modules pλ : M(λ) → S~ , induced by the projection
U~ (g) → U~ (b) orthogonal to U~ (g) · u− . If λ ∈ X, then pλ is also a morphism of T -modules
M(λ) → S~ ⊗C−λ . An important role below will be played by the morphism of graded S~ -modules
κalg
V,λ defined (for V in Rep(G) and λ ∈ X) as the following composition:
B T idV ⊗pλ
V ⊗ M(λ) ֒→ V ⊗ M(λ) / (V ⊗ S~ ⊗ C−λ )T = Vλ ⊗ S~ . (2.1.3)
B
Note that V ⊗ M(λ) has a natural structure of Z~ (g)-module, induced by the (right) action of
Z~ (g) ⊂ U~ (g) on M(λ). With this definition, κalg V,λ is also a morphism of Z~ (g)-modules, where
W
Z~ (g) = S~ acts on S~ via the restriction of the right action of S~ on S~ hhλii.
2.2. The affine Grassmannian: equivariant cohomology of cofibers. Write Gm for the multi-
plicative group. Let Ǧ be the Langlands dual group of G. The group Ǧ comes equipped with the
maximal torus Ť ⊂ Ǧ, with opposite Borel subgroups B̌ = Ť · Ǔ and B̌ − = Ť · Ǔ − , and with a
canonical isomorphism X = Hom(Gm , Ť ), the cocharacter lattice of Ť . (To be completely precise,
one should first choose Ǧ, B̌, Ť , and then use the affine Grassmannian of Ǧ to define G, B, T by
Tannakian formalism; see §6.1 for details.)
Let K = C((z)), resp. O = C[[z]]. Let GrǦ := Ǧ(K)/Ǧ(O), resp. GrŤ := Ť (K)/Ť (O), be the
affine Grassmannian associated with the group Ǧ, resp. Ť . (We will consider the reduced ind-
scheme structure on these affine Grassmannians.) Thus, one has X = GrŤ and there is a natural
embedding X = GrŤ ֒→ GrǦ . For λ ∈ X, we let λ be the image of λ and let iλ : {λ} ֒→ GrǦ denote
the one point embedding. The group Ǧ(K) ⋊ Gm acts on GrǦ on the left, where the factor Gm acts
by rotation of the loop.
For the rest of this section, we will use simplified notation Gr := GrǦ . The following subsets of
the affine Grassmannian will play an important role. For λ ∈ X+ , we let Grλ := Ǧ(O) · λ. This is
6
a finite dimensional (Ǧ(O) ⋊ Gm )-stable locally closed subvariety of Gr. One has a stratification
Gr = ⊔λ∈X+ Grλ . Further, for any λ ∈ X, following Mirković-Vilonen one puts Tλ := Ǔ − (K) · λ.
We let tλ : Tλ ֒→ Gr be the inclusion.
Let A := Ť × Gm , a toral subgroup of Ǧ(K) ⋊ Gm . The Mirković-Vilonen space Tλ is A-stable.
Further, the set X ⊂ Gr is known to be equal to the set of A-fixed points in Gr. Therefore, for
any object F of the equivariant derived category DA b (Gr), there are well defined A-equivariant
! q !
cohomology groups HA (Tλ , tλ F), resp. HA (iλ F). These are graded modules over the graded
q
algebra HA (pt) ∼ = S(t)[~] = S~ .
q
Let PervǦ(O) (Gr), resp. PervǦ(O)⋊Gm (Gr), be the category of Ǧ(O)-equivariant, resp. Ǧ(O)⋊ Gm -
equivariant, perverse sheaves on Gr. Let also PervǦ(O)-mon (Gr) be the category of perverse sheaves
on Gr which are constructible with respect to the stratification by Ǧ(O)-orbits. Recall that all three
of these categories are semisimple, with simple objects parametrized by X+ . In particular, the
forgetful functors
Perv Ǧ(O)⋊Gm (Gr) → PervǦ(O) (Gr) → PervǦ(O)-mon (Gr)
are equivalences of categories (see [MV, Appendix A] for a similar result in a much more general
situation). Let
∼
S : Perv Ǧ(O) (Gr) −
→ Rep(G)
be the geometric Satake equivalence. By the remark above, any object of PervǦ(O) (Gr) can be
considered naturally as an object of DA b (Gr).
The following lemma is a simple consequence of results of Kazhdan–Lusztig [KL] and Mirko-
vić–Vilonen [MV], cf. also [YZ, BrF]. It will be proved in §6.2.
Lemma 2.2.1. For any F in PervǦ(O) (GrǦ ) and λ ∈ X, one has

(1) The graded S~ -module HA (i!λ F), resp. the graded S(t)-module HŤ (i!λ F), is free.
q q
(2) There is a canonical isomorphism of graded S~ -modules, resp. of graded S(t)-modules

HA (Tλ , t!λ F) ∼
= S(F) λ ⊗ S~ hλ(2ρ̌)i, HŤ (Tλ , t!λ F) ∼
= S(F) λ ⊗ S(t)hλ(2ρ̌)i. (2.2.2)
q q
λ λ ı t
One may factor the embedding iλ : {λ} ֒→ Gr as a composition {λ} −→ Tλ −→ Gr. Hence, there
is a push-forward morphism
(ıλ )! : HA (i!λ F) = HA (ı!λ t!λ F) −→ HA (Tλ , t!λ F).

q q q
Let κtop
F ,λ be the following composite morphism
(ıλ )! (2.2.2)
κtop !
F ,λ : HA (iλ F)
/ H q (Tλ , t! F) / S(F) λ ⊗ S~ hλ(2ρ̌)i. (2.2.3)
q
A λ ∼
Thus, we get a diagram of morphisms of graded S~ -modules
B κalg
S(F ),λ
hλ(2ρ̌)i κtop
F ,λ
S(F) ⊗ M(λ) hλ(2ρ̌)i / S(F) ⊗ S~ hλ(2ρ̌)i o HA (i!λ F).
q
(2.1.3) λ (2.2.3)
One of our key results (to be proved in §7.6) reads as follows.
Theorem 2.2.4. For any F in PervǦ(O) (Gr) and λ ∈ X, the morphisms κalg top
S(F ),λ hλ(2ρ̌)i and κF ,λ are
injective and have the same image. Thus, there is a natural isomorphism of graded S~ -modules ζF ,λ that fits
7
into the following commutative diagram
B ζF ,λ
S(F) ⊗ M(λ) hλ(2ρ̌)i / H q (i! F)
_ ∼ A λ _
κalg
S(F ),λ
hλ(2ρ̌)i κtop
F ,λ
(2.2.5)

S(F) λ ⊗ S~ hλ(2ρ̌)i S(F) λ ⊗ S~ hλ(2ρ̌)i.
Remark 2.2.6. We have defined in §2.1 an action of Z~ (g) on (S(F) ⊗ M(λ))B . On the other hand,
it is explained in [BF, §2.4] that HA (i!λ F) also has a natural action of Z~ (g) = SW
~ coming from the
q
natural map Gr = (Ǧ(K)⋊ Gm )/(Ǧ(O)⋊ Gm) → pt/(Ǧ(O)⋊ Gm). We claim that our isomorphism
ζF ,λ is also Z~ (g)-equivariant.
First, the action of S~ ⊗Z~ (g) on HA (i!λ F) factors through an action of S~ ⊗C[~] Z~ (g) = C[t∗ ×
q
(t∗ /W ) × A1 ]. Then, the action of S~ ⊗C[~] Z~ (g) factors through the natural action of the algebra
HA (λ). Finally, it is explained in [BF, §3.2] that the S~ ⊗C[~] Z~ (g)-algebra HA (λ) is isomorphic to
q q
the direct image under the natural quotient map of OΓλ , where
Γλ := {(η1 , η2 , z) ∈ t∗ × t∗ × A1 | η2 = η1 + zλ}.
The claim easily follows from these remarks and the S~ -equivariance of ζF ,λ .
Remark 2.2.7. Consider the case F = ICν is the IC-sheaf associated with the Ǧ(O)-orbit Grν for
some ν ∈ X, and λ = w0 ν (where w0 ∈ W is the longest element). Then V ν := S(ICν ) is a
simple G-module with highest weight ν, and λ is the lowest weight of V ν . In view of the right-
hand isomorphism in Lemma 2.4.1 below, the image of the morphism κalg V ν ,w0 ν is computed by
Kashiwara in [Ka]: namely, with our conventions, combining Theorem 1.7 and Proposition 1.8 in
loc. cit. we obtain that the image of κalg ν ∼
V ν ,w0 ν in Vw0 ν ⊗S~ = S~ is generated by the following element:
 
Y −ν(wY 0 α)−1
 (α̌ − j~) . (2.2.8)
α∈R+ j=0
The topological context is easy in this case. Namely we have

 
−ν(w0 α)−1
Y Y
Tw0 ν ∩ Grν = Tw0 ν ∩ Grν = Ǔ − (O) · (w0 ν) ∼
=  C−α̌+j~ 
α∈R+ j=0
as A-varieties. One can easily deduce that the image of κtop ICν ,w0 ν is also generated by (2.2.8), see
§6.2. Hence, in this particular case, Theorem 2.2.4 can be directly deduced from these remarks.
In the case λ = ν, one can also directly check that both κalg top
V ν ,ν and κICν ,ν are isomorphisms.
2.3. Classical analogue. We will also prove an analogue of Theorem 2.2.4 where one replaces A
by Ť . In this case the representation theory of the algebra U~ (g) has to be replaced by the geometry
of the algebraic variety g∗ .
We will identify t∗ with the subspace (g/u ⊕ u− )∗ ⊂ g∗ . This way we obtain a canonical mor-
phism q : S(g/u) → S(t) induced by restriction of functions. For V in Rep(G) and λ ∈ X, the
“classical analogue” of the morphism κalg alg
V,λ , which we will denote by κV,λ , is the composition
B T idV ⊗q⊗1
V ⊗ S(g/u) ⊗ C−λ ֒→ V ⊗ S(g/u) ⊗ C−λ −−− −−−→ (V ⊗ S(t) ⊗ C−λ )T = Vλ ⊗ S(t).
This morphism is S(t)-equivariant, where the S(t)-action on the left-hand side is induced by the
morphism (g/u)∗ → t∗ given by restriction of linear maps.
8
Now we consider perverse sheaves on Gr. For F in PervǦ(O) (Gr) and λ ∈ X, we will denote by
κtop
F ,λ the following composite morphism
(ıλ )! (2.2.2)
κtop !
F ,λ : HŤ (iλ F)
/ H q (Tλ , t! F) / S(F) λ ⊗ S(t)hλ(2ρ̌)i.
q
Ť λ ∼
Then the classical analogue of Theorem 2.2.4 (to be proved in §7.7) reads as follows.
Theorem 2.3.1. For any F in PervǦ(O) (Gr) and λ ∈ X, the morphisms κalg top
S(F ),λ hλ(2ρ̌)i and κF ,λ are
injective and have the same image. Thus, there is a natural isomorphism of graded S(t)-modules ζ F ,λ that
fits into the following commutative diagram
B ζ F ,λ
S(F) ⊗ S(g/u) ⊗ C−λ hλ(2ρ̌)i / H q (i! F)
_ ∼ Ť λ _
κalg hλ(2ρ̌)i κtop (2.3.2)
S(F ),λ F ,λ

S(F) λ ⊗ S(t)hλ(2ρ̌)i S(F) λ ⊗ S(t)hλ(2ρ̌)i.
2.4. Alternative descriptions: differential operators on G/U and intertwining operators for
Verma modules. An important role in our arguments will be played by two alternative descrip-
B
tions of the C[~]-modules V ⊗ M(λ) .
If X is a smooth algebraic variety, we write DX for the sheaf of differential operators on X. The
sheaf DX comes equipped with a natural filtration by the order of differential operator. We let
D~,X be the corresponding sheaf of asymptotic differential operators. As for enveloping algebras,
this algebra has an alternative description as the sheaf of graded C[~]-algebras generated (locally)
by OX in degree 0 and the left OX -module TX (the tangent sheaf) in degree 2, with relations
ξ · ξ ′ − ξ ′ · ξ = ~[ξ, ξ ′ ] for ξ, ξ ′ ∈ TX and ξ · f − f · ξ = ~ξ(f ) for ξ ∈ TX and f ∈ OX . As for
enveloping algebras, we will use this description of D~,X and still denote by ξ the image of an
element ξ ∈ TX . (If we were using the description provided by (2.1.1), this element should rather
be denoted ~ξ.) Note that D~,X acts on OX [~] via ξ · f = ~ξ(f ) for ξ ∈ TX and f ∈ OX .
We put D(X) = Γ(X, DX ), resp. D~ (X) = Γ(X, D~,X ), for the corresponding algebra of global
sections. The order filtration makes D(X) a filtered algebra; the associated Rees algebra D(X)~ is
canonically isomorphic to D~ (X). There is also a canonical injective morphism D~ (X)/~·D~ (X) →
Γ(X, D~,X /~ · D~,X ), which is not surjective in general. Note finally that there exists a canonical
algebra isomorphism D~,X /~ · D~,X ∼ = (pX )∗ OT ∗ X , where T ∗ X is the cotangent bundle to X and
∗
pX : T X → X is the projection.
Consider the quasi-affine variety X := G/U . There is a natural G × T -action on X defined as
follows: g × t : hU 7→ ghtU. The T -action on X also induces an action of T on D~ (X) L by algebra
automorphisms. In particular, this T -action gives a weight decomposition D~ (X) = λ∈X D~ (X)λ .
Thus, D~ (X)0 = D~ (X)T is the algebra of right T -invariant asymptotic differential operators.
Differentiating the T -action on X yields a morphism γ : S~ → D~ (X) of graded C[~]-algebras.
Using this we will consider D~ (X) as an S~ -module where t ∈ t acts by right multiplication by
t − ~ρ(t). Note also that differentiating the G-action on X we obtain a morphism Z~ (g) → D~ (X).
This defines a Z~ (g)-module structure on D~ (X) (induced by multiplication on the left).
If M is an S~ -module and ϕ an algebra automorphism of S~ , we denote by ϕM the S~ -module
which coincides with M as a C-vector space,and where s ∈ S~ acts as ϕ(m) acts on M . If ϕ, ψ are
algebra automorphisms of S~ we have ψ ϕM = ϕ◦ψM . This construction provides an autoequiv-
alence of the category of S~ -modules, acting trivially on morphisms. We will use in particular this
notation when ϕ = w ∈ W (extended in the natural way to an automorphism of S~ ), and for the
9
∼
following automorphisms: if µ ∈ X, we denote by (µ) : S~ − → S~ the automorphism which sends
t ∈ t to t − ~µ(t) ∈ S~ . We will use similar notation for S(t)-modules.
If M, N are (S~ , U~ (g))-bimodules, we will write Hom(S~ ,U~ (g)) (M, N ) for the space of morphisms
of bimodules from M to N . It is an S~ -module and a Z~ (g)-module in a natural way. If M, N are
graded and M is finitely generated as a bimodule then this space is a graded S~ -module and a
Z~ (g)-module.
The following simple result will be proved in §3.1 (for the first isomorphism) and §4.3 (for the
second isomorphism).
Lemma 2.4.1. For any V in Rep(G) and λ ∈ X, there are canonical isomorphisms of graded S~ -modules
and Z~ (g)-modules.
G B
V ⊗ (λ)D~ (X)λ ∼
= V ⊗ M(λ) ∼
= Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)).
From Theorem 2.2.4 and Lemma 2.4.1 we deduce:
Corollary 2.4.2. For any F ∈ PervǦ(O) (Gr) and λ ∈ X, there are natural isomorphisms of graded S~ - and
Z~ (g)-modules
G
S(F) ⊗ (λ)D~ (X)λ hλ(2ρ̌)i ∼= HA (i!λ F) ∼= Hom(S~ ,U~ (g)) (M(0), S(F) ⊗ M(λ))hλ(2ρ̌)i.
q
One can also give an alternative description of equivariant cohomology of cofibers in the “clas-
sical case” of §2.3, as follows. Let B := G/B. For λ ∈ X, we denote by OB (λ) the line bundle
on B associated with the character −λ of B (so that ample line bundles correspond to dominant
weights). For any variety X over B, and any λ ∈ X, we will denote by OX (λ) the pull-back to
X of OB (λ). In particular, consider the G-varieties G/T × t∗ and e g := G ×B (g/u)∗ , which are
both equipped with a natural morphism to B. Both varieties are equipped with a natural action
of G × C× , where the action of G is induced by left multiplication on G, and any z ∈ C× acts by
multiplication by z −2 on (g/u)∗ or t∗ .
Consider the morphism
a : G/T × t∗ → e
g, (gT, η) 7→ (g ×B η)
where on the right-hand side η is considered as an element of g∗ trivial on u ⊕ u− . For any λ ∈ X
we have a canonical isomorphism a∗ Oeg (λ) ∼ = O ∗ (λ), so that we obtain a pull-back morphism
∗
∗
G/T ×t
a :Γ e g, Oeg (λ) → Γ G/T × t , OG/T ×t∗ (λ) .

Note that we have a canonical isomorphism Γ G/T × t∗ , OG/T ×t∗ (λ) ∼ = IndGT (−λ) ⊗ S(t) where
Ind is the usual induction functor for representations of algebraic groups, as defined e.g. in [J1,
§I.3.3]. Hence if V is in Rep(G), using the tensor identity and Frobenius reciprocity we obtain an
isomorphism
G T
V ⊗ Γ(G/T × t∗ , OG/T ×t∗ (λ)) ∼ = V ⊗ S(t) ⊗ C−λ = Vλ ⊗ S(t). (2.4.3)
By a similar argument, there exists a canonical isomorphism
G B
V ⊗ Γ(e g, Oeg (λ)) ∼ = V ⊗ S(g/u) ⊗ C−λ . (2.4.4)
Then one can easily check that, under isomorphisms (2.4.3) and (2.4.4), the morphism κalg
V,λ identi-
fies with the morphism
G G
V ⊗ Γ(eg, Oeg (λ)) → V ⊗ Γ(G/T × t∗ , OG/T ×t∗ (λ))
induced by a∗ .
Using (2.4.4), from Theorem 2.3.1 we deduce the following description.
10
Corollary 2.4.5. For any F ∈ Perv Ǧ(O) (Gr) and λ ∈ X, there exists a natural isomorphism of graded
S(t)-modules
G
HŤ (i!λ F) ∼
= S(F) ⊗ Γ(eg, Oeg (λ)) hλ(2ρ̌)i.
q
2.5. Weyl group symmetries. Each of the spaces in Corollaries 2.4.2 and 2.4.5 exhibits a kind of
symmetry governed by the Weyl group W . These symmetries play a technical role in our proofs
of Theorem 2.2.4 and 2.3.1. But we will also show that they are respected by the isomorphisms in
Corollaries 2.4.2 and 2.4.5. Some of our constructions are based on isomorphisms which do not
respect the gradings; hence for simplicity we just forget the gradings in this subsection.
The constructions on the side of the group G depend on the choice of root vectors for all simple
roots. For these constructions to match with the constructions in perverse sheaves on Gr, one has
to choose the root vectors provided by the Tannakian construction of G from the tensor category
PervǦ(O) (Gr); see §6.5 for details.
The symmetry in the case of equivariant cohomology of cofibers of perverse sheaves is easy to
∼
construct. Namely, the normalizer NǦ (Ť ) of Ť in Ǧ acts naturally on Gr; we denote by mg : Gr − →
Gr the action of g ∈ NǦ (Ť ). If we denote by g 7→ g the projection NǦ (Ť ) ։ NǦ (Ť )/Ť ∼
= W , for
any λ ∈ X we have mg ◦ iλ = igλ (where we identify the one-point varieties {λ} and {gλ}). If F is
in PervǦ(O) (Gr), we deduce an isomorphism of graded S~ -modules
∼ !
→ gHA (igλ
HA (i!λ m!g F) − F).
q q
On the other hand, since F is Ǧ(O)-equivariant (hence in particular NǦ (Ť )-equivariant), there ex-
∼
ists a canonical isomorphism m!g F ∼ = F. Hence we obtain an isomorphism HA (i!λ F) − → gHA (i!gλ F).
q q
Using classical arguments one can check that this isomorphism only depends on g (and not on g);
we will denote by
∼
ΞF ,λ q !
w : HA (iλ F) − → wHA (i!wλ F)
q
the isomorphism associated with w ∈ W . These isomorphisms generate an action of W in the

sense that for any F in PervǦ(O) (Gr), λ ∈ X and x, y ∈ W we have

y
ΞF
x
,yλ
◦ ΞF
y
,λ
= ΞF ,λ
xy . (2.5.1)
Now, let us consider D-modules on X. In §3.2 we recall a construction of Gelfand–Graev (see
[KLa, BBP]) based on “partial Fourier transforms” for D-modules which provides an action of W
on D~ (X) by algebra automorphisms such that, for any w ∈ W and λ ∈ X the action of w restricts
to an isomorphism of S~ -modules
(λ) ∼
→ w (wλ)D~ (X)wλ .
D~ (X)λ − (2.5.2)
For V in Rep(G), we will denote by
G ∼ G
ΦV,λ
w : V ⊗
(λ)
D~ (X)λ → w V ⊗ (wλ)D~ (X)wλ
−
the induced isomorphism. This collection of isomorphisms satisfies the relations
y

ΦV,yλ
x ◦ ΦV,λ
y = ΦV,λ
xy . (2.5.3)
Finally, in §4.7 we will define, for any V in Rep(G), λ ∈ X and w ∈ W , an isomorphism of
S~ -modules
∼
ΘV,λ → w Hom(S~ ,U~ (g)) (M(0), V ⊗ M(wλ)).
w : Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) −
This collection of isomorphisms naturally appears in the construction of the dynamical Weyl group,
see §2.6 below. As above, it satisfies the relations
y

ΘV,yλ
x ◦ ΘV,λ
y = ΘV,λ
xy . (2.5.4)
11
Our second main result, which will be proved in §7.6, is the following.
Theorem 2.5.5. The isomorphisms of Corollary 2.4.2 are such that the following diagram commutes for
any F in PervǦ(O) (Gr), λ ∈ X and w ∈ W :
G ∼ ∼
S(F) ⊗ (λ)D~ (X)λ / H q (i! F)
A λ
/ Hom(S ,U (g)) (M(0), S(F) ⊗ M(λ))
~ ~
≀ ΦS(F ),λ
w ≀ ΞF
w
,λ
≀ ΘS(F ),λ
w

w
G ∼ ∼

S(F) ⊗ (wλ)D~ (X)wλ / wH q (i! F)
A wλ
/ w Hom
(S~ ,U~ (g)) (M(0), S(F) ⊗ M(wλ)).
One can also give a “classical” analogue of Theorem 2.5.5. First, the same construction as above
provides, for any F in PervǦ(O) (Gr), λ ∈ X and w ∈ W , an isomorphism of graded S(t)-modules
F ,λ ∼
ξw : HŤ (i!λ F) −
→ wHŤ (i!wλ F).
q q
This collection of isomorphisms satisfies relations

y F ,yλ
F ,λ F ,λ
ξx ◦ ξy = ξxy .
On the other hand, let g∗r ⊂ g∗ be the open set consisting of regular elements, and let e
gr be the
inverse image of g∗r under the morphism π : e g → g∗ defined by π(g ×B η) = g · η. Then there
gr , which induces for any λ ∈ X and w ∈ W an isomorphism of
exists a canonical action of W on e
S(t)-modules and G-modules
∼ w
Γ e g, Oeg (λ) −→ Γ eg, Oeg (wλ) ,
see §5.2 for details. Hence we obtain, for any V in Rep(G), an isomorphism
V,λ
G ∼ w G
σw g, Oeg (λ)) −
: V ⊗ Γ(e → V ⊗ Γ(e g, Oeg (wλ)) .
This collection of isomorphisms again satisfies relations
y V,yλ

σx ◦ σyV,λ = σxy
V,λ
. (2.5.6)
We have the following compatibility property, to be proved in §7.7.
Theorem 2.5.7. The isomorphism of Corollary 2.4.5 is such that the following diagram commutes for any
F in PervǦ(O) (Gr), λ ∈ X and w ∈ W :
G ∼
S(F) ⊗ Γ(e
g, Oeg (λ) / H q (i! F)
Ť λ
S(F ),λ
≀ σw F ,λ
≀ ξw

w
G ∼
S(F) ⊗ Γ(e
g, Oeg (wλ) / wH q (i! F).
Ť wλ
2.6. Applications: dynamical Weyl groups and Brylinski–Kostant filtration. The first applica-
tion of our results concerns a geometric realization of the dynamical Weyl group due to Braverman–
Finkelberg ([BrF]). Let Q~ be the field of fractions of S~ . If M is a Q~ -module, we define the Q~ -
module wM by a similar formula as above. In §8.3 we will recall the definition of the dynamical
Weyl group, a collection of isomorphisms of Q~ -modules
∼
DWalg → wQ~ ⊗ Vwλ
V,λ,w : Q~ ⊗ Vλ −
for all V in Rep(G), λ ∈ X and w ∈ W .

12
Let O− := C[z −1 ], and let Ǧ(O− )1 be the kernel of the morphism Ǧ(O− ) → Ǧ given by evalu-
ation at z = ∞. For λ ∈ X we set Wλ := Ǧ(O− )1 · λ; this is a locally closed (ind-)subvariety of Gr
which is a transversal slice to the orbit Grλ at λ. We denote by
sλ : Wλ ∩ Tλ ֒→ Gr
the inclusion. We also set X
nλ := |λ(α̌)|
α∈R+
The following result is proved in [BrF]; we reproduce the proof in §8.4 since we will need some of
the details.
Lemma 2.6.1. For any F in PervǦ(O) (Gr) and λ ∈ X there exists a canonical isomorphism of S~ -modules

HA (Wλ ∩ Tλ , s!λ F) ∼
= S(F) λ ⊗ S~ hnλ i.
q
The inclusion {λ} ֒→ Wλ ∩ Tλ induces a morphism of Q~ -modules

∆F ,λ : Q~ ⊗S~ HA (i!λ F) → Q~ ⊗S~ HA (Wλ ∩ Tλ , s!λ F)
q q
which is an isomorphism due to the localization theorem in equivariant cohomology (see e.g. [EM,
Theorem B.2]). For F in PervǦ(O) (Gr), λ ∈ X and w ∈ W we define the morphism

DWgeom
F ,λ,w :=
w
∆F ,wλ ◦ ΞF ,λ
w ◦ (∆
F ,λ −1
) :
∼
→ w Q~ ⊗S~ HA (Wwλ ∩ Twλ , s!wλ F) .
Q~ ⊗S~ HA (Wλ ∩ Tλ , s!λ F) −
q q
The following result (which is a consequence of Theorem 2.5.5) is equivalent to the main result
of [BrF]. Our proof, given in §8.5, cannot really be considered as a new proof since it is based
on a similar strategy (namely reduction to rank 1), but we believe our point of view should help
understanding this question better.
Proposition 2.6.2. For any F in PervǦ(O) (Gr) and λ ∈ X dominant, the following diagram is commuta-
tive, where the vertical isomorphisms are induced by those of Lemma 2.6.1.
DWgeom
F ,λ,w
Q~ ⊗S~ HA (Wλ ∩ Tλ , s!λ F) / w Q~ ⊗S~ HA (Wwλ ∩ Twλ , s!wλ F)
q q
≀ ≀
DWalg
S(F ),λ,w

Q~ ⊗ S(F) λ / w Q~ ⊗ S(F) wλ
The second application of our results is a new proof of a result of the first author ([G1]) giving a
geometric construction of the Brylinski–Kostant filtration. Namely, let e ∈ u be a regular nilpotent
element which is a sum of (non-zero) simple root vectors. If V is in Rep(G) and λ ∈ X, the
Brylinski–Kostant filtration on Vλ associated with e is defined by
FBK
i (Vλ ) = {v ∈ Vλ | e
i+1
· v = 0} for i ≥ 0
and FBK
i (Vλ ) = 0 for i < 0. This filtration is independent of the choice of e (since all the choices for
e are conjugate under the action of T ).
On the other hand, for any ϕ ∈ t∗ we consider the specialized equivariant cohomology
Hϕ (i!λ F) := Cϕ ⊗S(t) HŤ (i!λ F),
q
a filtered vector space. Assume that ϕ ∈ t∗ r{0} satisfies (ad∗ e)2 (ϕ) = 0, where ad∗ is the coadjoint
representation. Then ϕ is regular, hence by the localization theorem in equivariant cohomology,
the morphism
Hϕ (i!λ F) → Cϕ ⊗S(t) (S(F))λ ⊗ S(t) = S(F) λ
13
induced by κtop F ,λ is an isomorphism. The left-hand side is equipped with a natural filtration; we
geom

denote by F q (S(F))λ the resulting filtration on S(F) λ .
The following result was first proved in [G1] (see also [AR] for a different proof). We observe in
§8.6 that it is an immediate consequence of Theorem 2.3.1.
Proposition 2.6.3. For any F in Perv Ǧ(O) (Gr), λ ∈ X and i ∈ Z we have

FBK
i (S(F))λ = Fgeom geom
2i+λ(2ρ̌) (S(F))λ = F2i+1+λ(2ρ̌) (S(F))λ .
Remark 2.6.4. Assume G is quasi-simple. Then it follows from [Ko, Corollary 8.7] that the 3-
dimensional representation of any sl2 -triple through e occurs only once in g, hence in g∗ . It follows
that ϕ ∈ t∗ is uniquely defined, up to scalar, by the condition (ad∗ e)2 (ϕ) = 0. Using this remark
one can easily check (for a general reductive G, and independently of Proposition 2.6.3) that the
filtration Fgeom
q is independent of the choice of ϕ once e is fixed. On the other hand all the possible
choices for e are conjugate under the action of T , hence our filtration is independent of any choice
(other than T and B).
3. D IFFERENTIAL OPERATORS ON THE BASIC AFFINE SPACE AND PARTIAL F OURIER TRANSFORMS
In Sections 3–5 we fix a complex connected reductive group G, and we use the notation of §2.1.
We also choose for any simple root α a non-zero vector eα ∈ gα . We denote by fα ∈ g−α the unique
vector such that [eα , fα ] = α̌.
3.1. Structure of D~ (X). The results in this subsection are taken from [BGG, Sh]. Below we will
use the two natural actions of G on C[G] induced by the actions of G on itself. The action given by
(g · f )(h) = f (g −1 h) for f ∈ C[G] and g, h ∈ G will be called the left regular representation; it is a left
action of G. The action given by (f · g)(h) = f (hg −1 ) for f ∈ C[G] and g, h ∈ G will be called the
right regular representation; it is a right action of G.
First we begin with the description of D~ (G). Differentiating the right regular representation
defines an anti-homomorphism of algebras U~ (g) → D~ (G). Then it is well known that multipli-
cation in D~ (G) induces an isomorphism of C[G]-modules
∼
C[G] ⊗ U~ (g) −
→ D~ (G). (3.1.1)
The left regular representation induces an action on D~ (G), which will be called simply the left ac-
tion below. Through isomorphism (3.1.1), it is given by the left regular representation of G on C[G]
(and the trivial action on U~ (g)). Similarly, the action induced by the right regular representation
(which will be called simply the right action below) corresponds, under isomorphism (3.1.1), to the
right action on C[G] ⊗ U~ (g) which is the tensor product of the right regular representation and
the action on U~ (g) which is the composition of the adjoint action with the anti-automorphism
g 7→ g −1 . There is also a natural morphism U~ (g) → D~ (G) obtained by differentiation of the left
regular representation. Under isomorphism (3.1.1), it is given by the map which sends m ∈ U~ (g)
to the function G ∋ g 7→ g−1 · m ∈ U~ (g), considered as an element in C[G] ⊗ U~ (g). In particular,
this morphism restricts to the morphism m 7→ 1 ⊗ m on Z~ (g) ⊂ U~ (g).
Let us recall the standard description of D~ (X) based on quantum hamiltonian reduction. As X
is a quasi-affine variety, the natural morphism
D~ (X) → EndC[~] (C[X][~])
is injective. The algebra C[X] identifies with the subalgebra of C[G] given by the elements fixed by
the right action of U ⊂ G. Any element D in D~ (G) induces a morphism C[X][~] → C[G][~]. This
morphism is trivial iff D ∈ C[G] ⊗ u · U~ (g) , and its image is contained in C[X][~] iff the image of
14

D in C[G] ⊗ U~ (g)/u · U~ (g) is U -invariant for the right action. In this way we obtain a canonical
isomorphism
U
D~ (X) ∼ = C[G] ⊗ U~ (g)/u · U~ (g) right . (3.1.2)
In this description, the action induced by the left G-action on X is induced by the left regular
representation on C[G]. The morphism Z~ (g) → D~ (X) obtained from the differentiation of this
left action of G on X corresponds to the morphism m 7→ 1 ⊗ (m mod u · U~ (g)).
Recall that there is also a T -action on X defined by t · gU = gtU . (Note that this action is not
induced by the right action of G on itself considered above, but rather by its composition with t 7→
t−1 .) This action provides a T -action on D~ (X) (where the action of t ∈ T is induced
Lby the right
action of t−1 ∈ G on D~ (G) described above) and a weight decomposition D~ (X) = λ∈X D~ (X)λ .
This T -action also defines a morphism γ : S~ → D~ (X) which, under isomorphism (3.1.2), is given
by γ(m) = 1 ⊗ (m mod u · U~ (g)). Recall that we consider D~ (X) as an S~ -module where t ∈ t acts
by right multiplication by γ(t) − ~ρ(t) · 1. Under isomorphism (3.1.2), this action is given by the
S~ -action on U~ (g)/u · U~ (g) where t ∈ t acts by left multiplication by t − ~ρ(t).
From this description (and isomorphism (2.1.2)) one easily obtains the following result.
Lemma 3.1.3. For any λ ∈ X there exists a canonical isomorphism of graded S~ -modules and Z~ (g)-
modules
(λ)
D~ (X)λ ∼= IndG B (M(λ)).
Proof of the first isomorphism in Lemma 2.4.1. We have

G G B
V ⊗ (λ)D~ (X)λ ∼
= V ⊗ IndG ∼
= V ⊗ M(λ)
B (M(λ))
where the first isomorphism follows from Lemma 3.1.3 and the second one from the tensor identity
and Frobenius reciprocity.
3.2. Partial Fourier transforms for D~ (X). Let us recall a construction due to Gelfand–Graev,
and studied by Kazhdan–Laumon [KLa] in the ℓ-adic setting and by Bezrukavnikov–Braverman–
Positselskii [BBP] in our D-module setting. We choose a reductive group Gsc with simply-con-
nected derived subgroup and a surjective group morphism Gsc ։ G with finite central kernel
denoted Z. We denote by T sc , B sc the inverse images of T , B in Gsc , and let U sc be the unipo-
tent radical of B sc . We set Xsc := Gsc /U sc . Note that Z acts naturally on Xsc , with quotient X.
Note also that for any simple root α there exists a unique injective morphism of algebraic groups
ϕα : SL(2, C) → Gsc such that

× z 0 0 0 0 1
∀z ∈ C , ϕα = α̌(z) and d(ϕα ) = fα , d(ϕα ) = eα
0 z −1 1 0 0 0
(where we identify the Lie algebras of Gsc and G.)
Let α be a simple root, let Pαsc be the minimal parabolic subgroup of Gsc containing B sc associ-
ated with α, and let Qsc sc sc
α := [Pα , Pα ]. Consider the projection
τα : Xsc → Gsc /Qsc
α.
It is explained in [KLa, §2.1] that τα is the complement of the zero section of a Gsc -equivariant
vector bundle
τα′ : Vα → Gsc /Qsc
α
of rank 2. Moreover, there exists a canonical Gsc -equivariant symplectic form on this vector bundle
(which depends on ϕα , i.e. on the choice of fα ). Hence the constructions recalled in §B.2 provide
an automorphism of D~ (Vα ) as a C[~]-algebra. As the complement of Xsc in Vα has codimension
∼
2, restriction induces an isomorphism D~ (Vα ) − → D~ (Xsc ). Hence we obtain an automorphism
15
∼
Fsc sc
α : D~ (X ) − → D~ (Xsc ). This automorphism is Z-equivariant (since Z acts on Vα by sym-
plectic automorphisms), and we have D~ (Xsc )Z = D~ (X) in a natural way. Hence we obtain a
G-equivariant C[~]-algebra automorphism
∼
Fα : D~ (X) −
→ D~ (X).
on Z-fixed points. Using the fact that any two simply connected covers of a connected semi-
simple group are isomorphic (as covers of the given group), one can check that the automorphism
Fα does not depend on the choice of Gsc .
Lemma 3.2.1. (1) The automorphisms Fα , α a simple root, generate an action of W on D~ (X).
(2) For any simple root α and any λ ∈ X, setting s = sα , Fα restricts to an isomorphism of G-modules
and of S~ - and Z~ (g)-modules
∼
→ s D~ (X)sλ .
D~ (X)λ −
Proof. We observe that there are natural isomorphisms

D~ (X)[~−1 ] ∼
= Γ X, D~,X [~−1 ] ∼= Γ X, DX ⊗ C[~, ~−1 ] ∼= D(X) ⊗ C[~, ~−1 ].
Moreover, under these isomorphisms, the automorphism induced by Fα coincides with the tensor
product of the similar automorphism of D(X) considered in [BBP] with idC[~,~−1] . Hence the lemma
follows from [BBP, Proposition 3.1 and Lemma 3.3].
By (1), we can define a group morphism w 7→ Fw from W to the group of C[~]-algebra auto-
morphisms of D~ (X), such that Fsα = Fα for any simple root α. And by (2) these isomorphisms
restrict to isomorphisms of G-modules and of S~ - and Z~ (g)-modules
∼
Fλw : D~ (X)λ −
→ w
D~ (X)w(λ)
which satisfy the relations
y

Fyλ
x ◦ Fλy = Fλxy .
Using the relation (λ)(wM ) = w((wλ)M ) we obtain isomorphism (2.5.2), which allows to define the
collection of isomorphisms ΦV,λ
w of §2.5.
Remark 3.2.2. (1) As explained in §B.1, the construction of isomorphisms ΦV,λ

w is not compati-
ble with the natural gradings on (V ⊗ (λ)D~ (X)λ )G and (V ⊗ (wλ)D~ (X)wλ )G . However, it
follows from Theorem 2.5.5 that it induces an isomorphism of graded modules
G ∼ G
V ⊗ (λ)D~ (X)λ hλ(2ρ̌)i −→ w V ⊗ (wλ)D~ (X)wλ h(wλ)(2ρ̌)i.
This property is also observed in [LS, Proposition 2.9].
(2) Statement (2) of Lemma 3.2.1 (and the fact that Fα is an involution) can also be proved
directly as follows: by Corollary 3.3.3 below and the injectivity of the morphisms RV,λ G,L
considered in this statement (see Lemmas 4.6.4 and 4.8.2), it is enough to prove the claim
in the case G has semisimple rank one, which can be treated by explicit computation (see
e.g. the proof of Lemma A.2.1 below).
As a consequence of these constructions we also obtain the following result, which will be
needed later. This result is also proved (using different methods) in [Sh].
Proposition 3.2.3. For any λ ∈ X, the graded S~ -module D~ (X)λ is free.
16
Proof. Using isomorphisms Fλw defined above we can assume that λ is dominant. Then using
Lemma 3.1.3 it is equivalent to prove that IndG
B (M(λ)) is free over S~ .
We claim that, if λ is dominant,
IndG
B (S(g/u)[~] ⊗ C−λ ) is free over S~ and R
>0
IndG
B (S(g/u)[~] ⊗ C−λ ) = 0. (3.2.4)
Indeed, it is sufficient to prove that the S(t)-module IndGB (S(g/u) ⊗ C−λ ) is free and that we have
R>0 IndG
B (S(g/u) ⊗ C −λ ) = 0. Consider the vector bundles
qe :Ne → B and qeg : e g → B.
N
Here N e := G ×B (g/b)∗ is the Springer resolution, and e g is defined in §2.4. There is a natural
e
inclusion of vector bundles N ֒→ g, and the quotient is the trivial vector bundle t∗ ×B. Hence there
e
is a Z≥0 -filtration on (qeg )∗ Oeg (as a sheaf of S(t) ⊗ OB -modules) with associated graded (qNe )∗ ONe ⊗
S(t). By [Bro, Theorem 2.4], we have H>0 (N e , O e (λ)) = 0. It follows that H>0 (e
g, Oeg (λ)) = 0, and
N
0
that H (e 0 e
g, Oeg (λ)) has a filtration with associated graded H (N , ONe (λ)) ⊗ S(t). Now it follows from

definitions that, for any i ≥ 0, Hi (e g, Oeg (λ)) ∼
= Ri IndG B S(g/u) ⊗ C−λ , which finishes the proof of
(3.2.4).
For i ≥ 0, let Mi := S~ ·vλ ·U~≤i (g) ⊂ M(λ), where U~≤ (g) is the PBW filtration of U~ (g). Then M q
q
is a B-stable and S~ -stable exhaustive filtration of M(λ), and the associated graded is isomorphic
to S(g/u)[~] ⊗ C−λ . From the second claim in (3.2.4) it follows that IndG B (M(λ)) has a filtration
with associated graded IndG B (S(g/u)[~] ⊗ C −λ ), and then the corollary follows from the first claim
in (3.2.4).
Remark 3.2.5. The arguments in the proof of Proposition 3.2.3 also prove that, when λ is dominant,
we have R>0 IndGB (M(λ)) = 0.
3.3. Restriction to a Levi subgroup. Fix a subset I of the set of simple roots and let l be the
Levi subalgebra containing t that has the set I as simple roots. Note that our choice of a Borel
subgroup, a maximal torus and simple root vectors for g determines a similar choice for l, hence
the constructions of the present section make sense both for g and for l.
We put
M M M
uL := gα , bL := t ⊕ uL , n+
L := gα , n−L := gα .
α∈R+ α∈R+ α∈R−
α∈ZI α∈ZI
/ α∈ZI
/
−
Thus, one has a triangular decomposition g = n+ L ⊕ l ⊕ nL , and uL = u ∩ l is the nilradical
of the Borel subalgebra bL = b ∩ l of l. Further, let p± := l ⊕ n± ±
L and b± := bL ⊕ nL , resp.
±
u± := uL ⊕ nL . Thus, p± is a pair of opposite parabolic subalgebras of g such that p+ ∩ p− = l,
and b± is a pair of Borel subalgebras of g such that b+ ∩ b− = bL , with respective nilpotent
radicals u± . Let L, P± , B± , U± , NL± , BL , UL be the subgroups of G corresponding to the Lie alge-
bras l, p± , b± , u± , n±
L , bL , uL , respectively. By definition, we have XG = G/U+ and XL = L/UL .
(Observe that B+ , U+ , b+ , u+ coincide with the objects denoted by B, U, b, u in the preceding
sections.)
Now we construct a morphism of L-modules
rLG : D~ (XG ) → D~ (XL )
as follows. Note that the right U~ (l)-action on U~ (g)/u · U~ (g) descends to a well-defined action on
U~ (g)/(u · U~ (g) + U~ (g) · n− L ). Using this, from the diagram g ←֓ p− ։ l of natural Lie algebra
morphisms, one obtains the following morphisms of right U~ (l)-modules:
L = U~ (p− )/(uL · U~ (p− ) + U~ (p− ) · n− ) ∼
U~ (g)/(u · U~ (g) + U~ (g) · n− ) ∼ L = U~ (l)/uL · U~ (l). (3.3.1)
17
All the above maps are bijections since the linear maps
∼ ∼
g/(u ⊕ n−
L)
o p− /u− / l/uL
are clearly vector space isomorphisms. We deduce the following chain of maps
U U
C[G] ⊗ (U~ (g)/u · U~ (g)) ֒−→ C[G] ⊗ (U~ (g)/u · U~ (g)) L
UL
−→ C[G] ⊗ U~ (g)/(u · U~ (g) + U~ (g) · n−
L)
U
−→ C[G] ⊗ (U~ (l)/uL · U~ (l)) L
U
−→ C[L] ⊗ (U~ (l)/uL · U~ (l)) L ,
where the third morphism is induced by (3.3.1), and the last one is induced by restriction of func-
tions C[G] → C[L]. Using isomorphism (3.1.2), this allows to define the desired morphism
U U
= C[G] ⊗ (U~ (g)/u · U~ (g) → C[L] ⊗ (U~ (l)/uL · U~ (l)) L ∼
rLG : D~ (XG ) ∼ = D~ (XL ).
We also define an automorphism of L-modules
∼
tG
L : D~ (XL ) −
→ D~ (XL )
as follows. The linear map ρG − ρL on t extends naturally to a Lie algebra morphism l → C, which
we denote similarly. Then the assignment l ∋ x 7→ x + ~(ρG − ρL )(x) defines a graded C[~]-algebra
automorphism ıG L of U~ (l), which descends to a UL -equivariant automorphism of the quotient
UL
∼
U~ (l)/(uL · U~ (l)). Using the isomorphism D~ (XL ) = C[L] ⊗ (U~ (l)/uL · U~ (l)) as above, we
obtain the wished for automorphism tG L . This morphism can also be described in more geometric
terms as follows: the linear form ρG − ρL on t can be considered as a character of T , which extends
in a natural way to a character of L, and then descends to an invertible function fLG on L/UL . Then
one can easily check that tG G −1 · D · f G .
L is the automorphism of D~ (XL ) sending D to (fL ) L
Finally we define the morphism of L-modules
resG G G
L := tL ◦ rL : D~ (XG ) → D~ (XL ).
One can easily check that resG

L is S~ -equivariant, and T -equivariant for the T = {1} × T -actions.
The following result (in which we use superscripts on the left to indicate which reductive group
we consider) will be proved in §3.4 below.
Proposition 3.3.2. Let α ∈ I. Then the following diagram commutes:

GF
α
D~ (XG ) / D~ (XG )
resG
L resG
L
LF
α
D~ (XL ) / D~ (XL ).
As a corollary we obtain the following result. For λ ∈ X and V in Rep(G), we denote by

G L
RV,λ
G,L : V ⊗
(λ)
D~ (XG )λ → V|L ⊗ (λ)D~ (XL )λ
the morphism induced by resG

L.
18
Corollary 3.3.3. For any λ ∈ X, V in Rep(G) and w ∈ WL ⊂ WG , the following diagram commutes:
G GΦV,λ
w G
V ⊗ (λ)D~ (XG )λ / w V ⊗ (wλ)D~ (XG )wλ
RV,λ
G,L
wRV,wλ
G,L
V ,λ
L LΦ |L
w L
V|L ⊗ (λ)D~ (XL )λ / w V|L ⊗ (wλ)D~ (XL )wλ .
3.4. Proof of Proposition 3.3.2. If G has simply-connected derived subgroup, then so does L, as
well. Hence we can assume that G = Gsc .
In the next lemma, if X is any variety we consider C[X][~] as the algebra of functions on X
with values in C[~]. The subset P− · P+ ⊂ G is an open subvariety, so that if x is in U~ (g) and
f is in C[P− · P+ ][~], it makes sense to consider x · f ∈ C[P− · P+ ][~], and also the restriction
(x·f )|L ∈ C[L][~]. (Here we consider the right action of U~ (g) on C[G][~] obtained by differentiating
the right regular representation of §3.1.)
Lemma 3.4.1. Let f ∈ C[P− · P+ ][~] be a left NL− -invariant function. Then for any x ∈ U~ (g) · n−
L we
have
(x · f )|L = 0.
Proof. For any y ∈ U~ (g) the function y · f is again left NL− -invariant, so that we can assume that
x ∈ n− − −
L . Then the result follows from the observation that for g ∈ L we have g · NL = NL · g.
In the next lemma we use the embedding XL = L/UL = P+ /U+ ֒→ P− · P+ /U+ ⊂ G/U+ = XG .
Lemma 3.4.2. For any left NL− -invariant function f ∈ C[P− · P+ /U+ ][~] and any D ∈ D~ (XG ) we have
D(f )|XL = (rLG D)(f|XL ).
Proof. The element D ∈ D~ (XG ) induces a morphism C[P− · P+ /U+ ][~] → C[P− · P+ ][~]. Similarly,
the restriction morphism C[P− · P+ /U+ ][~] → C[XL ][~] is the restriction to right U+ -invariants
of the restriction morphism C[P− · P+ ][~] → C[P+ ][~]. Hence it is enough to show that if f ∈
C[P− · P+ ][~] is left NL− -invariant then the morphism D~ (G) → C[L][~] sending D ′ to D ′ (f )|L
factors (via isomorphism (3.1.1)) through the quotient

C[G] ⊗ U~ (g) → C[G] ⊗ U~ (g)/(U~ (g) · n−
L) .
This fact follows from Lemma 3.4.1.
As a corollary of Lemma 3.4.2 we obtain the following description of the morphism rLG . We
denote by εL : D~ (NL− ) → C[~] the morphism sending a differential operator D to the value at
1 ∈ NL− of the function D(1N − ), where 1N − is the constant function with value 1.
L L
Corollary 3.4.3. The morphism rLG coincides with the composition

∼ D − ⊗D7→εL (D − )·D
→ D~ (NL− ) ⊗C[~] D~ (XL ) −−−−−−−−−−−−→ D~ (XL )
D~ (XG ) ֒→ D~ (P− · P+ /U+ ) −
where the first morphism is restriction to the open subset P− · P+ /U+ ⊂ XG , and the second morphism
uses the isomorphism P− · P+ /U+ ∼ = NL− × XL induced by the action of NL− on XG .
Proof. As the action of D~ (XL ) on C[XL ][~] is faithful, it is enough to check the claim after acting
on any f ∈ C[XL ]. However one can write f = f˜|XL where f˜ = 1N − ⊗ f ∈ C[P− · P+ /U+ ], and then
L
the claim follows from Lemma 3.4.2.
Finally we can give the proof of Proposition 3.3.2.
19
Proof of Proposition 3.3.2. It is enough to prove the commutativity of the following two diagrams:
GF LF
α α
D~ (XG ) / D~ (XG ) D~ (XL ) / D~ (XL )
G
rL G
rL tG tG
L L
LF LF
α α
D~ (XL ) / D~ (XL ), D~ (XL ) / D~ (XL ).
The commutativity of the right-hand diagram follows from Lemma B.2.2 below, hence we only
have to consider the left-hand diagram. Now we observe that (since the construction of partial
Fourier transform is local on the base of the vector bundle) the automorphism GFα extends to an
automorphism of D~ (P− · P+ /U+ ) denoted similarly, which makes the following diagram com-
mutative, where vertical morphisms are induced by restriction:
GF
α
D~ (X G ) / D~ (XG )
_
_
GF
α
D~ (P− · P+ /U+ ) / D~ (P− · P+ /U+ ).
Next, by construction the following diagram commutes:

GF
α
D~ (P− · P+ /U+ ) / D~ (P− · P+ /U+ )
≀ ≀

id⊗LFα
D~ (NL− ) ⊗C[~] D~ (XL ) / D~ (N − ) ⊗C[~] D~ (XL ).
L
Then the commutativity follows from Corollary 3.4.3.
3.5. Classical analogue. Consider the sheaf of algebras

AX := D~,X /~ · D~,X
on X. This sheaf is canonically isomorphic to (pX )∗ OT ∗ X , where pX = T ∗ X → X is the natural
projection. We also set A(X) := Γ(X, AX ). If α is a simple root, we define similarly the sheaf of
rings AVα on Vα , and the ring A(Vα ). In §3.2 we have defined an automorphism of D~,Vα as a
sheaf of C[~]-algebras. Specializing to ~ = 0 we deduce an automorphism of AVα . By the same
∼
arguments as in §3.2, restriction induces an isomorphism A(Vα ) − → A(X), hence we obtain an
algebra automorphism
∼
Fα : A(X) −
→ A(X)
such that the following diagram commutes, where vertical arrows are induced by the natural
quotient morphism D~,X → AX :
Fα
D~ (X) ∼
/ D~ (X)
(3.5.1)
Fα

A(X) / A(X).
∼
The following lemma will be used to deduce Theorem 2.3.1 from Theorem 2.2.4.
Lemma 3.5.2. The natural morphism D~ (X) → A(X) induces an isomorphism
∼
D~ (X)/~ · D~ (X) −
→ A(X).
20
In other words, for any V in Rep(G) and λ ∈ X, the morphism
B B
V ⊗ M(λ) → V ⊗ S(g/u) ⊗ C−λ
induced by the quotient morphism M(λ) → M(λ)/(~ · M(λ)) ∼ = S(g/u) ⊗ C−λ induces an isomorphism
B ∼ B
V ⊗ M(λ) /~ − → V ⊗ S(g/u) ⊗ C−λ .
Proof. By Lemma 3.1.3, both statements are equivalent to the fact that for any λ ∈ X the morphism

IndG G
B (M(λ))/~ → IndB S(g/u) ⊗ C−λ
is an isomorphism. In the case λ is dominant, this property follows from the exact sequence of
B-modules
~ / / S(g/u) ⊗ C−λ
M(λ) / M(λ)
and the cohomology vanishing R1 IndG B (M(λ)) = 0, see Remark 3.2.5. Since Fα induces an iso-
∼
morphism D~ (X)λ − → D~ (X)sα λ (see Lemma 3.2.1), using diagram (3.5.1) we deduce the general
case from the case λ is dominant.
It follows in particular from (3.5.1) and Lemma 3.5.2 (using Lemma 3.2.1) that the assignment
sα 7→ Fα defines an action of W on A(X) by algebra automorphisms, which we denote by w 7→ Fw .
∼
Moreover, for any w ∈ W and λ ∈ X, Fw defines an isomorphism of S(t)-modules A(X)λ − →
wA(X) .
wλ
3.6. Complementary results on the structure of D(X). In this subsection we observe that Lemma
3.5.2 has some interesting consequences on the structure of D(X). These results will not be used
in the rest of the paper.
We begin with the following direct consequence of Lemma 3.5.2 (using the natural isomorphism
A(X) ∼= C[T ∗ X]), which appears to be new.
Corollary 3.6.1. The canonical graded algebra morphism gr D(X) → C[T ∗ X] is an isomorphism.
This corollary allows to give new proofs of some results of [LS] and [BBP]. These proofs use the
following simple lemma.
Lemma 3.6.2. The C-algebra C[T ∗ X] is finitely generated, hence noetherian.
Proof. First we observe that there exists a canonical isomorphism T ∗ X ∼
= G ×U (g/u)∗ . Hence T ∗ X
is a T -torsor over e
g, which implies that we have a natural algebra isomorphism
M
C[T ∗ X] ∼
= Γ e g, Oeg (λ) .
λ∈X
By the same observation, there exists a natural morphism T ∗ X → g∗ ×t∗ /W t∗ , g ×U η 7→ (g ·η, η|t ),
which induces an algebra morphism S(g) ⊗S(t)W S(t) → C[T ∗ X]. Note also that if λ ∈ X+ , then
Γ(B, OB (λ)) identifies naturally with a subspace of C[X], hence also defines a subspace Xλ of
C[T ∗ X] using the projection T ∗ X → X. Pr
Let λ1 , · · · , λr be a finite collection of dominant weights such that X+ = i=1 Z≥0 λi . We
∗
claim that C[T X] is generated (as an algebra) by S(g) ⊗S(t)W S(t) and the G-modules Xλi for
i ∈ {1, · · · , r}, together with the images of these subspaces under the automorphisms Fw for all
w ∈ W . This claim clearly implies the statement of the lemma.
To prove the claim we first observe that if λ ∈ X+ the morphism

g, Oeg (λ)
S(g) ⊗S(t)W S(t) ⊗ Xλ → Γ e
21
induced by the product in C[T ∗ X] is surjective. Indeed by the graded Nakayama lemma it is
enough to prove surjectivity after tensoring with the trivial S(g)⊗S(t)W S(t)-module, hence it is also
enough to prove surjectivity after tensoring with the trivial S(t)-module. However the arguments
in the proof of Proposition 3.2.3 imply that the natural morphism

g, Oeg (λ) → Γ N
C ⊗S(t) Γ e e , O e (λ)
N
induced by restriction is an isomorphism, hence the latter surjectivity statement follows from
[Bro, Proposition 2.6]. From this observation, together with the fact that if λ, µ ∈ X+ the natural
morphism
Γ(B, OB (λ)) ⊗ Γ(B, OB (µ)) → Γ(B, OB (λ + µ))
is surjective (see e.g. [BK, Theorem 3.1.2]), it follows that the subalgebra of C[T ∗ X] generated by
S(g) ⊗S(t)W S(t) and the modules Xλi contains
M
g, Oeg (λ) .
Γ e
λ∈X+
The claim follows, using the W -action and the fact that every weight is W -conjugate to a dominant
weight.
In the following corollary, statement (1) is due to Levasseur–Stafford (see [LS, Theorem 3.3]).
The present proof is suggested in [LS, Remark 3.4], but the authors didn’t have Corollary 3.6.1 to
complete the argument. Statement (2) is due to Bezrukavnikov–Braverman–Positselskii (see [BBP,
Theorem 1.1]).
Corollary 3.6.3. The algebra D(X) is

(1) Auslander–Gorenstein and Cohen–Macaulay;
(2) left and right noetherian.
P some elements µ̌i ∈ X∗ (T ), i = 1, · · · s, such that if λ(µ̌i ) = 0 for all i then λ = 0,

Proof. (1) Choose
and set kλk := i |λ(µ̌i )|. Consider the filtration on D(X) defined for n ≥ 0 by
 
M
Fn D(X) := D(X)≤n ∩  D(X)λ  ,
kλk≤n
where D(X)≤ is the filtration by order of differential operators. Then this filtration is connected
q
(in the sense that F0 D(X) = C) and the associated graded is canonically isomorphic to C[T ∗ X] by
Corollary 3.6.1. Hence the claim follows from [YeZ, Corollary 0.3], using the simplicity of D(X)
proved in [LS, Proposition 3.1] and Lemma 3.6.2.
(2) It is enough to prove that D~ (X) is (left and right) noetherian. But this follows from [ATV,
Lemma 8.2] (for g = ~) and Lemma 3.6.2.
4. M ORPHISMS BETWEEN ASYMPTOTIC UNIVERSAL V ERMA MODULES
In this section we will use the following convention. If P(µ) is a property depending on µ ∈ t∗ ,
we will say that P(µ) holds “for µ ∈ t∗ sufficiently large” is there exists n ∈ Z≥0 such that P(µ)
holds for any µ ∈ t∗ satisfying |µ(α̌)| ≥ n for all simple roots α. We will use similar conventions
for subsets of t∗ (e.g. X).
22
4.1. Reminer on Verma modules. The results in this subsection are well known, see e.g. [TV, EV].
We include (short) proofs for the reader’s convenience. We will use the “dot-action” defined by
w • µ = w(µ + ρ) − ρ.
For any µ ∈ t∗ we consider the Verma module
V(µ) := U (g) ⊗U (b) Cµ
(a left U (g)-module). We also set 1µ := 1 ⊗ 1 ∈ V(µ). We set V(µ)− := U (u− )u− · 1µ , so that we
have V(µ) = C · 1µ ⊕ V(µ)− .
If V is in Rep(G), λ ∈ X, µ ∈ t∗ , and if φ ∈ HomU (g) (V(µ), V ⊗ V(µ − λ)), then one can write
φ(1µ ) = u ⊗ 1µ−λ + x
for unique u ∈ Vλ and x ∈ V ⊗ V(µ − λ)− . We set u := EV,λ µ (φ); it is called the expectation value of φ.
Let µ ∈ t∗ and let α be a simple root. If n := µ(α̌) + 1 ∈ Z≥0 , then as in [Hu, §1.4] there exists a
unique embedding of U (g)-modules
V(sα • µ) ֒→ V(µ)
)n
which sends 1sα •µ to (fn! α
· 1µ . Iterating, we perform the following construction. Let µ ∈ X+ − ρ,
and let w ∈ W . Choose a reduced decomposition w = sk · · · s1 , where α1 , · · · , αk is a sequence of
simple roots (possibly with repetitions) and si is the reflection associated with αi . Then for any
i = 1, · · · , k, ni := (si−1 · · · s1 • µ)(α̌i ) + 1 ∈ Z≥0 . Moreover the collection (n1 , · · · , nk ) and the
product (fαk )nk · · · (fα1 )n1 do not depend on the reduced decomposition of w (see [TV, Lemma
4]). Hence there exists a unique embedding of graded right U (g)-modules
(fαk )nk (fα1 )n1
ιw w
µ : V(w • µ) ֒→ V(µ) such that ιµ (1w•µ ) = ··· · 1µ .
nk ! n1 !
Lemma 4.1.1. Let V in Rep(G), and λ ∈ X.
(1) For µ ∈ t∗ sufficiently large, the morphism

EV,λ
µ : HomU (g) V(µ), V ⊗ V(µ − λ) → Vλ
is an isomorphism.
(2) For w ∈ W and µ ∈ X+ sufficiently large, the morphism

HomU (g) V(w • µ), V ⊗ V(w • (µ − λ)) → HomU (g) V(w • µ), V ⊗ V(µ − λ)
defined by φ 7→ (idV ⊗ ιw
µ−λ ) ◦ φ is an isomorphism
Proof. (1) It is well known (see e.g. [Hu, Theorem 3.6]) that there exists an enumeration ν1 , · · · , νk
of the T -weights of V and a filtration (as a U (g)-module)
{0} = M0 ⊂ M1 ⊂ · · · ⊂ Mk = V ⊗ V(µ − λ)
where for all i = 1, · · · , k, Mi /Mi−1 ∼
= Vνi ⊗ V(µ − λ + νi ). Consider, for any i = 1, · · · , n, the
associated exact sequence

0 → HomU (g) V(µ), Mi−1 → HomU (g) V(µ), Mi

→ HomU (g) V(µ), V(µ − λ + νi ) ⊗ Vνi → Ext1U (g) V(µ), Mi−1 .
It is also well known that

HomU (g) V(η1 ), V(η2 ) = Ext1U (g) V(η1 ), V(η2 ) = 0

unless η1 ∈ W • η2 , and HomU (g) V(η), V(η) = C. Now if µ ∈ t∗ is sufficiently large, the property
µ ∈ W • (µ − λ + νi ) implies that w = 1 and λ = νi , and the result follows.
23
(2) First we remark that the morphism under consideration is indeed well defined if µ ∈ X+ is
sufficiently large. As ιwµ−λ is injective, our morphism is injective. Hence it is enough to prove that
both sides have the same dimension for µ ∈ X+ sufficiently large.
By (1), if µ is sufficiently large the left-hand side is isomorphic to Vwλ . Now similar arguments,
using the property that that if η ∈ X+ − ρ then
HomU (g) (V(w • η), V(η)) = C,
show that, again if µ is sufficiently large, the right-hand side is isomorphic to Vλ . As dimC (Vλ ) =
dimC (Vwλ ), this finishes the proof.
4.2. Asymptotic Verma modules. If µ ∈ t∗ , we denote by C[~]hhµii the graded right S~ -module
where any t ∈ t acts by multiplication by ~µ(t).
For µ ∈ t∗ we define the graded right U~ (g)-module
M(µ) := C[~]hh0ii ⊗S~ M(µ).
We also set vµ := 1 ⊗ vµ ∈ M(µ). Any t ∈ t acts on vµ by multiplication by ~(µ + ρ)(t), and M(µ)
admits a basis (as a C[~]-module) such that, for any vector v in this basis, there exists γ ∈ Z≥0 R+
such that v · t = (µ + ρ + γ)(t) for any t ∈ t. We set M(µ)− := vµ · u− U~ (u− ). Then we have
M(µ) = C[~] · vµ ⊕ M(µ)− .
Note also that for any λ, µ ∈ t∗ there exists a natural isomorphism
C[~]hhµii ⊗S~ M(λ) ∼
= M(λ + µ) (4.2.1)
sending 1 ⊗ vλ to vλ+µ .
It will be convenient to invert ~. To simplify notation we set
Mloc (µ) := C[~, ~−1 ] ⊗C[~] M(µ), Uloc (g) := C[~, ~−1 ] ⊗C[~] U~ (g).
We define Mloc (µ)− in the obvious way. There exists an isomorphism of graded C[~, ~−1 ]-algebras
∼ 1
U (g) ⊗C C[~, ~−1 ] −
→ Uloc (g)op , g ∋ x 7→ − x.
~
(In the left-hand side, U (g) is in degree 0.) Using this isomorphism one can regard Mloc (µ) as a
(left) module over the algebra U (g) ⊗C C[~, ~−1 ]. With this structure it is isomorphic to V(−µ −
ρ) ⊗C C[~, ~−1 ].
Let µ ∈ t∗ and let α be a simple root. If n := −µ(α̌) ∈ Z≥0 , then as in §4.1 there exists a unique
embedding of graded Uloc (g)-modules
Mloc (sα µ) ֒→ Mloc (µ)
)n
which sends vsα µ to vµ · (−f α
~n n! . Iterating, we perform the following construction. Let µ ∈ X , and
−
let w ∈ W . Choose a reduced decomposition w = sk · · · s1 , where si is the reflection associated

with the simple root αi . Then for any i = 1, · · · , k, ni := −(si−1 · · · s1 µ, α̌i ) ∈ Z≥0 , hence we can
define
(−fαi1 )n1 (−fαik )nk
vwµ := v µ · · · · ∈ Mloc (λ),
~n1 · n1 ! ~nk · nk !
and there exists a unique embedding of graded right Uloc (g)-modules
iw
µ : Mloc (wµ) ֒→ Mloc (µ) such that iw w
µ (vwµ ) = vµ .
24
4.3. Specialization. Let λ ∈ X. Recall that in §2.1 we have defined a structure of B-module and of
(U~ (b), U~ (g))-bimodule on M(λ). In particular, if V is in Rep(G), we have a (diagonal) B-module
structure on V ⊗ M(λ).
Lemma 4.3.1. For V in Rep(G) and λ ∈ X, the assignment φ 7→ φ(v0 ), induces an isomorphism
∼ B
Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) −
→ V ⊗ M(λ) .
Proof. By definition, M(0) is an induced right U~ (g)-module. Hence we have
Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) ∼
= Hom(S~ ,U~ (b)) (S~ hh0ii, V ⊗ M(λ)).
It follows that the left-hand side is isomorphic to the space of vectors x in V ⊗ M(λ) which satisfy
x · b = (b + ~ρ(b)) · x
for all b ∈ b ⊂ U~ (b). This is exactly the condition for being U (b)-invariant (i.e. B-invariant), see
§2.1.
If µ ∈ t∗ , we denote by P 7→ P (µ) the unique algebra morphism S~ → C[~] which sends x ∈ t to

~µ(x). Note that if P ∈ S~ and if for some w ∈ W we have P (µ) = 0 for all µ ∈ w(X− ) sufficiently
large, then P = 0.
If λ, µ ∈ t∗ , we denote by
Spµ : M(λ) → Mloc (λ + µ)
the natural morphism induced by isomorphism (4.2.1). The following lemma follows from the
fact that M(λ) is a free S~ -module and the remarks above on the map P 7→ P (µ).
Lemma 4.3.2. Let λ ∈ X and w ∈ W . Let m ∈ M(λ), and assume that Spµ (m) = 0 for all µ ∈ w(X− )
sufficiently large. Then m = 0.
We will derive several useful corollaries.
Corollary 4.3.3. Let V in Rep(G), λ ∈ X and w ∈ W . Let m ∈ V ⊗ M(λ), and assume that for all
µ ∈ w(X− ) sufficiently large we have
(idV ⊗ Spµ )(m) · b = ~(µ + ρ)(b) · (idV ⊗ Spµ )(m)
for all b ∈ b. Then there exists a unique φ ∈ Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) such that m = φ(v0 ).
Proof. By Lemma 4.3.1 we only have to prove that m is fixed by B, i.e. that for any b ∈ b we have
m · b − (b + ~ρ(b)) · m = 0
(see the proof of Lemma 4.3.1). However, our assumption implies that this vector is annihilated
by (idV ⊗ Spµ ) for all µ ∈ w(X− ) sufficiently large. Hence we conclude using Lemma 4.3.2.
If φ ∈ Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) we denote by Spµ (φ) : Mloc (µ) → V ⊗ Mloc (λ + µ) the
morphism obtained by tensoring with C[~]hhµii, using isomorphism (4.2.1), and inverting ~. This
construction induces a morphism of C[~, ~−1 ]-modules
Spµ : Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) → Hom−Uloc (g) (Mloc (µ), V ⊗ Mloc (λ + µ)).
Corollary 4.3.4. Let V in Rep(G), λ ∈ X, w ∈ W , and φ ∈ Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)). If
Spµ (φ) = 0 for all µ ∈ w(X− ) sufficiently large, then φ = 0.
25
Proof. This follows from the commutativity of the following diagram
φ7→φ(v0 )
Hom(S~ ,U~ (g)) (M(0), V ⊗ M(λ)) / V ⊗ M(λ)
Spµ idV ⊗Spµ

φ7→φ(vµ )

Hom−Uloc (g) (Mloc (µ), V ⊗ Mloc (λ + µ)) / V ⊗ Mloc (λ + µ)
and Lemma 4.3.2.
4.4. Intertwining operators. Let V in Rep(G), and let φ ∈ Hom−Uloc (g) (Mloc (µ), V ⊗ Mloc (λ + µ))
be an intertwining operator. Then we can write
φ(vµ ) = u ⊗ f (~) · vλ+µ + x
for unique u ∈ Vλ , f (~) ∈ C[~, ~−1 ] and x ∈ V ⊗ M(λ + µ)− . The vector u ⊗ f (~) ∈ Vλ ⊗ C[~, ~−1 ]
is called the expectation value of φ. This way we have defined a morphism of graded C[~, ~−1 ]-
modules −1
EV,λ
µ : Hom−Uloc (g) Mloc (µ), V ⊗ Mloc (λ + µ) → Vλ ⊗ C[~, ~ ].
Lemma 4.4.1. Let V in Rep(G), and λ ∈ X.

(1) For µ ∈ t∗ sufficiently large, EV,λ
µ is an isomorphism.
(2) For w ∈ W and µ ∈ X− sufficiently large, the morphism

Hom−Uloc (g) Mloc (wµ), V ⊗ Mloc (w(λ + µ)) → Hom−Uloc (g) Mloc (wµ), V ⊗ Mloc (λ + µ)
defined by φ 7→ (idV ⊗ iw
λ+µ ) ◦ φ is an isomorphism
Proof. By the remarks above we have an isomorphism Uloc (g)op ∼ = U (g) ⊗ C[~, ~−1 ], which induces
an isomorphism
Hom−U (g) (Mloc (µ), V ′ ⊗ Mloc (ν)) ∼
loc
= HomU (g) (V(−µ − ρ), V ′ ⊗ V(−ν − ρ)) ⊗ C[~, ~−1 ]
for any V in Rep(G) and µ, ν ∈ t∗ . Moreover, under these isomorphisms the morphisms consid-
′
ered in the lemma are induced by those of Lemma 4.1.1. Hence the claims follow from Lemma
4.1.1.
Fix V in Rep(G), λ ∈ X and w ∈ W . For any µ ∈ X− sufficiently large, we define the morphism
of graded C[~, ~−1 ]-modules

ΨV,λ
w,µ : Hom−Uloc (g) Mloc (µ), V ⊗ Mloc (µ + λ) → Hom−Uloc (g) Mloc (wµ), V ⊗ Mloc (w(µ + λ)) ,

in such a way that for any φ ∈ Hom−Uloc (g) (Mloc (µ), V ⊗ Mloc (µ + λ) we have
φ ◦ iw w V,λ
µ = (iλ+µ ⊗ idV ) ◦ Ψw,µ (φ).
This morphism is well defined by Lemma 4.4.1.
Let now α be a simple root, and consider the case w = sα . Then, even if µ is not in X− , one can
define a morphism

ΨV,λ
sα ,µ : Hom−Uloc (g) Mloc (µ), V ⊗ Mloc (µ + λ) → Hom−Uloc (g) Mloc (sα µ), V ⊗ Mloc (sα (µ + λ))
with the same properties as above as soon as µ ∈ X is sufficiently large and µ(α̌) < 0. With
this extension of the definition, consider again some w ∈ W , and let w = sk · · · s1 be a reduced
decomposition. Note that if µ ∈ X− , for any i = 1, · · · , k we have (si−1 · · · s1 µ)(α̌i ) ≤ 0. Then, by
definition, for µ ∈ X− sufficiently large we have
V,s ···s (λ) V,s (λ)
ΨV,λ k−1 1 1 V,λ
w,µ = Ψsk ,sk−1 ···s1 (µ) ◦ · · · ◦ Ψs2 ,s1 (µ) ◦ Ψs1 ,µ . (4.4.2)
26
4.5. Simple reflections. In this subsection we fix a simple root α, and set s = sα .
Proposition 4.5.1. Let V in Rep(G) and λ ∈ X. There exists a unique morphism of graded S~ -modules

ΘV,λ
s : Hom(S~ ,U~ (g)) M(0), V ⊗ M(λ) hλ(2ρ̌)i → s Hom(S~ ,U~ (g)) M(0), V ⊗ M(sλ) h(sλ)(2ρ̌)i
such that for any µ ∈ X sufficiently large such that µ(α̌) < 0 we have
Spsµ ◦ ΘV,λ
s =
(
(−~)λ(α̌) (−µ(α̌))(−µ(α̌) − 1) · · · (−µ(α̌) − λ(α̌) + 1) · ΨV,λ
s,µ ◦ Spµ if λ(α̌) ≥ 0,
1 V,λ (4.5.2)
(−~)−λ(α̌) (−µ(α̌)−λ(α̌))···(−µ(α̌)+1)
· Ψs,µ ◦ Spµ if λ(α̌) ≤ 0.
Proof. Unicity follows from Corollary 4.3.4. Let us prove existence. Choose an enumeration of
positive roots α1 , · · · , αn such that αn = α and, for any i = 1, · · · , n − 1, a non-zero vector fi ∈
g−αi . (These choices can be arbitrary.) For any multi-index k = (k1 , · · · , kn−1 ) we set f k :=
kn−1
f1k1 · · · fn−1 ∈ Uloc (g). Then any vector of M(λ) can be written in a unique way in the form
X
Pk,i ⊗ f k fαi
k,i
where Pk,i ∈ S~ (and where only finitely many terms are non-zero). Choose also a basis {uj , j ∈ J}
of V . To simplify notation, for µ ∈ t∗ we set n(µ) := −µ(α̌).
For P in S~ we set
P P (P − ~) · · · (P − (m − 1)~)
:= .
m ~ m(m − 1) · · · 1
With this notation, for µ ∈ X such that n(µ) ≥ m we have

α̌ m n(µ)
(sµ) = ~ . (4.5.3)
m ~ m
Let φ : M(0) → V ⊗ M(λ) be a morphism of bimodules. Write
X
φ(v0 ) = uj ⊗ Pj,k,i ⊗ f k fαi
j,k,i
where Pj,k,i ∈ S~ . Then for µ ∈ X such that n(µ) ≥ 0 we have

X
Spµ (φ)(v µ ) = uj ⊗ Pj,k,i (µ) ⊗ f k fαi ∈ V ⊗ Mloc (λ + µ),
j,k,i
hence
X (−f )n(µ)
α
Spµ (φ)(v sµ ) = uj ⊗ Pj,k,i (µ) ⊗ f k fαi · n(µ)
j,k,i
~ n(µ)!
 
n(µ) X
(−1)n(µ)  X n(µ)
= n(µ) (−~)m fαm · uj ⊗ Pj,k,i (µ) ⊗ f k fαi+n(µ)−m  .
~ n(µ)! m
m=0 j,k,i
By Lemma 4.4.1(2) (or more precisely an obvious generalization, when w = s, to the case µ(α̌) ≤ 0
instead of µ ∈ X− ), if µ if sufficiently large then the terms for which i + n(µ) − m < n(λ + µ)
vanish. Hence we obtain that in this case Spµ (φ)(v sµ ) equals
 

(−1) n(µ)
 X n(µ) m 
n(µ)  (−~)m fα · uj ⊗ Pj,k,i (µ) ⊗ f k fαi+n(µ)−m  ,
~ n(µ)! j,k,i,m
m
0≤m≤i−n(λ)
27
i.e. equals
X
(−~)n(λ) n(λ + µ)! n(µ) m
(idV ⊗ isλ+µ ) (−~) m
fα · uj ⊗ Pj,k,i (µ) ⊗ f k fαi−n(λ)−m .
n(µ)! j,k,i,m
m
0≤m≤i−n(λ)
Note that in this sum the indices do not depend on µ. Hence we can consider the element
X
α̌
x := (−1)m fαm · uj ⊗ s(Pj,k,i ) ⊗ f k fαi−n(λ)−m ∈ V ⊗ M(sλ). (4.5.4)
j,k,i,m
m ~
0≤m≤i−n(λ)
By construction (and using (4.5.3)), for any µ ∈ X sufficiently large with n(µ) ≥ 0, (idV ⊗Spsµ )(x)
is a multiple of Spµ (φ)(vsµ ); hence it satisfies
(idV ⊗ Spsµ )(x) · b = ~(sµ + ρ)(b) · (idV ⊗ Spsµ )(x)
for all b ∈ b. Hence by Corollary 4.3.3 there exists a unique ψ ∈ Hom(S~ ,U~ (g)) (M(0), V ⊗ M(sλ))
such that x = ψ(v0 ). We set ΘV,λ
s (φ) := ψ. With this definition, for any φ and any sufficiently large
µ with n(µ) ≥ 0 we have
(−~)n(λ) n(λ + µ)!
Spµ (φ) ◦ isµ = (idV ⊗ isλ+µ ) ◦ Spsµ (ΘV,λ
s (φ)),
n(µ)!
which implies (4.5.2).
4.6. Restriction to a Levi subgroup. Fix a subset I of the set of simple roots, and recall the notation
of §3.3. As in §3.3 we can consider our constructions both for G and for L; we add super- or
subscripts to indicate which reductive group we consider.
−
The projection g ։ l, along n+ L ⊕ nL , is L-equivariant and it induces a morphism of graded
− ∼
C[~]-modules πLG : U~ (g) ։ U~ (g)/(n+L · U~ (g) + U~ (g) · nL ) = U~ (l). We consider the morphism of
graded S~ -modules
MG (λ) → ML (λ) (4.6.1)
G G G

sending p ⊗ u ∈ M (λ) to p ⊗ ıL ◦ πL (u) . One can easily check that (4.6.1) is also a morphism of
BL -modules, so that it induces a morphism of graded C[~]-modules
B B
RV,λ
G,L : V ⊗ M (λ)
G
→ V|L ⊗ ML (λ) L (4.6.2)
for any V in Rep(G) and λ ∈ X.
Note that if K ⊂ L is a smaller Levi subgroup containing T , then for any V in Rep(G) and λ ∈ X
we have
V|L ,λ
RL,K ◦ RV,λ V,λ
G,L = RG,K . (4.6.3)
If L = T , then we have MT (λ) = S~ hhλii, where BT = T acts via −λ. Hence we have an isomor-
phism of graded S~ -modules
B
V|T ⊗ MT (λ) T = Vλ ⊗ S~ .
Under this isomorphism, one can easily check that RV,λ alg
G,T = κV,λ .
Using Lemma 4.3.1, RV,λG,L induces a morphism of graded S~ -modules

RV,λ G G L L
G,L : Hom(S~ ,U~ (g)) M (0), V ⊗ M (λ) → Hom(S~ ,U~ (l)) M (0), V|L ⊗ M (λ) .
Lemma 4.6.4. For all V in Rep(G) and λ ∈ X, the morphism RV,λ

G,L is injective.
28
Proof. By (4.6.3), it is enough to prove the claim when L = T . Then by construction, if we identify
T T
Hom−Uloc (t) Mloc (0), V|T ⊗ Mloc (λ) with Vλ ⊗ C[~, ~−1 ] in the natural way, for any µ ∈ t∗ we have
SpTµ ◦ RV,λ V,λ
G,T = Eµ ◦ Spµ .
G

Let φ ∈ Hom(S~ ,U~ (g)) MG (0), V ⊗ MG (λ) be such that RV,λ G,T (φ) = 0. Then for any µ ∈ t
∗
we have EV,λ G
µ (Spµ (φ)) = 0. Using Lemma 4.4.1(1) we deduce that for µ sufficient large we have
SpG
µ (φ) = 0. By Corollary 4.3.4 we deduce that φ = 0, which finishes the proof.
The following result follows from construction (see the proof of Proposition 4.5.1), using the
fact that if α is a simple root of L then [fα , n− −
L ] ⊂ nL . For simplicity, in this statement we neglect
the gradings.
Lemma 4.6.5. Let V in Rep(G) and λ ∈ X. If α ∈ I the following diagram commutes:
G ΘV,λ
sα
Hom(S~ ,U~ (g)) MG (0), V ⊗ MG (λ) / sα Hom
(S~ ,U~ (g)) MG (0), V ⊗ MG (sα λ)
RV,λ
G,L
sαRV,sα λ
G,L
V ,λ
L Θ |L
sα
Hom(S~ ,U~ (l)) ML (0), V|L ⊗ ML (λ) / sα Hom
(S~ ,U~ (l)) ML (0), V|L ⊗ ML (sα λ) .
4.7. Definition of operators Θw .

Lemma 4.7.1. Let V in Rep(G), and λ ∈ X. For any simple root α, we have
sα

ΘV,s
sα
αλ
◦ ΘV,λ
sα = id

as endomorphisms of Hom(S~ ,U~ (g)) M(0), V ⊗ M(λ) hλ(2ρ̌)i. In particular, ΘV,λ
s is an isomorphism.
Proof. Let Lα be the Levi subgroup containing T with roots {α, −α}. Using Lemma 4.6.5 and
Lemma 4.6.4 it is enough to prove the equality when G = Lα . In this case it is checked in Corollary
A.3.2 below.
Let again V in Rep(G) and λ ∈ X. Let w ∈ W , and choose a reduced expression w = sk · · · s1 .
We define the isomorphism of graded S~ -modules

ΘV,λ
w : Hom(S~ ,U~ (g)) M(0), V ⊗ M(λ) hλ(2ρ̌)i
∼
−→ w Hom(S~ ,U~ (g)) M(0), V ⊗ M(wλ) h(wλ)(2ρ̌)i
by the formula
(sk−1 ···s1 ) V,sk−1 ···s1 (λ)
ΘV,λ
w := Θsk ◦ · · · ◦ s1ΘV,s
s2
1λ
◦ ΘV,λ
s1 .
Lemma 4.7.2. The operator ΘV,λ

w does not depend on the choice of the reduced decomposition.
Proof. Using (4.5.2) and (4.4.2) we obtain that for µ ∈ X− sufficiently large we have
Spwµ ◦ ΘV,λ V,λ
w = n(w, λ, µ) · Ψw,µ ◦ Spµ
where Q λ(α̌) (−µ(α̌)) · · · (−µ(α̌) − λ(α̌) + 1)
α̌>0,w(α̌)<0,λ(α̌)>0 (−~)
n(w, λ, µ) = Q −λ(α̌) (−µ(α̌) − λ(α̌)) · · · (−µ(α̌) + 1)
.
α̌>0,w(α̌)<0,λ(α̌)<0 (−~)
The right-hand side is independent of the reduced decomposition, hence we conclude by Corol-
lary 4.3.4.
By construction and Lemma 4.7.1, our collection of isomorphisms satisfies condition (2.5.4) for
all V in Rep(G), λ ∈ X and x, y ∈ W .
29
4.8. Relation with the operators Φ. As explained in §3.1, for any V in Rep(G) and λ ∈ X there
exists a canonical isomorphism of graded S~ -modules
G B
V ⊗ (λ)D~ (X)λ ∼
= V ⊗ M(λ) .
Using Lemma 4.3.1 we deduce a canonical isomorphism
G
V ⊗ (λ)D~ (X)λ ∼
= Hom(S~ ,U~ (g)) M(0), V ⊗ M(λ) . (4.8.1)
The following result is clear from definitions.
Lemma 4.8.2. Let V in Rep(G), λ ∈ X, and L ⊂ G be a Levi subgroup containing T . The following
diagram commutes:
(4.8.1) G
Hom(S~ ,U~ (g)) MG (0), V ⊗ MG (λ) ∼
/ V ⊗ (λ)D~ (XG )λ
RV,λ
G,L
RV,λ
G,L

(4.8.1)
Hom(S~ ,U~ (l)) ML (0), V ⊗ ML (λ) / V ⊗ (λ)D~ (XL )λ L .
|L ∼ |L
In the following proposition we assume that the vectors fα of Section 3 are the same as the
vectors fα of the present Section 4.
Proposition 4.8.3. For any V in Rep(G), λ ∈ X and w ∈ W the following diagram commutes:
G (4.8.1)
V ⊗ (λ)D~ (X)λ ∼
/ Hom
(S~ ,U~ (g)) M(0), V ⊗ M(λ)
ΦV,λ
w ΘV,λ
w

w
G (4.8.1)
V ⊗ (wλ)D~ (X)wλ ∼
/ w Hom
(S~ ,U~ (g)) M(0), V ⊗ M(wλ) .
Proof. By (2.5.3) and (2.5.4) it is enough to prove commutativity when w is a simple reflection.
By Corollary 3.3.3, Lemma 4.6.5 and Lemma 4.8.2, our constructions are compatible with re-
striction to a Levi subgroup. Using Lemma 4.6.4, we deduce that it is enough to prove the claim
when G has semisimple rank one. In this case it is proved in Lemma A.4.1 below.
It follows from this proposition that for any V in Rep(G), λ ∈ X and w ∈ W there exists a unique
isomorphism
B ∼ w B
ΩV,λ
w : V ⊗ M(λ) −
→ V ⊗ M(wλ)
which corresponds to ΦV,λ V,λ
w under the left-hand isomorphism of Lemma 2.4.1, and to Θw under
the right-hand isomorphism of Lemma 2.4.1.
5. G EOMETRY OF THE G ROTHENDIECK –S PRINGER RESOLUTION
5.1. The W -action on the regular part of e

g. In this section we are interested in the geometry of
g (see §2.4). We have a standard commutative diagram
the Grothendieck–Springer resolution e
π / g∗
e
g
δ

t∗ / t∗//W ∼
= g∗//G
where the right vertical and the lower horizontal maps are the natural quotient maps, π is defined
in §2.5, and δ is defined by δ(g ×B η) = η|b ∈ (b/u)∗ ∼
= t∗ .
30
It is well known that there exists an action of W on e gr , which commutes with the natural G-
action, and such that the restrictions of π and δ are W -equivariant, where W acts naturally on t∗
∼
and trivially on g∗ (see e.g. [BR, Proposition 1.9.2]). We denote by θw : e gr −→e gr the action of w.
In the whole section we will use [BR] as a convenient reference for the properties of this action,
though much of this material was known before.
Recall that for any λ ∈ X we have a line bundle Oeg (λ) on eg and its restriction Oegr (λ) to e
gr .
Lemma 5.1.1. For any w ∈ W and λ ∈ X, there exists an isomorphism of G×Gm -equivariant line bundles
(θw−1 )∗ Oegr (λ)hλ(2ρ̌)i ∼
= Oegr (wλ)h(wλ)(2ρ̌)i.
Proof. It is enough to prove the isomorphism when w = sα is a simple reflection. Next, by com-
∗
patibility of θw −1 with tensor products (and passing to a simply connected cover if necessary), it
is enough to prove the isomorphism when λ(α̌) ∈ {−1, 0, 1}. Finally using duality one can as-
sume that λ(α̌) ∈ {−1, 0}. In these cases the isomorphism follows from [BR, Lemma 1.11.3(2) and
Remark 1.11.7(2)].
5.2. Construction of the operators σ. The isomorphism of Lemma 5.1.1 is unique up to a scalar
since we have
Γ(egr , Oegr )G×Gm ∼ g, Oeg )G×Gm = C
= Γ(e
(see e.g. equation (5.2.2) below). We will need to fix a normalization of this isomorphism, using
our choice of vectors fα . Let us denote by η0 ∈ g∗ the element which is zero on t and on any gβ
where β is not opposite to a simple root, and such that η0 (fα ) = 1 for any simple root α. Then
(1 ×B η0 ) ∈ e
gr , and this point is W -invariant. For any λ ∈ X, Oeg (λ) is the sheaf of sections of the
line bundle

L(λ) := G ×B (g/u)∗ × C−λ
over eg. The fiber of this line bundle over (B/B, η0 ) can be canonically identified with C through
the morphism x 7→ (1 ×B (η0 , x)). Then there exists a unique isomorphism of G × Gm -equivariant
line bundles
∼
(θw−1 )∗ Oegr (λ)hλ(2ρ̌)i −
→ Oegr (wλ)h(wλ)(2ρ̌)i (5.2.1)
whose restriction to (B/B, η0 ) ∈ e
gr is idC via the identifications above.
Let jr : e
gr ֒→ e
g be the inclusion. As the codimension of e gre gr in e
g is at least 2 (see e.g. [BR,
Proposition 1.9.3]), for any λ ∈ X the morphism Oeg (λ) → (jr )∗ Oegr (λ) induced by adjunction is an
isomorphism (see [Ma, Theorem 11.5.(ii)]). We deduce that the restriction induces an isomorphism
∼
g, Oeg (λ)) −
Γ(e → Γ(e
gr , Oegr (λ)). (5.2.2)
As θw−1 is an isomorphism, the adjunction morphism Oegr (λ) → (θw−1 )∗ (θw−1 )∗ Oegr (λ) is also an
isomorphism; in particular there is a canonical isomorphism
∼ ∗
gr , Oegr (λ)) −
Γ(e → Γ(e
gr , θ w −1 Oe
gr (λ)).
Putting these remarks together with isomorphism (5.2.1), we obtain for any w ∈ W and λ ∈ X a
canonical isomorphism of graded S(t)-modules and G-modules
∼ w
g, Oeg (λ))hλ(2ρ̌)i −
Γ(e → g, Oeg (wλ))h(wλ)(2ρ̌)i.
Γ(e (5.2.3)
Hence for any V in Rep(G) we obtain an isomorphism of graded S(t)-modules
V,λ
G ∼ G
σw : V ⊗ Γ(e → w V ⊗ Γ(e
g, Oeg (λ)) hλ(2ρ̌)i − g, Oeg (wλ)) h(wλ)(2ρ̌)i.
By construction, this collection of isomorphisms satifies relations (2.5.6).
31
As explained above, for any µ ∈ X we can describe Γ(e g, Oeg (µ)) as the space of sections of the
g. In particular, for any η ∈ (g/u)∗ there exists a unique morphism
line bundle L(µ) over e
evµη : Γ(e
g, Oeg (µ)) → C
(“evaluation at (1, η)”) such that if f ∈ Γ(e g → L(µ) we have
g, Oeg (µ)) is considered as a section e
f (1 ×B η) = 1 ×B (η, evµη (f )).
With this definition, (5.2.1) can be characterized as the unique isomorphism of G × C× -equiva-
∼
riant line bundles (θw−1 )∗ Oegr (λ)hλ(2ρ̌)i −
→ Oegr (wλ)h(wλ)(2ρ̌)i such that the following diagram
commutes:
(5.2.3)
g, Oeg (λ))
Γ(e ∼
/ Γ(eg, Oeg (wλ))
❖❖❖ ♥♥
❖❖❖
❖❖❖ ♥♥♥
♥ ♥ (5.2.4)
evλ
η0
❖❖❖ ♥♥♥♥evwλ
' w♥♥ η0
C.
Indeed the diagram commutes by construction. To prove unicity if suffices to prove that the mor-
phism evλη0 is non-zero. However, if λ is dominant this property follows from the fact that Oeg (λ)
is globally generated (which itself follows from the similar claim for B), and the general case fol-
lows from commutativity of (5.2.4) (and the fact that every weight it W -conjugate to a dominant
weight).
Below we will need a refinement of this characterization in the case w = sα for a simple root α.
We denote by ηα ∈ g∗ the element which is zero on t and on any gβ with β 6= −α, and such that
ηα (fα ) = 1.
Lemma 5.2.5. When w = sα , (5.2.1) is the unique isomorphism of G × C× -equivariant line bundles
(θsα )∗ Oegr (λ)hλ(2ρ̌)i → Oegr (sα λ)h(sα λ)(2ρ̌)i such that the following diagram commutes:
(5.2.3)
g, Oeg (λ))
Γ(e ∼ g, Oeg (sα λ))
/ Γ(e
❖❖❖ ♥♥
❖❖❖ ♥♥♥
❖❖❖
♥♥♥♥s λ
evλ
ηα
❖❖❖
'
♥
w♥♥♥ evηα
α
C
Proof. As for the similar claim concerning diagram (5.2.4), we only have to check that the diagram
∼
commutes. In this proof we denote the isomorphism (5.2.3) by ϑα : Γ(e g, Oeg (λ)) −
→ Γ(e
g, Oeg (sα λ)).
Choose a coweight µ̌ ∈ X∗ (T ) such that α(µ̌) = 0 and β(µ̌) > 0 for all simple roots β 6= α. Then
limz→0 µ̌(z) · η0 = ηα , so that it is enough to prove that the following diagram commutes for all
z ∈ C× :
ϑα
g, Oeg (λ))
Γ(e ∼ g, Oeg (sα λ))
/ Γ(e
❖❖❖ ♥♥
❖❖❖ ♥ ♥
❖❖❖ ♥♥
evλ ❖❖❖ ♥♥♥♥sα λ
µ̌(z)·η0 ' w ♥
♥ ♥ ev µ̌(z)·η0
C
However for ν ∈ X and f ∈ Γ(e g, Oeg (ν)) we have

f 1 ×B (µ̌(z) · η0 ) = f (µ̌(z) ×B η0 ) = µ̌(z) · (µ̌(z −1 ) · f )(1 ×B η0 )

= µ̌(z) · 1 ×B η0 , evλη0 (µ̌(z −1 ) · f ) = µ̌(z) ×B η0 , evλη0 (µ̌(z −1 ) · f )

= 1 ×B µ̌(z) · η0 , z λ(µ̌) · evλη0 (µ̌(z −1 ) · f ) .
32
On the other hand we have

f 1 ×B (µ̌(z) · η0 ) = 1 ×B µ̌(z) · η0 , ev λµ̌(z)·η0 (f ) ,
which implies that
evλµ̌(z)·η0 (f ) = z λ(µ̌) · evλη0 (µ̌(z −1 ) · f ).
Hence we obtain
evsµ̌(z)·η
αλ
0
(ϑα f ) = z (sα λ)(µ̌) · evsηα0 λ (µ̌(z −1 ) · (ϑα f )) = z λ(µ̌) · evλη0 (µ̌(z −1 ) · f ) = evλµ̌(z)·η0 (f )
since (sα λ)(µ̌) = λ(µ̌) and ϑα is G-equivariant. This finishes the proof.
5.3. Restriction to a Levi subgroup. Fix a subset I of the set of the set of simple roots, and recall
the notation of §3.3. We can consider the Grothendieck–Springer resolution el associated with L,
and there exists a natural morphism ̟LG : G×Lel = G×BL (l/uL )∗ → e g induced by the identification
∗ ∼ + − ∗ ∗
l = (g/(nL ⊕ nL )) . In particular for L = T we have t = t , and the morphism ̟TG : G ×T et =
e
G/T × t∗ → e g identifies with the morphism denoted “a” in §2.4. The following diagram commutes
by construction:
G
̟T
/ G ×L el */ e (5.3.1)
G ×T et L G
g.
G×L ̟T ̟L
We have (̟LG )−1 (e gr ) ⊂ G ×L elr . (Note that this inclusion is strict in general.) This open subset
is WL -invariant (for the WL -action on G ×L elr induced by the action on elr ), and the morphism
(̟LG )−1 (e gr induced by ̟LG is WL -equivariant.
gr ) → e
Adjunction for the morphism ̟LG induces an injective morphism

Γ(eg, Oeg (λ)) ֒→ Γ(G ×L el, OG×Lel (λ)) ∼
= IndG e
L Γ(l, Oel (λ)) . (5.3.2)
For simplicity, in the next statement we forget about the gradings (i.e. the C× -actions).
Lemma 5.3.3. The following diagram commutes for any V in Rep(G), λ ∈ X and w ∈ WL , where vertical
maps are induced by (5.3.2):
G V,λ
σw w
G
V ⊗ Γ(e
g, O
e
g (λ)) / V ⊗ Γ(e
g, O
e
g (wλ))
_ _
V|L ,λ
L σw L
V|L ⊗ Γ(el, Oel (λ)) / w V|L ⊗ Γ(el, Oel (wλ)) .
Proof. It is enough to prove the result when w = sα is a simple reflection associated with a simple
root α ∈ I. Let sl2,α ⊂ g be the Lie subalgebra generated by gα and g−α , and consider the open
subset
gα−r := {(g ×B η) ∈ e
e g | η|sl2,α 6= 0} ⊂ e
g.
By [BR, Lemma 1.9.1] we have e gr ⊂ e gα−r , and it follows from [BR, Lemma 2.9.1] that the action
∼
map θsα is the restriction of an isomorphism (denoted similarly) θsα : e gα−r −
→egα−r . Moreover,
isomorphism (5.2.1) is (a shift of) the restriction of an isomorphism of G × C× -equivariant line
bundles
∼
ςsG,λ
α
: (θsα )∗ Oegα−r (λ) −
→ Oegα−r (sα λ)h−2λ(α̌)i.
The same assertions are of course true for the Levi L, and we obtain a similar isomorphism ςsL,λ
α .
33
The morphism ̟LG restricts to a morphism G ×L elα−r → e gα−r (denoted similarly) which satisfies
G L G G
̟L ◦ (G ×L θsα ) = θsα ◦ ̟L . What we have to prove is that the two isomorphisms
∼
(̟LG )∗ ςsG,λ
α
, IndG L,λ L ∗
L (ςsα ) : (G ×L θsα ) OG×Lelα−r (λ) −
→ OG×Lelα−r (sα λ)
coincide. As both isomorphisms are G × C× -equivariant, and as

× ×
Γ(G ×L elα−r , OG×Lelα−r )G×C ∼= Γ(G ×L el, OG×Lel )G×C ∼= C,
we know that these isomorphisms coincide up to multiplication by a scalar c ∈ C× . Consider the
following diagram:
∼
g, O eg (λ))
Γ(e g, Oeg (s
/ Γ(e
_ α λ))
_
(5.3.2) (5.3.2)

∼
Γ(G ×L el, OG×Lel (λ)) / Γ(G ×L el, O (s λ))
G× el α
(5.3.4)
evλ,G
ηα L evsηα
α
λ,G
′ ev λ,L ′ ev sα λ,L
ηα ηα

*C C, t
where the upper horizontal morphism is induced by ςsG,λ α , the middle horizontal morphism is
G L,λ
induced by IndL (ςsα ), and the lower vertical maps are induced by evaluation at (1 ×BL ηα ). We
know that the upper square commutes up to multiplication by c, that the lower square and the
exterior square both commute and finally that the morphism evλ,G ηα is non-zero (see Lemma 5.2.5
and its proof). We deduce that c = 1, which finishes the proof.
5.4. Relation to the operators Φ. Recall that if λ ∈ X there is a natural morphism (λ)D (X) →
~ λ
Γ(eg, Oeg (λ)) sending ~ to 0 (see e.g. §3.5).
Proposition 5.4.1. Let V in Rep(G), λ ∈ X and w ∈ W . The following diagram commutes, where vertical
maps are the natural morphisms sending ~ to 0:
G ΦV,λ
w G
V ⊗ (λ)D~ (X)λ / w V ⊗ (wλ)D~ (X)wλ

G V,λ
σw w
G
V ⊗ Γ(e
g, Oeg (λ)) / V ⊗ Γ(e
g, Oeg (wλ))
Proof. First, using relations (2.5.3) and (2.5.6) it is enough to prove the lemma when w = sα is a
simple reflection. Then using the compatibility of our constructions with restriction to a Levi sub-
group (see Corollary 3.3.3 and Lemma 5.3.3) and the injectivity of morphism (5.3.2), it is enough
to prove the lemma when G has semisimple rank one (with unique simple root α). In this case it
is proved in Corollary A.5.2 below.
Remark 5.4.2. By Lemma 3.5.2, the vertical arrows in the diagram of Proposition 5.4.1 induce iso-
morphisms
G ∼ G G ∼ G
V ⊗ (λ)D~ (X)λ /h~i −
→ V ⊗ Γ(e g, Oeg (λ)) , V ⊗ (sλ)D~ (X)sλ /h~i −→ V ⊗ Γ(e g, Oeg (sλ)) .
V,λ
Hence the proposition implies that the operators σw can be completely recovered from the oper-
V,λ V,λ
ators Φw (or equivalently the operators Θw , see Proposition 4.8.3).
34
5.5. Geometric interpretation: W -action on the regular part of T ∗ X. The results in this subsection
will not be used in the rest of the paper.
Consider the natural morphism T ∗ X = G ×U (g/u)∗ → g∗ , and denote by (T ∗ X)r the inverse
image of the open subset of regular elements in g∗ . Note that the G × T -action on X defined in §2.4
induces an action on T ∗ X which stabilizes (T ∗ X)r , and also a moment map T ∗ X → t∗ .
The existence of the collection of isomorphisms (5.2.1) has the following quite surprising con-
sequence. This construction will be re-interpreted and studied further in [GK].
Proposition 5.5.1. There exists an action of W on (T ∗ X)r (which depends on the choice of the eα ’s),
denoted ⊙, which satisfies the following properties:
(1) for any w ∈ W the morphism w ⊙ (−) is G-equivariant;
(2) for x ∈ (T ∗ X)r and t ∈ T we have w ⊙ (t · x) = w(t) · (w ⊙ x);
(3) the natural morphism (T ∗ X)r → e gr is W -equivariant;
(4) the restriction (T ∗ X)r → t∗ of the moment map is W -equivariant.
Proof. The morphism pr : (T ∗ X)r → e gr is a T -torsor; in particular it is affine. Hence to prove the
proposition it is enough to construct of collection of isomorphisms of G-equivariant sheaves of
algebras
(θw−1 )∗ (pr )∗ O(T ∗ X)r ∼= (pr )∗ O(T ∗ X)r
for all w ∈ W , which are compatible with composition in W . Now we have a natural isomorphism
of G-equivariant sheaves of algebras
M
(pr )∗ O(T ∗ X)r ∼
= Oegr (λ),
λ∈X
hence it is enough to construct a collection of isomorphisms of G-equivariant line bundles

(θw−1 )∗ Oeg (λ) ∼
r = Oeg (wλ)r
compatible with tensor product and composition in W . However one can easily check that the
collection constructed in (5.2.1) satisfies these requirements.
6. R EMINDER ON THE S ATAKE EQUIVALENCE

In Sections 6–8 we let Ǧ be a complex connected reductive group. We choose a maximal torus
Ť ⊂ Ǧ (with Lie algebra ť), and a Borel subgroup B̌ ⊂ Ǧ. We let Ǔ, resp. Ǔ − , be the unipotent
radical of B̌, resp. of the opposite Borel subgroup (with respect to Ť ). We set X := X∗ (Ť ), and let
X+ ⊂ X be the sub-semigroup of dominant coweights of Ť (where positive roots of Ǧ are those
appearing in Lie(Ǔ )). We also set S~ := S(ť∗ )[~], considered as a graded algebra where elements of
ť∗ and ~ are in degree 2. Finally we denote by 2ρ̌ ∈ X ∗ (Ť ) the sum of the positive roots.
6.1. Satake equivalence. Recall (see §2.2) that the affine Grassmannian attached to Ǧ is the ind-
variety
GrǦ := Ǧ(K)/Ǧ(O)
(equipped with the reduced scheme structure). This ind-variety is equipped with an action of the
group scheme Ǧ(O). Recall (see [G1, MV]) that the category
PervǦ(O) (GrǦ )
of Ǧ(O)-equivariant perverse sheaves on GrǦ (with coefficients in C) can be endowed with a nat-
ural convolution product ⋆ which makes it a tensor category, and that the functor
FǦ := H (GrǦ , −) : PervǦ(O) (GrǦ ) → Vect(C)
q
35
(where Vect(C) is the category of finite dimensional C-vector spaces) is a tensor functor. (As usual,
perverse sheaves on GrǦ are assumed to be supported on a finite union of Ǧ(O)-orbits.) We let
G := Aut⋆ (FǦ )
be the C-group scheme of automorphisms of this tensor functor. It is well known (see [G1, MV])
that G is a (complex) connected reductive group, with root datum dual to that of Ǧ. Moreover, the
functor FǦ lifts to an equivalence of tensor categories
∼
SǦ : Perv Ǧ(O) (GrǦ ) −
→ Rep(G)
known as the geometric Satake equivalence.
In §2.2 we have defined the embedding X = GrŤ ֒→ GrǦ , the points λ (for λ ∈ X), the orbits GrλǦ
(for λ ∈ X+ ), the semi-infinite orbits Tλ (for λ ∈ X), and the morphisms iλ and tλ .
Using the identification of GrŤ with X, the group T (of automorphisms of the tensor functor FŤ )
is identified with the torus HomZ (X, C× ). In particular, the character lattice X ∗ (T ) is canonically
identified with X, hence the category Rep(T ) identifies with the category of finite dimensional
X-graded vector spaces. Define the functor
M
FX := Hλ(2ρ̌) Tλ , t!λ (−) : Perv Ǧ(O) (GrǦ ) → Rep(T ).
λ∈X
By [MV, Theorems 3.5 and 3.6], we have a canonical isomorphism

∼
ForT ◦ FX −
→ FǦ , (6.1.1)
where ForT : Rep(T ) → Vect(C) is the forgetful functor. Moreover, FX is a tensor functor, and
(6.1.1) is an isomorphism of tensor functors. So FX is the composition of SǦ with a tensor functor
Rep(G) → Rep(T ) compatible with forgetful functors. By [DM, Corollary 2.9], the latter functor is
induced by a group morphism T → G. It is proved in [MV] that this morphism is injective, and
identifies T with a maximal torus of G. Hence from now on we will consider T as a subgroup
of G. By construction, if λ ∈ X and if F is in PervǦ(O) (GrǦ ) we have a canonical isomorphism of
C-vector spaces

Hλ(2ρ̌) Tλ , t! F ∼
λ = S (F) . Ǧ λ
(6.1.2)
The choice of B̌ ⊂ Ǧ determines a set of positive roots for Ǧ, hence also a set of positive roots
for G. We denote by B the Borel subgroup of G containing T associated with this set of positive
roots.
6.2. Equivariant cohomology. For the results stated below, see e.g. [Lu, §1].
For any complex algebraic variety X endowed with an algebraic action of an algebraic group
H, recall that the equivariant cohomology (respectively Borel–Moore homology) is defined by
q−2 dim(X)
HH (X) := ExtDH (X) (CX , CX ), HHq (X) := ExtDH (X) (CX , DX ),
q q
where DH (X) is the H-equivariant derived category of X, CX is the (equivariant) constant sheaf
on X and DX is the (equivariant) dualizing sheaf. Then HH (X) is a (graded-commutative) algebra
q
H
for the Yoneda product (or cup product) and H q (X) is a right module over this algebra, again for
the Yoneda product. If X is smooth, then this module is free of rank 1.
Let now K be a torus, with Lie algebra k. Let λ ∈ X ∗ (K), and consider the 1-dimensional
K-module Cλ , considered as a K-variety. Consider the K-equivariant morphisms
ι π
{pt} −
→ Cλ −
→ {pt},
36
(where ι(pt) = 0), and the induced morphisms in equivariant homology
ι! (π ∗ )−1
HKq (pt) / HK
q+2 (Cλ ) ∼
/ HKq+2 (pt),
which are morphisms of right HK (pt)-modules. Since the right HK (pt)-module HKq (pt) is free
q q
of rank 1, there exists a unique element c(λ) ∈ H2K (pt) such that the composition above is the
action of c(λ), and c : X ∗ (K) → HK (pt) is a morphism of abelian groups. There exists a unique
q
∼
isomorphism a : k∗ − → H2K (pt) such that the following diagram commutes:
X ∗ (K)
❖❖❖ c
d ssss ❖❖❖
sss ❖'
ys a
k∗ ∼
/ H2 (pt),
K
where d is the differential. Moreover, this isomorphism extends to an isomorphism of graded

C-algebras
∼
S(k∗ ) −
→ HK (pt), (6.2.1)
q
where in the left-hand side k∗ is in degree 2. We will use this isomorphism throughout the paper
without further details. In particular we can identify
H (pt) ∼
= S~ , H (pt) ∼= S(t),
q q
A Ť
where ~ ∈ S~ corresponds to the natural generator of H2C× (pt) = ∼ C.

If V is any K-module, with K-weights λ1 , · · · , λn (counted with multiplicities), then if as above
ι denotes inclusion of 0 and π projection to 0, via isomorphism (6.2.1), the composition
ι! (π ∗ )−1
HKq (pt) / HK
q+2n (V ) ∼
/ HKq+2n (pt)
identifies with the action of d(λ1 ) · · · d(λn ). Note that the action map induces an isomorphism
∼
of HK (pt)-modules HK 0 (V ) ⊗ HK (pt) − → HKq (V ), and that the forgetful map HK 0 (V ) → H0 (V )
q q
is an isomorphism. Hence we obtain a canonical isomorphism HKq (V ) ∼ = H0 (V ) ⊗ HK (pt). The
q
Borel–Moore homology H0 (V ) contains the canonical class [V ], which can therefore be viewed
as a generator of HKq (V ). Similarly we have the canonical class [pt] ∈ HKq (pt). Then we have
π ∗ [pt] = [V ], hence the morphism ι! has the property that
ι! ([pt]) = [V ] · d(λ1 ) · · · d(λn ).
We can now give a proof of Lemma 2.2.1.
Proof of Lemma 2.2.1. To fix notation we treat the case of A; the case of Ť is similar. Let F in
PervǦ(O) (GrǦ ) and λ ∈ X.
(1) Using the Leray–Serre spectral sequence for an appropriate fibration, there exists a spectral
sequence which computes HA (i!λ F) and with E2 -term
q
E2p,q = HpA (pt) ⊗ Hq (i!λ F) (6.2.2)

(see e.g. [Lu, Proof of Proposition 7.2]). It is well known also (see [KL, Theorem 5.5] or [Sp, Corol-
laire 2.10]) that the ordinary cohomology H (i!λ F) is concentrated in degrees of constant parity.
q
Hence the spectral sequence (6.2.2) degenerates at E2 . It follows that there exists a non-canonical
isomorphism of graded S~ -modules
HA (i!λ F) ∼
= HA (pt) ⊗ H (i!λ F).
q q q
In particular, the left-hand side is free over S~ = HA (pt).

q
37
(2) By [MV, Theorem 3.4], H (Tλ , t!λ F) is concentrated in degree λ(2ρ̌). Hence the same spectral
q
sequence argument as in (1) shows that HA (Tλ , t!λ F) is free over HA (pt). Moreover, the morphism
q q
HA (Tλ , t!λ F) → H (Tλ , t!λ F)

q q
induced by the forgetful functor is an isomorphism in degree λ(2ρ̌). Hence we obtain an inclusion
(6.1.2) (λ,2ρ̌)
SǦ (F) λ ∼= H (Tλ , t!λ F) ֒→ HA (Tλ , t!λ F). Using again the spectral sequence argument, the
q
morphism

HA (pt) ⊗ SǦ (F) λ → HA (Tλ , t!λ F)
q q
induced by the cup product is an isomorphism of HA (pt)-modules, which finishes the proof.
q
We deduce the following result from Lemma 2.2.1.

Corollary 6.2.3. For λ ∈ X and F in PervǦ(O) (GrǦ ), The morphism κtop
F ,λ is injective.
Proof. By the localization theorem in equivariant cohomology, the morphism

Q~ ⊗S~ HA (i!λ F) → Q~ ⊗S~ HA (Tλ , i!λ F)
q q
induced by (ιλ )! is an isomorphism (since λ is the only A-fixed point in Tλ ). As HA (i!λ F) is free
q
over S~ (see Lemma 2.2.1(1)), we deduce that (ιλ )! is injective, which implies that κtopF ,λ is also
injective.
We will also need the following result, which again follows from the fact that the spectral se-
quence (6.2.2) degenerates.
Lemma 6.2.4. Let F in Perv Ǧ(O) (GrǦ ) and λ ∈ X.
The forgetful morphism HA (i!λ F) → HŤ (i!λ F) induces an isomorphism
q q
∼
HA (i!λ F) ~ · HA (i!λ F) −→ HŤ (i!λ F).
q q q
6.3. Restriction to a Levi subgroup. Below we will make extensive use of the geometric descrip-
tion of the functor of restriction to a Levi subgroup, due to Mirković–Vilonen [MV] in the (crucial)
case of the maximal torus, and to Beilinson–Drinfeld [BD] in the general case.
Let P̌ ⊂ Ǧ be a parabolic subgroup containing B̌, and let Ľ ⊂ P̌ be the (unique) Levi factor
containing Ť . Let also P̌ − be the opposite parabolic subgroup. Note that Ť is also a maximal torus
of Ľ, and B̌Ľ := B̌ ∩ Ľ is a Borel subgroup containing Ť . We have dual groups G and L defined as
in §6.1. Consider the diagram
qP̌ − iP̌ −
GrĽ o GrP̌ − / Gr
Ǧ (6.3.1)
where qP̌ − is induced by the projection P̌ − ։ Ľ whose kernel is the unipotent radical of P̌ − , and
iP̌ − is induced by the embedding P̌ − ֒→ Ǧ. Define the functor
′
RǦ
Ľ
:= (qP̌ − )∗ ◦ (iP̌ − )! : Dcb (GrǦ ) → Dcb (GrĽ ),
where Dcb (GrǦ ) is the derived category of constructible complexes of C-vector spaces on the ind-
variety GrǦ which are supported on a finite union of Ǧ(O)-orbits, and similarly for Dcb (GrĽ ).
The functor ′ RǦ Ľ
does not map the subcategory Perv Ǧ(O) (GrǦ ) of Dcb (GrǦ ) into the subcategory
PervĽ(O) (GrĽ ) of Dcb (GrĽ ); however, the following modification of this functor has this property.
Recall that the connected components of GrĽ are parametrized by the quotient X/(ZŘĽ ), see
[BD, Proposition 4.5.4]. (Here ŘĽ denotes the coroots of Ľ, and ZŘĽ is the lattice they generate.)
38
If M is in Dcb (GrĽ ) and χ ∈ X/(ZŘĽ ), we denote by Mχ the restriction of M to the corresponding
connected component. Define the functor RǦ Ľ
: Dcb (GrǦ ) → Dcb (GrĽ ) by the formula
M
′ Ǧ
RǦ
Ľ
(M ) = RĽ
(M ) χ
[χ(2ρǦ − 2ρĽ )],
χ∈X/(ZŘĽ )
where ρǦ and ρĽ are the half sums of positive roots of Ǧ and Ľ. It is proved in [BD, Proposition
5.3.29] that RǦ
Ľ
restricts to a functor
RǦ
Ľ
: PervǦ(O) (GrǦ ) → PervĽ(O) (GrĽ ).
Moreover, it is explained in [BD, §5.3.30] that this functor is a tensor functor.
Using the base change theorem one can easily construct an isomorphism of tensor functors
∼
→ FĽ ◦ RǦ
FǦ − Ľ
,
see e.g. [AHR, §4.1] for details. We deduce a morphism of algebraic groups
L = Aut⋆ (FĽ ) → Aut⋆ (FĽ ◦ RǦ ) ∼ Aut⋆ (FǦ ) = G.
Ľ =
It is known that this morphism is injective, and identifies L with the Levi subgroup of G whose
root system is the system of coroots of Ľ. Hence we will consider L as a subgroup of G. By
construction the following diagram is commutative:
RǦ
Ľ
PervǦ(O) (GrǦ ) / Perv
Ľ(O) (GrĽ )
SǦ SĽ
V 7→V|L
Rep(G) / Rep(L)
Note that by the constructions of §6.1 we have identified T with a subgroup of L, but also with
a subgroup of G. These identifications are compatible with the identification of L as a subgroup
of G, see e.g. [AHR, §4.1]. Moreover, by the base change theorem there exists a canonical isomor-
phism of functors
RĽ ◦ RǦ ∼
= RǦ .
Ť Ľ Ť
(6.3.2)
6.4. Restriction to a Levi subgroup: cofibers. Let P̌ , P̌ − , Ľ be as in §6.3. Let also λ ∈ X and F in
PervǦ(O) (GrǦ ). We want to compare the S~ -modules

HA (iǦ !
λ) F and HA (iĽ ! Ǧ
λ ) RĽ (F) ,
q q
where iǦ Ľ P̌
λ and iλ are the inclusions of λ is GrǦ and GrĽ respectively. Let tλ be the inclusion of
(qP̌ − )−1 (λ) in GrǦ . Then by the base-change theorem there is a canonical isomorphism of graded
S~ -modules

HA (iĽ ! Ǧ ∼ q −1 P̌ !
λ ) RĽ (F) = HA (qP̌ − ) (λ); (tλ ) F hλ(2ρ̌L − 2ρ̌G )i.
q
As {λ} is closed in (qP̌ − )−1 (λ), the (! , ! )-adjunction for the inclusion {λ} ֒→ (qP̌ − )−1 (λ) induces a
canonical morphism of S~ -modules

HA (iǦ ! Ľ ! Ǧ
λ ) F → HA (iλ ) RĽ (F) hλ(2ρ̌G − 2ρ̌L )i. (6.4.1)
q q
If P̌ = B̌ (so that Ľ = Ť ) we have canonical isomorphisms

HA (iŤλ )! RǦ (F) ∼= HA (t!λ F)h−λ(2ρ̌G )i ∼
= SǦ (F) λ ⊗ S~ .
q q
Ť
39
Via this isomorphism, (6.4.1) identifies with the morphism κtop
F ,λ defined in (2.2.3). Moreover, one
can easily check that the following diagram is commutative:
Ǧκtop
F ,λ
-
HA (iǦ
λ)F
! / H q (iĽ )! RǦ (F) hλ(2ρ̌G − 2ρ̌L )i / S (F) ⊗ S~ hλ(2ρ̌G )i
q
(6.4.1) A λ Ľ Ľκtop Ǧ λ
F ,λ
(where for simplicity we neglect shifts of morphisms). As the morphism Ǧκtop

F ,λ is injective (see
Corollary 6.2.3), we deduce that (6.4.1) is also injective.
We will also need the following result, which is clear by construction.
Lemma 6.4.2. Let λ ∈ X and F in PervǦ(O) (GrǦ ). For any w ∈ WL ⊂ WG the following diagram
commutes:
(6.4.1)
HA (iǦ !
λ)F
/ H q (iĽ )! RǦ (F)
q
A λ Ľ
ǦΞF ,λ RǦ (F ),λ
w ĽΞ Ľ
w

wH q
(6.4.1)
A (iǦ !
wλ ) F
/ H q (iĽ )! RǦ (F) .
A λ Ľ
6.5. Construction of root vectors. There are several ways to construct simple root vectors in g :=
Lie(G) out of the equivalence S using [G1] and [MV], see e.g. [YZ, Ba]. Here we recall a simple
version essentially explained in [Va], which will be sufficient for our purposes.
Let α be a simple root. Let P̌ α be the minimal parabolic subgroup of Ǧ containing B̌ associated
with α, and let Ľα be the Levi factor of P̌ α containing Ť . Then as explained in §6.3, the “dual”
group Lα associated with Ľα can be canonically identified with the Levi subgroup of G whose Lie
algebra is g−α ⊕ t ⊕ gα , so to construct a root vector in gα it is enough to construct a root vector in
the α-weight space of the Lie algebra lα of Lα .
Let Gr◦Ľα be the connected component of 0 in GrĽα . Then the subcategory PervĽα (O) (Gr◦Ľα ) of
PervĽα (O) (GrĽα ) is closed under convolution. If we denote by M α the group of automorphisms of
the restriction F◦Ľα of the fiber functor FĽα to this subcategory, then by definition we have a natural
morphism Lα → M α , which induces a morphism lα → mα whose restriction to α-weight spaces
is an isomorphism. Hence to construct a root vector in gα it is enough to construct a root vector in
the α-weight space of mα .
Now let L be the positive generator of the Picard group of Gr◦Ľα . (See [Va, §1.4] or [Ba, §3.3]
for the explicit construction of this line bundle.) The cup product with the first Chern class of L
defines an endomorphism of the functor F◦Ľα . By the arguments of [YZ, §3.4], this endomorphism
defines an element in the α-weight space of mα , which finishes the construction.
This construction of the root vectors is clearly compatible with restriction to a Levi subgroup in
the sense of §6.3.
7. P ROOFS OF THE MAIN RESULTS
In Sections 7–8 we will use the results of Sections 3–5 for the datum T ⊂ B ⊂ G constructed in
§6.1, and for the root vectors constructed in §6.5. Note that all the objects which are denoted by
the same symbol in Sections 3–5 and in Section 6 (e.g. X, X+ , S~ ) get identified canonically.
7.1. Preliminaries on Verma modules. Recall the Verma modules V(µ) defined in §4.1.
40
Let µ ∈ t∗ , and assume that for all α ∈ R, µ(α̌) ∈
/ Z. We claim that for ν ∈ t∗ we have
HomU (g) (V(ν), V(µ)) = Cδµ,ν . (7.1.1)
Indeed, if there exists a non-zero morphism V(ν) → V(µ) then by [Hu, §3.4] ν must be in W[µ] • µ,
where W[µ] = {w ∈ W | w(µ) − µ ∈ ZR}. However, by [Hu, Theorem 3.4] and our assumption on
µ, W[µ] = {1}, which implies (7.1.1).
If λ ∈ X, then the t-weights of the U (b)-module V(µ + λ) ⊗ C−µ (where the U (b)-action is
diagonal, the action on V(µ + λ) being the restriction of the U (g)-action) are in X, and the action of
U (u) is locally finite. We deduce that the U (b)-action can be integrated to a B-action.
Lemma 7.1.2. Let µ ∈ t∗ be such that µ(α̌) ∈
/ Z for all α ∈ R, and let λ ∈ X. Then there exists a natural
isomorphism of B-modules
∼
V(µ + λ) ⊗ C−µ − → IndBT (λ).
Proof. First, let us explain how this morphism in constructed. The projection of t-modules V(µ +
λ) → Cµ+λ (with kernel V(µ + λ)− ) induces a morphism of T -modules V(µ + λ) ⊗ C−µ → Cλ . By
Frobenius reciprocity we deduce a morphism of B-modules as in the statement of the lemma.
Now we prove that this morphism is injective. For this, it suffices to prove that its restriction
to the socle of the left-hand side (as a B-module, or equivalently as a U (b)-module) is injective.
We claim that this socle is isomorphic to Cλ , and has a basis consisting of the vector 1λ+µ ⊗ 1 ∈
V(µ + λ) ⊗ C−µ . Indeed, this is equivalent to saying that the socle of V(µ + λ) is isomorphic to
Cµ+λ . However, if there exists a non-zero morphism Cν → V(µ+λ) for some ν ∈ t∗ then we obtain
a non-zero morphism of U (g)-modules V(ν) → V(µ + λ). By (7.1.1), this implies that ν = µ + λ,
proving the claim and the injectivity of the morphism.
Now, it is easy to see that the T -modules V(µ + λ) ⊗ C−µ and IndB T (λ) have the same weights,
with the same (finite) multiplicities. Hence our morphism must be an isomorphism.
Below we will use the standard order on t∗ , defined by ν ≤ µ iff µ − ν ∈ Z≥0 R+ .
Let α be a simple root, P α ⊂ G the corresponding minimal parabolic subgroup containing B,
and Lα the Levi factor of P α containing T . Let also B α := Lα ∩ B, and bα := Lie(B α ). Let µ ∈ t∗
/ Z for any β ∈ R r {±α}. We claim that for ν ∈ t∗ we have
be such that µ(β̌) ∈


C if ν = µ;
∼
HomU (g) (V(ν), V(µ)) = C if ν = sα • µ and ν < µ; (7.1.3)


0 otherwise.
Indeed, if there exists a non-zero morphism V(ν) → V(µ) with ν 6= µ then, with the same notation
as above we must have ν ∈ W[µ] • µ and ν < µ. Again by [Hu, Theorem 3.4], this implies that
ν = sα • µ. On the other hand, if ν = sα • µ and ν < µ, then the ν-weight space of V(µ) is
one-dimensional, and consists of singular vectors by [Hu, Proposition 1.4].
In the following lemma, for µ ∈ t∗ we denote by Vα (µ) the Verma module associated with µ for
the reductive group Lα with Borel subgroup B α .
Lemma 7.1.4. Let µ ∈ t∗ be such that µ(β̌) ∈
/ Z for all β ∈ R r {±α}, and let λ ∈ X. Then there exists a
natural isomorphism of B-modules
∼
V(µ + λ) ⊗ C−µ − → IndB α
B α V (µ + λ) ⊗ C−µ .
Proof. First, let us explain the construction of this morphism. As in the proof of Lemma 7.1.2
V(µ + λ) ⊗ C−µ has a natural structure of B-module, and Vα (µ + λ) ⊗ C−µ has a natural structure
of B α -module. The subspace of V(µ + λ) ⊗ C−µ spanned by weight spaces whose weight is not
in λ + Zα is stable under the action of B α , and the quotient by this subspace is clearly isomorphic
41
to Vα (µ + λ) ⊗ C−µ . Hence we have constructed a morphism of B α -modules V(ν + λ) ⊗ C−µ →
Vα (µ + λ) ⊗ C−µ . Using Frobenius reciprocity we obtain the desired morphism of B-modules.
Now we prove that this morphism is injective. As in Lemma 7.1.2, it is enough to prove that its
restriction to the socle of the left-hand side is injective. But it follows from (7.1.3) that this socle
has dimension 1 or 2, and injectivity is clear by construction.
Finally, one can deduce surjectivity as in the proof of Lemma 7.1.2 by comparing characters.
7.2. Generic and sub-generic situations: classical case. Let t∗rs and (g/u)∗rs be the sets of elements
in t∗ and (g/u)∗ which are regular semisimple (as elements of g∗ ). Then (g/u)∗rs = (g/u)∗ ×t∗ t∗rs ,
and the action of B on (g/u)∗ induces an isomorphism of B-varieties
∼
→ (g/u)∗rs
B/T × t∗rs − (7.2.1)
where the B-action on the left-hand side is trivial on t∗rs ,
and given by left multiplication on B/T
(see e.g. [J2, p. 188]). In particular, we deduce that for any λ ∈ X there is a natural isomorphism of
B-modules
C[t∗rs ] ⊗S(t) S(g/u) ⊗ C−λ ∼
= C[t∗rs ] ⊗ IndB
T (−λ). (7.2.2)
Now let α be a simple root, and let P α , Lα , B α be defined as in §7.1. Let also lα be the Lie algebra
of Lα , and uα := lα ∩ u. Let t∗α−rs be the complement in t∗ of the collection of hyperplanes defined
by the equations β̌ for β ∈ R r {±α}. Let also (g/u)∗α−rs = (g/u)∗ ×t∗ t∗α−rs , and (lα /uα )∗α−rs =
(lα /uα )∗ ×t∗ t∗α−rs .
Lemma 7.2.3. The (coadjoint) action of B on (g/u)∗ induces an isomorphism of B-varieties
∼
B ×B α (lα /uα )∗α−rs −
→ (g/u)∗α−rs .
Proof. Both varieties under consideration are smooth complex varieties, hence it is enough to
prove that the map is bijective. We use the Killing form to get identifications (g/u)∗ ∼= b, (lα /uα )∗ ∼
=
α α α α
b (where b := Lie(B )), and define bα−rs , bα−rs in an obvious way.
First, let x ∈ bα−rs , and consider the Jordan decomposition x = s + n. There exists u ∈ B such
that u · s ∈ t, and we must have u · s ∈ tα−rs . Then u · n ∈ Zb (u · s) ⊂ bα , which implies that
u · x ∈ bαα−rs . This proves surjectivity of our map.
Next, we prove injectivity. Let u1 , u2 ∈ B and x1 , x2 ∈ bαα−rs , and assume that u1 · x1 = u2 · x2 .
Consider the Jordan decompositions x1 = s1 +n1 , x2 = s2 +n2 , so that u1 ·s1 = u2 ·s2 . Conjugating if
necessary x1 by an element of B α (and modifying u1 accordingly), one can assume that s1 ∈ tα−rs .
Then the fact that (u−1 α −1 α
2 u1 ) · s1 = s2 ∈ b implies that u2 u1 ∈ B . We deduce that
(u1 ×B α x1 ) = (u2 (u−1 −1

2 u1 ) ×B α (u1 u2 ) · x2 ) = (u2 ×B α x2 )
in B ×B α bαα−rs , which finishes the proof.

In particular it follows from Lemma 7.2.3 that for any λ ∈ X there is a natural isomorphism of
B-modules

C[t∗α−rs ] ⊗S(t) S(g/u) ⊗ C−λ ∼
= IndB ∗ α α
B α C[tα−rs ] ⊗S(t) S(l /u ) ⊗ C−λ . (7.2.4)
Remark 7.2.5. The same arguments as in Lemma 7.2.3 can be used to prove the following more
general claim, which will not be used in this paper. Let I be a set of simple roots, and consider
the associated Levi subgroup LI containing T . Let B I = B ∩ LI , uI = Lie(U ∩ LI ). Let t∗I−rs be
the complement in t∗ of the collection of hyperplanes defined by the coroots associated to roots in
R r ZI. Let (g/u)∗I−rs = (g/u)∗ ×t∗ t∗rs , and similarly for (lI /uI )∗I−rs . Then the coadjoint action of B
∼
induces an isomorphism of varieties B ×B I (lI /uI )∗I−rs − → (g/u)∗I−rs .
42
7.3. Generic and sub-generic situations: quantum case. Set a := Lie(A) = t×A1 . We will identify
a∗ with t∗ × A1 in the natural way. We denote by C[a∗rs ] the localization of S~ with respect to the
collection {α̌ + n~ | α ∈ R, n ∈ Z}. If (ν, a) ∈ a∗ , we denote by CS~ (ν, a) the one-dimensional
S~ -module associated with (ν, a).
Lemma 7.3.1. For any λ ∈ X, there exists a natural isomorphism of B-modules and C[a∗rs ]-modules
∼
C[a∗rs ] ⊗S~ M(λ) −
→ C[a∗rs ] ⊗ IndB
T (−λ).
Proof. Recall that there exists a morphism of T -modules and S~ -modules MG (λ) → MT (λ) =
S~ ⊗ C−λ , see (4.6.1). Using Frobenius reciprocity we deduce a morphism as in the statement of
the lemma.
First, we claim that for any (ν, a) ∈ a∗ such that ν(α̌) ∈
/ aZ for all α ∈ R, the induced morphism
of B-modules
CS~ (ν, a) ⊗S~ M(λ) → IndB T (−λ) (7.3.2)
is an isomorphism. Indeed, if a = 0, then ν ∈ t∗rs , and morphism (7.3.2) can be identified with
the specialization of isomorphism (7.2.2) at ν, hence we are done. Now, assume a 6= 0. Then
there is an algebra isomorphism U~ (g)op ⊗C[~] Ca ∼ = U (g) which maps x ∈ g ⊂ U~ (g) to −ax
(see §4.2). Using this isomorphism, the left-hand side of (7.3.2) identifies with the U (b)-module
V(− a1 ν − λ − ρ) ⊗ C 1 ν+ρ . By our assumption − a1 ν − ρ satisfies the assumptions of Lemma 7.1.2,
a
hence (7.3.2) is an isomorphism in this case also.
Now we deduce that our morphism of C[a∗rs ]-modules is an isomorphism. As this morphism is
B-equivariant, it is sufficient to prove that its restriction to each T -weight space is an isomorphism.
It is easily checked that for any λ ∈ X the λ-weight spaces of both modules are free C[a∗rs ]-modules
of the same rank. Choose a basis of each of these spaces, and consider the determinant dλ ∈ C[a∗rs ]
of the restriction of our morphism in these bases. Then for each (ν, a) ∈ a∗ such that ν(α̌) ∈ / aZ
for all α ∈ R, we have dλ (ν, a) 6= 0. However a polynomial P ∈ S~ which does not vanish on
any hyperplane α̌ + n~ = 0 (with α ∈ R, n ∈ Z) is necessarily a scalar multiple of a product of
polynomials of the form α̌ + n~. Hence dλ is invertible in the algebra C[a∗rs ], which proves that our
morphism is indeed an isomorphism.
Let now α be a simple root, and let Lα be the Levi subgroup defined in §7.1, with its Borel
subgroup B α . For λ ∈ X, we denote by Mα (λ) the asymptotic universal Verma module associated
with λ for the group Lα and its Borel subgroup B α . Let C[a∗α−rs ] be the localization of S~ with
respect to the collection {β̌ + n~ | β ∈ R r {±α}, n ∈ Z}.
Lemma 7.3.3. For any λ ∈ X, there exists a natural isomorphism of B-modules and C[a∗α−rs ]-modules

C[a∗α−rs ] ⊗S~ M(λ) → IndB ∗ α
B α C[aα−rs ] ⊗S~ M (λ)
where in the right-hand side B α acts trivially on C[a∗α−rs ].

Proof. As in Lemma 7.3.1, the morphism in question is constructed using the morphism M(λ) →
Mα (λ) of (4.6.1) and Frobenius reciprocity.
To prove that this morphism is an isomorphism it is enough to prove that for any (ν, a) ∈ a∗
such that ν(β̌) ∈
/ aZ for all β ∈ R r {±α}, the natural morphism of B-modules

CS~ (ν, a) ⊗S~ M(λ) → IndB α
B α CS~ (ν, a) ⊗S~ M (λ) (7.3.4)
is an isomorphism. (Note that the functor IndB B α is exact by [J1, Corollary I.5.13], hence specializa-
tion commutes with induction here.)
If a = 0, then morphism (7.3.4) can be identified with the specialization of isomorphism (7.2.4)
at ν, hence we are done. If a 6= 0, by the same arguments as in the proof of Lemma 7.3.1, morphism
43
(7.3.4) can be identified with the morphism

1 B α 1
V − ν − λ − ρ ⊗ C 1 ν+ρ → IndB α V − ν − λ − ρ ⊗ C 1 ν+ρ
a a a a
considered in Lemma 7.1.4. Hence it is an isomorphism in this case also.
7.4. Generic and sub-generic situations: isomorphism. Fix λ ∈ X and F in Perv Ǧ(O) (GrǦ ). To
simplify notation we set V := SǦ (F).
As a first step towards proving Theorem 2.2.4, we construct a canonical isomorphism of C[a∗rs ]-
modules
B
C[a∗rs ] ⊗S~ V ⊗ M(λ) ∼
= C[a∗rs ] ⊗S~ HA (i!λ F).
q
(7.4.1)
In fact, we deduce this isomorphism from the fact that the morphism
B
C[a∗rs ] ⊗S~ HA (i!λ F) → C[a∗rs ] ⊗ Vλ , resp. C[a∗rs ] ⊗S~ V ⊗ M(λ) → C[a∗rs ] ⊗ Vλ (7.4.2)
q
induced by κtop alg

F ,λ , resp. κV,λ , is an isomorphism. For the first morphism this follows from the
localization theorem in equivariant cohomology. Let us now prove that the second morphism is
also an isomorphism. In fact we have a series of isomorphisms
B B B
C[a∗rs ] ⊗S~ V ⊗ M(λ) ∼
= V ⊗ (C[a∗rs ] ⊗S~ M(λ)) ∼
= V ⊗ C[a∗rs ] ⊗ IndBT (−λ)
B
∼
= C[a∗rs ] ⊗ IndB ∼
= C[a∗rs ] ⊗ Vλ .
T (V ⊗ C−λ )
Here the second isomorphism follows from Lemma 7.3.1, the third one from the tensor identity,
B
and the last one from the isomorphism IndB T (V ⊗ C−λ ) = (V ⊗ C−λ )T given by Frobenius
reciprocity. By construction the composition of these isomorphisms is precisely the right-hand
morphism in (7.4.2).
Now, let α be a simple root. As a second step towards proving Theorem 2.2.4, we want to show
that (7.4.1) restricts to a canonical isomorphism of C[a∗α−rs ]-modules
B
C[a∗α−rs ] ⊗S~ V ⊗ M(λ) ∼
= C[a∗α−rs ] ⊗S~ HA (i!λ F).
q
(7.4.3)
Let P̌ α be the minimal parabolic subgroup of Ǧ containing B̌ associated with α, and let Ľα be the
unique Levi factor of P̌ α containing Ť . By the constructions of §6.3, these data determine a Levi
factor Lα in G, hence the corresponding minimal parabolic subgroup P α containing B.
Consider first the right-hand side of (7.4.3). We will use the constructions of §6.4 for the Levi
subgroup Ľα . By the localization theorem in equivariant cohomology, the morphism
α !
C[a∗α−rs ] ⊗S~ HA (i!λ F) → C[a∗α−rs ] ⊗S~ HA (iĽ Ǧ
λ ) RĽα (F)
q q
induced by (6.4.1) is an isomorphism.

Consider now the left-hand side of (7.4.3). Here we will use the constructions of §7.3. We have
a series of isomorphisms
B B B
C[a∗α−rs ] ⊗S~ V ⊗ M(λ) ∼
= V ⊗ (C[a∗α−rs ] ⊗S~ M(λ)) ∼
= V ⊗ IndB ∗ α
B α (C[aα−rs ] ⊗S~ M (λ))
B B α
∼
= C[a∗α−rs ] ⊗S~ IndB α ∼
= C[a∗α−rs ] ⊗S~ V|Lα ⊗ Mα (λ)
B α (V|Lα ⊗ M (λ)) .
Here the second isomorphism follows from Lemma 7.3.3, the third one from the tensor identity,
B B α
and the last one from the isomorphism IndB α ∼
= V|Lα ⊗ Mα (λ)
B α (V|Lα ⊗ M (λ)) given by
Frobenius reciprocity.
44
Using these isomorphisms, to construct (7.4.3) we only have to construct a canonical isomor-
phism of S~ -modules
B α α !
V|Lα ⊗ Mα (λ) ∼
= HA (iĽ Ǧ
λ ) RĽα (F) .
q
As Ľα has semisimple rank one, this isomorphism is constructed in §A.7 below.
One can check that the isomorphism obtained from (7.4.3) by extension of scalars to C[a∗rs ] coin-
cides with (7.4.1). In other words, (7.4.3) is the restriction of (7.4.1) to
B B
C[a∗α−rs ] ⊗S~ V ⊗ M(λ) ⊂ C[a∗rs ] ⊗S~ V ⊗ M(λ) .
7.5. Equivariance. Recall the operators ΩV,λ

w defined in §4.8.
Lemma 7.5.1. Let λ ∈ X and F in PervǦ(O) (GrǦ ). Isomorphism (7.4.1) is W -equivariant, in the sense
that for any w ∈ W the following diagram commutes:
B (7.4.1)
C[a∗rs ] ⊗S~ S(F) ⊗ M(λ) / C[a∗ ] ⊗S H q (i! F)
rs ~ A λ
S(F ),λ
C[a∗rs ]⊗S~ Ωw C[a∗rs ]⊗S~ ΞF
w
,λ

B (7.4.1)

C[a∗rs ] ⊗S~ w S(F) ⊗ M(wλ) / C[a∗ ] ⊗S w HA (i!wλ F)
q
rs ~
Proof. It is enough to prove the commutativity when w is a simple reflection. So let α be a simple
root. Recall that isomorphism (7.4.1) is the restriction of isomorphism (7.4.3) to a∗rs . Now iso-
morphism (7.4.3) is deduced from the isomorphism of Theorem 2.2.4 for Ľα , proved (directly) in
§A.7 below. Moreover, by Corollary 3.3.3 (or Lemma 4.6.5) and Lemma 6.4.2, the operators Ω and
Ξ for the group Ǧ can also be constructed from the similar operators for the group Ľα . Hence
the commutativity for sα follows from Theorem 2.5.5 for Ľα , which is proved (directly) in §A.9
below.
7.6. Proof of the main results: quantum case. If α ∈ R is any root, we define the localization
C[a∗α−rs ] of S~ by the same recipe as for simple roots (see §7.3). Then for any w ∈ W we have
w(C[a∗α−rs ]) = C[a∗w(α)−rs ] as subalgebras of C[a∗rs ] To finish the proofs we will need the following
obvious lemma.
Lemma 7.6.1. Let M and N be free S~ -modules of finite rank, and let
∼
ϕ : C[a∗rs ] ⊗S~ M −
→ C[a∗rs ] ⊗S~ N
be an isomorphism of C[a∗rs ]-modules. Assume that, for any α ∈ R, ϕ restricts to an isomorphism of
∼ ∼
C[a∗α−rs ]-modules C[a∗α−rs ] ⊗S~ M −
→ C[a∗α−rs ] ⊗S~ N . Then ϕ restricts to an isomorphism M −
→ N.
Proof of Theorem 2.2.4. Let λ ∈ X and F in Perv Ǧ(O) (GrǦ ), and set V := S(F).
Injectivity of κalg top
V,λ follows from Lemma 4.6.4, while injectivity of κF ,λ is proved in Corollary
q !
6.2.3. By Lemma 2.2.1(1), HA (iλ F) is a free S~ -module. It follows from Proposition 3.2.3 and
B
the first isomorphism in Lemma 2.4.1 that the same is true for V ⊗ M(λ) . In (7.4.1) we have
constructed an isomorphism between the extensions of scalars of these S~ -modules to C[a∗rs ]. By
(7.4.3) this isomorphism restricts to an isomorphism bewteen extensions of scalars to C[a∗α−rs ] for
any simple root α. From Lemma 7.5.1 we deduce that the same property is true for any root. Hence
the isomorphism follows from Lemma 7.6.1.
Proof of Theorem 2.5.5. As our S~ -modules are free, it is enough to prove commutativity after re-
striction to a∗rs ⊂ a∗ . In this setting the claim follows from Lemma 7.5.1.
45
7.7. Classical analogues.
Proof of Theorem 2.3.1. One can prove Theorem 2.3.1 using exactly the same strategy as for Theo-
rem 2.2.4. Alternatively, isomorphism ζ F ,λ can be deduced from isomorphism ζF ,λ of Theorem
2.2.4 using Lemma 3.5.2 and Lemma 6.2.4. (The statements about injectivity are easy.)
Proof of Theorem 2.5.7. One can prove Theorem 2.5.7 using exactly the same strategy as for The-
orem 2.5.5. Alternatively, one can deduce Theorem 2.5.7 from Theorem 2.5.5 using Proposition
5.4.1, Lemma 3.5.2 and Lemma 6.2.4 (see also Remark 5.4.2).
8. C OMPLEMENTARY RESULTS AND APPLICATIONS
8.1. Convolution. In this subsection we construct the morphisms Conv considered in §1.3, and
prove their compatibility with the isomorphisms of Corollary 2.4.2, thereby finishing the proof of
the theorem stated in §1.5.
First, consider the “geometric” setting. Multiplication induces a morphism D~ (X)⊗C[~] D~ (X) →
D~ (X). One can check that for λ, µ ∈ X this morphism induces a morphism of graded S~ -modules
(λ) (λ+µ)
D~ (X)λ ⊗S~ D~ (X)µ → (λ+µ)D~ (X)λ+µ .
Hence for V, V ′ in Rep(G) we obtain a morphism of graded S~ -modules
G G
Convgeom
V,V ′ ,λ,µ :
(λ)
V ⊗ D~ (X)λ hλ(2ρ̌)i ⊗S~ (λ+µ)
V ′ ⊗ D~ (X)µ hµ(2ρ̌)i
G
→ (λ+µ) V ⊗ V ′ ⊗ D~ (X)λ+µ h(λ + µ)(2ρ̌)i.
Now, consider the “algebraic” setting. Note that if λ, µ ∈ X we have a canonical isomorphism
of (S~ , U~ (g))-bimodules
S~ hhλii ⊗S~ M(µ) ∼
= M(λ + µ)
which sends 1⊗vµ to vλ+µ . In particular if V is in Rep(G) and φ : M(0) → V ⊗M(µ) is a morphism
of (S~ , U~ (g))-bimodules, then we can consider S~ hhλii ⊗S~ φ as a morphism of bimodules M(λ) →
V ⊗ M(λ + µ).
Fix V, V ′ in Rep(G) and λ, µ ∈ X. Then following [ABG, §8.4] we define the morphism of graded
S~ -modules

ConvalgV,V ′ ,λ,µ : Hom M(0), V ⊗ M(λ) hλ(2ρ̌)i ⊗S~
(λ)
Hom M(0), V ′ ⊗ M(µ) hµ(2ρ̌)i

→ Hom M(0), V ⊗ V ′ ⊗ M(λ + µ) h(λ + µ)(2ρ̌)i
(where as usual we consider morphisms of (S~ , U~ (g))-bimodules) which sends a pair (φ, ψ) to the
following composition:
φ idV ⊗(S~ hhλii⊗S~ ψ)

M(0) / V ⊗ M(λ) / V ⊗ V ′ ⊗ M(λ + µ).
Finally we consider the “topological” setting. Here again we follow [ABG, §8.7] (though we
have to be more careful because we use equivariant cohomology.) Let F, G in PervǦ(O) (Gr), and let
λ, µ ∈ X. The convolution F ⋆ G is defined in terms of the “convolution diagram”
mult : Ǧ(K) ×Ǧ(O) Gr → Gr
e G) where F ⊠
induced by left multiplication of Ǧ(K) on Gr. More precisely, F ⋆ G = mult∗ (F ⊠ eG
is the twisted external product, as defined e.g. in [MV, §4].
46
Let ν := λ + µ. Consider the cartesian square
jν
mult−1 (ν)
/ Ǧ(K) ×
Ǧ(O) Gr
multν mult
iν
ν / Gr.
This diagram is A-equivariant if we consider Ǧ(K) ×Ǧ(O) Gr as an A-variety where Ť acts by left
multiplication on Ǧ(K), and C× acts diagonally by loop rotation. By the base-change theorem we
have an isomorphism
e G) ∼
i!ν (F ⋆ G) = i!ν mult∗ (F ⊠ e G),
= (multν )∗ jν! (F ⊠
so that we obtain an isomorphism

HA i!ν (F ⋆ G) ∼= HA mult−1 (ν), jν! (F ⊠
e G) .
q q
Now let kλ,µ : {λ ×Ǧ(O) µ} ֒→ Ǧ(K) ×Ǧ(O) Gr be the obvious embedding. The (! , ! )-adjunction for
the embedding {λ ×Ǧ(O) µ} ֒→ mult−1 (ν) induces a morphism

q !
HA kλ,µ e G) → HAq mult−1 (ν), jν! (F ⊠
(F ⊠ e G) ,
which can be reinterpreted as a morphism

q !
HA kλ,µ e G) → HAq i!λ+µ (F ⋆ G) .
(F ⊠ (8.1.1)
Lemma 8.1.2. There exists a canonical isomorphism of graded S~ -modules

q !
HA kλ,µ (F ⊠e G) ∼= HA (i!λ F) ⊗S~ (λ)HA (i!µ G).
q q
Proof. Choose some closed finite union of Ǧ(O)-orbits Y ⊂ Gr such that G is supported on Y ,
and a closed normal subgroup H ✁ Ǧ(O) of finite codimension c, which acts trivially on Y . Let
f : Ǧ(K)/H × Y → Ǧ(K) ×Ǧ(O) Y and g : Ǧ(K)/H → Gr be the natural projections. Then by

definition of F ⊠ eG ∼
e G we have a canonical isomorphism f ∗ F ⊠ = (g∗ F) ⊠ G hence (since both f
and g are smooth morphisms of relative dimension c) a canonical isomorphism

f! F ⊠eG ∼= (g! F) ⊠ G. (8.1.3)
We will consider Ǧ(K)/H × Y as an A-variety where any t ∈ Ť , resp. a ∈ C× acts by
t · (gH, x) := (tgt−1 H, t · x), resp. a · (gH, x) := (a · g)λ(a)−1 H, λ(a) · (a · x)).
With this definition, the morphism f is A-equivariant, and the point (λ̃, µ) ∈ Ǧ(K)/H × Y is
A-stable. (Here λ̃ := λ̂H/H, where λ̂ is λ considered as an element of Ǧ(K).) Now kλ,µ factors as
the composition
∼ lλ,µ f
{(λ, µ)} / {(λ̃, µ)} / Ǧ(K)/H × Y / Ǧ(K) ×
Ǧ(O) Y,
hence using (8.1.3) we obtain an isomorphism

q !
HA kλ,µ e G) ∼
(F ⊠
q !
= HA lλ,µ (g! F ⊠ G) .
Now we have
(g! F ⊠ G) ∼
q !
HA lλ,µ = HA (i!λ F) ⊗S~ (λ)
HA (i!µ G)
q q
(where we use the fact that (λ)HA (i!µ G) is free over S~ by Lemma 2.2.1, so that we have a Künneth
q
formula in equivariant cohomology.) This finishes the proof.
47
Using the isomorphism of Lemma 8.1.2 we can reinterpret (8.1.1) as a morphism

Convtop
q ! (λ) q !
F ,G,λ,µ : HA (iλ F) ⊗S~ HA (iµ G) → HA i!λ+µ (F ⋆ G) .
q
Remark 8.1.4. In case λ, µ are antidominant, the morphism Convtop

F ,G,λ,µ can also be described in
terms of Wakimoto sheaves, as in [ABG, diagram on p. 652].
The following result will be proved in §8.2
Proposition 8.1.5. Let F, G be in PervǦ(O) (Gr) and λ, µ ∈ X. Under the isomorphisms of Corollary 2.4.2,
the morphisms Convgeom alg top
S(F ),S(G),λ,µ , ConvS(F ),S(G),λ,µ and Conv F ,G,λ,µ match.
Remark 8.1.6. (1) Note that, unlike most of the constructions in the paper, the morphisms Conv
are not compatible with restriction to a Levi subgroup (see §3.3, §4.6 or §6.4) in the obvious
strong sense. We will use a weaker compatibility result in the proof of Proposition 8.1.5.
(2) Proposition 8.1.5 has an obvious “classical analogue”, that we do not state for simplicity,
but which can be proved by the same methods.
(3) Using similar constructions one can define, for V, V ′ in Rep(G) and λ, µ ∈ X a “geometric”
convolution morphism
G G G
V ⊗ M(λ) ⊗S~ (λ)
Vµ′ ⊗ S~ → (V ⊗ V ′ )λ+µ ⊗ S~
(where for simplicity we disregard the grading) and, for F, G in PervǦ(O) (Gr) and λ, µ ∈ X,
a “topological” convolution morphism

HA (i!λ F) ⊗S~ (λ)HA (t!µ G) → HA t!λ+µ (F ⋆ G) .
q q q
(These maps “extend” Convalg top

V,V ′ ,λ,µ and Conv F ,G,λ,µ in the natural sense, using the inclu-
sions κalg alg
V ′ ,µ , κV ⊗V ′ ,λ+µ , (ıµ )! and (ıλ+µ )! .) Then the same arguments as for Proposition 8.1.5
show that the following diagram commutes:
G G G
V ⊗ M(λ) ⊗S~ (λ) Vµ′ ⊗ S~ / (V ⊗ V ′ )λ+µ ⊗ S~
Th. 2.2.4 & Lem. 2.2.1 ≀ Lem. 2.2.1 ≀

HA (i!λ F) ⊗S~ (λ)H q (t! G) / H q t! (F ⋆ G) .
q
A µ A λ+µ
In particular, for V = V ′ = C[G] (the regular representation of G, considered as an ind-

object in Rep(G)) and taking the direct sum over λ and µ, one can check that composing the
“algebraic” map with the morphism induced by the multiplication morphism m : C[G] ⊗
C[G] → C[G] one obtains an action of the algebra D~ (X) on the C[~]-algebra Drel
~ of relative
asymptotic differential operators along the fibers of the projection G → G/T . This action
can be realized “topologically” using the composition
S(m)
HA (i!λ R) ⊗S~ HA (t!µ R) −→ HA t!λ+µ (R ⋆ R) −−−→ HA t!λ+µ R ,
(λ)
q q q q
where R := S(C[G]).
In this construction one can also replace T by a Levi factor of a parabolic subgroup;
details are left to the reader.
48
8.2. Proof of Proposition 8.1.5. For λ, µ ∈ X we set
Tλ,µ := {(nλ̂ ×Ǧ(O) mµ) | n, m ∈ Ň − (K)} ⊂ Ǧ(K) ×Ǧ(O) Gr,
and we denote by tλ,µ : Tλ,µ ֒→ Ǧ(K) ×Ǧ(O) Gr the inclusion. (The notation λ̂ is defined in the
proof of Lemma 8.1.2.) Then we have decompositions
G G
Ǧ(K) ×Ǧ(O) Gr = Tλ,µ , mult−1 (Tν ) = Tλ,µ .
λ,µ∈X λ+µ=ν
∼
In fact, Tλ,µ is the inverse image of Tλ × Tλ+µ ⊂ Gr × Gr under the isomorphism Ǧ(K) ×Ǧ(O) Gr −
→
Gr × Gr sending (g1 ×Ǧ(O) g2 Ǧ(O)) to (g1 Ǧ(O), g1 g2 Ǧ(O)).
Lemma 8.2.1. For any F, G in Perv Ǧ(O) (Gr), there exists a natural isomorphism of graded S~ -modules

q e G) ∼
HA Tλ,µ , t!λ,µ (F ⊠ = HA (Tλ , t!λ F) ⊗S~ (λ)HA (Tµ , t!µ G).
q q
e λ := λ̃ · Ň − (O− )1 ⊂ Ǧ(K) (where Ň − (O− )1 ⊂ Ň − (O− ) is the kernel of the

Proof. If we set T
e λ ֒→ Ǧ(K) ։ Gr induces an isomorphism T eλ −∼
evaluation at z = ∞), then the composition T → Tλ .
Then the same arguments as in the proof of Lemma 8.1.2 prove our claim.
Now we fix ν ∈ X and F, G in PervǦ(O) (Gr). For λ ∈ X we set
G
T≥λ
ν := Tλ′ ,µ ⊂ mult−1 (Tν ),
λ′ +µ=ν
λ′ ≥λ
−1
and denote by t≥λ ≥λ ≥λ
ν : Tν ֒→ Ǧ(K) ×Ǧ(O) Gr the inclusion. Then Tν is closed in mult (Tν ), and
Tλ,ν−λ is open in T≥λν . It follows in particular from Lemma 8.2.1 and Lemma 2.2.1 that for any

! e
λ, µ such that λ + µ = ν, the cohomology HA Tλ,µ , tλ,µ (F ⊠ G) is concentrated in degrees of the
q
same parity as ν(2ρ̌). From this parity vanishing observation, one can deduce that the long exact
sequence associated with the decomposition T≥λ ν = T>λ >λ
ν ⊔ Tλ,ν−λ (where Tν has the obvious
definition) for the object (t≥λ ! e
ν ) (F ⊠ G) breaks into a family of short exact sequences. And then
(using the base change theorem) we deduce that the graded S~ -module

HA Tν , t!ν (F ⋆ G)
q
admits a deacreasing X-filtration with part bigger than λ isomorphic to HA (T≥λ

q ≥λ ! e
ν , (tν ) (F ⊠ G)),
and with associated graded
M M
q e G) ∼
HA Tλ,µ , t!λ,µ (F ⊠ = HA (Tλ , t!λ F) ⊗S~ (λ) HA (Tµ , t!µ G).
q q
λ+µ=ν λ+µ=ν
(Here the isomorphism is provided by Lemma 8.2.1.)

Lemma 8.2.2. Under the isomorphisms

HA Tν , t!ν (F ⋆ G) ∼= S(F) ⊗ S(G) ν ⊗ S~ hν(2ρ̌)i,
q
and
HA (Tλ , t!λ F) ⊗S~ HA (Tµ , t!µ G) ∼
(λ)
= (S(F)λ ⊗ S(G)µ ) ⊗ S~ hν(2ρ̌)i
q q

! (F ⋆ G) considered above is induced by
provided by Lemma 2.2.1, the “topological” filtration on H T , t
q
A ν ν
the filtration on S(F) ⊗ S(G) ν by the subspaces
M
S(F)λ′ ⊗ S(G)ν−λ′ .
λ′ ≥λ
49
Sketch of proof. By construction of the isomorphisms in Lemma 2.2.1 it is sufficient to prove the
analogous claim for ordinary cohomology; in other words one can forget about A-equivariance.
Now recall the construction of the tensor structure in [MV, Proposition 6.4]; in particular let X be
a smooth (algebraic) curve, and consider the local analogues of Gr and the convolution diagram
over X 2 as in [MV, Equation (5.2)]. In the proof of [MV, Proposition 6.4], the authors define a
global counterpart Tν (X 2 ) ⊂ GrX 2 of Tν . Then one can consider

mult−1 Tν (X 2 ) ⊂ GrX ×Gr
e X
in the “global analogue” of Ǧ(K)×Ǧ(O) Gr. One can define locally closed ind-subvarieties Tλ,µ (X 2 )
inside this inverse image (when λ + µ = ν) which, over points in the diagonal copy of X in X 2 ,
coincide with our subvarieties Tλ,µ and which, over points outside the diagonal, coincide with
Tλ × Tµ . Then one has a filtration as above, but this time globally over X 2 . Over points in the
diagonal, this filtration coincides with the one considered above
F by construction. And over points
2
outside of the diagonal the variety Tν (X ) is a disjoint union λ+µ=ν Tλ × Tµ , and the filtration is
obtained from the decomposition of the appropriate cohomology sheaves as a direct sum as e.g. in
[MV, Equation (6.25b)]. This implies the claim.
Using these remarks we are now ready to give a proof of Proposition 8.1.5.
Proof of Proposition 8.1.5. First it is easy to check, by explicit computation, that the isomorphism
between the left-hand side and the right-hand side of the equation in Corollary 2.4.2 is compatible
with the morphisms Convgeom and Convalg .
Set ν := λ + µ, V := S(F), V ′ := S(G). To finish the proof we have to prove that the left square
in the following diagram commutes, where the isomorphisms are as in Theorem 2.2.4:
Conv top
F ,G,λ,µ (ıν )!
HA (i!λ F) ⊗S~ (λ)HA (i!µ G) / H q (i! (F ⋆ G)) / H q (Tν , t! (F ⋆ G))
q q
A ν A ν
≀ ≀ ≀
Convalg κalg
B B V,V ′ ,λ,µ B V ⊗V ′ ,ν
V ⊗ M(λ) ⊗S~ (λ) V ′ ⊗ M(µ) / V ⊗ V ′ ⊗ M(ν) / (V ⊗ V ′ )ν ⊗ S~ .
Of course it is sufficient to prove that the outer square commutes.

Now recall the filtrations on HA (Tν , t!ν (F ⋆ G)) and (V ⊗ V ′ )ν ⊗ S~ considered above. Then
q
each of our morphisms factors through the part of the filtration bigger than λ: for the first line
this follows from the fact that (λ̂ ×Ǧ(O) µ) ∈ T≥λ ν , and for the second line this can be checked by
explicit computation. Moreover, the restriction to the image of these morphisms of the projection
to HA (Tλ , t!λ F) ⊗S~ (λ) HA (Tµ , t!µ G), resp. (Vλ ⊗ Vµ ) ⊗ S~ , is injective. Hence it is enough to check
q q
that the corresponding diagram:
HA (i!λ F) ⊗S~ (λ)H q (i! G) / H q (Tλ , t! F) ⊗S (λ) H q (T , t! G)

q
A µ A λ ~ A µ µ
≀ ≀

B B
V ⊗ M(λ) ⊗S~ (λ) V ′ ⊗ M(µ) / (Vλ ⊗ Vµ ) ⊗ S~
commutes. However the upper line is induced by (ıλ )! ⊗ (ıµ )! , and the bottom line is induced by
κalg alg
V,λ ⊗ κV ′ ,µ , hence this claim is clear.
50
8.3. Reminder on dynamical Weyl groups. Let us fix V in Rep(G), λ ∈ X and w ∈ W . For any
µ ∈ X− sufficiently large we consider the morphism of C[~, ~−1 ]-modules
DWalg −1 −1
V,λ,w,µ : C[~, ~ ] ⊗ Vλ → C[~, ~ ] ⊗ Vwλ , DWalg V,λ V,λ V,λ −1
V,λ,w,µ := Ewµ ◦ Ψw,µ ◦ (Eµ ) .
This morphism is well defined by Lemma 4.4.1(1). We will sometimes extend this morphism to a
morphism of C(~)-modules C(~) ⊗ Vλ → C(~) ⊗ Vwλ in the obvious way (and denote the extension
also by DWalg
V,λ,w,µ ).
Recall that for µ ∈ t∗ we have defined a morphism P 7→ P (µ) in §4.3. We denote similarly the
induced morphisms S~ ⊗Vλ → C[~, ~−1 ] ⊗ Vλ or S~ ⊗Vwλ → C[~, ~−1 ] ⊗ Vwλ .
Lemma 8.3.1. There exists a unique isomorphism of C(~)-modules

∼
DWalg
V,λ,w : Q~ ⊗ Vλ −
→ Q~ ⊗ Vwλ
such that for any x ∈ Q~ ⊗ Vλ the following property holds: for any µ ∈ X− sufficiently large such that
x(µ) is defined, we have

DWalg alg
V,λ,w (x) (wµ) = DWV,λ,w,µ x(µ)
(in particular, the left-hand side is defined). This morphism induces an isomorphism of Q~ -modules
∼
DWalg → wQ~ ⊗ Vwλ .
V,λ,w : Q~ ⊗ Vλ −
Proof. This claim is proved in [TV, EV]. For later use, let us explain how it can be deduced from
our constructions. Unicity is clear (see e.g. §4.3). Let us prove existence.
It is enough to treat the case where w = s is the reflection associated with a simple root α
(provided we only require µ ∈ X to satisfy µ(α̌) ≤ 0, not necessarily to be antidominant), see in
particular (4.4.2). Define the isomorphism of C(~)-modules
∼
′
DWalg
V,λ,s : Q~ ⊗ Vλ −
→ Q~ ⊗ Vsλ
to be the composition
κalg
V,λ B ΩV,λ B
Q~ ⊗ Vλ ←−−− Q~ ⊗S~ V ⊗ M(λ) −−s−→ Q~ ⊗S~ V ⊗ M(sλ)
∼ ∼
κalg
V,sλ q⊗v⊗p7→s(q)p⊗v
s
−−−→ Q~ ⊗S~ Vsλ ⊗ S~ −−−−−−−−−−→ Q~ ⊗ Vsλ .
∼ ∼
(The fact that the first and third arrows are invertible was proved in the course of the proof of
Theorem 2.2.4, see §7.4.) Then for x ∈ Q~ ⊗ Vλ we set
( alg
1 ′
(−α̌)(−α̌+~)···(−α̌+(λ(α̌)−1)~) · DW V,λ,s (x) if λ(α̌) ≥ 0;
DWalg
V,λ,s (x) = ′ alg
(−α̌ − ~)(−α̌ − 2~) · · · (−α̌ + λ(α̌)~) · DWV,λ,s (x) if λ(α̌) ≤ 0
in Q~ ⊗Vsλ . It follows from Proposition 4.5.1 that the morphism DWalg

V,λ,s satisfies our requirements.

The operators DWalg

V,λ,w form the dynamical Weyl group as considered (in the non-asymptotic case)
in [TV, EV] and (in the asymptotic case) in [BrF].
51
8.4. Transverse slices and semi-infinite orbits. Recall the definition of Wλ , sλ , and nλ in §2.6. The
following result contains in particular Lemma 2.6.1. It is essentially proved in [BrF]; we explain
the details for the reader’s convenience.
Lemma 8.4.1. Let F in Perv Ǧ(O) (Gr) and λ ∈ X.
(1) There exists a canonical isomorphism of graded S~ -modules

H (Wλ ∩ Tλ , s! F) ∼ = S(F) ⊗ S~ hnλ i.
q
A λ λ
(2) Under the isomorphism of (1) and (2.2.2), the natural morphism
HA (Wλ ∩ Tλ , s!λ F) → HA (Tλ , t!λ F)
q q
induced by the (closed) inclusion Tλ ∩ Wλ ֒→ Tλ identifies with multiplication by

−λ(α̌)−1
Y Y
(−α̌ + j~)
α>0, j=0
(λ,α̌)<0

on S(F) λ ⊗ S~ .
Proof. There exists a closed subvariety V ⊂ Ǧ(O) isomorphic to a (finite dimensional) affine space,
which is stable under conjugation by A, and which satisfies the following conditions:
• the morphism V → Grλ , u 7→ u · λ, is an open embedding;
• the morphism V × Wλ → Gr induced by the Ǧ(O)-action on Gr is an open embedding.
Indeed, by NǦ (Ť )-equivariance it is enough to treat the case λ is dominant. And in this case one
can e.g. take as V the subvariety J λ considered in [NP, Lemme 2.2]. We will identify V × Wλ with
its image in Gr, and denote it by O.
We have canonical isomorphisms
= H (ı∗ t! F) ∼
H (Tλ , t! F) ∼ = H (({λ} ֒→ O ∩ Tλ )∗ (O ∩ Tλ ֒→ Gr)! F).
q q q
A λ A λ λ A
Here the first isomorphism follows from [FW, Proposition 2.3] (since λ is an attractive fixed point
of A on Tλ ), and the second one from the fact that O is open in Gr.
As F is Ǧ(O)-equivariant, there exists a canonical isomorphism
(O ֒→ Gr)∗ F ∼ = C [2λ(2ρ̌)] ⊠ F(λ) where F(λ) := (Wλ ֒→ Gr)∗ F[−2λ(2ρ̌)].
V (8.4.2)
As V is smooth of dimension dim(Grλ ) = λ(2ρ̌), we also have canonically
∼ (Wλ ֒→ Gr)! F.
F(λ) =
By definition we have
Tλ = {x ∈ Gr | lim 2ρ(s) · x = λ}.
s→∞
(Here ρ is considered as a cocharacter of Ť .) If follows that we have an isomorphism
O ∩ Tλ ∼ = (V ∩ Tλ ) × (Wλ ∩ Tλ ). (8.4.3)
λ
The variety V ∩ Tλ is open in Gr ∩ Tλ , and by [MV, Equation (3.6)] we have
−(λ,α̌)−1
Y Y
λ
Gr ∩ Tλ = Ǔ (O) · λ ∼
−
= CA (−α̌ + k~) (8.4.4)
α>0, k=0
(λ,α̌)<0
as A-varieties. In particular, V ∩ Tλ is smooth.

Using (8.4.2) we obtain (under the identification (8.4.3)) a canonical isomorphism
(O ∩ Tλ ֒→ Gr)! F ∼ = C [2 dim(V ∩ Tλ )] ⊠ (Wλ ∩ Tλ ֒→ Wλ )! F(λ) .
V ∩Tλ
52
We deduce a canonical isomorphism
HA (Tλ , t!λ F) ∼
= HA (({λ} ֒→ Wλ ∩ Tλ )∗ s!λ F)h−2 dim(V ∩ Tλ )i.
q q
Again by [FW, Proposition 2.3], we have a canonical isomorphism

HA (Wλ ∩ Tλ , s!λ F) ∼
= HA (({λ} ֒→ Wλ ∩ Tλ )∗ s!λ F).
q q
Using the first isomorphism in (2.2.2), we obtain finally an isomorphism

HA (s!λ F) ∼
= S(F) λ ⊗ S~ hλ(2ρ̌) + 2 dim(V ∩ Tλ )i.
q
Now the dimension dim(V ∩ Tλ ) can be computed using (8.4.4), and (1) follows.
To prove (2) we have to understand the natural morphism
HAq ({λ}) → HAq (V ∩ Tλ )h−2 dim(V ∩ Tλ )i.
However this morphism factorizes as the following composition:
∼
HAq ({λ}) → HAq (Grλ ∩ Tλ )h−2 dim(Grλ ∩ Tλ )i −
→ HAq (V ∩ Tλ )h−2 dim(V ∩ Tλ )i,
where the first morphism is induced by the inclusion {λ} ֒→ Grλ ∩ Tλ , and the second morphism
is given by restriction to the open subvariety V ∩ Tλ . Now the first morphism can be computed
using (8.4.4) and the reminder in §6.2, and the result follows.
8.5. Geometric realization of dynamical Weyl groups. Now we are in a position to prove Propo-
sition 2.6.2.
Proof of Proposition 2.6.2. Fix F in Perv Ǧ(O) (Gr), and set V := S(F). Then it is enough to prove that
for any simple root α and any λ ∈ X such that λ(α̌) ≥ 0 the following diagram commutes, where
the vertical isomorphisms are induced by those of Lemma 8.4.1(1) and where s := sα :
DWgeom
F ,λ,s
Q~ ⊗S~ HA (Wλ ∩ Tλ , s!λ F) / Q~ ⊗S H q (Wsλ ∩ Tsλ , s! F)
q
~ A sλ
≀ ≀
DWalg
V,λ,s

Q~ ⊗ Vλ / Q~ ⊗ Vsλ
Consider the isomorphism of C(~)-modules

∼
′
DWgeom !
→ Q~ ⊗S~ HA (Tsλ , t!sλ F)
F ,λ,s : Q~ ⊗S~ HA (Tλ , tλ F) −
q q
defined by the composition

(ıλ )! ΞV,λ
Q~ ⊗ HA (Tλ , t!λ F) ←−−− Q~ ⊗ HA (i!λ F) −−s−→ Q~ ⊗ sHA (i!sλ F)
q q q
∼ ∼
(ısλ )! s⊗1
−−−→ Q~ ⊗S~ sHA (Tsλ , t!sλ F) −−→ Q~ ⊗S~ HA (Tsλ , t!sλ F).
q q
∼ ∼
(Here the first and third arrows are invertible by the localization theorem in equivariant coho-
mology.) Using the notation of §8.3, it follows from Theorems 2.2.4 and 2.5.5 that the following
diagram commutes:
′ DWgeom
F ,λ,s
Q~ ⊗S~ HA (Tλ , t!λ F) / Q~ ⊗S H q (Tsλ , t! F)
q
~ A sλ
≀ (2.2.2) (2.2.2) ≀
′ DWalg
V,λ,s
Q ~ ⊗ Vλ / Q~ ⊗ Vsλ
53
As explained in the proof of Lemma 8.3.1, for any x ∈ Q~ ⊗ Vλ we have

′
DWalg alg
V,λ,s (x) = −α̌ −α̌ + ~ · · · −α̌ + (λ(α̌) − 1)~ · DWV,λ,s (x)
in Q~ ⊗ Vsλ . On the other hand, it follows from Lemma 8.4.1(2) that if we identify the Q~ -modules
Q~ ⊗S~ HA (Tλ , t!λ F) and Q~ ⊗S~ HA (Wλ ∩ Tλ , s!λ F) with Q~ ⊗ Vλ , and Q~ ⊗S~ HA (Tsλ , t!sλ F) and
q q q
!
Q~ ⊗S~ HA (Wsλ ∩ Tsλ , ssλ F) with Q~ ⊗ Vsλ by the isomorphisms of Lemma 2.2.1(2) and Lemma
q
8.4.1(1), we have for any x ∈ Q~ ⊗ Vλ

′
DWgeom geom
V,λ,s (x) = −α̌ −α̌ + ~ · · · −α̌ + (λ(α̌) − 1)~ · DWV,λ,s (x)
in Q~ ⊗ Vsλ . The proposition follows.
8.6. Brylinski–Kostant filtration. We have a natural isomorphism of algebras S(g/u) = C[(g/u)∗ ]

so that for any ϕ ∈ t∗ there is a natural surjective algebra morphism
S(g/u) → C[ϕ + (g/b)∗ ] (8.6.1)
associated with the inclusion ϕ + (g/b)∗ ֒→ (g/u)∗ . (Here, as usual we consider ϕ as a linear form
on g trivial on u ⊕ u− .)
Lemma 8.6.2. Let V in Rep(G) and λ ∈ X. For any ϕ ∈ t∗rs , the morphism
B B
Cϕ ⊗S(t) V ⊗ S(g/u) ⊗ C−λ → V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ
induced by (8.6.1) is an isomorphism.
Proof. We have
B B
Cϕ ⊗S(t) V ⊗ S(g/u) ⊗ C−λ ∼
= Cϕ ⊗C[t∗rs ] C[t∗rs ] ⊗S(t) V ⊗ S(g/u) ⊗ C−λ
B (7.2.2) B
∼
= Cϕ ⊗C[t∗rs ] V ⊗ C[t∗rs ] ⊗S(t) S(g/u) ⊗ C−λ ∼
= Cϕ ⊗C[t∗rs ] V ⊗ C[t∗rs ] ⊗ IndB
T (−λ)
B
∼
= V ⊗ C−λ ⊗ C[B/T ] .
On the other hand we have ϕ + (g/b)∗ = B · ϕ ∼ = B/T as a B-variety, hence there is a natural
isomorphism of B-modules C[ϕ + (g/b)∗ ] ∼ = C[B/T ]. This implies our claim.
Let us now fix λ ∈ X and F in PervǦ(O) (Gr). To simply notation we set V := S(F).
Recall that for ϕ ∈ t∗ we have defined the filtered vector space Hϕ (i!λ F) in §2.6. On the other
hand, the algebra C[ϕ + (g/b)∗ ] is also naturally filtered: the filtration is defined so that for any
ψ ∈ ϕ + (g/b)∗ , the algebra isomorphism C[ϕ + (g/b)∗ ] ∼ = C[(g/b)∗ ] induced by the isomorphism
∼
→ ϕ + (g/b)∗ ,
(g/b)∗ − f 7→ ψ + f
is an isomorphisms of filtered algebras, where the filtration on C[(g/b)∗ ] is the one induced by
the grading such that the vectors in g/b ⊂ C[(g/b)∗ ] are in degree 2. Hence we have an induced
B
filtration on the vector space V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ .
Combining Theorem 2.3.1 and Lemma 8.6.2, one obtains the following.
Corollary 8.6.3. For any ϕ ∈ t∗rs , there exists a canonical isomorphism
B
Hϕ (i!λ F) ∼
= V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ .
This isomorphism is an isomorphism of filtered vector spaces, where the filtration on the right-hand side is
shifted by λ(2ρ̌), i.e. for any j ∈ Z it restricts to an isomorphism
B
Fj Hϕ (i!λ F) ∼ = Fj−λ(2ρ̌) V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ .
54
Now we are in a position to give a proof of Proposition 2.6.3. In fact, this proposition is a
consequence of Corollary 8.6.3 and the following result (which is essentially proved in [Bry]; we
reproduce the proof for the reader’s convenience).
Proposition 8.6.4. Let e ∈ u be a sum of non-zero simple root vectors. If ϕ ∈ t∗ r{0} satisfies (ad∗ e)(ϕ) =
0, then evaluation at ϕ induces an isomorphism
B ∼
V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ − → Vλ . (8.6.5)
For any j ∈ Z, this isomorphism restricts to an isomorphism
B B ∼ BK
F2j V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ = F2j+1 V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ −
→ Fj (Vλ ).
Proof. Let us denote morphism (8.6.5) by Λλ . The fact that Λλ is an isomorphism follows from
the proof of Lemma 8.6.2. What remains is to prove that this isomorphism is compatible with
filtrations. Note that Λλ is the restriction to the λ-weight spaces of the isomorphism of T -modules
∼
Λ : (V ⊗ C[ϕ + (g/b)∗ ])U −
→V
given again by evaluation at ϕ. By construction the filtration on the left-hand side of (8.6.5) has
jumps only in even degrees, which justifies the equality F2j = F2j+1 . For j ∈ Z we set

∗
Ffib
j (V λ ) := Λλ F 2j (V ⊗ C[ϕ + (g/b) ] ⊗ C −λ )B
.
Hence what we have to check is that Ffib q coincides with FBK

q .
It will be convenient to work in g rather than g∗ . Hence we choose a G-equivariant isomorphism
= g∗ ; it restricts to a B-equivariant isomorphism (g/u)∗ ∼
g∼ = b. Let h ∈ b be the image of ϕ, so that
∗ ∼
we obtain an identification ϕ + (g/b) = h + u. Consider the isomorphisms
B B
V ⊗ C[ϕ + (g/b)∗ ] ⊗ C−λ ∼
= V ⊗ C[h + u] ⊗ C−λ ∼
= HomB V ∗ ⊗ Cλ , C[h + u] .
Under this identification, Λλ sends a morphism f : V ∗ ⊗ Cλ → C[h + u] to the linear form on (Vλ )∗
given by ψ 7→ f (ψ ⊗ 1)(h). In fact the image of Λλ (f ) in V is Λ(f ), which can be described as the
linear form on V ∗ given by ψ 7→ f (ψ ⊗ 1)(h).
Fix some f ∈ HomB V ∗ ⊗ Cλ , C[h + n] , and let v = Λλ (f ) be its image in Vλ (or in V ). By
definition, v is in Ffib ∗
j (Vλ ) iff for any x ∈ u and any ψ ∈ V , the polynomial in t given by
f (ψ ⊗ 1)(h + tx)
has degree ≤ j. Fix ψ ∈ V ∗ , and choose x ∈ u such that this degree is maximal. By density of
regular nilpotent elements, one can assume that x is regular. Then there exists b ∈ B such that
x = b · e. (Indeed, it is well known that there exists g ∈ G such that x = g · e. Then x ∈ g · b; as
a regular nilpotent element is contained in only one Borel subalgebra we deduce that b = g · b,
which implies that g ∈ B.) And we have
f (ψ ⊗ 1)(h + tx) = f (ψ ⊗ 1)(b · (b−1 · h + te)) = f (b−1 · (ψ ⊗ 1))(b−1 · h + te).
Now, by [Bry, Lemma 4.2], the degree of the polynomial on the right-hand side is the same as the
degree of the polynomial f (b−1 · (ψ ⊗ 1))(h + te). Moreover, we have
f (b−1 · (ψ ⊗ 1))(h + te) = f (b−1 · (ψ ⊗ 1))(exp(te) · h) = f (exp(−te)b−1 · (ψ ⊗ 1))(h).
Hence, finally, v is in Ffib ∗
j (Vλ ) iff for any ψ ∈ V , the polynomial
f (exp(−te) · (ψ ⊗ 1))(h)
has degree ≤ j.
Now, the linear form on V ∗ given by ψ 7→ f (exp(−te) · (ψ ⊗ 1))(h) is exp(te) · v. Hence v is in
Fj (Vλ ) iff ej+1 · v = 0, i.e. iff v ∈ FBK
fib
j (Vλ ).
55
8.7. Ordinary cohomology of cofibers. As a consequence of Corollary 2.4.5 we can prove the
following result, which is equivalent to [ABG, Theorem 8.5.2]. (More precisely, [ABG, Theorem
8.5.2] also contains a claim about compatibility with convolution, which can be deduced from
Proposition 8.1.5.)
Proposition 8.7.1. For F in PervǦ(O) (Gr) and λ ∈ X+ , there exists a canonical isomorphism of graded
vector spaces
G
H (i! F) ∼= S(F) ⊗ Γ N e , O e (λ) hλ(2ρ̌)i.
q
λ N
Proof. By the arguments in the proof of Lemma 2.2.1 the forgetful functor induces an isomorphism
∼
C ⊗S(t) HŤ (i!λ F) −
→ H(i!λ F)
q
(where, in the left-hand side, S(t) acts trivially on C). On the other hand, as observed in the proof
of Lemma 3.6.2, restriction induces an isomorphism
∼
C ⊗S(t) Γ eg, Oeg (λ) −→Γ N e , O e (λ) .
N
Then the result follows from Corollary 2.4.5, using the fact that the functor of G-fixed points is
exact.
Remark 8.7.2. One can obtain in a similar way a description of HC× (i!λ F) (in case λ is dominant) in
q
terms of asymptotic D-modules on B. We omit the details.
8.8. Equivariant cohomology of spherical perverse sheaves. In this subsection we explain the
relation between our description of equivariant cohomology of cofibers of spherical perverse
sheaves on GrǦ and the description of the full equivariant cohomology of these perverse sheaves
given in [BF]. Details (and generalization to all reductive groups) will be discussed in a future
publication.
From now on, for simplicity we assume that Ǧ is quasi-simple and simply connected, so that G
∼
is simple (of adjoint type). The Killing form determines an isomorphism of G-modules κ : g − → g∗ .
Let us choose an element e ∈ u which is a sum of non-zero simple root vectors. Then e is regular
nilpotent, and it can be completed to an sl2 -triple (e, ρ̌, f ). (Note that f is uniquely determined by
e, and that the different choices for e are all conjugate under the adjoint action of T .) We consider
the Kostant slice
Σe := κ(e + gf ) ⊂ g∗ .
It is well known that Σe is included in g∗r , and that the (co-)adjoint quotient g∗ → g∗ /G ∼ = t∗ /W
∼ ∗ e e the inverse image of Σe under the
restricts to an isomorphism Σe − → t /W . We denote by Σ
∗ e
g → g . Then Σe ⊂ e
projection π : e e
gr , and the morphism Σe → Σe is W -equivariant, where W acts
e
on Σe by the restriction of the action on e gr , and trivially on Σe . It is also known that the natural
morphism e g → g∗ ×t∗ /W t∗ restricts to an isomorphism of algebraic varieties e gr → g∗r ×t∗ /W t∗
(see [G2, Remark 4.2.4(i)]). We deduce that the morphism δ : e ∗
g → t restricts to an isomorphism
e ∼ ∗
Σe − →t .
Lemma 8.8.1. The projection T ∗ X → e g admits a canonical section ωe : Σe e ֒→ T ∗ X over Σ
e e . This section
is independent of the choice of e in the following sense: if t ∈ T then the following diagram commutes:
t·(−)
e e
Σ e t·e
/Σ
_ ∼ _
ωe ωt·e
t·(−)
T ∗X ∼
/ T ∗ X.
56
The morphism ωe is also W -equivariant, if W acts on (T ∗ X)r via the action of §5.5, for the choice of simple
root vectors given by the components of e on each gα .
Proof. Given a Borel subalgebra b0 ⊂ g with unipotent radical u0 , the “universal Cartan subal-
gebra” b0 /u0 acts on (u0 /[u0 , u0 ])∗ ; under this action (u0 /[u0 , u0 ])∗ decomposes as a direct sum
of 1-dimensional eigenspaces, and the eigenvalues can be naturally identified with the nega-
tive simple roots under the canonical isomorphism t∗ ∼ = (b0 /u0 )∗ . Following [GK] we denote
∗
by O(b0 ) ⊂ (u0 /[u0 , u0 ]) the subset of vectors whose component in each eigenspace is non-zero.
We also set B e := {(b0 , x) | b0 ∈ B, x ∈ O(b0 )}. Then our choice of simple root vectors defines a
∼ e
G-equivariant isomorphism X − →B sending U/U to (b, ψ), where ψ := κ(f )|u ∈ O(b).
To define the section we need, given some η ∈ Σe and b0 ∈ B with unipotent radical u0 such that
η|u0 = 0, to define a lift of b0 to X, or equivalently to B. e However, by [GK] the Borel subalgebras
b0 and b are in general relative position. Hence the Killing form induces a non-degenerate pairing
∼
between u and u0 , hence an isomorphism u − → u∗0 . If we denote by η0 ∈ u∗0 the image of e under
this isomorphism, then one can check that η0 ∈ O(b0 ), hence the pair (b0 , η0 ) provides the desired
e
lifting of b0 to B.
From now on we fix a choice for e. Let use denote by

i : GrŤ ֒→ GrǦ and t : GrB̌ − ֒→ GrǦ
F F
F We have GrŤ = λ∈XF
the inclusions. {λ}, and i identifies with λ∈X iλ . Similarly one can identify
GrB̌ − with λ∈X Tλ , and then t with λ∈X tλ . Let F be in PervǦ(O) (GrǦ ) and consider the diagram

HŤ (GrB̌ − , t! F) o ? _ H q (Gr , i! F) / H q (Gr , F).
q
Ť Ť Ť Ǧ
Here the left-hand morphism is induced by the inclusion GrŤ ֒→ GrB̌ − , and the right-hand mor-
phism is induced by i as in §2.2. By the localization theorem in equivariant cohomology, both
morphisms become isomorphisms when we extend scalars from S(t) to its fraction field Q, which
provides a canonical isomorphism
Q ⊗S(t) HŤ (GrǦ , F) ∼
= Q ⊗S(t) HŤ (GrŤ , t! F). (8.8.2)
q q
By Lemma 2.2.1 there exists a canonical isomorphism of S(t)-modules HŤ (t! F) ∼

= S(F) ⊗ S(t);
q
hence we obtain a canonical isomorphism
∼
Q ⊗ HŤ (GrǦ , F) −
→ Q ⊗ S(F). (8.8.3)
q
On the other hand, let V in Rep(G), and consider the diagram

G
V ⊗ S(t) o ? _ V ⊗ C[T ∗ X] e e ].
/ V ⊗ C[Σ
Here the left-hand morphism is induced by inverse image with respect to the natural morphism
G × t∗ → T ∗ X, and the right-hand morphism is induced by inverse image with respect to the mor-
e e ֒→ T ∗ X of Lemma 8.8.1. One can check that both morphisms in this diagram become
phism ωe : Σ
isomorphisms when we extend scalars from S(t) to Q, so that we obtain a canonical isomorphism
∼
Q ⊗S(t) V ⊗ C[Σ e e] −→ Q ⊗ V. (8.8.4)
As far as we understand, the isomorphism like our (8.8.2) which is implicitly used in [BF] (see
in particular the proof of Theorem 6 in loc. cit.) is the one we are defining here. With this interpre-
tation, the “quasi-classical limit” (or “classical analogue”) of [BF, Theorem 6] says the following.
57
Proposition 8.8.5. Let F be in Perv Ǧ(O) (GrǦ ).
The image of HŤ (GrǦ , F) in Q ⊗ S(F) under isomorphism (8.8.3) coincides with the image of S(F) ⊗
q
e e ] under isomorphism (8.8.4), which provides a canonical isomorphism
C[Σ
HŤ (GrǦ , F) ∼ e e ].
= S(F) ⊗ C[Σ (8.8.6)
q
Combined with Lemma 2.2.1 and Corollary 2.4.5, this implies that we have the following com-
mutative diagram:
HŤ (GrB̌ − , t! F) o
q ? _ H q (Gr , i! F) / H q (Gr , F)
Ť Ť Ť Ǧ
Lem. 2.2.1 ≀ Cor. 2.4.5 ≀ (8.8.6) ≀

S(F) ⊗ S(t) o ? _ S(F) ⊗ C[T ∗ X] G e e ].
/ S(F) ⊗ C[Σ
Isomorphism (8.8.6) is W -equivariant, where the W -action on the left-hand side is defined in a
way similar to the definition of isomorphisms ΞF ,λ
w in §2.5, and the action on the right-hand side
is induced by the W -action on Σe e . Hence taking fixed points we obtain the following result, also
proved in [BF] (see in particular [loc. cit., Lemma 9]).
Corollary 8.8.7. There exists a canonical isomorphism HǦ(O) (GrǦ , F) ∼
= S(F) ⊗ C[Σe ].
q
A PPENDIX A. C OMPUTATIONS IN RANK ONE
In §§A.1–A.5 we use the notation of Sections 3–5, assuming in addition that G has semisimple
rank one. We let α be the unique simple root, and set s := sα , e := eα , f := fα .
A.1. Asymptotic universal Verma modules. For any ν ∈ X+ , we denote by V ν the correspond-
ing simple G-module. We choose a basis (vνν , vν−α ν ν
, · · · , vν−ν(α̌)α ) of V ν such that the following
formulas are satisfied for k = 0, · · · , ν(α̌):
ν
ν ν ν
α̌ · vν−kα = ν(α̌) − 2k vν−kα , e · vν−kα = kvν−(k−1)α . (A.1.1)
Such a basis exists and is unique up to a constant. We also set vλν = 0 if λ is not a weight of V ν .
The following commutation relation in the asymptotic enveloping algebra U~ (g) is easily checked
by induction:
f k e = ef k − k~ · α̌f k−1 − k(k − 1)~2 · f k−1 . (A.1.2)
Lemma A.1.3. Let ν ∈ X+ .
B
(1) If λ is not a weight of V ν , then V ν ⊗ M(λ) = 0.
B
(2) If k ∈ {0, · · · , ν(α̌)}, the S~ -module V ν ⊗ M(ν − kα) is free of rank one, and generated by

ν ν k k ν
xν−kα := vν ⊗ 1 ⊗ f − vν−α ⊗ α̌ + ν(α̌) − k ~ ⊗ f k−1
1

k ν

+ · vν−2α ⊗ α̌ + ν(α̌) − k ~ α̌ + ν(α̌) − k − 1 ~ ⊗ f k−2 + · · ·
2

k k ν

+ (−1) · vν−kα ⊗ α̌ + ν(α̌) − k ~ · · · α̌ + (ν(α̌) − 2k + 1)~ ⊗ 1.
k
58
Proof. (1) follows from the injectivity of κalg
V ν ,λ , see Lemma 4.6.4.
(2) We have to decide when an element of the form
ν
vν−kα ⊗ Pk (h, ~) ⊗ 1 + · · · + vνν ⊗ P0 (h, ~) ⊗ f k
is annihilated by e ∈ b ⊂ U (b). (Here the action of U (b) is the differential of the B-action.)
However, the image by e of such an element is given by
k−1
X
ν ν 1 ν 1
e · vν−kα ⊗ Pk ⊗ 1 + e · vν−iα ⊗ Pi ⊗ f k−i − vν−iα ⊗ Pi ⊗ f k−i e − vνν ⊗ P0 ⊗ f k e,
~ ~
i=1
i.e. (using (A.1.1) and (A.1.2)) by
k−1
X
ν ν

kvν−(k−1)α ⊗ Pk ⊗ 1 + ivν−(i−1)α ⊗ Pi ⊗ f k−i + (k − i)vν−iα
ν
⊗ Pi α̌ + (ν(α̌) − k − i)~ ⊗ f k−i−1
i=1

+ kvνν ⊗ P0 α̌ + (ν(α̌) − k)~ ⊗ f k−1 .
Hence the fact that this element is zero amounts to the following equations:

k · Pk = − α̌ + (ν(α̌) − 2k + 1)~ · Pk−1 ,
..
.
i · Pi = −(k − i + 1) α̌ + (ν(α̌) − k − i + 1)~ · Pi−1 ,
..
.
P1 = −k α̌ + (ν(α̌) − k)~ · P0 .
The result follows.
If λ ∈ X is not a weight of V ν , we set xνλ = 0. As an immediate consequence of Lemma A.1.3 we
obtain the following result.
Corollary A.1.4. For ν ∈ X+ and k ∈ {0, · · · , ν(α̌)}, the image of the morphism
B
κalg ν
V ν ,ν−kα : V ⊗ M(ν − kα)
ν
→ Vν−kα ⊗ S~ ∼
= S~

is generated by α̌ + ν(α̌) − k ~ · · · α̌ + ν(α̌) − 2k + 1 ~ .
A.2. Operators Φ. In this subsection we consider the constructions of Section 3, with the choice
of root vector f . In particular, for ν ∈ X+ and λ ∈ X, we have an isomorphism
ν G ∼ s ν (sλ) G
ΦVs ,λ : V ν ⊗ (λ)D~ (X)λ − → V ⊗ D~ (X)sλ .
Recall that, by the first isomorphism in Lemma 2.4.1, xνλ defines an element
G
yλν ∈ V ν ⊗ (λ)D~ (X)λ .
Lemma A.2.1. For ν ∈ X+ and λ ∈ X we have
ν ,λ
ΦVs (yλν ) = ysλ
ν
.
Proof. If λ is not a weight of V ν , then sλ is not either. Hence yλν = 0 = ysλ ν , and the result is clear.
V ν ,λ
Now assume λ = ν −kα for some k ∈ {0, · · · , ν(α̌)}. As Φs is an isomorphism of S~ -modules,
V ν ,λ ν ×
we know that Φs (yλ ) = c · ysλ for some c ∈ C , and we have to prove that c = 1. Let zλν be the
ν
image of yλν under the composition

G
V ν ⊗ (λ)D~ (X)λ ֒→ V ν ⊗ D~ (X) → V ν ⊗ A(X),
59
and similarly for zsλν . Then by definition (see §3.5) we have (id ν ⊗ F )(z ν ) = c · z ν .
V α λ sλ
Since, in the notation of §3.2, one can take as Gsc a product of SL(2, C) and a torus, it suffices
to prove the claim in the case G = SL(2, C), with B the subgroup of upper-triangular matrices,
T the subgroup of diagonal matrices, and f = ( 01 00 ). Let v1 , v2 be the natural basis of C2 and
∼
η1 , η2 the dual basis. Then the morphism gU 7→ g · v1 induces an isomorphism G/U − → C2 r {0},
and Vα = C2 , with the symplectic form ω defined by ω(v1 , v2 ) = 1. Hence we have an canonical
isomorphism A(X) ∼ = C[T ∗ (C2 )] = C[C2 × (C2 )∗ ], and the automorphism Fα can be computed
using Example B.2.1 below: it is defined by
η1 7→ v2 , η2 7→ −v1 , v1 7→ −η2 , v2 7→ η1 .
Consider the morphism ev : A(X) → C given by evaluation of functions at (v1 , η2 ) ∈ T ∗ (C2 ).
The formulas above imply that ev ◦ Fα = ev. Hence to conclude we only have to check that
ν
(idV ν ⊗ ev)(zλν ) = (idV ν ⊗ ev)(zsλ ).
However we have
(idV ν ⊗ ev)(zλν ) = vνν ⊗ 1 = (idV ν ⊗ ev)(zsλ
ν
),
which finishes the proof.
A.3. Operators Θ. In this subsection we consider the constructions of Section 4, with the choice
fα := f . In particular, for ν ∈ X+ and λ ∈ X, we have a morphism
ν
ΘVs ,λ : Hom(S~ ,U~ (g)) M(0), V ν ⊗ M(λ) → s Hom(S~ ,U~ (g)) M(0), V ν ⊗ M(sλ) .
By Lemma 4.3.1, xνλ defines a morphism of (S~ , U~ (g))-bimodules
ϕνλ : M(0) → V ν ⊗ M(λ),

which is a generator of the S~ -module Hom(S~ ,U~ (g)) M(0), V ν ⊗ M(λ) .
ν ,λ
ΘVs (ϕνλ ) = ϕνsλ .
Proof. If λ is not a weight of V ν , then sλ is not either. Hence ϕνλ = 0 = ϕνsλ , and the result is clear.
Now assume λ = ν − kα for some k ∈ {0, · · · , ν(α̌)}. Then sλ = ν − (ν(α̌) − k)α. By Lemma
ν
A.1.3 we know that ΘsV ,ν−kα(ϕνν−kα ) is a multiple of ϕνν−(ν(α̌)−k)α (by an element of S~ ). Hence
to prove the lemma it is enough to check that the coefficient of vνν ⊗ 1 ⊗ f ν(α̌)−k in the element of
B ν
V ν ⊗ M(ν − (ν(α̌) − k)α) corresponding to ΘsV ,ν−kα (ϕνν−kα ) is the same as the coefficient of
vνν ⊗ 1 ⊗ f ν(α̌)−k in xνν−(ν(α̌)−k)α , i.e. 1. However this coefficient can be computed using formula
(4.5.4), and the result is 1 as expected. This concludes the proof.
Corollary A.3.2. For any V in Rep(G) and λ ∈ X we have
s

ΘV,sλ
s ◦ ΘV,λ
s = id

as endomorphisms of Hom(S~ ,U~ (g)) M(0), V ⊗ M(λ) . In particular, each ΘV,λ
s is an isomorphism.
Proof. By complete reducibility it is enough to prove the result when V = V ν for some ν ∈ X+ . In
this case, by Lemma A.3.1 we have
s ν ν
ΘVs ,sλ ◦ ΘVs ,λ (ϕνλ ) = ϕνλ ,

which proves the claim since ϕνλ is a generator of Hom(S~ ,U~ (g)) M(0), V ν ⊗ M(λ) over S~ by
Lemma A.1.3.
60
A.4. Comparison of Θ and Φ.
Lemma A.4.1. For any V in Rep(G) and λ ∈ X, the following diagram commutes:
G (4.8.1)
V ⊗ (λ)D~ (X)λ ∼
/ Hom(S ,U (g)) M(0), V ⊗ M(λ)
~ ~
ΦV,λ
s ΘV,λ
s

s
G (4.8.1)
V ⊗ (sλ)D~ (X)sλ ∼
/ s Hom
(S~ ,U~ (g)) M(0), V ⊗ M(sλ) .
Proof. By complete reducibility it is enough to prove the lemma when V = V ν for some ν ∈ X+ .
In this case it follows from Lemma A.2.1 and Lemma A.3.1.
A.5. Operators σ. If ν ∈ X+ and λ ∈ X, we denote by

G
zλν ∈ V ν ⊗ Γ(e
g, Oeg (λ))
the image of yλν under the natural morphism
G G
V ν ⊗ (λ)D~ (X)λ → V ν ⊗ Γ(e
g, Oeg (λ))
sending ~ to 0. These elements can be naturally identified with the ones denoted similarly in the
proof of Lemma A.2.1. By Lemma 3.5.2 and Lemma A.1.3, zλν is a generator of the S(t)-module
G
g, Oeg (λ)) , and zλν 6= 0 iff λ is a weight of V ν .
V ν ⊗ Γ(e
ν ,λ
σsV ν
(zλν ) = zsλ .
Proof. The statement is clear if λ is not a weight of V ν . Now assume that λ = ν − kα for some
ν ν
k ∈ {0, · · · , ν(α̌)}. As σsV ,λ is an isomorphism of S(t)-modules we must have σsV ,λ (zλν ) = c · zsλ
ν
×
for some c ∈ C . Now recall the commutative diagram (5.2.4). We have
(idV ν ⊗ evλη0 )(zλν ) = vνν = (idV ν ⊗ evsλ ν
η0 )(zsλ ).
This proves that c = 1, and finishes the proof.
Corollary A.5.2. Let V in Rep(G) and λ ∈ X. The following diagram commutes, where vertical maps are
the natural morphisms sending ~ to 0:
G ΦV,λ
s G
V ⊗ (λ)D~ (X)λ / V ⊗ (sλ)D~ (X)sλ

G σsV,λ G
V ⊗ Γ(e
g, Oeg (λ)) / V ⊗ Γ(e
g, Oeg (sλ))
Proof. By complete reducibility it is enough to prove the claim when V = V ν for some ν ∈ X+ . In
this case it follows by comparing Lemma A.2.1 and Lemma A.5.1.
A.6. Satake equivalence. From now on we use the notation of Sections 6–8, assuming in addition
that Ǧ has semisimple rank one. We denote by α the unique positive coroot of Ǧ. Note that α̌ and
~ can be considered either as characters of A = Ť × C× or as elements of a∗ = t ⊕ C.
To simplify the statements of the next results we introduce the following notation. For λ ∈ X
and k ≥ 1 we define the A-module
Vkλ := C−α̌−(λ(α̌)+1)~ ⊕ C−α̌−(λ(α̌)+2)~ ⊕ · · · ⊕ C−α̌−(λ(α̌)+k)~ .
61
λ .
We also set V0λ = {0}. Note that dim(Vkλ ) = k, and that there are natural inclusions Vkλ ⊂ Vk+1
Lemma A.6.1. Let λ ∈ X.
(1) Assume λ(α̌) ≥ 0. For ν ∈ X+ we have isomorphisms of A-varieties


{λ} if ν = λ;
ν ∼ λ λ
Tλ ∩ GrǦ = Vk r Vk−1 if ν = λ + kα for some k ∈ Z>0 ;


∅ otherwise.
(2) Assume λ(α̌) < 0. For ν ∈ X+ we have isomorphisms of A-varieties

λ
V−λ(α̌)
 if ν = λ + (−λ(α̌))α;
ν ∼ λ λ
Tλ ∩ GrǦ = Vk r Vk−1 if ν = λ + kα for some k ∈ Z>−λ(α̌) ;


∅ otherwise.
Proof. Each connected component of GrǦ is isomorphic to a connected component of GrPGL(2,C) ,
and the action of ker(α̌) = Z(Ǧ) ⊂ Ť is trivial. Hence it is enough to prove the isomorphism when
Ǧ = PGL(2, C). In this case we can identify X with Z through µ 7→ µ(α̌), so that ν and λ can be
considered as integers.
First case: even weights. Write λ = 2ℓ with ℓ ∈ Z. Then λ is the class of the matrix
ℓ
z 0
,
0 z −ℓ
and Tλ is given by the classes of matrices of the form

zℓ 0
M (Q) =
Q(z) z −ℓ
where Q(z) ∈ z −ℓ−1 C[z −1 ]. If Q(z) = 0, then this point is in Grλ . Otherwise, write Q(z) =
az −m + · · · , where a 6= 0, m > ℓ, and “· · · ” means terms of degree between −m + 1 and −ℓ − 1.
Assume first that m + ℓ ≥ 0. (This condition is always satisfied when ℓ ≥ 0. Note also that it
implies m ≥ 0.) Then we have
m
zℓ 0 z m−ℓ R(z) 1 z m+ℓ R(z) z 0
· = ·
Q(z) z −ℓ −z m Q(z) 0 0 1 0 z −m
where R(z) ∈ O is the inverse to z m Q(z). This equality implies that M (Q) ∈ Gr2m .
If ℓ < 0, then we also have to consider the case ℓ < m < −ℓ. However, in this case M (Q) is in
λ+(−λ(α̌))α
Ǧ(O) · λ = GrǦ . This settles the first case.
Second case: odd weights. Write λ = 2ℓ + 1 with ℓ ∈ Z. Then λ is the class of the matrix
ℓ+1
z 0
,
0 z −ℓ
and Tλ is given by the classes of matrices of the form
ℓ+1
z 0
N (Q) =
Q(z) z −ℓ
where Q(z) ∈ x−ℓ−1 C[z −1 ]. If Q(z) = 0, then this point is in Grλ . Otherwise, write as above
Q(z) = az −m + · · · , where a 6= 0 and m > ℓ. We have the following equality:
ℓ+1 m+1
z 0 z m−ℓ R(z) 1 z ℓ+m+1 R(z) z 0
· = ·
Q(z) z −ℓ −z m Q(z) 0 0 1 0 z −m
62
where as above R(z) ∈ O is the inverse to z m Q(z). This equality implies that N (Q) ∈ Gr2m+1 if
ℓ + m + 1 ≥ 0.
If ℓ < 0 and ℓ + m + 1 < 0, then N (Q) is in Ǧ(O) · λ = Grλ+(−λ(α̌))α . This settles the second case,
and finishes the proof.
The following result is a direct consquence of Lemma A.6.1. It is stated without proof in [BF,
BrF].
Corollary A.6.2. Let ν ∈ X+ .
(1) If λ ∈ X is not of the form ν − kα for some k ∈ {0, · · · , ν(α̌)}, then Tλ ∩ GrνǦ = ∅.
(2) If k ∈ {0, · · · , ν(α̌)} and λ = ν − kα, then there exists an isomorphism of A-varieties
Tλ ∩ Grν ∼ = V ν−kα
Ǧ k
sending λ to 0.
By the constructions of §6.1 we have a dual group G (which is also of semisimple rank 1) and a
∗ + ν
maximal torus T ⊂ G such that X = X (T ). If ν ∈ X , we set ICν := IC GrǦ , CGrν .
Ǧ
Corollary A.6.3. For ν ∈ X+ and k ∈ {0, · · · , ν(α̌)}, the image of the morphism

κtop
q ! ∼
ICν ,ν−kα : HA (iν−kα ICν ) → SǦ (ICν ) ν−kα ⊗ S~ = S~

is generated by α̌ + ν(α̌) − k ~ · · · α̌ + ν(α̌) − 2k + 1 ~ .
Proof. It is well known that GrνǦ is rationally smooth. This implies that
ICν ∼= C ν [ν(α̌)] ∼
GrǦ = D ν [−ν(α̌)].
GrǦ
We deduce isomorphisms
H (i! ICν ) ∼
= HAq −ν(α̌) ({λ}), HA (t!ν−kα ICν ) ∼
= HAq−ν(α̌)+2k (Tν−kα ∩ GrνǦ ),
q q
A ν−kα
and the morphism HA (i!ν−kα ICν ) → HA (t!ν−kα ICν ) identifies with the morphism
q q
HAq−ν(α̌) ({λ}) → HAq−ν(α̌)+2k (Tν−kα ∩ GrνǦ )

given by proper push-forward in equivariant Borel–Moore homology. Hence we deduce the result
from the description of Tν−kα ∩ GrνǦ in Corollary A.6.2 and the considerations on equivariant
(co)homology in §6.2.
A.7. Proof of Theorem 2.2.4 for Ǧ. By semisimplicity of the category Perv Ǧ(O) (Gr), it is enough
to prove the theorem when F = ICν for some ν ∈ X+ . Then V ν := SǦ (ICν ) is the G-module
with highest weight ν. Comparing Corollary A.1.4 and Corollary A.6.3 we observe that indeed
the images of κtop alg
ICν ,λ and κV ν ,λ coincide, which implies the existence of the isomorphism ζICν ,λ .
A.8. Root vector and Mirković–Vilonen basis. We denote by e ∈ gα the vector constructed in
§6.5.
Let ν ∈ X+ and k ∈ {0, · · · , ν(α̌)}. Let V ν := SǦ (ICν ), a simple G-module with highest weight
ν. As in the proof of Corollary A.6.3, we have canonical isomorphisms
(6.1.2)
ν
= Hν(α̌)−2k (t!ν−kα ICν ) ∼
∼ = HA ν
Vν−kα 0 (Tν−kα ∩ GrǦ ).
The right-hand side is 1-dimensional, and has a canonical generator, namely the fundamental class
[Tν−kα ∩ GrνǦ ]. We denote by vν−kα
ν ν
∈ Vν−kα the vector corresponding to this generator.
63
Remark A.8.1. Beware that the basis we consider here is not the same as the one used in [BF, §5.2]
or in [BrF, §2.2], but rather the basis which is dual in the sense of Poincaré duality.
The following lemma is a special case of [Ba, Théorème 2].
Lemma A.8.2. For any ν ∈ X+ and k ∈ {0, · · · , ν(α̌)} we have
ν ν
e · vν−kα = kvν−(k−1)α .
It follows from this lemma that, for the choice of root vector e ∈ gα , the basis just constructed
satisfies conditions (A.1.1). Hence we can use the results and notation of §§A.1–A.5 for this choice.
We can now give a more precise version of Theorem 2.2.4 for Ǧ. As in the proof of Corollary
A.6.3 there exists a canonical isomorphism
H (i! ICν ) ∼
= HAq ({λ}).
q
A ν−kα −ν(α̌)
The right-hand side has a canonical generator, namely the unique element
ν(α̌)
cνν−kα ∈ HA (i!ν−kα ICν )
whose image in Hν(α̌) (i!ν−kα ICν ) ∼
= H0 ({λ}) is the fundamental class [{λ}]. Then (see §6.2) we
have
ζICν ,ν−kα (xνν−kα ) = cνν−kα . (A.8.3)
A.9. Proof of Theorem 2.5.5 for Ǧ. We have already proved in Lemma A.4.1 that the operators
Φ and Θ match under the natural isomorphisms. Hence we only have to compare them with
operators Ξ. Moreover, it is enough to prove the theorem in the case F = ICν for some ν ∈ X+ .
By construction (see §2.5) we have ΞsICν ,ν−kα (cνν−kα ) = cνν−(ν(α̌)−k)α . Hence the claim follows from
(A.8.3) and Lemma A.2.1 or Lemma A.3.1.
A PPENDIX B. F OURIER TRANSFORM FOR DIFFERENTIAL OPERATORS
In this appendix we briefly explain how to adapt some classical constructions of Fourier trans-
form for D-modules (see e.g. [BMV]) to the asymptotic setting.
B.1. Partial Fourier transform. Let X be a smooth complex algebraic variety, and let p : E → X be
a rank r algebraic vector bundle. Let p̌ : E ∗ → X be the dual vector bundle, and let E := E ×X E ∗
be the total space of the direct sum E ⊕ E ∗ , a vector bundle on X of rank 2r. The canonical pairing
of E and E ∗ gives a regular function f : E ×X E ∗ → C. We define a connection ∇ : OE → Ω1E
by ∇ = d − df . This connection is flat and makes P = (OE , ∇) a holonomic left DE -module, with
irregular singularities. Explicitly, we have P = DE /J where J ⊂ DE is the left ideal generated
by the elements ξ − ξ(f ) for all vector fields ξ on E. (Note that this ideal is the annihilator of the
function exp ◦f on E, the classical kernel for the Fourier transform.)
Now, let ǫ : E = E ×X E ∗ ֒→ E × E ∗ be the natural closed embedding and let ǫ∗ P be a
direct image of the D-module P. Thus, ǫ∗ P is a holonomic left DE×E ∗ -module supported on the
subvariety E. Write KY for the canonical bundle on a smooth variety Y and let
Q := (OE ⊠ p̌∗ KX ) ⊗OE×E ∗ ǫ∗ P.
The sheaf Q has the structure of a module of the algebra
−1

DE ⊠ p̌∗ KX ⊗OE ∗ DE ∗ ⊗OE ∗ p̌∗ KX ,
and this module has a canonical section 1Q ∈ Q that corresponds to the section 1 mod J ∈ DE /J .
Furthermore, a local computation shows that Q is a rank one free module
(with generator 1Q ) over
∗ ∗ −1
the ring DE ⊠ 1, as well as over the ring 1 ⊠ p̌ KX ⊗ DE ⊗ p̌ KX .
∗
64
Let p : E → X be the natural projection. Then p∗ Q is a rank one free module both over p∗ DE
−1
and over KX ⊗OX p̌∗ DE ∗ ⊗OX KX , with a canonical generator 1p∗ Q . Therefore, there is a uniquely
−1
determined morphism F : p∗ DE → KX ⊗ p̌∗ DE ∗ ⊗ KX such that one has u · 1p∗ Q = F (u) · 1p∗ Q .
It is immediate to check that this morphism is an anti-isomorphism of rings, i.e. it induces a ring
isomomorphism
∼ −1 op
p∗ DE − → (KX ⊗OX p̌∗ DE ∗ ⊗OX KX ) . (B.1.1)
∗
On the other hand, we have KE ∗ = p̌ (det(E) ⊗OX KX ), where det(E) denotes the sheaf of sec-
tions of the line bundle ∧r E on X. Hence, using the well-known isomorphism Dop ∼
E ∗ = KE ⊗OE ∗
∗
−1
DE ∗ ⊗OE ∗ KE ∗ , we compute:

−1 op ∼
KX ⊗OX p̌∗ DE ∗ ⊗OX KX = KX −1
⊗OX p̌∗ (Dop
E ∗ ) ⊗OX KX

∼
= K ⊗O p̌∗ KE ∗ ⊗O ∗ DE ∗ ⊗O ∗ K−1∗ ⊗O KX
−1
X X E E E X
∼
= det(E) ⊗OX p̌∗ DE ∗ ⊗OX det(E)−1 .
Thus, from (B.1.1) we deduce a canonical isomorphism of sheaves of algebras on X, called
Fourier isomorphism:
∼
→ det(E) ⊗OX (p̌∗ DE ∗ ) ⊗OX det(E)−1 .
F : p∗ DE −
Now, in a Rees algebra setting, we define J~ to be the left ideal of D~,E generated by the elements
ξ−ξ(f ) for all vector fields ξ on E, viewed as degree 2 homogeneous elements of the graded algebra
D~,E . (Here we use the notational conventions of §2.4. Note that the ideal J~ is the annihilator
of the “function” exp( 1~ f ), considered as an element of some completion of OX [~, ~−1 ].) We put
P~ := D~,E /J~ , a D~,E -module. Note that the ideal J~ is not homogeneous so the module P~ has
no natural grading.
Under the specialization ~ = 0, we have D~,E /(~) = (pE )∗ OT ∗ E , where pE : T ∗ E → E is the
cotangent bundle. The differential of the function f gives a section df : E → T ∗ E. The image of
this section is a smooth closed lagrangian subvariety Λ ⊂ T ∗ E, so the sheaf (pE )∗ OΛ has a natural
structure of a (pE )∗ OT ∗ E -module. Then, it follows from definitions that the projection D~,E ։ P~
induces an isomorphism of (pE )∗ OT ∗ E -modules:
P~ /(~) ∼= (pE )∗ OΛ .
We also consider the cotangent bundle q : T ∗ (E × E ∗ ) → E × E ∗ and let T ∗ (E × E ∗ )|E denote the
total space of the restriction of the cotangent bundle to E ⊂ E × E ∗ , a closed subvariety. One has
a natural diagram
ε pr
T ∗ E × T ∗ (E ∗ ) T ∗ (E × E ∗ ) o ? _ T ∗ (E × E ∗ )|E / / T ∗E
Here, the isomorphism on the left involves a sign and the map pr on the right is a smooth mor-
phism.
The following result is easily verified by a local computation.
Lemma B.1.2. The variety Z := ε(pr−1 (Λ)) is a smooth Lagrangian subvariety of T ∗ (E × E ∗ ). Further-
more, this subvariety is the graph of an isomorphism T ∗ E →
∼
T ∗ (E ∗ ), of algebraic varieties over X.
To proceed further we observe that, for any smooth variety Y , the sheaf KY [~] has a canonical
right D~,Y -action such that a vector field ξ ∈ TY acts on KY by β 7→ −~ · Lξ β, where Lξ stands for
the Lie derivative. We write KY~ for the resulting right D~,Y -module. Then, one has a canonical
isomorphism
Dop ∼ ~ ~ −1
~,Y = KY ⊗OY [~] D~,Y ⊗OY [~] (KY ) .
Note that this isomorphism specializes at ~ = 0 to the identity map (pE )∗ OT ∗ Y → (pE )∗ OT ∗ Y .
65
Next, mimicing the corresponding constructions for D-modules, one can define a direct image
ǫ∗ P~ , a left D~,E×E ∗ -module. Further, we put Q~ := (OE ⊠ p̌∗ KX
~ )⊗
OE×E ∗ [~] ǫ∗ P~ . Then, one checks
that there is a natural isomorphism
Q~ /(~) ∼= q∗ OZ ,
of q∗ OT ∗ (E×E ∗ ) -modules. Furthermore, repeating earlier constructions, one obtains a canonical
isomorphism
∼
→ det(E)[~] ⊗OX [~] (p̌∗ D~,E ∗ ) ⊗OX [~] det(E)−1 [~].
F~ : p∗ D~,E − (B.1.3)
This is an isomorphism of sheaves of C[~]-algebras on X. This isomorphism does not respect the
natural gradings on each side in (B.1.3) unless r = 0 and it specializes, at ~ = 0, to the isomorphism
∼
p∗ OT ∗ E → p̌∗ OT ∗ (E ∗ ) that results from Lemma B.1.2.
The above isomorphism can be described locally as follows. Let U ⊂ X be an open subvariety
over which E is trivializable, and let us choose an isomorphism of vector bundles E|U ∼ = Cr × U .
Then for i = 1, · · · , r we have a function xi on E|U given by the projection on the i-th copy of C, and
the corresponding vector fields ∂xi , so that we have an isomorphism of sheaves of C[~]-algebras

p∗ D~,E ∼
= D~,U ⊗C[~] Chxi , ∂x i/[∂x , xi ] = ~ . (B.1.4)
|U i i
(Here i runs over {1, · · · , r}.) Our isomorphism E|U ∼ = Cr

× U also defines a canonical section τ of
det(E) over U , the dual section τ of det(E) , and an isomorphism (E ∗ )|U ∼
∨ −1
= Cr × U , so that we
∗
obtain functions ξ1 , · · · , ξr and vector fields ∂ξ1 , · · · , ∂ξr on (E )|U , and an isomorphism

p∗ D~,E ∗ ∼
= D~,U ⊗C[~] Chξi , ∂ξ i/[∂ξ , ξi ] = ~ . (B.1.5)
|U i i
Then using isomorphisms (B.1.4) and (B.1.5), the restriction of isomorphism (B.1.3) to U can be
described as follows: it sends any P ∈ D~,U to τ ⊗ P ⊗ τ ∨ , xi to τ ⊗ −∂ξi ⊗ τ ∨ , and ∂xi to τ ⊗ ξi ⊗ τ ∨ .
B.2. “Symplectic” partial Fourier transform. Now we assume that E is a symplectic vector bundle
with symplectic form ω over a smooth complex algebraic variety X. Then we have an isomor-
phism of vector bundles
∼
E−→ E∗, v 7→ ω(v, −)
over X, hence an induced isomorphism p∗ D~,E ∼= p̌∗ D~,E ∗ . Moreover, ω defines a trivialization of
det(E). Hence isomorphism (B.1.3) provides an automorphism of p∗ D~,E . We denote by
∼
FE : D~ (E) −
→ D~ (E)
the induced automorphism. One can easily check that FE is equivariant under the natural action
of the group of symplectic automorphisms of E, and that we have FE ◦ FE = idD~ (E) .
Example B.2.1. If X = pt, then E is simply a symplectic vector space. For instance, assume that
E = C2 = Cv1 ⊕ Cv2 , equipped with the symplectic form such that ω(v1 , v2 ) = 1. Let (η1 , η2 ) be
the basis of E ∗ dual to (v1 , v2 ). Then D~ (E) is generated by η1 , η2 (considered as functions on E)
and v1 , v2 (considered as vector fields on E), and FE is defined by
η1 7→ v2 , η2 7→ −v1 , v1 7→ −η2 , v2 7→ η1 .
The following result (which can easily be checked using local trivializations) is used in §3.4.
Lemma B.2.2. Let f ∈ C[X] be an invertible function, which we consider as a function on E via the
projection E → X. Then the automorphism of D~ (E) given by D 7→ f −1 · D · f commutes with FE .
66
I NDEX OF NOTATION
hmi, 4 Fgeom , 14 N sλ , 13
1µ , 23 Fλw , 16 n−
L , 17 Spµ , 25
Fw , 21 n+
L , 17 Spµ , 25
A nλ , 13
A, 7 G Ne , 17 T
a, 43 g∗r , 12 t, 57
AX , 20 γ, 9 O t∗rs , 42
A(X), 20 Grλ , 6 O, 6 ΘV,λ
w , 11
e
g, 10 O− , 13 θw , 31
B
e
gr , 12 ΩV,λ
w , 30 Tλ , 7
B, 10
tλ , 7
H P
C
Hϕ , 13 ΦV,λ
w , 11 U
C[a∗α−rs ], 43 ϕM , 9 U~ (k), 5
C[a∗rs ], 43
I ϕνλ , 60
C[~]hhµii, 24 V
i, 57 π, 12
cνλ , 64 V(µ), 23
ICν , 63 ̟LG , 33
ConvalgV,V ′ ,λ,µ , 46 iλ , 7 pλ , 6 vλ , 6
top vµ , 24
ConvF ,G,λ,µ, 48 ıλ , 7 ΨV,λ
w,µ , 26
ıG Vkλ , 61
L , 18
D iw Q V ν , 58
µ , 24
∆F ,λ , 13 Ind, 10 Q, 57 vλν , 58
DWalg
V,λ,w , 12 ιw Q~ , 12
µ , 23 W
DWgeom
F ,λ,w , 13
R Wλ , 13
DX , 9 K
D(X), 9 K, 6 Rep(G), 6
X
D~,X , 9 κalg RǦ , 39
V,λ , 6 Ľ X, 5, 9
D~ (X), 9 κalg RV,λ
G,L , 28 ΞF ,λ
w , 11
V,λ , 8 F ,λ
E κtop
F ,λ , 7
RV,λ
G,L , 28
ξw , 12
xνλ , 58
evµη , 32 top
κF ,λ , 9 RV,λ
G,L , 18
EV,λ
µ , 23 Y
V,λ L S yλν , 59
Eµ , 26
L(λ), 31 S, 7
F S~ , 5 Z
Fα , 16 M S~ hhλii, 5 ζF ,λ , 7
V,λ
Fα , 20 M(λ), 5 σw , 12 ζ F ,λ , 9
FBK , 13 M(µ), 24 Σe , 56 Z~ (g), 5
FǦ , 35 Mloc (µ), 24 e e , 56
Σ zλν , 61
R EFERENCES
[AHR] P. N. Achar, A. Henderson, S. Riche, Geometric Satake, Springer correspondence, and small representations II, preprint
arXiv:1205.5089 (2012).
[AR] P. N. Achar, S. Riche, Constructible sheaves on affine Grassmannians and geometry of the dual nilpotent cone, preprint
arXiv:1102.2821 (2011), to appear in Israel J. Math.
[ABG] S. Arkhipov, R. Bezrukavnikov, V. Ginzburg, Quantum groups, the loop Grassmannian, and the Springer resolution,
J. Amer. Math. Soc. 17 (2004), 595–678.
67
[ATV] M. Artin, J. Tate, M. Van den Bergh, Some algebras associated to automorphisms of elliptic curves, in The Grothendieck
Festschrift, Vol. I, 33–85, Progr. Math., 86, Birkhäuser, 1990.
[Ba] P. Baumann, Propriétés et combinatoire des bases de type canonique, mémoire d’habilitation, available on
http://tel.archives-ouvertes.fr/tel-00705204.
[BD] A. Beilinson, V. Drinfeld, Quantization of Hitchin’s integrable system and Hecke eigensheaves, preprint available at
http://www.math.uchicago.edu/∼mitya/langlands.html.
[BGG] I. Bernstein, I. Gel’fand, S. Gel’fand, Differential operators on the base affine space and a study of g-modules, in Lie
groups and their representations (Proc. Summer School, Bolyai János Math. Soc., Budapest, 1971), 21–64, Halsted, 1975.
[BBP] R. Bezrukavnikov, A. Braverman, L. Positselskii, Gluing of abelian categories and differential operators on the basic
affine space, J. Inst. Math. Jussieu 1 (2002), 543–557.
[BF] R. Bezrukavnikov, M. Finkelberg, Equivariant Satake category and Kostant–Whittaker reduction, Mosc. Math. J. 8
(2008), 39–72.
[BFM] R. Bezrukavnikov, M. Finkelberg, I. Mirković, Equivariant homology and K-theory of affine Grassmannians and Toda
lattices, Compos. Math. 141 (2005), 746–768.
[BR] R. Bezrukavnikov, S. Riche, Affine braid group actions on Springer resolutions, Ann. Sci. Éc. Norm. Supér. 45 (2012),
535–599.
[BrF] A. Braverman, M. Finkelberg, Dynamical Weyl groups and equivariant cohomology of transversal slices in affine Grass-
mannians, Math. Res. Lett. 18 (2011), 505–512.
[BK] M. Brion, S. Kumar, Frobenius splitting methods in geometry and representation theory, Progr. Math. 231, Birkhäuser,
2004.
[Bro] B. Broer, Line bundles on the cotangent bundle of the flag variety, Invent. Math. 113 (1993), 1–20.
[Bry] R. Brylinski, Limits of weight spaces, Lusztig’s q-analogs, and fiberings of adjoint orbits, J. Amer. Math. Soc. 2 (1989),
517–533.
[BMV] J. L. Brylinski, B. Malgrange, J. L. Verdier, Transformation de Fourier géométrique. II, C. R. Acad. Sci. Paris Sér. I
Math. 303 (1986), 193–198.
[DM] P. Deligne, J. Milne, Tannakian categories, in Hodge cycles, motives, and Shimura varieties, Lecture Notes in
Math. 900, Springer, 1982.
[Do] C. Dodd, Equivariant coherent sheaves, Soergel bimodules, and categorification of affine Hecke algebras, preprint
arXiv:1108.4028 (2011).
[EV] P. Etingof, A. Varchenko, Dynamical Weyl groups and applications, Adv. Math. 167 (2002), 74–127.
[EM] S. Evens, I. Mirković, Fourier transform and the Iwahori–Matsumoto involution, Duke Math. J. 86 (1997), 435–464.
[FW] P. Fiebig, G. Williamson, Parity sheaves, moment graphs and the p-smooth locus of Schubert varieties, preprint
arXiv:1008.0719 (2010), to appear in Ann. Inst. Fourier (Grenoble).
[G1] V. Ginzburg, Perverse sheaves on a loop group and Langlands’ duality, preprint arXiv:alg-geom/9511007 (1995).
[G2] V. Ginzburg, Variations on themes of Kostant, Transform. Groups 13 (2008), 557–573.
[GK] V. Ginzburg, D. Kazhdan, A class of symplectic varieties associated with the space G/U , in preparation.
[Hu] J. E. Humphreys, Representations of semisimple Lie algebras in the BGG category O, Graduate Studies in Mathemat-
ics, 94. American Mathematical Society, Providence, RI, 2008.
[J1] J. C. Jantzen, Representations of algebraic groups, second edition, Mathematical surveys and monographs 107, Amer.
Math. Soc., 2003.
[J2] J. C. Jantzen, Nilpotent orbits in representation theory, in Lie theory, 1–211, Progr. Math. 228, Birkhäuser, 2004.
[Ka] M. Kashiwara, The universal Verma module and the b-function, in Algebraic groups and related topics (Kyoto/Nagoya,
1983), 67–81, Adv. Stud. Pure Math. 6, North-Holland, 1985.
[KLa] D. Kazhdan, G. Laumon, Gluing of perverse sheaves and discrete series representation, J. Geom. Phys. 5 (1988),63–120.
[KL] D. Kazhdan, G. Lusztig, Schubert varieties and Poincaré duality, in Geometry of the Laplace operator (Proc. Sympos.
Pure Math., Univ. Hawaii, Honolulu, Hawaii, 1979), 185–203, Proc. Sympos. Pure Math. XXXVI, Amer. Math. Soc.,
1980.
[Ko] B. Kostant, The principal three-dimensional subgroup and the Betti numbers of a complex simple Lie group, Amer. J.
Math. 81 (1959), 973–1032.
[LS] T. Levasseur, J. T. Stafford, Differential operators and cohomology groups on the basic affine space, in Studies in Lie
theory, 377–403, Progr. Math., 243, Birkhäuser, 2006.
[Lu] G. Lusztig, Cuspidal local systems and graded Hecke algebras. I. Inst. Hautes Études Sci. Publ. Math. 67 (1988),
145–202.
[Ma] H. Matsumura, Commutative ring theory, Cambridge University Press, 1986.
[MV] I. Mirković, K. Vilonen, Geometric Langlands duality and representations of algebraic groups over commutative rings,
Ann. of Math. 166 (2007), 95–143.
68
[NP] B. C. Ngô, P. Polo, Résolutions de Demazure affines et formule de Casselman-Shalika géométrique, J. Algebraic Geom.
10 (2001), 515–547.
[Sh] N. Shapovalov, Structure of the algebra of differential operators on the basic affine space, Funct. Analysis Applic. 8
(1974), 37–46.
[Sp] T. Springer, Quelques applications de la cohomologie d’intersection, Astérisque 92–93 (1982), 249–273.
[TV] V. Tarasov, A. Varchenko, Difference equations compatible with trigonometric KZ differential equations, Internat. Math.
Res. Notices 2000, no. 15, 801–829.
[Va] E. Vasserot, On the action of the dual group on the cohomology of perverse sheaves on the affine Grassmannian, Compo-
sitio Math. 131 (2002), 51–60.
[YeZ] A. Yekutieli, J. J. Zhang, Dualizing complexes and tilting complexes over simple rings, J. Algebra 256 (2002), 556–567.
[YZ] Z. Yun, X. Zhu, Integral homology of loop groups via Langlands dual group, Represent. Theory 15 (2011), 347–369.
D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF C HICAGO , C HICAGO , IL 60637, USA.

E-mail address: [email protected]
U NIVERSIT É B LAISE PASCAL - C LERMONT-F ERRAND II, L ABORATOIRE DE M ATH ÉMATIQUES , CNRS, UMR 6620,
C AMPUS UNIVERSITAIRE DES C ÉZEAUX , F-63177 A UBI ÈRE C EDEX , F RANCE .
E-mail address: [email protected]
69

1306.6754v2

Uploaded by

Copyright:

Available Formats

1306.6754v2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1306.6754v2

Uploaded by

Copyright:

Available Formats

DIFFERENTIAL OPERATORS ON G/U AND THE AFFINE GRASSMANNIAN

VICTOR GINZBURG AND SIMON RICHE

where we consider morphisms in the category of (S~ , U~ (g))-bimodules.

Here, W λ is the Wakimoto sheaf associated with λ, RG is an ind-perverse sheaf on Gr correspond-

Theorem. For F in PervǦ(O) (Gr) and λ ∈ X, there exist canonical isomorphisms

1.10. Conventions. Throughout,

2. S TATEMENT OF THE MAIN RESULTS

L of the subalgebra t ⊂ b is semisimple. Therefore, one has a weight decompo-

Lemma 2.2.1. For any F in PervǦ(O) (GrǦ ) and λ ∈ X, one has

(ıλ )! : HA (i!λ F) = HA (ı!λ t!λ F) −→ HA (Tλ , t!λ F).

Thus, we get a diagram of morphisms of graded S~ -modules

One of our key results (to be proved in §7.6) reads as follows.

The topological context is easy in this case. Namely we have

the isomorphism associated with w ∈ W . These isomorphisms generate an action of W in the

This collection of isomorphisms satisfies relations

for all V in Rep(G), λ ∈ X and w ∈ W .

The inclusion {λ} ֒→ Wλ ∩ Tλ induces a morphism of Q~ -modules

Proof of the first isomorphism in Lemma 2.4.1. We have

Remark 3.2.2. (1) As explained in §B.1, the construction of isomorphisms ΦV,λ

One can easily check that resG

Proposition 3.3.2. Let α ∈ I. Then the following diagram commutes:

As a corollary we obtain the following result. For λ ∈ X and V in Rep(G), we denote by

the morphism induced by resG

Corollary 3.4.3. The morphism rLG coincides with the composition

Next, by construction the following diagram commutes:

Then the commutativity follows from Corollary 3.4.3.

3.5. Classical analogue. Consider the sheaf of algebras

Corollary 3.6.3. The algebra D(X) is

P some elements µ̌i ∈ X∗ (T ), i = 1, · · · s, such that if λ(µ̌i ) = 0 for all i then λ = 0,

4. M ORPHISMS BETWEEN ASYMPTOTIC UNIVERSAL V ERMA MODULES

let w ∈ W . Choose a reduced decomposition w = sk · · · s1 , where si is the reflection associated

If µ ∈ t∗ , we denote by P 7→ P (µ) the unique algebra morphism S~ → C[~] which sends x ∈ t to

Spµ idV ⊗Spµ

and Lemma 4.3.2.

Lemma 4.4.1. Let V in Rep(G), and λ ∈ X.

where Pj,k,i ∈ S~ . Then for µ ∈ X such that n(µ) ≥ 0 we have

Lemma 4.6.4. For all V in Rep(G) and λ ∈ X, the morphism RV,λ

4.7. Definition of operators Θw .

Lemma 4.7.2. The operator ΘV,λ

5. G EOMETRY OF THE G ROTHENDIECK –S PRINGER RESOLUTION

5.1. The W -action on the regular part of e

coincide. As both isomorphisms are G × C× -equivariant, and as

hence it is enough to construct a collection of isomorphisms of G-equivariant line bundles

6. R EMINDER ON THE S ATAKE EQUIVALENCE

By [MV, Theorems 3.5 and 3.6], we have a canonical isomorphism

where d is the differential. Moreover, this isomorphism extends to an isomorphism of graded

where ~ ∈ S~ corresponds to the natural generator of H2C× (pt) = ∼ C.

E2p,q = HpA (pt) ⊗ Hq (i!λ F) (6.2.2)

In particular, the left-hand side is free over S~ = HA (pt).

HA (Tλ , t!λ F) → H (Tλ , t!λ F)

We deduce the following result from Lemma 2.2.1.

Proof. By the localization theorem in equivariant cohomology, the morphism

If P̌ = B̌ (so that Ľ = Ť ) we have canonical isomorphisms

(where for simplicity we neglect shifts of morphisms). As the morphism Ǧκtop

7. P ROOFS OF THE MAIN RESULTS

(u1 ×B α x1 ) = (u2 (u−1 −1

in B ×B α bαα−rs , which finishes the proof.

where in the right-hand side B α acts trivially on C[a∗α−rs ].

considered in Lemma 7.1.4. Hence it is an isomorphism in this case also.

induced by κtop alg