LOCALIZATION AND THE LONG INTERTWINING OPERATOR FOR REPRESENTATIONS OF AFFINE KAC-MOODY ALGEBRAS S. ARKHIPOV AND D. GAITSGORY Introduction This text is a raft draft, untouched since January 2009. What prevented its completion at the time was the absence of the documented formalism of loop groups acting on categories. Once this formalism is written down (hopefully, soon), this draft will be turned into an actual paper. Here is an overview of what’s done in this paper: The first author introduced a certain functor, denoted Φ, from the category of Kac-Moody representations at the negative level to that at the positive level. We interpret this functor as a composition of two contravariant functors: one is the usual contragredient duality (from the negative level to itself), and then Verdier duality from the negative level to the positive level (that comes from the fact that the two categories are dual to each other in the sense of Lurie). According to Kashwara-Tanisaki, the category at the negative level localizes onto the “thin” affine flag space, and the category at the negative level localizes onto the “thick” affine flag space. We show that in terms of the localization functors, Arkhipov’s functor corresponds to the “long intertwining operator” from the category of (twisted) D-modules on thin flags to that on thick flags. In the process we will need to consider the two version of the category of D-modules on Bun G , recently studied in [DrGa], denoted D-mod(Bun G ) and D-mod(Bun G ) co . We show that the category at the positive level naturally localizes on the former, whereas the category at the negative level naturally localizes on the latter. Date : May 15, 2015. 1
46
Embed
LOCALIZATION AND THE LONG INTERTWINING OPERATOR FOR …people.math.harvard.edu/~gaitsgde/GL/Arkh.pdf · FOR REPRESENTATIONS OF AFFINE KAC-MOODY ALGEBRAS S. ARKHIPOV AND D. GAITSGORY
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
LOCALIZATION AND THE LONG INTERTWINING OPERATOR
FOR REPRESENTATIONS OF AFFINE KAC-MOODY ALGEBRAS
S. ARKHIPOV AND D. GAITSGORY
Introduction
This text is a raft draft, untouched since January 2009. What prevented its completion atthe time was the absence of the documented formalism of loop groups acting on categories.Once this formalism is written down (hopefully, soon), this draft will be turned into an actualpaper.
Here is an overview of what’s done in this paper:
The first author introduced a certain functor, denoted Φ, from the category of Kac-Moodyrepresentations at the negative level to that at the positive level. We interpret this functor asa composition of two contravariant functors: one is the usual contragredient duality (from thenegative level to itself), and then Verdier duality from the negative level to the positive level(that comes from the fact that the two categories are dual to each other in the sense of Lurie).
According to Kashwara-Tanisaki, the category at the negative level localizes onto the “thin”affine flag space, and the category at the negative level localizes onto the “thick” affine flagspace. We show that in terms of the localization functors, Arkhipov’s functor corresponds tothe “long intertwining operator” from the category of (twisted) D-modules on thin flags to thaton thick flags.
In the process we will need to consider the two version of the category of D-modules onBunG, recently studied in [DrGa], denoted D-mod(BunG) and D-mod(BunG)co. We show thatthe category at the positive level naturally localizes on the former, whereas the category at thenegative level naturally localizes on the latter.
Date: May 15, 2015.
1
2 S. ARKHIPOV AND D. GAITSGORY
1. The finite-dimensional case
1.1. The categories.
1.1.1. Let G be an algebraic group and g its Lie algebra. According to our conventions, weshall denote by g-mod the canonical object of DGCat corresponding to the derived categoryof g-modules, and by H(g-mod) its heart, i.e., the abelian category of g-modules.
By Sect 10.2.3, the category g-mod identifies with its own dual (g-mod)∨, using the anti-automorphism
τ : U(g)x 7→−x, x∈g−→ U(g).
Namely, the pairing g-mod⊗ g-mod→ Vect is given by
(1.1) M,N 7→ 〈M,N〉g := M ⊗U(g)
N ' H•(g,M ⊗N).
Thus, we have an equivalence
Cang : g-mod→ (g-mod)∨,
and in particular,
(g-mod)c → ((g-mod)∨)c ' (g-mod)c,o.
We shall denote the resulting contravariant functor g-modc → g-modc by Dg. Explicitly, it isgiven by
Dg(M) ' Homg-mod(M,U(g)),
viewed as a right, and via τ as a left, g-module.
1.1.2. Let Y be a smooth scheme. Let DY-mod (resp., Let DopY -mod) be the (canonical object
of DGCat corresponding to) the derived category of left (resp., right) D-modules on Y, andlet H(DY-mod) and H(Dop
Y -mod) denote their respective hearts, the corresponding abeliancategories.
By Sect 10.2.4, the categories DY-mod and DopY -mod are naturally dual to each other with
the pairing given by
(1.2) F1,F2 7→ 〈F1,F2〉Y := ΓDR(Y,F1 ⊗OY-mod
F2) ' Γ(Y,F1 ⊗DY-mod
F2),
where F1 ⊗DY-mod
F2 is regarded as a sheaf of k-vector spaces in the Zariski topology.
Thus, we obtain an identification
CanY : (DY-mod)∨ ' Dop
Y -mod,
and in particular
(DY-mod)c →
((Dop
Y -mod)∨)c ' (Dop
Y -mod)c,o
.
We shall denote the resulting contravariant functor (DY-mod)c →
(Dop
Y -mod)c
by DY; this isthe usual Verdier duality functor.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 3
1.1.3. Let Y be as above and suppose that it is acted on by G. In this case we have a naturalfunctor
Γ(Y,−)l : DY-mod→ g-mod,
and its left adjoint, denoted Locg,Y given by M 7→ DY ⊗U(g)
M .
In addition, we can consider the functor Γ(Y,−)r : DopY -mod→ g-mod.
Proposition 1.1.4. Under the identifications
g-mod ' (g-mod)∨
and DopY -mod ' (DY-mod)∨,
the functors Γ(Y,−)r and Locg,Y are naturally mutually dual.
Proof. The assertion of the proposition is equivalent to the existence of a canonical isomorphism
〈M,Γ(Y,F)〉g ' 〈Locg,Y(M),F〉Y,which follows from the definitions. �
From Lemma 10.2.1, we obtain:
Corollary 1.1.5. The left adjoint of the functor Γ(Y,−)r is the ind-extension of the functor
1.1.6. Assume for a moment that the structure sheaf OY admits a right D-module structure,which is moreover, G-equivariant. This defines an identification DY-mod → Dop
Y -mod, which
intertwines the functors Γ(Y,−)l and Γ(Y,−)r. Under these identifications, the isomorphism ofCorollary 1.1.5 can be reformulated as the commutativity of the next diagram
(1.3)
(g-mod)cDg−−−−→ (g-mod)c
Locg,Y
y Locg,Y
y(DY-mod)
c DY−−−−→ (DY-mod)c.
1.1.7. A variant. Suppose that in the above situation Y carries an action of another algebraicgroup, denoted T . In particular, we have a weak action of T on the categories DY-mod andDop
Y -mod. Let DY-modT,w and DopY -modT,w denote the corresponding weakly equivariant cat-
egories. From Lemma 10.4.2, we obtain:
Lemma 1.1.8. The categories DY-modT,w and DopY -modT,w are compactly generated.
Hence, by Sect 10.4.1, we obtain that the duality functor (1.2), which naturally gives rise toa functor
(1.4) DY-modT,w ⊗DopY -modT,w → VectT,w = Rep(T )
InvT→ Vect
(here InvT is the functor of T -invariants) makes the categories DY-modT,w and DopY -modT,w
mutually dual. We shall denote by
CanT,wY : DY-modT,w → DopY -modT,w
the resulting equivalence, and by DT,wY the functor(DY-modT,w
)c→(Dop
Y -modT,w)c.
These functors commute with CanY and DY via the tautological forgetful functors to DY-modand Dop
Y -mod, respectively.
4 S. ARKHIPOV AND D. GAITSGORY
By Sect 10.5.2, the t-structures on DY-mod and DopY -mod induce t-structures on the cat-
egories DY-modT,w and DopY -modT,w, whose hearts are the usual categories of weakly T -
equivariant D-modules. We have:
Lemma 1.1.9. The categories DY-modT,w and DopY -modT,w identify with the derived categories
of their hearts.
1.1.10. Assume now that the action of T on Y commutes with that of G. In this case the functorΓ(Y,−)l naturally lifts to a functor
DY-modT,wΓ(Y,−)l→ g-mod⊗ Rep(T ),
and we define the functor ΓT,w(Y,−)l : DY-modT,w → g-mod as the composition
DY-modT,wΓ(Y,−)l→ g-mod⊗ Rep(T )
InvT→ g-mod.
The functor Locg,Y naturally lifts to a functor g-mod → DY-modT,w and is the left adjointof ΓT,w(Y,−)l.
1.1.11. Let us return to the setting of Sect 1.1.3. The statement and proof of Proposition 1.1.4extends to the present situation, i.e., under the identifications
(g-mod)∨ ' g-mod and
(DY-modT,w
)∨' Dop
Y -modT,w,
the functors Locg,Y and ΓT,w(Y,−)r and mutually dual, and the left adjoint to ΓT,w(Y,−)r isthe left adjoint of the functor
DY ◦ Locg,Y,T ◦Dg : (g-mod)c →(Dop
Y -modT,w)c.
1.2. The equivariant situation. Let B ⊂ G be a subgroup. All of the above categories carrya strong action of G, and in particular, of B.
1.2.1. Let g-modB denote the corresponding (strongly) B-equivariant category. By Sect 10.5.3,
the category g-modB inherits a t-structure from g-mod, whose heart is the abelian category ofHarish-Chandra modules with respect to the pair (g, B). By Sect 10.5.4, we have:
Lemma 1.2.2.
(a) The category g-modB identifies with the derived category of its heart, i.e., g-modB '(g, B)-mod.
(b) The category g-modB is compactly generated, and (g-modB)c is the preimage of g-modunder the forgetful functor.
From Sect 10.4.3, we obtain that the pairing (1.1) gives rise to a functor
g-modB ⊗ g-modB → VectBH•B→ Vect,
(here H•B denotes the functor of equivariant cohomology), which identifies g-modB with its owndual, i.e.,
(1.5) (g-modB)∨ ' g-modB .
We shall denote byCanBg : g-modB → (g-modB)∨,
and by DBg : (g-modB)c → (g-modB)c the resulting functors. They commute with Cang and
Dg via the tautological forgetful functor g-modB → g-mod.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 5
1.2.3. The pairing (1.5) can be viewed in terms of the equivalence given by Lemma 1.2.2 as fol-
lows: By the above lemma, g-modB identifies with (canonical object of DGCat correspondingto) the derived category (g, B)-mod of Harish-Chandra modules with respect to the pair (g, B).Choose a splitting B ←↩ T . Then the pairing
(g, B)-mod⊗ (g, B)-mod→ Vect,
corresponding to (1.5) is given by
(1.6) M,N 7→ 〈M,N〉(g,B) :=
(M ⊗
g;TN
)⊗ det(t[1]).
1.2.4. Let DY-modB and DY-modB;T,w denote the B-equivariant categories corresponding toDY-mod and DY-modT,w, respectively.
Lemma 1.2.5. The categories DY-mod and DY-modT,w are compactly generated.
By Sect 10.4.3, the pairing (1.2) gives rise to a functor
DY-modB ⊗DopY -modB → VectB
H•B→ Vect,
and defines an equivalence
(1.7) CanBY :(DY-modB
)∨' Dop
Y -modB ; DBY :(DY-modB
)c'(Dop
Y -modB)c,
and similarly
(1.8) CanB;T,wY :
(DY-modB;T,w
)∨'(Dop
Y -modB;T,w),
(1.9) DB;T,wY :
(DY-modB;T,w
)c'(Dop
Y -modB;T,w)c
1.2.6. The functors Γ(Y,−)l and Locg,Y are naturally compatible with the strong G-actions,and hence with the B-actions. By Sect 10.4, we obtain the mutually adjoint functors
Locg,Y : DY-modB � g-modB : Γ(Y,−)l,
which are compatible with the dualities in the same sense as in Proposition 1.1.4 and Corol-lary 1.1.5. If the line bundle KY admits a G-equivariant trivialization, then the isomorphismof (1.3) holds for B-equivariant categories as well. Similarly, the assertions of Sect 1.1.11 carryover to the B-equivariant context.
1.3. The enhanced affine space. From now on, G will be a reductive group, B ⊂ G a Borelsubgroup, Y = G/N and T be the Cartan group acting on G/N on the right.
1.3.1. Let χ be a k-point of Spec(Zg), where Zg := Z(U(g)). Let g-modχ be the full subcategoryof g-mod consisting of objects, whose localization off χ is 0, i.e.,
g-modχ ⊗Zg-mod
Zg-modχ,
where Zg-modχ denotes corresponding the full sub-category of Zg-mod. From Sect 10.1.1, weobtain that g-modχ is compactly generated.
By Sect 10.2.2, we obtain that the duality (1.1) restricts to a pairing g-modχ ⊗ g-modτ(χ),which defines an identification
Cang,χ : (g-modχ)∨ ' g-modτ(χ).
By Sect 10.5.1, the t-structure on g-mod induces a t-structure on g-modχ. We have:
6 S. ARKHIPOV AND D. GAITSGORY
Lemma 1.3.2. The natural functor makes g-modχ equivalent to the derived category of itsheart.
The above lemma implies that g-modχ identifies with the (canonical object of DGCatcorresponding to) derived category of g-modules, on which Zg acts with this generalized centralcharacter.
1.3.3. The fact that the weak action of T on DG/N -mod canonically extends to a strong one
implies, by Sect 10.4.4, that DG/N -modT,w is a category over the scheme t∗. Let λ ∈ t∗ be a
weight. Let DG/N -modT,w,λ be the full subcategory of DG/N -modT,w, whose localization off λis zero, i.e.,
DG/N -modT,w,λ := DG/N -modT,w,λ ⊗Sym(t)-mod
Sym(t)-modχ.
From Sect 10.1.1, we obtain that the category DG/N -modT,w,λ is compactly generated.
Note that the canonical line bundle KG/N admits a canonical trivialization as a G-equivariantline bundle, and the failure of this trivialization to be T -equivariant is given by the character−2ρ : T → Gm.
By Sect 10.2.2 and Sect 1.1.11, we obtain that the restriction of the pairing (1.4) identifies:
CanT,w,λG/N :(DG/N -modT,w,λ
)∨' DG/N -modT,w,−λ−2ρ.
We shall denote by DT,w,λG/N the resulting contravariant functor(DG/N -modT,w,λ
)c→(DG/N -modT,w,−λ−2ρ
)c.
By Sect 10.5.1, the t-structure on DG/N -modT,w induces a t-structure on DG/N -modT,w,λ.We have:
Lemma 1.3.4. The category DG/N -modT,w,λ identifies with the derived category of its heart.
1.3.5. Let $ : t∗ → Spec(Zg) be the Harish-Chandra map. We normalize it so that $(λ) =$(w(λ+ ρ)− ρ).
If χ = $(λ), it is easy to see that the functor ΓT,w(G/N,−)l sends DG/N -modT,w,λ to
g-modχ, and that the functor Locg,G/N sends g-modχ to DG/N -modT,w,λ.
We have the following theorem of [BB]:
Theorem 1.3.6. Assume that λ is such that λ+ρ is regular. Then the functors ΓT,w(G/N,−)l
and Locg,G/N define mutually inverse equivalences:
DG/N -modT,w,λ � g-modχ.
If λ + ρ is, moreover, dominant, then the above functors are exact, i.e., compatible with thet-structures.
Combining Theorem 1.3.6 with Proposition 1.1.11 we obtain:
Corollary 1.3.7. The functors ΓT,w(G/N,−)l and Locg,G/N intertwine the functors
DT,w,λG/N :(DG/N -modT,w,λ
)c→(DG/N -modT,w,−λ−2ρ
)cand
Dg,χ : (g-modχ)c → (g-modτ(χ))c.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 7
1.3.8. Let us now consider the B-equivariant situation. By Sect 10.4.5, the following commu-tative diagrams of categories are in fact a pull-back squares:
1.4. Contragredient duality and Arkhipov’s functor. Let G be a reductive group andB ⊂ G be the Borel subgroup.
8 S. ARKHIPOV AND D. GAITSGORY
1.4.1. Consider the category g-modB . Let us recall the definition of the contragredient dualityfunctor Contrg : (g-modB) ' (g-modB)∨. It will correspond to a contravariant functor
(1.14) M 7→M∗ : (g-modB)c → (g-modB)c.
We shall use the description of g-modB as (g, B)-mod given by Lemma 1.2.2. So, it issufficient to define the functor (1.14) as an exact contravariant self-equivalence on the categoryof finitely generated (g, B)-modules.
The latter is defined as follows. Let us choose a splitting B ←↩ T . For M ∈ H((g, B)-mod),let M ' ⊕
νM(ν) be the decomposition into weight spaces with respect to T . Since M is
assumed finitely generated, all M(ν) are finite-dimensional. Set M∗ := ⊕νM(ν)∗, where M(ν)∗
denotes the linear dual of M(ν). We define an action of g on M∗ by conjugating the naturalaction by (a choice of the representative of) the element w0 ∈ W . It is clear, however, thatthe construction of the functor M 7→M∗ does not depend, up to a canonical isomorphism, oneither the choice of T or that of a representative of w0.
It is clear also that for M = ∆λ–the Verma module with h.w. λ, we have
(∆λ)∗ = ∇−w0(λ),
where for a weight µ, we denote by ∇µ the dual Verma module with h.w. µ.
This implies that the functor (1.14) sends finitely generated Harish-Chandra modules tofinitely generated ones. The fact that (M∗)∗ 'M implies that an (anti)-self-equivalence.
1.4.2. We define the functor Φ : g-modB → g-modB as the composition
Φ = CanBg ◦Contrg .
We define the functor Ψ as the inverse of Φ, i.e., as the composition in the other orderContrg ◦CanBg . Our goal to describe this functor explicitly.
Recall that for any category C acted on by G and any F ∈ DG-modB,B there exists acanonical functor M 7→ F ?M : CB → CB . Let jw0,!, jw0,∗ denote the standard and co-standard
objects of DG-modB,B , corresponding to the element w0 of the Weyl group.
The main theorem in the finite-dimensional case reads as follows:
Note that since jw0,! ◦ jw0,∗ ' δB ' jw0,∗ ◦ jw0,!, and hence the functors of convolution withjw0,! and jw0,∗ are mutually inverse, the two assertions in the theorem are equivalent to eachother.
1.4.4. Let us give two tautological reformulations of Theorem 1.4.3. We remark that in theinfinite-dimensional case, Theorem 1.4.3 will not admit a direct analogue, whereas these refor-mulations will.
Consider jw0,! and jw0,∗ as B × B-equivariant D-modules on G. We will not distinguishbetween left and right D-modules on G due to the existence of a G×G-equivariant trivializationof the line bundle KG. Consider the objects
Note that this proposition provides a more comprehensible definition of the contragredientduality functor.
Proof. It is enough it establish the isomorphism at the level of complexes, where both M and Nadmit Verma flags. We claim that in this case both sides are acyclic off cohomological degree 0,and it would be enough to establish a canonical isomorphism between their 0-th cohomologies.Note that acyclicity for the LHS is evident: there are no higher Exts from a Verma module toa dual Verma module.
Let us make the choices B ←↩ T and w0 as in the definition of the functor Contrg. Bydefinition, the LHS is given by the relative Lie algebra cohomology of g, relative to t withcoefficients in Homk(M ⊗N, k), where the action of g on N is twisted by w0 (we shall denotethe resulting module by Nw0). Thus, the LHS calculates to
Proof. Denote DBg (M) =: M1 and DBg (N) =: N1. The LHS is then 〈M1, N1〉(g,B)⊗ det(b∗[−1])and the RHS, by (1.3), is isomorphic to
HomDG-modB×B (δB [−dim(G)],Locg⊕g,G(M1 ⊗N1)) ,
which can be rewritten as
HomVectB (k,M1 ⊗gN1) ' Hom(g,B)-mod(k,M1 ⊗N1),
and we are done by (1.6). �
2. Representations of the Kac-Moody algebra
2.1. The category of modules.
2.1.1. We fix a level κ : g ⊗ g → k and consider the Kac-Moody extension gκ. The object ofDGCat, denoted gκ-mod, has been introduced in [FG2]. We remind that, by definition, gκ-mod
is generated by compact objects of the form Indgκk (k), where k ⊂ g[[t]] is a lattice subalgebra.
2.1.2. The category gκ-mod is acted on strongly by the group ind-scheme G((t)) at level κ. Inparticular, if K is a pro-unipotent open-compact subgroup of G[[t]] we have a full subcategory
gκ-modK ⊂ gκ-mod of K-equivariant objects.
Lemma 2.1.3. Each of the categories gκ-modK is compactly generated;
(gκ-modK)c = gκ-modK ∩ (gκ-mod)c.
2.1.4. For two open compact subgroups K1 ⊂ K2, let eK2,K1 denote the tautological inclusion
functor gκ-modK2 → gκ-modK1 , and let AvK2,K1be its right adjoint.
By the general formalism of categories acted on by G((t)),
(2.1) gκ-mod ' lim−→K
gκ-modK ,
where the colimit is taken with respect to the functors e, and also
(2.2) gκ-mod ' lim←−K
gκ-modK ,
where the limit is taken with respect to the functors Av.
2.2. Duality in the Kac-Moody case. Let κ′ be the opposite level, i.e., κ′ = −κ − κKil.Our current goal is to construct a pairing
(2.3) 〈−,−〉g : gκ-mod⊗ gκ′-mod→ Vect,
which would identify (gκ-mod)∨ with gκ′ -mod.
The construction of the pairing (2.3) depends on an additional choice of trivializing a certaingerbe. This choice can be made by fixing an open-compact subgroup K0.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 11
2.2.1. The pairing (2.3) is defined as follows. We consider DG pairing
C+(gκ-mod)⊗C+(gκ′ -mod)→ C(Vect)
given by
(2.4) M,N 7→ C∞2 (g, k0,M ⊗N),
where C∞2 is the semi-infinite Chevalley complex taken with respect to the lattice k0 :=
Lie(K0) ⊂ g.
It is easy to see that if either M or N is acyclic, then so is C∞2 (g, k0,M ⊗N).
Since (gκ-mod)c (resp., (gκ′ -mod)c) is a obtained as a quotient category of a subcategory ofC+(gκ-mod) (resp., C+(gκ′ -mod)) by its intersection with the subcategory of acyclic complexes,we obtain that (2.4) gives rise to a pairing
(gκ-mod)c ⊗ (gκ′ -mod)c → Vect,
which we then ind-extend to obtain the pairing (2.3).
Theorem 2.2.2. The pairing (2.3) defines a perfect duality between gκ-mod and gκ′ -mod.
2.2.3. Let K1 ⊂ K2 be two open-compact subgroups. We observe:
Lemma 2.2.4. For M1 ∈ gκ-modK1 , M2 ∈ gκ-modK2 ,
〈eK2,K1(M2),M1〉g ' 〈M2,AvK2,K1
(M1)〉g.
Therefore, to prove Theorem 2.2.2, it suffices to show that the pairing (2.3) restricts to aperfect pairing
The existence of such RK,G((t))κ,κ′ will imply the duality assertion.
2.2.6. Recall that Dκ,κ′
G((t))-mod denotes the category of twisted D-modules on G((t)). It has a
natural forgetful functor Γ(G((t)),−) to gκ-mod⊗ gκ′ -mod. The category Dκ,κ′
G((t))-mod is acted
on strongly by the group G((t)) × G((t)) at the level (κ, κ′) and the above forgeftul functor iscompatible with the action. In particular, we obtain a functor
Γ(G((t)),−) : Dκ,κ′
G((t))-modK,K → gκ-modK ⊗ gκ′-modK .
Let δκ,κ′
G((t)),K be the canonical object in Dκ,κ′
G((t))-modK,K , corresponding to distributions on K
inside G((t)). We set
RK,G((t))κ,κ′ := Γ(G((t)), δκ,κ
′
G((t)),K).
12 S. ARKHIPOV AND D. GAITSGORY
Proposition 2.2.7. The object RK,G((t))κ,κ′ satisfies (2.7) and (2.8).
Proof. We shall prove (2.7), since (2.8) is similar. It is enough to consider the case M ∈(gκ-modK)c. By definition, it is enough to construct a functorial quasi-isomorphism of com-plexes
C∞2 (g, k0,M ⊗ RK,G((t))
κ,κ′ ) 'M ⊗ det. rel.(k, k0).
However, the latter has been carried out in [FG1].�
2.2.8. Let Cang denote the resulting equivalence
(gκ-mod)∨ ' gκ′ -mod,
and let us denote by
Dg : (gκ-mod)c → (gκ′ -mod)c
the corresponding contravariant equivalence.
Here is an example of a calculation of this functor. Let M = Indgκk (k). We claim that
D(M) ' Indgκ′k (k)⊗ det. rel.(k, k0). Indeed, we have
2.2.9. From now on we shall fix K0 to be the Iwahori subgroup I.
Consider now the category gκ-modI , which can be identified with (gκ-modK)I/K for any Kwhich is a normal subgroup of I. We have:
Lemma 2.2.10. The category gκ-modI is compactly generated. Its compact objects are thosewhich become compact under the forgetful functor to gκ-mod.
The pairing (2.3), being G((t))-invariant gives rise to a pairing
gκ-modI ⊗ gκ′ -modI → VectI ,
which, when composed with HI : VectI → Vect defines a pairing
(2.9) gκ-modI ⊗ gκ′ -modI → Vect .
By Sect 10.4.3, the pairing (2.9) is also perfect. We shall denote by CanIg the resultingequivalence
(gκ-modI)∨ ' (gκ′ -modI),
and by DIg the corresponding contravariant functor
(gκ-modI)c ' (gκ′ -modI)c;
these functors are compatible with Cang and Dg via the natural forgetful functors to (gκ-mod)c
and (gκ′ -mod)c,o, respectively.
We remark that the canonical object in gκ-modI ⊗ gκ′ -modI that defines CanIg is
(2.10) RI,G((t))κ,κ′ := Γ(G((t)), δI)
κ,κ′ .
For example, for M = ∆κ,µ–the Verma module with h.w. µ, we have DIg(M) ' ∆κ′,−µ.
2.3. Affine contragredient duality and affine Arkhipov’s functor.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 13
2.3.1. The affine algebra gκ we have been dealing with was the extension of the loop algebrag((t)). In what follows, we shall sometimes add a subscript gκ,0, where 0 ∈ P1 with t being aglobal coordinate on P1.
We shall also consider the affine algebra gκ,0, which is the central extension of the loopalgebra g((t−1)). Of course, one could identify gκ,0 ' gκ,∞, but we prefer not to do that.
2.3.2. From now on, we shall make a distinction positive and negative levels. We shall say thatthe level κ is negative if (on every simple factor of g), κ + κKil
2 = c · κKil, with c /∈ Q≥0. We
shall say that κ is positive if κ + κKil2 = c · κKil, with c /∈ Q≤0. Evidently, κ is negative if
and only if κ′ is positive. We shall say that κ is irrational if c /∈ Q (i.e., irrational=positive ∩negative). We shall say that κ is critical if c = 0 (i.e., κcrit = −κKil2 ).
2.3.3. Let κ be negative. We shall now define the contragredient duality functor
(2.11) Contr0→∞κ→κ′ : gκ,0-modI0 ' (gκ,∞-modI∞)∨.
This will be based on the following:
Lemma 2.3.4. Let κ be negative.
(1) The category H(gκ-modI) is Artinian, i.e., every finitely generated object is of finite length.
(2) The triangulated category (gκ-modI)c is equivalent to Db(H(gκ-modI)f.g.).
As in Sect 1.4.1, it suffices to define the corresponding exact contravariant
(2.12) M 7→M∗ : H(gκ,0-modI0)f.g. ' H(gκ,∞-modI∞)f.g..
For M ∈ H(gκ,0-modI0)f.g., write M ' ⊕µ,n
M(µ, n), where µ refers to the grading by means
of T ⊂ B ⊂ I, and n ∈ k is the degree with respect of the Sugawara L0 operator. Since Mis finitely generated, and κ is negative, each M(µ, n) is finite-dimensional. We set the vectorspace underlying M∗ to be ⊕
µ,nM(µ, n)∗. The action of the affine algebra gκ,∞ is defined as
follows: an element x⊗ tn acts as Adw0(x)⊗ tn; the fact that this action extends continuously
to power series follows from the definitions.
It is clear that (∆κ,µ,0)∗ ' ∇κ,−w0(µ),∞, where the latter denotes the dual affine Vermamodule, which is known to be finitely generated. This implies that (2.12) constructed abovehas the required properties, and in particular, gives rise to a functor (2.11), as required.
2.3.5. Let us again assume that κ is negative. We define the functor
Φ0→∞κ→κ′ : gκ,0-modI0 ' gκ′,∞-modI∞
to be the composition
(2.13) gκ,0-modI0Contr0→∞
κ→κ′−→ (gκ,∞-modI∞)∨CanIg−→ gκ′,∞-modI∞ .
We define the functorΨ∞→0κ′→κ : gκ′ -modI∞ ' gκ,0-modI0
to be the inverse of Φ0→∞κ→κ′ .
The goal of the rest of this paper is describe explicitly the kernels
S0→∞κ′,κ′ ∈ gκ′,0-modI0 ⊗ gκ′,∞-modI∞ and S∞→0
κ,κ ∈ gκ,0-modI0 ⊗ gκ,∞-modI∞
that give rise to the functors Φ0→∞κ→κ′ and Ψ∞→0
κ′→κ via the pairing 〈−,−〉g. In addition, we willshow how the functors Φ0→∞
κ→κ′ and Ψ∞→0κ′→κ are compatible with the localization functors.
14 S. ARKHIPOV AND D. GAITSGORY
2.3.6. Tautologically, we have the following descriptions:
(2.14) S0→∞κ′,κ′ = (Id⊗Φ) (RI0,G((t))
κ′,κ ) and S∞→0κ,κ = (Ψ⊗ Id) (RI∞,G((t−1))
κ′,κ ).
Proposition 2.3.7. The objects S0→∞κ′,κ′ and S∞→0
κ,κ belong to the hearts of the correspondingt-structures.
Proof. We have:
(2.15) Φ0→∞κ→κ′(∇κ,µ,0) = ∆κ′,w0(µ),∞
and, correspondingly,
(2.16) Ψ∞→0κ′→κ(∆κ′,w0(µ),∞) ' ∇κ,µ,0,
to prove the proposition, it suffices to observe that RI0,G((t))κ′,κ , regarded merely as an object of
H(gκ,0-modI0), has a filtration with subquotients isomorphic to dual Verma modules, and that
RI∞,G((t−1))κ′,κ , regarded merely as an object of H(gκ′,∞-modI∞) has a filtration with subquotients
isomorphic to Verma modules.�
3. D-modules on BunG
3.1. The spherical case.
3.1.1. Let X be a smooth complete curve. Let BunG(X) denote the moduli stack of principalG-bundles on X. In this section we shall establish some basic facts about the category ofD-modules on X.
By definition, the category of (left) D-modules, denoted DBunG(X)-mod on BunG(X) islim←−U
DU -mod, where the limit is taken over the partially ordered set of open substacks of finite
type U ⊂ BunG(X) and the functors DU2-mod→ DU1 -mod for j1,2 : U1 ↪→ U2 are j∗1,2 ' j!1,2.
More generally, a choice of a level κ, defines a TDO DκBunG(X) on BunG(X), and we set
DκBunG(X)-mod! := lim
←−U
DκU -mod.
In the sequel we will establish the following result:
Theorem 3.1.2. The category DκBunG(X)-mod! is compactly generated.
From the definitions (and independent of Theorem 3.1.2) we have the following descriptionof compact objects of Dκ
BunG(X)-mod!:
Lemma 3.1.3. An object F ∈ DκBunG(X)-mod! is compact if and only if there exists an open
sub-stack U ⊂ BunG(X) of finite type and FU ∈ (DκU -mod)c, such that F|U = FU and such that
the following equivalent conditions hold:
• For any F1 ∈ DκBunG(X)-mod!, supported on BunG(X)− U , we have Hom(F,F1) = 0.
• For any Uj↪→ U1, we have:
F|U1 ' j!(FU ),
in particular, the object j!(FU ) is defined.
We remark that the proof of Theorem 3.1.2 will provide an even more explicit description ofcompact objects in Dκ
BunG(X)-mod!.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 15
3.1.4. We define the category DκBunG(X)-mod∗ as
lim−→U
DκU -mod,
where for U1 ⊂ U2, the functor DκU1
-mod→ DκU2
-mod is (j1,2)∗.
Observe that for every open substack U ⊂ BunG(X) of finite type, we have:
(3.1) (DκU -mod)
∨ ' Dκ,opU -mod and Dκ,op
U -mod ' Dκ′
U -mod.
Hence, from Lemma 10.1.2 we obtain:
Corollary 3.1.5. There exists a canonical equivalence:(Dκ
BunG(X)-mod∗
)∨' Dκ′
BunG(X)-mod!.
3.1.6. Note that there is a naturally defined functor
(3.2) Q∗→! : DκBunG(X)-mod∗ → Dκ
BunG(X)-mod!.
Indeed, having such a functor amounts to a compatible collection of functors
DκU1
-mod→ DκU2
-mod
for all pairs U1, U2 of open subsets of finite type of BunG(X). The functors in question aredefined by
DκU1
-modj∗12,1→ Dκ
U1∩U2-mod
(j12,2)∗→ DκU2
-mod,
where ji,ij : Ui ∩ Uj ↪→ Ui.
We shall also prove the following:
Proposition 3.1.7. If κ is irrational, then the functor Q∗→! of (3.2) is an equivalence.
In fact, we shall prove a stronger assertion:
Proposition 3.1.8. Let κ be irrational. Then there exists an open substack U ⊂ BunG(X) offinite type, such that no non-zero object of Dκ
BunG(X)-mod! is supported on BunG(X)− U .
3.1.9. Let(Dκ
BunG(X)-mod)c,!∗
denote the small category
lim←−U
(DU -mod)c,
where the limit is taken over the poset of open sub-stacks U of finite type.
By construction, it admits a fully faithful functor into DκBunG(X)-mod!, and receives a fully
faithful functor from(Dκ
BunG(X)-mod!
)c. In addition, the Verdier duality functor defines a
contravariant equivalence:(Dκ
BunG(X)-mod)c,!∗
→(Dκ′
BunG(X)-mod)c,!∗
.
16 S. ARKHIPOV AND D. GAITSGORY
3.1.10. From Theorem 3.1.2 and Corollary 3.1.5, by duality we obtain a fully faithful functor(Dκ
BunG(X)-mod∗
)c↪→(Dκ
BunG(X)-mod)c,!∗
.
It is easy to see that the ind-extension of the composition(Dκ
BunG(X)-mod∗
)c↪→(Dκ
BunG(X)-mod)c,!∗
↪→ DκBunG(X)-mod!,
is the functor Q∗→! defined above.
3.2. Level structure.
3.2.1. Let x := x1, ..., xn be a finite collection of points on X. Consider the group-scheme
G(Ox) := G(Ox1 × ...× Oxn). Let K be a group sub-scheme of G(Ox) of finite codimension.
Let U ⊂ BunG(X) be an open sub-stack of finite type. Let UK be the stack (of infinite type)that classifies pairs (PG, α), where PG is a point of U , and α is a structure of level K on PG atx. Note that for a fixed U and K small enough, UK is a scheme of finite type. If K is normal
in G(Ox), we have an action of G(Ox)/K on UK and (UK)/(G(Ox)/K) ' U .
Set
DκBunG(X,K)-mod! := lim
←−U
DκUK -mod.
Note that unlike the case of K = G(Ox) (and an appropriate Iwahori variant discussedbelow), we do not know whether the category Dκ
BunG(X,K)-mod! is compactly generated or even
dualizable. Whether or not or compact objects generate DκBunG(X,K)-mod!, they are described
by the corresponding version of Lemma 3.1.3.
If K1 is a normal subgroup of K2, we obtain that DκBunG(X,K1)-mod! is acted on strongly by
K2/K1 and (Dκ
BunG(X,K1)-mod!
)K2/K1
' DκBunG(X,K2)-mod!.
In particular, we have the tautological forgetful functor
π!K1,K2
: DκBunG(X,K2)-mod! → Dκ
BunG(X,K1)-mod!,
(where πK1,K2denotes the map BunG(X,K1)→ BunG(X,K2)), and its right adjoint
(πK1,K2)∗ : Dκ
BunG(X,K1)-mod! → DκBunG(X,K2)-mod!.
3.2.2. We define the category DκBunG(X,K)-mod∗ as
lim−→U
DκU -mod,
where for U1 ⊂ U2, the functor DκUK1
-mod → DκUK2
-mod is (jK1,2)∗, where jK1,2 denotes the
corresponding open embedding UK1 ↪→ UK2 .
The equivalence (3.1) continues to hold. In partcular, we obtain a pairing
(3.3) 〈−,−〉BunG(X,K) : DκBunG(X,K)-mod! ⊗Dκ′
BunG(X,K)-mod∗ → Vect .
By Lemma 10.1.2 we obtain:
Corollary 3.2.3. If one of the categories DκBunG(X,K)-mod∗ or Dκ′
BunG(X,K)-mod! is dualizable,
then so is the other one, and in this case (3.3) identies them as each other’s duals.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 17
As in the case of BunG(X) we can introduce the small category, there exists a naturallydefined functor
QK∗→! : DκBunG(X,K)-mod∗ → Dκ
BunG(X,K)-mod!.
However, for general K, the functor QK∗→! is not an equivalence even for irrational κ.
3.2.4. Additionally, as in the case of BunG(X) we can introduce the small category(Dκ
BunG(X,K)-mod)c,!∗
:= lim←−U
(DκUK -mod)c,
equipped with fully faithful functors(Dκ
BunG(X,K)-mod!
)c↪→(Dκ
BunG(X,K)-mod)c,!∗
↪→ DκBunG(X,K)-mod!,
and Verdier duality defines an contravariant equivalence(Dκ
BunG(X,K)-mod)c,!∗
'(Dκ′
BunG(X,K)-mod)c,!∗
.
However, the conclusion of Sect 3.1.10 holds only conditionally:
Lemma 3.2.5. If Dκ′
BunG(X,K)-mod! is compactly generated, and hence the functor(Dκ
BunG(X,K)-mod∗
)c→(Dκ
BunG(X,K)-mod)c,!∗
is well-defined, then the ind-extension of the composition(Dκ
BunG(X,K)-mod∗
)c→(Dκ
BunG(X,K)-mod)c,!∗
↪→ DκBunG(X,K)-mod!
identifies with the functor QK∗→!.
3.3. The Iwahori case.
3.3.1. Let x := x1, ..., xn be a finite collection of points on X. Let BunG(X, Ix) (resp.,
BunG(X,◦Ix)) be the moduli stack of principal G-bundles equipped with a reduction of their
fibers at x1, ..., xn to B (resp., N). The stack BunG(X,◦Ix) is acted on by the torus Tn =
T × ...× T .
By the above, we have the categories DκBunG(X,Ix)-mod! and Dκ
BunG(X,◦Ix)
-mod!, the latter
being acted on strongly by Tn and we have
(Dκ
BunG(X,◦Ix)
-mod!)Tn ' Dκ
BunG(X,Ix)-mod!.
In addition, we can consider the category
Dκ
BunG(X,◦Ix)
-modTn,w
! := (Dκ
BunG(X,◦Ix)
-mod!)Tn,w,
which is easily seen to be equivalent to
lim←−U
DκU ′ -modT
n,w,
where U ′ denotes the preimage of U in BunG(X,◦Ix).
18 S. ARKHIPOV AND D. GAITSGORY
For a collections of weights λ = λ1, ..., λn, we can consider the corresponding monodromic
category (Dκ
BunG(X,◦Ix)
-mod!)Tn,w,λ, and it is easy to see that we have an equivalence:
Dκ
BunG(X,◦Ix)
-modTn,w,λ
! := (Dκ
BunG(X,◦Ix)
-mod!)Tn,w,λ ' lim
←−U
DκU ′ -modT
n,w,λ.
Proceeding as above, we define also the corresponding categories as the corresponding co-limits:
DκBunG(X,Ix)-mod∗, D
κ
BunG(X,◦Ix)
-mod∗, Dκ
BunG(X,◦Ix)
-modTn,w∗ , Dκ
BunG(X,◦Ix)
-modTn,w,λ∗ .
3.3.2. We shall prove the following:
Theorem 3.3.3. The category (Dκ
BunG(X,◦Ix)
-mod!)Tn,w,λ is compactly generated.
By Corollary 3.2.3, we obtain:
Corollary 3.3.4. (Dκ
BunG(X,◦Ix)
-modTn,w,λ∗
)∨' Dκ′
BunG(X,◦Ix)
-modTn,w,−λ
! .
3.3.5. The construction of Sect 3.1.6 renders to the present situation, i.e., we have a naturallydefined functor
QTn,w,λ∗→! : Dκ
BunG(X,◦Ix)
-modTn,w,λ∗ → Dκ
BunG(X,◦Ix)
-modTn,w,λ
! .
We will prove the following generalization of Proposition 3.1.7:
Proposition 3.3.6. Assume that κ is irrational and λ1, ..., λn are rational. Then the functor
QTn,w,λ∗→! is an equivalence.
As in the case of BunG(X), Proposition 3.3.6 follows from the next assertion:
Proposition 3.3.7. Let κ and λ be as in Proposition 3.3.6. Then there exists a sub-stack of
finite type U ⊂ BunG(X), such that no non-zero object of Dκ
BunG(X,◦Ix)
-modTn,w,λ
! is supported
on the complement to U ′ ⊂ BunG(X,◦Ix).
3.3.8. Finally, as in Sect 3.1.9, we can introduce a small category
(Dκ
BunG(X,◦Ix)
-modTn,w,λ
)c,!∗,
equipped with the corresponding functors, and the conclusion of Sect 3.1.10 holds.
3.4. The full level structure.
3.4.1. We are now going to define the category
DκBunG(X,x)
-mod!
of twisted D-modules over ”BunG(X) with a full level structure” at x.
We set
DκBunG(X,x)
-mod! := colim−→K
DκBunG(X,K)-mod!,
where the co-limit is taken with respect to the pull-back functors π!K1,K2
.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 19
By Lemma 10.1.2, we have:
DκBunG(X,x)
-mod! ' lim←−K
DκBunG(X,K)-mod!,
where the limit is taken with respect to the functors (πK1,K2)∗.
The latter implies that we also have the equivalence:
DκBunG(X,x)
-mod! ' lim←−U
DκU x
-mod,
where
DκU x
-mod := colim−→,π!
K
DκUK -mod ' lim
←−,π∗K
DκUK -mod.
Although each of the categories DκU x
-mod is compactly generated, we do not know whether
DκBunG(X,x)
-mod! is.
The category DκBunG(X,x)
-mod! is naturally acted on by G(Ox), and for K as above we have:
(DκBunG(X,x)
-mod!)K ' Dκ
BunG(X,K)-mod!.
3.4.2. We define the category DκBunG(X,x)
-mod∗ as
colim−→U
DκU x
-mod,
with respect to the functors
(jx1,2)∗ : Dκ
U x1-mod→ Dκ
U x2-mod.
By definition, DκBunG(X,x)
-mod∗ is equivalent to
colim−→K
DκBunG(X,K)-mod∗,
with respect to the functors (πK1,K2)!, and also
lim←−K
DκBunG(X,K)-mod∗,
with respect to the functors (πK1,K2)∗.
The assertion of Lemma 3.2.3 and the constructions of Sect 3.2.4 render ditto into the presentcontext.
3.4.3.
Theorem-Construction 3.4.4.
(1) The action of G(Ox) on the categories
DκBunG(X,x)
-mod! and DκBunG(X,x)
-mod∗
naturally extends to an action of the group ind-scheme G(Kx) at level κ.
(2) The pairing
DκBunG(X,x)
-mod! ⊗Dκ′
BunG(X,x)-mod∗ → Vect
is G(Kx)-invariant.
20 S. ARKHIPOV AND D. GAITSGORY
Proof. Fill in.�
4. Localization on BunG
4.1. Definition of localization.
4.1.1. Let gκ,x be the central extension at level κ of
(g⊗ Kx1)⊕ ...⊕ (g⊗ Kxn).
Let gκ,x-mod be the corresponding category of modules. It is naturally acted on by the
group ind-scheme G(Kx) at level κ.
Theorem-Construction 4.1.2.
(1) There exists a canonical functor
Locgκ,x,BunG(X,x),! : gκ,x-mod→ DκBunG(X,x)
-mod!,
compatible with the action of G(Ox).
(2) The data compatibility of Locgκ,x,BunG(X,x),! with G(Ox) can be naturally lifted to a data of
compatibility with the action of G(Kx).
4.1.3. Proof of Theorem 4.1.2(1). The data of the functor Locgκ,x,BunG(X,x),! amounts, by def-
inition, to a compatible family of functors
Locgκ,x,U x : gκ,x-mod→ DκU x
-mod.
Recall that (gκ,x-mod)c is a full subcategory in Db(H(gκ,x-mod)). Hence, it is sufficient toconstruct a compatible family of functors
LLocgκ,x,U x : Db(H(gκ,x-mod))→ gκ,x-mod→ DκU x
-mod.
The latter are obtained as the left derived functor of the naive localization functor
H(Locgκ,x,U x) : H(gκ,x-mod)→ H(DκU x
-mod).
Explicitly, let K ⊂ G(Ox) be small enough, so that k ∩ Stabg(Kx)
(u) = 0 for any u ∈ U x.
Then
LLocgκ,x,U x(Indgκ,xk (k)) ' H Locgκ,x,U x(Ind
gκ,xk (k)) ' Dκ
UK ,
where DκUK ∈ Dκ
UK -mod is regarded as an object of DκU x
-mod via the pull-back functor
DκUK -mod-mod→ Dκ
U x-mod.
The compatibility of the functors Locgκ,x,U x among themselves and with the action of G(Ox)
follows from the construction.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 21
4.1.4. In the course of the proof we have shown that for a fixed U , the functor Locgκ,x,U x sends
(gκ,x-modK)c → (DκU x
-modK)c,
and, hence,
(gκ,x-mod)c → (DκU x
-mod)c.
In particular, the restriction of the functor Locgκ,x,BunG(X,x),! to (gκ,x-mod)c factors as
(4.1) (gκ,x-mod)cLocgκ,x,BunG(X,x),!∗
−→(Dκ
BunG(X,x)-mod
)c,!∗↪→ Dκ
BunG(X,x)-mod!,
and
(4.2) (gκ,x-modK)cLocK
gκ,x,BunG(X,x),!∗−→
(Dκ
BunG(X,K)-mod)c,!∗
↪→ DκBunG(X,K)-mod!,
4.1.5. Proof of Theorem 4.1.2(2). Fill in.�
4.2. The functor of sections.
4.2.1. Consider now the category DκBunG(X,x)
-mod∗.
Theorem-Construction 4.2.2.
(1) There exists a canonical functor
Γ(BunG(X, x),−)∗ : DκBunG(X,x)
-mod∗ → gκ,x-mod,
compatible with the action of G(Ox).
(2) The data compatibility of Γ(BunG(X, x),−)∗ with G(Ox) can be naturally lifted to a data of
compatibility with the action of G(Kx).
4.2.3. Proof of Theorem 4.2.2(1). The data of a functor Γ(BunG(X, x),−)∗ amounts to a com-patible family of functors
Γ(U x,−) : DκU x
-mod→ gκ,x-mod
for open sub-stacks U ⊂ BunG(X) of finite type.
Each of the latter is determined by the corresponding functor
Γ(U x,−) : (DκU x
-mod)c → gκ,x-mod.
The latter is constructed as the composition of the usual derived functor of global sections
RΓ(U x,−) : (DκU x
-mod)c → Db(H(gκ,x-mod)) ↪→ D+(H(gκ,x-mod)),
and the fully faithful embedding
D+(H(gκ,x-mod)) ↪→ gκ,x-mod
of [FG2].
4.2.4. Proof of Theorem 4.2.2(2). Fill in.�
4.3. Duality.
22 S. ARKHIPOV AND D. GAITSGORY
4.3.1. The functor Locgκ,x,BunG(X,x),! does not in general map (gκ,x-mod)c to the subcategory
of compact objects in DκBunG(X,x)
-mod. Therefore, it does not admit a colimit preserving right
adjoint.
However, we have:
Proposition 4.3.2. The pairings
〈−,−〉BunG(X,x) : DκBunG(X,x)
-mod! ⊗Dκ′
BunG(X,x)-mod∗ → Vect
and
〈−,−〉gx : gκ,x-mod⊗ gκ′,x-mod
are compatible with the functors Locgκ,x,BunG(X,x),! and Γ(BunG(X, x),−)∗.
Proof. The assertion of the proposition reads that for
By the construction of both functors, the assertion reduces to showing that for an opensub-stack U ⊂ BunG(X) of finite type, M ∈ (gκ,x-mod)c and and F ∈ (Dκ′
U x-mod)c, we have a
functorial isomorphism:
(4.4) 〈Locgκ,x,U x(M),F〉U x ' 〈M,Γ(U x,F)〉gx .
The latter isomorphism follows from the definitions.�
4.3.3. Recall the functor
Locgκ,x,BunG(X,x),!∗ : (gκ,x-mod)c →(Dκ
BunG(X,x)-mod
)c,!∗of (4.1).
Let DBunG(X,x) denote the contravariant equivalence(Dκ
BunG(X,x)-mod
)c,!∗→(Dκ′
BunG(X,x)-mod
)c,!∗.
As a formal corollary of Proposition 4.3.2, we obtain:
Corollary 4.3.4. The functors
DBunG(X,x) ◦ Locgκ,x,BunG(X,x),!∗ ◦Dgx and Locgκ′,x,BunG(X,x),!∗
mapping
(gκ′,x-mod)c →(Dκ′
BunG(X,x)-mod
)c,!∗are isomorphic.
5. Finiteness properties of the category of D-modules on BunG
5.1. The spherical case.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 23
5.1.1. The goal of this subsection is to prove Theorem 3.1.2. In fact, we shall prove a strongerassertion, explained to us by V. Drinfeld:
Theorem 5.1.2. The stack BunG(X) can be represented as a union of open sub-stacks of finite
type Uα, such that for every pair Uαjα,β↪→ Uβ, the functor (jα,β)∗ sends(
DκUα -mod
)c → (DκUβ
-mod)c.
Let us show how this theorem implies Theorem 3.1.2. Indeed, by Verdier duality we obtain:
Corollary 5.1.3. For every pair Uαjα,β↪→ Uβ, the functor left adjoint to
(jα,β)! : DκUα -mod→ Dκ
Uβ-mod,
is defined.
From Lemma 10.1.2, we then obtain:
Corollary 5.1.4.
(1) The category DκBunG(X)-mod! can be written as colim
−→α
DκUα
-mod, where for Uα ⊂ Uβ, the
functor DκUα
-mod→ DκUβ
-mod is (jα,β)!.
(2) The category DκBunG(X)-mod! is compactly generated. Its compact objects are of the form
(jα)!(F) for F ∈(DκUα
-mod)c
, and where (jα)! denotes the tautological functor
DκUα -mod→ Dκ
BunG(X)-mod!
from point (1).
5.1.5. For the proof of Theorem 5.1.2, we shall need to recall some reduction theory. Let ΛGdenote the coweight lattice of G, and set ΛQ := Q⊗
ZΛ. Let ψ denote the map
ΛG ↪→ ΛQG ' ΛQ
[G,G] ⊕ ΛQZ0(G) → ΛQ
[G,G] ↪→ ΛQG.
Let ΛposG ⊂ ΛG denote the positive span of positive simple roots, and Λpos,QG ⊂ ΛQG the
corresponding rational span. Let Λ+G (resp., Λ+,Q
G ) denote the dominant cones.
For a parabolic P with Levi factor M consider the quotient ΛG,P := ΛG/Λ[M,M ]sc ' π1(M).Let ψP denote the map
ΛG,P ↪→ ΛQG,P ' (Λ[G,G]/Λ[M,M ])
Q ⊕ ΛQZ0(G) → (Λ[G,G]/Λ[M,M ])
Q ↪→ ΛQG,P .
Let also φP denote the map
ΛG,P ↪→ ΛQG,P ' ΛQ
[G,G]∩Z0(M) ⊕ ΛQZ0(G) → ΛQ
[G,G]∩Z0(M) ↪→ ΛQG.
Let ΛposG,P (resp., Λpos,QG,P ) be the image of ΛposG (resp., Λpos,QG ) in ΛG,P (resp., ΛQG,P ).
24 S. ARKHIPOV AND D. GAITSGORY
5.1.6. Consider the stack BunP (X) and let pP and qP denote its natural projections to BunG(X)and BunM (X), respectively. Recall that the group ΛG,P enumerates the connected componentsof the stacks BunM (X) and BunP (X) (the latter are in bijection via the projection pQ). For
λP ∈ ΛG,P we shall denote by BunλPP (X) the corresponding connected component.
Recall that a G-bundle PG ∈ BunG(X) is called ”semi-stable” if for every parabolic P , such
that PG = pP (PP ) with PP ∈ BunλPP (X), then ψP (λP ) ∈ −Λpos,QG,P .
Equivalently, PG is semi-stable if for every reduction PB of PG to the Borel B, we have
PB ∈ BunλB with ψ(λ) ∈ −Λpos,QG .
It is known that semi-stable bundles form an open sub-stack of finite type, BunssG (X). LetBunssM (X) be the corresponding open sub-stack of BunM (X), and let BunssP (X) be the pre-image of BunssM (X) in BunssP (X).
5.1.7. For a parabolic P let λP be an element ΛG,P . We say that is dominant if φP (λP ) ∈ Λ+,QG .
Let Λ+G,P denote the cone of dominant elements. We say that λP ∈ Λ+
G,P is regular if it lies off
the walls of Λ+G,P , i.e., if 〈φP (λP ), αi〉 > 0 for every simple root αi /∈ [M,M ].
Theorem 5.1.8.
(1) Let λP ∈ ΛG/Λ[M,M ] be dominant and regular. Then the map pP defines an isomorphism
between BunλP ,ssP (X) and a locally closed sub-stack of finite type in BunG(X). (We shall denote
this sub-stack BunP,λPG (X).)
(2) The sub-stacks BunP,λPG (X) for various pairs (P, λP ) are pairwise non-intersecting, and their
union covers BunG(X), i.e., every geometric point of BunG belongs to exactly one BunP,λPG (X).
(3) If PG ∈ BunP,λPG (X) admits a reduction to a parabolic P1 with parameter λP1, then the
element φP (λP )− φP1(λP1) ∈ Λpos,QG .
(4) Substacks of the form ⋃(P,λP ), φP (λP )−λ0∈Λpos,QG
BunP,λPG (X),
for some fixed λ0 ∈ Λ+,Q are closed.
5.1.9. Let P be a parabolic and P− an opposite parabolic. We shall identify their Levi factorsvia embedding M ' P ∩ P− into both P and P−.
Lemma 5.1.10. There exists an element d ∈ Q≥0, which depends on the genus of X, with thefollowing property:
For every parabolic P and λP ∈ ΛG,P such that αi(ϕP (λP )) ≥ d for every simple root αi /∈[M,M ], and for every PM ∈ BunλP ,ssM (X) the following holds:
• (i) H1(X, (NP )PM ) = 0, where NP is the unipotent radical of P .• (ii) H1(X,Lie((NP ))PM ) = 0.
We call a stratum BunP,λPG (X) with λ satisfying the assumption of the lemma ”deep enough”.
The first of these conditions implies that the map BunλP ,ssM (X) → BunλP ,ssP (X), given by theembedding is smooth and surjective. The second conditions implies that the map
p−P : BunλP ,ssP− (X)→ BunG(X)
is smooth.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 25
5.1.11. Finally, we are ready to prove Theorem 5.1.2. We take the open subsets Uα to be open
subsets that are unions of the strata BunP,λPG (X), which include all strata that are not deepenough. To prove the theorem, it is enough to show that if U is an open subset of finite type of
BunG(X) and V := BunλP ,ssG (X) is a stratum, which is contained and is closed in U , and F is
an object of(DκU−V -mod
)c, then the ∗-extension of F to the entire U will belong to (Dκ
U -mod)c.
Let us denote the morphisms U − V ↪→ U and V ↪→ U by j and i respectively. Since theimage of (Dκ
U -mod)c
under j∗ generates(DκU−V -mod
)c, it sufficient to show that the functor i!
sends (DκU -mod)
cto (Dκ
V -mod)c.
Consider the morphism p−P : BunλP ,ssP− (X) → BunG(X). By assumption, it is smooth. Let
U and V be the preimages in BunλP ,ssP− (X) of U and V , respectively. Let j and i denote thecorresponding morphisms.
Lemma 5.1.12. For any P and dominant regular λP , the natural morphism
BunλP ,ssM (X)→ BunλP ,ssP− (X) ×BunG(X)
BunλP ,ssP (X)
is an isomorphism.
By the ”deep enough” assumtion, the map π : U → U is smooth, and its restriction to V ,
i.e., the map V → V is a tower of fibrations into affine spaces. Hence, it is enough to show that
i!(F) ∈(DκV
-mod)c,
where F is the pull-back of F to U − V .
5.1.13. However, the latter situation falls into the following paradigm: we have a base stack S
(in our case S = BunλP ,ssM (X)), and another stack S′ mapping to S (in our case, the fiber of
S′ over PM ∈ BunλP ,ssM (X) is H1(X, (UP )PM )), in such a way that the morphism S′ → S isaffine, and there exists a Gm-action on S′ that preserves S and which contracts S′ to a sectioni : S → S′ (in our case, the Gm action is given by a regular dominant co-character of Z0(M)).
In this case, the functor i! sends(DS′-modGm
)cto (DS-mod)
c.
5.1.14. Proof of Proposition 3.1.8. We claim that the open sub-stack in question is any open
union of the strata BunλP ,ssG (X), which contains all those that are not sufficiently deep. In fact,we claim that for any sufficiently deep stratum, the category Dκ
BunλP ,ss
G (X)-mod is zero.
The stratum in question is isomorphic, as a stack, to BunλP ,ssP (X). The group Z0(M) actson it via its adjoint action on P . Moreover, the twisting given by κ is equivariant with respectto Z0(M) against the character of Lie(Z0(M)) equal to −κ(λP ,−), which is non-integral. Tofinish the proof, it suffices to notice that the ”sufficiently deep” assumption implies that the
above action of Z0(M) on BunλP ,ssP (X) is in fact trivial.
5.2. The Iwahori case.
26 S. ARKHIPOV AND D. GAITSGORY
5.2.1. As in the spherical case, Theorem 3.3.3 follows from the next result:
Theorem 5.2.2. For Uα, Uβ as in Theorem 5.1.2 with α ”deep enough”, the functor
(j′α,β)∗ : DκU ′α
-modTn,w,λ → Dκ
U ′β-modT
n,w,λ
sends (DκU ′α
-modTn,w,λ)c to (Dκ
U ′β-modT
n,w,λ)c.
This theorem admits the same corollaries as in the spherical case:
Corollary 5.2.3.
(1) The functor (j′α,β)!, left adjoint to (j′α,β)! : DκU ′β
-modTn,w,λ → Dκ
U ′α-modT
n,w,λ is defined.
(2) The category Dκ
BunG(X,◦Ix)
-modTn,w,λ
! is compactly generated and is equivalent to
lim−→α
DκU ′α
-modTn,w,λ,
under the functors (j′α,β)!.
(3) The compact objects of Dκ
BunG(X,◦Ix)
-modTn,w,λ
! are of the form (j′α)!(F), where F is an
object of(DκU ′α
-modTn,w,λ
)c, and where (j′α)! denotes the tautological functor Dκ
U ′α-mod →
Dκ
BunG(X,◦Ix)
-modTn,w,λ
! .
5.2.4. Proof of Theorem 5.2.2. Let (P, λP ) be as in Sect 5.1.7. Let BunλP ,ssG (X, Ix) (resp.,
BunλP ,ssG (X,◦Ix)) denote the preimage of the corresponding stratum in BunG(X, Ix) (resp.,
BunG(X,◦Ix)).
In the present situation we modify the definition of pair (P, λP ) to be deep enough, bymaking it slightly stronger. Namely, we shall require that
• (i) H1(X, (NP )′PM ) = 0, where (NP )′ is the sub-sheaf of Maps(X,NP ) equal to thekernel of the homomorphism of the latter to Nn
P , corresponding to evaluation at x =x1, ..., xn.• (ii) H1(X,Lie((NP ))PM (−x)) = 0.
As in the BunG(X) case, it is easy to see that all but finitely many strata are ”deep enough”.The open sub-stacks Uα that appear in Theorem 5.2.2 are open unions of strata that containall strata that are not deep enough in the new sense. As in the proof of Theorem 5.1.2, for theproof of Theorem 5.2.2, it is sufficient to show that for an open sub-stack U ⊂ BunG(X) of
finite type that contains a stratum BunλP ,ssG (X), the !-pullback functor sends
(DκU ′ -modT
n,w,λ)c → (Dκ
BunλP ,ss
G (X,◦Ix)
-modTn,w,λ)c.
To prove this, we shall need to refine our stratification. By definition, we have:
BunλP ,ssG (X, Ix) ' BunλP ,ssP (X) ×(pt /P )n
(B\G/P )n,
where the map
BunP → (pt /P )n
corresponds to taking the fiber of PP at the points x1, ..., xn.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 27
Let w := w1, ..., wn be an n-tuple of elements of the set W/WM . Let (B\G/P )w be the corre-
sponding Schubert stratum in (B\G/P )n, and let BunλP ,ss,wG (X, Ix) (resp., BunλP ,ss,wG (X,◦Ix))
be its preimage in BunλP ,ssG (X, Ix) (resp., BunλP ,ssG (X,◦Ix)).
It is sufficient to show that for an open sub-stack of finite type U ⊂ BunG(X,◦Ix), which
contains a deep enough stratum BunλP ,ssG (X), and F ∈ (DκU ′ -modT
n,w,λ)c, the !-restriction of
F to the stratum BunλP ,ss,wG (X,◦Ix) belongs to (Dκ
BunλP ,ss,w
G (X,◦Ix)
-modTn,w,λ)c.
5.2.5. Let P− be the opposite parabolic. Consider the Cartesian products
BunλP ,ssP− (X) ×BunG(X)
BunG(X, Ix),
and
BunλP ,ssP− (X) ×BunG(X)
BunG(X,◦Ix),
and let BunλP ,ss,wP− (X, Ix) (resp., BunλP ,ss,wP− (X,◦Ix)) be their locally closed sub-stacks, respec-
tively, corresponding to the condition that the relative position of the reductions of the G-torsor(PG)xi to B and P− is wi. (Our conventions are such that given a G-torsor with transversalreductions to P and P−, there exists exactly one reduction to B, which is in position w withrespect to P and P−.)
are smooth, and their images contain the strata BunλP ,ss,wG (X, Ix) and BunλP ,ss,wG (X,◦Ix), re-
spectively.
Consider the Cartesian product
(5.1) BunλP ,ss,wP− (X,◦Ix) ×
BunG(X,◦Ix)
BunλP ,ss,wG (X,◦Ix).
By Lemma 5.1.12, the stack (5.1) identifies with BunλP ,ssM (X,◦Ix) and its projection onto
BunλP ,ss,wG (X,◦Ix) is a tower of affine fibrations.
Hence, we need to show that for
F ∈
(Dκ
BunλP ,ss,w
P−(X,◦Ix)
-modTn,w,λ
)c,
its !-restriction to the sub-stack (5.1) is a compact object of the corresponding category.
Let i denote the corresponding closed embedding
i : BunλP ,ssM (X,◦Ix) ↪→ BunλP ,ss,wP− (X,
◦Ix).
In addition, we have a natural projection
(5.2) π : BunλP ,ss,wP− (X,◦Ix) � BunλP ,ssM (X,
◦Ix).
By the ”sufficiently deep” assumption, the morphism π is representable and affine. The twostacks appearing in (5.2) are acted on by Z0(M)n (via the embedding Z0(M) ↪→ T and the
28 S. ARKHIPOV AND D. GAITSGORY
action of the latter on G/N), and another copy of Z0(M) induced by its adjoint action on P−.Consider the action of Z0(M) given by
(5.3) z 7→ (w−11 (z), ..., w−1
n (z); z−1).
It acts trivially on BunλP ,ssM (X,◦Ix) and contracts BunλP ,ss,wP− (X,
◦Ix) to the image of the section
i. Hence, we find ourselves in the paradigm of Sect 5.1.13 (in the monodromic rather thanequivariant situation, with the conclusion being the same).
5.2.6. Proof of Proposition 3.3.7. We will show that the category Dκ
BunλP ,ss,w
G (X,◦Ix)
-modTn,w,λ
is zero for any (P, λP ), which is deep enough and any w.
By the above, it is sufficient to show that the corresponding category
Dκ
BunλP ,ss
M (X,◦Ix)
-modTn,w,λ
is zero. For that it is sufficient that the corresponding category of twisted D-modules
Dκ,λ
BunλP ,ss
M (X,Ix)-mod
on BunλP ,ssM (X, Ix) is zero.
Consider again the action of Z0(M) given by (5.3). The required assertion follows from the
fact that the resulting twisting on BunλP ,ssM (X, Ix) is Z0(M)-equivariant against the character
w−11 (λ1) + ...+ w−1
n (λn) + κ(λP ,−), which is non-integral.
6. Finiteness properties of the localization functor
6.1. Localization in the Iwahori case.
6.1.1. Let us return to the set-up of Sect 4.1. Consider the group-scheme K =◦Ix :=
◦Ix1×...×
◦Ix.
We obtain a functor
(6.1) Loc◦Ixgκ,x,BunG(X,x),!
: gκ,x-mod◦Ix → Dκ
BunG(X,◦Ix)
-mod!,
and the corresponding functors
(6.2) LocIxgκ,x,BunG(X,x),!
: gκ,x-modIx → DκBunG(X,Ix)-mod!,
(6.3) Loc◦Ix,T
n,w
gκ,x,BunG(X,x),!: gκ,x-mod
◦Ix,T
n,w := (gκ,x-mod◦Ix)T
n,w → Dκ
BunG(X,◦Ix)
-modTn,w
! ,
(6.4) Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!: gκ,x-mod
◦Ix,T
n,w,λ :=
(gκ,x-mod◦Ix)T
n,w,λ → DκBunG(X,Ix)-modT
n,w,λ! ,
for a collection λ = λ1, ..., λn of weights.
By Sect 4.1.4, each of the above functors, when restricted to the corresponding subcategoryof compact objects, factors as
Loc◦Ixgκ,x,BunG(X,x),!∗
: (gκ,x-mod◦Ix)c → (Dκ
BunG(X,◦Ix)
-mod)c,!∗,
LocIxgκ,x,BunG(X,x),!∗
: (gκ,x-modIx)c → (DκBunG(X,Ix)-mod)c,!∗,
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 29
6.1.3. By Proposition 4.3.2 and Theorem 3.3.3, we have:
Corollary 6.1.4. Under the equivalences(Dκ
BunG(X,Ix)-modTn,w,λ
!
)∨' Dκ′
BunG(X,Ix)-modTn,w,−λ
!
and (gκ,x-mod
◦Ix,T
n,w,λ
)∨' gκ′,x-mod
◦Ix,T
n,w,−λ,
the functors Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!and ΓT
n,w(BunG(X, x),−)∗ become the duals of one another.
Corollary 6.1.5. The functors
DBunG(X,x) ◦ Locgκ,x,BunG(X,x),!∗ ◦Dgx and Locgκ′,x,BunG(X,x),!∗
mapping
(gκ′,x-mod◦I,Tn,w,−λ)c →
(Dκ′
BunG(X,◦Ix)
-modTn,w,−λ
)c,!∗are isomorphic.
6.1.6. The main result of the present section is the following:
Theorem 6.1.7. Assume that κ is positive. Then the functor Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x)of (6.4) sends(
gκ,x-mod◦Ix,T
n,w,λ
)c→(Dκ
BunG(X,Ix)-modTn,w,λ
!
)c.
By Sect 3.3.8, this theorem can be reformulated as follows:
Theorem 6.1.8. Assume that κ is positive. Then functor
Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!∗: (gκ,x-mod
◦Ix,T
n,w,λ)c → (Dκ
BunG(X,◦Ix)
-modTn,w,λ)c,!∗
has its image in
(Dκ
BunG(X,◦Ix)
-modTn,w,λ
! )c ⊂ (Dκ
BunG(X,◦Ix)
-modTn,w,λ)c,!∗.
30 S. ARKHIPOV AND D. GAITSGORY
Further, by Corollary 6.1.5, we can reformulate Theorem 6.1.9 as follows:
Theorem 6.1.9. Assume that κ is negative. Then functor
Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!∗: (gκ,x-mod
◦Ix,T
n,w,λ)c → (Dκ
BunG(X,◦Ix)
-modTn,w,λ)c,!∗
has its image in
(Dκ
BunG(X,◦Ix)
-modTn,w,λ∗ )c ⊂ (Dκ
BunG(X,◦Ix)
-modTn,w,λ)c,!∗.
I.e., for M ∈ (gκ,x-mod◦Ix,T
n,w,λ)c, the !-stalks of Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!(M) are zero away from an
open sub-stack of finite type.
As a formal corollary of Theorems 6.1.7 and 6.1.9, we obtain:
Corollary 6.1.10.
(1) Let κ be negative. Then there exists a functor
Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),∗: gκ,x-mod
◦Ix,T
n,w,λ → Dκ
BunG(X,◦Ix)
-modTn,w,λ∗ ,
which is the left adjoint to ΓTn,w(BunG(X, x),−)∗. We have:
Q◦Ix,T
n,w,λ∗→! ◦ Loc
◦Ix,T
n,w,λ
gκ,x,BunG(X,x),∗' Loc
◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!.
(2) Let κ be positive. Then there exists a functor
ΓTn,w(BunG(X, x),−)! : Dκ
BunG(X,◦Ix)
-modTn,w,λ∗ → gκ,x-mod
◦Ix,T
n,w,λ,
right adjoint to Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!. We have:
ΓTn,w(BunG(X, x),−)! ◦Q
◦Ix,T
n,w,λ∗→! ' ΓT
n,w(BunG(X, x),−)∗.
6.2. Proof Theorem 6.1.7.
6.2.1. By definition, we have to show that for M among a set of compact generators of
gκ,x-mod◦Ix,T
n,w,λ there exists an open sub-stack of finite type UM ⊂ BunG(X) such that
for any F′ ∈ Dκ
BunG(X,◦Ix)
-modTn,w,λ supported outside of UM , we have
(6.9) HomDκ
BunG(X,◦Ix)
-modTn,w,λ
(Loc
◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!(M),F′
)= 0.
It is sufficient to take M to be the Verma module, i.e., ∆κ,µ,x, where µ = µ1, ..., µn is ann-tuple of weights such that µi − λi ∈ ΛG. The object
Loc◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!(∆κ,µ,x)
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 31
is the induced twisted D-module, i.e., for F′ ∈ Dκ
BunG(X,◦Ix)
-modTn,w,λ, we have:
HomDκ
BunG(X,◦Ix)
-modTn,w,λ
(Loc
◦Ix,T
n,w,λ
gκ,x,BunG(X,x),!(∆κ,µ,x),F′
)'
' Homtn
(kµ,Γ(BunG(X,
◦Ix),F′)
),
where kµ is the corresponding 1-dimensional representation of tn.
6.2.2. For (P, λP , w) let ιλP ,w denote the locally closed embedding of the corresponding stratum
into BunG(X,◦Ix). It is sufficient to take F′ from (6.9) to be of the form ιλP ,w∗ (F) for some
(6.10) F ∈ Dκ
BunλP ,ss,w
G (X,◦Ix)
-modTn,w,λ,
and to show that the vanishing occurs for all but finitely many strata. I.e., we have to showthat for all but finitely many strata
(6.11) Homtn
(kµ,Γ(BunG(X,
◦Ix), ιλP ,w∗ (F))
)= 0.
We can assume that F belongs to the heart of the t-structure. Moreover, we can assumethat F is an object of the corresponding category
Dκ,λ
BunλP ,ss,w
G (X,Ix)-mod
of twisted D-modules on BunλP ,ss,wG (X, Ix). Multiplying by the line bundle on BunG(X, Ix),
corresponding to λ− µ, we are reduced to calculating
(6.12) Γ(
BunG(X, Ix), ιλP ,w∗ (F))
for
(6.13) F ∈ Dκ,µ
BunλP ,ss,w
G (X,Ix)-mod,
and showing that it vanishes for all but finitely many strata.
6.2.3. With no restriction of generality we can assume that (P, λP ) is ”deep enough”. Recallthe projection
q−P : BunλP ,ssM (X, Ix)→ BunλP ,ss,wG (X, Ix).
It realizes BunλP ,ss,wG (X, Ix) as a quotient of BunλP ,ssM (X, Ix) with respect to the trivial actionof the unipotent group-scheme U, whose fiber at PM ∈ BunM (X) is
Maps(X,N(P )wPM ),
where N(P )w ⊂ N(P ) is the sub-sheaf, consisting of sections
{x 7→ n(x) |wi(n(xi)) ∈ n+}.
Let u denote the sheaf of Lie algebras corresponding to U.
Consider the pull-back of normal bundle to BunλP ,ssG (X) inside BunG(X); let E1 denote its
pull-back to BunλP ,ssM (X, Ix). This is a vector bundle, whose fiber at PM ∈ BunM (X) is thevector space H1(X, (g/p)PM ).
32 S. ARKHIPOV AND D. GAITSGORY
Consider now the normal bundle to BunλP ,ss,wG (X, Ix) insider
BunG(X, Ix) ×BunG(X)
BunλP ,ssG (X).
Let us denote by E2 its pull-back BunλP ,ssM (X, Ix). This is a vector bundle, whose fiber atPM ∈ BunM (X) is the vector space
⊕i=1,...,n
(g/n(P ))wi(PM )xi,
where (g/n(P ))w ⊂ g/n(P ) is the subspace , consisting of elements
{g |w0 · w(g) ∈ n+}.
For F in Sect 6.13, let F denote its pull-back to BunλP ,ssM (X, Ix). To prove the vanishingof (6.12), it is sufficient to show that for any integers k1, k2 and k3 the space of sections over
BunλP ,ssM (X, Ix) of the quasi-coherent sheaf
(6.14) F ⊗ Λtop(E1)⊗ Λtop(E2)⊗ Symk1(E1)⊗ Symk2(E2)⊗ Λk3(u∗)
vanishes.
We shall again use the torus Z0(M) that acts trivially on BunλP ,ssM (X,◦Ix), but with respect
to which the quasi-coherent sheaf (6.14) has an equivariant structure. Let us calculate thecharacters of this action.
6.2.4. The action of Z0(M) on E2 is trivial, and hence, so is the action on Λtop(E2)⊗Symk1(E2).
The action of Z0(M) on F is given by the character
w−11 (µ1) + ...+ w−1
1 (µn) + κ(λ,−),
if this character is integral (for otherwise, F = 0).
The action of Z0(M) on E1 consists of the characters α ∈ n(P ), each coming with multiplicity
equal to (1− g) + 〈α, λ〉. Therefore, the characters η1 on Symk1(E1) are sums of α’s from n(P ).
In particular, the action of Z0(M) on the line bundle Λtop(E1) is given by the character(Σ
α∈n(P )〈α, λ〉 · α
)+ 2(1− g) · ρP ,
where the first term is easily seen to be equal to κKil2 (λ,−), and where 2 · ρP := Σ
α∈n(P )α.
The action of Z0(M) on Λk3(u∗) has characters η3 equal to sums of α’s from n(P ).
6.2.5. Let ν : Gm → Z0(M) be a regular dominant co-character, i.e., 〈α, ν〉 > 0 for anyα ∈ n(P ). The resulting character of Gm on the space of sections of (6.14) is the sum of threeterms:
(1) (κ− κKil2 )(λ, ν)
(2) 〈η1, ν〉+ 〈η2, ν〉(3) 2(1− g)〈ρP , ν〉+ Σ
i=1,...,n〈w−1
i (µi), ν〉.
We obtain that the integers on the third line are bounded in a way independent of λP ; theintegers on the second line are positive, and the integer on the first line is arbitrarily large withλP as long as the form (κ− κKil
2 ) is positive definite.�
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 33
7. Arkhipov’s functor in the affine case
7.1. The kernel for Arkhipov’s functor.
7.1.1. Let κ be a negative level and recall the objects
The goal of this section is to describe S∞→0κ,κ and S0→∞
κ′,κ′ explicitly.
7.1.2. From now on we shall fix the global curve X to be P1 with two marked points x1 = 0and x2 =∞. We fix a coordinate t on P1 so that t(0) = 0.
Consider the stacks BunG(P1), BunG(P1, I0, I∞). Let BunssG (P1) ⊂ BunG(P1) be the semi-stable locus, which identifies with pt /G. Let BunssG (P1, I0, I∞) ⊂ BunG(P1, I0, I∞) be itspreimage under the forgetful map BunG(P1, I0, I∞)→ BunG(P1). We have an isomorphism:
BunssG (P1, I0, I∞) ' (G/B)× (G/B)/G.
Let
Bunss,genG (P1, I0, I∞) ⊂ BunssG (P1, I0, I∞)
be the open sub-stack corresponding to the open G-orbit on (G/B)× (G/B). We let
: Bunss,genG (P1, I0, I∞) ↪→ BunG(P1, I0, I∞)
denote the corresponding morphism.
Note that the TDO on BunG(P1), corresponding to any κ, trivializes over BunssG (P1). In par-ticular, the corresponding TDO on BunG(P1, I0, I∞) trivializes also over Bunss,genG (P1, I0, I∞).Let
Additionally, we shall consider its Verdier dual, i.e.,
(7.2) !(OBunss,genG (P1,I0,I∞)) ∈ (Dκ′
BunG(P1,I0,I∞)-mod!)c.
34 S. ARKHIPOV AND D. GAITSGORY
7.1.3. The following should be regarded as the main result of the present paper:
Theorem 7.1.4. We have a canonical isomorphism in gκ,0,∞-modI0,I∞ :
S∞→0κ,κ ' Γ
(BunG(P1, 0, ∞), ∗(OBunss,genG (P1,I0,I∞))
)∗.
In addition, we propose the following conjecture:
Conjecture 7.1.5. We have a canonical isomorphism in gκ′,0,∞-modI0,I∞ :
S0→∞κ′,κ′ ' Γ
(BunG(P1, 0, ∞), !(OBunss,genG (P1,I0,I∞))
)!.
Theorem 7.1.4 and Conjecture 7.1.5 can be regarded as affine analogs of Theorem 1.4.3 inits incarnation given by Corollary 1.4.5. Later on we shall establish an affine analog of thecommutative diagram of Corollary 1.4.6.
7.2. Proof of Theorem 7.1.4.
7.2.1. The main step in the proof of the theorem is the following proposition, which is an affineanalog of Proposition 1.5.1:
Proposition 7.2.2. For M ∈ (gκ,0-modI0)c, N ∈ (gκ,∞-modI∞)c there exists a canonicalisomorphism:
7.2.3. Proof of Proposition 7.2.2. As in the finite-dimensional case, it is enough to establishan isomorphism on the level of complexes, which consists of modules with a standard. Wewill show that if both M and N admit a standard flag, then both sides of the equation in theproposition are acyclic off degree 0 and are isomorphic to one another.
Indeed, since (∆κ,∞,µ)∗ ' ∇κ,0,−w0(µ), the acyclicity assertion is true for the LHS. Its 0-thcohomology for any M and N is given by
Hom(
(M ⊗Nw0)g[t,t−1] , k).
Hence, the full expression is given by
(7.3) Hom(H•(g[t, t−1];T,M ⊗Nw0), k),
since the vector space in (7.3) is acyclic off degree 0 if M and N admit standard flags, becausein this case the tensor product M ⊗N is free over both Lie sub-algebras
t · g[[t]]⊕ n and t−1 · g[[t−1]]⊕ n−.
Remark. The vector space in (7.3) identifies, up to a twist by w0 with the space of conformalblocks on P1 with coefficients in M and N . So, the above assertion is the well-known statementthat the Hom into the contragredient module can be calculated as conformal blocks.
The RHS of Proposition 7.2.2, by adjunction, is given by
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 35
and the stabilizer in g((t)) ⊕ g((t−1)) of the (unique) point in the corresponding T -torsor isg[t, t−1]. Hence the RHS of Proposition 7.2.2 is also given by (7.3).
�
7.2.4. Proposition 7.2.2 ⇒ Theorem 7.1.4. By definition, we need to show that for M ∈(gκ,0-modI0)c N ∈ (gκ′,∞-modI∞)c we have a canonical isomrphism
(7.4) Homgκ,0-modI0
(M, (Dg(N))∗
)'
' Homgκ,0-modI0
(M, 〈Γ
(BunG(P1, 0, ∞), ∗(OBunss,genG (P1,I0,I∞))
)∗, N〉g∞
).
By Proposition 7.2.2 , the LHS can be rewritten as
The latter is, however, tautologically isomorphic to the RHS of (7.4).�
8. Localization onto ”thick” flags
8.1. The ”thick” affine flag scheme.
8.1.1. By definition, the thick affine scheme (resp., enhanced thick affine scheme) is FlthickG :=
BunG(P1,◦I0, ∞) (resp., Fl thickG := BunG(P1, I0, ∞)). We will be concerned with the categories
of equivariant D-modules on it, i.e., one of the categories
Dκ′
FlthickG
-mod!, Dκ′
FlthickG
-modT,w! , Dκ′
FlthickG
-modT,w,λ! ,
defined as in Sect 3.4.
Viewing FlthickG as endowed with an action of G(K∞), following the recipe of Sect 4.1, wehave a functor
Locgκ′,∞,FlthickG ,! : gκ′,∞-mod→ Dκ′
FlthickG
-modT,w! .
Lemma 8.1.2. Under the equivalence:
Dκ′
FlthickG
-modT,w! ' Dκ′
BunG(P1,0,∞)-mod
◦I0
! ,
the functor Locgκ′,∞,FlthickGidentifies with the composition
gκ′,∞-mod→ gκ,0-mod◦I0 ⊗ gκ′,∞-mod
Loc◦I0,T,w
gκ′,0,∞,BunG(P1,0,∞),!
−→ Dκ′
BunG(P1,0,∞)-mod
◦I0,T,w! ,
where the first arrow is the functor M 7→ Indgκ,0
Lie(◦I0)
(k)⊗M .
36 S. ARKHIPOV AND D. GAITSGORY
8.1.3. We shall consider a particular case of the above set up, restricting ourselves to the I∞-equivariant situation. Namely, we will consider the category
Dκ′
FlthickG
-mod(T,w),I∞! .
Proposition 8.1.4. Every object F ∈ Dκ′
FlthickG
-mod(T,w),I∞! canonically splits as a direct sum
F0 ⊕ F 6=0, where
F0 ∈ Dκ′
FlthickG
-mod(T,w,0),I∞! ,
and
F 6=0 ∈ ⊕λ 6=0
Dκ′
FlthickG
-mod(T,w,λ),I∞! .
Proof. By the definition of Dκ′
FlthickG
-mod(T,w),I∞! , we have to prove the corresponding assertion
over each T -stable open sub-stack of finite type
U ⊂ (FlthickG )/I∞ ' BunG(P1,◦I0, I∞).
However, each such sub-stack is a finite union of locally-closed sub-stacks Y, each of theform pt /U for some unipotent group U. It is sufficient to prove the assertion for each of the
categories Dκ′
Y -modT,w. However, the category Dκ′
Y -mod is equivalent to Vect, with the actionof T being trivial as a weak action, and the strong action deviating the standard one by acharacter. Hence, the assertion of the proposition amounts to splitting an object of Rep(T )into isotypic components.
�
8.2. The Kashiwara-Tanisaki equivalence.
8.2.1. Let κ be negative, and hence κ′ be positive. Consider the functor
Loc(T,w,0),I∞
gκ′,∞,FlthickG ,!: gκ′,∞-modI∞ → Dκ′
FlthickG
-mod(T,w,0),I∞!
equal to the composition of
LocI∞gκ′,∞,FlthickG ,!
: gκ′,∞-modI∞ → Dκ′
FlthickG
-mod(T,w),I∞! ,
followed by the functor F 7→ F0, given by Proposition 8.1.4.
Proposition 8.2.2. The above functor Loc(T,w,0),I∞
gκ′,∞,FlthickG ,!sends
(gκ′,∞-modI∞)c → (Dκ′
FlthickG
-mod(T,w),I∞! )c.
Proof. By Proposition 8.1.4, for M ∈ (gκ′,∞-modI∞)c the restriction of Loc(T,w,0),I∞
gκ′,∞,FlthickG ,!(M) to
every T -stable open sub-stack U ⊂ (FlthickG )/I∞ of finite type, belongs to (Dκ′
U -mod(T,w),I∞! )c.
We have to show that there exists a large enough open sub-stack of finite type UM such that
HomDκ′
FlthickG
-mod(T,w),I∞!
(Loc
(T,w,0),I∞
gκ′,∞,FlthickG ,!(M),F
)= 0
for any
F ∈ Dκ′
FlthickG
-mod(T,w),I∞!
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 37
supported off U . With restriction of generality, we can take F to be T -equivariant, i.e., anobject of
Dκ′
Fl thickG-modI∞
! ' Dκ′
BunG(P1,I0,I∞).
In this case we have:
HomDκ′
FlthickG
-mod(T,w),I∞!
(Loc
(T,w,0),I∞
gκ′,∞,FlthickG ,!(M),F
)'
' HomDκ′BunG(P1,I0,I∞)
-mod!
(LocI0,I∞
gκ′,0,∞,BunG(P1,0,∞),!(Ind
gκ′,0Lie(I0)(k)⊗M),F
).
Since Indgκ′,0Lie(I0)(k)⊗M ∈ (gκ′,0,∞-modI0,I∞)c, our assertion follows from Theorem 6.1.7.
�
8.2.3. Consider now the functor
ΓT,w(FlthickG ,−)! : Dκ′
FlthickG
-mod(T,w,0),I∞! → gκ′,∞-modI∞ ,
equal to the composition
ΓT,w(BunG(P1, 0, ∞),−)! : Dκ′
FlthickG
-mod(T,w,0),I∞! '
' Dκ′
BunG(P1,0,∞)-mod
(◦I0,T,w,0),I∞
! → gκ,0-mod◦I0,T,w,0 ⊗ gI∞κ′,∞,
(see Corollary 6.1.10) followed by the functor
gκ,0-mod◦I0,T,w,0 ⊗ gI∞κ′,∞ → gI∞κ′,∞
that sends M 7→M◦I0,T .
Proposition 8.2.4. The functors Loc(T,w,0),I∞
gκ′,∞,FlthickG ,!and ΓT,w(BunG(P1, 0, ∞),−)! are mutually
adjoint.
Proof. By definition, we have to establish a functorial isomorphism for M ∈∈ (gκ′,∞-modI∞)c
and F ∈ Dκ′
FlthickG
-mod(T,w),I∞!
(8.1) HomDκ′
FlthickG
-mod(T,w),I∞!
(Loc
(T,w,0),I∞
gκ′,∞,FlthickG ,!(M),F
)'
' HomgI∞κ′,∞
(M,ΓT,w(FlthickG ,F)!
).
The LHS in (8.1) is by definition
(8.2) HomDκ′
BunG(P1,◦I0,I∞)
-modT,w!
(Loc
(◦I0,T,w),I∞
gκ′,0,∞,BunG(P1,0,∞),!(Ind
gκ′,0Lie(I0)(k)⊗M),F
),
which by Proposition 8.2.2 can be rewritten as
(8.3) colim−→j
HomDκ′
BunG(P1,◦I0,I∞)
-modT,w,0!
(Loc
(◦I0,T,w),I∞
gκ′,0,∞,BunG(P1,0,∞),!(Ind
gκ′,0Lie(I0)(k)⊗M)/tj ,F
),
where tj denotes the j-th power of the maximal ideal in Sym(t).
38 S. ARKHIPOV AND D. GAITSGORY
Consider the modules Indgκ′,0Lie(I0)(k)/tj ∈ gκ′,0-mod
◦I0,T,w,0 for j = 0, 1, .... For any N ∈
gκ′,0-mod◦I0,T,w,0 we have:
Homgκ′,0-mod
◦I0,T,w
(Ind
gκ′,0Lie(I0)(k), N
)' colim
−→j
Homgκ′,0-mod
◦I0,T,w
(Ind
gκ′,0Lie(I0)(k)/tj , N
).
Hence, the RHS in (8.1) can be rewritten as
Homg
(◦I0,T,w)I∞κ′,0,∞
(Ind
gκ′,0Lie(I0)(k)⊗M,ΓT,w(FlthickG ,F)!
)'
' colim−→j
Homgκ′,0,∞-mod(
◦I0,T,w,0),
◦I∞
(Ind
gκ′,0Lie(I0)(k)/tj ⊗M,ΓT,w(FlthickG ,F)!
),
which by adjunction can be further rewritten as
(8.4) colim−→j
HomDκ′
BunG(P1,◦I0,I∞)
-modT,w,0!
(Loc
(◦I0,T,w,0),I∞
gκ′,0,∞,BunG(P1,0,∞),!(Ind
gκ′,0Lie(I0)(k)/tj ⊗M),F
),
which is the same as the expression in (8.3).�
8.2.5. The localization theorem at the positive level, due to Kashiwara-Tanisaki reads as follows:
Theorem 8.2.6. The functor ΓT,w(BunG(P1, 0, ∞),−)! realizes Dκ′
FlthickG
-mod(T,w,0),I∞! as a
direct summand of gκ′,∞-modI∞ .
We will show how deduce theorem from another localization result (also due to Kashiwaraand Tanisaki), but which takes place at the negative level.
9. Arkhipov’s functor and localization
9.1. From thin to thick flags.
9.1.1. Let FlthinG (resp., FlthinG ) be the ”thin” affine flag scheme (resp., the ”thin” enhancedaffine flag scheme), attached to point 0 ∈ P1, i.e.,
FlthinG := G(K0)/I0, G(K0)/◦I0.
Note that unlike the corresponding ”thick” spaces, the ”thin” ones are local, i.e., only dependon the formal neighborhood of 0 ∈ P1.
For a level κ we can consider the corresponding categories of twisted D-modules
DκFlthinG
-mod and DκFlthinG
-mod.
The ind-schemes FlthinG and FlthinG are acted on by G(K0), and the above categories carry astrong action of this group ind-scheme at level κ.
9.1.2. The rest of this sub-section is devoted to the proof of the following result:
Theorem-Construction 9.1.3. There exists a canonical equivalence
Φthin→thick : DκFlthinG
-modI0,(T,w,0) ' Dκ′
FlthickG
-mod(T,w,0),I∞!
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 39
9.1.4. To construct the functor Φthink→thick we recall that the involution g 7→ g−1 defines anequivalence
DκFlthinG
-modI0,(T,w,0) ' Dκ′
Fl thinG-mod
◦I0,T,w,0.
The action of G(K0) on the category Dκ′
BunG(X,0,I∞)-mod! defines the convolution functor
(9.1) Dκ′
FlthinG-modI0,T,w,0 ⊗Dκ′
BunG(X,0,I∞)-modI0
! → Dκ′
BunG(X,0,I∞)-mod
◦I0,T,w,0! '
' Dκ′
FlthickG
-mod(T,w,0),I∞! .
The sought-for functor Φthin→thick is given by (9.1) by inserting the object
!(OBunss,genG (P1,I0,I∞)) ∈ Dκ′
BunG(P1,I0,I∞)-mod! ' Dκ′
BunG(X,0,I∞)-modI0
! .
9.1.5. Note that the functor Φthin→thick admits also the following interpretation. We claimthat there exists a canonical map
FlthinGr→ FlthickG ,
compatible with the twistings. Here are two (of course, equivalent) ways to view it:
By definiton, FlthinG classifies the data (PG, β, α0), where PG is a principal G-bundle on P1,β is its trivialization on P1 − 0, and α is a reduction of (PG)0 to N .
By definition, FlthickG classifies the data (PG, γ, α0), where PG is a principal G-bundle on P1,β is its trivialization on the formal neighborhood D∞ of ∞ ∈ P1, and α is a reduction of (PG)0
to N .
The map r acts identically on the data of (PG, α0), and attaches to a datum of β its restrictionto D∞ ⊂ P1 − 0, twisted by w0.
The group-theoretic interpretation (although non-rigorous) is as follows:
FlthinG ' G(K0)/◦I0
id×w0↪→ G[t, t−1]\
(G(K0)/
◦I0 ×G(K∞)
)' G′out∞\G(K∞) ' FlthickG ,
where Gout∞ = G[t], and G′out∞ ⊂ Gout∞ is the sub-group of maps whose value at 0 belongsto N .
Let p denote the canonical projection
FlthickG → FlthickG /I∞ ' BunG(P1,◦I0, I∞).
Let q denote the composition
p ◦ r : FlthinG → BunG(P1,◦I0, I∞).
Since FlthinG is ind-proper, it is easy to see that the functor Φthin→thick is canonically iso-morphic to q!.
40 S. ARKHIPOV AND D. GAITSGORY
9.1.6. Let us show that Φthin→thick is an equivalence.
First, it is easy to see that Φthin→thick sends
(DκFlthinG
-modI0,(T,w,0))c → (Dκ′
FlthickG
-mod(T,w,0),I∞! )c.
Secondly, from the affine Bruhat decomposition, it is easy to see that the compact generators
of (Dκ′
FlthickG
-mod(T,w,0),I∞! are in the essential image of (Dκ
FlthinG
-modI0,(T,w,0))c.
Hence, it remains to show that Φthin→thick is fully faithful on (DκFlthinG
-modI0,(T,w,0))c. For
that end, it is enough to show that the corresponding functor
. In the latter case, the required assertion reduces to the
statement that for F ∈ DκFl thinG
-modI0 , the map
(9.2) Hδ1FlthinG
(FlthinG ,F)→ Hc(Flthin,genG ,F)
is an isomorphism, where Flthin,genG ⊂ FlthinG is the big affine Bruhat cell. The fact that (9.2) isanother instance of Sect 5.1.13.
9.2. The basic commutative diagram.
9.2.1. Assume again that κ is negative. Recall the functor
ΓT,w : DκFlthinG
-mod(T,w,0) → gκ,0-mod.
It admits a right adjoint, denoted Locgκ,0,FlthinG.
The following theorem is due to Kashiwara-Tanisaki (see, e.g., [FG3] for a proof):
Theorem 9.2.2. The functor
ΓT,w : DκFlthinG
-modI0,(T,w,0) → gκ,0-modI0
realizes the former category as a direct summand of the latter.
The right adjoint to the functor in Theorem 9.2.2 is denoted LocI0,(T,w,0)
gκ,0,FlthinG
. It is isomorphic
to the composition of
LocI0gκ,0,FlthinG
: gκ,0-modI0 → DκFlthinG
-modI0 ,
followed by the functor
DκFlthinG
-modI0 → DκFlthinG
-modI0,(T,w,0)
defined as the projection on the corresponding direct summand, in complete analogy withSect 8.2.1.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 41
9.2.3. The goal of this section is to prove the following theorem:
Theorem 9.2.4. We have a commutative diagram of functors:
DκFlthinG
-modI0,(T,w,0) Φthin→thick−−−−−−−−→ Dκ′
FlthickG
-mod(T,w,0),I∞!
LocI0,(T,w,0)
gκ,0,FlthinG
x Loc(T,w,0),I∞gκ,0,Flthick
G,!
xgκ,0-modI0
Φ0→∞κ→κ′−−−−→ gκ′,∞-modI∞ .
Of course, as the horizontal arrows appearing in the theorem are equivalences, its statementis equivalent to the following one:
Theorem 9.2.5. We have a commutative diagram of functors:
DκFlthinG
-modI0,(T,w,0) Ψthick→thin←−−−−−−−− Dκ′
FlthickG
-mod(T,w,0),I∞!
LocI0,(T,w,0)
gκ,0,FlthinG
x Loc(T,w,0),I∞gκ,0,Flthick
G,!
xgκ,0-modI0
Ψ∞→0κ′→κ←−−−− gκ′,∞-modI∞ ,
where Ψthick→thin is the inverse functor to Φthin→thick.
9.2.6. By Theorem 9.2.2, the functor LocI0,(T,w,0)
gκ,0,FlthinG
is also the left adjoint to
ΓT,w : DκFlthinG
-modI0,(T,w,0) → gκ,0-modI0 .
Combining this with Theorems 9.2.4 and 9.2.5, we obtain that the following:
Corollary 9.2.7. We have the following commutative diagrams of functors:
DκFlthinG
-modI0,(T,w,0) Φthin→thick−−−−−−−−→ Dκ′
FlthickG
-mod(T,w,0),I∞!
ΓT,w(FlthinG ,−)
y ΓT,w(FlthickG ,−)
ygκ,0-modI0
Φ0→∞κ→κ′−−−−→ gκ′,∞-modI∞ .
and
DκFlthinG
-modI0,(T,w,0) Ψthick→thin←−−−−−−−− Dκ′
FlthickG
-mod(T,w,0),I∞!
ΓT,w(FlthinG ,−)
y ΓT,w(FlthickG ,−)
ygκ,0-modI0
Ψ∞→0κ′→κ←−−−− gκ′,∞-modI∞ .
The above corollary theorem should be regarded as an affine analog of Corollary 1.4.6.
Remark. Note that the first commutative diagram of Corollary 9.2.7 follows tautologically fromConjecture 7.1.5. In its turn, the commutativity of this diagram would imply Theorems 9.2.4and 9.2.5 by Theorem 9.2.2.
9.2.8. Proof of Theorem 8.2.6. This follows as a combination of Corollary 9.2.7 and Theo-rem 9.2.2.
9.3. Proof of Theorem 9.2.5.
42 S. ARKHIPOV AND D. GAITSGORY
9.3.1. Let us describe the functor Ψthick→thin explicitly. As Φthin→thick is an equivalence,Ψthick→thin equals its right adjoint.
Here is a geometric description of Ψthick→thin in terms of Sect 9.1.5. It equals the compositionof
q! : Dκ′
FlthickG
-mod(T,w,0),I∞! → Dκ
FlthinG
-modT,w,0,
followed by the averaging functor
AvI0 : DκFlthinG
-modT,w,0 → DκFlthinG
-modI0,(T,w,0),
right adjoint to the forgetful functor
DκFlthinG
-modI0,(T,w,0) → DκFlthinG
-modT,w,0.
However, for the purpose of proof of Theorem 9.2.5, we shall need a description of the functor
Ψthick→thin in terms of the action of G(K0) on the category Dκ′
BunG(X,0,I∞)-mod! as in Sect 9.1.4.
9.3.2. Consider the following general paradigm. Let C be a category acted on by G(K0) at level
κ′. For two open-compact subgroups K ′,K ′′ ⊂ G(K0) we have the functors
ConvK′,K′′ : Dκ′
G(K0)-modK
′,K′′ ⊗ CK′′→ CK
′
and
co-ConvK′,K′′ : CK
′→ CK
′′⊗Dκ′
G(K0)-modK
′,K′′ ' CK′′⊗Dκ′
G(K0)/K′′-modK
′.
Let C∨ denote the category of all (co-limit preserving) functors C→ Vect, whether or not C
is dualizable. For F∨ ∈ (CK′′)∨ consider the functor
co-ConvK′,K′′
F∨ : CK′→ Dκ′
G(K0)/K′′-modK
′
equal to the composition of co-ConvK′,K′′ , followed by the functor
CK′′⊗Dκ′
G(K0)/K′′-modK
′ F∨(−)⊗Id−→ Dκ′
G(K0)/K′′-modK
′.
In what follows we will need the following properties of these functors. First, let A : C1 → C2
be a functor between categories acted on by G(K0), and let A∨ : C∨2 → C∨1 be its dual.
Lemma 9.3.3. For F∨ ∈ (CK′′
2 )∨ the composition
CK′′
1A→ CK
′′
2
co-ConvK′,K′′
F∨−→ Dκ′
G(K0)/K′′-modK
′
is canonically isomorphic to
co-ConvK′,K′′
A∨(F∨) : CK′′
1 → Dκ′
G(K0)/K′′-modK
′.
Secondly, let us assume that G(K0)/K ′′ and ind-proper, and that F is a compact object of
CK′′. Let F∨ be the corresponding object of C∨, i.e., Hom(F,−).
Lemma 9.3.4. Under the above circumstances, the functor ConvK′,K′′(−,F) sends compact
objects to compact ones, and its right adjoint is given by co-ConvK′,K′′
F∨ .
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 43
9.3.5. Returning to our situation, we obtain that the functor Ψthick→thin is given by
co-Conv(◦I0,T,w,0),I0∗(OBun
ss,genG
(P1,I0,I∞)),
which maps
Dκ′
FlthickG
-mod(T,w,0),I∞! ' Dκ′
BunG(X,◦I0,I∞)
-modT,w,0! ' Dκ′
BunG(X,0,I∞)-mod
◦I0,T,w,0! →
→ Dκ′
G(K0)/I0-mod
◦I0,T,w,0 ' Dκ′
FlthinG-mod
◦I0,T,w,0 ' Dκ
FlthinG
-modI0,(T,w,0).
By Theorem 7.1.4, the assertion of Theorem 9.2.5 reduces to the commutativity of thefollowing diagram of functors:
DκFlthinG
-modI0,(T,w,0)
co-Conv(◦I0,T,w,0),I0∗(OBun
ss,genG
(P1,I0,I∞))
←−−−−−−−−−−−−−−−−−−−− Dκ′
FlthickG
-mod(T,w,0),I∞!xLoc
I0,(T,w,0)
gκ,0,FlthinG
Loc(T,w,0),I∞gκ,0,Flthick
G,!
xgκ,0-modI0
〈Γ(
BunG(P1,0,∞),∗(OBunss,genG
(P1,I0,I∞)))∗,−〉g∞
←−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− gκ′,∞-modI∞ .
However, it is easy to see that it is sufficient to check the commutativity of the followingdiagram instead:
(9.3)
DκFlthinG
-modI0,(T,w)
co-Conv(◦I0,T,w,0),I0∗(OBun
ss,genG
(P1,I0,I∞))
←−−−−−−−−−−−−−−−−−−−− Dκ′
FlthickG
-mod(T,w),I∞!xLoc
I0,(T,w)
gκ,0,FlthinG
Loc(T,w),I∞gκ,0,Flthick
G,!
xgκ,0-modI0
〈Γ(
BunG(P1,0,∞),∗(OBunss,genG
(P1,I0,I∞)))∗,−〉g∞
←−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−− gκ′,∞-modI∞ .
9.3.6. Note that the counter-clockwise circuit in (9.3) is the composition
gκ′,∞-modI∞
Indgκ′,0
Lie(◦I0)
(k)⊗−
−→ gκ′,0-modI0 ⊗ gκ′,∞-modI∞Loc
I0,I∞gκ′,0,∞,BunG(X,0,∞),!
−→
−→ Dκ′
BunG(X,0,I∞)-mod
◦I0,T,w!
co-Conv(◦I0,T,w,0),I0∗(OBun
ss,genG
(P1,I0,I∞))
−→ DκFlthinG
-modI0,(T,w).
By Lemma 9.3.3, this can be rewritten as
gκ′,∞-modI∞
Indgκ′,0
Lie(◦I0)
(k)⊗−
−→ gκ′,0-modI0 ⊗ gκ′,∞-modI∞ −→
co-Conv(◦I0,T,w,0),I0
Γ
(BunG(P1,0,∞),∗(OBun
ss,genG
(P1,I0,I∞))
)−→ Dκ
FlthinG
-modI0,(T,w).
44 S. ARKHIPOV AND D. GAITSGORY
9.3.7. The commutativity of (9.3) follows now from the following observation. Let us return
to the paradigm of Sect 9.3.2. Let C := gκ′,0-mod. Let K ′,K ′′ ⊂ G(K0) be open-compact
subgroups (in our situation, we take K ′ = I0, K ′′ =◦I0).
Consider the functor
gκ,0-modK′′'(gκ′,0-modK
′′)∨→ Dκ′
G(K0)/K′′-modK
′
given by
M 7→ co-ConvK′,K′′
M (Indgκ′,0k′′ (k)).
Lemma 9.3.8. The above functor is canonically isomorphic to the localization functor
LocK′′
gκ,0,G(K0)/K′: gκ,0-modK
′′→ Dκ
G(K0)/K′-modK
′′' Dκ′
G(K0)/K′′-modK
′.
LOCALIZATION AND THE LONG INTERTWINING OPERATOR 45
10. Appendix: DG categories
10.1.
10.1.1.
Lemma 10.1.2.
10.2.
Lemma 10.2.1.
10.2.2.
10.2.3.
10.2.4.
10.3.
10.4.
10.4.1.
Lemma 10.4.2.
10.4.3.
10.4.4.
10.4.5.
10.5.
10.5.1.
10.5.2.
10.5.3.
10.5.4.
10.5.5.
46 S. ARKHIPOV AND D. GAITSGORY
References
[AG] S. Arkhipov, D. Gaitsgory, Differential operators on the loop group via chiral algebras.[BB] A. Beilinson, J. Bernstein, La localisation des D-modules.
[FG1]
[FG2] E. Frenkel, D. Gaitsgory, D-modules on the affine flag scheme and representations of the Kac-Moody Liealgebra.
[FG3]
[DrGa] V. Drinfeld and D. Gaitsgory, Compact generation of the category of D-modules on the stack of G-bundles on a curve, arXiv:1112.2402