DUAL PERFECT BASES AND DUAL PERFECT GRAPHS

BYEONG HOON KAHNG1 , SEOK-JIN KANG2 , MASAKI KASHIWARA3 ,

AND UHI RINN SUH1
arXiv:1405.1820v1 [math.RT] 8 May 2014

Dedicated to Professor Boris Feigin on the occasion of his sixtieth birthday

Abstract. We introduce the notion of dual perfect bases and dual perfect graphs.
We show that every integrable highest weight module Vq (λ) over a quantum gener-
alized Kac-Moody algebra Uq (g) has a dual perfect basis and its dual perfect graph
is isomorphic to the crystal B(λ). We also show that the negative half Uq− (g) has a
dual perfect basis whose dual perfect graph is isomorphic to the crystal B(∞). More
generally, we prove that all the dual perfect graphs of a given dual perfect space
are isomorphic as abstract crystals. Finally, we show that the isomorphism classes
of finitely generated graded projective indecomposable modules over a Khovanov-
Lauda-Rouquier algebra and its cyclotomic quotients form dual perfect bases for
their Grothendieck groups.

Introduction
In [1], Berenstein and Kazhdan introduced the notion of perfect bases for inte-
grable highest weight modules over Kac-Moody algebras. Using the properties of
perfect bases, they obtained Kashiwara’s crystal structure without taking quantum
deformation and crystal limits.
Their work was extended by Kang, Oh and Park to the integrable highest weight
modules Vq (λ) (λ ∈ P + ) over a quantum generalized Kac-Moody algebra Uq (g) and
to the negative half Uq− (g) [7, 8]. It was shown that the upper global bases (or dual
canonical bases) B(λ) and B(∞) are perfect bases of Vq (λ) and Uq− (g), respectively.
They also showed that all the crystals arising from perfect bases of Vq (λ) and Uq− (g)
are isomorphic to the crystals B(λ) and B(∞), respectively.
The perfect basis theory plays an important role in the categorification of quan-
tum generalized Kac-Moody algebras. To be more precise, let A = Z[q, q −1 ] and let
UA− (g) be the integral form of Uq− (g). Let R be the Khovanov-Lauda-Rouquier alge-
bra associated with the Borcherds-Cartan datum for Uq (g). We denote by Rep(R)
the category of finite-dimensional graded R-modules and let [Rep(R)] denote its
Date: May 8, 2014.
1
This work was supported by NRF Grant # 2014-021261.
2
This work was supported by NRF Grant # 2014-021261 and by NRF Grant # 2013-055408.
3
This work was partially supported by Grant-in-Aid for Scientific Research (B) 22340005, Japan
Society for the Promotion of Science.
1
Grothendieck group. Then it was proved that [Rep(R)] is isomorphic to UA− (g)∨ , the
dual of UA− (g) with respect to a non-degenerate symmetric bilinear form on Uq− (g)
[8, 11, 12, 14]. Moreover, in [8] (see also [13]), it was shown that the isomorphism
classes of finite-dimensional graded irreducible R-modules form a perfect basis B of
Q(q) ⊗Z[q,q−1 ] [Rep(R)]. Thus B has a crystal structure which is isomorphic to B(∞).
Similarly, the cyclotomic Khovanov-Lauda-Rouquier algebra Rλ (λ ∈ P + ) provides
a categorification of Vq (λ) in the sense that [Rep(Rλ )] is isomorphic to VA (λ)∨ , the dual
of the integral form VA (λ) of Vq (λ) ([4, 5]). As in the case with Uq− (g), the isomorphism
classes of finite-dimensional graded irreducible Rλ -modules form a perfect basis Bλ of
Q(q) ⊗Z[q,q−1 ] [Rep(Rλ )] and Bλ has a crystal structure which is isomorphic to B(λ).
On the other hand, let Proj(R) (respectively, Proj(Rλ )) be the the category of
finitely generated graded projective R-modules (respectively, Rλ -modules). Then they
also provide a categorification of Uq− (g) and Vq (λ). That is, in [4, 5, 11, 12, 14], it was
shown that [Proj(R)] (respectively, [Proj(Rλ )]) is isomorphic to UA− (g) (respectively,
VA (λ)). Note that the isomorphism classes of finitely generated graded projective
indecomposable modules form a basis of [Proj(R)] and [Proj(Rλ )], respectively. To
describe their properties, we need the dual notion of perfect bases.
When the Cartan datum is symmetric, the isomorphism classes of finite-dimensional
graded irreducible modules correspond to Kashiwara’s upper global basis and the
isomorphism classes of finitely generated graded projective indecomposable modules
correspond to Kashiwara’s lower global basis(or Lusztig’s canonical basis) under the
categorification ([6, 15, 16]). However, when the Cartan datum is not symmetric, the
above statement is not true in general.
Thus it is an interesting problem to characterize the bases of Uq− (g) and Vq (λ)
that correspond to the isomorphism classes of finite-dimensional graded irreducible
modules and finitely generated graded projective indecomposable modules over R and
Rλ .
In this paper, as the first step toward this direction, we introduce the notion of
dual perfect bases and dual perfect graphs. The typical examples of dual perfect bases
are the lower global bases B(∞) and B(λ) of Uq− (g) and Vq (λ), receptively. It is
straightforward to verify that that their dual perfect graphs are isomorphic to the
crystals B(∞) and B(λ).
More generally, we show that all the dual perfect graphs of a given dual perfect
space are isomorphic as abstract crystals. Thus every dual perfect graph of Uq− (g) is
isomorphic to B(∞) and the same statement holds for Vq (λ).
Finally, we show that the dual basis of a perfect basis is a dual perfect basis.
Therefore, the isomorphism classes of finitely generated graded projective indecom-
posable modules form a dual perfect basis of Q(q) ⊗Z[q,q−1 ] [Proj(R)] and Q(q) ⊗Z[q,q−1 ]
[Proj(Rλ )], respectively.

2
 Acknowledgments. The second author would like to express his sincere gratitude
to RIMS, Kyoto University for their warm hospitality during his visit in March 2014.

1. Quantum generalized Kac-Moody algebras

We first recall the basic theory of quantum generalized Kac-Moody algebras. Let
I be an index set. An integral square matrix A = (aij )i,j∈I is called a symmetrizable
Borcherds-Cartan matrix if (i) aii = 2 or aii ≤ 0 for all i ∈ I, (ii) aij ≤ 0 for i 6= j, (iii)
aij = 0 if and only if aji = 0, (iv) there is a diagonal matrix D = diag(si ∈ Z>0 )i∈I
such that DA is symmetric.
An index i is real if aii = 2 and is imaginary if aii ≤ 0. We write I re := {i ∈ I |
aii = 2} and I im := {i ∈ I | aii ≤ 0}. In this paper, we assume that

aii ∈ 2Z for all i ∈ I.

A quadruple (A, P, Π, Π∨) consisting of

(a) a symmetrizable Borcherds-Cartan matrix A = (aij )i,j∈I ,
(b) a free abelian group P , which is called the weight lattice,
(c) the set of simple roots Π = {αi ∈ P | i ∈ I},
(d) the set of simple coroots Π∨ = {hi ∈ P ∨ | i ∈ I} ⊂ P ∨ := Hom(P, Z)
is called a Borcherds-Cartan datum if it satisfies the following properties:
(1) hhi , aj i = aij for i, j ∈ I,
(2) for any i, j ∈ I, there is Λi ∈ P such that hhj , Λi i = δij ,
(3) Π is a linearly independent set.
The subset P + := {λ ∈ P | λ(hi ) ∈LZ≥0 , i ∈ I} ⊂ P is called the set of dominant
integral weights. We denote by Q := i∈I Zαi the root lattice and denote by Q+ =
P ∨
i∈I Z≥0 αi the positive root lattice. We also call h:=Q⊗Z P the Cartan subalgebra.

We use the notation

 
q n − qi−n m1 [m1 ]i !
(1.1) [n]i := i , [m]i ! := [m]i [m − 1]i · · · [1]i , := ,
qi − qi−1 m2 i [m2 ]i ![m1 − m2 ]i !

where qi = q si for i ∈ I and [0]q ! := 1.

Definition 1.1. The quantum generalized Kac-Moody algebra Uq (g) associated with a
Borcherds-Cartan datum (A, P, Π, Π∨) is the associative algebra over Q(q) with unity
generated by ei , fi (i ∈ I) and q h (h ∈ P ∨ ) subject to the following defining relations:
′ ′
(1) q 0 = 1, q h q h = q h+h ,
(2) q h ei q −h = q αi (h) ei , q h fi q −h = q −αi (h) fi ,
Ki − Ki−1
(3) ei fj − fj ei = δij , where Ki = q si hi ,
qi − q−1
3
 1−aij
X  
k 1 − aij 1−aij −k
(4) (−1) ei ej eki = 0 if aii = 2 and i 6= j,
k i
k=0
1−aij
X  
1 − aij 1−aij −k
(5) (−1)k fi fj fik = 0 if aii = 2 and i 6= j,
k i
k=0
(6) ei ej − ej ei = 0, fi fj − fj fi = 0 if aij = 0.
For k ∈ Z>0 , let
 k  k
 ei re
if i ∈ I ,  fi if i ∈ I re ,
(k) (k)
ei = [k]i ! fi = [k]i !
 k  k
ei if i ∈ I im , fi if i ∈ I im .
Let Uq− (g) and Uq+ (g) be subalgebras of Uq (g) generated by fi ’s and ei ’s (i ∈ I),
respectively, and let Uq0 be the subalgebra of Uq (g) generated by q h ’s (h ∈ P ∨). Then
the algebra Uq (g) has the triangular decomposition:
Uq (g) = Uq− (g) ⊗ Uq0 (g) ⊗ Uq+ (g).
Fix i ∈ I. For any P ∈ Uq− (g), there exist unique elements Q, R ∈ Uq− (g) such that
Ki Q − Ki−1 R
ei P − P ei = .
qi − qi−1
We define the endomorphisms e′i , e′′i : Uq− (g) → Uq− (g) by
e′i (P ) = R, e′′i (P ) = Q.
Then we have the following commutation relations:
−aij
e′i fj = δij + qi fj e′i for i, j ∈ I.
As we can see in [3, 9], there exists a unique non-degenerate symmetric bilinear
form ( , ) on Uq− (g) satisfying
(1.2) (1, 1) = 1, (fi P, Q) = (P, e′i Q) for all P, Q ∈ Uq− (g).

We now turn to the representation theory of Uq (g). A Uq (g)-module V is called a

highest weight module with highest weight λ if there exists a nonzero vector vλ ∈ V
such that
(1) ei vλ = 0 for i ∈ I, (2) q h vλ = q λ(h) vλ for h ∈ P ∨ , (3) Uq (g)vλ = V.
The vector vλ is called a highest weight vector. For each λ ∈ P , there exists a unique
irreducible highest weight module Vq (λ) up to an isomorphism.
Consider the anti-involution φ on Uq (g) given by
q h 7→ q h , ei 7→ fi , fi 7→ ei for i ∈ I, h ∈ P ∨ .
4
Then there exists a unique non-degenerate symmetric bilinear form ( , ) on Vq (λ)
given by
(1.3) (vλ , vλ ) = 1, (P u, v) = (u, φ(P )v) for all P ∈ Uq (g) and u, v ∈ Vq (λ).

Definition 1.2. The category Oint consists of Uq (g)-modules V satisfying the follow-
ing properties:
M
(a) V has a weight space decomposition, i.e., V = Vµ , where
µ∈P

Vµ = {v ∈ V | q h v = q µ(h) v for all h ∈ P ∨ },

(b) there exist a finite number of weights λ1 , · · · , λt ∈ P such that
t
[
wt(V ) := {µ ∈ P | Vµ 6= 0} ⊂ (λj − Q+ ),
j=1

(c) if aii = 2, then the action of fi on V is locally nilpotent,

(d) if aii ≤ 0, then µ(hi ) ∈ Z≥0 for every µ ∈ wt(V ),
(e) if aii ≤ 0 and µ(hi ) = 0, then fi (Vµ ) = 0,
(f) if aii ≤ 0 and µ(hi ) = −aii , then ei (Vµ ) = 0.
The following proposition was proved in [3].
Proposition 1.3 ([3]).
(i) The category Oint is semisimple.
(ii) If λ ∈ P + , then Vq (λ) is a simple object of Oint .
(iii) Every simple object of the category Oint has the form Vq (λ) for some λ ∈ P + .

2. Lower crystal bases

Let V be a Uq (g)-module in the category Oint . It is straightforward to verify that
every vector v ∈ V can be uniquely written as
X (k)
v= fi vk ,
k≥0

where vk ∈ ker ei with vk = 0 for k ≫ 0. We define the Kashiwara operators e

ei and
e
fi (i ∈ I) by
N
X N
X
(k−1) e (k+1)
eei v = fi vk , fi v = fi vk .
k=1 k=0
Let
A0 := {f /g ∈ Q(q) | f, g ∈ Q[q], g(0) 6= 0}.
5
Definition 2.1. A free A0 -submodule L of V is called a lower crystal lattice of V if
it satisfies
(1) Q(q)L ⊗A0 L = V,
(2) L = λ∈P Lλ , where Lλ = L ∩ Vλ ,
(3) eei L ⊂ L, fei L ⊂ L for all i ∈ I.

Definition 2.2. Let L be a lower crystal lattice of V and let B be a Q-basis of L qL.
A pair (L, B) is called a lower crystal basis of V if it satisfies
F
(1) B = λ∈P Bλ , where Bλ = B ∩ (Lλ /qLλ ),
(2) eei B ⊂ B ∪ {0}, fei B ⊂ B ∪ {0},
(3) for any b, b′ ∈ B and i ∈ I, we have fei (b) = b′ if and only if e
ei b′ = b.
i
We define the I-colored arrows on B by setting b −→ b′ if and only if fei (b) = b′ . The
I-colored oriented graph (B, →) thus defined is called the crystal graph or simply the
crystal of V .
It is known that every Uq (g)-module in Oint has a lower crystal basis.
Proposition 2.3 ([3, 9]). Let Vq (λ) (λ ∈ P + ) be the irreducible highest weight module
in the category Oint with highest vector vλ . Then V (λ) has a unique lower crystal basis
(L(λ), B(λ)), where
L(λ) = A0 -submodule generated by {fei1 · · · feik vλ | is ∈ I, k ≥ 0},
(2.1)
B(λ) = {fei · · · fei vλ mod qL(λ) | is ∈ I, k ≥ 0} \ {0} ⊂ L(λ)/qL(λ).
1 k

Similarly, using the operator e′i in place of ei , we can develop the lower crystal basis
theory for Uq− (g). In particular, we have the following proposition.
Proposition 2.4 ([3, 9]). Let
L(∞) = A0 -submodule generated by {fei1 · · · feik 1 | is ∈ I, k ≥ 0},
B(∞) = {fei · · · fei 1 mod qL(∞) | is ∈ I, k ≥ 0} ⊂ L(∞)/qL(∞).
1 k

Then the pair (L(∞), B(∞)) is a unique lower crystal basis of Uq− (g).
Motivated by the properties of lower crystal bases, we define the notion of abstract
crystals and crystal morphisms as follows.
Definition 2.5. An abstract crystal is a set B together with the maps wt : B → P ,
ϕi , εi : B → Z ∪ {−∞} and eei , fei : B → B ∪ {0} with the following properties:
(1) ϕi (b) = εi (b) + hhi , wt(b)i .
(2) wt(eei b) = wt(b) + αi , wt(fei b) = wt(b) − αi .
(3) b = eei b′ if and only if fei b = b′ , where b, b′ ∈ B, i ∈ I.
(4) If ϕi (b) = −∞, then eei b = fei b = 0.
6
 (5) If b ∈ B and eei b ∈ B, then
( (
εi (b) − 1 if i ∈ I re , ϕi (b) + 1 if i ∈ I re ,
εi (e
ei b) = im
ϕi (e
ei b) =
εi (b) if i ∈ I , ϕi (b) + aii if i ∈ I im .

(6) If b ∈ B and fei b ∈ B, then

( (
re
ε (b) + 1 if i ∈ I , ϕi (b) − 1 if i ∈ I re ,
εi (fei b) = ϕi (fei b) =
i
im
εi (b) if i ∈ I , ϕi (b) − aii if i ∈ I im .

Definition 2.6. Let B1 and B2 be abstract crystals. A crystal morphism between

B1 and B2 is a map ψe : B1 → B2 ⊔ {0} with the following properties:
e ∈ B2 and i ∈ I, we have wt(ψ(b))
(1) For b ∈ B1 , ψ(b) e e
= wt(b), εi (ψ(b)) = εi (b)
e
and ϕi (ψ(b)) = ϕi (b).
e
(2) Suppose b, b′ ∈ B1 and ψ(b), e ′ ) ∈ B2 . If fei b = b′ , then fei (ψ(b))
ψ(b e e ′ ) and
= ψ(b
e )) = ψ(b)
ei (ψ(b
e ′ e for i ∈ I.

3. Lower global bases

Let V be a vector space over Q(q) and let
A := Q[q, q −1 ], A∞ := {f ∈ Q(q) | f is regular at q = ∞}.
Let LA (respectively, L0 and L∞ ) be a free A-submodule (respectively, free A0 -
submodule and free A∞ -submodule) of V such that
Q(q) ⊗A LA ∼ = Q(q) ⊗A∞ L∞ ∼
= Q(q) ⊗A0 L0 ∼ = V.
Definition 3.1.
(1) The triple (LA , L0 , L∞ ) is called
 a balanced triple of V if the canonical projec-
tion π : LA ∩ L0 ∩ L∞ → L0 qL0 is an isomorphism. 
(2) Assume that (L  A , L0 , L∞ ) is a balanced triple. Let B be a Q-basis of L0 qL0
and let G : L0 qL0 → LA ∩ L0 ∩ L∞ be the inverse of π. Then the set B :=
{G(b) | b ∈ B} is called the lower global basis of V corresponding to B.
Consider the Q-algebra automorphism − : Uq (g) → Uq (g) given by
q = q −1 , ei = ei , fi = fi , q h = q −h .
Define an involution − of Vq (λ) by
P vλ = P vλ ,
where P ∈ Uq (g) and vλ ∈ Vq (λ) is the highest weight vector.
7
 Let (L(λ), B(λ)) be the lower crystal basis of Vq (λ) given in Proposition 2.3 and
(n)
let UA− (g) be the A-subalgebra of Uq− (g) generated by fi (i ∈ I, n ≥ 0). Set
VA (λ) := UA− (g)vλ .
Then (VA (λ), L(λ), L(λ)) is a balanced triple of Vq (λ). Since B(λ) is a Q-basis of
L(λ) qL(λ), we obtain the lower global basis B(λ) = {G(b) | b ∈ B(λ)} of Vq (λ).
Similarly, (UA− (g), L(∞), L(∞)) is a balanced triple of Uq− (g) and we get the lower
global basis B(∞) = {G(b) | b ∈ B(∞)} of Uq− (g).
The lower global bases satisfy the following properties.
Proposition 3.2 ([2, 3, 9]).
(i) For any b ∈ B(λ) with λ ∈ P + , G(b) is a unique element in VA (λ) ∩ L(λ) such
that
G(b) = G(b), G(b) ≡ b mod qL(λ).
(ii) For any b ∈ B(∞), G(b) is a unique element in UA− (g) ∩ L(∞) such that
G(b) = G(b), G(b) ≡ b mod qL(∞).
Proposition 3.3 ([10]). The lower global bases B(λ) and B(∞) satisfy the following
properties.
(i) For any n ∈ Z≥0 and b ∈ B(λ) (respectively, b ∈ B(∞)), the subset {G(b) |
P (k)
εi (b) ≥ n} of B(λ) (respectively, of B(∞)) is an A-basis of k≥n fi VA (λ)
P (k)
(respectively, of k≥n fi UA− (g)).
(ii) For any i ∈ I and b ∈ B(λ) (respectively, b ∈ B(∞)), we have
( P
[1 + εi (b)]i G(fei b) + b′ Fb,b
i ′
′ G(b ) if i ∈ I re ,
fi G(b) =
G(fei b) if i ∈ I im ,
where the sum ranges over b′ such that εi (b′ ) > 1 + εi (b) and Fb,b
i
′ ∈ A.

4. Dual perfect bases

In this section, we introduce the notion of dual perfect bases. Fix a Borcherds-
Cartan datum (A, P, Π, Π∨ ) and let k be a field.
L
Definition 4.1. Let V = µ∈P Vµ be a P -graded k-vector space and let {fi }i∈I be
a family of endomorphisms of V . We say that (V, {fi }i∈I ) is a pre-dual perfect space
if it satisfies the following conditions.
S
(1) There exist finitely many elements λ1 , . . . , λk ∈ P such that wt(V ) ⊂ kj=1 (λj −
Q+ ).
(2) fi (Vµ ) ⊂ Vµ−αi for any i ∈ I and µ ∈ P .
8
 For each i ∈ I and v ∈ V \ {0}, define ℓi (v) to be the non-negative integer n such
that v ∈ fin V \ fin+1 V .
Definition 4.2. Let (V, {fi }i∈I ) be a pre-dual perfect space.
(i) A basis BFof V is called a dual perfect basis if
(a) B = µ∈P Bµ , where Bµ = B ∩ Vµ ,
(b) For any i ∈ I, there exists a map f i : B → B ∪ {0} such that for any b ∈ B,
there exists c ∈ k× satisfying
ℓ (b)+2
fi (b) − c fi (b) ∈ fi i V.
(c) If fi (b) = fi (b′ ) 6= 0, then b = b′ .
(ii) V is called a dual perfect space if it has a dual perfect basis.

Proposition 4.3. Every Uq (g)-module in Oint has a dual perfect basis.

Proof. It suffices to show that every irreducible highest weight module Vq (λ) (λ ∈ P + )
has a dual perfect basis. Let B(λ) = {G(b) | b ∈ B(λ)} be the lower global basis of
Vq (λ). Define
fi G(b) := G(fei b) for b ∈ B(λ).
Then by Proposition 3.3, B(λ) is a dual perfect basis. 

Let B be a dual perfect basis of a P -graded k-vector space V .

Lemma 4.4. Let i ∈ I.
(i) For any b ∈ B and n ∈ Z≥0 , there exists c ∈ k× such that
n+ℓi (b)+1
(4.1) fin b − c fin (b) ∈ fi V.
(ii) For any i ∈ I and n ∈ Z≥0 , we have
L
fin V = k b.
b∈f n
i B

(iii) For any b ∈ B, we have

ℓi (b) = max{n ∈ Z≥0 | b ∈ fin B}.
(iv) For any b ∈ B such that fi b 6= 0, we have
ℓi (fi b) = ℓi (b) + 1.
(v) For any n ∈ Z≥0 , the image of {b ∈ B | ℓi (b) = n} is a basis of fin V /fin+1V .
Proof. (i) By the definition, we have ℓi (f i b) ≥ ℓi (b) + 1 for any b ∈ B such that
f i b ∈ B. Hence we have
ℓi (f in b) ≥ ℓi (b) + n for any b ∈ B and n ∈ Z≥0 such that f ni b ∈ B.
9
We shall show (i) by induction on n. If n = 0, 1, (i) is obvious. Assume n > 1. Then
n+ℓ (b)
the induction hypothesis implies that fin−1 b − cf in−1b ∈ fi i V for some c ∈ k× .
n+ℓ (b)+1
Therefore, we have fin b − cfi f in−1 b ⊂ fi i V . Hence, if f in−1 b = 0, then we obtain
ℓ (f n−1 b)+2 n−1+ℓ (b)+2
(4.1). If f in−1 b ∈ B, then we have fi f in−1 b − c′ f ni b ∈ fi i i V ⊂ fi i
V for
′ ×
some c ∈ k . Hence we obtain (i).
L
(ii) By (i), we have f ni b ∈ fin V . Hence it is enough to show that fin V ⊂ k b. We
b∈f n
iB
have L
fin V ⊂ k b + fin+1 V,
b∈f n
iB

which easily follows from (i). Then (ii) follows from (fik V )λ = 0 for any λ ∈ P and
k ≫ 0.
(iii), (iv) and (v) easily follow from (ii). 
Let B be a dual perfect basis of a P -graded k-vector space V . For b ∈ B, we define
(
b′ if b′ ∈ B satisfies fi b′ = b,
ei b =
0 if there exists no b′ ∈ B such that fi b′ = b.
We also define the maps εi , ϕi : B → Z ∪ {−∞} by
(
ℓi (b) if i ∈ I re ,
εi (b) =
0 if i ∈ I im ,
ϕi (b) = εi (b) + hhi , wt(b)i ,
where the map wt : B → P is given by wt(b) = µ if b ∈ Bµ .
Then it is straightforward to verify that (B, ei , fi , εi , ϕi , wt) is an abstract crystal,
which will be called the dual perfect graph.
Proposition 4.5. Let B(λ) be the global basis of V (λ) (λ ∈ P + ). Then B(λ) is
isomorphic to B(λ) as an abstract crystal.
Proof. Recall that B(λ) becomes a dual perfect basis by defining fi G(b) = G(fei b) for
b ∈ B(λ). Hence for b, b′ ∈ B(λ), we have fei b = b′ if and only if G(b′ ) = G(fei b) =
fi G(b), which proves our claim. 

Remark 4.6. Let g be the generalized Kac-Moody algebra associated with a Borcherds-
Cartan datum. By taking the classical limit in Proposition 3.3, it follows that every
irreducible highest weight g-module V (λ) (λ ∈ P + ) has a dual perfect basis whose
dual perfect graph is isomorphic to B(λ) as an abstract crystal.

Proposition 4.7.
(i) The algebra Uq− (g) has a dual perfect basis.
10
 (ii) The lower global basis B(∞) is isomorphic to B(∞) as an abstract crystal.
Proof. As in Proposition 4.3 and Proposition 4.5, the lower global basis B(∞) satisfies
our assertions. 

5. Properties of dual perfect bases

Let k be a field, and B be a dual perfect basis of (V, {fi }i∈I ).
ℓi (b)
For b ∈ B and i ∈ I, we set e top i (b) = e i b ∈ B. More generally, for a sequence
i = (i1 , · · · , im ) ∈ I m (m ≥ 1), we set
e top top top
i b := e im · · · e i1 b.

We also use the notations

e il b := e lim
m
· · · e li11 b ∈ B ⊔ {0},
f il b := f li11 · · · f lim
m
b ∈ B ⊔ {0}
for l = (l1 , . . . , lm ) ∈ Zm
≥0 and set

f il := fil11 · · · filmm .
We say that a sequence i = (ik )k≥1 is good if {k ∈ Z>0 | ik = i} is an infinite set for
any i ∈ I. We say that a sequence l = (lk )k≥1 of non-negative integers is good if lk = 0
for k ≫ 0. For a good sequence i = (ik )k≥1 in I and a good sequence l = (lk )k≥1 of
non-negative integers, we set
X lk−1 1+lk
V >l, i = fil11 · · · fik−1 fik V,
k≥1
m
X
≥l, i lk−1 1+lk
V =V >l, i
+ fil11 · · · filmm V = fil11 · · · fik−1 fik V + fil11 · · · filmm V
k=1

for m ≫ 0.
Let ≺ be the lexicographic ordering on good sequences of integers, namely, l =
(lk )k≥1 ≺ l′ = (lk′ )k≥1 if and only if there exists k ≥ 1 such that ls = ls′ for any s with
1 ≤ s < k and lk < lk′ . Then for l = (lk )k≥1 and l′ = (lk′ )k≥1, we have
′
V ≥l, i ⊂ V ≥l , i if l′  l,
′
V >l, i ⊂ V ≥l , i if l′ ≺ l.
For any v ∈ V \{0} and a good sequence i = (ik )k≥1 in I, we define L(i, v) = (lk )k≥1
to be the largest sequence l = (lk )k≥1 such that v ∈ V ≥l . Note that such an L(i, v)
exists and is a good sequence (see Proposition 5.1 (ii)). Hence v ∈ V ≥l, i if and only
if l  L(i, v).
11
 Set BH = {b ∈ B | ℓi (b) = 0 for all i ∈ I}. For a good sequence i = (ik )k≥1, we set
e top
i = e top top
im · · · e i1 b for m ≫ 0. Note that it does not depend on m ≫ 0 and belongs
to BH (see Proposition 5.1 (ii)).
Proposition 5.1. Let i = (ik )k≥1 be a good sequence in I.
(i) For any b ∈ B and a sequence l = (l1 , . . . , lm ) of non-negative integers, there
exists c ∈ k× such that
m
X
fil11 · · · filmm b −c f li11 ··· f lim
m
b ∈ fil11 · · · fi1+l
k
k
V.
k=1

(ii) For each b ∈ B, define a sequence (bk )k≥0 by

b0 = b, bk = e top
ik bk−1 for k ≥ 1,

and let (Lk )k≥1 be the sequence of non-negative integers given by

Lk = lik (bk−1 ) for k ≥ 1.

Then we have
(a) (Lk )k≥1 is a good sequence,
(b) bk ∈ BH for k ≫ 0,
(c) L(i, b) = (Lk )k≥1 .
(iii) For any good sequence l = (lk )k≥1 of non-negative integers, we have
X
(5.1) V ≥l, i = k b,
{b∈B | lL(i,b)}
X
(5.2) V >l, i = k b.
{b∈B | l≺L(i,b)}

(iv) For any good sequence l = (lk )k≥1 of non-negative integers, we have an injective
map

(5.3) e l : {b ∈ B | L(i, b) = l} ֌ BH .

Proof. Let us first prove (i) by induction on m. The m = 1 case follows from
Lemma 4.4 (i). Assume that m > 1. Set b1 = f lim m
b. Then applying the induc-
×
tion hypothesis to b1 , there exists c ∈ k such that
m−1
X
lm−1 l
fil11 · · · fim−1 b1 −c f li11 ··· f im−1 b
m−1 1
∈ fil11 · · · fi1+l
k
k
V.
k=1
12
By Lemma 4.4 (i), there exists c′ ∈ k× such that filmm b − c′ b1 ∈ fi1+l m
m
V . Hence we
have
lm−1
fil11 · · · filmm b − cc′ f li11 · · · f lim
m
b = fil11 · · · fim−1 (filmm b − c′ b1 )
Xm
lm−1 lm−1
′ l1
+ c (fi1 · · · fim−1 b1 − cf i1 · · · f im−1 b1 ) ∈ fil11 · · · fi1+l
k
k
V.
k=1
P
Next we shall show (ii) (a) and (ii) (b). Since wt(b) − m
k=1 Lk αik ∈ wt(V ) for any
m, we have Lk = 0 for k ≫ 0. Hence bk does not depend on k ≫ 0. Thus ℓik+1 (bk ) = 0
for k ≫ 0 which implies (ii) (b).
To prove (ii) (c) and (iii), let L = (Lk )k≥1 be the sequence in (ii) which is uniquely
determined for each b ∈ B. For m ∈ Z≥0 , set L em (i, b) = (L1 , . . . , Lm ). We first
em (i, b), we have
observe that for any sequence l = (l1 , . . . , lm ) with l  L
m−1
X l k−1 1+lk em (i, b),
(5.4) b∈ fil11 · · · fik−1 fik V + fil11 · · · filmm V if (l1 , . . . , lm )  L
k=1

which immediately follows from (i).

Now we shall show
m−1
X X
lk−1 1+lk
(5.5) fil11 · · · fik−1 fik V + fil11 · · · filmm V = k b.
k=1 e m (i,b)}
{b∈B | (l1 ,...,lm )L

We have already seen that the right hand side is contained in the left hand side. Let
us show the converse inclusion by induction on m. In order to see this, it is enough
to show that X
fil11 · · · filmm V ⊂ k b.
e m (i,b)}
{b∈B | (l1 ,...,lm )L

Set i′ = (i2 , i3 , . . .). Then the induction hypothesis implies

X
fil22 · · · filmm V ⊂ k b.
e m−1 (i′ ,b)}
{b∈B | (l2 ,...,lm )L

Hence we have reduced (5.5) to

X
fil11 b0 ∈ kb
e m (i,b)}
{b∈B | (l1 ,...,lm )L

if (l2 , . . . , lm )  L em−1 (i′ , b0 ), which follows from the fact that f l1 b0 ∈ k f l1 b0 +

i1 i1
P em (i, f l1 b0 ). Thus the proof of (5.5) is complete.
ℓi (b)>l1 kb and (l 1 , . . . , l m )  L i1

Now, (ii) (c) follows from (5.5), and then (5.1) is nothing but (5.5) for m ≫ 0.
Equality (5.2) follows easily from (5.1).
13
 In order to prove (iv), observe that e l (b) = etop l
i (b), where e is the map defined
in (5.3) and L(i, b) = l. Since e has a left inverse fi |BH , we conclude that e l is
l l

injective. 
The following corollary easily follows from the preceding proposition.
Corollary 5.2. Let i = (ik )k≥1 be a good sequence in I and let l = (lk )k≥1 be a good
sequence of non-negative integers. Denote by p i,l : V ≥l,i → V ≥l,i /V >l,i the canonical
projection and set B i,l = {b ∈ B | L(i, b) = l}.
≥l, i
Then the image p i,l (B
i,l ) is a basis of V /V >l, i . Moreover, k× p i,l (B i,l ) is equal
× l
to k p i,l (f i BH ) \ {0} .
Here, for a subset S of a k-vector space V , we use the notation
k× S := {k× s | s ∈ S}.

Note that Proposition 5.1 (iii) with l = (0, 0, . . .) implies that

X L
fi V = kb.
i∈I b∈B\BH

Hence we conclude
P
Lemma 5.3. Set VH :=V /( i∈I fi V ) and let pH : V ։ VH be the canonical projection.
Then pH : BH → VH is injective and pH (BH ) is a basis of VH .

6. Uniqueness of dual perfect graphs

The purpose of this section is to prove that all the dual perfect graphs of a given
dual perfect space are isomorphic as abstract crystals.
Theorem 6.1. Let (V, {fi }i∈I ) be a dual perfect space and let B and B′ be its dual
perfect bases. Assume that pH (BH ) = pH (B′H ).
Then there is a crystal isomorphism ψ : B −∼ → B′ such that pH (b) = pH (ψ(b)) for
all b ∈ BH . Moreover, for any b ∈ B and a good sequence i in I, we have L(i, b) =
L(i, ψ(b)) and pH (e top top
i b) = pH (e i ψ(b)) ∈ pH (BH ).

Proof. For a good sequence i = (ik )k≥1 in I and a good sequence l = (lk )k≥1 , let
p i,l : V ≥l,i → V ≥l,i /V >l,i be the canonical projection. Set B i,l = {b ∈ B | L(i, b) = l}
and define B′i,l in a similar manner. Then by Corollary 5.2 and Lemma 5.3, we have
k× p i,l (B i,l ) = k× p i,l (B′i,l )
and both p i,l (B i,l ) and p i,l (B′i,l ) are bases of V ≥l,i /V >l,i . Hence for any b ∈ B, there
exists b′ ∈ B′ such that
(6.1) L(i, b) = L(i, b′ ) and b − cb′ ∈ V >L(i,b),i for some c ∈ k× .
14
 To prove our claim, it is enough to show that for another choice of a good sequence
′
i , (6.1) holds with the same b′ . Set v = b − cb′ . Since v ∈ V >L(i,b),i , v is a linear
combination of B \ {b}. Set l = L(i′ , b) and l′ = L(i′ , b′ ). Then b ∈ V ≥l,i and
′

′ ′ ′ ′
b′ ∈ V ≥l ,i . Assume that l ≺ l′ . Since b′ ∈ V >l,i , we have v − b ∈ V >l,i . Hence
v − b is a linear combination of B \ {b}, which is a contradiction to the fact that v
is a linear combination of B \ {b}. Hence we conclude that l′  l. Similarly, we have
l  l′ . Hence we obtain l = l′ . It follows that both p i′ ,l (b) and p i′ ,l (b′ ) belong to
k× p i′ ,l (B i,l ). If k× p i′ ,l (b) 6= k× p i′ ,l (b′ ), then v − b is a linear combination of B \ {b},
which is a contradiction. Hence k× p i′ ,l (b) = k× p i′ ,l (b′ ) and our assertion follows. 

Corollary 6.2. Let Uq (g) be the quantum generalized Kac-Moody algebra associated
with a Borcherds-Cartan datum and let Vq (λ) be the irreducible highest weight Uq (g)-
module with λ ∈ P + . Then the following statements hold.
(i) Every dual perfect graph of Vq (λ) is isomorphic to B(λ) as an abstract crystal.
(ii) Every dual perfect graph of Uq− (g) is isomorphic to B(∞) as an abstract crystal.

Remark 6.3. Theorem 6.1 also shows that every dual perfect graph of V (λ) over a
generalized Kac-Moody algebra g is isomorphic to B(λ) as an abstract crystal.

7. perfect bases and dual perfect bases

Now we will prove that the isomorphism classes of finitely generated graded projec-
tive modules over R and Rλ form a dual perfect basis of Q(q) ⊗Q[q,q−1 ] [Proj(R)] and
λ
Q(q) ⊗Q[q,q−1L ] [Proj(R )], respectively. We first recall the definition of perfect basis.
Let V = µ∈P Vµ be a P -graded k-vector space with a family of endomorphisms
{ei }i∈I satisfying the following conditions.
S
(i) There exist finitely many elements λ1 , . . . , λk ∈ P such that wt(V ) ⊂ kj=1(λj −
Q+ ).
(ii) ei (Vµ ) ⊂ Vµ+αi for any i ∈ I and µ ∈ P .
Set ℓ∨i (v) := max{n ∈ Z≥0 | eni v 6= 0} = min{n ∈ Z≥0 | en+1 i v = 0} for v ∈ V \ {0}.
Definition 7.1. A basis B of V is said to be perfect if
L
(1) B = µ∈P Bµ , where Bµ := B ∩ Vµ ,
(2) for any i ∈ I, there exists a map Ei : B → B ∪ {0} such that for any b ∈ B,
we have
(a) if ℓ∨i (b) = 0, then Ei (b) = 0,
(b) if ℓ∨i (b) > 0, then Ei (b) ∈ B and
ℓ∨ (b)−1
ei (b) − cEi (b) ∈ ker ei i for some c ∈ k× ,
(3) if Ei (b) = Ei (b′ ) 6= 0 for b, b′ ∈ B, then b = b′ .
15
 It was shown in [7] and [8] that every Uq (g)-module in the category Oint and the
negative half Uq− (g) have perfect bases.
For a perfect basis B, we define a map Fi : B → B ∪ {0} by
 ′
b if Ei (b′ ) = b,
Fi (b) =
0 if b ∈
/ Ei B.
Let us recall the following lemma.
Lemma 7.2 ([1]). Let B be a perfect basis of (V, {ei }i∈I ). Then
(i) for any n ∈ Z≥0 , we have
L
ker eni = kb.
b∈B, ℓ∨
i (b)<n

(ii) ℓ∨i (Ei b) = ℓ∨i (b) − 1 for any b ∈ B such that ℓ∨i (b) > 0.
ℓ∨ (b)
(iii) for b ∈ B and i ∈ I, we have b ∈ Ei B if and only if b ∈ ei V + ker ei i .
(iv) for any n ∈ Z≥0 and i ∈ I, the image of {b ∈ B | ℓ∨i (b) = n} is a basis of
ker en+1
i / ker eni .
Let (V, {fi }i∈I ) be a pre-dual perfect space such that dim Vλ < ∞ for any λ ∈ P .
We set
L ∨
(7.1) (V ∨ )λ = Homk (Vλ , k) for any λ ∈ P , and V ∨ = (V )λ .
λ∈P

Let h • , • i : V × V ∨ → k be the canonical pairing. We define ei : V ∨ → V ∨ as the

transpose of fi , so that we have hu, ei vi = hfi u, vi for any u ∈ V and v ∈ V ∨ .
Proposition 7.3. Let (V, {fi }i∈I ) be a a pre-dual perfect space such that dim Vλ < ∞
for any λ ∈ P . Let B be a basis of V and let B∨ ⊂ V ∨ be the dual basis of B. Then
(i) B is a dual perfect basis if and only if B∨ is a perfect basis.
(ii) Assume that B is a dual perfect basis. Denoting the canonical isomorphism
B −∼
→ B∨ by b 7→ b∨ , we have

ℓi (b) = ℓ∨i (b∨ ), and (e i b)∨ = Ei (b∨ ) for all b ∈ B and i ∈ I.

Here we understand 0∨ = 0.
L
Proof. Assume first that B∨ is a perfect basis. Then, ker eni = kb∨ .
b∨ ∈B∨ , ℓ∨ (b∨ )<n
n ⊥
Since fin Vcoincides with the orthogonal
L complement (ker ei ) :={u ∈ V | hu, ker eni i =
n n
0} of ker ei , we have fi V = kb. Hence b ∈ fin V if and only if ℓ∨i (b∨ ) ≥ n.
∨
ℓi (b )≥n
∨

Therefore we have ℓi (b) = ℓ∨i (b∨ ), and

L
fin V = kb.
ℓi (b)≥n
16
 We define f i : B → B ∪ {0} by (f i b)∨ = Fi (b∨ ). Let b ∈ B with ℓi (b) = n. We shall
show that fi b − cf i b ∈ fin+2 V for some c ∈ k× .
Recall that the image of B∨n :={b∨ ∈ B∨ | ℓ∨i (b∨ ) = n} forms a basis of ker en+1 i / ker eni .
Since ker en+1
i / ker eni is isomorphic to the dual (fin V /fin+1 V )∨ of fin V /fin+1V , the im-
age of Bn forms a basis of fin V /fin+1V dual to B∨n . By the hypothesis, the map
ei : ker en+2
i / ker en+1
i → ker en+1
i / ker eni gives an injection

Ei : b∨ ∈ B∨n+1 | Ei (b∨ ) 6= 0 → B∨n .
Hence fi : fin V /fin+1 V → fin+1 V /fin+2 V sends {b ∈ Bn | Fi (b∨ ) 6= 0} to Bn+1 up to a
constant multiple, and sends {b ∈ Bn | Fi (b∨ ) = 0} to {0}.
As a conclusion, for any b ∈ Bn , we have fi b ≡ c fi b mod fin+2 V for some c ∈ k× .
Thus B is a dual perfect basis.
The converse can be proved in a similar manner. 
Recall that the Khovanov-Lauda-Rouquier algebra R and its cyclotomic quotient
R provide a categorification of Uq (g) and Vqλ , respectively. That is, we have
λ

= Q(q) ⊗Z[q,q−1 ] [Rep(R)] ∼

Uq− (g) ∼ = Q(q) ⊗Z[q,q−1 ] [Proj(R)],
∼ Q(q) ⊗Z[q,q−1 ] [Rep(Rλ )] ∼
Vq (λ) = = Q(q) ⊗Z[q,q−1 ] [Proj(Rλ )].
It was shown in [8, 13] that the isomorphism classes of finite-dimensional graded ir-
reducible modules form a perfect basis of Q(q) ⊗Z[q,q−1 ] [Rep(R)] and Q(q) ⊗Z[q,q−1 ]
[Rep(Rλ )], respectively. Since the isomorphism classes of finite-dimensional graded ir-
reducible R-modules and those of finitely generated graded projective indecomposable
R-modules are dual to each other with respect to the perfect pairing given by
[Proj(R)] × [Rep(R)] −→ A,
([P ], [M]) 7−→ dimq HomR (P, M),
the following proposition follows immediately.
Proposition 7.4. The isomorphism classes of finitely generated graded projective
indecomposable modules form a dual perfect basis of Q(q) ⊗Z[q,q−1 ] [Proj(R)] and
Q(q) ⊗Z[q,q−1 ] [Proj(Rλ )].

References
[1] A. Berenstein, D. Kazhdan, Geometric and unipotent crystals. II: From unipotent bicrystals to
crystal bases, Contemp. Math. 433 (2007), 13–88.
[2] J. Hong, S.-J. Kang, Introduction to Quantum Groups and Crystal Bases, Graduate Studies in
Mathematics 42, Amer. Math. Soc., Providence, 2002.
[3] K. Jeong, S.-J. Kang, M. Kashiwara Crystal bases for quantum generalized Kac-Moody algebras,
Proc. Lond. Math. Soc. 90 (3) (2005), 395–438.
[4] S.-J. Kang, M. Kashiwara, Categorification of highest weight modules via Khovanov-Lauda-
Rouquier algebras, Invent. Math. 190 (2012), 699–742.
17
 [5] S.-J. Kang, M. Kashiwara, S.-j. Oh, Categorification of highest weight modules over quantum
generalized Kac-Moody algebras, Moscow Math. J. 13 (2013), 315–343.
[6] S.-J. Kang, M. Kashiwara, E. Park, Geometric realization of Khovanov-Lauda-Rouquier algebras
associated with Borcherds-Cartan data, Proc. London Math. Soc. (3) 107 (2013), 907-231.
[7] S.-J. Kang, S.-j. Oh, E. Park, Perfect bases for integrable modules over generalized Kac-Moody
algebras, Algebr. Represent. Theory 14 (2011), 571–587.
[8] , Categorification of quantum generalized Kac-Moody algebras and crystal bases, Int. J.
Math. 23 (2012), 1250116.
[9] , On crystal bases of the q-analogue of universal enveloping algebras, Duke Math. J. 63
(1991), 465–516.
[10] , Global crystal bases of quantum groups, Duke math. J. 69 (1993), 455–485.
[11] M. Khovanov, A. Lauda, A diagrammatic approach to categorification of quantum groups I,
Represent. Theory 13 (2009), 309–347.
[12] , A diagrammatic approach to categorification of quantum groups II, Trans. Amer. Math.
Soc. 363 (2011), 2685–2700.
[13] A. Lauda, M. Vazirani, Crystals from categorified quantum groups, Adv. Math. 228 (2011),
803–861.
[14] R. Rouquier, 2 Kac-Moody algebras, arXiv:0812.5023.
[15] , Quiver Hecke algebras and 2-Lie algebras, arXiv:1112.3619v1.
[16] M. Varagnolo, E. Vasserot, Canonical bases and KLR algebras, J. reine angew. Math. 659
(2011), 67–100.

Department of Mathematical Sciences, Seoul National University, 599 Gwanak-

Ro, Seoul 151-747, Korea
E-mail address: [email protected]

Department of Mathematical Sciences and Research Institute of Mathematics,

Seoul National University, 599 Gwanak-Ro, Seoul 151-747, Korea
E-mail address: [email protected]

Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-

8502, Japan, and Department of Mathematical Sciences, Seoul National University,
599 Gwanak-Ro, Seoul 151-747, Korea
E-mail address: [email protected]

Research Institute of Mathematics, Seoul National University, 599 Gwanak-Ro,

Seoul 151-747, Korea
E-mail address: [email protected]

18