Professional Documents
Culture Documents
On Tensor Categories of Lie Type E N A9: Article in Press
On Tensor Categories of Lie Type E N A9: Article in Press
http://www.elsevier.com/locate/aim
Abstract
Let V be the minuscule representation of the Lie algebra gEN ; N 6; 7: We show that the
centralizer of the action of the quantum group Uq gEN on V #n is generated by the Rmatrices and one additional element, appearing in the N 1st tensor power; a similar
description can be given for the classical limit. An analogous statement is true for a certain
direct summand of V #n for Uq gEN with N45; Na9; here V is the module whose highest
weight is the fundamental weight labeled by the vertex furthest from the triple point of the
graph EN : Moreover, we obtain a 2-parameter family of braid representations generalizing a
q-version of the Brauer centralizer algebras.
r 2002 Elsevier Science (USA). All rights reserved.
0. Introduction
A conceptual description of the centralizer algebras EndG V #n has been known
for the vector representations of the general linear groups (Schur) and of orthogonal
and symplectic groups (Brauer [2]) for some time. This is closely related to
determining all G-invariant polynomial maps on V : This problem has also been
solved for the smallest representation of gG2 and a few more representations (see
[22]) of classical Lie groups. Not that much more seems to be known for other
representations and other Lie types.
In this paper, we determine a minimal set of generators for EndG V #n for the
minuscule representations V for Lie types E6 and E7 : This will be done in the setting
of the DrinfeldJimbo quantum groups U Uq gEN ; and also includes the solution
$
0001-8708/03/$ - see front matter r 2002 Elsevier Science (USA). All rights reserved.
doi:10.1016/S0001-8708(02)00048-8
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
67
for the classical case for q 1: The quantum case turns out to be more manageable
due to the representations of braid groups, via R-matrices, in the centralizer of the
action of U on V #n : We actually consider the corresponding braid action on a path
space. For type A this would be the q-analog of Youngs orthogonal representation
of the symmetric group for Hecke algebras (see [8,27]); similar studies also appeared
in connection with the star triangle equation in connection with statistical mechanics
(see e.g. [9]). These representations are very suitable for determining whether the
image of the braid group generates the whole centralizer; whenever this is not the
case, we get a new generator for EndU V #n : This programme can be fully carried
#n
out for a certain summand Vnew
of V #n ; here V is the fundamental representation of
the KacMoody algebra gEN ; Na9 (or its quantum version) labeled by the vertex
furthest from the triple point of the graph EN : The terminology comes from the fact
that for nite EN (including N 8) it contains all the simple summands of V #n
which have not already appeared in smaller tensor powers; for N49 the term is a bit
of a misnomer as now this summand contains all simple representations which will
not appear in any higher tensor powers of V : At least for these special summands,
one observes a similar uniform decomposition behaviour for our E-series as it is
known for the classical type A and type BCD series. While this is different from the
proposed exceptional series by Vogel and Deligne, it may be helpful for their studies.
Let us now give a more detailed description of the contents of this paper. Roughly
speaking, it contains a combinatorial part and a more algebraic part. As already
mentioned, we study the new part of V #n : This can be considered a version of
Weyls approach of restricting the centralizer algebra of orthogonal groups to
traceless tensors. One of the obvious problems is to determine, for a given dominant
weight l; the smallest number n nl for which the simple highest weight module
Vl appears in V #n : This problem is solved for No9 in Section 2; the same formula
also gives the largest number n for which Vl appears in V #n in the KacMoody case
N49: Moreover, this leads to a uniform labeling of the dominant weights for all
KacMoody types EN ; Na9; which is also done in Section 2. This gives a
#n
combinatorial denition of a summand Vnew
also in the KacMoody case. We also
observe generic patterns in the tensor product rules for these summands. This can be
checked comparatively easily using Littelmanns path formalism.
A second factor in the denition of the generic labeling for dominant weights was
the gradation of a gEN module in terms of gDN1 submodules. One observes
easily that EndU V #n contains as a quotient algebra EndUq gDN1 V% #n ; where V% is
the vector representation of Uq gDN1 : This shows that besides the R-matrices we
need at least one more generator, an analogue of the Pfafan in the N 1st tensor
power. Our main result shows that for N 6; 7 we need no more generators.
To prove this we study the representations of the braid groups on the Littelmann
paths. Here the matrices have a natural block structure coming from the path
formalism. To determine these blocks one uses Drinfelds quantum Casimir in the
form of q-JucysMurphy elements (see also e.g. [14,15,21]). It turns out that in each
of these blocks at least one of the eigenidempotents has rank 1, and that the action of
the braid generator is determined by the action of this idempotent (this was already
ARTICLE IN PRESS
68
observed by Leduc and Ram [15] in the context of the q-Brauer algebra). The
idempotent itself can now, basically, be determined by additional equations coming
from the eigenidempotent property. In particular, the question whether the braid
group generates the whole centralizer depends on whether the diagonal entries of
these idempotents are zero or not; these entries are special cases of q-6j-symbols. The
basic equations and criteria for nonzero diagonal entries are derived in Section 4. It
also contains the proof of the main theorem about E6 and E7 in the quantum case.
Explicit formulas for the diagonal entries are proved in Section 5. The main theorem
for the classical case can then be easily derived from the quantum case. Moreover,
one also obtains a 2-parameter family of braid representations, dened over the eld
Qr; q of rational functions in 2 variables which generalizes the q-Brauer algebra
found in [1,18]. These representations should be useful for further studies of tensor
categories of exceptional Lie type. Some of the possible applications are discussed at
the end of this paper and in [29], where several of the results of this paper were
announced.
1. Preliminaries
Let g be a symmetrizable KacMoody algebra given by a Coxeter graph X with N
vertices, with generators ei and fi ; 0pipN 1: We denote the simple roots by
ai ; i 0; y; N 1: Fix an invariant bilinear form / ; S on h ; and dene a$ i
If all the roots have the same length, we assume / ; S to be normalized such
that a$ i ai : If / ; S is nondegenerate, we dene the fundamental weights Lj by
/$ai ; Lj S dij : If aAh ; the reection sa on h is dened by sa l l /l; a$ Sa:
2
/ai ;ai Sai :
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
69
#n
Endg V #n contains a subalgebra isomorphic to Endg0 VL#n
multVm% VL#n
% 0 for any weight m with levn m 0:
P
(c) If l has level i in V #n and l nj1 oj ; with oj a weight in V ; for j 1; 2; y; n;
then i is equal to the sum of the levels of oj in V :
Proof. If m has level i in V #n ; then so has m7aj for j40: Hence if v is a weight vector
of weight m; v as well as ej v and fj v are in V i: This shows (a).
By denition of level spaces, any level 0 vector is killed by e0 : Hence any g0 highest
weight vector vAV #n 0 is also a highest weight vector with respect to g: Statement
(b) follows from this. Using the notation
the level of a weight o
P levo lev1 o forP
of V ; we obtain i /nL l; L0 S nj1 /L oj ; L0 S
levoj : &
1.2. Littelmann paths
To get more precise information about weight multiplicities and tensor product
rules we will use Littelmann paths (see [16]); Littelmann calls these paths LS-paths in
honor of Lakshmibai and Seshadri. There are other methods available; indeed for
decomposing tensor products of minuscule and adjoint representations of
semisimple Lie algebras (which would sufce for a major part of this paper)
methods already known to Brauer would be sufcient (see e.g. [23, Chapter 7] for
theorems and history).
Let l be a dominant integral weight for g and let Vl be the corresponding
irreducible highest weight module. Then Littelmann showed that there exists a
labelling set for a basis for Vl consisting of paths which can be obtained by applying
certain root operators fa to the straight line from 0 to l; here a is a simple root.
Here is an outline of the denition of fa (see [16] for the precise denition). If
p : 0; 1-h is a piecewise linear path in h ; we dene the height function ha t by
ha t /pt; a$ S: Let ma mint ha t: If ha 1 ma o1; we dene fa p 0:
Otherwise, let t0 maxft; ha t ma g and let t1 minft; ha t ma 1g: Then
fa p is the path obtained by reecting pt0 ; t1 in the hyperplane given by x; a$ ma
and by dening fa pt pt a for tAt1 ; 1: Similarly, one denes root operators
ea ; which we will not do here in detail. They have the nice property that ea fa p p
for all paths p for which fa pa0; and that the algebra generated by root operators
acts irreducibly on the span of weight paths if and only if the corresponding Lie
algebra action is irreducible. Certain additional subtleties in the denitions in [16]
will not be relevant here. The reader who wants to get a feel for path bases is invited
to prove statements (a) and (b) below. We dene for a weight o of Vl the path po to
be the straight line from 0 to o:
(a) Let V be a minuscule module, i.e. all the weights of V are conjugate to the
highest weight. This implies that /o; a$ SAf71; 0g for all simple roots a: Then the
path basis for V is given by fpo g with o running through the weights of V :
ARTICLE IN PRESS
70
(b) If V is the adjoint representation of a simply laced Lie algebra, then the basis is
labeled by fpa g,fpi g; here a runs through the roots of g; and pi is the piecewise
linear path going from 0 to 1=2ai and back to 0, where ai is a simple root.
Lemma 1.2. Assume the notations of Section 1.1. Moreover, let g1 be the Lie
subalgebra of g0 whose generators belong to vertices not connected to vertex 0. Then
the highest weight vectors in the level space V i (with respect to g0 ) are labeled by paths
p satisfying
(i) p fa0 p0 ; where p0 is a highest weight path in V i 1 with respect to g1 ; and
(ii) eaj p 0 for all simple roots aj of g0 which are not in g1 (i.e. for which the
generators ej ; fj are not in g0 ).
Proof. If p labels a highest weight vector with respect to g0 which is not the highest
weight vector for g; then p0 ea0 p cannot be the zero path and fa0 p0 fa0 ea0 p p
(see [16]). Property (ii) follows as ea0 commutes with the root operators coming from
g1 : The other implication is shown similarly. &
Remark 1.3. In practice, it sufces to consider only those highest weights n in
V i 1 (with respect to g1 ) for which n a0 is a dominant weight for g0 : The latter
can often be quite easily checked using the fact that ea p 0 if ma 4 1:
For 2 given piecewise linear paths p1 : 0; n-h and p2 : 0; m-h we dene the
concatenation p1 p2 : 0; n m-h by
(
p1 t
if 0ptpn;
p1 p2 t
1:1
p1 n p2 t n if nptpn m:
Paths p representing basis vectors of V will always be parametrized via the interval
0; 1: For a given module V ; we dene a path of length n to be any piecewise linear
path which can be obtained by concatenating n basis paths. It is parametrized by the
interval 0; n: We refer to subsets pi; i 1; i 0; 1; y of the image of a path p as
segments of p: Then we have
Theorem 1.4 (Littelmann [16]). The multiplicity of a highest weight module Vl in V #n
is equal to the number of paths of length n terminating at l which are entirely in the
closure of the dominant Weyl chamber.
The last theorem suggests the following denitions, all with respect to a xed
simple highest weight module V : Let Pn denote the set of all paths of length n within
the closure of the dominant Weyl chamber, and let Pn m be the subset of all those
paths in Pn which end in m: It is easy to see that for V the vector representation of
GLN; a path of length n corresponds to a Young tableau of a Young diagram m
with n boxes. For this, and notational reasons, we shall in future denote paths in Pn
by letters t; s; y:
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
71
Corollary 1.5. There exists an assignment tAPn /pt ACn Endg V #n such that
pt V #n is an irreducible g-module with highest weight tn; and such that pt ps dts pt :
P
n
One checks easily that zm tAPn m pt is a central idempotent in Cn
Endg V #n : One also checks that for given sAPn1 we have
X
ps
pt ;
t;t0 s
will usually denote the latter algebra only by Cn1 : Let Wm be a simple Cn -module
labeled by the dominant weight m: Then we have the following isomorphism of Cn1 modules:
n1
Wmn D"l ml Wl
1:2
where l runs through all highest weights in V #n1 ; and the multiplicity ml of the
n1
simple Cn1 -module Wl
is equal to the number of paths of length 1 from l to m:
n
By denition, we can dene a basis vt tAPn m for the simple Cn;m -module Wm
such that vt spans the image of pt for each tAPn m: Let d; m be dominant weights for
which Vd CV #nk and Vm CV #n ; and let Pk d; m be the set of all paths of length k
from d to m within the closure of the dominant Weyl chamber. Let W d; m be the
vector space spanned by these paths. Then we obtain a representation of Ck
Endg V #k on W d; m by
aAEndV #k /pt #azn
m ;
1:3
here we used the obvious bijection between elements sAPk d; m and paths sAPn m
for which sj0;nk t:
1.3. Quantum groups
Let U Uq g be the DrinfeldJimbo quantum deformation of the universal
enveloping algebra of g; for the context of this paper, we also assume g to be a
simply-laced KacMoody algebra with invertible Cartan matrix. In this case, most
denitions of Uq g coincide, so we can e.g. assume the version of the quantum group
as in Lusztigs book [17], with q v: We assume as ground ring the eld Qq of
rational functions in the variable q; most of the results hold in greater generality (e.g.
if q is a complex number not equal to a root of unity or 0). It is well known that in
our setting the category RepU of integrable representations of U is semisimple, and
it has the same Grothendieck semiring as the original Lie algebra. Moreover, RepU
is a braided tensor category. This implies that for U-modules V ; W ; there are natural
braiding isomorphisms RVW : V #W -W #V which satisfy
RU#V ;W RUW #1V 1U #RVW ;
1:4
ARTICLE IN PRESS
72
If tAPn m; then so are the paths s in the equation above. Moreover, it follows from
Eq. (1.3) that the paths s differ from t only in the interval i 1; i 1: Because of
this we shall often only consider spaces Wi l; n with a basis consisting of paths of
length 2 from l ti 1 to n ti 1: Eq. (1.5) induces an obvious action of Ai
on Wi l; n: We will call the corresponding matrix block Ai l; n: If there is no
danger of confusion, we will often suppress the index i in Wi l; n and Ai l; n: We
shall also need the following theorem, due to Drinfeld [7].
Proposition 1.6. Let Vl ; Vm ; VL V be simple U-modules with highest weights l; m; L;
respectively, and such that Vm is a submodule of Vl #VL : Then
RVl VL RVL Vl jVm qcm cl cL 1Vm ;
where for any weight g the quantity cg is given by /g 2r; gS:
Observe that the braiding axiom also implies that the braiding matrices RV #n1 ;V
and RV ;V #n1 are equal to R1 R2 ?Rn1 and Rn1 ?R2 R1 ; respectively. Using the
notation introduced before the last proposition, we obtain
Corollary 1.7. The element Mn An An1 ?A21 ?An1 An acts on the path t via the
scalar qctn ctn1 cL ; where L is the highest weight of V :
1.4. Involutions
We x a simple highest weight module V VL of the quantum group U Uq g:
There exists an algebra homomorphism - on Qq uniquely determined by b% b for
bAQ and by q% q1 : Let Cn EndU V #n : It has been shown by Kirillov jr [13] (see
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
73
One can also show that zl zl for any central idempotent zl ACn : In particular, if
Q
the path idempotent pt can be obtained as pt ni1 zti ; we also have pt pt :
Similarly, we can dene a linear antihomomorphism T on Cn using Kashiwaras
inner product. In the following we use the notations for quantum groups as in [17,
Section 3], with some minor modications: We denote the generators of U by
ei ; fi ; ki71 : As we are in the simply laced case, we can set q vi and ki ki for all
indices i: Then one checks (see [17, 19.1.1]) that the maps
ei /qfi ki ;
fi /qei ki1 ;
ki /ki ;
Dki ki #ki ;
it is straightforward to check for the generators (and hence for the whole algebra U)
that
D3r r#r3D:
It is shown (see e.g. [17, 19.1.2]) that there exists an up to scalar multiples unique
symmetric bilinear form /; S on the highest weight U-module V such that
/uv; wS /v; ruwS for all v; wAV and uAU: Extending the bilinear form / ; S
to V #n in the obvious way, the same equation holds as well for v; wAV #n ; by : If
aT denotes the adjoint of aAEndV #n with respect to / ; S; we therefore obtain
uT ruAU for uAU; here we identify u with the linear operator by which it acts on
V #n : Hence, if cACn ; it also follows that cT ACn : Moreover, it follows from the
denition of / ; S on V #n that the operation T is functorial, i.e.
c1 #c2 T cT1 #cT2
and
on Cn ;
ARTICLE IN PRESS
74
(b) Both and T act trivially on central idempotents of Cn ; and RTi Ri and Ri
for 1pipn 1:
R1
i
Lemma 1.9. Consider the representation of Cn on the path space. Moreover, assume
that V #2 decomposes as a direct sum of mutually nonisomorphic modules. Then
(a) The basis vectors can be normalized over an algebraic extension of Qq in such a
way that any eigenidempotent pi of Ai (see Eq (1.5)) as well as Ai itself act as a
symmetric matrix.
(b) For any eigenidempotent Pi of Ai ; and any path idempotent pt we have pt Pi pt
apt ; with aAQq such that a% a:
Proof. Part (b) follows from the fact that pt Pi pt and pt are invariant under : For
part (a), observe that the space Wl spanned by highest weight vectors of V #n with
weight l is a simple Cn -module Wl (assuming it is nonzero). Hence we can dene an
orthogonal basis vt tAPn l with respect to the inner product on V #n ; where
vt Apt Wl : After adjoining suitable square roots to Qq; we can make this into an
orthonormal basis. By construction, the operation T coincides with the transpose of
the matrices acting on our vector space. By our assumptions, each spectral
idempotent of A is a central idempotent in C2 ; hence it is invariant under T ; and so is
A: This shows part (a). &
Remark. Part (a) of the last lemma holds for arbitrary Uq g-modules. This can be
derived from the explicit formula of the R-matrix. For nite-dimensional g; one can
always nd a module V generating the representation ring of Uq g satisfying the
condition in (a); the general case can be deduced from this using the braiding axioms.
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
75
There is also a simple way of describing the tensor product rules for O2N: here
Vl #V decomposes as a direct sum of simple modules Vm ; where m ranges over all
Young diagrams obtained by adding or subtracting a box to/from l:
It was shown by Brauer that the algebra EndO2N V #n is generated as an algebra
by the elements 1i #a#1ni2 ; where aAEndO2N V #2 EndSO2N V #2 (for
N41), and 1j is the identify on V #j : For EndSO2N V #n ; we need an additional
generator, called the Pfafan. It is the unique SO2N-invariant map on the N-fold
antisymmetrization of V which maps v1 4v2 4?4vN to vN1 4vN2 4?4v2N ;
where vi is an orthonormal basis for V :
Fig. 1.
ARTICLE IN PRESS
76
given by
8
>
< ei ei1
ai ei ei1
>
:1
29 N; 1; 1; y; 1; 1
if ipN 2;
if i N 1;
2:1
if i 0:
Observe that removing the vertex 0 results in the diagram DN1 ; hence ai N1
i1
denes a set of simple roots for type DN1 which lives in the subspace of RN
spanned by ei ; 1pipN 1: So in notations of Section 1.1 we have g0 so2N2 :
Moreover, if o is a weight of the root system of type EN ; the corresponding
weight o
% of g0 is obtained by removing the 0th coordinate of o in our notations.
We next describe the dominant weights. It is easy linear algebra to check that
the equation /Li ; aj S dij imply that L0 2; 0; y0; Li i; 1; y; 1; 0; y; 0;
with i 1s, for 0oioN 2; LN2 12N 1; 1; y; 1; 1 and LN1 12N
3; 1; y; 1; 1:
It follows that for any weight l the weight l% can be identied with the last N 1
coordinates of the vector representing l: We say that l is an integer or a half-integer
weight if all the coordinatesPof l% are integers or half-integers, respectively. Writing l
as a linear combination
mi Li one sees easily that l is an integer weight if
and only if mN2 mN1 is even. In the following we dene for any weight o
# by o
# N1 oN1 and o
# i oi (i.e. we change the sign of the
the weight o
N 1-coordinate).
Moreover, we dene for each weight o the number ko by one of the following 2
ways:
(a) ko /o; 2#a0 S; where a# 0 129 N; 1; 1; y; 1; 1 a0 eN1 ; or,
P
(b) if o i mi Li ; then ko mN1 mN2 2m0 :
The equivalence of statements (a) and (b) is easy to show; it sufces to check that
/Li ; 2#a0 S is equal to 0 if 0oioN 2; and that it is equal to 2, 1 and 1 for i 0;
N 2 and N 1; respectively. Also observe that we have
(c) ko joj
% are positive, and
% o0 ; if all entries of o
(d) klp2lN1 ; for any dominant weight l:
Here statement (d) follows from kl /l; 2#a0 S /l; 2a0 S 2lN1 :
We now dene a set G which will be useful for labelling dominant weights of EN
and for labelling representations of braid groups. The choice of notation is
motivated by the results on fundamental modules in the following section. As before
we will identify Young diagrams with dominant weights of DN without change of
notation.
Denition 2.1. (a) The set G consists of all triples n; m; i where nAN; m is a Young
diagram with jmjpn; and iAf0; 1; 2g; moreover, if i 0; n jmj is even, if i 1;
jmjpn 3; and if i 2; jmj n 6: For given nAN; the set Gn consists of all
elements of G whose rst coordinate is equal to n:
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
77
(b) The map F assigns to each integral dominant weight of EN ; Na9; an element
in G as follows:
(0) If l is integer and l0 jl% jX0; Fl l0 ; l% ; 0 if lN1 X0; and Fl
l0 ; l#% ; 0 if lN1 o0:
(1a) If l is half-integer with klp1 and lN1 o0; then Fl l# a# 0 ; 1
(1b) If l is half-integer with klp1 and lN1 40; then Fl l a#0 ; 1;
(2) If kl 2; then Fl l 2a#0 ; 2;
(42) If kl42; write kl 3k1 l k2 l; with 0pk2 lp2: Then Fl
l k1 l k2 l#a0 k1 leN ; k2 l; (i.e. the Nth row of the Young
diagram in Fl has k1 l boxes).
Remark 2.2. Observe that the level 0 weights l and l# get mapped to the same
#
element in G: As lal
for lN1 a0; F is not injective. This can be explained in terms
of centralizer algebras and their subalgebras generated by R-matrices similar to the
discussion in Section 1.5. In all other cases, F is injective.
Lemma 2.3. (a) A dominant weight l of type EN is of the form l l0 ; l% ; where l% is a
dominant weight of so2N2 such that /l% ; 2E Spl0 :
(b) F does indeed map into G: If noN 1; Gn is contained in the image of F:
(c) The image of F consists of triples n; m; iAG for which m has at most N rows; if it
has exactly N rows, jmj n 3i:
Proof. It follows from /l; ai SX0 for ioN that l% must be a dominant weight of type
DN1 : Observe that a0 can be written as a0 129 N; e ; where e is the highest
weight of a spinor representation. The rst statement of the claim now follows from
/l; a0 SX0:
To prove that Fl is indeed in G is now straightforward. E.g. to prove the
statement about jmj in case (1b), observe that
n jmj 2/n; m; a0 S 2/l a0 ; a0 S 4 2/l; a0 S;
this means that n jmjX4 and even for n; m; 1 Fl (in case (1b)). Similarly, one
shows that n jmjX3 and odd for n; m; 1 Fl in case (1a). Case (2) is proved
similarly. For case 42 observe that
2mN1 2lN1 k1 l k2 lXkl k1 l k2 l 2k1 l 2mN :
Hence m is indeed a Young diagram. Also observe that m has N nonzero rows if and
only if kl42: The statement about jmj in (c) now follows directly from the
denition of F: &
ARTICLE IN PRESS
78
2.2. Casimir
Let r r0 ; N 2; N 3; y; 1; 0; where r0 2 N 1N 2=2; and let
lARN : It is easy to check that /r; ai S 1 for i 0; 2; y; N 1: We dene the
scalar cl for any dominant weight l by
cl /l 2r; lS:
2:2
In view of our generic labelling set we will also pair r with vectors in Rk ; k4N: In
this case, the component rNi is dened to be equal to 1 i; for i 0; 1; y :
*
Lemma 2.4. Let l; n; b; g be weights. Moreover, define for any weight o the weight o
*
by o n l o: Then we have
(a) clb cl 2/l r; bS /b; bS; and cm cmb 2/m r; bS /b; bS;
(b) clb clg 2/l r; b gS /b; bS /g; gS;
(c) cmt cms cl cn 2/l r; bt b* s S 2/bs ; b* s S; where mt l bt and
ms l bs :
Proof. The proofs are straightforward computations.
&
3. Fundamental modules
3.1. Weights for low level spaces
Lemma 3.1. (a) Let V VL1 ; and let V i be as in Section 1.1. If o is a weight in V i;
2
then o
; and jjojj
% is a weight in VL% 1 #VE#i
% p1 i i2 N 9=4; here VL% 1 is the
vector representation and VE is a spinor representation of so2N2 :
More explicitly, we have, V 0DVL% 1 ; V 1DVE and V 2DVL% N6 "VL% N8 "?; a
direct sum of antisymmetrizations of the vector representation of so2N2 :
Moreover, if k is a weight of level iX3; kkp3i 6:
(b) Similarly, the gradation of VL0 in terms of so2N2 modules is given by VL0 0DV0
(the trivial representation), VL0 1DVE and VL0 2DVL% N5 "VL% N7 "y; a direct sum
of antisymmetrizations of the vector representation of so2N2 :
Proof. Observe that if o is a weight in V i; then o0 1 i9 N=2: If o is
2
2
conjugate to L1 ; and hence jjojj2 1=9 N 1; we have jjojj
% jjojj o20 =9
2
N 1 i i2 9 N=4: The general case follows from jjojj pjjL1 jj2 (see [10,
Proposition 11.4]). Moreover, by Lemma 1.2, any highest weight vector in V i 1
(with respect to g0 so2N2 ) has a weight of the form o0 a0 ; with o0 a weight of
V i: As a% 0 E ; the highest weight for a spinor representation, the claim follows
by induction on i:
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
79
1
o a0
j1
o a0 :
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
80
Corollary 3.3.
(a) The weights of V in levels 0,1 are of the form o 1; 7ei (for level 0) and
o 12N 7; 71; 71; y; 71; with an even number of minus signs (for level 1).
We call these weights together with level 2 weights of the form N
8; w1; 1; y; 1; 0; y; 0; with N 6 1s and 5 zeros, and with wASN1 (for
level 2) straight weights. Straight weights are all conjugate to the highest weight
L1 via the Weyl group.
(b) If b; g are straight weights, /b; gS is equal to 1=9 N 1 i for some
nonnegative integer i; with i 0 only if b g (see also [10, Proposition 11.4]).
(c) If b; g are straight weights with b g a a simple root and /l; aS 0; then
l g is not a dominant weight.
(d) If b; g are straight weights such that b g is a weight in VL0 conjugate to its
highest weight, then /b; gS 1=9 N 1:
(e) If b; g are as in (d), with l; n dominant weights such that n l b g; then
cgb cl cn =2 1=9 N is of the form l2N 3 m; with l and m only
depending on Fl:
Proof. Checking that the straight weights are conjugate to L1 is straightforward (e.g.
o0 12N 7; 1; y; 1 sa0 1; eN1 ; and N 8; 1; y; 1; 0; y; 0 sa0 12N
P
7; 1; y; 1 N2
iN5 ei ; the rest of the statement follows easily using the Weyl
group of type DN1 :
For part (b), observe that /L1 ; L1 S 1=9 N 1: We can always conjugate
the 2 weights by an element of the Weyl group such that one of them is equal to the
highest weight L1 withoutPchanging the value of the pairing. Hence /b; gS
/L1 ; L1 kS; where k j aij is a sum of simple roots. As V is an integrable
highest weight module, L1 k is a weight in V only if /L1 ; aij S40 for at least one
summand aij of k: This shows (b).
For (c), we have /l g; aS /g; bS /g; gSo0; by (b). For (d) it sufces to
observe that jjL0 jj2 4=9 N and jjbjj2 1 1=9 N jjgjj2 ; hence 4=9
N jjb gjj2 jjbjj2 2/b; gS jjgjj2 ; from which one easily deduces the claim.
To prove statement (e), observe that we can write a level 1 weight o in the form
o a0 1; eN1 an even number of ei ; and we can write a straight level 2
P
weight k in the form k 2a0 1; eN1 4j1 eij : Hence b g ja0 an even
number of summands 7ei : Let 2l be the sum of the signs of the summands ei : Then
one observes that /l r; b gS l2N 3 m;
with also m
independent of N:
The claim now follows from (d) and Lemma 2.4,(c). &
Corollary 3.4. The values of ko for the various types of weights of V are given as
follows:
(0) If o 1; 7ei is of level 0, then ko 0 or 2 depending on whether the sign
is positive or negative.
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
81
P
0
1
0
(1) Let o 2N 7; 1; 1; y; 1: If o o j eij with an even number 2j of
summands with a minus sign, then ko 3 2j:
(2) If o is a straight level 2 weight with N 6 1s in o;
% then ko 2; for weights o
of level 2 of other types, kop0:
3.2. Tensor products
Let us denote the rst coordinate of Fl by nl; i.e. Fl nl; m; i for
suitable m and i: Then, using notations of Denition 2.1, we can explicitly write out
nl as
8
if klp0 and l is integer;
>
< l0
nl l0 9 N=2
if klp1 and l half-integer;
3:1
>
:
l0 k1 l k2 l9 N=2 if klX2:
and even. The number k1 l and k2 l are dened accordingly such that kl
3:2
1
N 1; 1; y; 1- 3N 3; 3; y; 3:
2
ARTICLE IN PRESS
82
3:3
Similarly, the proof of part (b) follows as soon as one has shown
nl o nl 1
3:4
One of the less painful ways of checking the validity of Eq. (3.3) goes as follows: Let
o be a weight of level j such that kop3j 6: Observe that all nonstraight weights
satisfy this condition, by Lemma 3.1(a) and Corollary 3.4. If klX0 and kl
oX0; we get k1 l opk1 l j 2: As kl opkl 2; we get the inequality
in (3.3). Equality there can hold only if k2 l o k2 l 2 and ko 3j 6:
But then 3jkl and also 3jkl o; contradicting k2 l o 2: If klo0; but
kl oX0; we have k1 l k2 lpk1 l k2 l; except for kl 1: In the
latter case, one shows Eq. (3.3) directly as before; in all the other cases it follows
from what we have already proved before. The cases with kl oo0 are easy and
are left to the reader. Hence whenever o is not straight, nl opnl: This proves
part (b), assuming part (a) has been shown.
Moreover, by the last paragraph, it sufces to prove Eq. (3.3) with o a straight
weight. This can be checked in a straightforwad way, using Corollary 3.4. One can
even go a bit further to determine necessary conditions so that equality holds in (3.3),
as follows: One computes the values k1 l o and k2 l o for each case of
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
83
k2 l iAf0; 1; 2g; and for each value ko 3; 2; 1; 0; 1; 2 (these are the only
possibilities left for ko with o straight, by the results of the previous paragraph, if
klX0). The quantity j k1 l o k2 l o k1 l k2 l now determines
the level j at which o has to be so that equality holds in Eq. (3.3). To give an
example: let k2 l 1; and let ko run through the values 3; 2; 1; 0; 1; 2: Then
equality in 3.3 would hold only if j 1; 0; 1; 0; 1; 0: As there are no level 1
weights, we see that nl o nl 1 for a level 1 weight o only if ko 3 or
ko 1; i.e. o would have at most 2 negative entries (see Corollary 3.4). One also
sees that equality in 3.3 cannot hold in this case if o has level 2. For nl
o4nl 1; we would have to nd o of level oj; with j as above. It is easily
checked that this cannot happen, again using Corollary 3.4. One can treat the cases
k1 l 0 and k1 l 2 similarly, except that one has to be a little bit more careful if
k2 l 0; and klo0: in this case a level 1 weight o may have more than 4 negative
entries. This nishes the proof of Theorem 3.6. Moreover, we get the following
detailed information about when nl o nl 1:
Case (a): Assume that kl
2 mod 3: Then nl o nl 1 only if o
1; ej ; 1pjpN 2; or o o0 :
Case (b): Let kl
1 mod 3: Then nl o nl 1 only if o 1; 7ej ;
1pjpN 1; or if o is a level 1 weight with 0 or 2 negative entries (for klX1).
Case (c): Let kl
0 mod 3: Then nl o nl 1 only if o 1; 7ej ;
1pjpN 1; or if o is a level 1 weight with 0, 2 or 4 negative entries (if kl40) or if
o is a level 2 weight with ko 2: If klp0; level 1 weights may have more
negative entries. &
We dene levl levnl l (see Section 1.1). Using the last theorem and
Eq. (3.2), we obtain
Corollary 3.7. levl levnl l k1 l k2 l
Lemma 3.8. Let l; n be dominant weights with nn nl 2 and levn levl 2:
Then jn% l% jXN 5:
Proof. It follows from the denitions that m0 nm levm9 N=2; for any
dominant weight m: Hence n0 l0 N 7: Using Eq. (3.2), we get
nn nl n0 l0 k1 n k2 n k1 l k2 l9 N=2:
It follows that k1 n k2 n k1 l k2 l 2: Simple number theory shows that
this is possible only if knXkl 2: If n l b g; kb kgX2 implies the
claims, using Corollary 3.4. &
#n
Denition 3.9. We call Vnew
the maximum direct summand of V #n for which each
simple direct summand has as highest weight a weight l with nl n:
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
84
#1
#2
#n
Proposition 3.10. Assume N4n: Then the branching rules for Vnew
CVnew
C?CVnew
do not depend on N:
Proof. First of all observe that for noN the only dominant weights l for which
nl n are the ones for which klp2: Indeed, klX3 implies k1 lX1; and hence
the diagram m in Fl would have N rows; this would imply nlXjmjXN: We have
already seen in Lemma 2.3, (b) and its corollary that we can use the same labelling
#n
set for the highest weights in Vnew
for any sufciently large N: It sufces to translate
the branching rules into this generic setting, and to check that they do not depend on
N as well. This is straightforward if slightly tedious. The result is described in detail
in the following corollary. &
Corollary 3.11. The set G is a partially ordered set (poset), with the order defined as the
transitive closure of gAGnog0 AGn 1 if VF1 g0 appears in VFg #V for gEN modules with N sufficiently large. Explicitly this means
(0) We have g n; m; iog0 n 1; n; i iff either n m and i 1; or if n differs
from m by a single box.
(1) We have g n; m; iog0 n 1; n; i 1 iff n is a subdiagram of m with mi
ni p1 for all i; moreover, if i 1; jn% mj
% must be even.
(2) We have g n; m; 0og0 n 1; n; 2 iff n can be obtained from m by
subtracting 5 boxes in 5 different rows.
3.3. Regular paths
For this subsection l and n will always be dominant weights with nn nl 2:
Moreover, we assume that the set Pl; n P2 l; n is nonempty. It follows from
Theorem 3.6(b) that it only contains piecewise linear paths t of the form l-mt -n;
with both mt l and n mt being straight weights. The motivation of the following
denition will become clear in Section 4.
Let tAPl; n: We dene the quantity et by the equation
et cmt cl cn =2
1
:
9N
3:5
3:6
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
85
It will be important to determine when paths are regular. We have the following
results:
Lemma 3.13. Let l; n be as described at the beginning of this subsection. Then
(a) Vn appears in Vl #VL0 with multiplicity p1:
(b) If the multiplicity in (a) is zero, Pl; n only contains regular paths, and at most 2
of them, i.e. the multiplicity of Vn in Vl #V #2 is at most 2.
(c) Assume l and n have the same level. Then all paths in Pl; n are regular, except
possibly if l is a level 0 weight with lN1 0 and n l:
Proof. We will use the explicit description of weights of low level for VL1 and VL0 in
Lemma 3.1 and Corollary 3.3. Assume we can write n l as the sum of two level 0
weights b1 and b2 : Hence n l 2; g; where g 7ei 7ej : If ga0; then
necessarily b1 1; 7ei and b2 1; 7ej ; or vice versa. Hence we have at most
2 paths of length 2 from l to n; and n l is not a weight in VL0 :
If n l is the sum of a level 1 weight g and a level 0 weight b; we have jjn% l% jjN
1=2 or 3=2: One checks easily in the rst case that n l is a level 1 weight of VL0 ;
which has multiplicity 1. In the second case, there exists an index r for which br
71 and gr 71=2; with matching signs. Assuming signs, we necessarily get
b 1; er and g n l b; i.e. we have 2 paths at the most from l to n; and n l is
not a weight of VL0 (see Lemma 3.1(b)).
If n l is the sum of 2 level 1 weights, or one level 0 and one level 2 weight, we
have jn% l% jXN 5; by Lemma 3.8. If jjn% l% jjN 2; then n l is not a weight in
VL0 and one shows as before that the only possible segments are a level 0 weight
b 1; er and a level 2 weight k n l b; i.e. we have at most 2 paths from l
to n: Similarly, if jjn% l% jjN 1 and jjn% l% jj1 XN 3; then n l can only be written
as a sum of 2 level 1 weights, one of them necessarily o0 ; which also determines the
other weight. Again, one checks that in the remaining case jjn% l% jjN 1 and
jjn% l% jj1 N 5 the difference n l can only be a weight of VL0 of multiplicity 1.
If n l is written as a sum of a level 2 weight and a level 1 weight, one deduces
again from nm nl 2 that necessarily the level 1 weight has to be equal to o0 ;
and that it has to be added rst to l; i.e. there is at most one path from n to l:
Showing that all paths in case (b) are regular is straightforward: let n l b g;
with b; g weights in V : One checks easily that Pl; n only contains one path if b g:
If bag; one checks from the analysis above that /b; gS 1=9 N and hence
el b /l r; b gS: Moreover, one also reads off from our analysis before
that b g is 7 the sum of positive roots. Hence jel bjp1 only if /l; b gS 0
and b g is equal to 7 a simple root. This is ruled out by Corollary 3.3(c). The
only remaining part, checking the case with n% l% in case (c), is shown in a similar
way.
The following technical proposition will only be needed in Section 5 except for
Example 4.8, which could also be done by more elementary methods.
ARTICLE IN PRESS
86
Proposition 3.14. Assume the generic case with noN: Then there are at most 2
nonregular paths in Pl; n for any l; n with nn nl 2: If we have exactly 2
nonregular paths s and t; then et es 1:
Proof. Special cases have already been proved in Lemma 3.13. It will be shown
in Lemmas 3.15 and 3.16 in the next subsection that we have at most 2 nonregular paths in the remaining cases. Moreover, if we have exactly 2 nonregular
paths, one of them, say t; of the form l-l b-l b g n; the other one is
obtained by interchanging b with g: Moreover, n l has to be a weight of VL0 : It
follows from Corollary 3.3(d) that /b; gS 1=9 N 1; and hence et
/l r; b gS 1: If s is the other path, es /l r; g bS 1: It is now easy
to check that es and et can only have critical values at the same time if they both
are equal to 1. &
3.4. Lemmas for Proposition 3.14
We only need to look for possible nonregular paths in Pl; n if n l is a weight of
VL0 ; by Lemma 3.13(b). If n l b g; with b; g straight weights, we have
/b; gS 1=9 N 1; by Corollary 3.3(d). Hence if t is the path l-l b-n; we
obtain
et clb clg /l r; b gS:
Lemma 3.15. Assume that levn levl 1; and that nn nl 2: Then the only
possible nonregular paths in Pl; n have as level 0 segment b 1; ek1 ; where k1 is
the smallest index r for which nr olr : In particular, there are at most 2 nonregular paths
in Pl; n:
Proof. Let n l b g be a weight of VL0 ; and with b 1; 7er : Recall that
P
the level 1 weight g can be written as g a0 1; eN1 2d
j1 ekj : This implies
that
*
et /l r; b gS 1
l r; a0 7er eN1
2d
X
+
ek j :
j1
Case 1: Let b 1; er : Then we have r km for some m (as jnr lr j 1=2), and,
in particular, gao0 ; i.e. dX1: It follows easily from that b g is the sum of at
least 3 simple roots, except possibly if r N 1 and g o0 eN2 eN1 : In the
rst case, we get j/l r; b gSjX3; which implies t to be regular. In the 2nd case,
one checks that j/l r; b gSj 2 only if /l; eN2 eN1 S 0; i.e. lN2
lN1 0: One checks easily that the only possible paths t would have et 3:
Hence any path in Pl; n for which the level 0 segment is of the form b 1; er is
regular.
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
87
Case 2: Let b 1; er : If ni oli for more than one index i40; then dX1 in :
As br gr o0; we have rakj for all j: Let k1 be the smallest of the indices kj : If k1 or;
then b g a0 ek1 er ek2 eN1 possibly more positive roots. Hence
/l r; b gSX3; which implies el ba71ael g: So the only paths which
may not be regular are the ones for which b 1; er ; with r the smallest index for
which nr olr ; and g n l b: Finally, if ni oli for only one index i40; a segment
b 1; ek can only appear for k r: Hence again there are at most 2 nonregular
paths, by case 1. &
Lemma 3.16. Let levn levl 2 nn nl: Then there exist nonregular paths
P
only if n l 2o0 4j1 ekj ; and the segments are of the form o0 ek1 ek4 ;
respectively, o0 ek2 ek3 : In particular, there are at most 2 nonregular paths s from
l to n:
Proof. By the remarks at the beginning of this subsection we can assume jni li jp1
for all i40: By Lemma 3.8 the only possible level 0 segments b are of the form
b 1; er : One shows as in the proof of Lemma 3.15, case 1 that clb aclg for any
possible level 1 weight. If k is a level 2 weight, then clk aclg can be deduced from
this, using g k n l k n l g: Moreover, it is easy to show that jel
bj42 as well as jel kj42; with b and k as before. Hence paths in Pl; n
involving a level 0 and a level 2 segment are regular.
So it sufces to consider paths whose segments g and g0 are both level 1 weights such
P
that kg kg0 X2: Hence g g0 2o0 2d
j eij ; with 2dp4: If d 2; then one
shows easily that jel gjX3 if g or g0 is equal to o0 : So we only have to consider the
case when g o0 ek1 ek2 and g0 o0 el1 el2 ; where we have /g; g0 S
1 1=9 N; and g g0 el1 el2 ek1 ek2 : One checks that this or its negative is
the sum of at least 2 positive roots, except possibly if k1 ol1 ol2 ok2 ; or if
l1 ok1 ok2 ol2 : In the rst case, g g0 is the sum of only 2 simple roots only if
k1 1 l1 and k2 1 l2 and lk1 ll1 as well as lk2 ll2 : It follows that
l g0 is not dominant, while el g 3: This shows that all paths are regular in
the rst case.
For the second case observe that /g; g0 S 1 1=9 N: Hence we have 2
nonregular paths through l g and l g0 if /l r; g g0 S 0; and only one if
j/l r; g g0 Sj 2: The cases with d 0 or d 1 are easier, and can be checked
similarly. &
4. Generators of EndUq g V #n
4.1. Basics
#n
We will study the subalgebra of EndUq g Vnew
generated by the R-matrices. As the
#n
simple module of EndU Vnew labeled by lAGn has a basis labeled by Pn l; the set
ARTICLE IN PRESS
88
4:1
where Pi is the eigenprojection of Ai for the eigenvalue r1 q32N : In the following
lemma we will consider paths t such that ti is equal to l; mt and n for i
n 1; n; n 1; and for xed dominant weights l and n; as usual, the difference
ti ti 1 is a weight in V : Recall the denition of et (see (3.5)) and of Mi
Ai Ai1 ; y; A21 A2 ; y; Ai : It follows from Corollary 1.7 that Mn1 acts on t via the
scalar a2n2 qcmt cl cV ; and Mn acts on t via the scalar a2n qcn cmt cV (here cV is the
Casimir for V ).
Lemma 4.1. Let A ats and P pts be the matrices obtained from the action of An
and Pn on Pl; n; and let et; es be defined as in (3.5). Then
1 qetes ats q q1 dts r r1 q q1 pts :
4:2
This determines all entries of A in terms of entries of P; except for att ; where tAPl; n
with et 0: In this case, we have
att r1
X
sat
1 qetes
ds :
r r1 q q1
1
Proof. As Mn An Mn1 An ; we obtain A1
n Mn1 An Mn : It follows from this and
the formulas stated before that
4:3
Eq. (4.2) in statement now follows from (4.1). The rest of the statement will be
proved after Lemma 4.3. &
The following lemma essentially is the classical JucysMurphy approach. It was
used for constructing representations of Hecke algebras of type A; and, in a slightly
different way, for constructing representations of the symmetric group.
Lemma 4.2. Assume that Vn CVl #V #2 ; but n l is not a weight in VL0 : Then all offdiagonal entries of A Al; n are nonzero.
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
89
Proof. Let t and s be the 2 paths in Pl; n; with intermediate diagrams mt and ms ;
respectively. As A can be chosen to be symmetric by Lemma 1.9, it is diagonalizable
over a suitable algebraic extension of Qq: If A only had one eigenvalue, it would be
a multiple of the identity matrix. But then att ass would also imply et es and
cmt cms : But this is not possible by Lemma 3.13(b).
Hence we can assume that A has eigenvalues q and q1 : Using the formula for
the inverse of a 2 2 matrix and Eq. (4.3), we obtain ass =1 q2et att : Using
ass att q q1 and solving for att ; we obtain
att
q q1 q2et
:
1 q2et
The entry ass is computed in the same way. Moreover, detA 1 implies ast ats
ass att 1: It is checked easily that ass att 1 only if et 71: But this would
contradict regularity of t; shown in Lemma 3.13.
Lemma 4.3. Let r q2N3 ; with 2N 3a 1; and let A and P be as in Lemma 4.1.
Then the diagonal entry ds pss is equal to 0 only if es 71:
Proof. By Lemma 3.13(a), P is a rank 1 idempotent. By Lemma 1.9 we can assume P
to be of the form P vvT ; where v is an eigenvector of A with eigenvalue r1 : This
implies that if pss 0; then the sth row and sth column of P is equal to 0. The same
statement also holds for A by equation (4.2), except possibly for the diagonal entry
ass : Therefore ass must be an eigenvalue of A; i.e. ass Afq; q1 ; q32N g: If ass q;
Eq. (4.2) would imply
q1 q2es q q1 ;
from which one easily deduces that 2es 2: One shows similarly that ass q1
would imply es 1: So we only need to consider the case ass q2N3 r1 : But as
r1 only has multiplicity 1, this contradicts the fact that vs 0 for its eigenvector v:
Conclusion of proof of Lemma 4.1. In order to solve for ats in Eq. (4.2), we have to
check that et esa0: Assume to the contrary for 2 paths tas: Then Eq. (4.2)
would imply pts 0; and, as P is symmetric and rank 1, also ptt dt 0 and pss
ds 0: Hence this could only happen if both s and t were not regular. But then it
would follow from Proposition 3.14 that et es 2:
Hence we only have to consider the case et 0: Again by Proposition 3.14 this
can only happen for one path tAPl; n: But as we have already computed all other
entries of A in terms of the entries of P; we can compute att by considering the ttentry of the equation AP r1 P: &
Recall that representing the braid matrices on the path space via symmetric
matrices leads to square roots. We now show that these representations can already
be dened over Qq; and are uniquely determined by the diagonal entries of certain
ARTICLE IN PRESS
90
1
1 xi y j
Q
kol xk xl yk yl
Q
:
1pk;lpn 1 xk yl
4:4
Lemma 4.6. Let P vwT for 2 vectors v; w; and let ds vs ws be its s-diagonal entry.
Then we obtain, for each path t for which we know that dt a0 (so, in particular, for t
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
91
1
q q1 dts r r1 q q1 vt ws vt r1 vt :
1 xt xs
If t is regular, then ptt and hence also vt are nonzero; hence the stated equality holds.
Our denition of regularity also implies that cmt acmt for paths tat: This implies
xt axt: It follows that the square minor of the matrix for the system of linear
equations derived above whose rows and columns are labeled by paths t for which
vt a0 has nonzero determinant, by Eq. (4.4). &
Lemma 4.7. Let s : l-m-n be a nonregular path in Pl; n: Assume that there exists
a predecessor d of l such that Id; m (the set of all weights which can occur halfway in a
* where all diagonal entries of An ;
path in Pl; n) contains exactly 2 weights, l and l;
* n are known. Then we can compute the diagonal entry an
acting on W l;
ss of An via the
braid relations.
*
Proof. We also denote by s the path d-l-m-n: Consider the path t : d-l-m-n:
n1 n1
Observe that ats ast a0; by Proposition 4.5
Let us consider the matrix entry An An1 An tt An1 An An1 tt : By assumption,
n
n n
* n: Hence we can compute
the quantities att and a a are known for all tAPl;
tt
tt
An An1 An tt
n n1 n
att :
att att
n1 n n1
ass ast :
ats
n
ARTICLE IN PRESS
92
and
d2
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
93
l r; a0 2eN1
2d
X
+
ek j
j1
is positive (as /l r; eN1 SX0). This implies the regularity of the path, as either
/b; gS 1=9 N 1 (if n l is a weight of VL0 ), or n l is not a weight of
VL0 ; in which case the claim follows from Lemma 4.2.
n1
ARTICLE IN PRESS
94
#n
Proposition 4.10. The algebra EndU Vold
is generated by the restrictions of the
#n
#n1
#1 and of 1nk #p to Vold
; where p projects onto 1CV #k :
algebra EndU V
n
Proof. The proof follows a similar pattern as the one of the last theorem. Let Wn
be a simple Cn -module. It decomposes as a Cn1 -module into the direct sum of
simple Cn1 -modules
Wnn "m Wmn1 :
The claim is shown if we can nd for each summand m a nonzero idempotent
* n ; the algebra generated by Cn1 and
pm ACn1;m and elements u um and v vm in C
1nk #p such that uv pm and vu 1nk #p; here the equation is to be understood
n
dimq Vm
1l #p;
dimq Vl dimq V
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
95
2m1
X
eij ;
j1
ARTICLE IN PRESS
96
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
97
ARTICLE IN PRESS
98
and q: So we obtain equations for the diagonal entries of Pi as in Lemma 4.1, with
qetes replaced by a monomial in r and q: One can use the same arguments as in
Proposition 5.3 and its preceding lemmas to show that the matrices for the braid
generators are uniquely determined. This proves the uniqueness statement.
To prove the existence, we know that for N4n the R-matrices for Uq gEN ;
#n
restricted to Vnew
give us braid representations which satisfy conditions (1)(3) for
2N3
: As this holds, in particular, for innitely many values of N; the braid
rq
relations also must hold for the generic matrices compute above. This shows the
existence for generic r: &
Y xj 1 xi xj
:
xi xj
jai
Proof. We will use Cramers rule for solving this linear system. It follows from
Cauchys identity (4.4) that
Q
2
kol xk xl
Q
detC
:
5:1
1pk;lpn 1 xk xl
By symmetry, it sufces to compute the solution y1 detD=detC; where D
coincides with C except for the rst column, where di1 1 for i 1; 2; y; n: It
sufces to observe that D can be computed via Eq. (4.4) after setting y1 0 and
yj xj for j41: One obtains
Q
Qn
2
2pkolpn xk xl
j2 xj x1 xj
Q
Q
:
detD
n
2pkolpn 1 xk xl
j2 1 x1 xj
The statement now follows easily from Cramers rule.
&
Lemma 5.6. Let the vector b be given by bi 1=1 x2i : Then the equation Cy b
has the solution given by
yi xEi
Y 1 xi xj
;
x i xj
jai
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
99
Proof. We will use Cramers rule for solving this linear system as in Lemma 5.5.
Again, using symmetry, it sufces to compute the solution y1 detD=detC;
where D coincides with C except for the rst column, where now di1 1=1 x2i for
i 1; 2; y; n: The computation
of detD is done in several steps:
Q
Step 1: Let Dr;s rpkolps xi xj : Then D1;n D2;n divides detD:
To prove this, observe that after subtracting the jth row, the ith row of D is
equal to
!
x21 x2j
xl xi xj
; y; :
; y;
1 xl xi 1 xl xj
1 x21 1 x2j
Hence we can factor xi xj from the determinant. Moreover, if 2pioj; we get a
second factor xi xj in the ith column after subtracting the jth column from it.
Step 2: Now expand D with respect to the rst column, and write the Ci1 minor as
matrix 1=1 xk zl 2pk;lpn ; where
(
zl
xl
xl1
if l4i;
if lpi:
5:2
1
2
1
xk xl
1 xi
1pk;lpn
i1
Qn
Q
j2 1 x1 xj
2pkolpn xk xl
Q
1
x
k xl
1pk;lpn
2
3
n
2 Y
X
Y
1 x1
n
6
7
1i1
1 xi xj
xk xl 5:
4
2
j2
1
x
i
i1
1pkolpn
kaial
Step 3: Observe that the expression in straight brackets in the last formula is a
polynomial, which we will call F : One checks easily that the highest power of xi in F
is at most n if i 1 and at most n 1 if i41: On the other hand, it follows from step
1 that D1;n divides F : As the highest power of xi in D1;n is n 1 regardless of the
value of i; we can write
F ax1 bD1;n :
Looking at the denition of F ; we see that it is a linear combination of monomials in
the xi s with absolute degree equal to n 1n 2=2 2s; with sAN: As D1;n is
homogeneous of degree nn 1=2; it follows from the previous sentence that b 0
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
100
5:3
Step 4: It follows from the explicit formulas for detC (see (4.4)) and detD (see
the last the steps) that y1 detD=detC is as claimed, provided we can show that
the constants a and b in Eq. (5.3) are equal to 1. To prove this for n even, observe
that the coefcient Fn of xn1 in F is a homogeneous polynomial in the variables
x2 ; y; xn of degree n 1n 2=2: Hence it sufces to consider the monomials of
degree n 2 in
n
X
1i1
i2
n
n
n
Y
X
1 Y
i1
1
x
x
1
1 xi xj :
i
j
1 x2i j2
i2
j2; jai
n=21
These monomials are necessarily of the form xi
En=21 x2 ; y; x i ; y; xn ; where
Ek is the kth elementary symmetric polynomial. It follows from Weyls character
formula for type A that
Y
xk xl detxlrs ;
En=21 x2 ; y; x i ; y; xn
1pkolpn;
kaial
where the determinant is taken over the n 2 n 2 matrix with 2prpn; rai;
and 1pspn 2 with ls n 2 s for son=2 and ls n 3 s for sXn=2: It
follows from this and elementary rules for determinants that
Fn
n
X
n=2
i2
where now 2pp; qpn: This shows that a 1: One shows by the same method also
that b 1: Having computed E and hence also detD; we get the claim using
Cramers rule. &
Proposition 5.7. Let xi qei ; with ei formal variables for 1pipn; and let k
qk qk : Then, with notations of Lemmas 5.6 and 5.5, and with setting r q2N3 ; the
linear system
2N 3 1Cy q q1 b r1 c
has the solution
ys
Y es et =2
es 1
2N 3 1 tas es et =2
if n even;
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
ys
Y es et =2
es 1
2N 3 1 tas es et =2
101
if n odd:
In particular, after restricting the equations of Lemma 4.6 to paths t for which dt a0
and setting et et; we obtain for ys the nonzero diagonal entries ds :
Proof. Using Lemmas 5.6 and 5.5, we obtain
"
#
n
Y 1 xt xs
Y
1
n
1 e
1
1
q q xs r 1 xs xs
ds
xt
:
2N 3 1
xs xt
tas
t1
As ds is the diagonal entry of a -invariant idempotent, we have ds ds ; by Lemma
1.9(b). As the xi s and r are powers of q; we also have xi x1
and r% r1 : It is
i
straightforward to check that ds ds implies
(
n
Y
rq if n even;
xt
5:4
r
if n odd:
&
t1
Theorem 5.8. The statements of Theorems 4.9 and 4.11 also hold for qAC; q not a root
of unity. In the classical case q 1; Cn is generated by C2 ; the quasi-Pfaffian, and the
symmetric group Sn :
Proof. As the diagonal entries ds are quotients of products of q-integers (with
possibly q q1 added), they are zero for q 1 only if they are already zero as
functions in q: Hence the arguments in the proofs for the above stated corollaries
and theorem also go through for q 1 and for q not a root of unity. &
The results above can also be used to compute the diagonal entries, and hence the
braid matrices over Qr; q: More generally, we have
Theorem 5.9. Let now r; q be elements defining a field extension of Q: The
representations of the braid group Bn into the algebra In r; q are well-defined and
semi-simple if q is not a root of unity, and rl qm a71 for l; kAZ
Proof. There is obviously no problem with substituting elements of a eld extension
for the variable r and q for diagonal entries ds 0; as rational functions. The other
diagonal entries can be computed by the equations in Lemma 4.6, only running over
paths s for which ds a0: One can explicitly compute from Corollary 3.3(d), that there
exist integers lt and mt such that rlt qmt qet for r q2N3 ; where et was computed
for gEN : Hence we can compute the matrix entries in terms of rational functions of
r and q which are quotients of products of factors of the form rlt qmt rlt qmt ; with
possibly q q1 added. It is easy to check for both cases that these quantities are
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
102
0
zero only if rl qm 71 for suitable integers l 0 and m0 : Hence our representations are
0
0
well-dened and semisimple whenever rl qm a71: &
Remark 5.10. It is easy to check that for g n; m; 0 the braid representation pg
determines the direct summand of the q-version of Brauers centralizer algebra,
labeled by m; which was dened in [1,18]. For other elements in G; one does get new
braid representations. For the q-Brauer algebra one also obtains interesting braid
representations when it is not semisimple. Naively speaking, this can be done
by just removing paths for which the diagonal entry of one of the eigenprojections of
a braid generator becomes 0 (see Proposition 4.4). It would be interesting to study
similar braid representations in connection with the algebra In : Observe that one
does obtain such representations for the EN -specializations, i.e. with r q2N3 :
Results for the q-Brauer algebra and Hecke algebras also suggest further interesting semisimple quotients of In for q a root of unity. As one can compute the
quantities lt and mt in the proof of Theorem 5.9 explicitly (see e.g. Corollary 3.3(e)),
it should also be possible to determine such semisimple quotients. Similarly, it should
be possible to nd much stronger conditions under which In may not be semisimple,
for given n:
Remark 5.11. A denition of the algebra In via generators and relations seems to
be rather complicated. The relations already for I3 are rather intricate, and
certainly do not sufce to dene I4 : Another possibility of dening the algebras In
would be via a given basis of reduced words (where reduced would have to be
dened appropriately). One way might be to use canonical bases, as indicated
by Lusztig in [17, 27.3.10]. An explicit integral basis over a suitable ring has
been constructed for the q-Brauer algebra in a paper by Morton and Wassermann [19].
A possible reason why In is difcult to dene via generators and relations is the
fact that there exist larger quotients than In of the group algebra of the innite braid
group whose restriction to B2 would only give a three-dimensional algebra. For one,
observe that one can obtain additional irreducible representations of Bn after
interchanging the parameters r1 and q in the simple components of In :
A potentially interesting quotient algebra might be the algebra Jn : Let gi ; i
1; 2; y; n 1 be the images of the braid generators in Jn : Then Jn is dened by the
relations (besides the braid relations)
(R1) gi qgi q1 gi r1 0;
q
r1
q
r1
q2 pi g1
;
(R2) pi gi1 pi
i1 pi
l
ARTICLE IN PRESS
H. Wenzl / Advances in Mathematics 177 (2003) 66104
103
irreducible representations (3 one-dimensional, 3 two-dimensional and 1 threedimensional representation) to exhaust this dimension. This is also in compliance
with the fact that the additional relation s3i 1 generates a nite group of order 24
(see [4]). Now it only remains to observe that relation R2 kills the two-dimensional
representation in which the generators have eigenvalues r1 and q: Hence J3 has
dimension 20, while I3 only has dimension 19. It is also possible to compute the
dimension of J4 in a similar way (for this computer calculations by E. Rowell were
needed). It would be interesting to know for which n the algebra Jn has nite
dimension.
Remark 5.12. Another possible application of the algebras In would be to generalize
the approach in [12] for classifying (braided) tensor categories whose Grothendieck
semiring is the one of a Lie group. This was successfully carried out in [12] for Lie
type A; using Hecke algebra representations of braid groups. Observe that by the
uniqueness statement in Theorem 5.4 the braid representations are uniquely
determined by the restriction rule, which basically coincides with the tensor product
rules of our category (see [29] for further details).
5.4. VogelDeligne approach
In this approach (see [5,6,26]), the authors consider the decomposition of tensor
powers of the adjoint representation of an exceptional Lie algebra as well as of
several low rank classical Lie algebras. They observed a number of intriguing
similarities in the decomposition behaviour. Further computational evidence for
such patterns was given by Cohen and de Man [3].
In the paper [24,25], the rigidity of braid representations was used to show that the
dimension formulas for representations in the second tensor power of the adjoint
representation (given for the Lie group case in [5]) are forced by the Grothendieck
semiring and the braiding structure. Here we sketch how the approach in this paper
could be applied to tensor powers of adjoint representations, and we give one modest
application of it:
As in our setting, one observes that there are only 3 new representations in the
second tensor power of the adjoint representation. Hence, switching to the quantum
group setting, and restricting to the new part, we again get representations of braid
groups whose generators satisfy a cubic equation; for type E8 this coincides with the
representations we have studied here. For other types, however, we get more
representations than the ones obtained via specializations from In : So clearly, if there
exists an exceptional series, one would expect a larger quotient of the braid groups
than the one given by In : It is easy to check that this quotient would have to be a
quotient of the algebra Jn ; dened at the end of the last section. In fact, it can be
shown that the commutants of the action of an exceptional Lie group g on g#n
new ; as
described in [3], are obtained as classical limits q-1 of quotients of the algebras Jn
for np4; here the eigenvalues of the braid generators are powers of q which can be
computed using the formulas for the quantum Casimir (see Proposition 1.6).
ARTICLE IN PRESS
104
References
[1] J. Birman, H. Wenzl, Braids, link polynomials and a new algebra, Trans. AMS 313 (1989) 249273.
[2] R. Brauer, On algebras which are connected with the semisimple continuous groups, Ann. of Math.
63 (1937) 854872.
[3] A. Cohen, R. de Man, Computational evidence for Delignes conjecture regarding exceptional Lie
groups, C. R. Acad. Sci. Paris Ser. I Math. 322 (5) (1996) 427432.
[4] H.S.M. Coxeter, Factor groups of the braid group, Proceedings of the Fourth Canadian Mathematics
Congress, Banff 1957, University of Toronto Press, 1959, pp. 95122.
[5] P. Deligne, La serie exceptionnelle de groupes de Lie, C. R. Acad. Sci. Paris Ser. I Math. 322 (4)
(1996) 321326.
[6] P. Deligne, R. de Man, La serie exceptionnelle de groupes de Lie. II, C. R. Acad. Sci. Paris Ser.
I Math. 323 (6) (1996) 577582.
[7] V. Drinfeld, On almost cocommutative Hopf algebras, Leningrad Math. J. 1 (1990) 321343.
[8] P.N. Hoefsmit, Representations of Hecke algebras of nite groups with BN pairs of classical type,
Thesis, University of British Columbia, 1974.
[9] M. Jimbo, T. Miwa, M. Okado, Solvable lattice models related to the vector representations of
classical simple Lie algebras, Comm. Math. Phys. 116 (1988) 507525.
[10] V. Kac, Innite Dimensional Lie Algebras, 3rd Edition, Cambridge University Press, Cambridge.
[11] Ch. Kassel, Quantum Groups, Springer, Berlin, 1995.
[12] D. Kazhdan, H. Wenzl, Reconstructing monoidal categories, Adv. in Soviet Math. 16 (1993)
111136.
[13] A. Kirillov Jr., On an inner product in modular categories, J. AMS 9 (1996) 11351170.
[14] P. Kulish, N. Reshetikhin, E. Sklyanin, Yang-Baxter equation and representation theory, Lett. Math.
Phys. 5 (1981) 393.
[15] R. Leduc, A. Ram, A ribbon Hopf algebra approach to the irreducible representations of centralizer
algebras: the Brauer, BirmanWenzl, and Type A IwahoriHecke algebras, Adv. in Math. 125 (1997)
194.
[16] P. Littelmann, Paths and root operators in representation theory, Ann. of Math. (2) 142 (3) (1995)
499525.
[17] G. Lusztig, Introduction to Quantum Groups, Birkhauser, Basel, 1993.
[18] J. Murakami, The Kauffman polynomial of links and representation theory, Osaka J. Math. 24
(1987) 745758.
[19] H. Morton, A. Wassermann, A basis for the BirmanWenzl algebra, Preliminary preprint, University
of Liverpool, 1989/2000.
[20] R. Orellana, H. Wenzl, q-Centralizer algebras for spin groups, preprint.
[21] N. Reshetikhin, Quantized universal enveloping algebras, the YangBaxter equation and invariants
of links, LOMI preprint, 1988.
[22] G.W. Schwarz, Invariant theory of G2 and Spin7 ; Comment. Math. Helv. 63 (4) (1988) 624663.
[23] J. Stembridge, Computational aspects of root systems, Coxeter groups and Weyl characters, posted at
www.math.lsa.umich.edu/jrs/papers/carswc.ps.gz.
[24] I. Tuba, H. Wenzl, Representations of the braid groups B3 and of SL2; Z; Pacic J., 2001, to
appear.
[25] V. Turaev, H. Wenzl, Semisimple and modular categories from link invariants, Math. Ann. 309
(1997) 411461.
[26] P. Vogel, Algebraic structures on modules of diagrams, preprint.
[27] H. Wenzl, Hecke algebras of type An and subfactors, Inv. math. 92 (1988) 349383.
[28] H. Wenzl, C tensor categories from quantum groups, J. AMS 11 (1998) 261282.
[29] H. Wenzl, Tensor categories and braid representations, Proceedings of the Durham LMS Symposium
1999 (posted at math.ucsd.edu/ wenzl), to appear.