Representations of Semisimple Lie

Algebras in the BGG Category O

James E. Humphreys

Dept. of Mathematics & Statistics, U. Massachusetts, Amherst,

MA 01003-4515
E-mail address: [email protected]
1991 Mathematics Subject Classification. Primary 17B10;
Secondary 20G05, 22E47
Contents

Preface xi

Chapter 0. Review of Semisimple Lie Algebras 1

§0.1. Cartan Decomposition 1
§0.2. Root Systems 3
§0.3. Weyl Groups 4
§0.4. Chevalley–Bruhat Ordering of W 5
§0.5. Universal Enveloping Algebras 6
§0.6. Integral Weights 7
§0.7. Representations 8
§0.8. Finite Dimensional Modules 9
§0.9. Simple Modules for sl(2, C) 9

Part I. Highest Weight Modules 11

Chapter 1. Category O: Basics 13

§1.1. Axioms and Consequences 13
§1.2. Highest Weight Modules 15
§1.3. Verma Modules and Simple Modules 17
§1.4. Maximal Vectors in Verma Modules 18
§1.5. Example: sl(2, C) 20
§1.6. Finite Dimensional Modules 20
§1.7. Action of the Center 22
§1.8. Central Characters and Linked Weights 24

iii
iv Contents

§1.9. Harish-Chandra Homomorphism 25

§1.10. Harish-Chandra’s Theorem 26
§1.11. Category O is Artinian 28
§1.12. Subcategories Oχ 30
§1.13. Blocks 30
§1.14. Formal Characters of Finite Dimensional Modules 32
§1.15. Formal Characters of Modules in O 33
§1.16. Formal Characters of Verma Modules 34
Notes 35
Chapter 2. Characters of Finite Dimensional Modules 37
§2.1. Summary of Prerequisites 37
§2.2. Formal Characters Revisited 38
§2.3. The Functions p and q 38
§2.4. Formulas of Weyl and Kostant 40
§2.5. Dimension Formula 41
§2.6. Maximal Submodule of M (λ), λ ∈ Λ+ 43
§2.7. Related Topics 45
Notes 46
Chapter 3. Category O: Methods 47
§3.1. Hom and Ext 47
§3.2. Duality in O 49
§3.3. Duals of Highest Weight Modules 51
§3.4. The Reflection Group W[λ] 52
§3.5. Dominant and Antidominant Weights 54
§3.6. Tensoring Verma Modules with Finite Dimensional Modules 56
§3.7. Standard Filtrations 58
§3.8. Projectives in O 60
§3.9. Indecomposable Projectives 62
§3.10. Standard Filtrations of Projectives 64
§3.11. BGG Reciprocity 65
§3.12. Example: sl(2, C) 66
§3.13. Projective Generators and Finite Dimensional Algebras 68
§3.14. Contravariant Forms 68
§3.15. Universal Construction 70
Contents v

Notes 71
Chapter 4. Highest Weight Modules I 73
§4.1. Simple Submodules of Verma Modules 74
§4.2. Homomorphisms Between Verma Modules 75
§4.3. Special Case: Dominant Integral Weights 76
§4.4. Simplicity Criterion: Integral Case 77
§4.5. Existence of Embeddings: Preliminaries 78
§4.6. Existence of Embeddings: Integral Case 79
§4.7. Existence of Embeddings: General Case 81
§4.8. Simplicity Criterion: General Case 82
§4.9. Blocks of O Revisited 83
§4.10. Example: Antidominant Projectives 84
§4.11. Application to sl(3, C) 85
§4.12. Shapovalov Elements 86
§4.13. Proof of Shapovalov’s Theorem 88
§4.14. A Look Back at Verma’s Thesis 90
Notes 91
Chapter 5. Highest Weight Modules II 93
§5.1. BGG Theorem 93
§5.2. Bruhat Ordering 94
§5.3. Jantzen Filtration 95
§5.4. Example: sl(3, C) 97
§5.5. Application to BGG Theorem 98
§5.6. Key Lemma 98
§5.7. Proof of Jantzen’s Theorem 100
§5.8. Determinant Formula 102
§5.9. Details of Shapovalov’s Proof 103
Notes 106
Chapter 6. Extensions and Resolutions 107
§6.1. BGG Resolution of a Finite Dimensional Module 108
§6.2. Weak BGG Resolution 109
§6.3. Exactness of the Sequence 110
§6.4. Weights of the Exterior Powers 111
§6.5. Extensions of Verma Modules 113
vi Contents

§6.6. Application: Bott’s Theorem 115

§6.7. Squares 116
§6.8. Maps in a BGG Resolution 118
§6.9. Homological Dimension 120
§6.10. Higher Ext Groups 122
n
§6.11. Vanishing Criteria for Ext 123
§6.12. Computation of ExtnO (M (µ), M (λ)∨ ) 124
§6.13. Ext Criterion for Standard Filtrations 125
§6.14. Characters in Terms of Ext•O 126
§6.15. Comparison of Ext•O and Lie Algebra Cohomology 127
Notes 128
Chapter 7. Translation Functors 129
§7.1. Translation Functors 130
§7.2. Adjoint Functor Property 131
§7.3. Weyl Group Geometry 132
§7.4. Nonintegral Weights 134
§7.5. Key Lemma 135
§7.6. Translation Functors and Verma Modules 137
§7.7. Translation Functors and Simple Modules 138
§7.8. Application: Category Equivalences 138
§7.9. Translation to Upper Closures 140
§7.10. Character Formulas 142
§7.11. Translation Functors and Projective Modules 143
§7.12. Translation From a Facet Closure 144
§7.13. Example 145
§7.14. Translation From a Wall 146
§7.15. Wall-Crossing Functors 148
§7.16. Self-Dual Projectives 149
Notes 152
Chapter 8. Kazhdan–Lusztig Theory 153
§8.1. The Multiplicity Problem for Verma Modules 154
§8.2. Hecke Algebras and Kazhdan–Lusztig Polynomials 156
§8.3. Examples 157
§8.4. Kazhdan–Lusztig Conjecture 159
Contents vii

§8.5. Schubert Varieties and KL Polynomials 160

§8.6. Example: W of Type C3 161
§8.7. Jantzen’s Multiplicity One Criterion 162
§8.8. Proof of the KL Conjecture 165
§8.9. Outline of the Proof 166
§8.10. Ext Functors and Vogan’s Conjecture 168
§8.11. KLV Polynomials 169
§8.12. The Jantzen Conjecture and the KL Conjecture 171
§8.13. Weight Filtrations and Jantzen Filtrations 172
§8.14. Review of Loewy Filtrations 173
§8.15. Loewy Filtrations and KL Polynomials 175
§8.16. Some Details 177
Part II. Further Developments 179
Chapter 9. Parabolic Versions of Category O 181
§9.1. Standard Parabolic Subalgebras 182
§9.2. Modules for Levi Subalgebras 183
§9.3. The Category Op 184
§9.4. Parabolic Verma Modules 186
§9.5. Example: sl(3, C) 188
§9.6. Formal Characters and Composition Factors 189
§9.7. Relative Kazhdan–Lusztig Theory 190
§9.8. Projectives and BGG Reciprocity in Op 191
§9.9. Structure of Parabolic Verma Modules 192
§9.10. Maps Between Parabolic Verma Modules 193
§9.11. Parabolic Verma Modules of Scalar Type 195
§9.12. Simplicity of Parabolic Verma Modules 196
§9.13. Jantzen’s Simplicity Criterion 198
§9.14. Socles and Self-Dual Projectives 199
§9.15. Blocks of Op 200
§9.16. Analogue of the BGG Resolution 201
§9.17. Filtrations and Rigidity 203
§9.18. Special Case: Maximal Parabolic Subalgebras 204
Notes 206
Chapter 10. Projective Functors and Principal Series 207
viii Contents

§10.1. Functors on Category O 208

§10.2. Tensoring With a Dominant Verma Module 210
§10.3. Proof of the Theorem 211
§10.4. Module Categories 212
§10.5. Projective Functors 213
§10.6. Annihilator of a Verma Module 215
§10.7. Comparison of Hom Spaces 216
§10.8. Classification Theorem 218
§10.9. Harish-Chandra Modules 219
§10.10. Principal Series Modules and Category O 221
Notes 222
Chapter 11. Tilting Modules 223
§11.1. Tilting Modules 224
§11.2. Indecomposable Tilting Modules 225
§11.3. Translation Functors and Tilting Modules 227
§11.4. Grothendieck Groups 229
§11.5. Subgroups of K 230
§11.6. Fusion Rules 231
§11.7. Formal Characters 232
§11.8. The Parabolic Case 234
Chapter 12. Twisting and Completion Functors 235
§12.1. Shuffling Functors 236
§12.2. Shuffled Verma Modules 237
§12.3. Families of Twisted Verma Modules 239
§12.4. Uniqueness of a Family of Twisted Verma Modules 240
§12.5. Existence of Twisted Verma Modules 242
§12.6. Twisting Functors 242
§12.7. Arkhipov’s Construction of Twisting Functors 243
§12.8. Twisted Versions of Standard Filtrations 244
§12.9. Complete Modules 245
§12.10. Enright’s Completions 247
§12.11. Completion Functors 248
§12.12. Comparison of Functors 249
Chapter 13. Complements 251
Contents ix

§13.1. Primitive Ideals in U (g) 252

§13.2. Classification of Primitive Ideals 253
§13.3. Structure of a Fiber 254
§13.4. Kostant’s Problem 255
§13.5. Kac–Moody Algebras 256
§13.6. Category O for Kac–Moody Algebras 258
§13.7. Highest Weight Categories 259
§13.8. Blocks and Finite Dimensional Algebras 260
§13.9. Quiver Attached to a Block 261
§13.10. Representation Type of a Block 263
§13.11. Soergel’s Functor V 264
§13.12. Coinvariant Algebra of W 265
§13.13. Application: Category Equivalence 266
§13.14. Endomorphisms and Socles of Projectives 267
§13.15. Koszul Duality 268
Bibliography 271
Frequently Used Symbols 283
Index 287
Preface

Representation theory plays a central role in Lie theory and has developed in
numerous specialized directions over recent decades. Motivation comes from
many areas of mathematics and physics, notably the Langlands program.
The methods involved are also diverse, including fruitful interactions with
“modern” algebraic geometry. Here we focus primarily on algebraic methods
in the case of a semisimple Lie algebra g over C with universal enveloping
algebra U (g), where the prerequisites are relatively modest.
The category Mod U (g) of all U (g)-modules is much too large to be
understood algebraically. Fortunately, many interesting Lie group repre-
sentations can be studied effectively in terms of a more limited subcate-
gory where modules are subjected to appropriate finiteness conditions: the
BGG category O introduced in the early 1970s by Joseph Bernstein, Israel
Gelfand, and Sergei Gelfand. Their papers, stimulated in part by D.N.
Verma’s 1966 thesis, have led to far-reaching work involving a growing list
of researchers. In this book we discuss systematically the early work leading
to the Kazhdan–Lusztig Conjecture and its proof around 1980. This is at
the core of more recent developments, some of which we go on to introduce
in the later chapters. Taken on its own, the study of category O offers a
rewarding tour of the beautiful terrain that lies just beyond the classical
Cartan–Weyl theory of finite dimensional representations of g.
Part I (comprising Chapters 1–8) is written in textbook style, at the level
of a second year graduate course in a U.S. university. The emphasis here is
on highest weight modules, starting with Verma modules and culminating
in the determination of formal characters of simple highest weight modules
in the setting of the Kazhdan–Lusztig Conjecture (1979). The proof of this
conjecture requires sophisticated ideas from algebraic geometry which go

xi
xii Preface

well beyond the algebraic framework of earlier chapters. Thus Chapter 8

marks a shift toward the survey style used in the remainder of the book.
The chapters in Part II can to a large extent be read independently.
They supplement the more unified theme of Part I in a variety of ways,
often motivated by problems arising in Lie group representations. The book
ends with an introduction to the influential work of Beilinson, Ginzburg,
and Soergel on Koszul duality.
I have tried to keep prerequisites to a minimum. The reader needs to be
comfortable with the basic structure theory of semisimple Lie algebras over
C (summarized in Chapter 0) as well as with standard algebraic methods
including elementary homological algebra.
Exercises are scattered throughout the text (mainly in Part I) where I
thought they would do the most good. Some of the more straightforward
ones are used later in the development. At any rate, the most important
exercise for the reader is to engage actively with the ideas presented. Exam-
ples are also interspersed, though unfortunately it is difficult to gain much
direct insight from low rank cases of the sort which can be done by hand.
The deeper parts of the theory have required some imaginative leaps not
based on examples alone.
The substantial reference list includes all source material cited, together
with related books and survey articles. I have added a somewhat arbitrary
sample of other research papers to point the reader in directions such as
those sketched in Chapter 13. There is also a list of frequently used symbols,
most of which are introduced early in the book. Anyone who consults the
literature will encounter a wide array of notational choices; here I have tried
to keep things simple and consistent to the extent possible.
The mathematics presented here is not original, though parts of the
treatment may be. Many people have provided helpful feedback on ear-
lier versions of the chapters, including Troels Agerholm, Henning Ander-
sen, Brian Boe, Tom Braden, Jon Brundan, Walter Mazorchuk, Wolfgang
Soergel, Catharina Stropple, and Geordie Williamson. I am especially in-
debted to Jens Carsten Jantzen for his detailed suggestions at many stages
of the writing. His ideas have left a lasting imprint on the study of category
O. Naturally, the final choices made are my own responsibility. Corrections
and suggestions from readers are welcome.

J.E. Humphreys
February 2008
[email protected]
 Chapter 0

Review of Semisimple
Lie Algebras

The main technical prerequisite for this book is a good working knowledge
of the structure of semisimple Lie algebras over an algebraically closed field
of characteristic 0 such as C. (In fact, it is enough to work over a splitting
field of characteristic 0, as indicated below.) In this preliminary chapter
we summarize results with which the reader should be familiar, coupled
with some explicit references to the textbook literature. The basic structure
theory through the classification by Dynkin diagrams is treated in a very
large number of sources, of which we cite only a few here, e.g., Bourbaki
[45, 46], Carter [60], Humphreys [126, 130], Jacobson [144].
The subject can be approached from a number of angles, including the
traditional theory of Lie groups and the theory of linear algebraic groups;
but group theory generally remains in the background here. Since notation
varies considerably in the literature, the reader needs to be aware of our
conventions (which are often closest to those in [126]); these are intended to
steer something of a middle course among the available choices. Frequently
used notations are listed at the end of the book.

0.1. Cartan Decomposition

The basic object of study here is a semisimple Lie algebra g over a field of
characteristic 0, having a Cartan subalgebra h which is split: the eigenvalues
of ad h are in the field, for all h ∈ h. Write ` := dim h. For convenience we
normally take C for the field, unless the contrary is stated.

1
2 0. Review of Semisimple Lie Algebras

Denote by Φ ⊂ h∗ the root system of g relative to h (which by conven-

tion does not contain 0), whose abstract properties will be recalled in the
following section. To each root α ∈ Φ corresponds a nonzero 1-dimensional
subspace of g called a root space:
gα = {x ∈ g | [hx] = α(h)x for all h ∈ h}.
Usually we fix a simple system ∆ ⊂ Φ having ` elements and correspond-
ing positive system Φ+ ⊂ Φ having m elements. This defines
L a Cartan
decomposition g = n− ⊕ h ⊕ n, where n := α>0 gα and n− := α<0 gα .
L

The corresponding standard Borel subalgebra is b := h ⊕ n, whose oppo-

site Borel subalgebra is b− := h ⊕ n− . These are maximal solvable subalge-
bras of g; more generally, any such subalgebra is called a Borel subalgebra.
Any subalgebra p containing a Borel subalgebra is called parabolic; those
containing b are standard. There are 2` of these, in natural bijection with
subsets I ⊂ ∆ (where we interpret I as the Greek letter “iota” for notational
consistency). A parabolic subalgebra decomposes as p = l ⊕ u, where the
Levi factor l is reductive (the direct sum of its semisimple derived algebra
and an abelian subalgebra) while u is the largest nilpotent ideal of p. When
p = pI ⊃ b, the center of lI lies in h and uI ⊂ n; here the root system
ΦI = Φ ∩ ZI of (the derived algebra of) lI lies in Φ and has I as a simple
gα with α running over those α ∈ Φ+ not in ΦI .
L
system, whereas uI =
The Lie algebra g acts on itself by derivations ad x, where (ad x)(y) :=
[xy]. Let G be the adjoint group, generated by all automorphisms exp(ad x)
with x ∈ g nilpotent. All Cartan subalgebras of g are conjugate under
G; their common dimension ` is the rank of g. The adjoint representation
ad : g → Der g is a first example of a representation of g. The associ-
ated Killing form (x, y) := Tr(ad x ad y) is nondegenerate. The algebra
g decomposes uniquely (up to order of summands) into the direct sum of
simple ideals. In turn, the simple Lie algebras are uniquely determined
by their (irreducible) root systems; these are classified explicitly as types
A` , B` , C` , D` , E6 , E7 , E8 , F4 , G2 . This often makes case-by-case proofs pos-
sible.
Subalgebras of type A1 in g play a special role. Such an algebra is
isomorphic to sl(2, C), which has a basis (h, x, y):
     
1 0 0 1 0 0
h := , x := , y := .
0 −1 0 0 1 0
Each α ∈ Φ+ determines such a subalgebra (call it sα ) with basis denoted
(hα , xα , yα ).
We always work with a standard basis of g, consisting of root vectors
xα ∈ gα and yα ∈ g−α (for α > 0) together with hα = [xα yα ] for α ∈ ∆
so that all α(hα ) = 2. There are many ways to choose such a basis. For
0.2. Root Systems 3

example, a Chevalley basis has very simple structure constants, all in Z,

though it is not usually essential to make this kind of refined choice; see [46,
VIII,2.4] (where the convention is that [xα yα ] = −hα ), [126, §25].

0.2. Root Systems

The occurrence of root systems in the Lie algebra setting has led to a some-
what more widely applicable notion of abstract root system Φ in a vector
space over R. Ultimately the classification turns out to be the usual one
in terms of Dynkin diagrams, but the axiomatic treatment suggests useful
generalizations and is logically independent of the Lie algebra theory.
In the structure theory of g, the root system spans a Q-form E0 of the
dual space h∗ , where the Killing form is nondegenerate. Thus E := R ⊗Q E0
has a natural structure of euclidean space. Humphreys [126, §9] then defines
an abstract root system in a finite dimensional euclidean space E over R,
with inner product denoted (λ, µ), to be a finite set Φ of vectors spanning
E and not containing 0. It is required that for each α ∈ Φ, Rα ∩ Φ = {±α}.
Moreover, the euclidean reflection sα defined by λ 7→ λ − 2(λ, α)/(α, α)α
sends Φ to itself. Further, the Cartan invariant hβ, α∨ i := 2(β, α)/(α, α)
lies in Z for all α, β ∈ Φ. Here α∨ := 2α/(α, α) is the coroot of α. The
Z-span Λr of Φ in E is called the root lattice. (The notation in [126] is a
little different.)
It is this formulation which we adopt, identifying E0 with the Q-span of
the roots in h∗ , which also contains the root lattice Λr .
In Bourbaki [45, VI, §1] (where the notation and initial viewpoint are
different), one starts just with a finite subset Φ as above in a finite dimen-
sional vector space E over R. It is required that each α ∈ Φ determine a
unique coroot α∨ in E ∗ , so that the “reflection” sα,α∨ permutes Φ. Here
sα,α∨ (λ) = λ − hλ, α∨ iα. Further, the values of each α∨ on Φ must lie in Z.
In either version one gets a dual root system Φ∨ := {α∨ | α ∈ Φ} (living
in the euclidean space E in our version, but in the dual space E ∗ in the
Bourbaki version). This is a root system whose Dynkin diagram is dual to
that of Φ; in particular, if Φ is of type B` , then Φ∨ is of type C` . In the
Killing form identification of h and h∗ , the coroot α∨ corresponds to hα ∈ h:
hβ, α∨ i = β(hα ) for all β ∈ Φ.

Each choice of simple system ∆ in Φ defines a partition into subsets

of positive and negative roots (denoted respectively Φ+ and Φ− ). Here ∆
forms a basis of EP(or a Z-basis of Λr ), while each β ∈ Φ+ can be written
+
uniquelyPas β = α∈∆ cα α with cα ∈ Z . Define the height of β to be
ht β := α∈∆ cα ; so ht β = 1 if and only if β ∈ ∆.
4 0. Review of Semisimple Lie Algebras

Here are a few useful facts which emerge from the study of root systems
with a given simple system ∆:
(1) The set ∆∨ forms a simple system in Φ∨ .
(2) For any α, β ∈ Φ, the roots of the form α + kβ form an unbroken
root string α − rβ, . . . , α − β, α, α + β, . . . , α + sβ, which involves
at most four roots.
(3) If β is positive but not simple, there exists a simple root α for which
hβ, α∨ i > 0; thus sα β ∈ Φ+ and ht sα β < ht β.

0.3. Weyl Groups

The natural symmetry group attached to a root system Φ is its Weyl group
W , the (finite!) subgroup of GL(E) generated by all reflections sα with
α ∈ Φ (or just the simple reflections sα with α ∈ ∆ when ∆ is a fixed simple
system). Evidently the root lattice Λr is stable under the action of W .
Abstractly, W is a finite Coxeter group, having generators sα (α ∈ ∆)
and defining relations of the form (sα sβ )m(α,β) = 1. Moreover, W satisfies
the crystallographic restriction m(α, β) ∈ {2, 3, 4, 6} when α 6= β. (In the
Lie algebra setting there is a natural identification of W with a finite group
of automorphisms of h in g; this induces an action on h∗ .) The Weyl group
of Φ∨ is naturally isomorphic to W . Moreover, (wα)∨ = wα∨ when α ∈ Φ
and w ∈ W .
The subgroup of W fixing a point λ ∈ E is itself a reflection group, being
generated by those sα for which hλ, α∨ i = 0 (see for example [130, 1.12]).
Write `(w) = n if w = s1 . . . sn with si simple reflections and n as small
as possible; such an expression is called reduced . Standard facts about the
length function on W include:
(1) The number of α ∈ Φ+ for which wα < 0 is precisely `(w). In
particular, when α ∈ ∆ (equivalent to `(sα ) = 1), we have sα β > 0
for all β 6= α in Φ+ . Moreover, w is uniquely determined by the set
of α > 0 for which wα < 0.
(2) If w ∈ W , then `(w) = `(w−1 ). Thus `(w) = |Φ+ ∩ w(Φ− )|.
(3) There is a unique element w◦ ∈ W of maximum length |Φ+ |, send-
ing Φ+ to Φ− . Moreover, `(w◦ w) = `(w◦ ) − `(w) for all w ∈ W .
(4) If α > 0 and w ∈ W satisfy `(wsα ) > `(w), then wα > 0, while
`(wsα ) < `(w) implies wα < 0. It follows that `(sα w) > `(w) ⇔
w−1 α > 0.
Given a subset I ⊂ ∆, the subgroup WI it generates is called a “para-
bolic” subgroup of W . It is a Coxeter group in its own right and its length
function `I agrees with the restriction of `.
0.4. Chevalley–Bruhat Ordering of W 5

For these and other facts about Weyl groups (or Coxeter groups in gen-
eral), see for example Bourbaki [45, VI, §1], Humphreys [130, Chap. 1].

0.4. Chevalley–Bruhat Ordering of W

There is a subtle (and very useful) way to partially order W . The discussion
here applies equally well to an arbitrary Coxeter group, but in the case of
Weyl groups the ordering arises first in the work of Chevalley and others
on inclusions among closures of Bruhat cells for a semisimple group having
Lie algebra g. The ordering is most often referred to just as the Bruhat
ordering; see Björner–Brenti [29, Chap. 2], Humphreys [130, 5.9–5.11].
If S denotes the set of simple reflections sα (with α ∈ ∆), the set T of all
reflections in W is defined by T := w∈W wSw−1 . This is known to consist
S
t
of the reflections sα with α ∈ Φ+ . For w, w0 ∈ W and t ∈ T , write w0 → w
t
if w = tw0 and `(w0 ) < `(w). In turn, write w0 → w if w0 → w for some
t ∈ T . Extend this relation to a partial ordering of W by defining w0 < w
to mean that w0 = w0 → w1 → . . . → wn = w for some w1 , . . . , wn−1 .
The identity element of W is the unique minimal element for this order-
ing. It is clear from the definition that w0 < w ⇒ `(w0 ) < `(w). (Caution:
In the literature the ordering is sometimes defined with 1 as maximal rather
than minimal element.)
In rank 2 cases, W is a dihedral group and the ordering is easy to
describe: w0 < w if and only if `(w0 ) < `(w). But things get considerably
more complicated in higher ranks. For a picture of the Bruhat ordering in
the symmetric group S4 (the Weyl group of type A3 ), see [29, Fig. 2.4].
Although a parallel definition could be given using right rather than left
multiplication by elements of T , the two definitions actually lead to the same
ordering. Other important features can be summarized as follows.

Proposition. The Bruhat ordering of a Coxeter group (W, S) satisfies:

(a) w0 ≤ w if and only if w0 occurs as a subexpression in one (hence
any) reduced expression s1 . . . sn (with si ∈ S) for w. Here a subex-
pression is a product si1 . . . sik with 1 ≤ i1 < · · · < ik ≤ n.
(b) Adjacent elements in the Bruhat ordering differ in length by 1.
(c) If w0 < w and s ∈ S, then w0 s ≤ w or w0 s ≤ ws (or both).
(d) If `(w1 ) + 2 = `(w2 ), the number of elements w ∈ W satisfying
w1 < w < w2 is 0 or 2.
(e) If I ⊂ ∆, the Bruhat ordering of the Coxeter group WI is induced
by the Bruhat ordering of W .
6 0. Review of Semisimple Lie Algebras

Part (b) implies that the same ordering would result if we had defined
t
w0 → w more restrictively by requiring that `(w) = `(w0 ) + 1. (This version
will be useful in Chapter 6.)
Part (d) expresses a feature of the Möbius function for an arbitrary
Coxeter group (which can be derived in various ways); this plays an essential
role in 6.7.

0.5. Universal Enveloping Algebras

An essential tool in the construction and study of representations is the
universal enveloping algebra U (a) of a Lie algebra a. This is an associative
algebra with 1, infinite dimensional if a 6= 0 and noncommutative if a is
not abelian. The algebra U (a) is left and right noetherian and has no zero-
divisors. Any Lie algebra homomorphism a1 → a2 induces an associative
algebra homomorphism between the enveloping algebras. The adjoint action
of a on itself induces an action of a on U (a) by derivations: if x ∈ a, then
ad x sends u 7→ xu − ux for all u ∈ U (a).
Given an ordered basis (x1 , . . . , xn ) of a, the monomials xt11 . . . xtnn with
ti ∈ Z+ form a basis of U (a). This is called a PBW (Poincaré–Birkhoff–
Witt) basis. In particular, a may be identified with a subspace of U (a). By
extending a basis of a Lie subalgebra of a to a basis of a, one gets a natural
embedding of its enveloping algebra into U (a).
In the case of a semisimple Lie algebra g = n− ⊕ h ⊕ n, write a standard
basis in the order y1 , . . . , ym (with yi ∈ n− ), h1 , . . . , h` (with hi = hαi for
simple roots αi ), x1 , . . . , xm . We say that a corresponding basis element
h1 . . . hs` ` xt11 . . . xtmm
r m s1
y1r1 . . . ym
is written in a standard PBW ordering. (Often it is convenient to let
y1 , . . . , y` correspond to negatives of simple roots.) Then U (g) is the di-
rect sum of subspaces U (g) Pν with ν ∈ Λr : here U (g)ν is the span of those
monomials for which ν = i (ti − ri )αi .
Denote the center of U (g) by Z(g). This subalgebra plays an important
role in representation theory, sometimes acting by scalars even on infinite
dimensional modules to which Schur’s Lemma does not apply. In Chapter 1
we will work out in some detail the structure of Z(g): it turns out to be just
a polynomial algebra in ` indeterminates, though this is far from obvious.
In the classical theory one constructs directly a special element of Z(g)
(unique up to scalar multiples) which is called a Casimir element. It plays a
key role in algebraic proofs of Weyl’s complete reducibility theorem (recalled
below) and related study of finite dimensional representations. The defini-
tion uses the nondegeneracy of the Killing form on g. For example, when
0.6. Integral Weights 7

g = sl(2, C), with standard basis (h, x, y), a Casimir element is h2 +2xy+2yx.
Rewritten in a standard PBW ordering, it becomes h2 + 2h + 4yx.
There is a standard anti-involution τ : g → g which interchanges xα
with yα for all α ∈ Φ+ and fixes all h ∈ h; it extends canonically to an
anti-automorphism of U (g) (see for example Jantzen [148, 1.2]). We may
call τ the transpose map, since in the case when g = sl(n, C) it is just the
usual transpose map for matrices. The source of τ is an involution equal to
−τ which is easily constructed using the Serre relations for g. (In some of
the literature, such as [145, 148] and [231], the letter σ is used instead of
τ .)
Later it will be shown that τ fixes Z(g) pointwise, using properties of the
Harish-Chandra homomorphism: see Exercise 1.10. (There does not seem
to be a more elementary proof.)

0.6. Integral Weights

The representations we study will involve weights relative to the action of
h, as recalled in the following section. Here we assemble some basic facts
which can be developed abstractly within the theory of root systems (see
for example [126, §13]).
There is a natural dual lattice in E to the root lattice, defined by
Λ := {λ ∈ E | hλ, α∨ i ∈ Z for all α ∈ Φ}.
Here it is enough to let α run over ∆. We call Λ the integral weight lattice
associated to Φ. It lies in the Q-span E0 of the roots in h∗ and includes
the root lattice Λr as a subgroup of finite index. (The index is given by the
determinant of the matrix of Cartan invariants relative to ∆.) In its action
on E, the Weyl group W keeps Λ stable.
When the simple system ∆ is fixed, there is a natural partial ordering
on Λ defined by µ ≤ λ if and only if λ − µ ∈ Γ, where Γ ⊂ Λ is defined to
be the set of all Z+ -linear combinations of simple roots.
Write ∆ = {α1 , . . . , α` }. Then the group Λ is free abelian of rank `,
with a basis consisting of fundamental weights $1 , . . . , $` . These satisfy
h$i , αj∨ i = δij . The subset Λ+ := Z+ $1 + · · · + Z+ $` is called the set of
dominant integral weights. From the fact that hβ, α∨ i = β(hα ) when β ∈ Φ,
one shows easily that hλ, α∨ i = λ(hα ) for all λ ∈ Λ.
The special weight ρ := $1 + · · · + $` ∈ Λ+ can also be characterized as
half the sum of positive roots. It satisfies hρ, α∨ i = 1, or sα ρ = ρ − α, for
all α ∈ ∆.
Here are two useful facts ([126, 13.2]):
(1) The number of dominant weights ≤ λ for a given λ ∈ Λ+ is finite.
8 0. Review of Semisimple Lie Algebras

(2) Given λ ∈ Λ+ , all wλ ≤ λ for w ∈ W .

For a simple system ∆, there is a natural fundamental domain C ⊂ E
for the action of W . Relative to the inner product on E,
C := {λ ∈ E | (λ, α) > 0 for all α ∈ ∆},
while its closure is obtained by changing > to ≥. We call C a Weyl chamber
in E. The Weyl chambers are in natural bijection with simple systems in Φ
and are permuted simply transitively by W . They can be characterized as
the connected components in the complement of the union of all hyperplanes
orthogonal to roots.
For use in 7.9 we need an elementary lemma on the partial ordering
of weights relative to a Weyl chamber (taken from Bourbaki [45, VI, §1,
Prop. 18]).
Lemma. Let λ ∈ Λ and fix a chamber C ⊂ E for the action of W corre-
sponding to a simple system ∆. The following conditions are equivalent:
(a) λ ∈ C.
(b) λ ≥ sα λ for all α ∈ ∆.
(c) λ ≥ wλ for all w ∈ W .

Proof. Evidently (a) is equivalent to (b), while (c) implies (b). To show
that (a) implies (c), assume λ ∈ C and use induction on `(w) to derive (c).
The case `(w) = 0 is automatic. If `(w) > 0, write w = w0 sα for some
α ∈ ∆ so that `(w0 ) < `(w). This implies in particular that wα < 0 and
thus w0 α > 0 (0.2). Now consider
λ − wλ = (λ − w0 λ) + w0 (λ − sα λ).
The first summand on the right is ≥ 0 by induction, whereas the second
equals −hλ, α∨ iw0 α. The assumption that λ ∈ C forces hλ, α∨ i ≥ 0. Since
w0 α > 0, the summand in question is ≥ 0. This shows that λ ≥ wλ as
required. 

0.7. Representations
Our main concern here is with representations of g (or U (g)), not necessarily
finite dimensional. We mostly use the language of modules rather than
representations: simple, semisimple, indecomposable, projective, etc. So
the object of study is the category Mod U (g) of all (left) U (g)-modules, or
rather a well-behaved subcategory.
In the finite dimensional case, the theory of integral weights outlined
above is sufficient. In the infinite dimensional setting, we have to broaden
0.9. Simple Modules for sl(2, C) 9

the idea of “weight”. If M is arbitrary, it still makes sense to define its

weight spaces relative to the action of h. For each λ ∈ h∗ , let
Mλ := {v ∈ M | h · v = λ(h)v for all h ∈ h}.
If Mλ 6= 0, we say that λ is a weight of M . The multiplicity of λ in M is then
defined to be dim Mλ (possibly ∞). Define Π(M ) := {λ ∈ h∗ | Mλ 6= 0}, the
set of all weights occurring in M . A partial ordering µ ≤ λ of h∗ is defined
just as in the integral case (0.6).
It is easily checked that weight Pvectors for distinct weights in M are
linearly independent. Thus the sum λ∈h∗ Mλ is direct, though it may be 0
and moreover M need not be the direct sum of these subspaces. Our main
focus will be on the case when M actually is the direct sum of its weight
spaces, i.e., h acts semisimply on M . Then we call M a weight module.
When ∆ is fixed, the corresponding fundamental dominant weights in
Λ form a convenient basis of h∗ . If λ = `i=1 ci $i , then λ ∈ Λ (resp. Λ+ )
P
precisely when all ci ∈ Z (resp. Z+ ).

0.8. Finite Dimensional Modules

The theory of finite dimensional modules for g has been well studied. In
1.6 and Chapter 2 many of the main results will be recovered in the more
general setting of category O. However, one key fact due to Weyl has no
analogue in the infinite dimensional theory and will only be quoted here. It
is usually referred to as Weyl’s Complete Reducibility Theorem :
Every finite dimensional U (g)-module is isomorphic to a
direct sum of simple modules, the multiplicities of the lat-
ter being uniquely determined.
In some treatments of the basic structure theory (such as [126]), Weyl’s
Theorem and its consequences play a key role from an early stage.
When dim M < ∞, M is always a weight module. This follows from
Weyl’s Theorem: more precisely, elements of h act on M via semisimple
matrices while elements of n or n− act via nilpotent matrices. This gen-
eralizes the usual notion of Jordan decomposition, which arises intrinsi-
cally in g, independent of any specific linear realization. All weights of M
are in fact integral. The set Π(M ) of weights of M is W -invariant, with
dim Mλ = dim Mwλ .

0.9. Simple Modules for sl(2, C)

While we do not presuppose any detailed knowledge of finite dimensional
simple modules for an arbitrary semisimple Lie algebra g, we do need to
refer at times to the results for g = sl(2, C). These are easy to derive from
10 0. Review of Semisimple Lie Algebras

scratch and play a key role in most developments of the structure theory
outlined earlier.
Here is a quick summary (see for example [126, §7]). Thanks to Weyl’s
theorem on complete reducibility, the problem is to classify and construct
(up to isomorphism) all simple finite dimensional modules. Fix a standard
basis (h, x, y) for g, with [hx] = 2x, [hy] = −2y, [xy] = h. Since dim h = 1,
weights λ ∈ h∗ may be identified with complex numbers. In turn, the integral
weight lattice Λ is identified with Z and Λr with 2Z.
The simple modules are in natural bijection with dominant integral
weights λ ∈ Λ+ and may be denoted L(λ). It is not difficult to con-
struct explicit matrix realizations for each of these. Here L(λ) has 1-
dimensional weight spaces, with weights λ, λ−2, . . . , −(λ−2), −λ. In par-
ticular, dim L(λ) = λ + 1. For example, L(0) is the trivial module, L(1)
is the natural 2-dimensional realization of g, and L(3) is the adjoint mod-
ule. Basis vectors vi (0 ≤ i ≤ λ) for L(λ) can be chosen so that (setting
v−1 = vλ+1 = 0):
h · vi = (λ − 2i)vi ,
x · vi = (λ − i + 1)vi−1 ,
y · vi = (i + 1)vi+1 .
Here v0 is (up to scalars) the unique weight vector in L(λ) killed by x. In
this case W has order 2 and permutes weights of L(λ) by sending λ−2i to
its negative.
Part I. Highest Weight
Modules

11
 Chapter 1

Category O: Basics

We begin by laying the foundations for study of the BGG category O, start-
ing with the axioms and their immediate consequences in 1.1. This category
encompasses all finite dimensional U (g)-modules, the traditional starting
point for representation theory of semisimple Lie algebras. In 1.6 the finite
dimensional simple modules are identified as those having dominant integral
highest weights. Then in Chapter 2 we shall recover the classical theorems
of Weyl and others in the category O setting. But our main concern is the
study of O itself, which has wider implications in representation theory.
The most accessible infinite dimensional modules in O are the Verma
modules (1.3), which arise first as an auxiliary tool for the construction of
simple modules. Later their study deepens considerably, as we shall see in
Chapters 4–5.
Some basic tools introduced in this chapter are central characters and
related subcategories Oχ (1.7–1.13), along with formal characters of weight
modules (1.14–1.16).
One important aspect of category O is deferred until Chapter 9, largely
for pedagogical reasons: the study of subcategories Op attached to arbitrary
parabolic subalgebras p (not just the Borel subalgebra b). On one hand, Op
requires a cumbersome extra layer of notation; on the other hand, the results
there are sometimes easier to derive once the corresponding results for O
are in hand.

1.1. Axioms and Consequences

The BGG category O is defined to be the full subcategory of Mod U (g)
whose objects are the modules satisfying the following three conditions:

13
14 1. Category O: Basics

(O1) M is a finitely generated U (g)-module.

L
(O2) M is h-semisimple, that is, M is a weight module: M = λ∈h∗ Mλ .
(O3) M is locally n-finite: for each v ∈ M , the subspace U (n) · v of M is
finite dimensional.
As a matter of notation, we usually write HomO (M, N ) rather than
HomU (g) (M, N ) or Homg (M, N ) to emphasize that M and N lie in O.
From 0.7 it is clear that all finite dimensional modules lie in O. Before
constructing further examples in 1.3, we explore the implications of the
axioms. To begin, we deduce from the axioms two features of the weight
structure of an arbitrary M in O:
(O4) All weight spaces of M are finite dimensional.
(O5) In the notation of 0.7, the set Π(M ) of all weights of M is contained
in the union of finitely many sets of the form λ − Γ, where λ ∈ h∗
and Γ is the semigroup in Λr generated by Φ+ .
First, it is obvious from (O2) that a finite generating set in (O1) can always
be taken to consist of weight vectors. To verify (O4) and (O5) it will suffice
to let M be generated by a single weight vector v of weight λ. Thanks to
the PBW Theorem (0.5), we can write U (g) = U (n− )U (h)U (n). Applying a
factor from U (n) to v we get a finite dimensional subspace V of M (thanks to
(O3)) spanned by weight vectors having weights of type λ + sum of positive
roots. Now V is stable under U (h), while the action of U (n− ) on V produces
only weights lower than these. Moreover, only a finite number of standard
basis monomials y1i1 . . . ym
im in U (n− ) can yield the same weight when applied

to a weight vector in V .
It is easy to derive a few further consequences of the axioms.

Theorem. Category O satisfies:

(a) O is a noetherian category, i.e., each M ∈ O is a noetherian U (g)-
module.
(b) O is closed under submodules, quotients, and finite direct sums.
(c) O is an abelian category.
(d) If M ∈ O and L is finite dimensional, then L ⊗ M also lies in O.
Thus M 7→ L ⊗ M defines an exact functor O → O.
(e) If M ∈ O, then M is Z(g)-finite: for each v ∈ M , the span of
{z · v | z ∈ Z(g)} is finite dimensional.
(f) If M ∈ O, then M is finitely generated as a U (n− )-module.

Proof. (a) Since U (g) is a noetherian ring (0.5), this follows from (O1).
1.2. Highest Weight Modules 15

(b) Closure under taking quotients or finite direct sums is immediate.

Since U (g) is noetherian, any submodule of a finitely generated U (g)-module
is also finitely generated, whereas (O2) and (O3) are automatic for a sub-
module of a module in O.
(c) Since Mod U (g) itself is an abelian category, we only need to check
that O is closed under finite direct sums and under taking kernels and cok-
ernels of homomorphisms. This follows from (b).
(d) It is clear that L ⊗ M satisfies axioms (O2) and (O3). To check
finite generation, let v1 , . . . vn be a basis of L and let w1 , . . . wp generate
the module M . We claim that the vi ⊗ wj generate L ⊗ M . If N is the
submodule they generate, then certainly all v ⊗ wj with v ∈ L lie in N . In
turn, for each x ∈ g we have x · (v ⊗ wj ) = x · v ⊗ wj + v ⊗ x · wj ∈ N , and
then also v ⊗ x · wj ∈ N . Iteration shows that v ⊗ u · wj ∈ N for all PBW
monomials u ∈ U (g). Thus L ⊗ M ⊂ N .
(e) Since each v ∈ M is a sum of weight vectors, we may assume that
v ∈ Mλ for some λ ∈ h∗ . The fact that z ∈ Z(g) commutes with the action of
h then implies that z · v ∈ Mλ . Weight spaces of M being finite dimensional
by (O4), the span of {z · v | z ∈ Z(g)} is therefore finite dimensional.
(f) The axioms imply that M is generated by a finite dimensional U(b)-
submodule N . Thanks to the PBW Theorem, a basis of N then generates
M as a U (n− )-module. 

For some applications it is enough to work in the subcategory Oint whose

objects all have integral weights: this encompasses for example all finite
dimensional modules (1.6). In any case, proofs are often easier to carry out
initially in Oint . Going in the opposite direction, we sometimes consider the
larger category C of U (g)-modules whose objects are the weight modules
with finite dimensional weight spaces.
Exercise. (a) If M ∈ O and [λ] := λ + Λr is any coset of h∗ modulo Λr , let
M [λ] be the sum of all weight spaces Mµ for which µ ∈ [λ]. Prove that M [λ]
is a U (g)-submodule of M and that M is the direct sum of (finitely many!)
such submodules.
(b) Deduce that all weights of an indecomposable module M ∈ O lie in
a single coset of h∗ modulo Λr .

1.2. Highest Weight Modules

Still lacking is a concrete construction of modules in O, apart from familiar
finite dimensional examples such as the adjoint module g. In this direction
we introduce some helpful terminology. First define a nonzero vector v + in
a U (g)-module M to be a maximal vector of weight λ ∈ h∗ if v + ∈ Mλ
16 1. Category O: Basics

and n · v + = 0. (Such a vector is also referred to in the literature as a

primitive vector of weight λ.) Thanks to axioms (O2) and (O3), every
nonzero module in O has at least one maximal vector.
Next define a U (g)-module M to be a highest weight module of
weight λ if there is a maximal vector v + ∈ Mλ such that M = U (g) · v + .
By the PBW Theorem, such a module satisfies M = U (n− ) · v + . It is easy
to spell out the basic properties of an arbitrary highest weight module by
specializing the discussion in the previous section:

Theorem. Let M be a highest weight module of weight λ ∈ h∗ , generated

by a maximal vector v + . Fix an ordering of the positive roots as α1 , . . . , αm
and choose corresponding root vectors yi in g−αi . Then:

(a) M is spanned by the vectors

P y1i1 . . . ym
im · v + with i ∈ Z+ , having
j
respective weights λ − ij αj . Thus M is a semisimple h-module.
(b) All weights µ of M satisfy µ ≤ λ: µ = λ − (sum of positive roots).
(c) For all weights µ of M , we have dim Mµ < ∞, while dim Mλ = 1.
Thus M is a weight module, locally finite as n-module, and M ∈ O.
(d) Each nonzero quotient of M is again a highest weight module of
weight λ.
(e) Each submodule of M is a weight module. A submodule generated
by a maximal vector of weight µ < λ is proper; in particular, if M
is simple its maximal vectors are all multiples of v + .
(f) M has a unique maximal submodule and unique simple quotient.
In particular, M is indecomposable.
(g) All simple highest weight modules of weight λ are isomorphic. If
M is one of these, dim EndO M = 1.

Proof. Part (a) follows from the standard commutation relations involving
elements of h and n− , together with the fact that weight vectors having
distinct weights are linearly independent. Then (b) is immediate.
For (c), it is clear that only finitely many choices of i1 , . . . , im yield the
same weight, by expressing sums of positive roots in terms of simple roots.
Local finiteness of the action of n follows from the fact that gα maps Mµ
into Mµ+α , coupled with (b).
Part (d) is clear, while part (e) follows from the fact that M lies in O
(forcing its submodules to lie in O) along with part (b).
Thanks to (e), each proper submodule of M is a weight module. It
cannot have λ as a weight, since the 1-dimensional space Mλ generates M .
Therefore the sum of all proper submodules is still proper, whence (f).
1.3. Verma Modules and Simple Modules 17

For (g), suppose M1 and M2 are simple highest weight modules of weight
λ, with respective maximal vectors v1+ , v2+ . Set M := M1 ⊕ M2 and v + :=
(v1+ , v2+ ). Evidently v + is a maximal vector in M , so the submodule N it
generates is a highest weight module of weight λ. In turn, the two projections
N → M1 and N → M2 are surjective. As simple quotients of a highest
weight module, M1 and M2 must then be isomorphic, thanks to (f). [In fact,
they are both isomorphic to N under the projections, since N is semisimple.]
Finally, since M is simple any nonzero endomorphism ϕ must be an
isomorphism and also take v + to a multiple cv + . Since v + generates M , it
follows that ϕ is just multiplication by c. (This property is analogous to
Schur’s Lemma, even though M need not be finite dimensional.) 

In a loose sense, highest weight modules are the building blocks for all
objects in O. Since we cannot expect modules here to be semisimple, we
look instead for a finite filtration: a chain of submodules 0 ⊂ M1 ⊂ M2 ⊂
· · · ⊂ Mn = M for which the quotients Mi+1 /Mi have a known structure.
Corollary. Let M be any nonzero module in O. Then M has a finite filtra-
tion 0 ⊂ M1 ⊂ M2 ⊂ · · · ⊂ Mn = M with nonzero quotients each of which
is a highest weight module.

Proof. Since M can be generated by finitely many weight vectors and is

moreover locally n-finite, the n-submodule V generated by such a generating
family of weight vectors is finite dimensional. If dim V = 1, it is clear that
M itself is a highest weight module. Otherwise proceed by induction on
dim V .
Start with a nonzero weight vector v ∈ V of weight λ which is maximal
among all weights of V and is therefore a maximal vector in M . It generates
a submodule M1 , while the quotient M := M/M1 again lies in O and is gen-
erated by the image V of V . Since dim V < dim V , the induction hypothesis
can be applied to M , yielding a chain of highest weight submodules whose
pre-images in M are the desired M2 , . . . , Mn . 

1.3. Verma Modules and Simple Modules

We can construct a large family of highest weight modules by exploiting the
technique of induction. The idea, which arises in similar ways in other
types of representation theory, is to start with an easily constructed family
of modules for a subalgebra and then “induce” to g.
Here we start with the Borel subalgebra b corresponding to a fixed choice
of positive roots, which in turn has an abelian quotient algebra b/n isomor-
phic to h. Any λ ∈ h∗ then defines a 1-dimensional b-module with trivial
n-action, denoted Cλ . Now set M (λ) := U (g) ⊗U (b) Cλ , which has a natural
18 1. Category O: Basics

structure of left U (g)-module. This is called a Verma module and may

also be written as Indgb Cλ to emphasize the functorial nature of induction.
Thanks to the PBW Theorem, U (g) ∼ = U (n− ) ⊗ U (b), which allows us to
∼
write M (λ) = U (n ) ⊗ Cλ as a left U (n− )-module. Therefore M (λ) is a
−

free U (n− )-module of rank one. In particular, the vector v + := 1 ⊗ 1 in

the definition of M (λ) is nonzero and is acted on freely by U (n− ), while
n · v + = 0 and h · v + = λ(h)v + for all h ∈ h. Thus v + is a maximal vector
and generates the U (g)-module M (λ). Moreover, the set of weights of M (λ)
is visibly λ − Γ. It follows that M (λ) lies in O.
Remark. More generally, one can start with an arbitrary finite dimensional
U (b)-module N on which h acts semisimply and get an induced module
U (g) ⊗U (b) N in O. This defines an exact functor from such U (b)-modules
to U (g)-modules, since (as above) an induced module is free as a U (n− )-
module.

One can alternatively describe M (λ) by generators and relations. From

the PBW Theorem, we see that the left ideal I of U (g) which annihilates
v + is generated by n together with all h − λ(h) · 1 with h ∈ h. Thus
M (λ) ∼= U (g)/I. Obviously I also annihilates a maximal vector of weight λ
generating an arbitrary highest weight module M . So M (λ) maps naturally
onto M and therefore plays the role of universal highest weight module
of weight λ.
Using Theorem 1.2(f), we can write unambiguously L(λ) (resp. N (λ))
for the unique simple quotient (resp. unique maximal submodule) of M (λ).
Since every nonzero module in O has at least one maximal vector, we con-
clude from this discussion and Theorem 1.2(g):
Theorem. Every simple module in O is isomorphic to a module L(λ) with
λ ∈ h∗ and is therefore determined uniquely

up to isomorphism by its highest
weight. Moreover, dim HomO L(µ), L(λ) = δλµ . 
Exercise. Show that M (λ) has the following property: For any M in O,
HomU (g) M (λ), M = HomU (g) Indgb Cλ , M ∼ g




= HomU (b) Cλ , Resb M ,
where Resgb is the restriction functor. [Use the universal mapping property
of tensor products.] In the context of induced modules for group algebras
of finite groups, this adjointness property is known as Frobenius reciprocity.

1.4. Maximal Vectors in Verma Modules

In order to explore the submodule structure of a Verma module M (λ), it is
natural to begin by looking for maximal vectors of weight µ < λ. This theme
will be developed further in Chapter 4, but for our immediate purposes it is
1.4. Maximal Vectors in Verma Modules 19

enough to carry out just the most elementary construction. Fix a standard
basis (for example, a Chevalley basis) for g as in 0.1, consisting of h1 , . . . , h`
together with root vectors xα ∈ gα and yα ∈ g−α for α ∈ Φ+ (abbreviated
to xi , yi for a numbering of simple roots α1 , . . . , α` ).
Proposition. Given λ ∈ h∗ and a fixed simple root α, suppose n := hλ, α∨ i
lies in Z+ . If v + is a maximal vector of weight λ in M (λ), then yαn+1 · v +
is a maximal vector of weight µ := λ − (n + 1)α < λ. Thus there exists
a nonzero homomorphism M (µ) → M (λ) whose image lies in the maximal
submodule N (λ).

Here it is essential that α be simple; in this case it is easily seen that

all weight spaces M (λ)λ−kα with k ∈ Z+ are 1-dimensional and spanned
by a vector yαk · v + . Otherwise a less direct approach to the existence of
such a maximal vector is required (to be discussed in Chapter 4). The
problem is that for a nonsimple positive root α with hλ, α∨ i ∈ Z>0 , a weight
space M (λ)λ−kα is typically of dimension > 1 and spanned by the images
a(β)
of v + under a number of monomials in U (n)− involving powers yβ with
P
a(β) = kα. To write down explicitly a maximal vector in such a weight
space is quite challenging.
The proof of the proposition depends on some standard commutation
formulas in U (g) (see for example [126, 21.2]):
Lemma. Let xi , yi be standard basis vectors as above, corresponding to sim-
ple roots α1 , . . . , α` . Then for all k ≥ 0 and 1 ≤ i, j ≤ `, we have:
(a) [xj , yik+1 ] = 0 whenever j 6= i.
(b) [hj , yik+1 ] = −(k + 1)αi (hj )yik+1 .
(c) [xi , yik+1 ] = −(k + 1)yik (k · 1 − hi ).

Proof. We outline the steps, which are elementary, leaving the details as an
exercise for the reader. When k = 0, (a) follows from the fact that αj − αi is
not a root if j 6= i, while (b) just expresses the fact that [hj , yi ] = −αi (hj )yi
and (c) is obvious. Now proceed by induction. For example, in (c) rewrite
the left side in U (g):
[xi , yik+1 ] = xi yik+1 − yik+1 xi = [xi , yi ]yik + yi [xi , yik ] = hi yik + yi [xi , yik ].
Then apply the induction hypothesis, together with (b) (replacing k + 1
there by k). 

Proof of Proposition. Recalling the isomorphism of U (n− )-modules be-

tween U (n− ) and M (λ), part (c) of the lemma translates into the statement
of the proposition in view of part (b). 
20 1. Category O: Basics

Corollary. With λ as in the proposition, let v + be instead a maximal vector

of weight λ in L(λ). Then yαn+1 · v + = 0.

Proof. Since λ − (n + 1)α < λ, no maximal vector of this weight can exist
in the simple module L(λ) (by Theorem 1.2(e)). So the maximal vector of
this weight in M (λ) constructed in the proposition must lie in the maximal
submodule and map to 0 in the quotient L(λ). 

1.5. Example: sl(2, C)

The description of Verma modules and simple modules in the case g =
sl(2, C) serves as a (much simplified) prototype of the general theory to be
developed in the following sections. The reader is encouraged to fill in the
following outline now, although some steps will become considerably easier
later on. (The finite dimensional case has already been reviewed in 0.9.)
Fix a standard basis (h, x, y) for g, with [hx] = 2x, [hy] = −2y, [xy] = h.
Since dim h = 1, weights λ ∈ h∗ may be identified with complex numbers.
In turn, the integral weight lattice Λ is identified with Z and Λr with 2Z.
• M (λ) has weights λ, λ−2, λ−4, . . . , each with multiplicity one. Basis
vectors vi (i ≥ 0) for M (λ) can be chosen so that (setting v−1 = 0):
h · vi = (λ − 2i)vi ,
x · vi = (λ − i + 1)vi−1 ,
y · vi = (i + 1)vi+1 .
(This is not the only way to choose a basis, but is convenient when
working with integral weights and then reducing modulo a prime.)
• dim L(λ) < ∞ if and only if λ ∈ Z+ . In this case, the maximal
submodule of M (λ) is isomorphic to L(−λ − 2).
• M (λ) is simple if and only if λ is not in Z+ .
Exercise. When g = sl(2, C), show that M (λ) ⊗ M (µ) cannot lie in O.

1.6. Finite Dimensional Modules

Relying on the representation theory of sl(2, C), we can now determine pre-
cisely which of the simple modules L(λ) with λ ∈ h∗ are finite dimensional.
For each α ∈ Φ+ , let sα be the copy of sl(2, C) in g spanned by hα , xα , yα .
Given an enumeration of simple roots as α1 , . . . , α` , abbreviate sαi by si ,
hαi by hi , and sαi by si .
For use in the proof below, observe that if M ∈ O and N ⊂ M is a
finite dimensional si -submodule of M generated by a weight vector, then
h automatically stabilizes N . Indeed, if v ∈ N has weight µ, then the
1.6. Finite Dimensional Modules 21

commutation relations in g force h · (xi · v) = xi · (h · v) + αi (h)xi · v. Since,

h · v = µ(h)v ∈ N , the left side also lies in N . (Similarly for yi · v.)

Theorem. The simple module L(λ) in O is finite dimensional if and only

if λ ∈ Λ+ . This is the case if and only if dim L(λ)µ = dim L(λ)wµ for all
µ ∈ h∗ and all w ∈ W .

Proof. Proceed in steps.

(1) In one direction, assume that dim L(λ) < ∞. For each fixed i, the
structure of the si -module L(λ) is then transparent. In particular, λ(hi ) ∈
Z+ , forcing λ ∈ Λ+ (0.7).
(2) In the other direction, assume that λ ∈ Λ+ . To show that dim L(λ) <
∞ requires more work. Again we look at the si -module structure of L(λ)
for each 1 ≤ i ≤ `. Since n := hλ, αi∨ i ∈ Z+ , Corollary 1.4 implies that
the si -submodule generated by v + is finite dimensional (and is in fact the
unique simple module of dimension n + 1).
(3) In turn, we claim that for each fixed i, L(λ) is the sum of all its
finite dimensional si -submodules; call this sum M . By step (2), M 6= 0.
On the other hand, take any finite dimensional si -submodule N of L(λ); so
N ⊂ M . Then g ⊗ N is a finite dimensional si -module, where we use the
adjoint action of si on g. It is routine to check that the map g ⊗ N → L(λ)
sending x ⊗ v to x · v is an si -module homomorphism, so its image lies in
M . Thus M is a nonzero g-submodule of the simple module L(λ), forcing
L(λ) = M .
(4) Since each v ∈ L(λ) lies in a finite dimensional si -submodule, by
(3), it follows that each xi or yi acts on L(λ) as a locally nilpotent op-
erator. If we denote the representation afforded by L(λ) as ϕ : U (g) →
End L(λ), it now makes sense to patch together the unipotent operators
exp ϕ(xi ) exp ϕ(−yi ) exp ϕ(xi ) which act on finite dimensional si -submodules.
This yields a well-defined automorphism ri of L(λ).
(5) If µ is any weight of L(λ), we claim that ri (L(λ)µ ) = L(λ)si µ for all
i. Indeed, the weight space L(λ)µ lies in a finite dimensional si -submodule
by step (3). This in turn is a weight module, thanks to the observation just
before the theorem. Now the claim follows from the sl(2, C) theory in 0.9.
(6) Because W is generated by the simple reflections si , step (5) implies
that all weight spaces L(λ)wµ have the same dimension. Since λ is dominant
integral, by assumption, all weights of L(λ) are therefore W -conjugates of
dominant integral weights µ ≤ λ. But there exist only finitely many of
these (0.6). It follows that L(λ) has only finitely many weights, thus is
finite dimensional.
22 1. Category O: Basics

(7) From the proof it also follows that dim L(λ)µ = dim L(λ)wµ for all
w ∈ W whenever dim L(λ) < ∞. 

Variants of this proof are found in numerous texts, including Bourbaki

[46, Chap. VIII, §7], Carter [60, Chap. 10], Humphreys [126, 21.2].
Exercise. Let M be a highest weight module of weight λ ∈ Λ+ . If all xi
and yi (1 ≤ i ≤ `) act locally nilpotently on M (which is automatic for the
xi ), rework steps in the proof above to show that dim M < ∞. Conclude
that M ∼ = L(λ). (This situation comes up in 2.6 below.)
Example. It is easy to recover from step (2) of the proof the familiar fact
that L(0) is the trivial one-dimensional g-module: the subalgebras si (which
generate g) act trivially on v + .

Coupled with Weyl’s complete reducibility theorem (0.7), the theorem

gives a complete parametrization of the finite dimensional modules in O.
Of course, there are many concrete features yet to be investigated: notably,
formal characters and dimensions. These will be worked out in Chapter
2 after we assemble more information about O. Here we just note a few
standard facts for later use in the case λ ∈ Λ+ :
• Let µ := wλ with w ∈ W . For any α ∈ Φ, not both µ − α and µ + α
can occur as weights of L(λ). [Recall that W permutes the weights
of L(λ). If β := w−1 α, not both λ + β and λ − β can be < λ.]
• If µ and µ + kα (with k ∈ Z, α ∈ Φ) are weights of L(λ), then so are
all intermediate weights µ + iα. [View L(λ) as an sl(2, C)-module
and use the theory in 0.9.]
• The dual space L(λ)∗ , with the standard action (x·f )(v) = −f (x·v)
for x ∈ g, v ∈ L(λ), f ∈ L(λ)∗ , is isomorphic to L(−w◦ λ) (where
w◦ ∈ W is the longest element). [Observe that L(λ)∗ is again
simple; its weights relative to h are the negatives of those for L(λ).
On the other hand, w◦ λ is the lowest weight of L(λ).]

1.7. Action of the Center

Our discussion of Verma modules has shown the existence of a unique max-
imal submodule and simple quotient, but beyond this (and Proposition 1.4)
we have no insight into the submodule structure. Without using deeper in-
formation about the universal enveloping algebra we cannot say much more.
Indeed, our arguments so far have relied mainly on the triangular decom-
position g = n− ⊕ h ⊕ n and the resulting PBW Theorem. But this kind
of set-up is found much more generally in infinite dimensional Kac–Moody
algebras (see Carter [60], Kac [166], Moody–Pianzola [223]); here there
1.7. Action of the Center 23

is a good analogue of the category O, but for example “Verma modules”

typically fail to have finite Jordan–Hölder length.
To prove a finite length theorem in our situation, we have to look closer
at the action of the center Z(g) of U (g). By contrast, the study of finite di-
mensional modules (including the proof of Weyl’s complete reducibility the-
orem), requires only the action of the Casimir element of Z(g) (0.5). Recall
from Theorem 1.1(e) that every M ∈ O is locally finite as a Z(g)-module.
The situation simplifies when M is a highest weight module, generated by
a maximal vector v + of weight λ. If z ∈ Z(g) and h ∈ h, we have
h · (z · v + ) = z · (h · v + ) = z · (λ(h)v + ) = λ(h)z · v + .
Since dim Mλ = 1, this forces z · v + = χλ (z)v + for some scalar χλ (z) ∈ C.
In turn, z acts on an arbitrary element u · v + of M (with u ∈ U (n− )) by the
same scalar, since zu = uz.
For fixed λ, the function z 7→ χλ (z) defines an algebra homomorphism
χλ : Z(g) → C, whose kernel is a maximal ideal of Z(g). We call it the
central character associated with λ. More generally, we call any algebra
homomorphism Z(g) → C a central character. The set of all these is in
natural bijection with the set Max Z(g) of maximal ideals.
To describe χλ more concretely, write z ∈ Z(g) as a linear combination
of PBW monomials based on the decomposition g = n− ⊕ h ⊕ n. Any
monomial having a nonzero factor from n will kill v + . In turn, factors from
h just multiply v + by scalars, while n− then takes v + to vectors of lower
weight. The upshot is that z · v + depends just on the PBW monomials with
factors in h. Denoting by pr : U (g) → U (h) the projection onto the subspace
U (h) given by setting all other monomials equal to 0, we see that
χλ (z) = λ(pr(z)) for all z ∈ Z(g).
Here the linear function λ is extended canonically to an algebra homomor-
phism λ : U (h) → C.
T
Using the obvious fact that λ∈h∗ Ker λ = 0, it follows that the restric-
tion of pr to Z(g) is an algebra homomorphism. Denote this restriction by
ξ; it is called the Harish-Chandra homomorphism.
Remark. To understand more intrinsically why ξ is a homomorphism (as
in the original study of the center by Harish-Chandra), without resort to
representation theory, look at the centralizer U (g)0 of h in the algebra U (g).
This subalgebra obviously includes Z(g) as well as U (h). As the notation
suggests, it is actually the 0-graded subspace of U (g) relative to the natural
grading induced by the extension to U (g) of the adjoint representation of g
(0.5). In turn, L := U (g)n ∩ U (g)0 is seen to coincide with n− U (g) ∩ U (g)0 .
This is a two-sided ideal of U (g), complementary to the subspace U (h) of
24 1. Category O: Basics

U (g)0 , and the resulting projection homomorphism U (g)0 → U (h) restricts

to ξ : Z(g) → U (h).

Two natural questions arise at this point: (1) Is the Harish-Chandra

homomorphism injective? (2) What is the image of ξ in U (h)?

1.8. Central Characters and Linked Weights

Since χλ = χµ whenever L(µ) occurs as a subquotient of M (λ), it is impor-
tant to determine precisely when this equality of central characters occurs.
It is also natural to ask whether there are any other central characters be-
sides the χλ . To deal with both of these issues we have to formalize better
the relationship here between λ and µ, then bring the Harish-Chandra ho-
momorphism into the picture.
A valuable clue is provided by Proposition 1.4: If n := hλ, α∨ i ∈ Z+ for
some α ∈ ∆, then M (λ) has a maximal vector of weight λ − (n + 1)α < λ.
In this case, we have χλ = χµ where µ := λ − (n + 1)α. To express this more
suggestively, recall from 0.6 that sα ρ = ρ − α whenever α is simple. Thus
sα (λ + ρ) = λ − hλ, α∨ iα + ρ − α = µ + ρ, or sα (λ + ρ) − ρ = µ.

Definition. For w ∈ W and λ ∈ h∗ , define a shifted action of W (called

the dot action) by w · λ = w(λ + ρ) − ρ. If λ, µ ∈ h∗ , we say that λ and µ
are linked (or W -linked) if for some w ∈ W , we have µ = w · λ. Linkage is
clearly an equivalence relation on h∗ . The orbit {w · λ | w ∈ W } of λ under
the dot action is called the linkage class (or W -linkage class) of λ.

For example, the fixed point −ρ under the dot action lies in a linkage
class by itself (and no other weight does). The usual notion of regular
weight for λ ∈ h∗ , requiring that the isotropy group of λ in W be trivial, or
|W λ| = |W |, is replaced in the setting of linkage classes by a shifted notion
of regular weight (which might also be called dot-regular): the weight
λ ∈ h∗ is regular if |W · λ| = |W |, or in other words, hλ + ρ, α∨ i = 6 0 for
all α ∈ Φ. Weights which are not regular may be called singular. In this
sense, 0 is regular while −ρ is singular.
In general, each λ + ρ ∈ Λ is W -conjugate (by 0.6) to a unique element
in the closure C of the Weyl chamber

C := {µ ∈ E | hµ, α∨ i > 0 for all α ∈ ∆}.

So the linkage class of λ has a unique element in C − ρ; this weight is then

the unique maximal element in its linkage class. (For arbitrary λ ∈ h∗ , we
shall fine-tune the parametrization in Chapter 3.)
1.9. Harish-Chandra Homomorphism 25

Exercise. Unlike the usual action of W on h∗ , the dot action is not additive.
If λ, µ ∈ h∗ and w ∈ W , verify that
w · (λ + µ) = w · λ + µ
w · λ − w · µ = w(λ − µ).

This language allows us to reformulate Proposition 1.4:

If α ∈ ∆, λ ∈ h∗ and hλ + ρ, α∨ i = 0, then M (sα · λ) =
M (λ). If hλ + ρ, α∨ i ∈ Z>0 , then there exists a nonzero
homomorphism M (sα · λ) → M (λ) with image in N (λ).
Now we can strengthen somewhat our earlier conclusion:
Proposition. If λ ∈ Λ and µ is linked to λ then χλ = χµ .

Proof. Start as before with a simple root α. Since λ ∈ Λ, n := hλ, α∨ i ∈ Z.

In case n ∈ Z+ , we already saw that χλ = χµ with µ = sα · λ. Suppose on
the other hand n ∈ Z<0 . If n = −1, then sα · λ = λ and there is nothing to
prove. If n < −1, then setting µ := sα · λ (which forces sα · µ = λ), we have
hµ, α∨ i = −n − 2 ≥ 0. So by the first case, χλ = χµ .
Now W is generated by the simple reflections (0.3), while the linkage
relation is transitive, so the proposition follows by induction on `(w). 

1.9. Harish-Chandra Homomorphism

In order to remove the restriction in Proposition 1.8 that λ be an integral
weight, we can invoke a simple density argument in affine algebraic geom-
etry. Here we view h∗ as the affine space A` over C, while U (h) = S(h)
(identified in turn with P (h∗ )) is the algebra of polynomial functions in `
variables acting naturally on A` . To make this concrete, take for example
the fundamental weights $1 , . . . , $` as a basis for h∗ , along with generators
h1 , . . . , h` for S(h).
Whereas Λ is a discrete subspace of h∗ in the usual topology, it becomes
dense in the Zariski topology; here the closed sets are the zero sets of finite
sets of polynomials. Recall the argument: In A` we can identify Λ with Z` .
Now use induction on ` to see that a polynomial function f on A` vanishing
on Z` must be zero: If ` = 1, use the fact that a nonzero polynomial can
have only finitely many roots. If ` > 1, write f as a polynomial in the last
variable. Substituting fixed integers for the first ` − 1 variables produces a
polynomial in one variable vanishing on Z (therefore zero). So the induction
hypothesis for `−1 can be applied, showing that f = 0. From this argument
we conclude that Z` is dense in A` .
We know that χλ = χw·λ for all w ∈ W when λ ∈ Λ. Since χλ (z) =
λ(ξ(z)) for z ∈ Z(g), this translates into the statement that the polynomial
26 1. Category O: Basics

functions ξ(z) and w · ξ(z) agree on Λ. By density, this forces them to agree
everywhere on h∗ . In other words, Proposition 1.8 is true for all λ ∈ h∗ .
Using this we can say more about the image of ξ. To take account
of the ρ-shift involved in the dot action, first compose ξ with the algebra
automorphism of S(h) induced by the substitution p(λ) 7→ p(λ−ρ). Call this
composite homomorphism ψ : Z(g) → S(h) the twisted Harish-Chandra
homomorphism. Now we have
χλ (z) = (λ + ρ)(ψ(z)) for all z ∈ Z(g).
We have shown that the polynomial functions on h∗ in the image of ψ are
constant on W -orbits (linkage classes), in other words lie in the algebra
S(h)W of W -invariants. To summarize:
Theorem. (a) Whenever λ, µ ∈ h∗ are W -linked, χλ = χµ .
(b) The image of the twisted Harish-Chandra homomorphism ψ : Z(g) →
U (h) = S(h) lies in the subalgebra S(h)W of W -invariant polyno-
mials. 
Example. Let g = sl(2, C) and identify h∗ as usual with C via λ 7→ λ(h).
Recall from 0.5 that a Casimir element in Z(g) is given explicitly by c =
2xy + h2 + 2yx = h2 + 2h + 4yx. Recalling that ρ(h) = 1, we get ψ(c) =
(h − 1)2 + 2(h − 1) = h2 − 1, which visibly lies in S(h)W . In turn, χλ (c) =
(λ + ρ)(h2 − 1) = (λ + 1)2 − 1. It follows that χλ = χµ if and only if λ and
µ are linked: µ = λ or µ = −λ − 2.
Exercise. Show that the homomorphism ψ is independent of the choice of a
simple system in Φ. [Any simple system has the form w∆ for some w ∈ W .]

1.10. Harish-Chandra’s Theorem

What we really need is the converse of Theorem 1.9(a), in order to get
a workable necessary condition for L(λ) and L(µ) to occur as composition
factors of a Verma module. This is contained in the following basic theorem.
Theorem (Harish-Chandra). Let ψ : Z(g) → S(h) be the twisted Harish-
Chandra homomorphism.
(a) The homomorphism ψ is an isomorphism of Z(g) onto S(h)W .
(b) For all λ, µ ∈ h∗ , we have χλ = χµ if and only if µ = w · λ for some
w ∈ W.
(c) Every central character χ : Z(g) → C is of the form χλ for some
λ ∈ h∗ .

(a) This is the key point. Its proof depends mainly on arguments involv-
ing the enveloping algebra, independent of category O. Here we just outline
1.10. Harish-Chandra’s Theorem 27

the basic ideas, referring in the Notes below to several thorough discussions
in the literature.
We have seen that ψ maps Z(g) into the algebra S(h)W . Instead of
trying to prove injectivity and surjectivity directly, it is better to make a
comparison with a similar but somewhat more transparent map. Consider
the algebra P (g) of polynomial functions on the vector space g, which is
naturally isomorphic to S(g∗ ). The restriction map θ : P (g) → P (h) is
an algebra homomorphism. Next introduce the adjoint group G ⊂ Aut g
generated by the operators exp ad x with x ∈ g nilpotent. (Since ad x is
nilpotent, the exponential power series is just a polynomial.) The group G
is a Lie group (or a semisimple algebraic group if C is replaced by another
algebraically closed field of characteristic 0). It acts naturally on P (g) as
a group of automorphisms, with algebra of fixed points denoted P (g)G .
Similarly, W acts on P (h).
In this setting Chevalley proved a fundamental Restriction Theorem:

θ maps P (g)G isomorphically onto P (h)W .

Since the Killing form is nondegenerate on g and restricts to a nondegenerate

form on h, we can further identify P (g) with S(g) and P (h) with S(h).
The proof of Chevalley’s theorem involves some use of finite dimensional
representation theory: P (g)G is generated by traces of powers of operators
representing elements of g in such representations.
At this point the rewritten isomorphism S(g)G → S(h)W somewhat
resembles the Harish-Chandra homomorphism ξ. In fact, S(g)G can even be
identified naturally as a vector space with Z(g), though not as an algebra.
Even though the Chevalley map θ does not coincide with ξ (as illustrated
when g = sl(2, C) in [126, 23.3]), one gets enough information from the
comparison to see that ξ is bijective and thus to complete the proof. The
trick is to see that the various maps here are compatible with the natural
filtrations of the algebras, then observe that the induced graded versions of
ξ and the Chevalley isomorphism agree.
(b) Suppose that for some λ, µ ∈ h∗ , the linkage classes of λ and µ are
disjoint. Using Lagrange interpolation, find a polynomial function f on h∗
which takes value 1 on W · λ and 0 on W · µ. Then replace f by its “average”
1 X
wf
|W |
w∈W

to get a W -invariant g having the same values on these dot orbits. Using the
assumption about ψ, take any pre-image z ∈ Z(g) of g under ψ. Then check
that χλ (z) = λ(ψ(z)) = g(λ) = 1, whereas χµ (z) = µ(ψ(z)) = g(µ) = 0.
This forces χλ 6= χµ .
28 1. Category O: Basics

(c) Here the argument relies on some standard commutative algebra. To

the given χ corresponds (via ψ) a homomorphism ϕ : S(h)W → C. Since W
is a finite group, it is easy to see that S(h) is an integral extension
Q of the ring
of invariants: any f ∈ S(h) is a root of the monic polynomial w∈W (t − wf )
(with t an indeterminate), whose coefficients clearly lie in S(h)W . So a stan-
dard consequence of the Going Up Theorem in commutative algebra can be
invoked (our field being algebraically closed): ϕ extends to a homomorphism
b : S(h) → C.
ϕ
This translates into the desired conclusion as follows. Since S(h) is the
algebra of polynomial functions on h∗ , any homomorphism S(h) → C such
b is the evaluation map f 7→ f (λ) for some λ ∈ h∗ . (The kernel of the
as ϕ
homomorphism is the obvious maximal ideal vanishing at the point λ.) In
turn, χ(z) = λ(ψ(z)) = χλ (z) for all z ∈ Z(g), forcing χ = χλ . 
Remark. Chevalley made more precise the structure of the algebra S(h)W
in the context of his study of invariants under arbitrary finite reflection
groups: S(h)W is isomorphic to a polynomial algebra in ` variables, which
can be generated by homomogenous elements of the polynomial algebra
S(h) of degrees d1 , . . . , d` satisfying d1 · · · d` = |W |. The smallest degree is
2. (These di have other interesting interpretations as well.) See for example
Bourbaki [45, V, §5], Humphreys [130, Chap. 3].
Exercise. Prove that the transpose map τ (see 0.5) fixes Z(g) pointwise.
[Check that τ commutes with the Harish-Chandra homomorphism ξ and use
the fact that ξ is injective.]

1.11. Category O is Artinian

Now it is easy to see that category O is artinian as well as noetherian: each
M ∈ O is both artinian and noetherian.
Theorem. Each M ∈ O is artinian (as well as noetherian). Moreover,
dim HomO (M, N ) < ∞ for all M, N ∈ O.

Proof. Thanks to Corollary 1.2, it is enough P to prove the first statement

when M = M (λ) is a Verma module. Let V := w∈W M (λ)w·λ , so dim V <
∞. Suppose N ⊃ N 0 is a proper inclusion of submodules of M (λ), so Z(g)
acts on the quotient N/N 0 by the character χλ . The nonzero module N/N 0
has a maximal vector of some weight µ ≤ λ, so χµ = χλ . But then by
Theorem 1.10(b), µ = w · λ for some w ∈ W , forcing N ∩ V 6= 0 and
dim(N ∩ V ) > dim(N 0 ∩ V ). It follows that any properly descending chain
of submodules in M (λ) terminates after finitely many steps.
Now let M, N ∈ O. Starting with the fact that dim EndO L(λ) = 1
for all λ ∈ h∗ (Theorem 1.2(g)), induction on the length of M shows
1.11. Category O is Artinian 29

that dim HomO M, L(λ) < ∞. In turn, a similar induction shows that
dim HomO (M, N ) < ∞. 

As a consequence of the theorem, each M ∈ O possesses a composition

series with simple quotients isomorphic to various L(λ). The multiplicity
of L(λ) is independent of the choice of composition series and is denoted
by [M : L(λ)]. For example, [M (λ) : L(λ)] = 1 in view of the fact that
dim M (λ)λ = 1. We call the common length of all such composition series
the length of M (also known as Jordan–Hölder length).

Example. As we observed in 1.8, the weight −ρ lies in a linkage class by

itself. By the theorem, M (−ρ) has no composition factor of lower weight;
thus M (−ρ) = L(−ρ).

Since O is artinian and noetherian, a standard theorem (associated in

various contexts with the names Azumaya, Krull, Remak, Schmidt) is valid
here: each M ∈ O can be written as a direct sum of indecomposable mod-
ules, with the summands being unique up to isomorphism and order. This
will be made more precise in the following section.
Similarly, the chain conditions imply a straightforward description of
the Grothendieck group K(O), defined to be the quotient of the free
abelian group on symbols [M ] (for M ∈ O) by the subgroup generated by
all [B]−[A]−[C] for which 0 → A → B → C → 0 is a short exact sequence in
O. Whenever M, N ∈ O, it follows from the construction that [M ] = [N ] if
and only if M and N have the same composition factor multiplicities. Note
too that any exact functor O → O induces an endomorphism of K(O): for
example, tensoring with a fixed finite dimensional module (Theorem 1.1(d)).
In our case, K(O) has a Z-basis consisting of the symbols [L(λ)], where
λ ∈ h∗ . Indeed, for any M ∈ O, the corresponding element [M ] of K(O) can
be written uniquely as a (finite!) Z+ -linear combination of this basis; here
the coefficient of [L(λ)] is the composition factor multiplicity [M : L(λ)].
Although we often have to settle for formal information about modules
in O, the fact that modules have finite length encourages us to look for
some natural submodules. For example, the socle of a module M is the
sum (automatically direct) of all simple submodules of M , denoted Soc M ;
it is the largest semisimple submodule of M . The radical of M is the
intersection of all maximal submodules in M , denoted Rad M . It may also be
characterized as the smallest submodule for which M/ Rad M is semisimple;
this quotient is called the head, denoted Hd M . For example, Rad M (λ) =
N (λ) and Hd M (λ) ∼ = L(λ).
30 1. Category O: Basics

1.12. Subcategories Oχ
The action of Z(g) on an arbitrary M ∈ O is usually complicated. But
in view of Corollary 1.2 we can expect this action to involve only a finite
number of central characters. To make this precise, define a subspace of M
for each fixed χ by

n
M χ := {v ∈ M | z − χ(z) · v = 0 for some n > 0 depending on z}.
In words, each element of Z(g) acts on v as the scalar χ(z) plus a nilpotent
operator. It is clear that M χ is a U (g)-submodule of M , while the subspaces
M χ are independent.
Now Z(g) stabilizes each weight space Mµ , since Z(g) and U (h) com-
mute. Since we are working over an algebraically closed field, a standard
result from
L linear algebra on families of commuting operators implies that
χ
Mµ = χ (Mµ ∩ M ). Because M is generated by finitely many weight
vectors, it must therefore be the direct sum of finitely many nonzero sub-
modules M χ . Thanks to Harish-Chandra’s Theorem, each central character
χ occurring in this way must be of the form χλ for some λ ∈ h∗ .
Denote by Oχ the (full) subcategory of O whose objects are the modules
M for which M = M χ . From the discussion above we conclude:
Proposition. Category O is the direct sum of the subcategories Oχ as χ
ranges over the central characters of the form χλ . Therefore each inde-
composable module lies in a unique Oχ . In particular, each highest weight
module of weight λ lies in Oχλ . 
L
The decomposition O = χ Oχ is a major step in the direction of reduc-
ing the study of category O to the study of more manageable subcategories.
In view of the proposition, each indecomposable module lies in a single Oχ .
Moreover, Oχ involves only finitely many simple modules and the corre-
sponding Verma modules. (In Chapter 3 projectives and injectives will be
added to this list.)
Exercise. Fix a central character χ and let {V (λ) } be a collection of mod-
ules in Oχ indexed by the weights λ for which χ = χλ and satisfying: (1)
(λ)
dim Vλ = 1; (2) µ ≤ λ for all weights µ of V (λ) . Then the symbols
[V (λ) ] form a Z-basis of the Grothendieck group K(Oχ ). For example, take
V (λ) = M (λ) or L(λ).

1.13. Blocks
In module categories which fail to be semisimple, the notion of block often
helps to organize modules. Blocks play a major role, for example, in the
modular representation theory of finite groups as developed by Brauer and
1.13. Blocks 31

others. In that theory the blocks are as efficient as possible in separating

indecomposable modules which are unrelated homologically.
In general, one looks at the natural Ext functors on a module category
which is both artinian and noetherian. If two simple modules M1 and M2
can be extended nontrivially, i.e., if there is a nonsplit short exact sequence
0 → Mi → M → Mj → 0 with {i, j} = {1, 2}, then we place M1 and M2
in the same block. More generally, if simple modules M and N belong to a
finite sequence M = M1 , M2 , . . . , Mn = N such that adjacent pairs belong
to the same block, we place M and N in the same block. This partitions
the simple modules into blocks.
Now if M is arbitrary, we say it belongs to a block if all its composition
factors do. A formal argument (which we leave to the reader) shows that
M decomposes uniquely as a direct sum of submodules, each belonging
to a single block. In particular, each indecomposable module belongs to a
single block. (See for example Jantzen [153, II.7.1]; his finite dimensionality
assumption can be replaced here by the chain conditions on modules.)
Our decomposition of O using central characters obviously goes in this
direction: in the partition of modules just described, each block of O will
lie in some Oχ . When the weights are all integral, these are precisely the
blocks:
Proposition. If λ ∈ Λ, the subcategory Oχλ is a block of O.

Proof. It just has to be shown that all L(w · λ) lie in the same block. If
α ∈ ∆, set µ := sα · λ. Unless µ = λ, we may assume µ < λ (otherwise
reverse the roles of λ and µ); cf. the proof of Proposition 1.8. We know from
Proposition 1.4 that there is a nonzero homomorphism M (µ) → N (λ) ⊂
M (λ). Factoring out N (µ) (and its image N in M (λ)), this induces an
embedding of L(µ) into M (λ)/N , which is a highest weight module with
quotient L(λ). Since highest weight modules are indecomposable, it follows
that L(λ) and L(µ) lie in the same block.
Now iterate this procedure, using a reduced expression for w ∈ W , to
conclude that L(λ) and L(w · λ) lie in the same block. 

When χ = χ0 , Oχ is called the principal block (sometimes denoted

O0 ), by analogy with the case of modules for a finite group.
This neat description of blocks breaks down when nonintegral weights
are involved, as the example g = sl(2, C) shows: a typical Verma module
M (λ) with λ positive but not integral will be simple, even though −λ − 2 is
linked to λ (1.5). The problem here is that the linkage class intersects more
than one coset of Λr in h∗ . The following easy exercise gives a refinement of
the linkage classes in the nonintegral case, which will be studied further in
32 1. Category O: Basics

3.5 below. But a definitive description of blocks requires more information

about the submodule structure of Verma modules (4.9).

Exercise. Suppose λ 6∈ Λ, so the linkage class W · λ is the disjoint union

of its nonempty intersections with various cosets of Λr in h∗ . Prove Lthat
each M ∈ Oχλ has a corresponding direct sum decomposition M = Mi ,
in which all weights of Mi lie in a single coset. [Recall Exercise 1.1(b).]

Caution: In the category O literature the term “block” is sometimes

defined more broadly to mean a subcategory Oχ .

1.14. Formal Characters of Finite Dimensional Modules

Beyond the parametrization of simple modules L(λ) in O by highest weights,
how much can one expect to say about their internal structure? In the tradi-
tional finite dimensional representation theory of finite groups, compact Lie
groups, or associated Lie algebras, a natural goal is to write down explicitly
the matrices giving the action of individual group or Lie algebra elements
on a conveniently chosen basis. This is especially useful for applications in
classical physics or chemistry, but can be carried out effectively mainly in
low ranks or other special cases such as symmetric or exterior powers of a
natural representation.
A more modest goal in finite dimensional representation theory is to
determine the characters of irreducible representations: the traces of repre-
senting matrices (which are independent of the choice of basis). Here the
theory is well developed. In well-behaved cases, the character of a represen-
tation determines it uniquely up to equivalence. For a semisimple Lie algebra
g over C, the notion of character is formulated more abstractly: here it is
enough to keep track of multiplicities of weights, which indirectly determine
traces of matrices representing group elements in associated highest weight
representations. (In a compact Lie group, each element is conjugate to a
semisimple one.)
It is easy to introduce a suitable formal character in the finite dimen-
sional setting. Start with the integral group ring ZΛ of the weight lattice
Λ. Although the group structure of Λ is usually written additively, this
would confuse the situation in the group ring where the elements are formal
Z-linear combinations of group elements. So we associate to each weight λ
a symbol e(λ), suggestive of an exponential. Given a (finite dimensional!)
U (g)-module M , which is automatically a weight module, its formal char-
acter can then be defined as
X
ch M := dim Mλ e(λ).
λ∈Λ
1.15. Formal Characters of Modules in O 33

Using the ring structure of ZΛ, we have ch(M ⊗ N ) = ch M ch N . Thanks

to Weyl’s complete reducibility theorem, all such characters are known once
we find ch L(λ). In Chapter 2 the classical results on these characters will
be recovered in the more general setting of category O.

1.15. Formal Characters of Modules in O

In order to make effective use of Verma modules, we need to introduce the
language of formal characters in the infinite dimensional setting. The idea is
obvious: write down a formal sum as above, which no longer need be finite.
To manipulate such formal sums rigorously, we reformulate the definition
as follows. Specifying dimensions of weight spaces amounts to assigning a
nonnegative integer to each weight of a module M ∈ O. Thus we view ch M
as a Z+ -valued function chM on h∗ . (More generally, any module in the
larger category of weight modules having finite dimensional weight spaces
can be assigned a formal character in this way.)
In place of the basis element e(λ) of ZΛ we now have a characteristic
function eλ which takes value 1 at λ and value 0 at µ 6= λ. In the finite di-
mensional case the functions corresponding to elements of ZΛ vanish outside
a finite subset of Λ, allowing the multiplication in ZΛ to be reinterpreted as
the convolution product:
X
(f ∗ g)(λ) := f (µ) g(ν).
µ+ν=λ

To deal with modules in O, we simply carry over this definition to the

additive group X of functions f : h∗ → Z whose support (outside which f
is 0) lies in a finite union of sets of the form λ − Γ. Obviously all functions
eλ (with λ ∈ h∗ ) lie in X . It is easy to check that X is a commutative ring
under convolution, in which e0 plays the role of multiplicative identity while
eλ eµ = eλ+µ for λ, µ ∈ h∗ .
For typographical reasons we continue to write ch M (rather than chM )
when M is an arbitrary weight module, viewing this as the function whose
value at λ ∈ h∗ is dim Mλ . Similarly, we may write e(λ) in place of eλ .
When convenient we also express ch M as a (possibly infinite) formal linear
combination of the e(λ). For example,
X
ch L(λ) = e(λ) + mλµ e(µ) for some mλµ ∈ Z+ .
µ<λ

Denote by X0 the additive subgroup of X generated by all ch M .

Proposition. Formal characters of modules in O have the following prop-

erties.
34 1. Category O: Basics

(a) If 0 → M 0 → M → M 00 → 0 is a short exact sequence in O, we have

ch M = ch M 0 + ch M 00 . Thus any ch M is determined uniquely by
the formal characters of the composition factors of M , taken with
multiplicity.
(b) X0 can be identified naturally with the Grothendieck group K(O) by
sending ch M 7→ [M ].
(c) If M ∈ O and dim L < ∞, then ch(L ⊗ M ) = ch L ∗ ch M .

Proof. (a) This follows from the fact that dim Mµ = dim Mµ0 + dim Mµ00 .
Then (b) is an easy consequence.
(c) Recall from Theorem 1.1(d) that L ⊗ M also lies in O, so its formal
character is defined. The stated formula is clear from the distributive law
when formal characters are thought of as formal sums, as in the finite di-
mensional setting. To make
P this precise in general using convolution, just
observe that (L ⊗ M )λ = µ+ν=λ Lµ ⊗ Mν . 
Remark. In the setting of ZΛ, the natural action of W on Λ extends to an
action on ZΛ. Each character of a finite dimensional module such as L(λ)
with λ ∈ Λ+ is then W -invariant (Theorem 1.6): w ch L(λ) = ch L(λ). This
is no longer true in general for X , which creates an obstacle to using Verma
modules as a tool in the proof of Weyl’s Character Formula (Chapter 2).

1.16. Formal Characters of Verma Modules

What is the formal character ch M (λ) of a Verma module? This is easily
computed. In fact, all such characters can be described uniformly in terms
of a single function p ∈ X : letPp(γ) equal the number of families (cα )α>0
with cα ∈ Z+ for which γ = − α>0 cα α. Call this the Kostant function
(though it is actually the negative of the “partition function” P originally
introduced by Kostant in his study of weight multiplicities for finite di-
mensional modules). By definition, p vanishes outside Λr . Now p is just
the formal character
Q of M (0): indeed, dim M (0)ν is the number of distinct
PBW monomials α>0 yαcα in U (n− ) (for a fixed ordering of Φ+ ) which take
a maximal vector v + into the weight space M (0)ν . Similar reasoning shows
in general:
Proposition. For any λ ∈ h∗ , we have ch M (λ) = p ∗ e(λ). In particular,
ch M (0) = p. 

While the formal characters of Verma modules are quite transparent, it

is far from clear how to determine ch L(λ) for arbitrary λ ∈ h∗ . This is a
fundamental problem in the study of O. To go beyond the classical finite
dimensional case treated below in Chapter 2, we will need to undertake
a deeper study of Verma modules in Chapters 4–5. But ultimately the
Notes 35

computation of formal characters of all simple modules in O turns out to

require a transition to a more sophisticated geometrically defined category
(Chapter 8 below).
Working in X0 does make it possible to reformulate the problem in a
useful way, taking advantage of Harish-Chandra’s Theorem: first write
X
(1) ch M (λ) = a(λ, µ) ch L(µ) with a(λ, µ) ∈ Z+ and a(λ, λ) = 1.
µ

Here µ ranges over weights ≤ λ and linked to λ by W , while a(λ, µ) =

[M (λ) : L(µ)]. The partial ordering permits us to invert the resulting trian-
gular system of equations:
X
(2) ch L(λ) = b(λ, µ) ch M (µ) with b(λ, µ) ∈ Z and b(λ, λ) = 1.
µ
The sum is again taken over weights µ ≤ λ linked to λ. To make the role of
W more explicit we can recast this as:
X
(3) ch L(λ) = b(λ, w) ch M (w · λ) with b(λ, w) ∈ Z, b(λ, 1) = 1.
w·λ≤λ
This last format turns out to be most useful in practice, though the deter-
mination of the coefficients is typically quite subtle. Even in the case λ = 0,
where the left side is known to be equal to e(0), it is nontrivial to fill in the
right side.
Remark. In the simplest case, when g = sl(2, C), it is easy to see that for
λ ≥ 0 we have ch L(λ) = ch M (λ) − ch M (sα · λ), where sα · λ = −λ − 2 (see
1.5 and Example 1.9). Thus b(λ, w) = ±1 is independent of the choice of
dominant λ. On the other hand, for λ < 0 we have b(λ, w) = 1 for w = 1
and 0 otherwise; here again the result is independent of λ. This raises the
question in general as to how b(λ, w) depends on each of λ and w. For
λ ∈ Λ+ the answer will be worked out in the next chapter, but in other
cases deeper ideas are involved.
Exercise. Show that the formal characters ch L(λ) with λ ∈ h∗ are linearly
independent in X and form a Z-basis of the additive subgroup X0 of X
generated by all ch M , while the formal characters of Verma modules provide
an alternative Z-basis. Similar statements hold in the Grothendieck group
K(O) for the corresponding elements [L(λ)] and M (λ)].

Notes
The axiomatic treatment of category O originates in the papers [25, 26,
27] of Bernstein–Gelfand–Gelfand (abbreviated as BGG in what follows).
The letter chosen to label the category is the first letter of a Russian word
meaning “basic”.
36 1. Category O: Basics

Concerning the notion of “induced module” (1.3) see Dixmier [84, 5.1].
Step (3) in the proof of Theorem 1.6 is adapted from Dixmier [84,
Prop. 1.7.9], as suggested by Jantzen.
The structure of Z(g), together with applications to representation the-
ory (1.7–1.10), is explained in numerous sources, starting with the funda-
mental paper of Harish-Chandra [122]. See the texts (in roughly chrono-
logical order): Humphreys [126, §23], Dixmier [84, 7.4–7.5], Varadarajan
[250, 4.10], Bourbaki [46, VIII, §8], Carter [60, Chap. 11].
(1.7) We have emphasized highest weight modules and their central char-
acters. In fact, every simple U (g)-module M satisfies dim EndU (g) M = 1
and therefore has a central character. This is a general fact (Quillen’s
Lemma) for a finite dimensional Lie algebra over an algebraically closed
field; see Dixmier [84, 2.6.8].
 Chapter 2

Characters of Finite
Dimensional Modules

As promised in Chapter 1, we pause here to look more closely at the subcat-

egory of O consisting of finite dimensional modules. These are direct sums
of various L(λ) with λ ∈ Λ+ , so the main question at first is to work out
the formal characters and dimensions of these simple modules. To derive
the classical formulas of Weyl and Kostant, we follow the approach of BGG
[25] in studying L(λ) formally as a Z-linear combination of Verma modules.

2.1. Summary of Prerequisites

In addition to the general background recalled in Chapter 0, our treatment
of the finite dimensional theory requires most of the machinery developed
in Chapter 1—but only in the case of integral weights.
The construction of simple modules as quotients of Verma modules is
essential, together with the language of formal characters.
From 1.6 one has to know that finite dimensional simple modules have
highest weights in Λ+ . This is the easier half of the classification theorem,
but of course the other half is needed to insure the existence of enough such
modules to make the theory nontrivial.
One substantial shortcut could be taken here to bypass the use of Harish-
Chandra’s Theorem on central characters. As explained in detail in the
Appendix to §23 in [126] (second and later printings), the study by V. Kac of
“integrable” modules for Kac–Moody Lie algebras relies just on the Casimir
elements associated with representations to separate highest weights. There
the full center of the enveloping algebra is both unwieldy and unnecessary for

37
38 2. Characters of Finite Dimensional Modules

this special class of modules. While the method of Kac can be used equally
well in the finite dimensional situation and cuts down the prerequisites, it
would be artificial in the wider context of category O to delay the discussion
of central characters.

2.2. Formal Characters Revisited

Beyond rank 1, it becomes impractical to write down simultaneously all
weight multiplicities of an arbitrary finite dimensional L(λ). Weyl’s classical
character formula captures the essential information indirectly but in closed
form by expressing ch L(λ) as a quotient in which the denominator depends
only on root system data. While this approach leaves the individual weight
multiplicities concealed, it has conceptual elegance and also leads to an easily
computable formula for dim L(λ).
The approach by BGG to Weyl’s formula requires some “thinking outside
the box”, the box in this case being the category of finite dimensional U (g)-
modules. The idea is already implicit in the formulation given in 1.16(3):
X
ch L(λ) = b(λ, w) ch M (w · λ), with b(λ, w) ∈ Z and b(λ, 1) = 1.
w∈W,w·λ≤λ

In the finite dimensional case, the sum here is taken over all of W : since
λ ∈ Λ+ (and ρ ∈ Λ+ ), we have w · λ ≤ λ for all w ∈ W (0.6).
In the simplest case g = sl(2, C), with Λ+ ∼
= Z+ , the situation is quite
transparent (0.9): here |W | = 2 and we find that
ch L(λ) = ch M (λ) − ch M (−λ − 2), with sα · λ = −λ − 2.
But in general it turns out to be a deep problem to determine all coefficients
occurring for arbitrary λ (even when λ ∈ Λ).
We know from Theorem 1.6 that the character of a finite dimensional
module is actually W -invariant, but in general this fails for modules in O. In
particular, we cannot apply W to the character of a Verma module without
leaving O; so there is no easy way to exploit the W -invariance in the formula
above. Instead a less direct way must be developed.

2.3. The Functions p and q

To investigate the formal character of a finite dimensional L(λ), we work
with two special functions in the ring X of formal characters.
Recall from 1.16 the Kostant function, defined by p(ν) = number of
tuples (cα )α>0 with cα ∈ Z+ and − α>0 cα α = ν; so p(ν) = 0 outside Λr .
P
There we proved:
p = ch M (0) and p ∗ e(λ) = ch M (λ).
2.3. The Functions p and q 39

Q apply the convolution operation ∗ in X to longer products,

Below we shall
writing just for convenience.
To express p in terms of the individual positive roots, we introduce
further elements fα ∈ X , α > 0. By definition,
(
1 if λ = −kα for some k ∈ Z+ ,
fα (λ) :=
0 otherwise.
Symbolically, fα = e(0) + e(−α) + e(−2α) + · · · . It is clear that fα ∈ X .
Lemma (A). With the above notation,
Q
(a) p = α>0 fα .


(b) e(0) − e(−α) ∗ fα = e(0).

Proof. (a) follows at once from the definition of convolution.

(b) Symbolically, e(0) − e(−α) ∗ e(0) + e(−α) + e(−2α) + . . . = e(0),
which is clear after cancelling other terms in pairs. (A formal proof is
easy.) 

Besides the function p, we need a function q whose purpose is not im-

mediately clear; it soon turns out to interact usefully with p. Define
Y

q := e(α/2) − e(−α/2) .
α>0

A moment’s thought shows X . Note that q can be rewritten

P that this lies in Q
using the fact that ρ = α α/2, or e(ρ) = α>0 e(α/2):
Y
Y

q= e(α/2) ∗ e(0) − e(−α) = e(ρ) ∗ e(0) − e(−α) .
α>0 α>0

This makes it obvious that q 6= 0: for example, q(ρ) = 1.

Q
Exercise.
Q Rewrite q as e(−ρ) ∗ α>0 (e(α) − 1), using the fact that e(−ρ) =
α>0 e(−α/2).

Now the special feature of q is that it is an alternating function relative

to the action of W :
Lemma (B). For all w ∈ W , we have wq = (−1)`(w) q.

Proof. In case `(w) = 0, w = 1 and there is nothing to prove. When

`(w) = 1, w = sα for some α ∈ ∆. Recall (0.3) that sα sends α to −α but
keeps all other positive roots in Φ+ . This shows from the rewritten formula
for q that sα q = −q. Now continue by induction on `(w). 

It is easy now to see how p and q interact:

40 2. Characters of Finite Dimensional Modules

Lemma (C). For arbitrary λ ∈ h∗ ,

q ∗ ch M (λ) = q ∗ p ∗ e(λ) = e(λ + ρ).

Proof. We just have to prove the second equality, which is equivalent to

q ∗ p = e(ρ). Combining Lemmas A and B, we compute
Y
Y


q∗p = e(0) − e(−α) ∗ e(ρ) ∗ p = e(0) − e(−α) ∗ fα ∗ e(ρ) = e(ρ).
α>0 α>0

2.4. Formulas of Weyl and Kostant

With these lemmas in hand it is easy to complete the proof of Weyl’s Char-
acter Formula, in tandem with Kostant’s Weight Multiplicity Theorem:
Theorem (Weyl). Let λ ∈ Λ+ , so dim L(λ) < ∞. Then
X
(−1)`(w) e w(λ + ρ) ,


q ∗ ch L(λ) =
w∈W
Q

where q = α>0 e(α/2) − e(−α/2) . In particular, when λ = 0,
X
q= (−1)`(w) e(wρ).
w∈W

Proof. This is just a matter of putting the pieces together, starting as in

2.2 with
X
ch L(λ) = b(λ, w) ch M (w · λ), where b(λ, w) ∈ Z and b(λ, 1) = 1.
w∈W

Writing ch M (w · λ) as p∗e(w ·λ) and multiplying both sides of the equation

by q, Lemma 2.3(C) yields
X
q ∗ ch L(λ) = b(λ, w)e(w · λ + ρ).
w∈W

Here b(λ, 1) = 1. We just have to show that b(λ, w) = (−1)`(w) . Recalling

from Lemma 2.3(B) that q is an alternating function under W , whereas
characters of finite dimensional modules are W -invariant, the application
to both sides of a simple reflection sα ∈ W changes the sign of the left
side while replacing w · λ + ρ on the right by sα w · λ + ρ. Now compare
the coefficients to get b(λ, sα w) = −b(λ, w) for all w. Induction on length
implies that b(λ, w) = (−1)`(w) as required.
Since ch L(0) = e(0) (Example 1.6), the last statement of the theorem
follows immediately (and could in fact be proved independently). 

In the course of proving the character formula, we have also proved

Kostant’s formula for the multiplicity of an individual weight µ of L(λ).
2.5. Dimension Formula 41

This was originally derived in [196] using Weyl’s formula but was then
observed to be essentially equivalent to it.
Corollary (Kostant). If λ ∈ Λ+ and µ ≤ λ, then
X
dim L(λ)µ = (−1)`(w) p(µ − w · λ)
w∈W
X
(−1)`(w) p (µ + ρ) − w(λ + ρ)


= 
w∈W

While this formulation is quite elegant, it is usually not suitable for

computations due to the size of the Weyl group. From the computational
point of view, storing all data about W typically consumes an enormous
amount of space. In practice, computer methods based on the older recursive
method of Freudenthal are usually implemented.
Exercise. Fix ν ∈ Γ. Prove that for n  0, dim L(nρ)nρ−ν = p(−ν).
Remarks. (1) By now there are a number of quite different settings for
Weyl’s formula, ranging from compact Lie groups and Haar integrals (which
Weyl himself used) to complex or algebraic geometry related to flag varieties
for semisimple algebraic groups. A special advantage of the BGG approach
presented here is that it generalizes well to algebraic situations having an
infinite dimensional flavor, especially the theory of Kac–Moody Lie algebras
and its offshoots. (Carter [60, 19.3], Jantzen [154], Kac [166, Chap. 10],
Moody–Pianzola [223, 6.4].) See 13.5–13.6 below.
(2) The equality of the multiplicative and additive formulas for the Weyl
function q goes by the name Weyl Denominator Formula due to its role
in the character formula. Here the sum is taken over W and the product over
Φ+ . In the Kac–Moody setting, where W is typically an infinite Coxeter
group, formal identities of this type involving infinite products and sums turn
out to recapture many classical identities involving special functions such
as the Dedekind η-function, e.g., Jacobi’s Triple Product Identity. Large
families of new identities emerge from this kind of representation theory.
(3) To emphasize the dot action of W , Weyl’s Character Formula could
be restated in the form:
(−1)`(w) e(w · λ)
P
ch L(λ) = Pw∈W `(w) e(w · 0)
w∈W (−1)

2.5. Dimension Formula

As remarked earlier, Weyl’s formula yields a practical closed formula for
dim L(λ) when λ ∈ Λ+ . Though its derivation is unrelated to category O
methods, the formula itself is an essential byproduct of Weyl’s theory.
42 2. Characters of Finite Dimensional Modules

The initialPidea is straightforward: thinking of the character of L(λ) as

the finite sum µ dim L(λ)µ e(µ) in ZΛ, add the coefficients to get dim L(λ).
In the setting of X , this translates into adding up the values of ch λ. The
copy of ZΛ in X is the subring Y generated by all e(µ) with µ ∈ Λ. Here
one has a well-defined homomorphism v : Y → Z assigning to f the sum of
its values. So the problem is to compute v(ch L(λ)). Unfortunately, Weyl’s
formula expresses this character as a quotient of two alternating sums over
W , on each of which vPclearly takes the value 0: that is, applying v to both
sides of q ∗ ch L(λ) = w∈W e(w(λ + ρ)) fails to give any information about
the value of v on the character.
For example, when g = sl(2, C) and Λ is identified with Z, we know (0.9)
that when λ ≥ 0,

e(λ + 1) − e(λ − 1)
ch L(λ) = e(λ) + e(λ − 2) + · · · + e(−λ) = .
e(1) − e(−1)

So dim L(λ) = λ + 1, but applying v to the numerator or denominator of

the fraction produces 1 − 1 = 0.
This is roughly analogous to the problem of evaluating at t = 1 the
quotient (tλ+1 − t−λ−1 )/(t − t−1 ) (or an equivalent rational function in t).
In elementary calculus one can often transform such a problem by replacing
both the numerator and the denominator by their derivatives. Here we adopt
a similar formal strategy. But in rank ` the formulas involve ` variables,
corresponding to coordinates of a weight relative to fundamental weights.
So we have to use formal analogues of partial differentiation. The end result
is an elegant formula, which only requires products over the positive roots
(rather than sums taken over W ):

Theorem (Weyl). If λ ∈ Λ+ , then

∨
Q
α>0 hλ + ρ, α i
dim L(λ) = Q ∨
.
α>0 hρ, α i

Proof. We outline the formal steps, leaving routine details to the reader.

• For each α > 0, define an operator ∂α : Y → X0 by extending

linearly the assignment ∂α e(µ) = hµ, α∨ ie(µ) (where µ ∈ Λ). It is
routine to check that ∂α is a derivation of Y: ∂α (f ∗ g) = (∂α f ) ∗
g + f ∗ (∂α g). These
Q operators commute for different roots, so we
can form ∂ := α>0 ∂α and view it as a differential operator (no
longer a derivation in rank > 1).
2.6. Maximal Submodule of M (λ), λ ∈ Λ+ 43

• Recall from 2.3 (and the exercise there), along with Theorem 2.4,
that the Weyl function q can be written as either a sum or a prod-
uct:
Y X
(−1)`(w) e(wρ).


q = e(−ρ) ∗ e(α) − 1 =
α>0 w∈W

• Using first the product form of q, apply ∂ to q ∗ch L(λ), followed by

the homomorphism v. While this looks extremely complicated to
write out in full, it simplifies considerably because v(e(α) − 1) = 0.
All that survives is v(∂q) v(ch L(λ)).
`(w) e(wρ). By linearity,
P
• Now compute v(∂q) using q = w∈W
(−1)
we only need to determine v ∂e(wρ)

Q . Clearly ∂α e(ρ) = hρ, α∨ ie(ρ),
∨
which yields at once v ∂e(ρ) = α>0 hρ, α i. The result is similar
for wρ in place of ρ, but here we get
Y Y Y
hwρ, α∨ i = hρ, w−1 α∨ i = hρ, (w−1 α)∨ i.
α>0 α>0 α>0

• The number of positive roots sent to negative roots by w−1 is

`(w−1 ) = `(w) (see 0.3). Since hρ, −α∨ i = −hρ, α∨ i when α > 0,
we can rewrite the previous step to conclude:
X X Y Y
(−1)`(w) v ∂e(wρ) = hρ, α∨ i = |W | hρ, α∨ i.


v(∂q) =
w∈W w∈W α>0 α>0

• The same method, applied to the right side of the character formula
(where λ + ρ rather than ρ occurs), yields
X Y
(−1)`(w) e w(λ + ρ) = |W | hλ + ρ, α∨ i.



v ∂
w∈W α>0

Now the theorem follows by cancelling the factors |W |. 

+
Exercise. If k ∈ Z>0 , exhibit a simple module L(λ) of dimension k |Φ | .

2.6. Maximal Submodule of M (λ), λ ∈ Λ+

It is an elementary fact that a Verma module or other highest weight module
has a unique maximal submodule (Theorem 1.2(f)); the resulting quotient
is L(λ). But it is not so elementary to describe this maximal submodule
explicitly. In the special case λ ∈ Λ+ , the maximal submodule N (λ) of
M (λ) does turn out to have a simple description.
The proof of the theorem below requires a formal identity in U (g), which
is valid more generally in arbitrary associative algebras over fields of char-
acteristic 0.
44 2. Characters of Finite Dimensional Modules

Lemma. Let A be an associative
algebra over a field of characteristic 0. Let

k ∈ Z>0 and write as usual ki for a binomial coefficient. If a, b ∈ A, then
   
k k k−1 k
[a , b] = [a, b]a + [a, [a, b]]ak−2 + · · · + [a, [a, . . . , [a, b] . . . ].
1 2

Proof. In case k = 1, this is just the identity [a, b] = [a, b]. Then proceed
by induction on k. 
Theorem. Enumerate the simple roots as (α1 , . . . , α` ), abbreviating si :=
sαi . If λ ∈ Λ+ , the maximal submodule N (λ) of M (λ) is the sum of the
submodules M (si · λ) for 1 ≤ i ≤ `.

Proof. In 1.3, we saw that M (λ) ∼ = U (g)/I, where I is the left ideal gen-
erated by n along with all h − λ(h) · 1 (h ∈ h). Here I is precisely the
annihilator in U (g) of a maximal vector v + of M (λ).
In our situation ni := hλ, αi∨ i ∈ Z+ . Consider the left ideal J of U (g)
generated by I together with all yini +1 (1 ≤ i ≤ `). Thanks to Corollary
1.4, J annihilates a maximal vector of L(λ). We claim it is the precise
annihilator.
Indeed, set M := U (g)/J, which has L(λ) as a quotient. If we can show
M is finite dimensional, our claim will follow from complete reducibility: a
direct sum of more than one simple finite dimensional module clearly cannot
be a highest weight module. As observed in Exercise 1.6, it will be enough
to show that all yi act locally nilpotently on M (which the xi obviously
do). In this situation, steps in the proof of Theorem 1.6 will imply the
finite dimensionality based on the fact that M is a sum of finite dimensional
si -submodules for each fixed i.
By construction a maximal vector (coset of 1) in M is killed by yini +1
for each i. On the other hand, M is spanned by the cosets of all possible
monomials (∗) yi1 . . . yit with ij ∈ {1, . . . , `}. Now apply the above lemma,
with A = U (g) and a, b root vectors y, z belonging to two negative roots.
Since root strings have length at most 4 (0.2), (ad y)4 (z) = 0. So if the coset
of a monomial (∗) is killed by yik , the identity in the lemma shows that the
coset of the longer monomial yi0 yi1 . . . yit is killed by yik+3 . Induction on t
(starting with t = 0) implies the local nilpotence of the action of yi on M .
To translate the conclusion that U (g)/J ∼ = L(λ) into the assertion made
in the theorem, recall from 1.4 that the submodule M (si · λ) of M (λ) is
generated by yini +1 · v + . So the left ideal in U (g) annihilating the sum N of
these submodules is precisely J. This forces N = N (λ). 

The theorem may be restated in a suggestive way. Recalling that the

elements of length 1 in W are precisely the simple reflections, we see that
2.7. Related Topics 45

the following sequence is exact whenever λ ∈ Λ+ :

M
M (w · λ) → M (λ) → L(λ) → 0.
w∈W, `(w)=1

The alternating sum format derived for ch L(λ) in terms of characters of

Verma modules resembles an Euler characteristic, so it is tempting to extend
this sequence to the left. The kth term in the sequence would be of the form
M
M (w · λ),
w∈W, `(w)=k

ending with `(w◦ ) = |Φ+ |, where w◦ denotes the longest element of W .

Somewhat miraculously, this kind of resolution of L(λ) does exist. But to
develop it will require further study of maps between Verma modules in
Chapter 4. The BGG resolution itself will then be taken up in Chapter 6.

2.7. Related Topics

There is a rich literature involving finite dimensional representations of g
and their applications in both mathematics and physics. Most of this is
technically unrelated to category O, but the reader should be aware of some
of the possible directions for further study:
• When λ, µ ∈ Λ+ , the decomposition into simple summands L(ν) of
L(λ)⊗L(µ) has been looked at from many angles. The multiplicities
with which these summands occur can be computed in a variety of
ways, especially in type A` or other classical types: Littlewood–
Richardson Rule, etc. They turn out to be closely related to a
number of apparently remote questions in algebraic geometry and
combinatorics.
• The restriction of a simple module to a semisimple (or reductive)
subalgebra of g is a natural problem (leading to branching rules),
which can often be studied together with the tensor product de-
composition problem.
• Fundamental representations L($1 ), . . . , L($` ) can be constructed
explicitly for classical types, by exploiting the exterior powers of
the “natural” representation of g as a classical Lie algebra.
• In the study of multiplicities in L(λ) ⊗ L(µ), a special role is played
by the dominant weights which are W -conjugate to the weights
λ + wµ (with w ∈ W ): which of these occur as highest weights of
summands, and then with what multiplicity? The work of Kumar
and others on the PRV (Parasarathy, Ranga Rao, Varadarajan)
Conjecture is a high point.
46 2. Characters of Finite Dimensional Modules

• In type A` (and other classical types), representations and multi-

plicities have been explored extensively using combinatorial meth-
ods, as just indicated for the tensor product problem. There are
deep connections with representations of symmetric groups, classi-
cal invariant theory, and the like, going back to Schur and Weyl.
• While the algorithms for explicit computation of weight multiplic-
ities dim L(λ)µ are lengthy in practice, there is a reasonable theo-
retical picture of the weight diagram of L(λ) as a whole. Starting
with the extremal weights wλ (w ∈ W ), take all other µ ∈ Λ+
satisfying µ ≤ λ, together with their W -conjugates. These are
precisely the weights occurring with positive multiplicity. Equiva-
lently, the weight diagram can be characterized as consisting of all
points µ ∈ Λ which satisfy µ ≤ λ and lie in the convex hull of the
extremal weights.
• Taking b− = h ⊕ n− , the U (b− )-submodule of L(λ) generated by
one of its extremal weight vectors is called a Demazure module.
Its formal character has been explicitly determined, following De-
mazure’s conjecture; this module turns out to be intimately related
to a Schubert variety for an associated semisimple group.
• Work of Lusztig and Kashiwara involving quantized enveloping al-
gebras has revealed unexpectedly the existence of canonical bases
for the finite dimensional modules L(λ).
Exercise. Prove that every finite dimensional simple U (g)-module occurs as
a direct summand in a suitable tensor product of copies of the fundamental
modules L($1 ), . . . , L($` ).

Notes
The proof of Weyl’s Character Formula given here follows the 1971 paper by
BGG [25], who were inspired in part by Verma’s 1966 thesis [251]. Previous
algebraic treatments (such as Freudenthal’s proof in [144]) were much less
conceptual, relying on algebraic analogues of older analytic methods. A
number of texts have followed the BGG argument, e.g., Bourbaki [46, VIII,
§9], Carter [60, Chap. 12], Humphreys [126, §24].
Weyl’s Dimension Formula (2.5) is derived by a standard algebraic vari-
ant of the analytic proof based on the expression of characters in terms of
actual exponentials.
There are countless research papers and books treating aspects of finite
dimensional representation theory for semisimple Lie algebras. For example,
Fulton–Harris [98] emphasize the down-to-earth combinatorics involved in
working out weight multiplicities and realizations of special representations.
 Chapter 3

Category O: Methods

Here we introduce some standard tools which will be applied throughout

the following chapters:

• The functors Hom and Ext (3.1)

• A duality functor M 7→ M ∨ on O (3.2–3.3)
• Reflection groups W[λ] (3.4), dominant and antidominant weights
(3.5)
• Tensoring Verma modules with finite dimensional modules (3.6)
• “Standard” filtrations having Verma modules as subquotients (3.7)
• Projective objects in O and BGG reciprocity (3.8–3.13)
• “Contravariant” forms on modules (3.14–3.15)

3.1. Hom and Ext

A basic tool in the study of various objects in O such as Verma modules is
the
Hom functor. We already saw (Theorem 1.3) that dim HomO L(µ), L(λ) =
δλµ . At first glance it is far from obvious which computations are easy
and
which are not, for example the computation of HomO M (µ), M (λ) . As
the list of interesting objects grows, this sort of question will always be in
the background.
It is well known that Hom is left exact but often fails to be right exact.
This leads to right derived functors Extn , which will be studied in greater
generality in Chapter 6. For the moment we limit ourselves to the use of
Ext0 = Hom and Ext1 (often just written Ext).

47
48 3. Category O: Methods

For arbitrary U (g)-modules A, C, the vector space ExtU (g) (C, A) classi-
fies up to a natural notion of equivalence the possible short exact sequences
(∗) 0→A→B→C→0
of U (g)-modules. The zero element corresponds to the class of split exact
sequences. In case A, C ∈ O, there is no need here for B to lie in O. For
example, there exist nonsplit extensions when A = C is a Verma module; as
shown below, these cannot be realized in O. So it is essential to write ExtO
when working in this category.
Exercise. Let g = sl(2, C), and identify λ ∈ h∗ with a scalar as usual. Let
N be
 a 2-dimensional
 U (b)-module defined by letting x act as 0 and h act
λ 1
as . Show that the induced U (g)-module M := U (g) ⊗U (b) N fits
0 λ
into an exact sequence which fails to split:
0 → M (λ) → M → M (λ) → 0.
(Here M cannot lie in O, thanks to part (a) of the proposition below.)

The most basic general tool available for the study of the derived functors
n
Ext of Hom in a module category is the long exact sequence associated to a
short exact sequence. This comes in two flavors, since Hom is contravariant
when the second variable is fixed but covariant when the first variable is
fixed. For example, taking D as the second variable, we get from (∗) above
0 → Hom(C, D) → Hom(B, D) → Hom(A, D) →

→ Ext1 (C, D) → Ext1 (B, D) → Ext1 (A, D) → . . .

The sequence continues with Ext2 , etc.
In this chapter a number of specific computations will appear, but mainly
of the easier sort. For example:
Proposition. Let λ, µ ∈ h∗ .
(a) If M is a highest weight module of weight µ, with λ 6< µ, then
ExtO (M (λ), M ) = 0. In particular,



ExtO M (λ), L(λ) = 0 = ExtO M (λ), M (λ) .


(b) If µ ≤ λ, then ExtO M (λ), L(µ) = 0.
(c) If µ < λ and N (λ) is the maximal submodule of M (λ), then
HomO N (λ), L(µ) ∼



= ExtO L(λ), L(µ) .


(d) ExtO L(λ), L(λ) = 0.
3.2. Duality in O 49

Proof. (a) Given an extension 0 → M → E → M (λ) → 0 in O, the

hypothesis guarantees that any preimage in Eλ of a maximal vector of weight
λ in M (λ) is also a maximal vector. So it generates a submodule mapping
isomorphically onto M (λ), which splits the sequence.
(b) follows immediately from (a).
(c) Starting with the exact sequence 0 → N (λ) → M (λ) → L(λ) → 0,
one gets as part of a long exact sequence:




. . . → HomO M (λ), L(µ) → HomO N (λ), L(µ) → ExtO L(λ), L(µ) →


→ ExtO M (λ), L(µ) → . . .
Since µ < λ, the first term vanishes because L(λ) is the unique simple
quotient of M (λ) and the last term vanishes by part (b).
(d) Replace µ by λ in the proof of (c). Using (a) and the fact that L(λ)
is not a composition factor of N (λ), the second and fourth terms of the
resulting sequence vanish. 

3.2. Duality in O
There is a useful duality functor on O, obtained by modifying the usual
notion of dual module for U (g) (which arises naturally from the Hopf algebra
structure of U (g)). Recall that for any U (g)-module M , the dual vector
space M ∗ becomes a U (g)-module if we define
(x · f )(v) := −f (x · v) for all v ∈ M, f ∈ M ∗ , x ∈ g.
But if M lies in O, it will usually not be the case that M ∗ also does. For one
thing, M ∗ will be “too large” whenever dim M = ∞; so it will be necessary to
cut down to a suitable subspace in order to get a finitely generated module.
Moreover, the standard action of U (g) creates intrinsic problems and must
therefore be modified.
Twisting the action of g improves the situation. Recall from 0.5 the
transpose map τ on g, which fixes h pointwise but sends gα to g−α for all
roots α. A new action of g on M ∗ is defined by (x · f )(v) := f (τ (x) · v).
We always work with this action on dual vector spaces in what follows. In
particular, (h · f )(v) = f (h · v) for all h ∈ h, f ∈ M ∗ , v ∈ M (where M can
be any U (g)-module).
To get started, consider the category C of U (g)-modules which are weight
modules with finite dimensional weight spaces. (This includes O.) If M ∈ C
and λ ∈ h∗ , the dual space (Mλ )∗ of Mλ can be identified naturally with
the set of those f ∈ M ∗ which vanish on all summands Mµ of M for which
µ 6= λ. On the other hand, any f ∈ M ∗ which has weight λ (under the
50 3. Category O: Methods

twisted action of h) vanishes at all v ∈ Mµ : whenever λ(h) 6= µ(h) we get

µ(h)f (v) = f (µ(h)v) = (h · f )(v) = λ(h)f (v).
Thus f ∈ (Mλ )∗ , identifying (M ∗ )λ with (Mλ )∗ . This allows us to drop the
parentheses. Now we can define
M
M ∨ := Mλ∗ for any M ∈ C
λ∈h∗
and call this the dual in C.
Since the notion of formal character makes sense for C, it follows from
this discussion that ch M ∨ = ch M . Moreover, M 7→ M ∨ defines an exact
contravariant functor on C: an exact sequence 0 → L → M → N → 0
in C yields an exact sequence 0 → N ∨ → M ∨ → L∨ → 0. This implies
that M is simple if M ∨ is simple. It is also obvious that (M ⊕ N )∨ ∼ =
M ∨ ⊕ N ∨ . Finally, the functor sending M 7→ M ∨∨ is naturally isomorphic
to the identity functor on C.
It still has to be checked that O is closed under duality: When M ∈ O, is
M ∨ locally finite for n as well as finitely generated? If M ∈ O and f ∈ Mµ∨
with µ ∈ h∗ , the twisted action of g on M ∨ insures that xα · f ∈ Mµ+α∨ . This

implies that the action of n on M is locally finite. Observe next that each
L(λ)∨ is simple because its dual L(λ)∨∨ ∼ = L(λ) is simple; its highest weight
is λ because ch L(λ) = ch L(λ). Therefore L(λ)∨ is isomorphic to L(λ).
∨

Since M has finite length, we conclude from the exactness of duality that
M ∨ has finite length (with the same composition factors) and is therefore
finitely generated. So M ∨ ∈ O.
How does duality interact with the subcategories Oχ ? Recall from
Exercise 1.10 that τ fixes Z(g) pointwise: τ commutes with the Harish-
Chandra homomorphism ξ (which is injective), while U (h) is commutative.
If M ∈ Oχ , it follows from the definition of M ∨ that Z(g) also acts on M ∨
by the generalized character χ; in particular, z ∈ Z(g) acts on any maximal
vector as scalar multiplication by χ(z).
To summarize:
Theorem. The duality functor on O sending M to M ∨ is exact and con-
travariant, with the properties:
(a) Duality induces a self-equivalence on category O: its square is nat-
urally isomorphic to the identity functor.
(b) For any M ∈ O and any central character χ, (M ∨ )χ ∼ = (M χ )∨ ; in
particular, if M ∈ Oχ , then M ∨ ∈ Oχ .
(c) If M ∈ O, then ch M = ch M ∨ ; so M and M ∨ have the same com-
position factor multiplicities and therefore define the same element
of K(O). In particular, L(λ)∨ ∼ = L(λ).
3.3. Duals of Highest Weight Modules 51

(d) If M, N ∈ O, then (M ⊕ N )∨ ∼
= M ∨ ⊕ N ∨ ; so if M is indecompos-
∨
able, M is also indecomposable.
(e) ExtO (M, N ) ∼ ∨ ∨

O (N , M ) for all
M, N ∈ O. In particular,
= Ext
ExtO L(λ), L(µ) ∼ = ExtO L(µ), L(λ) . 
Exercise. Show that (L ⊗ M )∨ ∼
= L∨ ⊗ M ∨ if M ∈ O and dim L < ∞.

3.3. Duals of Highest Weight Modules

It is easy to apply the duality functor in O to simple modules and Verma
modules:
Theorem. Let λ, µ ∈ h∗ .
(a) L(λ)∨ ∼
= L(λ).
(b) M (λ)∨ has L(λ) as its unique simple submodule. Its other compo-
sition factors L(µ) satisfy µ < λ.
(c) Any nonzero homomorphism M (µ) → M (λ)∨ has the simple
sub-
module L(λ) as its image.
Thus dim HomO M (λ), M (λ)∨ = 1,
while HomO M (µ), M (λ)∨ = 0 when µ 6= λ.
(d) ExtO M (µ), M (λ)∨ = 0 for all λ, µ.

Proof. (a) This was already observed in 3.2.

(b) Applying duality to the short exact sequence
0 → N (λ) → M (λ) → L(λ) → 0
produces an exact sequence
0 → L(λ)∨ → M (λ)∨ → N (λ)∨ → 0.
Thanks to (a), L(λ)∨ ∼
= L(λ) occurs as a simple submodule. No other simple
submodule can occur in M (λ)∨ , since its dual would have to be a simple
quotient of M (λ). Since M (λ)∨ has the same formal character as M (λ),
hence the same composition factor multiplicities, (b) follows.
(c) The image M in M (λ)∨ of a nonzero homomorphism M (µ) → M (λ)∨
is a highest weight module of weight µ; by (b), it has a submodule isomorphic
to L(λ)∨ ∼ = L(λ). Therefore µ ≥ λ. On the other hand, (b) implies that
µ ≤ λ. Thus µ = λ. In this case, the universal property of M (λ) provides
a nonzero homomorphism M (λ) → L(λ) ,→
M (λ)∨ . Since dim M (λ)∨ λ = 1,
∨
Exercise 1.3 forces dim HomO M (λ), M (λ) = 1.
(d) We have to show that any short exact sequence
0 → M (λ)∨ → M → M (µ) → 0
splits. If µ happens to be a maximal weight of M , the pre-image of a maximal
vector of weight µ in M (µ) will be a maximal vector in M ; then the universal
52 3. Category O: Methods

property of Verma modules will insure a splitting. Otherwise we must have

µ < λ. But then we can dualize and use the fact that M (λ)∨∨ ∼ = M (λ) to
get another such short exact sequence
0 → M (µ)∨ → M ∨ → M (λ) → 0
Since λ > µ, this sequence splits; as a result, the original sequence splits. 
Exercise. Does part (d) of the theorem remain true if M (µ) and M (λ)∨
are interchanged?

3.4. The Reflection Group W[λ]

What do we know at this point about the possible composition factors L(µ)
of a Verma module M (λ)? The most obvious condition is (1) µ ≤ λ; in
particular, µ must lie in the same coset of h∗ /Λr as λ. From 1.7 we know
that in addition, (2) µ = w · λ for some w ∈ W . These two conditions
narrow the choices for µ considerably, but turn out to be far from optimal.
This will become clear in 5.1.
To focus the problem better when λ ∈ h∗ is not assumed to be integral,
Jantzen [148, 1.3] introduced for each λ an associated subgroup of W which
can replace W in (2). Since sα λ = λ − hλ, α∨ iα, the scalar hλ, α∨ i plays a
crucial role here. Note that it lies in Z if and only if hλ + ρ, α∨ i does.
Start by defining
Φ[λ] := {α ∈ Φ | hλ, α∨ i ∈ Z}
and
W[λ] := {w ∈ W | wλ − λ ∈ Λr }.
The notation [λ] is motivated by the fact that
Φ[λ] = Φ[µ] and W[λ] = W[µ] whenever λ ≡ µ (mod Λ).
Since ρ ∈ Λ, one could equally well write
Φ[λ] = {α ∈ Φ | hλ + ρ, α∨ i ∈ Z}, W[λ] = {w ∈ W | w · λ − λ ∈ Λr }.
It follows that Φ[λ] = Φ[µ] and W[λ] = W[µ] whenever µ ∈ W[λ] · λ.
It is clear that Λ = {λ ∈ h∗ | Φ[λ] = Φ}. Moreover,
Φ[λ] = {α ∈ Φ | sα ∈ W[λ] },
thanks to the fact that the only scalar multiples of α in Λr are Zα (a fact
which is obvious for simple roots and hence for their W -conjugates).
In terms of this subgroup of W we can refine the above necessary con-
ditions for L(µ) to occur as a composition factor of M (λ):
(3) µ ≤ λ and µ = w · λ for some w ∈ W[λ] .
3.4. The Reflection Group W[λ] 53

In fact W[λ] also plays the role of a Weyl group, acting on a subspace of
E = R ⊗Z Λ:
Theorem. Let λ ∈ h∗ . Then:
(a) Φ[λ] is a root system in its R-span E(λ) ⊂ E.
(b) W[λ] is the Weyl group of the root system Φ[λ] . In particular, it is
generated by the reflections sα with α ∈ Φ[λ] .

For the proof we need some standard facts about root systems and re-
flection groups:
(A) Associated to W is the affine Weyl group Wa associated with W ,
which acts in the affine space A associated with E as a group generated by
affine reflections. Abstractly Wa is the semidirect product W n Λr . (See
Bourbaki [45, VI,§2], Humphreys [130, Chap. 3].) Here W acts naturally
on Λr , regarded as an additive group of translations in E. It is a Coxeter
group, generated by the affine reflections
sα,n : µ 7→ sα µ + nα with α ∈ Φ, n ∈ Z, µ ∈ E.
Now if λ ∈ E, its isotropy group in W is itself the Weyl group of the root
system spanned by all α ∈ Φ orthogonal to λ. In particular, it is a reflection
group. Similarly, the isotropy group in Wa of a point in A is generated by
the reflections sα,n it contains. (See [45, V, §3, Prop. 1] or [130, 1.12, 4.8].)
(B) If Φ1 ⊂ Φ is a root system in the subspace E1 of E which it spans,
while Φ ∩ E1 = Φ1 , then there is a basis ∆ of Φ for which ∆ ∩ Φ1 is a basis
of Φ1 . (See [45, VI,§1,Prop. 24].)
Remark. The basic idea in the proof of the theorem is suggested by Exercise
1 in Bourbaki [45, VI, §2], where the setting is the real euclidean space E.
But Jantzen’s proof in [148, 1.3] requires additional technical steps to pass
from R to an arbitrary splitting field for g. We present a simplified version
of his proof when the ground field is C; here we may write h∗ = E + iE with
i2 = −1 in C. (For another version see Bernstein–Gelfand [24, App. 1].)

Proof of Theorem. For (a), the only non-obvious point to check is that
sα Φ[λ] = Φ[λ] for all α ∈ Φ[λ] . Let β ∈ Φ[λ] and note that (sα β)∨ = sα β ∨
(since β and sα β have the same length). Then
hλ, (sα β)∨ i = hλ, sα β ∨ i = hsα λ, β ∨ i = hλ, β ∨ i − hλ, α∨ ihα, β ∨ i ∈ Z,
since both α and β are assumed to lie in Φ[λ] .
The proof of (b) is more subtle. The basic trick is to interpret the
defining condition for W[λ] as requiring that λ be fixed by a subgroup of
the affine Weyl group Wa . Then (A) above applies. This can only be done
directly in case λ ∈ E(λ), the R-span of Φ[λ] in E. Now the condition that
54 3. Category O: Methods

sα ∈ Φ[λ] , or sα λ − λ ∈ Λr , translates into the condition that sα,n λ = λ

for some integer n. From (A) we conclude that the isotropy group of λ in
Wa is generated by precisely these reflections. The natural projection of Wa
onto W maps this isotropy group onto W[λ] , which is therefore the reflection
subgroup of W generated by the sα with α ∈ Φ[λ] .
It remains to consider the general case, with λ ∈ h∗ arbitrary. As re-
marked above, we assume that h∗ = E ⊕ iE. Write λ uniquely in the form
λ = λ0 + iλ1 , where λk ∈ E. Then wλ − λ = (wλ0 − λ0 ) + i(wλ1 − λ1 ), with
each expression in parentheses still lying in E. It follows that w ∈ W[λ] if
and only if wλ0 − λ0 ∈ Λr (forcing λ0 ∈ E(λ)) and wλ1 = λ1 . Denote by
W1 the isotropy group of λ1 in W . Thanks to (A) above, this is itself the
Weyl group of the root system Φ1 (in the subspace of E it spans) consisting
of all α ∈ Φ orthogonal to λ1 . Thus W[λ] = W[λ0 ] ∩ W1 .
Any root in the R-span E1 of Φ1 is orthogonal to λ1 and hence lies in
Φ1 . From (B) above it follows at once that E1 ∩ Λr is precisely the root
lattice of Φ1 . Thus W[λ] consists of those w ∈ W1 for which wλ − λ lies in
the root lattice associated with W1 . Now we can apply to W1 and λ0 the
earlier argument in the real case. 

When dealing with nonintegral weights, these reflection subgroups W[λ]

enable us to refine the study of a single linkage class, by partitioning it
among various cosets of Λr in h∗ . Each such subset is itself the linkage class
of a weight λ under the reflection group W[λ] . Eventually we will show in
4.9 that the subcategories corresponding to these orbits W[λ] · λ are precisely
the blocks of O in the sense of 1.13.
Exercise. Show that Φ[λ] ∩ Φ+ is a positive system in the root system Φ[λ] ;
but the corresponding simple system (call it ∆[λ] ) may be unrelated to ∆.
For a concrete example, take Φ to be of type B2 with short simple root α
and long simple root β. If λ := α/2, check that Φ[λ] contains just the four
short roots in Φ.

3.5. Dominant and Antidominant Weights

When λ ∈ Λ, we have W[λ] = W . In that case, the linkage class of λ can be
parametrized by any weight in the class: for example, a weight in Λ+ − ρ.
In the case of a nonintegral weight λ, it is not so obvious at first sight how
to select an optimal representative of the orbit W[λ] · λ. It turns out that the
lowest weight in the orbit is most convenient in the further study of Verma
modules, while for other purposes the highest weight may be preferred.
Call λ ∈ h∗ antidominant if hλ + ρ, α∨ i 6∈ Z>0 for all α ∈ Φ+ . For ex-
ample, −ρ is antidominant. Similarly, call λ dominant if hλ + ρ, α∨ i 6∈ Z<0
for all α ∈ Φ+ . Warning: This differs from the usual notion of dominance in
3.5. Dominant and Antidominant Weights 55

Λ: now the set of “dominant” weights in Λ becomes Λ+ − ρ, with −ρ being

both dominant and antidominant. To avoid confusion we usually emphasize
just the notion of antidominance for h∗ , the other case being symmetric.
Notice that any linkage class W · λ contains at least one antidominant
weight: If a weight µ in the class is minimal relative to the standard partial
ordering, but some hµ + ρ, α∨ i ∈ Z>0 , then the linked weight sα · µ < µ
contradicts the minimality. Similarly, for any λ ∈ h∗ , the orbit W[λ] · µ
contains at least one antidominant weight.
Proposition. Let λ ∈ h∗ , with corresponding root system Φ[λ] and Weyl
group W[λ] . Let ∆[λ] be the simple system corresponding to the positive
system Φ[λ] ∩ Φ+ in Φ[λ] . Then λ is antidominant if and only if one of the
following three equivalent conditions holds:
(a) hλ + ρ, α∨ i ≤ 0 for all α ∈ ∆[λ] .
(b) λ ≤ sα · λ for all α ∈ ∆[λ] .
(c) λ ≤ w · λ for all w ∈ W[λ] .
Therefore there is a unique antidominant weight in the orbit W[λ] · λ.

Proof. Observe first that λ is antidominant if and only if (a) holds. The
“only if” part is obvious. In the other direction, assume (a). If hλ + ρ, α∨ i
lies in Z for some α ∈ Φ+ , then α ∈ Φ[λ] ∩ Φ+ and is therefore a Z+ -linear
combination of ∆[λ] . It follows from (a) that hλ + ρ, α∨ i ≤ 0. So λ must be
antidominant.
For a simple root α (in any root system, such as Φ[λ] ) we have
sα · λ = sα (λ + ρ) − ρ = λ + ρ − hλ + ρ, α∨ iα − ρ = λ − hλ + ρ, α∨ iα.
From this the equivalence of (a) and (b) is clear.
Since (c) obviously implies (b), it remains to show that (b) implies (c).
Here we use induction on length in W[λ] , relative to the simple system ∆[λ] .
When `(w) = 0, we have w = 1 and there is nothing to prove. When
`(w) > 0, write w = w0 sα in W[λ] , with α ∈ ∆[λ] and `(w0 ) = `(w) − 1. From
0.3(4) we get wα < 0. Now
λ − w · λ = (λ − w0 · λ) + w0 · (λ − sα · λ).
By induction, λ − w0 · λ ≤ 0. On the other hand,
w0 · (λ − sα · λ) = w(sα · λ − λ) = −hλ + ρ, α∨ iwα ≤ 0,
using the equivalence of (a) and (b) along with the fact that wα < 0. 
Exercise. If λ ∈ h∗ is antidominant, prove that M (λ) = L(λ). [This and
its converse will be discussed further in Chapter 4.]
56 3. Category O: Methods

Remark. Recall from Exercise 1.13 the refined decomposition of Oχ when

χ is a central character belonging to a linkage class of nonintegral weights.
This decomposition is indexed by the nonempty intersections of the linkage
class with cosets in h∗ /Λr . From now on we can use the antidominant weight
λ in such an intersection to parametrize the corresponding subcategory of
Oχ . Call this subcategory Oλ . (This conflicts with the special notation O0
often used for the principal block, introduced in 1.13. But since λ = 0 is
never antidominant, there should be no confusion in practice.) In 4.9 we
will be able to show that the Oλ are precisely the blocks of O.
Notice that the various subsets W[λ] · λ of the given linkage class (with
λ antidominant and lying in one coset modulo Λr ) have the same cardinal-
ity. Indeed, if µ is another such antidominant weight and w · µ = λ, then
W[µ] = w−1 W[λ] w. Moreover, in each case the structure of partially ordered
set is the same, with a unique maximal (dominant) and a unique minimal
(antidominant) element.

3.6. Tensoring Verma Modules with Finite Dimensional

Modules
As we saw in Theorem 1.1(d), category O is closed under tensoring with
finite dimensional modules. As a tool in the study of projective modules,
and in preparation for the introduction of “translation functors” in Chapter
7, we examine here an interesting special situation.
Theorem. Let M be a finite dimensional U (g)-module. For any λ ∈ h∗ ,
the tensor product T := M (λ) ⊗ M has a finite filtration with quotients
isomorphic to Verma modules of the form M (λ + µ). Here µ ranges over
the weights of M , each occurring dim Mµ times in the filtration.

Proof. A direct computational proof is outlined in the exercise below, re-

quiring some bookkeeping with formal characters; but here we place the
theorem in a more general context which is useful elsewhere.
Let M be a U (g)-module and L a U (b)-module. Then L ⊗ M is also a
U (b)-module (with the extra U (g)-module structure on M forgotten). The
Tensor Identity provides a natural isomorphism
U (g) ⊗U (b) L ⊗ M ∼



= U (g) ⊗U (b) L ⊗ M ,
where the left side is a tensor product of U (g)-modules and the right side
is an induced module viewed as left U (g)-module. This is valid in a general
setting for Lie algebras and subalgebras (see Knapp [189, Prop. 6.5]) and is
implicated in ideas such as Frobenius Reciprocity in representation theory.
As pointed out in Remark 1.3, N 7→ U (g) ⊗U (b) N is an exact functor
from U (b)-modules to U (g)-modules whenever dim N < ∞.
3.6. Tensoring Verma Modules with Finite Dimensional Modules 57

Now assume that L and M are finite dimensional, and set N := L ⊗ M .

Order a basis of N consisting of weight vectors v1 , . . . , vn (with weights
ν1 , . . . νn ) so that i ≤ j whenever νi ≤ νj . This yields a filtration 0 ⊂ Nn ⊂
· · · ⊂ N1 = N , with Nk the U (b)-submodule spanned by vk , . . . , vn . An easy
induction on dim N then shows that the induced module on the right side
of the Tensor Identity has a corresponding filtration with quotients which
are Verma modules of highest weights equal to the weights of N .
In our situation the U (b)-module L is Cλ , so induction yields M (λ). On
the other hand, L ⊗ M has the same dimension as M , but λ is added to
all weights µ of M . Thus we can identify the tensor product T with the
induced module U (g) ⊗U (b) (L ⊗ M ). By the argument above, this module
has the asserted type of filtration by Verma modules. 
Example. A special case of the induced module construction above will be
needed in Chapter 6. There one takes M to be the kth exterior power of
the vector space g/b, viewed as a U (b)-module. (What are its weights?)
Exercise. Without invoking the Tensor Identity, give a direct proof of the
theorem by filling in these steps:
(a) Check that the formal character of T agrees with the
P sum of formal
characters of the indicated Verma modules: Write ch M = µ mµ e(µ), with
mµ := dim Mµ . Then Proposition 1.16 implies that
X
ch T = ch M (λ) ∗ ch M = p ∗ e(λ) ∗ mµ e(µ)
µ
X
X
= mµ p ∗ e(λ + µ) = mµ ch M (λ + µ).
µ µ

(b) To construct the desired filtration, denote by v + a maximal vector in

M (λ) and by v1 , . . . , vn a basis of M consisting of vectors having respective
weights µ1 , . . . , µn (with repetitions allowed). Order this basis in such a way
that i ≤ j whenever µi ≤ µj . Then set ti := v + ⊗ vi in T and let Ti be the
submodule generated by ti , . . . , tn . So 0 ⊂ Tn ⊂ Tn−1 ⊂ · · · ⊂ T1 ⊂ T .
(c) Prove that T = T1 , using induction with respect to the PBW filtra-
tion of U (g). This starts with the calculation
(x · v + ) ⊗ v = x · (v + ⊗ v) − v + ⊗ x · v for x ∈ g, v ∈ M.
Evidently the vector on the right side lies in T1 .
(d) The vector ti has weight λ + µi , as does its image ti in Ti /Ti+1 ,
which clearly generates the quotient module. Using the chosen ordering of
weights, show that ti is either 0 or else a maximal vector of weight λ + µi .
In either case the universal property of Verma modules gives a surjection
ϕi : M (λ + µi ) → Ti /Ti+1 .
58 3. Category O: Methods

(e) Each ϕi is an isomorphism, by the formal character comparison

above.
Remark. Either proof of the theorem shows more precisely that T has a
submodule isomorphic to M (λ + µ), where µ is any maximal weight of M .
Similarly, T has a quotient isomorphic to M (λ + ν), where ν is any minimal
weight of M .

While it is easy to keep track of formal characters, the detailed submod-

ule structure of the tensor product in the theorem is not so easy to study.
But the decomposition of O into the direct sum of subcategories Oχ does
provide some further information due to the fact that each Verma module
corresponds to a single central character. It is instructive to look at the
rank 1 case, though even here the full picture requires development of some
further theory. Let g = sl(2, C) and identify h∗ with C. The reader should
be able to describe for each λ ∈ N the decomposition of M (0)⊗L(λ) relative
to central characters χ.

3.7. Standard Filtrations

The type of filtration described in Theorem 3.6 plays a prominent role in
the study of projective modules in O as well as the later study of “tilting”
modules (Chapter 11). We say that M ∈ O has a standard filtration
(also sometimes called a Verma flag) if there is a sequence of submodules
0 = M0 ⊂ M1 ⊂ M2 ⊂ · · · ⊂ Mn = M for which each M i := Mi /Mi−1 (1 ≤
i ≤ n) is isomorphic to a Verma module. While the Mi are not in general
unique, comparison of formal characters shows that the filtration length n
is well defined, as is the multiplicity with which each Verma module occurs
as a subquotient. By convention, the 0 module may also be said to have a
standard filtration of length 0.
Denote the multiplicity with which each
Verma module M (λ) occurs in
a standard filtration of M by M : M (λ) , reserving the earlier notation
[M : L(λ)] for the multiplicity of L(λ) in a Jordan–Hölder series of M . (The
notation has to be handled with care, for example when M (λ) = L(λ).) It is
clear that a direct sum M ⊕ N of modules with standard filtrations also
has
such a filtration;

by comparing

formal characters, we get (M ⊕ N : M λ) =
M : M (λ) + N : M (λ) .
Here are some less obvious, but still elementary, observations about stan-
dard filtrations:
Proposition. Let M ∈ O have a standard filtration.
(a) If λ is maximal among the weights of M , then M has a submodule
isomorphic to M (λ) and M/M (λ) has a standard filtration.
3.7. Standard Filtrations 59

(b) If M = M 0 ⊕ M 00 in O, then M 0 and M 00 have standard filtrations.

(c) M is free as a U (n− )-module.

Proof. (a) The assumption on λ implies that M has a maximal vector of

weight λ, which (by the universal property of Verma modules) insures the
existence of a nonzero homomorphism ϕ : M (λ) → M . We claim it is
injective. In the

given standard filtration, let i be the smallest index for
i
which ϕ M (λ) ⊂ Mi . Thus the induced map ψ : M (λ) → M = Mi /Mi−1
is nonzero. But Mi /Mi−1 ∼ = M (µ) for some µ, which forces λ ≤ µ. By the
maximality of λ, we have λ = µ. This forces ψ to be an isomorphism. In
turn, ϕ itself is injective. Write briefly M (λ) ⊂ M . Now M (λ) ∩ Mi−1 =
Ker ψ = 0, resulting in an induced short exact sequence

0 → Mi−1 → M/M (λ) → M/Mi → 0.

Both Mi−1 and M/Mi have standard filtrations, which combine to produce
a standard filtration for M/M (λ).
(b) There is nothing to prove if M is a Verma module, so we can proceed
by induction on the filtration length of M . Let λ be maximal among the
weights of M ; without loss of generality, say Mλ0 6= 0. As before, there
is a maximal vector in M 0 of weight λ and an associated homomorphism
M (λ) → M 0 ,→ M . Applying part (a) to the composite
map, we see that
M (λ) → M 0 is injective, while M/M (λ) ∼ = M 0 /M (λ) ⊕ M 00 admits a
standard filtration. The induction hypothesis, applied to M/M (λ), shows
that M 0 /M (λ) and M 00 admit standard filtrations (so M 0 also does).
(c) Suppose M has a standard filtration. To show that M is U (n− )-free,
use induction on the standard filtration length starting with the obvious case
when M ∼ = M (λ). For the induction step, use part (a) to find a submodule
N isomorphic to some M (λ) for which M/N has a standard filtration. This
module is U (n− )-free by induction, so M is. 

Exercise. (a) If a module M has a standard filtration and there exists an

epimorphism ϕ : M → M (λ), prove that Ker ϕ admits a standard filtration.
(b) Show by example when g = sl(2, C) that the existence of a monomor-
phism ϕ : M (λ) → M when M has a standard filtration fails to imply that
Coker ϕ also has a standard filtration.

For a module with a standard filtration, we can characterize the resulting

filtration multiplicities in a useful way (needed in 3.11 below):

Theorem. If M has a standard filtration, then for all λ ∈ h∗ we have

M : M (λ) = dim HomO M, M (λ)∨ .

60 3. Category O: Methods

Proof. The proof is by induction on the filtration length of

M . In the case
of length 1, combine the obvious fact that M
(µ) : M (λ) = δλµ with the
earlier computation dim HomO M (µ), M (λ)∨ = δλµ in Theorem 3.3(c).
For the induction step, look at the beginning of the long exact sequence
associated with a short exact sequence 0 → N → M → M (µ) → 0:
0 → HomO M (µ), M (λ)∨ → HomO M, M (λ)∨ → HomO N, M (λ)∨ →

→ ExtO M (µ), M (λ)∨ → . . .

∨


The Ext term vanishes by Theorem 3.3(d), while
dim HomO M (µ), M (λ) =
δλµ as just recalled. By induction, N
: M (λ) is the
dimension of the asso-
ciated Hom space. Since M : M (λ)
= N : M (λ) + δλµ by the choice of
µ, the desired value of M : M (λ) follows. 

3.8. Projectives in O
In order to make effective use of homological methods, we want to show
that O has enough projectives: for each M ∈ O, there is a projective object
P ∈ O and an epimorphism P → M . Recall that an object P in an abelian
category is called projective if the left exact functor Hom(P, ?) is also right
exact. Equivalently, given an epimorphism π : M → N and any morphism
ϕ : P → N , there is a lifting ψ : P → M such that πψ = ϕ. Dually, an
object Q is injective if the right exact functor Hom(?, Q) is also left exact.
To say that O has enough injectives is to say that for each M ∈ O, there is
an injective object Q and a monomorphism M → Q.
In category O, the existence of enough projectives will automatically
imply the existence of enough injectives: given a projective P , it follows
immediately from the definitions that the duality functor (which is exact and
contravariant) sends P to an injective P ∨ . So we focus here on projectives.
Some examples are easy to find. Recall from 3.5 that λ ∈ h∗ is called
“dominant” if −λ is antidominant.
Proposition. Category O contains some projective objects:
(a) Suppose λ ∈ h∗ is dominant, hence maximal in its W[λ] -orbit, for
example λ ∈ Λ+ − ρ. Then M (λ) is projective in O.
(b) If P ∈ O is projective, while dim L < ∞, then P ⊗ L is projective
in O.

Proof. (a) Given an epimorphism π : M → N in O and a homomorphism

ϕ : M (λ) → N (which may be assumed to be nonzero), we need to find a
lifting ψ : M (λ) → M . Since M (λ) ∈ Oχ (where χ = χλ ), the image of ϕ
lies in N χ . So we may as well assume that M, N ∈ Oχ .
3.8. Projectives in O 61

Now if v + is a maximal vector in M (λ), its image v := ϕ(v + ) is a

maximal vector in N of weight λ. Since π is surjective, there exists a vector v
of weight λ in M such that ϕ(v + ) = π(v). The n-submodule of M generated
by v must contain a maximal vector having a weight linked to λ. But
by assumption there is no such weight > λ. This forces n · v = 0. In
turn, the universal property of M (λ) implies the existence of the desired
homomorphism ψ sending v + to v.
(b) It has to be shown that HomO (P ⊗ L, ?) is right exact. In general,
we have
HomO (P ⊗ L, M ) ∼


= HomO P, Hom(L, M ) .
This starts with the standard vector space isomorphism, which is then easily
checked to be compatible with the U (g)-actions. Here the vector space
Hom(L, M ) has the usual structure of U (g)-module. Since dim L < ∞, we
can further replace Hom(L, M ) by L∗ ⊗ M , where again the U (g)-actions
are compatible. Since P is projective, while tensoring with L∗ is exact in
O, this shows that P ⊗ L is also projective. 

Part (a) of the proposition shows that at least some of the Verma mod-
ules have categorical meaning. In a similar but more artificial vein, one can
use the same argument to show that an arbitrary M (λ) is a projective object
in the full subcategory of O whose objects are the modules with all weights
bounded above by λ.

Example. In case λ = −ρ, we know already from Example 1.11 that

M (−ρ) = L(−ρ). Now part (a) of the proposition shows that M (−ρ) is
projective.

Exercise. If Q is injective in O, while dim L < ∞, prove that Q ⊗ L is

injective in O.

By combining the two parts of the proposition, we can construct a large

collection of projectives. Although these are typically hard to analyze in
detail, they do insure the existence of enough projectives:

Theorem. Category O has enough projectives (and enough injectives).

Proof. The first goal is to find for each λ ∈ h∗ a projective object in O

mapping onto L(λ). It is clear that µ := λ + nρ is dominant for sufficiently
large n. Therefore M (µ) is projective, by the proposition. In turn, nρ ∈ Λ+ ,
so dim L(nρ) < ∞ and P := M (µ) ⊗ L(nρ) is projective. Obviously the
lowest weight of L(nρ) is −nρ = w◦ (nρ). So Remark 3.6 shows that the
tensor product has a quotient isomorphic to M (µ−nρ) = M (λ). This shows
that M (λ) and hence also L(λ) is the quotient of a projective module.
62 3. Category O: Methods

For use in 3.10 below, note that Theorem 3.6 insures the existence of
a standard filtration in P . The subquotients are the M (µ + ν), where ν
runs over the weights of L(nρ) (counting multiplicity). In particular, M (λ)
occurs just once, with all other µ + ν > λ.
Now an induction based on length will complete the proof that an arbi-
trary nonzero M ∈ O is the quotient of a projective module. Assuming M
has length > 1, it has a simple submodule L(λ), giving rise to a short exact
sequence 0 → L(λ) → M → N → 0. Here the quotient N has smaller length
than M , so by induction there is a surjection Q → N for some projective
Q in O. Lifting this map to Q → M by the projective property, either
Q → M is surjective (and we are done) or else the map gives a splitting
M ∼ = L(λ) ⊕ N . In the latter case, a direct sum of projectives mapping
respectively onto L(λ) and N will map onto M . 

3.9. Indecomposable Projectives

As in other artinian module categories, it follows from the existence of
enough projectives that each M ∈ O has a projective cover π : PM → M .
Here π is an epimorphism and is essential, meaning that no proper submod-
ule of the projective module PM is mapped onto M . Up to isomorphism,
the module PM is then the unique projective up to isomorphism having this
property. (See for example Curtis–Reiner [74, §6C].)
In particular, for each λ ∈ h∗ denote by πλ : P (λ) → L(λ) a fixed
projective cover. If N were a maximal submodule of P (λ) different from
the maximal submodule Ker πλ , then N would be mapped onto L(λ) by πλ
contrary to the map being essential. Thus P (λ) has a unique maximal sub-
module and is therefore indecomposable. Moreover, thanks to the projective
property, one actually has epimorphisms P (λ) → M (λ) → L(λ). (This in
turn implies that P (λ) is a projective cover of M (λ).)
Are there any other indecomposable projectives in O? If P is one of
these, it has at least one simple quotient L(λ), which defines an epimorphism
ϕ : P → L(λ). The projective property of P yields a homomorphism ψ :
P → P (λ) for which πλ ψ = ϕ. Since πλ is essential, ψ must be surjective. In
turn the projective property of P (λ) yields a map ψ 0 : P (λ) → P for which
ψψ 0 is the identity on P (λ). Thus P (λ) is isomorphic to a direct summand
of P . Since P is indecomposable, P ∼ = P (λ). (This argument makes no use
of the special properties of O.)
We have proved part (a) of the following theorem.
Theorem. The modules P (λ) with λ ∈ h∗ satisfy:
(a) In category O, every indecomposable projective module is isomor-
phic to some projective cover P (λ).
3.9. Indecomposable Projectives 63

(b) When a projective module P ∈ O is written as a direct sum of

indecomposables, the number

of those isomorphic to P (λ) is equal
to dim HomO P, L(λ) .
(c) For all M ∈ O, dim HomO (P (λ), M ) = [M : L(λ)]. In particular,
dim EndO P (λ) = [P (λ) : L(λ)].

Proof. (b) Each projective in O is a direct sum of copies of various P (λ),

while P (λ) has L(λ) as its unique simple quotient. Since

dim EndO L(λ) = 1
(Theorem 1.2(g)), we see that dim HomO P, L(λ) is the number of sum-
mands isomorphic to P (λ) in any decomposition of P into indecomposables.
(c) Comparison of formal characters shows that the right side is “addi-
tive” in M : given a short exact sequence 0 → M1 → M → M2 → 0,
[M : L(λ)] = [M1 : L(λ)] + [M2 : L(λ)].
But since HomO (P (λ), ?) is exact, we get
0 → HomO (P (λ), M1 ) → HomO (P (λ), M ) → HomO (P (λ), M2 ) → 0.
Thus the dimension on the left side is also additive in M . Using additivity,
the desired equality now follows by induction on the length of M , starting
with the case when M = L(µ):
dim HomO (P (λ), M ) = δλµ = [M : L(λ)]. 
Applying Theorem 3.2(d), we see that the indecomposable injectives in
O are precisely the modules Q(λ) := P (λ)∨ . Here Q(λ) is the injective
hull of L(λ) ∼
= L(λ)∨ .
Remark. Before taking up the main results of BGG [27], we pause to
illustrate some of the techniques developed so far by proving a special result
which will be useful in Chapter 10. The reader is encouraged to work out
the proof independently as an exercise; but for completeness we give the
details: Let λ, µ ∈ h∗ , with λ dominant, and let dim L < ∞. Then


(∗) dim HomO M (λ) ⊗ L, M (µ) = dim Lµ−λ .
Recalling the discussion in the proof of Proposition 3.8(b), we can use the
assumption dim L < ∞ to replace the Hom space on the left by
HomO M (λ), L∗ ⊗ M (µ) .

Since λ is dominant, Proposition 3.8(a) says that M (λ) is projective; thus

M (λ) = P (λ). In turn, part (c) of the above theorem shows that
dim HomO M (λ), L∗ ⊗ M (µ) = [L∗ ⊗ M (µ) : L(λ)].

Recall from 1.6 that the weights of L∗ are just the negatives of the weights of
L. By Theorem 3.6, L∗ ⊗ M (µ) has a standard filtration, with each quotient
M (−ν+µ) occurring dim Lν times. So L(λ) occurs as a composition factor of
64 3. Category O: Methods

L∗ ⊗ M (µ) at least dim Lµ−λ times. On the other hand, since λ is dominant,
L(λ) cannot occur as a composition factor of any other Verma module in
the filtration.

3.10. Standard Filtrations of Projectives

Having found enough projectives in O, our next goal is to show that each
such module has a standard filtration (3.7).

Theorem. Each projective module in O has a standard
filtration. In a

standard filtration of P (λ),
the multiplicity P (λ) : M (µ) is nonzero only
if µ ≥ λ while P (λ) : M (λ) = 1.

Proof. The idea is to exploit the tensor product construction in Proposition

3.8(b). We can extract P (λ) as a direct summand of a projective of the
form M (µ) ⊗ L with µ dominant and dim L < ∞. As noted in the proof of
Theorem 3.8 (based on Theorem 3.6), this tensor product has a standard
filtration involving M (λ) once along with other Verma modules of highest
weight > λ. Thanks to Proposition 3.7(b), the direct summand P (λ) inherits
the same type of standard filtration. 

Corollary. Each projective module P ∈ O is determined up to isomorphism

by its formal character: if P 0 is also projective and ch P = ch P 0 , then
P ∼
= P 0.

Proof. ThanksL to Theorem 3.9(b), P determines uniquely integers dλ > 0

∼
for which P = λ dλ P (λ). So it suffices to show that ch P determines the
dλ . Use induction on the length of P , the length 1 case P = P (λ) = L(λ)
being obvious.
P
By Exercise 1.16, ch P = µ cµ ch M (µ) with the cµ ∈ Z determined
by P . The theorem shows that all cµ ≥ 0. Let λ be minimal among the
weights for which cλ > 0. Then the theorem implies that P (λ) occurs
precisely cλ times in the direct sum decomposition of P : that is, cλ = dλ .
After discarding these summands, the induction hypothesis takes over for
the remaining summand. 

Exercise. Extend Exercise 1.16 by showing that the [P (λ)] form a Z-basis
of the subgroup they span in K(O).

Remark. Note that if P (λ) → M (λ) is the canonical map, the existence of
a standard filtration in P (λ) implies that the kernel of this map also has a
standard filtration: see Exercise 3.7.
3.11. BGG Reciprocity 65

3.11. BGG Reciprocity

At this point we have attached to each λ ∈ h∗ three indecomposable mod-
ules related by epimorphisms: P (λ) → M (λ) → L(λ). Though the Verma
modules have no obvious categorical meaning, they play some kind of inter-
mediate role here. This is made precise in the following fundamental result
discovered by Bernstein–Gelfand–Gelfand [27]:
Theorem (BGG Reciprocity). Let λ, µ ∈ h∗ . Then


P (λ) : M (µ) = [M (µ) : L(λ)],
which is the same as [M (µ)∨ : L(λ)].

For example, in the special case when λ is maximal in its linkage class,
we have P (λ) = M (λ) thanks to Proposition 3.8(a); here both sides in the
theorem are equal to δλµ .

Proof. Unless λ and µ are linked (and also lie in the same coset of h∗
modulo Λr ) there is nothing to prove. Fix a linkage class and corresponding
central character χ. Most of the groundwork for the proof has already been
laid in the previous sections.
Since ch M (µ)∨ = ch M (µ), both have the same composition factor mul-
tiplicities. For technical reasons it is better to work here with the dual. The
strategy is to show separately that each side of the equality in the statement
of the theorem agrees with dim HomO P (λ), M (µ)∨ .

(a) The equality dim HomO P (λ), M (µ)∨ = [M (µ)∨ : L(λ)] is true

more generally when M (µ)∨ is replaced by an arbitrary M ∈ O, thanks to

Theorem 3.9(c).
(b) It remains to check that P (λ) : M (µ) = dim HomO P (λ), M (µ)∨ .

Since P (λ) has a standard filtration, this is a special case of Theorem 3.7. 
Exercise. Prove that M (λ) is projective only when λ ∈ h∗ is dominant.
Remark. The original proof of the theorem in [27] used the same strat-
egy of comparing multiplicities with dimensions of a Hom space; but the
details were more complicated, going back to a somewhat elaborate indirect
construction of projective objects in O. The modern proof, which relies
essentially on the duality functor, is developed by Irving [133, §4] in an
unpublished manuscript.
Irving emphasizes the module structure in explaining why BGG reci-
procity is natural. Roughly speaking, each occurrence of L(λ) in a compo-
sition series of M (µ) yields a quotient M of M (µ) with head L(µ) and socle
L(λ). The resulting dual module M ∨ is then a homomorphic image of P (λ).
The “pre-image” of the socle L(µ) of M ∨ in P (λ) is then shown to be the
66 3. Category O: Methods

head of a copy of M (µ) in a suitable standard filtration of P (λ). This leads

to the desired equality of multiplicities in the theorem.

By now we have seen the main features of projectives in O. However,

many fine points are yet to be investigated, some requiring fairly sophisti-
cated techniques:
• Which simple modules occur in Soc P (λ), and with what multiplic-
ity? More generally, one can ask for a detailed description of the
socle or radical series of P (λ).
• For which λ is P (λ) self-dual (isomorphic to the injective hull Q(λ)
of L(λ))?
• What is the structure of the algebra EndO P (λ)? For example,
when is this algebra commutative?

3.12. Example: sl(2, C)

By now we have enough theory in hand to sort out the indecomposable
modules in O when g = sl(2, C). As usual h∗ is identified with C, while Λ
is identified with Z. Some initial steps were already outlined in 1.5.
Since nonintegral weights (as well as −ρ) are linked to no comparable
weights in the natural ordering, the only interesting subcategories are those
of the form Oχ with χ = χλ and λ ∈ Z (but not equal to −ρ). So consider
a typical linked pair λ, µ = −λ − 2 with λ ≥ 0. Because no lower weight is
linked to µ, we must have L(µ) = M (µ). On the other hand, dim L(λ) < ∞,
while there exists a maximal vector of weight µ in M (λ) (by Proposition 1.4).
It must be unique (up to scalars) and generate a submodule isomorphic to
L(µ) in view of the fact that all weight spaces in M (λ) and L(µ) are 1-
dimensional. So M (λ) has just these two composition factors.
By Proposition 3.8(a), M (λ) = P (λ) is projective. Its dual is the injec-
tive module Q(λ), having socle L(λ) and head L(µ). What is the structure
of P (µ)? It has a standard filtration with M (µ) = L(µ) as top quotient
(Theorem 3.10). Thanks to BGG Reciprocity (3.11) and the fact that L(µ)
occurs just once as a composition factor of M (λ) and M (µ), any standard
filtration involves these Verma modules with multiplicity 1. Therefore we
must have an embedding M (λ) ⊂ P (µ), forcing L(µ) = Soc P (µ). In turn,
the extension 0 → M (λ) → P (µ) → M (µ) → 0 is nonsplit. Moreover,
P (µ) ∼= Q(µ) is self-dual.

Exercise. Use the formulas in 0.9 to compute explicitly how the Casimir
element z := h2 + 2h + 4yx ∈ Z(g) acts on these modules. In particular, z
acts on each of M (λ) and M (µ) by the scalar χλ (z) = λ2 + 2λ (call it c),
3.12. Example: sl(2, C) 67

but z − c 6= 0 on P (µ); instead (z − c)2 = 0. What is the structure of the

algebra EndO P (µ)?

So far we have found five nonisomorphic indecomposable modules in Oχ :

L(λ), L(µ) = M (µ), M (λ) = P (λ), M (λ)∨ = Q(λ), P (µ) = Q(µ).

Are there any others?

Proposition. If g = sl(2, C) and {λ, µ} is a linkage class of integral weights

with λ ≥ 0, µ = −λ − 2, and χ = χλ , then every indecomposable module in
Oχ is isomorphic to one of the five modules listed above.

Proof. We use a
mixture of general and ad
hoc reasoning. Recall that

ExtO M (ν), M (ν) = 0 = ExtO L(ν), L(ν) for all ν (Proposition 3.1).
In
our special situation, we can also show at this point that ExtO L(λ), M (λ) =
0, by inspecting part of a long exact sequence:




. . . → HomO L(λ), M (λ) → HomO M (λ), M (λ) → HomO N (λ), M (λ) →



→ ExtO L(λ), M (λ) → ExtO M (λ), M (λ) → . . .
The first term is clearly 0, while the second is C since dim M (λ)λ = 1. In the
submodule N (λ) ∼ = L(µ) and HomO L(µ), M (λ) ∼


third term, the maximal

=
HomO L(µ), L(µ) = C. As just noted, the fifth term is 0, which forces the
fourth term to be 0 as claimed.
Now suppose M ∈ Oχ is an indecomposable module but not isomor-
phic to any of those listed. Its only composition factors are L(λ) and L(µ).
Clearly each nonzero vector in Mλ is a maximal vector and therefore gener-
ates a submodule isomorphic to L(λ) or M (λ). Using the above vanishing
results for Ext, together with the standard fact that Ext commutes with di-
rect sums in either variable, we see by induction that the submodule N ⊂ M
generated by Mλ is a direct sum of copies of L(λ) and/or M (λ). So N must
be a proper submodule of M . In turn, M/N is a direct sum of copies of L(µ).
At least one L(µ) must extend N nontrivially, lest M split into a direct sum
of known indecomposables. So L(µ) must extend some summand L(λ) or
M (λ) nontrivially. In these respective cases, we get a copy of the injective
module Q(λ) or Q(µ) as a submodule of M . By the injective property, it
splits off and must therefore be all of M , a contradiction. 

Beyond rank 1, the study of indecomposable modules in categories Oχ

is far more complicated: this becomes especially evident when some of the
Verma modules have multiple composition factors and hence infinitely many
different submodules.
68 3. Category O: Methods

3.13. Projective Generators and Finite Dimensional

Algebras
Enough features of the subcategories Oχ are now visible to permit a com-
parison with the representation theory of certain finite dimensional algebras.
The latter subject is well developed and sometimes suggests methods which
are useful in the study of category O.
If χ = χλ is a central character corresponding to the linkage class
∗
L an antidominant weight λ ∈ h , consider a projective module P :=
of
w nw P (w · λ). Here the sum is taken over coset representatives in W
of the isotropy group of λ, while all nw > 0. The special property of such
a module is that a large enough direct sum P ⊕ · · · ⊕ P covers any given
M ∈ Oχ . This is an obvious consequence of the fact that the P (w · λ) rep-
resent all isomorphism classes of indecomposable projectives in Oχ . Call P
a projective generator of the category Oχ . In turn, define A := EndO P .
This is a finite dimensional algebra (Theorem 1.11).
Proposition. Let P as above be a projective generator of Oχ , with A =
EndO P . Then the functor HomO (P, ?) defines a category equivalence be-
tween Oχ and the category of finite dimensional right A-modules.

Proof. It is easy to get started. Given M ∈ Oχ , the right A-module struc-

ture on HomO (P, M ) is the obvious one induced by composing η : P → P
with ϕ : P → M . The further details of the argument are standard, as given
for example in Bass [14, Thm. II.1.3] in the general setting of an abelian
category with suitable finiteness conditions. 

While the choice of P (and hence of A) is not unique, in practice this

makes little difference. Comparison of Oχ with the module category over a
finite dimensional algebra illuminates some of the categorical features and
suggests further problems. On the other hand, this broader viewpoint is
not immediately helpful in working out the combinatorical aspects of O, for
example the formal characters of simple modules.

3.14. Contravariant Forms

Following work of Jantzen and Shapovalov (see the Notes), we introduce
here a basic tool in the study of category O. This uses the transpose map τ
of U (g) (0.5). If M is any U (g)-module, a symmetric bilinear form (v, v 0 )M
on M is called contravariant if
(u · v, v 0 )M = (v, τ (u) · v 0 )M for all u ∈ U (g), v, v 0 ∈ M.
Obviously a scalar multiple of the form will have the same property. It
turns out that nonzero contravariant forms always exist for highest weight
3.14. Contravariant Forms 69

modules (and some others) in O. But first we look at some elementary

consequences of the definition.
Proposition. Contravariant forms have the following properties:
(a) If the U (g)-module M has a contravariant form, then (Mλ , Mµ )M =
0 whenever λ 6= µ in h∗ .
(b) Suppose M = U (g) · v + is a highest weight module generated by a
maximal vector v + of weight λ. If M has a nonzero contravariant
form, then the form is uniquely determined up to a scalar multiple
by the (nonzero!) value (v + , v + )M .
(c) If U (g)-modules M1 , M2 have contravariant forms (v, v 0 )M1 and
(w, w0 )M2 , then M := M1 ⊗ M2 also has a contravariant form,
given by (v ⊗ w, v 0 ⊗ w0 )M := (v, v 0 )M1 (w, w0 )M2 . In case both of
the forms are nondegenerate, so is the product form.
(d) If M has a contravariant form and N is a submodule, the orthogonal
space N ⊥ := {v ∈ M | (v, v 0 )M = 0 for all v 0 ∈ N } is also a
submodule.
(e) If M ∈ O has a contravariant form, then the summands M χ for
distinct central characters χ are orthogonal.

Proof. (a) Given weight vectors v, v 0 of respective weights λ, µ, use the fact
that τ (h) = h for h ∈ h:
λ(h) (v, v 0 )M = (h · v, v 0 )M = (v, h · v 0 )M = (v, µ(h)v 0 )M = µ(h)(v, v 0 )M .
When λ 6= µ this forces (v, v 0 )M = 0.
(b) In view of (a) it suffices to look at values of the form on a weight
space Mµ . Typical vectors v, v 0 ∈ Mµ can be written as u · v + , u0 · v + for
suitable u, u0 ∈ U (n− ). Note that since u takes Mλ into Mµ , the element
τ (u) ∈ U (n) takes Mµ into Mλ (which is spanned by v + ). Then (v, v 0 )M =
(u·v + , u0 ·v + )M = (v + , τ (u)u0 ·v + )M , which is a scalar multiple of (v + , v + )M
depending just on the action of U (n− ) and not on the choice of the form.
(c) and (d) follow readily from the definitions. Note that the nonde-
generacy in (c) is a general
P fact unrelated to contravariance. To prove it
directly, suppose that i,j aij vi ⊗ wj 6= 0 lies in the radical of the form on
M ; take the {vi } and {wj } to be independent sets, so for fixed i or j some
(j)
aij 6= 0. Then for each fixed j, set v0 := i aij vi 6= 0 and find v (j) for
P
(j)
which cj := (v0 , v (j) )M1 6= 0. This forces j cj wj 6= 0 to lie in the radical
P
of the form on M2 , a contradiction.
(e) Abbreviate by writing (v, w) in place of (v, w)M . Let χ1 , χ2 be dis-
tinct central characters. Thanks to (a), weight spaces of M belonging to
distinct weights are orthogonal; so it is enough to show for each weight λ of
70 3. Category O: Methods

M that (Mλχ1 , Mλχ2 ) = 0. Fix z ∈ Z(g) for which χ1 (z) = c1 6= c2 = χ2 (z).

In the special case when v, w ∈ M are both eigenvectors for z, we get (as in
a familiar linear algebra argument):
c1 (v, w) = (z · v, w) = (v, z · w) = c2 (v, w),
using the contravariance and the fact that τ (z) = z (Exercise 1.10). This
forces (v, w) = 0. In general, to v ∈ Mλχ1 corresponds the least e ∈ Z+ such
that (z − c1 )e · v = 0, and similarly for w ∈ Mλχ2 . Use induction on the
sum of these two exponents. The induction hypothesis applies to each of
the pairs (z − c1 ) · v, w and v, (z − c2 )w, yielding
c1 (v, w) = (z · v, w) = (v, z · w) = c2 (v, w),
forcing (v, w) = 0 as before. 
Exercise. Show that any contravariant form on a highest weight module
must induce the zero form on its maximal submodule. In case g = sl(2, C),
use the explicit formulas for the action of g on L(λ) in 0.9 (or 1.5) to con-
struct a nondegenerate contravariant form on L(λ).

3.15. Universal Construction

We still have to show the existence of (nonzero!) contravariant forms on
special modules in O such as highest weight modules.
Rather than look immediately at M (λ), we start with a sort of universal
construction involving U (g); this will later play an essential role in the study
of Verma modules (5.8). Let ε+ : U (n) → C be the augmentation map
sending all nonconstant PBW basis elements to 0 (and define ε− similarly).
Using the PBW Theorem, consider the linear map ϕ := ε− ⊗ id ⊗ε+ :
U (g) ∼
= U (n− ) ⊗ U (h) ⊗ U (n) → U (h). Next define a bilinear form on U (g)
by the recipe C(u, u0 ) := ϕ(τ (u)u0 ). Since τ is an anti-automorphism and
interchanges U (n) and U (n− ), it is clear that ϕ(τ (u)) = ϕ(u) for all u. This
implies that our form is symmetric:
C(u, u0 ) = ϕ(τ (u)u0 ) = ϕ τ (τ (u)u0 ) = ϕ τ (u0 )τ 2 (u) = C(u0 , u).

Moreover, C(1, 1) = 1. A further property, related to contravariance, follows

immediately from the definition and the anti-automorphism property of τ :
(∗) C(u◦ u, u0 ) = C(u, τ (u◦ )u0 ) for all u◦ , u, u0 ∈ U (g).
This universal form yields C-valued forms as follows. Each λ ∈ h∗ defines
a representation of the abelian Lie algebra h and thus induces an algebra
homomorphism λ : U (h) → C. (Since U (h) = S(h), we can also view this
as evaluation of polynomial functions on h∗ .) Now compose λ with ϕ to get
a linear map ϕλ : U (g) → C. This defines in turn a C-valued symmetric
bilinear form C λ (u, u0 ) on U (g).
Notes 71

With this form in hand, it is easy to construct a contravariant form

on a highest weight module M of weight λ generated by a maximal vector
v + . Here a typical vector can P
be written (in general non-uniquely) as u · v +
− 0
with u ∈ U (n ). Set M = µ6=λ Mµ and let N be the unique maximal
submodule of M (so N ⊂ M 0 ). Writing u ∈ U (g) in the above PBW format,
observe that u · v + ≡ ϕλ (u)v + (mod M 0 ). From this it is easy to verify the
equivalence of the following statements:
(1) u · v + ∈ N,
(2) U (g)u · v + ⊂ N
(3) ϕλ (U (g)u) = 0,
(4) C λ (U (g), u) = 0,
(5) u lies in the radical of the form C λ .
In particular, if u1 , u2 ∈ U (n− ) satisfy u1 · v + = u2 · v + , then u1 − u2 is in
the radical of C λ . This allows us to transfer the form to M itself, by defining
(unambiguously) (v, v 0 )M := C λ (u, u0 ) whenever v = u · v + and v 0 = u0 · v +
with u, u0 ∈ U (n− ). From the above equivalent statements we then conclude
that N is the radical of the resulting form.
It still has to be checked that the form on M is contravariant:
(u◦ u · v + , u0 · v + )M = (u · v + , τ (u◦ )u0 · v + )M
for all u◦ ∈ U (g) and all u, u0 ∈ U (n− ). This follows immediately from (∗)
above. In view of Proposition 3.14, we have shown:
Theorem. If M is a highest weight module of weight λ, there exists a
(nonzero) contravariant form (v, v 0 )M on M . The form is defined uniquely
up to scalar multiples by the value (v + , v + )M . Its radical is the unique max-
imal submodule of M . In particular, the form is nondegenerate if and only
if M ∼
= L(λ). 
In principle the theorem affords a strategy for deciding when M (λ) =
L(λ). But for this one has to be able to determine the radical of the form,
which is not easy to do. We return to the simplicity question in Chapter 4.
Exercise. Recall that U (g) is the graded direct sum of subspaces U (g)ν with
ν ∈ Λr (0.5). In particular, restriction of C to the finite dimensional space
U (n− )−ν with ν ∈ Γ defines a U (h)-valued bilinear form Cν . Prove that this
form is nondegenerate, so its determinant relative to a PBW basis is nonzero.
[Use the fact that M (−ρ) is simple (Example 1.11), so its contravariant form
defined above is nondegenerate on each weight space.]

Notes
(3.3) In his study of parallel module categories, Donkin has popularized an
alternative notation: ∆(λ) in place of M (λ) and ∇(λ) in place of M (λ)∨ .
72 3. Category O: Methods

Proposition 3.5 is based on Jantzen [148, 1.8].

Proposition 3.7(c) was suggested by Kåhrström (based on lectures of
Mazorchuk).
The basic source for the results in 3.8–3.11 is Bernstein–Gelfand–Gelfand
[27], but the proofs have been simplified over the years. We have drawn on
some of the arguments in Irving [133, §4] and Jantzen [149, 4.6–4.9] as well
as Irving [138] (where the proof of BGG Reciprocity is formalized in the
setting of “BGG algebras”: see 13.7 below).
The ideas in 3.13 are formulated by BGG [27, Thm. 3], with the same
reference to Bass [14] given here.
(3.14–3.15) Contravariant forms arose in work by Steinberg [243, pp. 228–
229] and Burgoyne and Wong in 1971–72 (see Humphreys [132, 2.4, 3.8,
4.2]). In the present setting the results are mainly due to Jantzen [145,
146, 147, 148], with independent contributions by Shapovalov [231], as
will be seen in Chapters 4–5.
 Chapter 4

Highest Weight
Modules I

In this and the next chapter we focus on Verma modules and their simple
quotients. A number of natural questions arise, which turn out to be of
widely varying difficulty:

• When is M (λ) simple?

• What are the composition factors of M (λ), apart from multiplicity?
• What are the precise multiplicities [M (λ) : L(µ)] for all λ ∈ h∗ ?
(For example, when is the multiplicity of L(µ) equal to 1?) The
answer would give a recursive method to compute ch L(λ) for all
λ ∈ h∗ .
• What is the submodule structure of M (λ)? For example, what are
its simple submodules?
• What can be said about HomO (M (µ), M (λ)) for arbitrary λ, µ?

Underlying everything here is the question: How much depends just on

the Weyl group, as opposed to the specific weights involved? In the case
of composition factor multiplicities in Verma modules, this line of inquiry
eventually culminates in the Kazhdan–Lusztig Conjecture (Chapter 8).
Our first goal is a necessary and sufficient condition for [M (λ) : L(µ)]
to be nonzero, which requires the treatment of some of the other questions
above as well. Along the way we will bring into play a variety of methods
for the study of highest weight modules: contravariant forms, Jantzen filtra-
tions, translation functors, etc. We start by sorting out some of the easier
matters, then move on to the more challenging ones.

73
74 4. Highest Weight Modules I

Several factors tend to complicate the logic of the exposition: (1) It is

more straightforward to deal with the case when the weights involved are
integral, which is richer than the general case and more often needed in
applications. This leads us in some cases to postpone treatment of the non-
integral case. (2) The proofs of some of the main theorems are intertwined
with other theorems and therefore require temporary detours. (3) Alterna-
tive approaches to some theorems have by now been developed, requiring us
to make choices. (See the Notes for comments on the history.)
Features of Verma modules discussed in this chapter include: simple
submodules (4.1), homomorphisms (4.2), simplicity criterion (4.4, 4.8), ex-
istence of embeddings M (sα · λ) ,→ M (λ) (4.5–4.7, 4.12–4.13).

4.1. Simple Submodules of Verma Modules

It is quite easy to show that an arbitrary Verma module has a unique simple
submodule, by taking advantage of the U (n− )-module isomorphism between
U (n− ) and M (λ). The absence of zero-divisors in the noetherian ring U (n− )
(see 0.5) can be exploited via the following elementary lemma:
Lemma. Let R be a left noetherian ring. If x ∈ R is not a right zero-divisor,
then the left ideal Rx meets every nonzero left ideal of R nontrivially. In
particular, if R has no right zero-divisors, then any two nonzero left ideals
intersect nontrivially.

Proof. Suppose I is a nonzero left ideal of R, but I ∩ Rx = 0. Construct

an ascending chain of left ideals:
0 ⊂ I ⊂ I + Ix ⊂ I + Ix + Ix2 ⊂ . . .
We claim that each sum here is direct; this insures that the inclusions are
proper, contrary to the noetherian hypothesis on R. Indeed, suppose that
a0 + a1 x + · · · + an xn = 0 for some ai ∈ I. Then
a0 = −(a1 + a2 x + · · · + an xn−1 )x ∈ I ∩ Rx = 0,
forcing a0 = 0 and a1 + a2 x + · · · + an xn−1 = 0 (because x is not a right
zero-divisor). Induction on n shows that all ai = 0. 
Proposition. For any λ ∈ h∗ , the module M (λ) has a unique simple sub-
module, which is therefore its socle.

Proof. If M (λ) has distinct simple submodules L and L0 , these must in-
tersect trivially. On the other hand, in a fixed isomorphism of M (λ) with
U (n− ) the submodules must correspond to distinct nonzero left ideals, which
intersect nontrivially because of the lemma. This is impossible. 
4.2. Homomorphisms Between Verma Modules 75

Thanks to 1.3, the simple submodule of M (λ) is isomorphic to some

L(µ) with µ ≤ λ. Moreover, µ = w · λ for some w ∈ W[λ] (3.4). In order to
say anything more precise about L(µ), we need some further theory.

Exercise. Let M be a nonzero submodule of a Verma module M (λ). If M

has a nondegenerate contravariant form (3.14), prove that M is the unique
simple submodule of M (λ).

4.2. Homomorphisms Between Verma Modules

Consider next HomO (M (µ), M (λ)), which carries some (though as it turns
out not all) information about the possible occurrences of L(µ) as a com-
position factor of M (λ). As in the previous section, we identify U (n− ) as a
left U (n− )-module with either of these Verma modules. Given ϕ : M (µ) →
M (λ), denote respective maximal vectors by vµ+ and vλ+ . Then ϕ(vµ+ ) = u·vλ+
for a unique u ∈ U (n− ). In turn, ϕ(u0 · vµ+ ) = u0 u · vλ+ for all u0 ∈ U (n− ).
This means that ϕ is uniquely determined by u.

Theorem. Let λ, µ ∈ h∗ .
(a) Any nonzero homomorphism ϕ : M (µ) → M (λ) is injective.
(b) In all cases, dim HomO (M (µ), M (λ)) ≤ 1.
(c) The unique simple submodule L(µ) of M (λ) in Proposition 4.1 is
itself a Verma module.

Proof. (a) As in 4.1, we use the fact that U (n− ) has no zero-divisors. We
just saw that ϕ corresponds to the map U (n− ) → U (n− ) given by u0 7→ u0 u
for some fixed u 6= 0. Since this algebra has no zero-divisors, Ker ϕ = 0.
(b) Suppose ϕ1 and ϕ2 are nonzero homomorphisms M (µ) → M (λ),
therefore injective by part (a). Thanks to Proposition 4.1, M (µ) has a
unique simple submodule L. Then ϕ1 and ϕ2 map L isomorphically onto
respective simple submodules L1 and L2 of M (λ). Again by Proposition
4.1, L1 = L2 . This means that for some nonzero c ∈ C, the module homo-
morphism ϕ1 − cϕ2 takes L to 0. From (a) we conclude that ϕ1 − cϕ2 = 0,
forcing ϕ1 and ϕ2 to be proportional.
(c) By the universal property of Verma modules, L(µ) is the image of a
homomorphism M (µ) → M (λ), which by part (a) must be injective. 

Whenever HomO (M (µ), M (λ)) 6= 0, the theorem allows us to write un-

ambiguously M (µ) ⊂ M (λ). In the following sections we investigate when
such embeddings exist. Along the way, we can also determine which Verma
modules are simple.
76 4. Highest Weight Modules I

4.3. Special Case: Dominant Integral Weights

We pause to look at the important special case when λ ∈ Λ+ − ρ (so λ
is “dominant” in the general sense introduced in 3.5). An extreme case is
λ = −ρ, but typically there are many other weights in Λ+ − ρ having a mix
of negative and nonnegative coordinates relative to fundamental weights.
In this case the inclusions among Verma modules embedded in M (λ) are
transparent:
Proposition. Suppose λ + ρ ∈ Λ+ . Then M (w · λ) ⊂ M (λ) for all w ∈ W ;
thus all [M (λ) : L(w · λ)] > 0. More precisely, if w = sn . . . s1 is a reduced
expression for w, with si the reflection relative to the simple root αi , there
is a sequence of embeddings
M (w · λ) = M (λn ) ⊂ M (λn−1 ) ⊂ · · · ⊂ M (λ0 ) = M (λ).
Here
λ0 := λ, λk := sk · λk−1 = (sk . . . s1 ) · λ, so λn = sn · λn−1 = w · λ.
∨ i ∈ Z+
In particular, w · λ = λn ≤ λn−1 ≤ · · · ≤ λ0 = λ, with hλk + ρ, αk+1
for k = 0, . . . , n − 1.

Proof. The proof uses induction on `(w) and relies on some bookkeeping
with indices along with repeated applications of Proposition 1.4 as reformu-
lated in 1.8.
The case `(w) = 0 is obvious. For the inductive step, fix 0 < k < n
and write w0 := sk . . . s1 . From 0.3 we know that (w0 )−1 αk+1 > 0. Since
λ + ρ ∈ Λ+ , we have:
∨
hλk + ρ, αk+1 i = hw0 · λ + ρ, αk+1
∨
i = hw0 (λ + ρ), αk+1
∨
i=
∨
= hλ + ρ, (w0 )−1 αk+1
∨
i = hλ + ρ, (w0 )−1 αk+1 i ∈ Z+ .

Applying the reformulated Proposition 1.4 to this situation, we get a homo-

morphism (embedding!) M (λk+1 ) ⊂ M (λk ), with λk+1 ≤ λk . 
Exercise. Assume λ + ρ ∈ Λ+ .
(a) Prove that the unique simple submodule of M (λ) is isomorphic to
M (w◦ · λ), where w◦ is the longest element of W .
(b) In case λ ∈ Λ+ , show that the inclusions obtained in the proposition
are all proper.

For arbitrary λ ∈ h∗ (even for λ ∈ Λ), it is not so easy at this point to

specify those µ = w · λ for which an embedding M (µ) ⊂ M (λ) exists. A
basic obstacle is the failure of Proposition 1.4 to carry over to nonsimple
positive roots: even when sα · λ < λ, there is no obvious way to construct
4.4. Simplicity Criterion: Integral Case 77

an explicit embedding M (sα · λ) ⊂ M (λ). (This is explored further in 4.5

below.)
Example. To illustrate the potential problem here, let g = sl(3, C), with
∆ = {α, β}. Fix λ ∈ Λ+ and set µ := sα · λ. If γ = α + β, then sγ · µ < µ.
Since the method of Proposition 1.4 does not apply here, it is nontrivial
to find an embedding M (sγ · µ) ,→ M (µ). On the other hand, there is a
reduced expression sγ = sα sβ sα , so sγ sα = sα sβ and sγ · µ = sα sβ · λ. From
the above proposition one gets embeddings M (sγ · µ) ,→ M (sβ · λ) ,→ M (λ).
But no similar chain of embeddings passes through M (µ).

4.4. Simplicity Criterion: Integral Case

There is a straightforward necessary and sufficient condition for M (λ) to
be simple. To state it, recall from 3.5 what it means for λ ∈ h∗ to be
antidominant:
hλ + ρ, α∨ i 6∈ Z>0 for all α ∈ Φ+ .
Theorem. Let λ ∈ h∗ . Then M (λ) = L(λ) if and only if λ is antidominant.

While this is easy to state in general, the proof requires more preparation
in case λ fails to lie in Λ. So we start with the integral case and return in
4.8 to the general case.

Proof (Integral Case). Assume λ ∈ Λ.

(a) Suppose M (λ) is simple, i.e., its maximal submodule N (λ) is 0. Re-
call that in the integral case, antidominance requires only that hλ+ρ, α∨ i ≤ 0
for all simple roots α. If this fails for some α ∈ ∆, then sα ·λ < λ and Propo-
sition 1.4 provides a nonzero homomorphism M (sα · λ) → M (λ) (which we
now know to be an embedding). But this forces M (sα · λ) ⊂ N (λ), which is
absurd.
(b) Conversely, suppose λ is antidominant. Thanks to 3.5, λ ≤ w · λ for
all w ∈ W . Since all composition factors of M (λ) are of the form L(w · λ)
with w · λ ≤ λ, it follows that only L(λ) can occur as a composition factor.
But it occurs just once, so M (λ) = L(λ). 
Exercise. If λ ∈ Λ is antidominant, prove that the socle of any projective
module P (w · λ) with w ∈ W is a direct sum of copies of L(λ).
Remark. Observe how part (b) of the proof of the theorem can be extended
to arbitrary λ, by exploiting the reflection group W[λ] studied in 3.4: The
only candidates for highest weights of composition factors of M (λ) are of
the form w · λ ≤ λ with w ∈ W[λ] . Since λ is assumed to be antidominant,
Proposition 3.5 insures that λ ≤ w · λ for all such w. To extend the first
part of the proof is more challenging, since we now have to consider arbitrary
78 4. Highest Weight Modules I

positive roots; here Proposition 1.4 is not applicable. So we turn next to the
study of embeddings M (sα · λ) ,→ M (λ) when α ∈ Φ+ . This starts again
with the easier case when λ ∈ Λ and then uses a Zariski density argument.

4.5. Existence of Embeddings: Preliminaries

Assuming that λ ∈ Λ while µ = sα ·λ ≤ λ for some α ∈ Φ+ , we want to prove
the existence of an embedding M (µ) ⊂ M (λ) without actually exhibiting
a maximal vector of weight µ in M (λ). Of course, we can assume µ < λ.
The proof of Proposition 4.3 might suggest a step-by-step approach using
a reduced expression for sα . But without the dominance assumption there
we cannot guarantee at each step that successive weights are lower in the
partial ordering (Example 4.3).
Instead we proceed indirectly, applying Proposition 4.3 to an element
w ∈ W which links µ to a dominant weight. The proof requires a couple of
preparatory results, after which it is mainly a matter of careful bookkeeping.
First we record an elementary commutation property, which is needed in the
proposition below:
Lemma. Let a be a nilpotent Lie algebra, with x ∈ a and u ∈ U (a). Given
a positive integer n, there exists an integer t depending on x and u such that
xt u ∈ U (a)xn .

Proof. Recall from 0.5 that the adjoint action of a on itself extends natu-
rally to an action by derivations on U (a) (which in this case are nilpotent).
If lx and rx denote respectively the left and right multiplication by x on
U (a), then ad x = lx − rx (with all these operators commuting).
Since ad x is nilpotent on U (a), there exists q > 0 such that (ad x)q u = 0.
Now choose any t ≥ q + n and compute:
t  
t t t
X t t−i
x u = lx u = (rx + ad x) u = r (ad x)i u =
i x
i=0
q  
X t
(ad x)i u xt−i ∈ U (a)xt−q .


i
i=0
By choice of t, the final exponent is ≥ n as required. 

The following proposition contains the key technical step in the proof to
be given in the next section, but without the distracting subscripts needed
there. For its proof we need a standard identity valid in any associative
algebra with elements x, y, h := [xy] satisfying the commutation relations
for sl(2, C); this is easily proved by induction on t (as in Lemma 1.4(c)).
(1) [xy t ] = ty t−1 (h − t + 1).
4.6. Existence of Embeddings: Integral Case 79

Proposition. Let λ, µ ∈ h∗ and α ∈ ∆, with n := hλ + ρ, α∨ i ∈ Z. Assume

that
M (sα · µ) ⊂ M (µ) ⊂ M (λ).
Then there are two possibilities for the position of M (sα · λ):
(a) If n ≤ 0, then M (λ) ⊂ M (sα · λ).
(b) If n > 0, then M (sα · µ) ⊂ M (sα · λ) ⊂ M (λ).

Proof. (a) The case n ≤ 0 is straightforward:

hsα · λ + ρ, α∨ i = hsα (λ + ρ), α∨ i = hλ + ρ, −α∨ i = −n ≥ 0.
Since α is simple, Proposition 1.4 implies that M (λ) ⊂ M (sα · λ).
(b) Since n > 0, Proposition 1.4 shows that M (sα · λ) ⊂ M (λ). The
proof of the remaining inclusion requires much more work. To keep track
of the various maximal vectors involved, denote by vλ+ a maximal vector of
weight λ in M (λ), so yαn · vλ+ is a maximal vector in M (sα · λ). Similarly, let
vµ+ be a maximal vector in M (µ). The assumption M (sα ·µ) ⊂ M (µ) implies
that yαs ·vµ+ is a maximal vector of M (sα ·µ), where s := h µ + ρ, α∨ i ∈ Z+ .
With this notation, the assumption M (µ) ⊂ M (λ) insures the existence
of u ∈ U (n− ) for which u · vλ+ = vµ+ . In this situation, we can apply the
above lemma: there exists t > 0 such that yαt u ∈ U (n− )yαn . Thus
yαt · vµ+ = yαt u · vλ+ ∈ U (n− )yαn · vλ+ ⊂ M (sα · λ).
Increasing t if needed, we can assume t ≥ s. If t > s apply the above identity
(1) to get:
(2) [xα yαt ] · vµ+ = tyαt−1 (hα − t + 1) · vµ+ = (s − t)tyαt−1 · vµ+ ,
thanks to µ(hα ) = hµ, α∨ i = s − 1. Then expand the bracket on the left:
(3) xα yαt · vµ+ − yαt xα · vµ+ = xα yαt · vµ+ ∈ M (sα · λ).
Comparison of (2) and (3) shows that yαt−1 · vµ+ ∈ M (sα · λ). Reducing t
stepwise in this fashion (always with the restriction t > s), we arrive at
yαs · vµ+ ∈ M (sα · λ). Since this vector generates M (sα · µ), the desired
embedding follows. 

4.6. Existence of Embeddings: Integral Case

With Proposition 4.5 in hand, the proof of the following theorem in the spe-
cial case λ ∈ Λ mainly requires keeping track of many subscripts. However,
further ideas will be needed to handle arbitrary λ in 4.7 below.
Theorem (Verma). Let λ ∈ h∗ . Given α > 0, suppose µ := sα · λ ≤ λ.
Then there exists an embedding M (µ) ⊂ M (λ).
80 4. Highest Weight Modules I

Proof (Integral Case). Assume λ ∈ Λ, so all weights involved here are

integral. We proceed in steps.
(1) Since µ ∈ Λ, there exists w ∈ W such that µ0 := w−1 · µ lies in
Λ+ − ρ. Thanks to Proposition 4.3, there is a sequence of (not necessarily
proper) embeddings relative to a reduced expression w = sn . . . s1 (where si
is the reflection relative to a simple root αi ):
M (µ0 ) = M (µ0 ) ⊃ M (µ1 ) ⊃ · · · ⊃ M (µn ) = M (µ),
where µk := sk · µk−1 = (sk . . . s1 ) · µ0 for k = 1, . . . , n. Moreover, µ0 ≥ µ1 ≥
· · · ≥ µ.
(2) Next define a parallel list of weights associated to λ0 := w−1 · λ,
starting with λ0 := λ0 . For k = 1, . . . , n − 1, define λk := sk · λk−1 =
(sk . . . s1 ) · λ, so that λn = λ.
(3) Without loss of generality, we can assume that µ < λ. In turn, this
forces all µk 6= λk . How are these λk related to the µk ? Since µ is linked
to λ by the reflection sα , a quick calculation (left to the reader) shows that
µk = sβk · λk for some root βk (with βn = α). The reflection here is the
conjugate wk−1 sα wk , where wk := sn . . . sk+1 (setting wn := 1). In particular,
µk − λk is an integral multiple of βk .
(4) Although all the λk and µk lie in the linkage class of λ, we have
little control over the inequalities among these weights. We do know that
µk ≥ µk+1 for all k. We also know that µ < λ, whereas µ0 > λ0 because µ0
is the unique dominant weight in this linkage class. Therefore there must
exist a least index k for which µk > λk but µk+1 < λk+1 .
(5) The objective now is to prove step-by-step that there exist embed-
dings M (µk+1 ) ⊂ M (λk+1 ), M (µk+2 ) ⊂ M (λk+2 ), . . . , culminating in the
desired M (µ) ⊂ M (λ).
(6) Rewrite the definitions (recalling Exercise 1.8): µk+1 − λk+1 =
sk+1 (µk − λk ). By choice of k, the left side is a negative multiple of βk+1
whereas µk −λk is a positive multiple of βk . Since sk+1 is a simple reflection,
the only positive root which it takes to a negative root is αk+1 , forcing βk
(and βk+1 ) to equal αk+1 . Now Proposition 1.4 yields M (µk+1 ) ⊂ M (λk+1 ).
(7) Combining the previous steps, we obtain embeddings:
M (µk+2 ) = M (sk+2 · µk+1 ) ⊂ M (µk+1 ) ⊂ M (λk+1 ).
This puts us precisely in the situation of Proposition 4.5, where either of
the alternatives (a) and (b) yields M (µk+2 ) ⊂ M (sk+2 · λk+1 ) = M (λk+2 ).
(8) Iterate the argument to get M (µ) = M (µn ) ⊂ M (λn ) = M (λ). 
Exercise. Work through the steps of this argument in the special case
discussed in Example 4.3.
4.7. Existence of Embeddings: General Case 81

4.7. Existence of Embeddings: General Case

Somewhat in the spirit of the density argument in 1.9, we can complete the
proof of Theorem 4.6 for arbitrary λ ∈ h∗ . The idea here is to fix a root
α > 0 and n ∈ Z+ , then consider the set X = Xα,n ⊂ h∗ consisting of all λ
for which an embedding M (λ − nα) ⊂ M (λ) exists. Write ν := nα. (More
generally, we could replace ν in the argument by any element of Γ, as the
reader can check.) Since an embedding of this type can exist only if λ − nα
is linked to λ, the set X is contained in the affine hyperplane H = Hα,n ⊂ h∗
defined by the condition hλ + ρ, α∨ i = n.
We have already proved that Λ ∩ H ⊂ X. On the other hand, an
argument like that in 1.9 shows that Λ ∩ H is Zariski dense in H. To
conclude that X = H it will therefore suffice to prove that X is Zariski
closed, i.e., the set of common zeros of certain polynomial functions on H.
To make the situation concrete, we use the fundamental weights $1 , . . . , $`
as a basis for h∗ and express polynomial functions in terms of the coordinates
of λ. For nα, this requires us to express α first as a Z-linear combination
of the $i . Fortunately, we do not need to write down explicit polynomials,
only to show their existence.
If v + is a maximal vector of weight λ in M (λ), any maximal vector
of weight λ − ν must be annihilated by n and have the form v := u · v +
for a nonzero u ∈ N := U (n− )−ν . This is the subspace of U (n− ) having
r1 rm
as basis the PBW monomials P y1 . . . ym which correspond to positive roots
α1 , . . . , αm for which ν = i ri αi . Here dim N = P(ν), where P is Kostant’s
partition function.
The idea now is to define linear maps g λ : N → U (n− )` (the direct sum
of ` copies of the vector space U (n− )) for all λ ∈ h∗ in such a way that
λ belongs to X if and only if rank g λ < dim N (that is, g λ has nonzero
nullity). If the matrices of such linear maps are written in terms of the
weight coordinates of λ, we know from linear algebra that the rank inequal-
ity translates into the fact that certain minors of the matrix for g λ have
determinant 0 (a polynomial condition on the entries). In turn these minors
depend polynomially on λ.
Take standard bases (hi , xi , yi ) for copies si of sl(2, C) corresponding to
simple roots. Then for u ∈ U (n− ), the commutation relations show that
[xi u] = ui + u0i hi , where ui , u0i ∈ U (n− ) depend linearly on u. Then each
λ ∈ h∗ defines linear maps fiλ : N → U (n− ), where fiλ (u) = ui + λ(hi )u0i .
These combine to give a single linear map g λ : N → U (n− )` , as promised.
Now the condition above that v = u · v + be a maximal vector of weight
λ − ν translates successively into equivalent conditions on u:

(a) n · v = nu · v + = 0.
82 4. Highest Weight Modules I

(b) x1 u · v + = x2 u · v + = · · · = x` u · v + = 0.
(c) [x1 u] · v + = · · · = [x` u] · v + = 0 (since all xi · v + = 0).
(d) f1λ (u) · v + = · · · = f`λ (u) · v + = 0.
(e) all fiλ (u) = 0 (since the action of U (n− ) on M (λ) is free).
Condition (e) amounts to the single requirement that rank g λ < dim N .
This completes the proof of Theorem 4.6 in general. 
Unlike the more limited formulation in which α is required to be simple,
the theorem allows one to iterate the procedure and construct a sequence
of embeddings M (λ) ⊃ M (sα · λ) ⊃ M (sβ sα · λ) ⊃ . . . provided the weights
involved decrease in the partial ordering. This yields in particular a more
precise sufficient condition for L(µ) to occur as a composition factor of M (λ)
(with multiplicity at least 1). In 5.1 we will go on to discuss the necessity
of this condition.

4.8. Simplicity Criterion: General Case

Having obtained the analogue of Proposition 1.4 for an arbitrary positive
root (and arbitrary highest weight), we can now fill in the proof of the
criterion stated earlier for simplicity of a Verma module (4.4) but proved
there only for integral weights:
Theorem. Let λ ∈ h∗ . Then M (λ) = L(λ) if and only if λ is antidominant.

Proof. (a) Suppose M (λ) is simple. Going back to the original definition
of “antidominant”, we have to show for all α ∈ Φ+ that hλ + ρ, α∨ i 6∈ Z>0 .
Suppose this fails for some α. Since sα · λ = λ − hλ + ρ, α∨ iα, we have
sα ·λ < λ. Thanks to Theorem 4.6, there is an embedding M (sα ·λ) ⊂ M (λ),
which is proper because of the strict inequality of weights. This contradicts
the simplicity of M (λ).
(b) Suppose λ is antidominant. As we already pointed out after the
proof of the integral case in 4.4, use of the group W[λ] allows that proof to
be carried over to the general case: since λ is minimal in its W[λ] dot-orbit,
Harish-Chandra’s Theorem insures that M (λ) is simple. 
Corollary. Let λ ∈ h∗ be antidominant. Then for all w ∈ W[λ] , the socle of
P (w · λ) is a direct sum of copies of L(λ).

Proof. By Theorem 3.10, there is a standard filtration

0 = P0 ⊂ P1 ⊂ · · · ⊂ Pn = P (w · λ),
with each Pi /Pi−1 isomorphic to some M (w0 · λ). If L is a simple summand
of Soc P (w · λ), let i be the least index for which L ⊂ Pi (so L ∩ Pi−1 = 0).
4.9. Blocks of O Revisited 83

Thus L embeds in some M (w0 · λ). Combining the above theorem with
Theorem 4.2(c), L ∼
= L(λ). 

It is not so easy to determine how many copies of L(λ) occur in this

socle. Using more sophisticated arguments (see 13.14 below), Stroppel [246,
Thm. 8.1] shows (for integral weights) that Soc P (w · λ) is a direct sum of r
copies of L(λ), where r = (P (w · λ) : M (w◦ · λ)). By BGG Reciprocity, this
filtration multiplicity is the same as [M (w◦ · λ) : L(w · λ)], the computation
of which in general requires Kazhdan–Lusztig theory (8.4).
The corollary has an interesting further consequence, which will be re-
fined in 4.10 below for integral weights and in 7.16 for arbitrary weights:
Exercise. Let λ ∈ h∗ . If P (λ) ∼ = P (λ)∨ is self-dual, i.e., P (λ) ∼
= Q(λ),
prove that λ must be antidominant. What can you say about the converse?

4.9. Blocks of O Revisited

In 1.13 we formulated the notion of “block” in category O and showed by
elementary arguments that the blocks involving integral weights are the
subcategories Oχλ containing the modules whose composition factors have
highest weights linked to λ ∈ Λ. Now we can formulate a more definitive
result using the subgroups W[λ] of W introduced in 3.4:
Theorem. The blocks of O are precisely the subcategories consisting of mod-
ules whose composition factors all have highest weights linked by W[λ] to an
antidominant weight λ. Thus the blocks are in natural bijection with an-
tidominant (or alternatively, dominant) weights.

Proof. Start with an arbitrary weight µ. Thanks to Proposition 4.1 and

Theorem 4.2(c), M (µ) has a unique simple submodule L(λ) = M (λ). But
the highest weight of a simple Verma module must be antidominant, by
Theorem 4.8. Thus all composition factors of M (µ) including L(µ) lie in
the block of L(λ). The highest weights involved are then linked to λ by W[λ] .
In the other direction, L(λ) is the unique simple submodule of any M (w · λ)
with w ∈ W[λ] . 

As in Remark 3.5, we denote the block associated to an antidominant

weight by Oλ (keeping in mind the mild inconsistency in our use of O0
for the block corresponding to the linkage class of the dominant integral
weight 0). It was observed there that each of the Oλ belonging to a fixed
central character involves the same number of linked weights (hence simple
modules). This is consistent with the deeper result that all such blocks are
equivalent as categories: see Mathieu [207, Lemma A.3], where the proof
requires a localization technique using nonintegral powers of root vectors.
(Is there a more direct proof of the equivalence?)
84 4. Highest Weight Modules I

Exercise. Refine Proposition 3.14(e) as follows: If M ∈ O has a contravari-

ant form, then its block summands in distinct blocks Oλ , Oµ are orthogonal.

4.10. Example: Antidominant Projectives

Here we examine an interesting special case which expands on Exercise 4.8
and makes use of much of the theory developed so far. It illustrates well
what can and cannot be said based on what we know at this point.
The goal here is to describe the projective cover of a simple module
whose highest weight is antidominant and integral. Start with a weight
λ ∈ Λ+ − ρ. Let χ = χλ be the corresponding central character. Now
dim L(λ + ρ) < ∞ (1.6), so tensoring it with a projective produces another
projective (Proposition 3.8(b)).
We know that P (−ρ) = M (−ρ) = L(−ρ) (Example 3.8); being self-
dual, this module is also injective. Now set T := M (−ρ) ⊗ L(λ + ρ). This
is projective (as well as injective) and has a standard filtration by Verma
modules with highest weights −ρ + µ, where µ ranges over the weights of
L(λ + ρ) (3.6). The same is true of its direct summands. In particular, the
extremal weights w(λ + ρ), with w ∈ W , yield Verma modules M (w · λ) as
filtration quotients (with multiplicity 1) of T χ . These are clearly the only
weights µ of L(λ + ρ) for which −ρ + µ is linked to λ. Thus the block
summand T χ is filtered only by the M (w · λ).
Thanks to Remark 3.6, T χ has M (w◦ · λ) as a quotient. This Verma
module is actually simple, since w◦ · λ is antidominant (4.4). Since T χ
is projective, the projective cover P (w◦ · λ) must occur as a summand.
In particular, a standard filtration of P (w◦ · λ) involves various modules
M (w · λ), each occurring only once. By BGG Reciprocity (3.11), we get
[M (w · λ) : L(w◦ · λ)] ≤ 1 for all w ∈ W . On the other hand, every M (w · λ)
has L(w◦ · λ) as its unique simple submodule, by Theorems 4.2(c) and 4.4.
This forces [M (w · λ) : L(w◦ · λ)] ≥ 1. Again invoking BGG Reciprocity, we
conclude that T χ = P (w◦ · λ).
Now Remark 3.6 shows that M (λ) occurs as a submodule of T χ , hence
also its unique simple submodule L(w◦ · λ). Since T χ is injective (and in-
decomposable), it must be the injective envelope of this simple module. In
other words, in this special case the projective cover is actually self-dual.
(This can happen only in case the highest weight of the simple module is
antidominant, as noted in Exercise 4.8.)
To summarize:

Theorem. Let λ + ρ ∈ Λ+ , so w◦ · λ is antidominant and integral. Then a

standard filtration of P (w◦ · λ) involves all of the distinct M (w · λ) exactly
4.11. Application to sl(3, C) 85

once. Thus [M (w · λ) : L(w◦ · λ)] = 1 for all w ∈ W . Moreover, P (w◦ · λ)

is self-dual and equal to the injective envelope of L(w◦ · λ). 

In the language of translation functors (7.1), we have shown in particular

that translation from −ρ to w◦ · λ takes M (−ρ) to P (w◦ · λ). (See the more
general formulation in 7.13.)
What can be said if w◦ · λ is replaced by an arbitrary antidominant
weight? Although the method of proof we have used here fails to generalize,
the essential features of the proposition remain true in general. For the proof
it seems necessary to work explicitly with translation functors: see 7.16.
Remark. Self-dual projectives P turn out to be of considerable theoretical
interest, as discussed later in 13.11. A natural but rather subtle problem is to
describe the structure of the algebra EndO P . For the easy case g = sl(2, C),
see Exercise 3.12.
Exercise. Under the hypothesis of the theorem, what is dim EndO P (w◦ ·λ)?

4.11. Application to sl(3, C)

It requires a lot of new ideas (going as far as Chapter 8) to work out the
composition factor multiplicities of arbitrary Verma modules. Along the
way it is worth keeping track of what can be said in the simplest nontrivial
case: g = sl(3, C). Denote the two simple roots by α, β. Here
W = {1, sα , sβ , sα sβ , sβ sα , w◦ }.
Consider first a linkage class {w·λ | w ∈ W }, where λ is regular, integral, and
antidominant. Since λ is antidominant, Theorem 4.4 insures that M (λ) =
L(λ), while no other M (w · λ) can be simple. Moreover, Theorem 4.2(c)
shows that L(λ) is the unique simple submodule of each M (w · λ). Even
more is true: by Theorem 4.10, [M (w · λ) : L(λ)] = 1 for all w.
This makes it easy to find the composition factors (or formal character)
of M (sα · λ): The only linked weight strictly below sα · λ is λ. Since both
L(sα · λ) and L(λ) occur just once as composition factors, it follows that
ch L(sα · λ) = ch M (sα · λ) − ch M (λ). (Of course, α can be replaced by β
here.) The remaining three Verma modules require more theory to handle:
see 5.4 below.
Exercise. In the case of sl(3, C), what can be said at this point about
Verma modules with a singular integral highest weight? Leaving aside the
trivial case of −ρ, a typical linkage class has three elements: for example, if
λ lies just in the α-hyperplane and is antidominant, the linked weights are
λ, sβ · λ, sα sβ · λ.
86 4. Highest Weight Modules I

4.12. Shapovalov Elements

Before discussing the further implications of Verma’s Theorem 4.6, we take
up the natural question of whether it is possible for arbitrary γ > 0 to
construct an explicit embedding M (sγ · λ) ,→ M (λ) whenever sγ · λ < λ
(as done for simple roots in Proposition 1.4). This amounts to finding an
element u ∈ U (n− ) such that u · v + is a maximal vector of weight sγ · λ in
M (λ). Theorem 4.6 guarantees the existence of u, which is unique up to
nonzero scalar multiples (Theorem 4.2).
Though it is sometimes possible to write down u in a concrete case, as in
the exercise below, this is usually a daunting task. Indeed, if ν = λ − sγ · λ =
rγ, where r = hλ + ρ, γ ∨ i ∈ Z>0 , then u lies in U (n− )−ν . This space has
dimension P(ν), withPa basis consisting of the various PBW monomials
y1r1 · · · ymrm such that
i ri αi = γ if the positive roots are enumerated as
α1 , . . . , αm . Unless γ is simple, this space is usually too large for explicit
computations.
Moreover, one wants to understand the construction conceptually. For
example, how does the choice of u depend on λ? It is instructive to look
closely at the easiest nontrivial case, where computations can be done by
hand:
Exercise. Let g = sl(3, C), with Φ+ = {α, β, γ} (where γ = α + β) and
standard basis
(xα , xβ , xγ , hα , hβ , yα , yβ , yγ ).
(a) Use the commutation relations in g to work out products of basis
elements in the PBW ordering, such as xα yγ = yγ xα − yβ .
(b) Setting ν := γ (= ρ), the elements yα yβ and yγ form a basis of
U (n− )−ν . Rewrite the left multiple of each of these by xα and by
xβ in the PBW ordering:
xα yα yβ = yα yβ xα + yβ hα + yβ
xβ yα yβ = yα yβ xβ + yα hβ
xα yγ = yγ xα − yβ
xβ yγ = yγ xβ + yα
Similarly, rewrite xα yγ hβ and xβ yγ hβ .
(c) Write λ ∈ h∗ as a$α + b$β . Then hλ + ρ, γ ∨ i = 1 precisely when
a+b = −1 (defining an affine hyperplane H in h∗ ∼
= C2 ). Whenever
λ lies in H, sγ · λ = λ − γ. From now on, assume that λ satisfies
this condition.
(d) If r, s ∈ C, set u := ryα yβ + syγ ∈ U (n− ), with not both r and s
zero. If v + is a maximal vector in M (λ), suppose that u · v + is a
4.12. Shapovalov Elements 87

maximal vector (of weight λ − γ). Show that r(a + 1) − s = 0 and

br +s = 0, forcing r 6= 0. But the choice of r and s (which is unique
up to a nonzero scalar multiple) depends on λ. For example, take
λ = −$α with r = 1, s = 0 or λ = −$β with r = 1, s = 1.
(e) Now set u := yα yβ − yγ hβ ∈ U (b− )−γ . Show that u · v + is always
a maximal vector in M (λ), independent of the choice of λ in H.

This small example already illustrates fairly well a general idea due to
Shapovalov [231], which will be essential later in the proof of Theorem 5.8.
The basic data here include a positive but not necessarily simple root γ and
an integer r > 0 (here r = 1). On the one hand, an element u ∈ U (n− )−γ
which produces a maximal vector u · v + in M (λ) of weight λ − γ is unique
(up to a nonzero scalar multiple) but typically depends on λ. The relevant
λ form an affine hyperplane in h∗ , defined by hλ + ρ, γ ∨ i = 1. On the other
hand, an element of U (b− ) might serve the same purpose but be independent
of λ. This element is no longer unique up to scalars, only unique modulo the
left ideal in U (b− ) generated by hγ + ρ(hγ ) − 1 (= hγ + 1). In our example,
we also see that the monomial yα yβ occurs without a factor from U (h).
To formulate the general result we start with a standard basis of g and
resulting PBW basis of U (g). As usual, y1 , . . . , ym correspond to the negative
roots −α1 , . . . , −αm , where α1 , . . . , α` are simple. By “degree” of a PBW
monomial y1r1 . . . ymrm we mean its degree when viewed as a polynomial in
P
m variables, namely i ri . This contrasts with the grading of U (g) by the
root lattice.
Theorem (Shapovalov). Fix γ ∈ Φ+ and an integer r > 0. There exists
an element θγ,r ∈ U (b− )−rγ having the following properties:
(a) For each root β > 0, the commutator [xβ , θγ,r ] lies in the left ideal
Iγ,r := U (g)(hγ + ρ(hγ ) − r) + U (g)n.
(b) If γ = `i=1 ai αi , we can write
P

`
Y X
θγ,r = yirai + pj q j ,
i=1 j

with pj ∈ U (n− )−rγ , qj ∈ U (h), and deg pj < r

P
i ai .

Moreover, θγ,r is unique (up to a nonzero scalar multiple) modulo the left
ideal Jγ,r := U (b− )(hγ + ρ(hγ ) − r).

In (b) note that the highest degree term does not depend on the ordering
of the simple roots α1 , . . . , α` . Condition (b) contributes to the proof of
uniqueness by removing ambiguity as to scalar multiples; at the same time
it helps in an inductive step of the proof and makes the description of θγ,r
88 4. Highest Weight Modules I

more precise. The theorem will be proved in the next section, but first we
emphasize that it is difficult to devise a constructive method for obtaining
θγ,r explicitly. Instead we settle for a somewhat roundabout approach to
the existence.

4.13. Proof of Shapovalov’s Theorem

It is easy to get started. We fix once and for all r > 0. If γ is simple,
then thanks to Proposition 1.4 we can just set θγ,r := yγr . This suggests
an induction based on ht γ. Assuming γ 6∈ ∆, there exists α ∈ ∆ such
that p := hγ, α∨ i > 0 (0.2). We keep γ and α fixed from now on. Thus
β := sα γ = γ − pα > 0 with ht β < ht γ. The induction hypothesis provides
an element θβ,r with the desired properties.
Unfortunately, there is no obvious way to produce θγ,r from this data.
The enveloping algebra setting is too abstract for this purpose, so we try to
exploit some of the Verma module theory already developed. Given λ ∈ h∗
and a known element θ ∈ U (b− ) (such as θβ,r ), write θ(λ) for the unique
element of U (n− ) congruent to θ modulo the left ideal I. In other words,
given a maximal vector v + of weight λ in M (λ), the element θ(λ) is the
unique element u ∈ U (n− ) such that θ · v + = u · v + . (This is independent of
the choice of v + .) The hope is to reconstruct the unknown θγ,r from such
specializations.
As in some earlier proofs, a Zariski-density argument will enable us to
focus just on selected weights. Since we are only interested here in those λ
for which M (λ) has a submodule isomorphic to M (λ − rγ), we work in the
hyperplane Hγ,r ⊂ h∗ defined by
Hγ,r := {λ ∈ h∗ | hλ + ρ, γ ∨ i = r}.
We know that Λ ∩ Hγ,r is Zariski-dense in Hγ,r . Intersecting this set with
the half-space
Hα := {λ ∈ h∗ | hλ + ρ, α∨ i < 0}
still yields a dense subset of the hyperplane; call it Θ.
The strategy now is to assign to each λ ∈ Θ an element θγ,r (λ) ∈
U (n− )−rγ satisfying analogues of (a) and (b) in the theorem together with a
polynomial condition on its coordinates relative to a PBW basis of U (n− ).
This will allow the construction to be extended to all λ ∈ Hγ,r by using the
density of Θ.
Recall that I denotes the annihilator in U (g) of a maximal vector v + in
M (λ). We require θγ,r (λ) to satisfy:
(a0 ) [xγ , θγ,r (λ)] ∈ I.
4.13. Proof of Shapovalov’s Theorem 89

(b0 ) Independent of the choice of λ ∈ Θ, the highest degree term in

θγ,r (λ) when written in a standard PBW basis is i yirai . (This
Q
condition is independent of the ordering of roots.)
(c0 ) The coefficients of θγ,r (λ) in the PBW basis (hence in any other
basis of U(n− )) depend polynomially on λ.
What (c0 ) means is this: relative to the fixed basis of U (n− ), there are
elements pj ∈ U (n− )−rγ and qj ∈ U (h) (independent of λ) so that θγ,r (λ) =
as polynomial functions on h∗ . According
P
j pj qj (λ). Here the qj are viewed
ra
to (b0 ), qj = 1 whenever pj = i yi i . Since Θ is Zariski-dense
Q
P in Hγ,r , parts
(a) and (b) of the theorem will follow if we set θγ,r := j pj qj ∈ U (b− ).
In view of (b), it follows from the proof that θγ,r is unique modulo the left
ideal Jγ,r of U (b− ), which specializes at each λ to the annihilator in U (b− )
of a maximal vector generating M (λ).
Recall the data established at the outset: r > 0 and γ > 0 of height > 1,
together with α ∈ ∆ such that p = hγ, α∨ i > 0 and β = sα γ = γ − pα > 0 of
lower height. The induction hypothesis provided θβ,r ∈ U (n− ) satisfying the
conditions of the theorem; in particular, all possible specializations θβ,r (µ)
are defined for µ in the hyperplane Hβ,r . Note that β = `i=1 bi αi with
P
bi = ai except when αQi = α, in which case bi = ai − p. Thus the highest
degree term of θβ,r is i yirbi .
Now consider λ ∈ Θ, so hλ + ρ, α∨ i = −q < 0 with q ∈ Z>0 . Define
µ := sα · λ = λ + qα > λ and observe that µ ∈ Hβ,r . A quick calculation
shows that sα · (µ − rβ) = λ − rγ, with hµ − rβ + ρ, α∨ i = q + rp (call this
positive integer n). So we have an embedding M (λ − rγ) ,→ M (µ − rβ).
Finally, we can embed M (µ − rβ) ,→ M (µ) because r > 0. The resulting
inclusions of Verma modules are pictured in Figure 1.
M (µ)
@
@
@
M (λ) M (µ−rβ)
@
@
@
M (λ−rγ)

Figure 1. Inclusions of Verma modules

Thanks to the induction hypothesis, in three of the four cases we have

in hand an explicit element in U (n− ) taking a given maximal vector of the
Verma module to a maximal vector for the embedded submodule: yαq , yαn ,
and the specialization θβ,r (µ). Since such elements are unique up to scalars,
90 4. Highest Weight Modules I

the embedding M (λ−rγ) ,→ M (λ) is induced by a unique element θγ,r (λ) ∈

U (n− ) satisfying:
(1) θγ,r (λ)yαq = yαn θβ,r (µ), with n = q + rp.
It is clear that this new element θγ,r (λ) satisfies (a0 ) above. To check
(b0 ), compare the P two sides of (1): POn the right,Pthe highest degree of a
monomial is n + r i bi = q + rp + r i bi = q + r i ai , since bi = ai unless
αi = α (in whichQcase bi = ai − p). Thus the monomial of highest degree in
θγ,r (λ) must be i yirai , as required.
For (c0 ), use an alternative PBW basis for U (n− ) in which the power of
yα in each monomial occurs at the right end. When the right side of (1)
is rewritten in this way, comparison with the left side shows that yαq must
be a factor in every monomial. Since U (n− ) has no zero divisors, we can
then cancel yαq from both sides to get an explicit expression for θγ,r (λ). The
rewriting process just involves commutation relations in g (independent of
the weights). Now property (c0 ) for θβ,r (µ), in tandem with the fact that λ
depends linearly on µ, insures that all coefficients of θγ,r (λ) (for any basis
of U (n− )) depend polynomially on λ.

4.14. A Look Back at Verma’s Thesis

Verma’s 1966 thesis [251] (and the announcement [252]) stimulated the
work of BGG and others in the 1970s. Although his results and methods
have since been greatly improved and extended, the thesis itself merits some
concluding comments in this chapter. We indicate briefly where Verma’s
main ideas fit into our treatment (using our notation and terminology rather
than his).
His results Lemma 2.8 and Theorem 3.10 on homomorphisms between
Verma modules are equivalent to our Theorem 4.2. However, the dimension
bound on the Hom space was originally proved by studying the growth in
dimensions of weight spaces using the partition function (Appendix B).
His Theorem 3.1 is our 1.4, while his Theorem 3.2 is our 4.3.
Verma made one crucial error: he thought he had proved that all sub-
modules of M (λ) are generated by maximal vectors; in turn, all composition
factors L(µ) would result from embeddings M (µ) ,→ M (λ) and thus occur
with multiplicity 1. The source is a subtle gap in the last paragraph of the
proof of his Lemma 3.7 (which propagates to Theorem 3.8 and eventually
Lemma 7.1, Theorem 7.2). Counterexamples to his assertion emerged in the
work of BGG [25] and of Conze–Dixmier [72, Ex. 2]. The latter paper gives
an example in type D4 of a Verma module M (λ) having infinitely many
submodules not generated by their Verma submodules: here λ + ρ is taken
to be the fundamental weight corresponding to the vertex of the Dynkin
Notes 91

diagram connected to all other vertices. A full understanding of the multi-

plicity problem only emerged with the formulation of the Kazhdan–Lusztig
Conjecture in 1979 (as explained in Chapter 8 below).
Verma’s §5, which is unaffected by the error, yields our Theorem 4.6 on
embeddings M (sα · λ) ,→ M (λ).
In §6 of the thesis there are a number of interesting examples and coun-
terexamples, leading to the elaboration in §7 of the conjecture later proved
in a stronger form by BGG (5.1). Here the role of the Bruhat ordering is
first made explicit.

Notes
The results in this chapter are mainly due to Verma [251], BGG [25, 26],
and Shapovalov [231], with further refinements by Dixmier [84].
When reading BGG or Dixmier, it has to be kept in mind that their
notational conventions differ from ours in one essential respect: their M (λ)
denotes the Verma module with highest weight λ − ρ, while M (wλ) is the
Verma module with highest weight linked to λ − ρ by w. Some improve-
ments were incorporated in the English translation of Dixmier’s book, later
reprinted by the AMS. (But this translation contains a large number of un-
corrected misprints, most of which are not in the short list given at the end
of the AMS reprint.)
For Lemma 4.1 see Dixmier [84, 3.6.9]. [This is incorrectly cited as
3.6.13 in the simplified proof due to W. Borho of [84, 7.6.3] in the English
translation.] In Proposition 4.1 we follow Borho’s line of argument.
Theorem 4.6 is due originally to Verma [251, §5]; here we mainly follow
the proof in Dixmier [84, 7.6.9–7.6.13].
Following the work of Shapovalov [231], a number of people have stud-
ied further the embeddings of Verma modules M (sα · λ) ,→ M (λ). Lutsytuk
[205] develops a recursive formula for the elements of U (n− ) inducing such
embeddings (but claims erroneously that these must coincide with Shapo-
valov’s elements). Franklin [97] shows how to construct the Shapovalov
elements in a Z-form of U (g) which is well suited to reduction modulo a
prime. This is inspired by work of Carter on raising and lowering operators.
Computational methods explored more recently by de Graaf [116] admit
natural generalizations to Kac–Moody algebras.
In the proof of Theorem 4.10 the multiplicity 1 property is a crucial
point. For nonintegral weights this remains valid (7.16) and can be ap-
proached in several ways, none quite as direct as the argument given here
and typically requiring more information than we have yet developed. See
Borho–Jantzen [42, 2.10], Conze–Dixmier [72, Lemme 1], Jantzen [148,
92 4. Highest Weight Modules I

2.23]. In the setting of Chapter 8, the multiplicity 1 in 4.10 corresponds to

the fact that a related Kazhdan–Lusztig polynomial is equal to 1 (see for
example [130, Exercise 7.14]).
 Chapter 5

Highest Weight
Modules II

Now we turn to developments after 1970 due to Bernstein–Gelfand–Gelfand

and Shapovalov, along with important refinements due to Jantzen (to be
studied further in Chapter 7). A central result is the BGG Theorem on
composition factors of Verma modules (5.1), which can be reformulated for
regular weights in terms of the Bruhat ordering (5.2). We emphasize the
alternative proof of the theorem by Jantzen in the framework of Jantzen
filtrations (5.3–5.7). In the proof given here, the determinant of a nonde-
generate contravariant form on a weight space plays an essential role. It was
computed independently by Shapovalov and Jantzen (5.8–5.9).

5.1. BGG Theorem

In Chapter 4 we raised the question of determining which simple modules
occur as composition factors of a given Verma module M (λ). Leaving aside
the much more delicate question of their exact multiplicities, we can now
state the definitive answer which was conjectured (in an oversimplified form)
by Verma following his proof of Theorem 4.6 and ultimately proved by
Bernstein–Gelfand–Gelfand [25, 26]. For this a special notation will be
useful: Let λ, µ ∈ h∗ and write µ ↑ λ if µ = λ or there is a root α > 0 such
that µ = sα · λ < λ; in other words, hλ + ρ, α∨ i ∈ Z>0 . More generally, if
µ = λ or there exist α1 , . . . , αr ∈ Φ+ such that

µ = (sα1 . . . sαr ) · λ ↑ (sα2 . . . sαr ) · λ ↑ . . . ↑ sαr · λ ↑ λ,

we say that µ is strongly linked to λ and write µ ↑ λ.

93
94 5. Highest Weight Modules II

Theorem. Let λ, µ ∈ h∗ .
(a) (Verma) If µ is strongly linked to λ, then M (µ) ,→ M (λ); in
particular, [M (λ) : L(µ)] 6= 0.
(b) (BGG) If [M (λ) : L(µ)] 6= 0, then µ is strongly linked to λ.

By iterating Theorem 4.6, we immediately obtain part (a).

Part (b) of the theorem lies somewhat deeper. Rather than present the
original proof by BGG [25, 26] (which is exposed thoroughly by Dixmier
[84, Chap. 7]), we shall develop the proof using methods of Jantzen [148]
which are of independent interest: the Jantzen filtration and contravariant
forms.
Example. As an application of the theorem, Proposition 4.3 can be ex-
tended to cover the case when λ ∈ h∗ is dominant in the sense of 3.5 but
not necessarily integral: For all w ∈ W[λ] , there is an embedding M (w · λ) ⊂
M (λ); in particular, [M (λ) : L(w · λ)] > 0 for all w ∈ W[λ] . The observation
before the statement of Proposition 4.3 remains valid when α ∈ ∆[λ] and
w ∈ W[λ] ; but here α need not lie in ∆ (Exercise 3.4). So the above theorem
replaces Proposition 1.4 in the proof of 4.3, insuring the existence of the
relevant embeddings of Verma modules.
Remarks. (1) Taken together, the two parts of the theorem imply the
following equivalence:
(∗) Let λ, µ ∈ h∗ . Then [M (λ) : L(µ)] 6= 0 ⇔ M (µ) ,→ M (λ).
This can be proved independently of the BGG Theorem, by some inductive
arguments involving BGG reciprocity and the Ext functor: see Moody–
Pianzola [223, 2.11]. The subtle point, however, is that not every occur-
rence of L(µ) as a composition factor of M (λ) need be accounted for by a
straightforward homomorphism M (µ) → M (λ). Multiplicities bigger than
1 actually begin to show up in rank 3.
(2) Verma only conjectured that strong linkage should be a necessary
condition for M (µ) ,→ M (λ), believing that he had already proved (∗) (as
discussed in 4.14).
Exercise. Let λ ∈ h∗ . If all [M (λ) : L(µ)] = 1, use (∗) to prove that
the maximal submodule N (λ) of M (λ) is the sum of the corresponding
embedded Verma modules M (µ). (Examples show however that the converse
is false.)

5.2. Bruhat Ordering

It is useful to reformulate Theorem 5.1 in terms of the Weyl group alone.
For this we have to invoke some basic facts about the Chevalley–Bruhat
5.3. Jantzen Filtration 95

ordering of W : see 0.4. To make the transition from weights to elements of

W , we consider only regular integral weights (1.8), meaning “dot-regular”.
For example, 0 is regular, but −ρ is not. The linkage class of such a weight
has size |W | and is indexed by its lowest weight λ, which is antidominant:
this means that hλ + ρ, α∨ i 6∈ Z>0 for all α > 0 (3.5).
Now suppose α > 0 and sα ·(w·λ) < w·λ for some w ∈ W , which accord-
ing to Theorem 4.6 insures that M (sα w · λ) ,→ M (w · λ). The assumption
is equivalent to hw · λ + ρ, α∨ i ∈ Z>0 . This pairing can be rewritten as
hw(λ + ρ), α∨ i = hλ + ρ, w−1 α∨ i = hλ + ρ, (w−1 α)∨ i.
Since λ is antidominant, the positivity assumption forces w−1 α < 0, or
(w−1 sα )α > 0. Writing w0 := sα w, this implies
`(w) = `(sα w0 ) = `((w0 )−1 sα ) > `((w0 )−1 ) = `(w0 )
(see 0.3 for the relationship between length and positivity). This means by
sα
definition (0.4) that w0 → w. Scrutiny of the steps shows that they are
reversible, so we conclude that
(sα w) · λ < w · λ ⇐⇒ sα w < w in the Bruhat ordering.
Iterating, we get for all w0 , w ∈ W :
w0 · λ < w · λ ⇐⇒ w0 < w.
Since the right hand side is independent of which (regular, antidominant) λ
we choose, the theorem implies:
Corollary. Let λ ∈ Λ be antidominant and regular, and let w, w0 ∈ W .
Then [M (w · λ) : L(w0 · λ)] 6= 0 if and only if w0 ≤ w. 

This consequence of the BGG Theorem raises a natural question: Can

the composition factor multiplicities here be specified purely in terms of W ?
(The answer requires some completely new ideas: see Chapter 8.)

5.3. Jantzen Filtration

Given an arbitrary Verma module M (λ) with λ ∈ g∗ , Jantzen introduced
a remarkable filtration of M (λ) which leads to a fairly conceptual proof of
the BGG Theorem. It also leads to new information about the multiplicities
of composition factors of M (λ) in special cases. We first state his theorem;
filling in the details of the proof will occupy the following sections.
Theorem (Jantzen). Let λ ∈ h∗ be arbitrary. Then M (λ) has a filtration
by submodules
M (λ) = M (λ)0 ⊃ M (λ)1 ⊃ M (λ)2 ⊃ . . .
with M (λ)i = 0 for large enough i, satisfying:
96 5. Highest Weight Modules II

(a) Each nonzero quotient M (λ)i /M (λ)i+1 has a nondegenerate con-

travariant form in the sense of 3.14.
(b) M (λ)1 = N (λ), the unique maximal submodule of M (λ).
(c) The formal characters satisfy
X X
ch M (λ)i = ch M (sα · λ).
i>0 α>0, sα ·λ<λ

The filtration itself is known as the Jantzen filtration, while the

formula in part (c) is called the Jantzen sum formula . Denote by
M (λ)i := M (λ)i /M (λ)i+1 the ith filtration layer.
In part (c), the summation on the right side is over a set of positive
roots Φ+ λ . It can be characterized in terms of the subset Φ[λ] of roots α for
which hλ + ρ, α∨ i ∈ Z (3.4): this is a root system, with a positive system
Φ[λ] ∩ Φ+ . In turn, write Φ+ λ for the subset of those α satisfying sα · λ < λ,
i.e., hλ + ρ, α∨ i ∈ Z>0 . Note that this set depends on λ, not just on its coset
modulo Λ. (Jantzen [148, 5.3] writes R+ (λ).)
If M (λ)n 6= 0 but M (λ)n+1 = 0, the sum formula shows that n = |Φ+ λ|
provided one knows that the unique simple submodule L(µ) of M (λ) occurs
just once as a composition factor of M (λ) (hence also of each submodule
M (sα · λ)). This is a general fact, as proved for λ ∈ Λ in 4.10 and in 7.16
below for arbitrary λ.

Exercise. Let λ be regular, antidominant, and integral. In the Jantzen

filtration of M (w · λ), prove that the number n above is just `(w), so there
are `(w) + 1 nonzero layers in the filtration (as illustrated in 5.4 below by
sl(3, C)). [Use 0.3(2) to describe Φ+w·λ .]
More generally, if λ ∈ h∗ is regular and antidominant, while w ∈ W[λ] ,
we have |Φ+
w·λ | = `λ (w). Here `λ is the length function on W[λ] determined
by ∆[λ] .

A number of natural questions arise from Jantzen’s theorem. Not all

have yet been completely answered, while in some cases the answers turn
out to be quite deep: see Chapter 8.
(1) Is the Jantzen filtration of M (λ) unique relative to properties such
as the sum formula and the existence of nondegenerate contravari-
ant forms on the quotients?
(2) What are the composition factor multiplicities in each filtration
layer M (λ)i ?
(3) Are the filtration layers semisimple? If so, does the filtration coin-
cide with one of the standard module filtrations (radical or socle)
5.4. Example: sl(3, C) 97

having semisimple quotients (see 8.15)? This would provide an in-

trinsic definition not requiring change of base as in 5.7 below. (In
any case, the sum formula would require further subtle arguments.)
(4) When M (µ) ,→ M (λ), how do the respective Jantzen filtrations
behave relative to the essentially unique homomorphism involved?
More precisely, suppose µ ↑ λ and set r := |Φ+ +
λ | − |Φµ |. One might
expect that M (µ) ⊂ M (λ) if i ≤ r while M (µ)∩M (λ)i = M (µ)i−r
i

if i ≥ r.
Most of these questions were raised by Jantzen in [148, 5.17]. All the
evidence he compiled in low ranks encouraged the hope that the filtrations
would behave reasonably. In particular, the hereditary property in (4) be-
came known as the Jantzen Conjecture. This eventually turns out to be
true, but has been proved only using deep geometric ideas in the framework
of Kazhdan–Lusztig theory (8.12–8.13).

5.4. Example: sl(3, C)

Let g = sl(3, C), with the notation of 4.11. Taking λ to be regular, an-
tidominant, and integral, the sum formula makes it easy to work out the
Jantzen filtration for each M (w · λ). From this in turn we can deduce that
all composition factors occur with multiplicity one. This is already clear
when w = 1, sα , sβ : the unique simple submodule M (λ) = L(λ) always
occurs just once as a composition factor of M (w · λ) as does L(w · λ).
Suppose w = sα sβ , so the four weights w0 · λ ≤ w · λ correspond to
w0 = 1, sα , sβ , w (which happen to be below w in the Bruhat ordering). We
know that L(w · λ) and L(λ) each occurs once as a composition factor of
M (w · λ). On the other hand, the sum formula gives
X
ch M (w · λ)i = ch M (sα · λ) + ch M (sβ · λ)
i>0

= ch L(sα · λ) + ch L(sβ · λ) + 2 ch L(λ).

It follows quickly that M (w · λ)2 = L(λ) and M (w · λ)i = 0 for i > 2. This
implies that M (w · λ) has four composition factors, each with multiplicity
one. Similar reasoning then applies to the cases w = sβ sα and w = w◦ .
Question: What can be said if λ is no longer regular or integral?

Exercise. Consider the principal block O0 for sl(3, C). It has six simple
modules Lw := L(w · (−2ρ)) and corresponding Verma modules Mw . Here
the character of L1 = M1 is given by the partition function p. When w 6= 1,
compute ch Lw and show that all weight spaces have dimension one.
98 5. Highest Weight Modules II

5.5. Application to BGG Theorem

As promised, we can deduce from Jantzen’s theorem the BGG Theorem 5.1:
If [M (λ) : L(µ)] 6= 0, then µ is strongly linked to λ.

Proof of BGG Theorem. Given λ ∈ h∗ , proceed by induction on the

number of linked weights µ ≤ λ. In case λ is minimal in its linkage class,
M (λ) = L(λ) and there is nothing to prove.
For the induction step, suppose [M (λ) : L(µ)] > 0, with µ < λ. This
means that [M 1 : L(µ)] > 0. The sum formula forces [M (sα · λ) : L(µ)] > 0
for some α ∈ Φ+ +
λ . By induction, there exist α1 , . . . , αr ∈ Φ such that
µ = (sα1 . . . sαr )sα · λ ↑ (sα2 . . . sαr )sα · λ ↑ . . . ↑ sαr sα · λ ↑ sα · λ.
Coupled with the given sα · λ < λ, we get the desired upward chain from µ
to λ. 
Example. Among the examples worked out directly by Verma [251], the
case when g = sl(4, C) is instructive. Only in rank 3 does one begin to
encounter weights λ such that w · λ < λ but w · λ fails to be strongly linked
to λ. To construct an example, number the simple roots and fundamental
weights in the usual way and abbreviate c1 $1 + c2 $2 + c3 $3 by (c1 , c2 , c3 ).
Start with λ = (1, −2, −1) ∈ Λ. Set w = s2 s3 s2 s1 s2 (a reduced expression
in W ). Here λ + ρ = α1 , while w(λ + ρ) = −α3 . It follows that λ − w · λ =
(λ + ρ) − w(λ + ρ) = α1 + α3 , proving that w · λ < λ. On the other hand, his
computations in U (n− ) (and its division ring of quotients) show that there
is no possible embedding of M (w · λ) into M (λ).
This example can be explained in terms of the Bruhat ordering and the
BGG Theorem (which Verma conjectured on the basis of this and similar
cases). The weight in Λ+ − ρ linked to λ is µ = (0, −1, 0). Here λ = s2 s3 · µ,
whereas w · λ = s2 s3 s2 s1 s2 s2 s3 · µ = s3 s2 s1 · µ. Inspection of the Bruhat
ordering for W = S4 shows that s3 s2 s1 is unrelated to s2 s3 .

5.6. Key Lemma

Next we turn to the proof of Jantzen’s theorem (5.3). Before constructing
his filtration, we isolate in an elementary lemma some of the technical ideas
needed (following [148, 5.1]).
Recall some familiar facts about free modules over principal ideal do-
mains. Let A be a principal ideal domain and M a free A-module of finite
rank r. We use letters such as e and f to denote elements of M . Now sup-
pose M has an A-valued symmetric bilinear form (e, f ), “nondegenerate”
in the sense that (e, f ) = 0 for all f ∈ M implies e = 0. This means that
the matrix of the form relative to a basis of M has nonzero determinant
5.6. Key Lemma 99

D. While D depends on the choice of the basis, it is determined up to a

unit in A: changing the basis is implemented by a matrix in GL(r, A) whose
determinant is a unit in A.
The dual module M ∗ := HomA (M, A) is again free of rank r. It has a
submodule M ∨ consisting of all functions e∨ : M → A defined by e∨ (f ) =
(e, f ) with e ∈ M . Nondegeneracy of the form implies that ϕ : M → M ∨
given by ϕ(e) = e∨ is an isomorphism. Thus M ∨ also has rank r.
Part of the standard structure theory for modules over principal ideal
domains shows that M ∗ has a basis e∗1 , . . . , e∗r such that {di e∗i } is a basis of
M ∨ for some nonzero di ∈ A. This basis of M ∗ is dual to a unique basis
{ei } of M : thus e∗i (ej ) = δij . On the other hand, there is another basis {fi }
of M mapped by ϕ to the elements fi∨ = di e∗i . Comparing these two bases
of M , we get:

(1) (ei , fj ) = fj∨ (ei ) = dj e∗j (ei ) = dj δij .

Q
In particular, D = i di (up to units).
To state the key lemma, we also need to fix a prime element p ∈ A, with
associated valuation vp (a) = n, where pn | a but pn+1 - a. Now M := M/pM
is a vector space over the field A := A/pA. For each n ∈ Z+ , define

M (n) := {e ∈ M | (e, M ) ⊂ pn A}.

These submodules of M form a descending chain, starting with M = M (0).

Their images under reduction modulo p are subspaces M (n) of M .

Lemma. Let A be a principal ideal domain, with prime element p. Suppose

M is a free A-module of rank r, with a nondegenerate symmetric bilinear
form (e, f ) having nonzero determinant D relative to some basis of M .

(a) Defining M (n) and M (n) as above, we have

X
vp (D) = dimA M (n).
n>0

(b) For each n, the modified bilinear form (e, f )n := p−n (e, f ) on M (n)
induces a nondegenerate form on M (n)/M (n + 1).

Proof. (a) The proof will show that the indicated sum is finite, with M (n) =
0 for sufficiently large n.
We just have to combine the abovePremarks using the notation intro-
duced there. If f ∈ M is written as j aj fj with aj ∈ A, then a quick
calculation using (1) above shows that (ei , f ) = ai di for each i. In turn, for
100 5. Highest Weight Modules II

each fixed n > 0,

f ∈ M (n) ⇔ vp ((ei , f )) ≥ n for all i
⇔ vp (ai di ) ≥ n for all i
⇔ vp (ai ) + vp (di ) ≥ n for all i
⇔ vp (ai ) ≥ n − ni ,
where ni := vp (di ). With this in hand, we see that M (n) is spanned by those
fi for which n ≤ ni together with those pn−ni fi for which n > ni . Thus the
fi with n ≤ ni form a basis of M (n), forcing dim M (n) = #{i | n ≤ ni }. In
particular M (n) = 0 for sufficiently large n.
Q
Finally, we use D = i di to compute
X X r
X
dim M (n) = #{i | n ≤ ni } = ni = vp (D).
n>0 n>0 i=1

(b) Clearly the form (e, f )n takes values in A. It has to be checked first
that it actually induces a form on M (n): in other words, any e ∈ M (n)
lying in pM satisfies (e, M (n))n ∈ pA. This follows from
(e, M (n))n ⊂ p−n (pM, M (n)) ⊂ p−n+1 (M, M (n)) ⊂ pn−n+1 A = pA.

Now the induced form (e, f )n makes sense. What is its radical? Clearly
each f ∈ M (n + 1) satisfies
(f, M (n))n = p−n (f, M (n)) ⊂ p−n (f, M ) ⊂ p−n+n+1 A = pA,
forcing f to lie in the radical. On the other hand, recalling (1) above, we
have (ei , fj )n = 0 if i 6= j but otherwise = di p−n whenever n ≤ ni . In
particular, the fi lie in the radical of the form just when n < ni . Since
the fi with n ≤ ni form a basis of M (n), we conclude that the form is
nondegenerate on the quotient M (n)/M (n + 1). 

5.7. Proof of Jantzen’s Theorem

Now we can complete the proof of Theorem 5.3, following [148, 5.3], modulo
a determinant computation to be carried out in 5.8–5.9. For this we work
with a fixed but arbitrary highest weight λ ∈ h∗ . To construct a filtration
of M (λ) with the required properties requires a somewhat roundabout ap-
proach, since there is no obvious way to specify directly the submodules
M (λ)i and then extract the sum formula.
Jantzen’s strategy is to extend the base field to K := C(T ), a field of
rational functions in one indeterminate T . It is not difficult to check that
the previous theory developed for g applies equally well to the extended Lie
algebra gK := K ⊗C g over K (which is again a splitting field for g). The
5.7. Proof of Jantzen’s Theorem 101

essential ingredients we need are the construction of Verma modules, the

properties of their contravariant forms (3.14–3.15), and the criterion for a
Verma module to be simple (Theorem 4.8).
Let A be the principal ideal domain C[T ] having K as field of fractions.
Then the theory over K adapts easily to the base ring A, with gA := A ⊗C g.
By using Lemma 5.6 to construct filtrations in the separate weight spaces
of a Verma module and then “reducing modulo T ”, we shall get the needed
data over the residue field C to assemble the Jantzen filtration and sum
formula there.
Given λ ∈ h∗ , set λT := λ + T ρ ∈ h∗K . Obviously hλT + ρ, α∨ i 6∈ Z for all
α ∈ Φ. In particular, λT is antidominant. Thanks to Theorem 4.8, M (λT )
is simple. Equivalently, the (unique up to scalars) contravariant form on
M (λT ) is nondegenerate (3.14).
From the natural A-form U (gA ) ∼ = A ⊗C U (g) in U (gK ) ∼ = K ⊗C U (g),
we get an A-form M (λT )A in which each weight space is a free A-module
of finite rank. Moreover, the contravariant form on M (λT ) restricts to an
A-valued nondegenerate symmetric bilinear form on each of these weight
spaces. A typical weight occurring is λT − ν with ν ∈ Γ, so for each ν we
can apply Lemma 5.6 to this A-form MλT −ν of M (λT )λT −ν . Then set
X
M (λT )iA := MλT −ν (i).
ν∈Γ

Recalling how the filtrations of weight spaces are defined, we see at once
that the M (λT )iA are U (gA )-submodules and form a decreasing filtration of
M (λT )A . Moreover, the original contravariance property adapts at once to
these A-modules.
With this set-up in place over A, setting T = 0 recovers M (λ) ∼ =
M (λT )A /T M (λT )A and with it a decreasing filtration M (λ)i . Thanks to
the lemma, the quotients in this filtration acquire nondegenerate contravari-
ant forms. As the proof of the lemma shows, the filtration of each individual
weight space in M (λ) eventually ends at 0. Since only finitely many weights
are linked to λ, it follows that for large enough i we have M (λ)i = 0.
Having proved part (a) of the theorem, we immediately get part (b):
the quotient M (λ)/M (λ)1 is a highest weight module with a nondegenerate
contravariant form, hence simple (3.14).
It remains to verify the sum formula (c):
X X
ch M (λ)i = ch M (sα · λ), with the sum taken over α ∈ Φ+
λ.
i>0

Applying part (a) of Lemma 5.6, we can express the sum on the left in terms
of the determinants of the contravariant forms on the filtration quotients
102 5. Highest Weight Modules II

for individual weight spaces in the A-form. For this we need to invoke an
explicit formula worked out independently by Shapovalov [231] and Jantzen
[145, 147]. Its proof involves technical steps unrelated to Jantzen’s theorem
which will be explained in the following sections. (The computation there
is done under more general hypotheses.) For our weight λT in the A-form,
the formula gives the determinant of the universal contravariant form (3.15)
specialized to the λT − ν weight space of M (λT ):
Y Y
P(ν−rα)
(1) Dν (λT ) = hλT + ρ, α∨ i − r
α>0 r>0
Here P is the Kostant partition function. The determinant itself depends on
the choice of a basis in U (n− ) and is therefore unique only up to a nonzero
scalar multiple. But all we need is the value of vT , which is easily computed
and independent of the scalar. Notice that
hλT + ρ, α∨ i − r = hλ + ρ, α∨ i − r + T hρ, α∨ i.
Evaluating vT for this term gives 0 unless hλ+ρ, α∨ i = r (and thus α ∈ Φ+
λ ),
in which case the value is 1. Thus the
contribution to the character sum
for fixed ν and α is P ν − hλ + ρ, α∨ iα e(λ − ν). To complete the proof,
we calculate as follows using formal sums, interchanging summation over
α ∈ Φ+ λ and ν ∈ Γ as needed:
X XX
ch M (λ)i = P ν − hλ + ρ, α∨ iα e(λ − ν)

i>0 ν α
XX
P(ν) e λ − hλ + ρ, α∨ iα − ν .


=
α ν
In the second equation we change variable while staying in Γ. Since sα · λ =
λ − hλ + ρ, α∨ iα, the sum on the right for fixed α is just ch M (sα · λ). This
completes the proof of the sum formula, modulo (1). 
Remark. We defined λT to be λ + T ρ, but of course ρ disappears when T is
specialized to 0. The evaluation of vT only requires the fact that hρ, α∨ i =
6 0.
One might ask what happens if another weight is substituted for ρ. Jantzen
comments briefly on this in [148, p. 149]; but only the parallel geometric
approach by Beilinson–Bernstein [18] has led to a definitive answer: see 8.13
below.

5.8. Determinant Formula

Recall from 3.15 the universal bilinear form C on U (g) with values in U (h).
If ν ∈ Γ, write Cν for its restriction to the finite dimensional space U (n− )−ν .
Relative to a chosen ordered basis for this space, the resulting Shapovalov
matrix Sν is symmetric and has determinant Dν (determined only up to a
nonzero scalar multiple). By specializing to a simple highest weight module
5.9. Details of Shapovalov’s Proof 103

such as M (−ρ), one sees easily that Dν 6= 0 (Exercise 3.15); but this also
follows from the main theorem. The proof, given in the following section,
depends essentially on Theorem 4.12.
Theorem. Fix ν ∈ Γ. Then the determinant of the Shapovalov matrix Sν
is nonzero and is given up to a nonzero scalar factor by
Y Y
P(ν−rα)
Dν = hα + hρ, α∨ i − r .
α>0 r>0

Thus Dν is factored completely into linear factors involving the various

hα with α > 0. The degree of the indicated product as a polynomial in the
hα is clearly
XX
(∗) P(ν − rα).
α r
Here is a simple illustration:
Exercise. Use the calculations in Exercise 4.12, with g = sl(3, C) and ν =
α+β, to write down the matrix of Cν relative to the ordered basis {yα yβ , yγ }
of U (n− )−ν (recalling that hγ = hα + hβ ):
 
h α hβ + hβ −hβ
.
−hβ h α + hβ
Check that det Sν = hα hβ (hα + hβ + 1), in agreement with the theorem.
Remark. In the proof of the theorem it is convenient to look at the degree
of Dν when viewed as a polynomial in m variables hα (α > 0). This in turn
determines the degree in terms of h1 , . . . , h` : Recall that hα for arbitrary
α > 0 corresponds to the coroot α∨ , i.e., β(hα ) = hβ, α∨ i for all β ∈ Φ (or
for any λ ∈ h∗ ). Thus rewriting hα as a linear combination of h1 , . . . , h` is
the same as writing α∨ in terms of the αi∨ for i = 1, . . . , `. Fortunately we
do not need to make this explicit for the degree calculation.

5.9. Details of Shapovalov’s Proof

Fix a standard basis y1 , . . . , ym , h1 , . . . , h` , x1 , . . . , xm of g, taking α1 , . . . , αm
as an enumeration of Φ+ with α1 , . . . , α` simple and h1 , . . . , h` the corre-
sponding basis of h. The basic strategy involves two steps: (1) Investigate
the degree of Dν as a polynomial in the hα with α > 0 to see that 5.8(∗)
is an upper bound. This much is elementary. (2) Show that each of the
(relatively prime!) linear polynomials hα + hρ, α∨ i − r actually divides Dν
with multiplicity P(ν − rα). This will be done by invoking the properties
of Shapovalov elements (4.12); these in turn depend on Theorem 4.6, giv-
ing embeddings M (sα · λ) ,→ M (λ) for sufficiently many specializations to
individual weights λ.
104 5. Highest Weight Modules II

(1) To compute explicitly the degree of Dν we work with an (arbitrarily

ordered) PBW basis of U (n− )−ν having N := P(ν) elements. Each m-
tuple ω = (r1 , . . . , rm ) ofP
nonnegative integers parametrizes a basis element
yω := y1r1 . . . ym
rm , where
P
ri αi = ν. Write d(ω) := i ri for the degree of
yω as a polynomial in y1 , . . . , ym . A typical entry of the N × N matrix Sν
of Cν is then c(ω, ω 0 ) := Cν (yω , yω0 ). Now c(ω, ω 0 ) is just the projection to
U (h) of the product
τ (yω )yω0 = xrmm . . . xr11 y1s1 . . . ym
sm
,
where ω = (r1 , . . . , rm ) and ω 0 = (s1 , . . . , sm ). So the use of commutator
relations in U (g) to rewrite it in standard PBW order would be a nontrivial
task. But we only need to isolate the degree of the pure U (h)-summand,
viewed as a polynomial in the hα (α > 0). This requires some elemen-
tary inductive arguments using the commutator relations in U (g). (Sample
computations as in Exercise 4.12 may help the reader to follow the steps.)
Lemma. Let ω be arbitrary.
(a) For any fixed k, we have xk yω ≡ ni=1 pi (ω)yωi modulo the left ideal
P
I := U (g)n, where pi (ω) is a polynomial in the hα of degree ≤ 1.
(b) deg c(ω, ω 0 ) ≤ min{d(ω), d(ω 0 )} for all ω 0 .

Proof. (a) Use induction on d(ω). When this is 0, there is nothing to prove
since xk ∈ I. If d(ω) > 0, let rj be the first nonzero entry in ω and define ω 0
to have the same entries ri as ω when i 6= j but rj − 1 in the jth position.
Thus d(ω 0 ) < d(ω). Now for the given index k compute
xk yω = xk yj yω0 = yj xk yω0 + [xk yj ]yω0 .
Apply the induction assumption to the first summand to get (modulo I) a
sum of terms of the form yj p(ω 0 )yω00 . Bringing yj past such a polynomial
just replaces it by another polynomial of degree ≤ 1, after which further
commutation in U (n− ) re-organizes the term as a combination of standard
PBW monomials (modulo I).
It remains to analyze the second summand, assuming [xk yj ] 6= 0. There
are three easy cases. If k = j, we get hαk yω0 , which is of the required form.
If αk − αj < 0, we get a product of terms in U (n− ). If αk − αj > 0, we
instead have to use the induction assumption for ω 0 .
(b) Since Sν is symmetric, it is enough to prove that deg c(ω, ω 0 ) ≤ d(ω).
Use induction on d(ω). When this is 0, clearly c(ω, ω 0 ) = 1, so its degree is
also 0.
Assuming d(ω) > 0, let ri be the first nonzero entry in ω and let ω ◦ be
an m-tuple agreeing with ω except in the ith coordinate, which is ri − 1;
thus d(ω ◦ ) < d(ω). Next rewrite τ (yω )yω0 as τ (yω◦ )xi yω0 . Apply part (a) to
5.9. Details of Shapovalov’s Proof 105

xi yω0 to get (modulo I) a linear combination of various yω00 multiplied on the

left by polynomials of degrees at most 1 in the hα . Since τ fixes h pointwise
and I goes to 0 under projection to U (h), the induction assumption can be
applied to ω ◦ . 

To estimate the degree of Dν , we now have to look at the degrees of the

N ! products of matrix entries c(ω, ω 0 ), one from each row and each column,
which enter into the standard
P determinant formula. In view of the lemma,
deg Dν , cannot exceed ω d(ω). For a fixed α = αi , the integer r = ri
occurs as the ith component in ω for precisely P(ν − rα) − P(ν − (r + 1)α)
different ω. So the power of hα occurring in the highest degree term of Dν
is at most
X
X
r P(ν − rα) − P(ν − (r + 1)α) = P(ν − rα).
r>0 r>0

Summed over α > 0, this gives the predicted degree 5.8(∗).

(The argument below will show that the estimate just made is exact.
Indeed, Shapovalov states a more precise version of (b) above, which the
reader might try as an exercise: strict inequality holds unless ω = ω 0 .)
(2) It remains to prove that (for fixed α > 0 and r > 0) each factor
hα + ρ(hα ) − r divides Dν at least P(ν − rα) times. Then the product of all
these relatively prime factors as α and r vary has degree 5.8(∗) and must
therefore differ from Dν only by a scalar thanks to the estimate in (1).
Here is where the advantage of working with Shapovalov elements θα,r in
U (b− ) becomes apparent. The idea is to replace the matrix Sν by a matrix
with entries in U (b− )−ν having the same determinant, then apply Theorem
4.12. We work with θα,r for fixed α > 0 (in place of γ there) and r > 0.
Write θα,r for the highest degree term `i=1 yirai in the definition of θα,r
Q
P`
(where α = i=1 ai αi ).
Starting with the space U (n− )−ν+rα , multiply it on the right by θα,r to
get a subspace Vα,r of U (n− )−ν . Since right multiplication in U (g) is always
injective, dim Vα,r = dim U (n− )−ν+rα = P(ν − rα). Note that Vα,r involves
all PBW monomials of highest degree. Now let T be the span of a set of
monomials for which U (n− )−ν = T ⊕ Vα,r .
Next carry out a parallel construction in U (b− ) by using right multi-
0
plication by θα,r in place of θα,r . Let Vα,r = U (n− )−ν+rα θα,r . Then set
V := T ⊕ Vα,r ⊂ U (b )−ν . Clearly dim V = dim U (n− )−ν = P(ν).
0 −

From the construction of V we deduce first that the natural linear map
V ⊗ U (h) → U (b− )−ν is an isomorphism. To compare V with U (n− )−ν ,
use an ordering of PBW monomials compatible with degrees and combine
0 . The description of θ
bases of T and Vα,r α,r in Theorem 4.12(b) shows that
106 5. Highest Weight Modules II

there exists a unipotent matrix U over U (h) so that the restriction CV of

the form C to V has a matrix SV related to Sν by SV = U t Sν U . So their
determinants are the same (up to a scalar).
Look at a column of SV indexed by a basis element from Vα,r 0 . Here

an entry has the form CV (yω , yθα,r ) = the projection to U (h) of τ (yω )yθα,r .
From Theorem 4.12(a) we deduce that this projection is always divisible by
hα + ρ(hα ) − r. There are precisely dim U (n− )−ν+rα = P(ν − rα) of these
columns. Thus Dν is divisible by (hα + ρ(hα ) − r)P(ν−rα) as required. 

Notes
The original proof of the BGG Theorem (5.1) occurs in Bernstein–Gelfand–
Gelfand [25, 26] and is explained carefully by Dixmier [84, Chap. 7]. Note
that the BGG convention about the Bruhat ordering is the reverse of ours:
1 is their unique maximal element. Our convention is more natural when
an antidominant rather than dominant weight is used as a starting point.
In [124], van den Hombergh provides a somewhat streamlined proof of the
theorem in the special case of integral weights.
(5.3–5.7) We follow mainly Jantzen [148, Chap. 5].
The determinant formula in (5.8) was discovered independently by Shapo-
valov [231] and Jantzen [145, II.1] while they were doing dissertation re-
search with different goals. Jantzen was mainly interested in representations
of algebraic groups in prime characteristic p. His formulations start over Z
using a Chevalley basis for g and resulting Kostant Z-form of U (g), followed
by reduction modulo p. (An added remark [145, p. 124] calls attention to
Shapovalov’s just published work.) Later Jantzen [147] generalized the de-
terminant calculation to the parabolic case. We follow mainly [231] here,
but with some steps drawn from [145].
Ostapenko [225] studies the inverse of the Shapovalov matrix, showing
that the poles involved are all simple. This is recovered in the setting of
Virasoro algebras by Brown [48].
 Chapter 6

Extensions and
Resolutions

Most of the features of category O explored in the previous chapters orig-

inated before 1980, though some of the proofs have since been improved.
The central problem at that time was to determine the formal characters
(or composition factor multiplicities) of Verma modules. The solution of
this problem in 1979–80 will be discussed in Chapter 8, but first we pause
to consider some broader homological aspects of category O which illumi-
nate the classical results on finite dimensional modules and also point to
further problems involving higher Ext groups, projective resolutions, etc.
In 6.1–6.8 we develop a “BGG resolution” of a finite dimensional module
L(λ); its terms are direct sums of Verma modules. This realizes the Weyl–
Kostant character formula as an Euler characteristic, while recovering in a
transparent way Bott’s classical theorem on n-cohomology. Note that step
(E) in the proof of Theorem 6.2 is simplified by a reference to Chapter 7,
where translation functors will be developed in a self-contained way. But
we also explain how to do step (E) directly.
Following a broader look at homological dimension in O (6.9), the re-
mainder of the chapter treats a range of issues involving higher Ext functors
(6.10). These include vanishing criteria involving mixtures of simple mod-
ules, Verma modules, and dual Verma modules (6.11–6.12); an Ext criterion
for a module to have a standard filtration (6.13); expressions for formal char-
acters in terms of Ext (6.14); and a comparison with Lie algebra cohomology
(6.15).

107
108 6. Extensions and Resolutions

A word about the notational conventions in this chapter: To emphasize

the treatment of finite dimensional modules in the first half, we usually
work with a weight λ ∈ Λ+ (which is regular for the dot-action of W ) and
its linkage class. This is sometimes awkward, since w0 · λ < w · λ translates
into w < w0 in the Bruhat ordering. As a result, we modify the set-up locally
in 6.7–6.8.

6.1. BGG Resolution of a Finite Dimensional Module

For any module M ∈ O, it makes sense to look for a projective resolution
and then to discuss projective dimension, etc. Resolutions involving the in-
termediate Verma modules might however be more accessible, as we discuss
here in a special case.
Fix λ ∈ Λ+ . Our objective is to find a resolution of the finite dimensional
module L(λ) whose Euler character “realizes” the alternating sum formula
ch L(λ) = w∈W (−1)`(w) ch M (w · λ). In other words, we seek a finite exact
P
sequence of the following form:

m δ 2 δ
1 δ ε
(∗) 0 → M (w◦ ·λ) = Cm → Cm−1 → · · · → C1 → C0 = M (λ) → L(λ) → 0.

Here m = `(w◦ ) = |Φ+ | (which is also the dimension of n, n− , and g/b),

while ε : M (λ) → L(λ) is the canonical epimorphism. The kth term is
defined by
M
Ck := M (w · λ), where W (k) := {w ∈ W | `(w) = k}.
w∈W (k)

We call any resolution of type (∗) a BGG resolution of L(λ). Our first
priority is to find such a resolution; a further natural question is to what
extent it is uniquely determined.

Exercise. In any BGG resolution of L(λ), the map δk must be nonzero on

each summand M (w · λ) with `(w) = k.

As we observed at the end of 2.6, it is easy to get started in the direction

of constructing (∗) by taking advantage of the description there of the unique
maximal submodule N (λ) ⊂ M (λ) as the sum of embedded Verma modules
M (sα · λ) with α simple (equivalently, `(sα ) = 1). Since the Hom spaces
between Verma modules have dimension at most 1 (4.2), the choice of further
mappings entering into a resolution of type (∗) is limited. But there is no
easy way to make choices which yield an exact sequence. So we start with
a less direct construction.
6.2. Weak BGG Resolution 109

6.2. Weak BGG Resolution

Here we outline (following [26]) a construction of a “weak” resolution whose
terms are not yet obviously direct sums of the required Verma modules.
This involves a number of independent steps, for which details will be given
in the following two sections. The goal is easily formulated:

Theorem. Let λ ∈ Λ+ . Then there is an exact sequence

λ λ
0 → M (w◦ · λ) = Dm → Dm−1 → . . . → D1λ → D0λ = M (λ) → L(λ) → 0,
where Dkλ has a standard filtration involving exactly once each of the Verma
modules M (w · λ) with `(w) = k.

Rather than work immediately with an arbitrary λ ∈ Λ+ , it is easier

to focus initially on the special case λ = 0. The construction can then be
readily extended. Here are the steps in the proof of the theorem:
(A) Start with the m-dimensional vector space g/b (which is a b-module
via the adjoint action) and the associated b-modules k (g/b), with 0 ≤
V
k ≤ m. The quotient g/b is isomorphic as h-module to n− , with a basis
consisting of the cosets of a standard basis y1 , . . . , ym of n− . So its weights
are the negative roots. In turn, the weights of h on k (g/b) are the sums of
V
k distinct negative roots: a basis here consists of the cosets of all yi1 ∧· · ·∧yik
with 1 ≤ i1 < · · · < ik ≤ m. (Among these weights will be the w ·0 = wρ−ρ,
with `(w) = k, as discussed below in 6.4.)
(B) Now the idea is to generalize the original construction of a Verma
module by forming the induced modules Dk := U (g) ⊗U (b) k (g/b). Each of
V

these has a standard filtration, as indicated in Example 3.6. SinceV 0 (g/b) is

V
m
the trivial b-module C, we get D0 = M (0). At the other extreme,
P (g/b) is
also 1-dimensional, but the b-action involves the weight − α>0 α = −2ρ =
w◦ · 0; so we have Dm = M (w◦ · 0). The intermediate Dk are much more
complicated.
(C) Next introduce homomorphisms of U (g)-modules ∂k : Dk → Dk−1
for k = 1, . . . , m, together with the natural map ε : D0 = M (0) → L(0). The
maps ∂k are relative versions of the maps used in the standard resolution to
compute the cohomology of a Lie algebra and will be recalled in 6.3 below.
A direct (but slightly tricky) computation shows that one gets a complex:
∂k ∂k+1 = 0 for k = 1, . . . , m − 1, while ε∂1 = 0. A less direct argument then
verifies the exactness of the sequence.
(D) The resolution of L(0) constructed in this way still has terms Dk
which are usually much too large. The exactness is preserved by cutting
down to the principal block component of each term (written Dk0 here in
place of the usual Dkχ with χ = χ0 ). In particular, Dk0 must involve in any
110 6. Extensions and Resolutions

standard filtration at least one copy of each M (w ·0). It still has to be shown
that no more than one copy can occur: see 6.4 below.
(E) Having arrived at a suitable resolution of the trivial module L(0), we
can modify the terms to get the required resolution of L(λ). In the language
of Chapter 7 this amounts to applying the translation functor T0λ to the
sequence, which remains exact since the functor is exact (7.1). Obviously
T0λ L(0) ∼
= L(λ), while Theorem 7.6 shows that T0λ M (w · 0) ∼ = M (w · λ).
Remark. A more direct argument can be given for (E) in this special case,
avoiding the heavier details required to prove Theorem 7.6: Start by tensor-
ing the resolution of L(0) with L(λ). The exactness of this functor insures
that the resulting sequence is a resolution of L(λ). Next extract the block
component for χ = χλ , again producing a resolution of L(λ). As in the

χ case
λ = 0, it has to be shown that the resulting terms Dkλ := Dk0 ⊗ L(λ) with
χ = χλ are not “too large”.
This requires a closer look at the Verma modules involved in a standard
filtration of each Dk0 ⊗L(λ) (see 3.6): their highest weights are the w·0+µ for
`(w) = k and µ a weight of L(λ) (counted with multiplicity). For example,
µ = wλ gives w · 0 + wλ = w · λ. For fixed w, we claim that only this choice
of µ gives a Verma module of highest weight linked to w · λ (equivalently,
linked to λ). Say w0 ·(w ·0+µ) = λ. This simplifies to w0 wρ−ρ+w0 µ = λ, or
w0 wρ + w0 µ = ρ + λ. Since both λ and ρ lie in Λ+ , the respective weights on
the left occur in L(ρ) and L(λ) and are therefore ≤ ρ, λ respectively. This
forces equality throughout, so w0 = w−1 and µ = wλ.

6.3. Exactness of the Sequence

The sequence of modules Dk in 6.2 is a relative version of the standard
resolution of the trivial module in Lie algebra cohomology, which is treated
for example in Cartan–Eilenberg [59, XIII, §7], Hilton–Stammbach [123,
VII.4], Knapp [189, IV.3], Weibel [259, 7.7]. We assume here that the
reader is already acquainted with this theory.
The standard resolution involves free U (g)-modules U (g) ⊗C k (g). In
V
our setting the modules Dk are free only over U (n− ), but the map ∂k : Dk →
Dk−1 resembles the standard one: for u ∈ U (g) and arbitrary representatives
z1 , . . . , zk ∈ g of cosets ξ1 , . . . , ξk in g/b, define
k
X
∂k (u ⊗ ξ1 ∧ · · · ∧ ξk ) := (−1)i+1 (uzi ⊗ ξ1 ∧ · · · ∧ ξbi ∧ · · · ∧ ξk ) +
i=1
X
+ (−1)i+j (u ⊗ [zi zj ] ∧ ξ1 ∧ · · · ∧ ξbi ∧ · · · ∧ ξbj ∧ · · · ∧ ξk ).
1≤i<j≤k
6.4. Weights of the Exterior Powers 111

Here the hat indicates as usual that the term is omitted, while [zi zj ] denotes
the coset in g/b containing [zi zj ].
Some routine bookkeeping is required to check that this formula is inde-
pendent of the choice of coset representatives z1 , . . . , zk . Since Dk involves
a tensor product over U (b), when a fixed zi is replaced by zi + b with b ∈ g,
one can rewrite the extra term involving ub in the first sum as
(−1)i+1 u ⊗ b · (ξ1 ∧ · · · ∧ ξbi ∧ · · · ∧ ξk ) .

Then use the fact that b acts on the wedge product as a derivation, while
b · ξj is by definition the coset [bzj ] for each j 6= i. Comparison with the
second sum in the definition of ∂k then allows pairwise cancellation of all
extra terms involving b.
Next one has to verify that the maps ∂k and ε : D0 → L(0) (where D0 =
M (0)) produce at least a complex. In principle this just requires a direct
computation, but working out all the case distinctions for the inequalities
among indices gets complicated. The proof can be organized more efficiently,
as shown in the cited literature. But at any rate no really new ideas are
needed for the relative version.
Why is the complex exact? As usual, this requires a less direct approach.
Here again the essential ideas are found in the classical work, but modified
slightly. As in the usual proof of the PBW Theorem, the strategy is to
construct a filtration and then pass to the associated graded structures. In
U = U (g) the usual ascending filtration by degree comes from the natural
filtration of the tensor algebra of g, so U (l) is the span of the PBW basis
elements (for any basis of g) whose degree is ≤ l. The associated graded
algebra is just the symmetric algebra S(g).
The details of this argument are provided by BGG [26, §9], who point
out that everything goes through for a pair consisting of an arbitrary Lie
algebra and subalgebra. They also point out (in their Proposition 9.2) that
the picture is simpler if the subalgebra in question has a vector space com-
plement which is also a subalgebra. This is true in our case, with g = n− ⊕b.
Then one gets a natural identification of the complex defined above for g/b
with the standard resolution for n− : when u, zi ∈ U (n− ), map u⊗z1 ∧· · ·∧zk
in the kth term of the standard complex to u ⊗ z1 ∧ · · · ∧ zk and apply the
PBW Theorem. Exactness is just a question about images and kernels of
the linear maps involved, so no new arguments are actually needed for the
relative version.

6.4. Weights of the Exterior Powers

As stated earlier, it is easy to specify all weights occurring in k (g/b): these
V
are the sums −(α1 +· · ·+αk ) with α1 , . . . , αk distinct positive roots. Among
112 6. Extensions and Resolutions

these are the special ones w · 0 for which w ∈ W (k) . Indeed, each w of length
k is uniquely characterized by the set of k positive roots which it sends to
negative roots (0.3); the sum of the negative roots sent by w−1 to Φ+ is
wρ − ρ = w · 0. For example, the unique element of length m is w◦ , which
sends Φ+ to Φ− ; here w◦ · 0 = w◦ ρ − ρ = −2ρ, which is the sum over Φ− .
Example. When g = sl(3, C), call the simple roots α and β. The weights
in k are listed in the middle column, while the weights w · 0 with w ∈ W (k)
V
appear in the last column.
0 0 0
1 −α, −β, −α − β −α, −β
2 −2α − β, −α − 2β, −α − β −2α − β, −α − 2β
3 −2α − 2β −2α − 2β

Returning to the general case, it remains

V to check that weights of the
form w · 0 occur with multiplicity 1 in (g/b). For this we have to use the
standard facts about the length function on W summarized in 0.3. When
w ∈ W , write Πw := Φ+ ∩ wΦ− . Thus `(w) = |Πw |. For any Π ⊂ Φ+ ,
denote by Π the sum of the roots in Π. In particular, Πw = −w · 0 = ρ − wρ.
Now we just need to prove:
Lemma. Using the above notation, suppose that Π ⊂ Φ+ satisfies Π =
ρ − wρ = Πw for some w ∈ W . Then Π = Πw .

Proof. Use induction on `(w), the case w = 1 being obvious since Πw = ∅

and Πw = 0. When `(w) = k > 0, there exists α ∈ ∆ with `(sα w) = k − 1.
In particular, w−1 α < 0, so by definition α ∈ Πw . (This leads us to expect
that α ∈ Π, to be shown below.) Setting w0 := sα w, we get (w0 )−1 α > 0
and thus α 6∈ Πw0 .
For use in the final step of the proof, we observe that
(1) Πw = sα Πw0 ∪ {α}.
This depends mainly on the fact that sα permutes the positive roots other
than α:
sα Πw0 = sα Φ+ ∩ (sα w0 )Φ− = (Φ+ \ {α}) ∪ {−α} ∩ wΦ− .

Since α ∈ Πw ⊂ wΦ− , we can discard α and get

sα Πw0 = sα (Φ+ \ {α}) ∩ wΦ− = (Φ+ \ {α}) ∩ wΦ− ,
from which (1) follows.
By assumption, we are given Π ⊂ Φ+ with Π = Πw = ρ − wρ. Thus
(2) sα Π = sα Π = (ρ − α) − sα wρ = (ρ − w0 ρ) − α.
6.5. Extensions of Verma Modules 113

Now we claim that α ∈ Π. Otherwise sα Π ⊂ Φ+ (since sα permutes the

positive roots other than α), so in turn sα Π ∪ {α} ⊂ Φ+ . Using (2), this
implies that
sα Π ∪ {α} = ρ − w0 ρ = Πw0 .
Induction would force sα Π∪{α} = Πw0 , leading to the contradiction α ∈ Πw0 .
With these preliminaries out of the way, we can deal with the induction
step. Set Π0 := sα (Π \ {α}). From (2) we get
Π0 = ρ − w 0 ρ = Πw 0 ,
which by induction forces Π0 = Πw0 . This in turn implies Π \ {α} = sα Π0 =
sα Πw0 , or Π = sα Πw0 ∪ {α}, which by (1) is Πw as required. 

6.5. Extensions of Verma Modules

Having produced a resolution of L(λ) involving terms Dkλ which have stan-
dard filtrations, we next claim that each Dkλ is actually a direct sum of
Verma modules. This amounts to proving that certain Ext groups vanish.
(Recall from 3.1 the basic facts about these groups.)
In the case of Verma modules and extensions in O, splitting of extensions
can be understood in terms of the strong linkage relation on their highest
weights (which resembles 5.1(∗) for the Hom functor). In our situation this
translates also into the Bruhat ordering of W .
Theorem. Let λ ∈ h∗ .
(a) If ExtO M (µ), M (λ) 6= 0 for µ ∈ h∗ , then µ ↑ λ but µ 6= λ.

(b) Let λ ∈ Λ+ and w, w0 ∈ W . If ExtO M (w0 · λ), M (w · λ) 6= 0,

then w < w0 in the Bruhat ordering. In particular, `(w) < `(w0 ).

Proof. (a) Proposition 3.1(a) insures that µ 6= λ. By assumption, there is

a nonsplit short exact sequence
(1) 0 → M (λ) → M → M (µ) → 0.
Since P (µ) is a projective cover of M (µ) (3.9), the canonical map P (µ) →
M (µ) lifts to a map ϕ : P (µ) → M . The image of ϕ must intersect the
submodule M (λ) nontrivially: otherwise the sequence (1) splits. Now recall
from 3.10 that P (µ) has a standard filtration 0 = P0 ⊂ P1 ⊂ · · · ⊂ Pn =
P (µ) with Pi /Pi−1 ∼
= M (µi ) for some µi when i > 0 (and µn = µ). Thanks
to BGG Reciprocity (3.11) and the BGG Theorem 5.1, µ ↑ µi for all i ≤ n.
Since the image of ϕ intersects M (λ) nontrivially, there is a least i for
which ϕ(Pi ) ∩ M (λ) 6= 0. This implies that M (λ) has a nonzero submodule
which is a homomorphic image of M (µi ); therefore [M (λ) : L(µi )] 6= 0.
114 6. Extensions and Resolutions

Applying the BGG Theorem again, we get µi ↑ λ. This combines with

µ ↑ µi to complete the proof of (a).
(b) Under these assumptions, part (a) shows that w0 · λ ↑ w · λ (and the
weights are unequal). Here all the linked weights in question are regular and
integral, so the strong linkage relation translates into the Bruhat ordering
of W as in 5.2. Thus w < w0 . 

The special case allows us to conclude (following Rocha [226]) that the
“weak” BGG resolution is actually a resolution of type 6.1(∗). Indeed, an
easy induction on the number of Verma subquotients shows that each Dkλ
must split into a direct sum of Verma modules: nonsplit extensions as in
the theorem require w < w0 and thus `(w) < `(w0 ).

Corollary. If λ ∈ Λ+ , the resolution of L(λ) in 6.2 is a BGG resolution. 

Remark. One might ask which infinite dimensional simple modules L(λ)
have analogous resolutions. For example, in the extreme case when λ is
antidominant, we know that L(λ) = M (λ) (Theorem 4.8); so 0 → M (λ) →
L(λ) → 0 already realizes the character formula. But in general there are se-
rious obstacles, starting with the fact that the maximal submodule of M (λ)
need not be a sum of embedded Verma submodules (unless all composition
factor multiplicities are 1: Exercise 5.1). This is seen in the example at the
end of BGG [25] in case g = sl(4, C) and λ = −$1 − $3 in the usual num-
bering of fundamental weights. Notice in this case that the integral weight
λ is dominant (for the dot-action of W ) but not regular.
Gabber–Joseph [106, 2.6] proved the existence of BGG resolutions under
fairly general hypotheses. To state their result using our notation, fix a
regular antidominant weight λ ∈ h∗ , with associated root system Φ[λ] , simple
system ∆[λ] , and Weyl group W[λ] . Let I ⊂ ∆[λ] and let t be the length of
the longest element wI of WI (relative to ∆[λ] ). Then the simple module
L(wI · λ) has a resolution of the form

0 → C0 → C1 → . . . → Ct → L(wI · λ) → 0,

where Ck is the direct sum of all M (w · λ) with w ∈ WI of length k. In

particular, the maximal submodule of M (wI · λ) is a sum of Verma modules.
When λ is integral, this recovers the finite dimensional case w◦ · λ as well as
the antidominant case. (See also Jantzen’s discussion of the Gabber–Joseph
result in [149, 4.18, 4A.6].)
In spite of these positive results, there seems to be no reasonable nec-
essary and sufficient condition on λ insuring that L(λ), λ ∈ h∗ has a finite
resolution by direct sums of Verma modules.
6.6. Application: Bott’s Theorem 115

6.6. Application: Bott’s Theorem

At the end of his influential 1957 paper on the cohomology of vector bundles
on homogeneous spaces such as flag varieties of semisimple Lie groups, Bott
[43, §15] obtained what he described as a “curious corollary”: an explicit
formula in terms of W for the dimensions of certain Lie algebra cohomology
groups. (He also remarked that this formula can be shown “without much
trouble” to imply Weyl’s character formula.) Although he worked in the set-
ting of compact Lie groups and their complexifications, the essential point
of the corollary is to describe the cohomology of a maximal nilpotent subal-
gebra of g (conjugate under the adjoint group to n or n− ) with coefficients
in a finite dimensional simple module L(λ). (Kostant [197] later developed
these ideas further in the setting of Lie algebra cohomology.)
Bott’s formula was recovered by BGG as an easy consequence of their
weak resolution of L(λ) in [26, §9], by using some standard homological
algebra for the Ext and Tor functors. We follow Jantzen’s version of the
argument [154, 1.7]. To exploit the BGG resolution, it is convenient to
state Bott’s formula as follows:
Theorem. If λ ∈ Λ+ , then dim H k n− , L(λ) = |W (k) |, where as before

W (k) denotes the set of elements in W having length k.

Proof. By definition, H k n− , L(λ) = Extkn− C, L(λ) , where C is has triv-

ial n− -module action. Note that this vector space has an h-module structure,
since h normalizes n− . Dualizing, we obtain an isomorphism of h-modules
Extk− C, L(λ) ∼ = Extk− L(λ)∗ , C .



n n

Here L(λ)∗ ∼
= L(λ∗ ), where λ∗ = −w◦ λ. To compute the right side, we ob-
serve that the BGG resolution provides a convenient resolution of L(λ∗ ) by
free U (n− )-modules: the direct sums of Verma modules M (w · λ∗ ). Now
Extkn− (L(λ∗ ), C) is the cohomology of the complex Homn− (M • , C) of h-
modules for which M k := ∗ −
L
w∈W (k) M (w · λ ). For any n -module M
− ∗
we can naturally identify Homn− (M, C) with (M/n M ) . In particular,
when M = M (µ) we have M/n− M ∼ = Cµ as h-modules, while the dual be-
comes C−µ . Applied to the direct sum M k , thisL shows that the kth term in
the complex Homn− (M • , C) is isomorphic to w∈W (k) C−w·λ∗ . Because the
w · λ∗ are pairwise distinct, all maps in the complex are zero, this computes
Extkn− (L(λ∗ ), C) and thus proves Bott’s dimension formula (with additional
information about the h-module structure of the cohomology).
The end result of this computation can be restated more neatly in terms
of the cohomology of n, by doing a little bookkeeping. If we change the
system of positive roots so that n corresponds to negative roots, we have to
replace λ by the new highest weight w◦ λ of L(λ) and also replace ρ by −ρ in
116 6. Extensions and Resolutions

order to apply the argument Labove. Now we get an h-module isomorphism

between H k n, L(λ) and . Since (w◦ λ)∗ = −λ,


w∈W (k) C
−w((w ∗
◦ λ) −ρ)−ρ
this expression simplifies: H k n, L(λ) ∼
L
= w∈W (k) Cw·λ . 

Exercise. Show that the freeness of each term in the BGG resolution as a
U (n− )-module does not require the splitting result in 6.5 above, since only
the action of n− is in question.

Remark. There are other alternating sum character formulas in represen-

tation theory besides Weyl’s character formula. For example, classical work
of Frobenius on the characters of finite symmetric groups expresses an ir-
reducible character as an alternating sum of characters of certain easily
constructed induced characters (from subgroups which are direct products
of smaller symmetric groups). Zelevinsky [261] exploited BGG resolutions
for the reductive Lie algebra gl(n, C) (whose representation theory is very
close to that of sl(n, C)) to realize this formula. Here Schur–Weyl duality
provides a natural transition between the two settings. (Independent work
along this line was done by Akin [2] from a somewhat different viewpoint.)
Zelevinsky also treats other alternating sum formulas arising in Lie theory.

6.7. Squares
Although we have shown the existence of at least one BGG resolution of the
finite dimensional module L(λ), we have avoided an explicit description of
the maps involved. But more precise information is needed for deciding to
what extent such a resolution might be unique. In this and the next section
we explore this question. (It was of more critical interest in [26]: there it
was not yet known that the “weak” resolution in 6.2 above is actually of the
“strong” type 6.1(∗), so a more direct construction was needed.)
L
Suppose we are given a BGG resolution with Ck = `(w)=k M (w · λ)
and maps δk : Ck → Ck−1 along with ε : M (λ) → L(λ). The map δ1 :
C1 → C0 = M (λ) was already described in 2.6: here each Verma module
M (sα · λ) with α ∈ ∆ embeds naturally in M (λ). Beyond this the details
are less obvious.
To analyze maps between Verma modules, we employ the Bruhat or-
dering in the spirit of 5.1. As noted in 0.4, we may limit the use of the
arrow w → w0 here to cases when `(w0 ) = `(w) + 1. Now the Verma module
maps are easiest to visualize as arrows when we use the antidominant weight
λ◦ := w◦ · λ as the starting point. Accordingly, we depart from the BGG
convention in [26] by rewriting the BGG resolution in an equivalent form:
ε εm−2 εm−1 ε
0 → M (λ◦ ) = C0◦ →
0
C1◦ → · · · → Cm−1
◦ ◦
→ Cm = M (λ) → L(λ) → 0.
6.7. Squares 117

For k = 0, . . . , m − 1, the kth term is defined by

M
Ck◦ := M (w · λ◦ ).
w∈W (k)

Consider the pairs (w, w0 ) with `(w
0 ) = `(w) + 1. In general we know

that dim HomO (M (w · λ◦ ), M w0 · λ◦ ) ≤ 1; the dimension is 1 precisely
when w < w0 (5.1–5.2). In our case this means w → w0 .
The map εk is uniquely determined by its restriction to the summands
M (w · λ◦ ) of Ck◦ , mapping M (w · λ◦ ) into various summands M (w0 · λ◦ )
◦ . Fix an embedding of each M (w · λ◦ ) into M (λ), which also fixes
of Ck+1
embeddings M (w · λ◦ ) ,→ M (w0 · λ◦ ) whenever w < w0 . (This amounts to
specifying a maximal vector of weight w · λ◦ in M (w · λ◦ ), expressed as u · v +
for some u ∈ U (n− ) when v + is a fixed maximal vector of weight λ in M (λ).)
Now maps between Verma modules are specified by scalars. In partic-
ular, the map M (w · λ◦ ) → M (w0 · λ◦ ) induced by εk is determined by a
scalar e(w, w0 ) (possibly zero!). If no map of this type exists, just define
e(w, w0 ) = 0. So the possible maps εk correspond to collections of scalars
e(w, w0 ). These scalars are not arbitrary, since εk+1 εk = 0. To analyze
this we focus on special configurations in the Bruhat ordering. If elements
wi ∈ W are related as shown in the figure below, we say they form a square.

w2

* H
 HH
j
w1 w4
HH 
*
j
H
w3 

Now we invoke a special feature of the Bruhat ordering stated in Propo-

sition 0.4(d): whenever `(w4 ) = `(w1 )+2 and some w satisfies w1 < w < w4 ,
there are precisely two elements w2 , w3 between w1 and w4 (thus completing
a square as above). This was proved inductively in [26, Lemma 10.3] (see
[84, 7.7.6]) for Weyl groups using the lemma above. (The reader might try
to do this as an exercise.) Later Verma showed that it resulted for arbitrary
Coxeter groups from the nature of the Möbius function. Another general
proof comes from Kazhdan–Lusztig theory. See [130, 8.5] for details.
This property of the ordering, combined with the condition εk+1 εk = 0,
translates immediately into the equation
(1) e(w2 , w4 ) e(w1 , w2 ) + e(w3 , w4 ) e(w1 , w3 ) = 0.
Exercise. Suppose that all scalars e(wi , wj ) in (1) lie in R. Prove that the
factors in one term have like signs (+ or −), while those in the other term
have opposite signs.
118 6. Extensions and Resolutions

For use in the following section, we prove a key lemma showing how
squares may be formed by combining two configurations:
Lemma. Let α ∈ ∆ and β > 0 with β 6= α. Given the diagram on the left,
the diagram on the right exists, and conversely.

β  w0 w0
α
* HH
sα w0
 j
H
sα w
HH *

αH
j
w w sβ
α

Proof. On the left we are given sβ sα w = w0 , or sα w = sβ w0 , while `(w0 ) =

`(w) = `(sα w)+1. We have to show that the diagram on the right is correct.
(It is left as an exercise for the reader to argue in the opposite direction.)
Set γ := sα β; this translates into sγ = sα sβ sα . So in the diagram on the
right, sγ w = sα w0 . Since `(sα w0 ) = `(w0 ) ± 1 and `(w) = `(w0 ), it suffices
to show that `(sα w0 ) > `(w0 ) (the length difference being 1). This requires
some standard facts from 0.3.
By assumption `(sα w) < `(w), or equivalently, `(sβ w0 ) < `(w0 ). From
0.3 we get (w0 )−1 β < 0. Since β = sα γ, we have (sα w0 )−1 γ = (w0 )−1 sα γ =
(w0 )−1 β < 0. This forces `(sγ sα w0 ) < `(sα w0 ). Substituting sγ = sα sβ sα ,
we get `(sγ sα w0 ) = `(sα sβ w0 ) = `(w) = `(w0 ) (using sβ w0 = sα w). Thus
`(w0 ) < `(sα w0 ), as required. 

6.8. Maps in a BGG Resolution

Now we are in a position to show that in any BGG resolution, all scalars
e(w, w0 ) which could be nonzero because w < w0 are in fact nonzero.
Theorem. Given a BGG resolution with maps εk : Ck◦ → Ck+1 ◦ , all scalars

e(w, w0 ) are nonzero when w → w0 with `(w) = k and `(w0 ) = k + 1.

Proof. First we observe (as indicated earlier in Exercise 6.1) that the re-
striction of εk to each summand M (w · λ◦ ) of Ck◦ must be nonzero. By
exactness, this is clear already for ε0 ; so assume k > 0. If M (w · λ◦ ) lies in
Ker εk , exactness forces it to be in the image of εk−1 . This is impossible:
all maps of Verma module summands of Ck−1 ◦ into M (w · λ◦ ) have proper
submodules as images, so the sum of images is also a proper submodule.
To complete the proof, use downward induction on k. The case k = m
is already clear, so let k < m. We need to apply Lemma 6.7 twice.
6.8. Maps in a BGG Resolution 119

(1) Since εk 6= 0 on at least one summand M (w · λ◦ ) of Ck◦ , there exists

β
β > 0 such that w → w0 with `(w0 ) = k + 1 and e(w, w0 ) 6= 0. On the other
α
hand, since k < m there must exist α ∈ ∆ with w → sα w, so `(sα w) > `(w).
In case β = α, we can skip to step (2) below. Otherwise we get the diagram
on the left side of the figure below, which by the lemma implies the diagram
on the right. Now the induction hypothesis insures that both e(w0 , sα w0 )

β  w0
* H α
 HHj
w sα w0
HH *

αH
j
sα w  sα β


and e(sα w, sα w0 ) are nonzero, while e(w, w0 ) was chosen to be nonzero. So

6.7(1) forces e(w, sα w) 6= 0.
γ
(2) At this point we know that e(w, sα w) 6= 0. Suppose w → w00 with
w 6= sα w and `(w00 ) = k + 1. A second application of Lemma 6.7 produces
00

another square.

γ  w00 α
* HH
 j
H
w sα w00
HH *

αH
j
sα w  sα γ


Now the proof can be completed as in step (1). 

Going back to the earlier notation for a BGG resolution, note that the
proof above only requires that the sequence 6.1(∗) be a complex, together
with the fact (based on exactness) that δk is nonzero on each Verma sum-
mand of Ck .
Remark. As indicated above, BGG developed some of these ideas in prepa-
ration for an explicit construction of a resolution of type 6.1(∗). Rocha [226,
§10–§11] instead uses the theorem above to streamline their construction
and to explain the sense in which a BGG resolution is unique. For this she
works in a suitable R-split form of g, without any significant loss of gener-
ality. In particular, the scalars e(w, w0 ) above are real, as well as nonzero
when w < w0 .
When e(w, w0 ) 6= 0, define e0 (w, w0 ) := e(w, w0 )/|e(w, w0 )| = ±1; other-
wise set e0 (w, w0 ) = 0. As pointed out in Exercise 6.7, it is clear that for any
120 6. Extensions and Resolutions

square in the Bruhat ordering, the factors in one term in 6.7(1) have like
signs while the factors in the other term have opposite signs. This carries
over immediately to the modified scalars e0 (w, w0 ), so 6.7(1) remains valid
for these scalars. This recovers a key lemma [26, Lemma 10.4]: it is possible
to assign scalars ±1 to all w → w0 in the Bruhat ordering in such a way that
this equation holds for each square.
Now use the scalars e0 (w, w0 ) with `(w) = k (some may be 0) to define
new homomorphisms ε0k : Ck◦ → Ck+1 ◦ . The analogue of 6.7(1) shows that

these maps together with ε : M (λ) → L(λ) form at least a complex. By

complicated arguments BGG proved that this complex is the sought-for
resolution of L(λ). But Rocha can prove the exactness more simply. Using
downward induction on k (with the step k = m being obvious), she shows
the existence of an automorphism σk of Ck◦ so that ε0k σk = σk+1 εk for all
k. This comparison with the known BGG resolution shows that the new
complex is exact and also clarifies the uniqueness question.

6.9. Homological Dimension

In the study of O or its blocks as abelian categories with finiteness con-
ditions, it is natural to look at projective (or injective) resolutions. By a
projective resolution of M ∈ O of length n we mean an exact sequence
0 → Pn → Pn−1 → . . . → P1 → P0 → M → 0,
where all Pi are projective and Pn 6= 0. The projective dimension pd M
of M is defined to be the smallest possible length of such a resolution,
or ∞ if there is no finite projective resolution; thus pd M = 0 just when
M is projective. In turn, define the global or homological dimension
gl. dim O ≤ ∞ to be sup{pd M } taken over all M ∈ O; define gl. dim Oλ
similarly. It is nontrivial to compute such dimensions, but our study of
highest weight modules allows us to say something definite at this point.
We shall need some standard facts, which are developed in texts on
homological algebra such as Weibel [259] or in the notes of Bass [14, I.6].
Lemma. Let 0 → A → B → C → 0 be a short exact sequence in a module
category or other abelian category admitting projective resolutions. Then:
(a) pd A ≤ max{pd B, pd C − 1}.
(b) pd B ≤ max{pd A, pd C}.
(c) pd C ≤ max{pd A + 1, pd B}.
With the lemma in hand, several inequalities can be obtained right away,
following Bernstein–Gelfand–Gelfand [27, §7]. We focus especially on the
principal block O0 (1.13), whose simple modules are of the form L(w · 0)
with w ∈ W . Set m := `(w◦ ); so `(w) ≤ m for all w ∈ W .
6.9. Homological Dimension 121

Proposition. In the principal block O0 , projective dimensions satisfy for

all w ∈ W :
(a) pd M (w · 0) ≤ `(w).
(b) pd L(w · 0) ≤ 2m − `(w).
As a result, gl. dim O0 ≤ 2m.

Proof. (a) Use induction on `(w). Since 0 ∈ Λ+ , we know that M (0) is

projective (Proposition 3.8); thus pd M (0) = 0 = `(1). For the induction
step, recall from the discussion in 3.9 that P (w · 0) is the projective cover of
M (w · 0) as well as L(w · 0). Consider the short exact sequence
0 → N → P (w · 0) → M (w · 0) → 0.
Here N has a standard filtration (3.7). Thanks to Theorem 3.10, the highest
weights of all Verma modules M (w0 · 0) involved in this filtration satisfy
w0 · 0 > w · 0. Thus `(w0 ) < `(w). Now part (c) of the lemma, applied to the
above sequence, combines with iteration of part (b) and induction:
pd M (w · 0) ≤ pd N + 1 ≤ max{pd M (w · 0) | `(w0 ) < `(w)} + 1 ≤ `(w).

(b) Now use downward induction on `(w), starting with w = w◦ . Here

Theorem 4.4 shows that L(w◦ · 0) = M (w◦ · 0). By part (a), pd L(w◦ · 0) ≤
m = 2m − `(w◦ ). For the induction step, consider the short exact sequence
0 → N (w · 0) → M (w · 0) → L(w · 0) → 0.
All composition factors L(w0 · 0) of N (w · 0) have weights below w · 0, forcing
`(w0 ) > `(w). Since pd M (w · 0) ≤ `(w) ≤ 2m − `(w), part (c) of the lemma
and induction yield
pd L(w · 0) ≤ max{pd L(w0 · 0) | `(w0 ) > `(w)} + 1 ≤ 2m − `(w).

To apply these estimates to the global dimension of O0 , use part (b) of

the lemma inductively to see that pd M for any M ∈ O0 is bounded by the
projective dimensions of its composition factors. Thanks to part (b) of the
proposition, 2m is always an upper bound for these. 

Notice that the proof of the proposition requires only a few basic facts.
By using more theory, one can see that equality actually holds in (a) and
(b); then gl. dim O0 = 2m. The idea is to build on the proposition, starting
with pd M (0) = 0 and again using induction on `(w). To bound pd M (w · 0)
below by `(w), use Corollary 5.2 of the BGG Theorem together with BGG
Reciprocity and the fact that adjacent elements in the Bruhat ordering differ
in length by 1. Then part (a) of the Lemma and the induction hypothesis
take over. A similar method yields the reverse inequality in part (b) of the
proposition. We leave the details as an exercise for the reader.
122 6. Extensions and Resolutions

Theorem. For the principal block O0 , projective and global dimensions are
given by:
(a) pd M (w · 0) = `(w).
(b) pd L(w · 0) = 2m − `(w).
(c) gl. dim O0 = 2m.

Here the weight 0 can be replaced by any λ ∈ Λ+ , which is again dot-

regular. Part (c) was stated by BGG in remarks at the end of [27]. Their
idea is to show that for
all λ ∈ Λ+ , the equality pd L(λ) = 2m holds: in fact,
dim Ext2mO L(λ), L(λ) = 1. This
results from an isomorphism between
•
the Ext algebra ExtO L(λ), L(λ) and the cohomology algebra of the flag
variety G/B associated with a semisimple group G and Borel subgroup B
whose respective Lie algebras are g and b.
Recently Mazorchuk [215] has recovered the above theorem and obtained
related results within a unified framework, relying on homological methods.
In the process he formulates dimension conjectures about some other classes
of modules in O.

6.10. Higher Ext Groups

In 3.1 we recalled the way in which right derived functors such as Ext =
Ext1 arise from the Hom functor. Although Extn usually lacks the explicit
interpretation of Ext, the study of these functors often helps to organize
better the interaction of objects in a module category. As in the case of
Ext, long exact sequences derived from short exact sequences play a basic
role.
It is a natural problem to determine Extn for various pairs of standard
modules in O such as L(λ), M (λ), M (λ)∨ . In some cases this turns out to
be quite difficult, but in other cases we can make effective use of the theory
developed so far. Recall first what is already known about Verma modules
and their duals, starting with the functor Hom = Ext0 .
• Any nonzero homomorphism

M (µ) → M (λ) is injective, while
dim HomO M (µ), M (λ) ≤ 1 (4.2).
• The precise criterion for existence of a nonzero homomorphism (em-
bedding) M (µ) → M (λ) is that µ ↑ λ (Theorem 5.1).
• For all λ, µ ∈ h∗ , dim HomO M (µ), M (λ)∨ = δλµ (Theorem 3.3(c)).

• For all λ, µ ∈ h∗ , ExtO M (µ), M (λ)∨ ) = 0 (Theorem 3.3(d)).

• For all λ, µ ∈ h∗ , ExtO M (µ), M (λ) 6= 0 only if µ ↑ λ but µ 6= λ

(Theorem 6.5).
6.11. Vanishing Criteria for Extn 123

6.11. Vanishing Criteria for Extn

It is easy, using some inductive arguments together with long exact se-
quences, to obtain vanishing criteria for Extn in the spirit of Theorem 6.5.
This requires some careful bookkeeping, but no really new ideas.
Theorem. Let λ ∈ Λ+ and w, w0 ∈ W .
(a) Unless w0 · λ ↑ w · λ with w0 6= w, we have for all n > 0:
ExtnO M (w0 · λ), M (w · λ) = 0 = ExtnO M (w0 · λ), L(w · λ) .

(b) If w0 · λ ≤ w · λ, then for all n > `(w0 ) − `(w):

ExtnO M (w0 · λ), M (w · λ) = 0 = ExtnO M (w0 · λ), L(w · λ) .

Proof. In the Bruhat ordering of W , the statement w0 ·λ ↑ w·λ is equivalent

to w ≤ w0 .
(a) Assuming that w 6< w0 , the first vanishing statement is true for n = 1
by Theorem 6.5, so we proceed by induction on n. Consider the short exact
sequence
0 → N → P (w0 · λ) → M (w0 · λ) → 0.
Thanks to Theorem 3.10, N has a standard filtration involving quotients
M (x · λ) with w0 · λ < x · λ; here x < w0 , so by assumption w 6< x. Part of
the resulting long exact sequence is
. . . → ExtnO N, M (w · λ) → Extn+1 M (w0 · λ), M (w · λ) →



O

→ Extn+1 P (w0 · λ), M (w · λ) → . . .

O
The last term is 0 since P (w0 · λ) is projective, while the first term vanishes
by the induction hypothesis coupled with a subsidiary induction on the
standard filtration length of N .
For the second vanishing statement, use downward induction on `(w),
starting with the case w = w◦ : here L(w◦ · λ) = M (w◦ · λ), permitting us to
apply the first vanishing statement for all n > 0. Now assume the result for
all w 6< w0 . Start with the short exact sequence
0 → N (w · λ) → M (w · λ) → L(w · λ) → 0.
All composition factors L(x · λ) of N (w · λ) satisfy x · λ < w · λ, or w < x;
in particular, x 6< w0 . Since `(x) > `(w), we know by induction that the
vanishing statement holds for L(x · λ). So a subsidiary
induction on the
length of N (w · λ) shows that ExtkO M (w0 · λ), N (w · λ) = 0 for all k > 0.
Finally consider part of the long exact sequence:
. . . ExtnO M (w0 · λ), M (w · λ) → ExtnO M (w0 · λ), L(w · λ) →

n+1
M (w0 · λ), N (w · λ) . . .


→ ExtO
124 6. Extensions and Resolutions

The first term is 0 by the first vanishing statement, while the third term has
just been seen to be 0, so the middle term is 0 as desired.
(b) Here the starting point is `(w) ≤ `(w0 ). The idea is to imitate the
strategy in part (a), but using induction on `(w0 ) − `(w). If this is 0, then
w0 = w and part (a) applies. We leave to the reader the task of filling in the
remaining details. 

6.12. Computation of ExtnO (M (µ), M (λ)∨ )

Here we extend Theorem 3.3(d) to all Extn , while broadening the formu-
lation to cover all modules in O having a standard filtration. In the proof
essential use is made of the fact that a projective module has a standard
filtration (Theorem 3.10).
Theorem. For all n > 0 and all λ, µ ∈ h∗ , we have ExtnO (M (µ), M (λ)∨ ) =
0. More generally, this is true when M (µ) is replaced by any module having
a standard filtration.

Proof. The idea is to use induction on n. For the induction step we actually
need the more general formulation.
(1) In case n = 1, Theorem 3.3(d) shows that ExtO (M (µ), M (λ)∨ ) = 0.
To extend the vanishing to an arbitrary M having a standard filtration,
use induction on the filtration length of M ; we have just treated the length
1 case. Because M has a standard filtration, there is a short exact sequence
0 → N → M → M (µ) → 0
for some µ ∈ h∗ , with N of smaller filtration length. Part of the resulting
long exact sequence is
. . . → Ext(M (µ), M (λ)∨ ) → ExtO (M, M (λ)∨ ) → ExtO (N, M (λ)∨ ) → . . .
We have seen that the first term vanishes, while induction shows the same
for the third term. This forces the middle term to vanish.
(2) For the induction step, assume the result for n. Given M (µ), we get
a short exact sequence
0 → N → P (µ) → M (µ) → 0,
where N has a standard filtration because P (µ) does. (See Remark 3.10.)
Part of the resulting long exact sequence is
→ ExtnO (N, M (λ)∨ ) → Extn+1 ∨ n+1 ∨
O (M (µ), M (λ) ) → ExtO (P (µ), M (λ) ) →
By induction, the first term vanishes, while the third term vanishes because
P (µ) is projective; so the middle term vanishes.
Finally, replace M (µ) by any module M having a standard filtration and
argue inductively as in the case n = 1 to complete the proof. 
6.13. Ext Criterion for Standard Filtrations 125

Exercise. If M, N have standard filtrations, prove that ExtnO (M, N ∨ ) = 0

for all n > 0. [Imitate the proof of the theorem, but dualize the short exact
sequence in each step.]

6.13. Ext Criterion for Standard Filtrations

While the generalization from M (µ) to M serves mainly a technical function
in the proof of Theorem 6.12, there is in fact a precise converse to the theo-
rem which characterizes modules with a standard filtration. These include
projective modules (a fact used in the proof below, via Theorem 6.12), as
well as the “tilting” modules introduced later in Chapter 11.
Theorem. For any M ∈ O, the following three conditions are equivalent.
(a) M has a standard filtration.
(b) ExtnO (M, M (λ)∨ ) = 0 for all n > 0 and all λ ∈ h∗ .
(c) Ext1O (M, M (λ)∨ ) = 0 for all λ ∈ h∗ .

Proof. It was already shown in Theorem 6.12 that (a) implies (b), while it
is obvious that (b) implies (c). To show that (c) implies (a), use induction
on the length of M . The length one case (with M simple) is covered by the
first three steps below, as the reader can check. We may assume M 6= 0,
since by definition the 0 module has a standard filtration.


(1) Let λ be minimal so that HomO M, L(λ) 6= 0; thus there is a
surjective map M → L(λ). The first problem is to see that this map factors
through the natural map M (λ) → L(λ), yielding a (necessarily surjective)
map M → M (λ).


(2) By the choice of λ, we have Hom O M, L(µ) = 0 for all µ < λ. We
claim that this forces ExtO M, L(µ) = 0 for all µ < λ. Using L(µ)∨ ∼
1


=
L(µ), we have a short exact sequence
0 → L(µ) → M (µ)∨ → M (µ)∨ /L(µ) → 0.
Part of the resulting long exact sequence is
. . . → HomO M, M (µ)∨ /L(µ) → Ext1O M, L(µ) → Ext1O M, M (µ)∨ → . . .

The term on the right is 0 by
(c), while the term on the left is 0 because
µ < λ forces HomO M, L(ν) = 0 for any simple submodule L(ν) of the
quotient module. So the middle term vanishes, proving our claim.
From this it follows by an easy induction on the length of M 0 that
Ext1O (M, M 0 )
= 0 whenever all composition factors L(ν) of M 0 satisfy ν < λ.
(3) Now we can prove that M → L(λ) lifts to M → M (λ) by analyzing:




HomO M, N (λ) → HomO M, M (λ) → HomO M, L(λ) →
126 6. Extensions and Resolutions

→ Ext1O M, N (λ) .

The first term is 0, since all weights in N (λ) are < λ, while the last term is
0 by step (2). The resulting isomorphism of the middle terms permits the
desired lifting.
(4) At this point we have constructed a short exact sequence
0 → N → M → M (λ) → 0.
For our induction step, we have to check that N satisfies the vanishing
condition in (c). Consider part of the long exact sequence when µ ∈ h∗ is
arbitrary:
. . . → Ext1O (M, M (µ)∨ ) → Ext1O (N, M (µ)∨ ) → Ext2O (M (λ), M (µ)∨ ) → . . .
By assumption, the first term vanishes, while by Theorem 6.12 the third
term is 0. This forces the middle term to be 0.
(5) Now the induction hypothesis, applied to N , shows that N has a
standard filtration. Then the definition shows that M also does. 

6.14. Characters in Terms of Ext•O

In his study of extensions in category O and their relationship with Lie
algebra cohomology, Delorme [75, Thm.3 (i)] works out an expression for the
formal character of an arbitrary N ∈ O in terms of higher Ext groups. This
uses some elementary homological algebra together with BGG Reciprocity
(3.11) and will play a role in Chapter 8 below. First we introduce a notation
for certain Euler characteristics:
X
χ Ext•O (M (µ), N ) := (−1)n dim ExtnO (M (µ), N ).

n≥0

The sum on the right is actually finite, since modules in O have finite pro-
jective (or injective) dimension by 6.9.
Proposition. For all N ∈ O we have
X
χ Ext•O (M (µ), N ) ch M (µ).


(∗) ch N =
µ∈h∗

Proof. Clearly only finitely many µ contribute nonzero terms to the right
side. Note too that both sides are additive in N : if (∗) holds for N1 and N2 ,
then it holds for N1 ⊕ N2 .
The strategy is to verify (∗) first for injectives in O, for which it is
enough to consider the various indecomposable modules I(λ) = P (λ)∨ . On
one hand, BGG Reciprocity gives
X
ch I(λ) = ch P (λ) = [M (µ) : L(λ)] ch M (µ).
µ
6.15. Comparison of Ext•O and Lie Algebra Cohomology 127

On the other hand, ExtnO M (µ), I(λ) = 0 for all n > 0 since I(λ) is

injective, whereas by definition

Ext0O M (µ), I(λ) = HomO M (µ), I(λ) ∼= HomO (P (λ), M (µ)∨ ).

Thanks to Theorem 3.9(c), the dimension of the right side is equal to

[M (µ)∨ : L(λ)] = [M (µ) : L(λ)]. Since the Euler characteristic collapses
in this case, comparison with ch I(λ) above yields (∗).
If N is arbitrary, it has a (finite) injective resolution
0 → N → I0 → I1 → . . . → In → 0.
Exactness then gives ch N = i (−1)i ch Ii . On the other hand, for each
P
fixed µ the Euler characteristic on the right side of (∗) can be computed
using the injective resolution of N . This yields

X
χ Ext•O (M (µ), N ) = (−1)i χ Ext•O (M (µ), Ii ) .

Finally, applying (∗) to each Ii (and interchanging summations over i and

µ) proves (∗) for N . 

6.15. Comparison of Ext•O and Lie Algebra Cohomology

To round out this chapter we outline briefly some connections between the
Extn functors for O and the more traditional Lie algebra cohomology, fol-
lowing Delorme [75, 76]. (He works over the complex numbers.) While this
comparison does not immediately lead to new computations, it makes the
functors ExtnO look less artificial and also permits interesting translations of
older cohomology results such as Bott’s Theorem (6.6).
The main idea is to view modules in O as “(g, h)-modules” in the spirit
of Harish-Chandra modules coming from Lie group representations. For
pairs such as (g, h) there is a well-developed theory of relative Lie algebra
cohomology (see for example Borel–Wallach [39, Chap. I]), which in this
case is related to n-cohomology. Delorme’s first two theorems can be stated
as follows, using our notation:
Theorem. Let M, N ∈ O. Then
∼ Extn (M, N ), which in turn is
(a) For all n ≥ 0, ExtnO (M, N ) =

(g,h)
isomorphic to H n g, h, Hom(M, N ) .
(b) If M = M (µ), then Extn M (µ), N ∼ = H n (n, N )µ , a weight space


O
for the natural action of h on n-cohomology.

In view of Bott’s
Theorem, (b) yields an interesting interpretation of

ExtkO M (w · λ), L(λ) when λ ∈ Λ+ and w ∈ W . More generally, (b)
128 6. Extensions and Resolutions

contributes to the formulation of Vogan’s version of the Kazhdan–Lusztig

Conjecture (8.10).
Using Proposition 6.14, together with Rocha’s vanishing criterion (6.5)
for ExtO (which can be avoided here with more work), Delorme arrives at
vanishing criteria for ExtnO which are more general that those in Theorem
6.11 above. These involve a kind of length function `(µ, λ) expressing the
distance between M (µ) and M (λ) in certain “strongly” standard filtrations:
ExtnO M (µ), M (λ) = 0 for all n > `(µ, λ),

ExtnO M (µ), L(λ) = 0 for all n > `(µ, λ),

Combined with the earlier comparisons, this recovers and generalizes some
vanishing criteria for n-cohomology with coefficients in L(λ) or M (λ) found
earlier by W. Schmid.

Notes
The BGG resolution goes back to the fundamental work of Bernstein–
Gelfand–Gelfand [26, Part II], culminating in their Theorem 10.10 . Af-
terwards the treatment was streamlined considerably by Rocha [226], who
developed the approach in 6.5 to show that their “weak” resolution is actu-
ally of the “strong” type they construct separately. She worked in the more
general setting of parabolic category O, as did Lepowsky [203] in an earlier
variant of the resolution using parabolic Verma modules.
The idea of a BGG resolution resonates in a number of directions, such
as the work of Gabber and Joseph [106] related to primitive ideals in U (g).
There are also generalizations to the setting of Kac–Moody algebras (13.5–
13.6 below) or quantum groups.
(6.9) Apart from the original source [27], unpublished work by König
[190] and a recent paper by Mazorchuk [215] go further into these questions.
In [138, §4] Irving places results such as those in 6.11 and 6.14 in the
axiomatic setting of “BGG algebras”.
(6.12) The results here are “well-known”; we have drawn on the lecture
notes of Gaitsgory [108].
(6.14–6.15) These sections are based on a 1978 preprint [75] by Delorme,
later published in a summary version [76] with the author identified only as
“M. Delorm”.
 Chapter 7

Translation Functors

Beyond the techniques involved in proving the theorems of Verma and BGG
discussed in Chapters 4 and 5, progress in the algebraic study of Verma
modules has been mainly due to the use of translation functors. These were
introduced in category O by Jantzen [148].
Foundations are laid in 7.1–7.4, after which the effect of the functors
on Verma modules is examined in 7.6. This already shows how to obtain
category equivalences between certain blocks (7.8).
It is a more subtle matter to apply translation functors to simple modules
(7.9), but then the effect on projectives becomes transparent (7.11). Deeper
ideas begin to emerge when we translate from a wall of a Weyl chamber
into the chamber or compose translation functors to create “wall-crossing”
functors (7.12–7.15). By combining these methods, we arrive at a general
proof that self-dual projectives are those corresponding to antidominant
weights (7.16); the case of integral weights was treated more directly in
Theorem 4.10.
Most of the ideas in this chapter were developed by Jantzen to aid in
the study of composition factor multiplicities in Verma modules. Though his
strong partial results in this direction were overtaken by the more definitive
Kazhdan–Lusztig Conjecture [171] (soon a theorem), translation functors
have continued to be a vital tool in the study of O and analogous categories.
In order to bypass the extra technical details needed to treat nonintegral
weights, the reader may want to focus at first just on the integral case.

129
130 7. Translation Functors

7.1. Translation Functors

The study of category O reduces to some extent to the study of the sub-
categories Oχ , with further refinements involving the dot-orbits of groups
W[λ] in cosets of h∗ modulo Λr . But it is natural to ask how the different
Oχ are related to each other as χ varies over the central characters. In
particular, which of these categories are equivalent to others? Here we in-
troduce “translation functors” on O, exact functors which take modules in
one Oχ to another. The idea has been developed extensively by Jantzen
in the algebraic setting (including the closely analogous study of algebraic
groups) but also arises independently in work of Zuckerman [262] on Lie
group representations.
More generally, as we shall see in later chapters, the structure of cate-
gory O is typically studied in terms of a variety of endofunctors: one can
learn much from seeing how they act on standard objects and how they
interact with other functors. We have already accumulated some elemen-
tary examples of exact functors on O: duality, tensoring with a fixed finite
dimensional module, projecting O onto one of the subcategories Oχ . By
composing such functors we get new ones whose effect on modules may be
much less obvious.
Translation functors are constructed as composites of projections to vari-
ous Oχ and tensoring with finite dimensional modules. Fix λ, µ ∈ h∗ , writing
prλ and prµ for the natural projections of O onto Oχλ and Oχµ . If L is finite


dimensional and M ∈ O, then M 7→ prµ L ⊗ (prλ M ) , followed by inclu-
sion into O, defines an exact functor on O. Its restriction to Oχλ (without
the inclusion) relates just the two subcategories in question. But when L is
arbitrary, examples quickly show that nothing very useful can be expected
from the construction. Instead, require that µ − λ ∈ Λ; in this case we say
that λ and µ are compatible, which is automatic if λ, µ ∈ Λ. (This implies
that Φ[λ] = Φ[µ] and W[λ] = W[µ] .) Now the W -orbit of ν := µ − λ contains a
unique weight ν in Λ+ ; set L := L(ν) and write Tλµ for the resulting functor
on O (or on Oχλ ). We call Tλµ a translation functor.
Here are a couple of initial observations which don’t yet require restric-
tions on the weights involved:

Proposition. If λ, µ ∈ h∗ , the exact functor Tλµ commutes with the duality

functor and takes projective modules to projective modules.

Proof. To see that Tλµ (M ∨ ) = ∼ (T µ M )∨ , use Theorem 3.2(b)(d) and first

λ
exercise in that section. The second assertion follows from the corresponding
properties of tensoring with a finite dimensional module (Proposition 3.8(b))
and projecting to a direct summand. 
7.2. Adjoint Functor Property 131

The rationale for requiring λ and µ to be compatible will soon become

more obvious. For now just observe that in the special case M = M (λ), a
standard filtration of L(ν)⊗M (λ) (as in 3.6–3.7) involves unique occurrences
of the Verma modules M (λ + wν), w ∈ W ; one of these highest weights is
µ. Thus M (µ) must be a subquotient of Tλµ M (λ). (This module also has a
standard filtration, thanks to Proposition 3.7(b).)

7.2. Adjoint Functor Property

How are the exact functors Tλµ and Tµλ related to each other? This is easy
to see, by invoking a basic adjointness property:
Lemma. Let L, M, N ∈ O with dim L < ∞. Then there is a natural iso-
morphism of vector spaces HomO (L ⊗ M, N ) ∼
= HomO (M, L∗ ⊗ N ).

Proof. Omitting the action of U (g), the vector space isomorphism is a stan-
dard fact from linear algebra (making no assumption about the dimensions
of M and N ). Compatibility with the U (g)-module structures is then routine
to check. 
Proposition. Let λ, µ ∈ h∗ be compatible. Then Tλµ is left and right adjoint
to Tµλ :
(a) For all M, N ∈ O,
HomO (Tλµ M, N ) ∼
= HomO (M, Tµλ N ).
(b) For all M, N ∈ O,
ExtnO (Tλµ M, N ) ∼
= ExtnO (M, Tµλ N ).
Similar isomorphisms hold with λ, µ interchanged.

Proof. By symmetry it is enough to check the displayed isomorphisms.

(a) With ν the W -conjugate in Λ+ of ν = µ − λ as above, recall that
L(ν)∗ ∼= L(−w◦ ν) (1.6). If ν = wν, then −w◦ ν = −w◦ wν = w◦ w(−ν) ∈
W (−ν) ∩ Λ+ . Thus −w◦ ν is the dominant weight conjugate to λ − µ. Com-
bine this observation with the lemma.
Then (b) follows from general properties of derived functors (and is easy
to verify directly). 

Like other exact functors, Tλµ preserves short exact sequences and there-
fore induces an endomorphism of the Grothendieck group K(O) (or equally
well, of the subgroup of X generated by formal characters of modules in O).
Since the [M (λ)] or the [L(λ)] form a basis for K(O), it is enough for many
purposes to know how translation functors act on Verma modules and sim-
ple modules. This will be explored below. But without further restrictions
132 7. Translation Functors

on the choice of λ and µ beyond the natural requirement that µ − λ ∈ Λ,

the effect of Tλµ on standard modules in hard to characterize.

Exercise. Let g = sl(2, C). Show that Tλµ need not take Verma modules to
Verma modules. [For example, let λ = 1 and µ = −3.]

7.3. Weyl Group Geometry

To deal with Tλµ when λ and µ are integral weights, we can take advantage
of the special geometry associated with the action of W on the euclidean
space E. This is developed by Bourbaki [45, Ch. V] in the general setting of
reflection groups, and in the case of finite reflection groups by Humphreys
[130, 1.15]. For the application to translation functors, see Jantzen [148,
Kap. 2] as well as [153, II.7].
Recall from 0.6 that the Weyl chambers in E are the connected com-
ponents in the complement of the union of all hyperplanes orthogonal to
roots; in particular, they are open sets (open cones if Φ is irreducible). The
chambers are permuted in simply transitive fashion by W , while the closure
of any fixed chamber is a fundamental domain for the action. The geom-
etry is unaffected by the shift of origin from 0 to −ρ, but now a typical
hyperplane becomes Hα := {λ ∈ E | hλ + ρ, α∨ i = 0}.
Under the simply transitive action of W , the chambers are in natural
bijection with simple systems in Φ: if C corresponds to ∆, then C lies on the
positive side of all hyperplanes Hα with α ∈ ∆ (hence all Hβ with β > 0).
Since a simple reflection sα permutes the positive roots other than α (0.3),
sα · C lies on the positive side of all hyperplanes Hβ with β > 0 but β 6= α.
In the study of translation functors we need a refined decomposition of E
into facets: a facet F is a nonempty subset of E determined by a partition
−
of Φ+ into disjoint subsets Φ0F , Φ+
F , ΦF :

∨ 0
hλ + ρ, α i = 0 when α ∈ ΦF

λ ∈ F ⇔ hλ + ρ, α∨ i > 0 when α ∈ Φ+ F
hλ + ρ, α∨ i < 0 when α ∈ Φ−


F

Clearly the closure F is obtained by replacing > by ≥ and < by ≤.

Remarks. (1) Given a facet F , it is easy to check that the roots α for which
±α ∈ Φ0F form a root system, whose Weyl group may be characterized as
the subgroup of W fixing F pointwise (such an isotropy group is generated
by the reflections it contains: see 0.3).
(2) Observe that if λ ∈ F and w · λ < λ for some w ∈ W , then the same
inequality holds for all points of F . Moreover, w · µ ≤ µ for all µ ∈ F .
7.3. Weyl Group Geometry 133

Next define the upper closure Fb of the facet F by the conditions


∨ 0
hλ + ρ, α i = 0 when α ∈ ΦF

λ ∈ Fb ⇔ hλ + ρ, α∨ i > 0 when α ∈ Φ+F
hλ + ρ, α∨ i ≤ 0 when α ∈ Φ−


F
The lower closure is then defined similarly.
Exercise. Let F be a facet in E.
(a) Prove that F is the union of the facets F 0 for which Φ+ +
F 0 ⊂ ΦF and
Φ− −
F 0 ⊂ ΦF .
(b) Prove that Fb is the union of the facets F 0 ⊂ F satisfying for every
α > 0 the condition:
hλ + ρ, α∨ i ≤ 0 for all λ ∈ F 0 ⇔ hλ + ρ, α∨ i ≤ 0 for all λ ∈ F.

Now a facet C is called a chamber if Φ0C = ∅. The closure C is a

fundamental domain for the dot-action of W . Those facets F of C for which
|Φ0F | = 1 are called walls; each of these lies in either the upper or the lower
closure of C. When ∆ is fixed, there is a (shifted) dominant chamber
whose upper closure is itself:
C◦ := {λ ∈ E | hλ + ρ, α∨ i > 0 for all α ∈ ∆}.
All other chambers are of the form C = wC◦ for w ∈ W . It is easy to check
−
that Φ+ + + +
C = Φ ∩ wΦ and ΦC = Φ ∩ wΦ .
−

A moment’s thought shows that a given (nonempty!) facet F lies in the

upper closure of the chamber C defined by Φ− 0 − +
C = ΦF ∪ ΦF and ΦC = ΦF .
+

This evidently characterizes C uniquely, proving part (a) of the following

proposition. Part (b) follows straightforwardly from the above description
of an upper closure:
Proposition. Upper closures of facets have the following properties:
(a) Each facet lies in the upper closure of a unique chamber.
(b) If the facet F lies in the upper closure of the chamber C, then
Fb ⊂ C.
b 

How does this geometry enter into the study of translation functors?
In the important special case when λ and µ lie in Λ (in which case their
difference also does), the work of Jantzen gives the most definitive results
when λ and µ both lie in the closure of a single facet F or when µ lies in the
upper closure of the facet containing λ. Moreover, the chamber geometry will
sometimes permit the use of inductive arguments, if we define the “distance”
between chambers C and C 0 to be the number of root hyperplanes separating
C 0 from C.
134 7. Translation Functors

7.4. Nonintegral Weights

Having refined the formulation of translation functors in the case of integral
weights by taking account of the facets in which weights lie, we now consider
the general case: λ, µ ∈ h∗ with µ − λ ∈ Λ. Recalling the discussion in 3.4
of the group W[λ] and the parametrization of its dot-orbits by antidominant
weights, we can use the subspace E(λ) of E spanned by Φ[λ] to capture
all of the “integral” behavior of the weight λ. To this end we associate
with an arbitrary λ ∈ h∗ the integral weight λ\ (relative to Φ[λ] ) in E(λ)
characterized uniquely by the requirement that
hλ\ + ρ, α∨ i = hλ + ρ, α∨ i for all α ∈ Φ[λ] .
Equivalently, hλ\ , α∨ i = hλ, α∨ i for α ∈ Φ[λ] .
Denote by Wλ◦ the stabilizer of λ in W[λ] , relative to the dot-action. It
depends on λ and not just its coset modulo Λ. (In [148] Jantzen uses the
notation Rλ rather than λ\ . Since he uses the notation Wλ in place of our
W[λ] , he writes Wλ◦ for the isotropy group.)
It is easy to check a few elementary facts:
Proposition. Let λ ∈ h∗ , with stabilizer Wλ◦ in W[λ] .
(a) For all w ∈ W[λ] , we have λ−w·λ = λ\ −w·λ\ , while (w·λ)\ = w·λ\ .
(b) The stabilizer of λ\ in W[λ] is Wλ◦ .
(c) Suppose Φ[λ] = Φ[µ] . Then (λ + µ)\ = λ\ + µ\ .

Proof. (a) Since the sα with α ∈ Φ[λ] generate W[λ] , it is enough to check
these assertions when w = sα . The first one follows immediately from the
way λ\ is defined. The second follows from the equality valid for all β ∈ Φ[λ] :
hsα (λ\ + ρ), β ∨ i = hλ\ + ρ, sα β ∨ i = hλ + ρ, sα β ∨ i = hsα (λ + ρ), β ∨ i.

(b) Part (a) shows that Wλ◦ stabilizes λ\ . In the other direction, we
know from the theory of reflection groups (0.3) that the stabilizer of λ\ in
W[λ] is generated by those sα for which hλ\ + ρ, α∨ i = 0, or equivalently,
hλ + ρ, α∨ i = 0. These reflections also stabilize λ, whence (b).
(c) It follows from the assumption that Φ[λ+µ] = Φ[λ] . In turn, for all
α ∈ Φ[λ] = Φ[µ] we compute
h(λ+µ)\ +ρ, α∨ i = hλ+µ+ρ, α∨ i = hλ\ +ρ, α∨ i+hµ\ , α∨ i = hλ\ +µ\ +ρ, α∨ i.

Remarks. (1) As noted in the proof of (b), Wλ◦ is generated by the re-
flections sα which fix λ\ (or λ). More precisely, if λ\ lies in the closure of
a chamber C in E(λ), the theory of reflection groups (7.3) shows that Wλ◦
7.5. Key Lemma 135

is generated by the reflections in the walls which contain λ\ . For exam-

ple, when λ is either dominant or antidominant, these reflections belong to
simple roots in ∆[λ] .
(2) Similarly, the isotropy group of λ in W is generated by those sα for
which hλ + ρ, α∨ i = 0. Since these are the reflections fixing λ\ , we see that
Wλ◦ is also the subgroup of W fixing λ.
Exercise. Show that Wλ◦ is the group which fixes pointwise the facet F ⊂
E(λ) to which λ\ belongs; in turn, Wλ◦ is the group fixing F pointwise.

How does the transition from λ to λ\ affect the study of translation

functors? Given λ, µ ∈ h∗ with λ and µ compatible, the study of Tλµ can often
be reduced to the study of how integral weights such as λ\ and µ\ are located
in the geometry of E(λ) = E(µ) relative to W[λ] = W[µ] . Here Jantzen’s
refinements in terms of the facets of E are replaced by similar refinements
involving the facets of E(λ). In particular, part (a) of the proposition shows
that the partial order of the linkage class of λ is mirrored by that of λ\ .

7.5. Key Lemma

Now we can begin to explore the effect of translation functors on both Verma
modules and simple modules in O. This leads to category equivalences
between certain blocks (as well as an alternative proof of the BGG Theorem
[148, 2.20]). The ideas here are due to Jantzen, as indicated in the Notes.
We start with the easier case of Verma modules, where Theorem 3.6 can
be applied directly. The following lemma captures, without extraneous no-
tation, the essential information about what happens when a Verma module
is tensored with a finite dimensional simple module and then projected to
an appropriate block.
Lemma. Let λ, µ ∈ h∗ be compatible, with ν := µ − λ ∈ Λ and ν the unique
W -conjugate of ν lying in Λ+ . Assume that λ\ lies in a facet F of E(λ),
while µ\ lies in the closure F . Then for all weights ν 0 6= ν of L(ν), the
weight λ + ν 0 is not linked by W[λ] = W[µ] to λ + ν = µ.

Proof. First we observe that the proof reduces to a corresponding statement

about linkage of integral weights in E(λ). Suppose there exists w ∈ W[λ]
such that w · (λ + ν 0 ) = λ + ν. Here w · (λ + ν 0 ) = w · λ + wν 0 , so we have
λ − w · λ = wν 0 − ν. Thanks to Proposition 7.4(a), the left side is the same
as λ\ − w · λ\ . Rewriting, we get w · (λ\ + ν 0 ) = λ\ + ν. So it is enough to
show that this forces ν 0 = ν.
Suppose on the contrary ν 0 6= ν with λ\ +ν 0 linked to λ\ +ν and eventually
reach a contradiction. The geometry comes into play by placing λ\ + ν,
λ\ + ν 0 in the closures of two chambers C, C 0 and using induction on the
136 7. Translation Functors

distance d(C, C 0 ) between them: this is a geometric version of the difference

between lengths in W and is defined to be the number of root hyperplanes
Hα separating C from C 0 (which is 0 just when C = C 0 ). Since we are
working with a hypothetical configuration of weights which turns out to be
impossible, it is unfortunately not easy to gain intuition into the proof by
drawing a sketch. But the individual steps are straightforward.
Suppose that ν 0 6= ν is a weight of L(ν) with λ\ + ν 0 linked to λ\ + ν.
Say the given facet F is contained in the closure C of some chamber C and
choose ν 0 to make d(C, C 0 ) as small as possible with λ\ + ν 0 ∈ C 0 . Using
induction on this distance, we will eventually arrive at a contradiction.
(1) The case d(C, C 0 ) = 0 (which means C = C 0 ) is impossible, since C
is a fundamental domain for the dot-action of W while λ\ + ν 0 6= λ\ + ν.
(2) Now d(C, C 0 ) > 0. Since a chamber is uniquely determined by the
hyperplanes which define its walls, there exists α > 0 such that Hα ∩ C 0
contains a wall of C 0 and Hα separates C 0 from C. Say C 0 lies on the positive
side and C on the negative side; so for all ξ ∈ F we have hξ + ρ, α∨ i ≤ 0.
(3) Set C 00 := sα · C 0 . Then d(C, C 00 ) < d(C, C 0 ). To see this, assume
without loss of generality that C 0 corresponds to the simple system ∆[λ]
(with α ∈ ∆[λ] ); thus C is separated from C 00 by the same hyperplanes
(except Hα ) which separate C from C 0 .
(4) Since λ\ + ν 0 ∈ C 0 , we have hλ\ + ν 0 + ρ, α∨ i ≥ 0 (thanks to (2)).
(5) Since λ\ ∈ F ⊂ C, we have hλ\ + ρ, α∨ i ≤ 0 (also thanks to (2)).
(6) From (4) we get sα · (λ\ + ν 0 ) = λ\ − hλ\ + ρ, α∨ iα + sα ν 0 ≤ λ\ + ν 0 .
(7) Combining (5) and (6), we get sα ν 0 ≤ sα ν 0 − hλ\ + ρ, α∨ iα ≤ ν 0 .
(8) Setting ν 00 := sα ν 0 − hλ\ + ρ, α∨ iα, we deduce from (6) and (7) that
sα · (λ\ + ν 0 ) = λ\ + ν 00 ∈ sα · C 0 = C 00 .
(9) Since ν 0 and sα ν 0 are weights of L(ν), while sα ν 0 ≤ ν 00 ≤ ν 0 by (7),
it follows from 1.6 that the intermediate weight ν 00 in this α-string is also a
weight of L(ν).
(10) The minimality assumed for ν 0 , coupled with (8) and (9), implies
that ν 00 = ν.
(11) In turn, the inequalities sα ν 0 ≤ ν ≤ ν 0 in (9) force sα ν 0 = ν: here
ν 6= ν 0 by assumption, whereas not both ν + α and ν − α can be weights of
L(ν) because ν is an extremal weight (see 1.6).
(12) But ν = ν 00 , which by definition is sα ν 0 − hλ\ + ρ, α∨ iα. So sα ν 0 = ν
in (11) forces hλ\ + ρ, α∨ i = 0.
(13) Since λ\ ∈ F , all ξ ∈ F must also satisfy hξ+ρ, α∨ i = 0, in particular
ξ = λ\ + ν. This implies that hν, α∨ i = 0.
7.6. Translation Functors and Verma Modules 137

(14) Combining sα ν 0 = ν from (11) with hν, α∨ i = 0 from (13), we reach

the contradiction ν = ν 0 . 
Exercise. How does the proof simplify if both λ\ and µ\ are assumed to lie
in C?

7.6. Translation Functors and Verma Modules

Now it is easy to see how Verma modules behave under translation from a
facet into its closure.
Theorem. Let λ, µ ∈ h∗ be antidominant and compatible. Assume that
λ\ lies in a facet F of E(λ) relative to the dot-action of W[λ] , while µ\
lies in F . Then Tλµ M (w · λ) ∼ = M (w · µ) for all w ∈ W[λ] . Similarly,
Tλµ M (w · λ)∨ ∼
= M (w · µ)∨ for all w ∈ W .
[λ]

Proof. The idea is to apply Lemma 7.5 to w · λ and w · µ in place of λ

and µ, letting ν := w · µ − w · λ. The assumption that λ\ ∈ F and µ\ ∈ F
transfers immediately to this setting, with F replaced by the facet w · F .
Now Tλµ involves tensoring M (w · λ) with the finite dimensional module
L(ν). Theorem 3.6 insures that the resulting module has a filtration whose
quotients are the Verma modules M (w · λ + ν 0 ), with ν 0 ranging over the
weights of L(ν) counted according to multiplicity; in particular, M (w · µ) =
M (w · λ + ν) occurs precisely once. After passing to the linkage class of w · µ,
the only question is whether other Verma modules can also be involved. But
this is ruled out immediately by the lemma.
The last statement of the theorem follows from the fact that translation
functors commute with taking duals (Proposition 7.1). 

It is not immediately obvious what Tλµ does to M (w·λ) if the assumptions

on λ and ν in the theorem are reversed: µ\ ∈ F while λ\ ∈ F . This will
be explored in 7.12 below; it turns out to be intimately involved with the
deeper study of composition factors of Verma modules in Chapter 8. But it
is easy to extend the theorem from Verma modules to other modules with
a standard filtration:
Corollary. Let λ, µ be as in the theorem. If M ∈ Oλ has a standard filtra-
tion, then Tλµ M has a standard filtration in Oµ .

Proof. Use induction on the filtration length of M . The theorem disposes

of the case when M is a Verma module. In general, we have a short exact
sequence 0 → N → M → M (w · λ) → 0. Since Tλµ is an exact functor, the
theorem yields 0 → Tλµ N → Tλµ M → M (w · µ) → 0. By induction, Tλµ N has
a standard filtration in Oµ ; so Tλµ M also does. 
138 7. Translation Functors

7.7. Translation Functors and Simple Modules

With Theorem 7.6 in hand, it is natural to ask under the same hypotheses
whether Tλµ permits a comparison of composition factor multiplicities in
Verma modules. In particular, what is the effect of Tλµ on a simple module
L(w · λ) (assuming for simplicity that λ ∈ Λ)? It is easy to get started in
this direction, but not so easy to formulate a definitive theorem.
Proposition. Let λ, µ ∈ h∗ be antidominant and compatible. Assume that
λ\ lies in a facet F of E(λ) relative to the dot-action of W[λ] , while µ\ lies
in F . If w ∈ W[λ] , then Tλµ L(w · λ) is either 0 or else isomorphic to L(w · µ).

Proof. Here the exactness of Tλµ can be invoked. Applied to the surjective
map M (w · λ) → L(w · λ), it produces (using Theorem 7.6) a surjective map
M (w · µ) → Tλµ L(w · λ). So M := Tλµ L(w · λ) is a highest weight module of
weight w · µ or else 0.
Suppose M 6= 0. Another application of the theorem shows that Tλµ ,
applied to the injective map L(w · λ) → M (w · λ)∨ , produces an injective
map M → M (w · µ)∨ . Then Theorem 3.3(c) forces M ∼ = L(w · µ). 
Exercise. In the proposition, assume that Tλµ L(w · λ) ∼ = L(w · µ). For
arbitrary M ∈ O, prove that [M : L(w · λ)] = [Tλµ M : L(w · µ)].

It remains to decide when translation gives a nonzero result. Here is a

cautionary example which points the way to a general theorem.
Example. Let g = sl(2, C) and consider just integral weights (identified
with integers). In case λ ≥ 0, it lies inside the facet F := Z+ . Start with
the exact sequence 0 → L(−λ − 2) → M (λ) → L(λ) → 0. We know that
L(−λ − 2) = M (−λ − 2). Now set µ = −1 (that is, −ρ), so M (µ) = L(µ).
Here µ lies in F as required in the proposition. By Theorem 7.6, applying
Tλµ produces the exact sequence 0 → M (µ) → M (µ) → Tλµ L(λ) → 0. This
forces Tλµ L(λ) = 0.
On the other hand, T µ L(−λ − 2) ∼
λ = L(µ). So the behavior of simple
modules under translation into the closure of a facet is more complicated
than the behavior of Verma modules. Note however that µ fails to lie in the
upper closure of F whereas it does lie in the upper closure of the opposite
facet (7.3).

7.8. Application: Category Equivalences

With the help of Theorem 7.6 and Proposition 7.7, we can exploit the exact-
ness properties of translation functors (7.2) to define an equivalence between
certain blocks of O. Recall from 4.9 that the blocks Oλ are parametrized
7.8. Application: Category Equivalences 139

by antidominant weights λ ∈ h∗ : the objects in Oλ are the modules whose

composition factors all have highest weights in W[λ] · λ.
As a step toward proving a category equivalence, consider first the as-
sociated Grothendieck groups (which are isomorphic to the groups spanned
by formal characters): see 1.11.
Proposition. Let λ, µ ∈ h∗ be antidominant and compatible. Assume that
λ\ , µ\ lie in the same facet of E(λ) = E(µ) relative to W[λ] = W[µ] . Then Tλµ
induces an isomorphism between the Grothendieck groups K(Oλ ) and K(Oµ )
of the associated blocks, sending [M (w · λ)] to [M (w · µ)] and [L(w · λ)] to
[L(w · µ)] for all w ∈ W[λ] .

Proof. Recall that the symbols [M (w·λ)] form a Z-basis of K(Oλ ), as do the
symbols [L(w · λ)]; there are similar bases for K(Oµ ). Thanks to Theorem
7.6, Tλµ M (w · λ) ∼= M (w · µ) and inversely using Tµλ . When an arbitrary
basis element [M ] of Oλ is written uniquely as a Z-linear combination of the
[M (w · λ)], it follows at once that the map induced by Tµλ Tλµ sends [M ] to
[M ] and similarly for the maps in reverse order on K(Oµ ). This gives an
isomorphism of Grothendieck groups.
Similarly, Proposition 7.7 implies that Tµλ Tλµ sends [L(w · λ)] to either
[L(w·λ)] or 0. But the latter is impossible because of the above isomorphism
of Grothendieck groups. In turn, Tλµ must send [L(w·λ)] to [L(w·µ)]; indeed,
Proposition 7.7 forces Tλµ L(w · λ) ∼
= L(w · µ). 

In view of the isomorphism (1.15) between K(O) and the group X0 of

associated formal characters induced by M 7→ ch M , the proposition implies
that composition factor multiplicities in Oλ match those in Oµ :
[M : L(w · λ)] = [Tλµ M : L(w · µ)] for all w ∈ W[λ] .
In the general setting of artinian abelian categories, Gaitsgory observes
in his lecture notes [108, Lemma 4.27] that isomorphism on the level of
Grothendieck groups is actually enough to imply category equivalence. Here
we show this directly in our situation:
Theorem. Let λ, µ ∈ h∗ be antidominant and compatible. In case λ\ and
µ\ lie in the same facet for the action of W[λ] = W[µ] on E(λ) = E(µ), the
functors Tλµ and Tµλ define inverse equivalences of categories between Oλ and
Oµ . In particular, Tλµ sends simple modules to simple modules.

Proof. It has to be shown that Tµλ Tλµ is naturally isomorphic to the identity
functor on Oλ (and similarly in the other order). For this we use induction
on the length of an arbitrary module M in Oλ to show that Tµλ Tλµ M ∼ = M.
µ λ
Coupled with the functorial properties of Tλ and Tµ , this will prove the
theorem.
140 7. Translation Functors

By the adjointness property (7.1), we have for all M, N ∈ O

HomO (T λ M, N ) ∼
µ
µ
= HomO (M, T N ).λ
In particular, for all M ∈ Oλ we get
HomO (Tµλ Tλµ M, M ) ∼ µ µ
= HomO (Tλ M, Tλ M ).
Let ϕM be the unique homomorphism on the left corresponding to the iden-
tity map on the right. By the naturality of the isomorphisms involved, each
short exact sequence 0 → N → M → L → 0 in Oλ with L simple induces a
commutative diagram
0 −−−−→ Tµλ Tλµ N −−−−→ Tµλ Tλµ M −−−−→ Tµλ Tλµ L −−−−→ 0
  
ϕN y ϕM y ϕL y
  

0 −−−−→ N −−−−→ M −−−−→ L −−−−→ 0

Now proceed by induction on length. In the case of length 1, the proposition
shows that ϕL is an isomorphism. By induction, ϕN is an isomorphism, so
ϕM is also an isomorphism. 
Remark. Under the hypotheses of the theorem, we know from Theorem 7.6
that Tλµ M (w · λ) ∼
= M (w · µ) for all w ∈ W[λ] . Each of the Verma modules
has a Jantzen filtration (5.3), leading Jantzen [148, 5.17] to ask whether Tλµ
induces isomorphisms M (w · λ)i ∼ = M (w · µ)i . (One can ask more generally
what happens if µ is only required to lie in the closure of the facet of λ\ .)
\

Exercise. Given compatible antidominant weights λ, µ, ν ∈ h∗ , assume that

µ\ lies in the closure of the facet containing λ\ while ν \ lies in the closure of
the facet containing µ\ . Prove that ch(Tµν Tλµ M ) = ch(Tλν M ) for all M ∈ Oλ .

7.9. Translation to Upper Closures

The next task is to determine what happens to simple modules under trans-
lation into the closure of a facet. Recall from 7.3 how to define the upper
closure Fb of a facet F for the dot-action of a reflection group (which will be
W[λ] in our situation) in terms of a partition of the positive roots.
Theorem. Let λ, µ ∈ h∗ be antidominant and compatible. Assume that
λ\ ∈ F for some facet F ⊂ E(λ), while µ\ ∈ F .
(a) For all w ∈ W[λ] , if w · µ\ lies in the upper closure of w · F , then
Tλµ L(w · λ) ∼
= L(w · µ).
(b) If w ∈ W[λ] but w · µ\ ∈
/w[· F , then Tλµ L(w · λ) = 0.

Proof. Proposition 7.7 shows that Tλµ L(w · λ) is either L(w · µ) or else 0,
while Theorem 7.6 shows that Tλµ M (w · λ) ∼
= M (w · µ). The exact functor
Tλµ must therefore take some composition factor L(w0 w · λ) to L(w · µ); here
7.9. Translation to Upper Closures 141

w0 w ·λ ≤ w ·λ. Exactness of Tλµ also insures that no other composition factor

can be translated to L(w · µ). On the other hand, Tλµ L(w0 w · λ) ∼ = L(w0 w · µ)
because it is nonzero. This forces w0 w · µ = w · µ, and similarly for µ\ . So
w0 lies in the subgroup W 0 of W[λ] generated by the reflections sα for which
α ∈ wΦ0F .
(a) Suppose w · µ\ lies in w [· F . Then each reflection sα with α ∈ wΦ0F
satisfies sα w · λ\ ≥ w · λ\ . By Remark (1) in 7.3, the group W 0 is itself the
Weyl group of a root system attached to F . Applying Lemma 0.6 to this
group, one gets w0 w · λ\ ≥ w · λ\ and similarly for λ in place of λ\ . Combined
with the choice of w0 above, this implies equality. Thus L(w · λ) = L(w0 w · λ)
is the unique composition factor of M (w · λ) taken to L(w · µ) by Tλµ .
(b) If w·µ\ fails to lie in the upper closure of w·F , then by definition it lies
in some hyperplane for sα ∈ W[λ] bounding w · F below. Thus sα w · λ < w · λ
but sα w · µ = w · µ. By Verma’s Theorem, there is a (proper) embedding of
M (sα w · λ) into M (w · λ). But Tλµ maps both of these Verma modules to
M (w · µ). In other words, given the short exact sequence
0 → M (sα w · λ) → M (w · λ) → Q → 0,
we get Tλµ Q = 0. In turn, the quotient Tλµ L(w · λ) of Tλµ Q is 0. 
Remark. The proof in (a) above is an adaptation of one given by Jantzen
for algebraic groups in [153, II.7.15]; it is more direct than his earlier proof
in [148, 2.11–2.13]. But either proof avoids use of the BGG Theorem, which
allows him to provide a new proof of this theorem using translation functors
(see [148, 2.20]). Here is an alternative approach to (a) which does use the
BGG Theorem:
Suppose w · µ\ lies in w[· F . In this case we claim that w0 w · λ < w · λ
is impossible, forcing equality to hold. Thus L(w · λ) = L(w0 w · λ) is the
unique composition factor of M (w · λ) taken to L(w · µ) by Tλµ .
If the strict inequality holds, the BGG Theorem (5.1) implies the exis-
tence of reflections s1 , . . . , sr ∈ W[λ] such that
w0 w · λ = s1 . . . sr w · λ < · · · < sr w · λ < w · λ.
The same inequalities hold when λ is replaced by λ\ and then apply equally
well to all elements of F . Since µ\ ∈ F , we get
w0 w · µ\ = s1 . . . sr w · µ\ ≤ · · · ≤ sr w · µ\ ≤ w · µ\ .
But w0 w·µ\ = w·µ\ , so equality holds everywhere in this chain. In particular,
w
[ · F meets the root hyperplane for sr , contrary to sr w · λ\ < w · λ\ .
Exercise. Let λ, µ satisfy the hypotheses of the above theorem. Prove that
for all w ∈ W[λ] , the projective module Tλµ P (w · λ) is nonzero.
142 7. Translation Functors

7.10. Character Formulas

The results on translation functors obtained so far in this chapter enable
us to refine further the problem of finding composition factor multiplicities
for arbitrary Verma modules. This problem has already been reduced to
the study of the blocks Oλ , where λ runs over the antidominant weights.
Thanks to the theorems of Verma and BGG (5.1), the formal character of
each L(w · λ) with w ∈ W[λ] can be expressed as a Z-linear combination of
the ch M (w0 · λ) for which w0 · λ ↑ w · λ. When λ is regular, this relationship
simplifies to w0 ≤ w in the Bruhat ordering of W[λ] (5.2).
In any case, it is equivalent to write [L(w · λ)] as a Z-linear combination
of the [M (w0 · λ)] in the Grothendieck group K(Oλ ):
X
(∗) [L(w · λ)] = bλw0 ,w [M (w0 · λ)],
w0 ∈W[λ] , w0 ·λ↑w·λ

where bλw,w = 1. This format for the coefficients emphasizes the role of W[λ]
while raising the question of how these integers depend on λ.
To exploit translation functors, look first at the case of integral weights;
here the formulations are a bit more straightforward. Since translation
functors are exact, they induce homomorphisms between the Grothendieck
groups K(Oλ ). Now in case λ and µ lie in the same facet for W , Proposition
7.8 insures that Tλµ sends [M (w · λ)] to [M (w · µ)] for all w ∈ W . This in
turn transforms the formula (∗) for [L(w · λ)] into one for [L(w · µ)] involving
precisely the same coefficients, which are thus seen to depend just on W .
So the category equivalence between Oλ and Oµ obtained in Theorem 7.8
behaves well with respect to formal characters.
More generally, if w · µ lies in the upper closure of the facet to which w · λ
belongs, then Theorem 7.9 shows that Tλµ L(w · λ) ∼ = L(w · µ). So again the
formula for [L(w · λ)] yields the formula for [L(w · µ)]. But here one has to
combine some coefficients, taking into account the isotropy group Wµ◦ . For
example, if λ is regular, then after translation the coefficient of [M (w0 · µ)]
in (∗) becomes
X
bλw0 z,w .
w0 z≤w, z∈Wµ◦

Example. Let g = sl(3, C), with ∆ = {α, β}. If C is the antidominant

Weyl chamber, suppose λ ∈ Λ is regular and antidominant, while µ ∈ Λ is
antidominant and lies in just the β-hyperplane. Setting w = sβ sα , we know
from 5.4 that

[L(w · λ)] = [M (w · λ)] − [M (sα · λ)] − [M (sβ · λ)] + [M (λ)].

7.11. Translation Functors and Projective Modules 143

Now w · µ lies in the α-hyperplane, in the upper closure of w · C. Since

Wµ = {1, sβ }, applying Tλµ involves some cancellation and results simply in
[L(w · µ)] = [M (w · µ)] − [M (sα · µ)]. (This is a typical pattern for sl(2, C).)

The general case involves two compatible antidominant weights λ, µ ∈ h∗

and the reflection subgroup W[λ] = W[µ] of W . Here the above discussion
adapts readily, taking into account the relative positions of λ\ and µ\ in the
alcove geometry of E(λ) = E(µ).

7.11. Translation Functors and Projective Modules

As observed in Proposition 7.1, translation functors take projectives to pro-
jectives. By combining Theorem 7.9 with the adjointness property of Tλµ
and Tµλ we can make a precise statement:
Theorem. Let λ, µ ∈ h∗ be antidominant and compatible. Further assume
that µ\ lies in the closure of the facet F to which λ\ belongs. If w · µ\ lies in
w[· F for some w ∈ W[λ] , then Tµλ P (w · µ) ∼= P (w · λ).

Proof. We know that P := Tµλ P (w · µ) is projective. Theorem 3.9(b) allows

us to determine the indecomposable summands P (w0 · λ) of P by looking at
homomorphisms:
dim HomO P, L(w0 · λ) = multiplicity of P (w0 · λ) as a summand of P.

In our situation, adjointness implies

HomO Tµλ P (w · µ), L(w0 · λ) ∼ µ
= HomO P (w · µ), Tλ L(w0 · λ) .

Thanks to Theorem 7.9, Tλµ L(w0 ·λ) is isomorphic to L(w0 ·µ) when w0 ·µ\ lies
\
in w 0 · F , but is otherwise 0. Since L(w · µ) is the unique simple quotient of

P (w·µ), we must get 0 unless L(w0 ·µ) ∼ = L(w·µ) (in which case w0 ·µ = w·µ).
The assumption on w·µ, together with Theorem 7.9, insures that P (w·λ)
occurs once as a summand of P . Suppose another summand P (w0 · λ) also
occurs: here w0 · µ\ = w · µ\ lies in the upper closure of the respective facets
w0 · F and w · F in which w0 · λ\ and w · λ\ lie. We have to show that
w0 · λ = w · λ.
By Proposition 7.3(a), the facet w · F lies in C b for a unique chamber
\
C. In turn, Proposition 7.3(b) implies that w · µ ∈ C. b Then C is also the
unique chamber whose upper closure contains the facet of w · µ\ = w0 · µ\ .
In turn, w0 · F must lie in C.
b But C is a fundamental domain for the action
of W[λ] , so w · λ = w · λ . This shows that w−1 w0 lies in the stabilizer of λ\
\ 0 \

in W[λ] , which is the same as the stabilizer Wλ◦ of λ in W[λ] (or in W ): see
Proposition 7.4(b). Thus w0 · λ = w · λ. 
144 7. Translation Functors

7.12. Translation From a Facet Closure

Now we study more closely translation functors Tµλ , with µ\ in the closure
of the facet of λ\ . The effect of such functors on Verma modules (or their
simple quotients) is potentially quite complicated, but also interesting for
the pursuit of composition factor multiplicities. Following Jantzen [148,
2.17], we can make an explicit statement about formal characters. Since
Tµλ M (w · µ) has a standard filtration (Corollary 7.6), this specifies the filtra-
tion multiplicities. Recall that notation such as Wµ◦ indicates the isotropy
group of µ relative to the dot-action of W . (This is the same when µ is
replaced by µ\ and W by W[µ] .)
Theorem. Let λ, µ be antidominant and compatible. Assume that µ\ lies
in the closure of the facet of λ\ . Then for all w ∈ W[λ] = W[µ] ,
X
(1) ch Tµλ M (w · µ) = ch M (ww0 · λ).
w0 ∈Wµ◦ /Wλ◦

In particular, all Verma modules which occur as quotients in a standard

filtration of Tµλ M (w · µ) have filtration multiplicity 1.

Proof. We take advantage of the characterization of standard filtration mul-

tiplicities given by Theorem 3.7, together with the adjointness property of
translation functors. Given w ∈ W[λ] , it is convenient to denote an arbitrary
weight linked to λ by ww0 · λ with w0 ∈ W[λ] .
Tµλ M (w · µ) : M (ww0 · λ) = dim HomO Tµλ M (w · µ), M (ww0 · λ)∨

= dim HomO M (w · µ), Tλµ M (ww0 · λ)∨

= dim HomO M (w · µ), (Tλµ M (ww0 · λ))∨

= dim HomO M (w · µ), M (ww0 · µ)∨ .

By Theorem 3.3(c), this last dimension is 0 unless ww0 · µ = w · µ (equivalent

to w0 ∈ Wµ◦ ), in which case it is 1. Since ww0 · λ = w · λ just when w0 ∈ Wλ◦ ,
the character equality in (1) follows. 

In the framework of 7.5, the theorem recovers indirectly the fact that the
only weights ν 0 of the finite dimensional module L(ν) with highest weight
W -conjugate to λ − µ for which µ + ν 0 ∈ W[µ] · λ are the extremal weights.
(In [148] this was the approach taken to the theorem, parallel to the proof
of Theorem 7.6.)
Corollary. Let λ, µ be as in the theorem. For arbitrary M ∈ O, we have
ch Tλµ Tµλ M = |Wµ◦ /Wλ◦ | ch M .

Proof. First write ch M as a Z-linear combination of characters of Verma

modules, then apply the theorem together with Theorem 7.6. 
7.13. Example 145

What can be said about the module structure of Tµλ M (w · µ) in the

situation of the theorem? Since we do not in general understand well the
submodule structure of a single Verma module (including its composition
factor multiplicities), there are obvious limits to what we can learn by ap-
plying translation functors. However, it is reasonable to use the adjoint
property as in the proof above to investigate the possible simple submod-
ules or simple quotients of Tµλ M (w · µ), using the results of 7.9. This may
help to decide, for example, whether Tµλ M (w · µ) is indecomposable. Rather
than attempt this in general, we turn to the most important special case.

7.13. Example
Theorem 7.12 can be used to recover Theorem 4.10 in a more general setting,
as follows. Keep the notation above: λ, µ are antidominant and compatible,
with µ\ is the closure of the facet of λ\ . Denote by wλ the longest element of
W[λ] = W[µ] , and let wµ◦ be the longest element of its (parabolic) subgroup
Wµ◦ . It follows from the definitions that λ◦ := wλ · λ and µ◦ := wλ · µ are
dominant. In particular, Proposition 3.8(a) shows that M (µ◦ ) is projective.
In turn, P := Tµλ M (µ◦ ) is projective, thanks to Proposition 7.1.
Now we can apply Theorem 7.12 with w taken to be wλ . The charac-
ter formula there applies to any standard filtration of P . To simplify the
notation we assume that λ is regular, leaving the details in the general case
to the reader. Now Wλ◦ = 1, so the sum in the theorem is taken over the
characters of all M (wλ w0 · λ) with w0 ∈ Wµ◦ .
Recall from 7.4 that the reflection group Wµ◦ is generated by the simple
reflections in W[λ] which fix µ; in particular, its intrinsic Bruhat ordering
is induced by the Bruhat ordering of W[λ] (0.4). Thus wµ◦ · λ ≥ w0 · λ for
all w0 ∈ Wµ◦ , a relationship which is reversed when wλ is applied to both
weights. Now Remark 3.6 implies that M (wλ wµ◦ · λ) occurs as a quotient of
P . This in turn forces P (wλ wµ◦ · λ) to occur as a summand of P .
We claim there are no other summands, which would necessarily be of the
form P (wλ w0 ·λ) with w0 ∈ Wµ◦ . From Verma’s Theorem 4.6 and the ordering
of weights, we know that M (wλ wµ◦ · λ) embeds in M (wλ w0 · λ); in particular,
L(wλ wµ◦ ·λ) is a composition factor of the latter. By BGG Reciprocity (3.11),
M (wλ w0 · λ) occurs in any standard filtration of P (wλ wµ◦ · λ). Comparison
with the character formula in Theorem 7.12 proves our claim.

Proposition. Let λ, µ ∈ h∗ be antidominant and compatible, with µ\ in the

closure of the chamber containing λ\ . Denote by wλ the longest element in
W[λ] = W[µ] and by wµ◦ the longest element in Wµ◦ . Then Tµλ M (wλ · µ) ∼
=
P (wλ wµ◦ · λ). Moreover, the Verma modules occurring as quotients in a
146 7. Translation Functors

standard filtration of P (wλ · λ) are the M (wλ w · λ) with w ∈ Wµ◦ ; each has
multiplicity 1. 

To recover the special case treated in 4.10, take λ to be integral and

λ takes M (−ρ) = P (−ρ) to P (λ).
µ = −ρ: then T−ρ

7.14. Translation From a Wall

As before, we assume that λ and µ are antidominant and compatible. But
from now on we focus on the situation when λ is regular, so Wλ◦ = {1} and
λ\ lies in a chamber C of E(λ) = E(µ). We then require µ\ to lie in a single
wall of C, the intersection of C with the root hyperplane Hα for some α > 0.
Write s = sα , so Wµ◦ = {1, s}. Thanks to Corollary 7.12, for all w ∈ W[λ]
we have ch Tλµ Tµλ L(w · µ) = 2 ch L(w · µ). But since simple modules cannot
extend themselves nontrivially (Proposition 3.1), this forces
(1) Tλµ Tµλ L(w · µ) ∼
= L(w · µ) ⊕ L(w · µ).

Now consider w ∈ W[λ] for which wα > 0; thus `(ws) > `(w). In
particular, w · µ\ lies in the upper closure of w · C. Thanks to Theorem 7.9,
Tλµ L(w · λ) ∼ µ
= L(w · µ), whereas Tλ L(ws · λ) = 0.
In this special situation we want to describe some of the structure and
composition factor multiplicities of Tµλ M (w · µ) and its quotient Tµλ L(w · µ).
If we express ch L(w · µ) as a Z-linear combination of various ch M (w0 · µ)
and then apply Tµλ , Theorem 7.12 yields a similar (but more complicated)
expression for ch Tµλ L(w·µ) as a Z-linear combination of characters of Verma
modules. This can be used effectively in the proof of the following theorem.

Theorem. Let λ, µ be antidominant and compatible. Assume that λ is reg-

ular, so λ\ lies in a chamber C, while µ\ lies in a single wall of C corre-
sponding to α > 0. Suppose w ∈ W[λ] satisfies wα > 0, so `(ws) > `(w)
with s = sα .
(a) There is a short exact sequence
0 → M (ws · λ) → Tµλ M (w · µ) → M (w · λ) → 0.

(b) Hd Tµλ M (w · µ) ∼
= L(w · λ). In particular, Tµλ M (w · µ) is indecom-
posable and the above exact sequence nonsplit.
(c) The module Tµλ L(w · µ) is self-dual, with head and socle both iso-
morphic to L(w · λ); thus Tµλ L(w · µ) is indecomposable.
(d) [Tµλ L(w · µ) : L(w · λ)] = 2.
(e) [Tµλ L(w · µ) : L(ws · λ)] = 1.
7.14. Translation From a Wall 147

(f) Let w0 ∈ W[λ] with w0 · λ 6= w · λ. If [Tµλ L(w · µ) : L(w0 · λ)] > 0,

then w0 s · λ < w0 · λ. Moreover, Tλµ L(w0 · λ) = 0.
(g) If w0 ∈ W[λ] satisfies w0 s · λ < w0 · λ, then
ExtO (L(w · λ), L(w0 · λ)) ∼
= HomO (Rad Tµλ L(w · µ), L(w0 · λ)).

Proof. (a) Theorem 7.12 implies that

ch Tµλ M (w · µ) = ch M (w · λ) + ch M (ws · λ).
Since the translated module has a standard filtration (Corollary 7.6), it
follows from Proposition 3.7(a) that M (ws · λ) occurs as a submodule.
(b) By adjointness,
HomO Tµλ M (w · µ), L(w0 · λ) ∼ µ
= HomO M (w · µ), Tλ L(w0 · λ) ,

which by Theorem 7.9 is nonzero only when w0 · µ = w · µ (hence w0 = w

or w0 = ws) and w0 · µ\ lies in the upper closure of the chamber containing
w0 · λ\ . The upper closure requirement is met only by w0 = w.
(c) By exactness of Tµλ , the module Tµλ L(w·µ) is a quotient of Tµλ M (w·µ).
So the simple quotients of the former are also simple quotients of the latter.
Thanks to (b), only L(w · λ) occurs (and has multiplicity 1). Since L(w · µ)
is self-dual in O and translation commutes with duality (Proposition 7.1),
the module Tµλ L(w · µ) is also self-dual and therefore has socle L(w · λ).
(d) From (c) we get [Tµλ L(w · µ) : L(w · λ)] ≥ 1, with equality holding
just when Tµλ L(w · µ) ∼ = L(w · λ). Why is this impossible? Applying Tλ
µ

would yield Tλµ Tµλ L(w · µ) ∼

= L(w · µ), which contradicts (1) above.
On the other hand, suppose [Tµλ L(w · µ) : L(w · λ)] ≥ 3. Then Exercise
7.7 would force [Tλµ Tµλ L(w · λ) : L(w · µ)] ≥ 3, again contradicting (1).
(e) From (a) we see that L(ws · λ) occurs just once as a composition
factor of Tµλ M (w · µ), with a maximal vector of weight ws · λ generating a
submodule M isomorphic to M (ws · λ). So [Tµλ L(w · µ) : L(ws · λ)] ≤ 1. If M
were in the kernel of the surjection Tµλ M (w · µ) → Tµλ L(w · µ), there would
just be a single composition factor L(w · λ) in the latter module, contrary
to (d). So M has nonzero image under the surjection, forcing Tµλ L(w · µ) to
have a composition factor isomorphic to L(ws · λ).
(f) In case w0 = ws (which satisfies the hypothesis by (e)), we have
w0 s · λ = w · λ < ws · λ = w0 · λ. Then Theorem 7.9(b) implies that
Tλµ L(w0 · λ) = 0.
In the contrary case, w0 6= w, ws, so L(w0 ·µ) is not isomorphic to L(w·µ).
By hypothesis, [Tµλ L(w · µ) : L(w0 · λ)] > 0. Thanks to Exercise 7.7, this
implies [Tλµ Tµλ L(w · µ) : Tλµ L(w0 · λ)] > 0. But Tλµ L(w0 · λ) is either 0 or
isomorphic to L(w0 · µ). Since L(w0 · µ) is not isomorphic to L(w · µ), the
148 7. Translation Functors

second alternative is impossible by (1) above. In turn, Theorem 7.9 forces

w0 s · λ < w0 · λ.
(g) Conversely, assume that w0 s·λ < w0 ·λ, so Tλµ L(w0 ·λ) = 0 by Theorem
7.9. By adjointness, this implies that both HomO (Tµλ L(w · µ), L(w0 · λ)) and
ExtO (Tµλ L(w · µ), L(w0 · λ)) are 0. Apply the functor HomO (?, L(w0 · λ)) to
the short exact sequence
0 → Rad Tµλ L(w · µ) → Tµλ L(w · µ) → L(w · λ) → 0.
The resulting long exact sequence together with the vanishing results above
complete the proof. 
Exercise. Keep the hypotheses of the theorem. For all w0 ∈ W[λ] , prove
that [Tµλ L(w · µ) : L(w0 · λ)] ≤ 2[M (ws · λ) : L(w0 · λ)].

How much does the theorem reveal about the structure and composition
factors of M := Tµλ L(w · µ)? This module is self-dual, with L(w · λ) as
head and socle. There is one occurrence of the composition factor L(ws · λ).
Since w · λ < ws · λ, it follows from (a) that any composition factor L(w0 · λ)
different from these must satisfy w0 · λ < ws · λ. Moreover, each composition
factor L(w0 · λ) with w0 6= w (including L(ws · λ)) satisfies w0 s · λ < w0 · λ,
forcing Tλµ L(w0 · λ) = 0. This last point is already evident from (1) above:
Tλµ M ∼= L(w · µ) ⊕ L(w · µ). Beyond this we still know very little about the
potentially complicated structure of Rad M/ Soc M .
In the simplest nontrivial case, M just has Loewy length 3: If w = s,
then M (λ) = L(λ) and Tλµ L(λ) ∼ = L(µ) = M (µ). Then Tµλ L(µ) has only
L(λ) and L(s·λ) as possible composition factors. But in general the structure
of M cannot be worked out using the tools we have developed so far.

7.15. Wall-Crossing Functors

To round out this chapter, we introduce some important functorial language
whose significance will become clearer in the following chapters. In the
above setting, with λ regular antidominant and µ\ lying only in the s-wall
of the chamber containing λ\ , the functor Θs := Tµλ Tλµ may be thought of as
“translation across a wall”. Call it a wall-crossing functor on O or Oλ .
It is sometimes referred to as an operator of “coherent continuation”. (The
notation θs is somewhat more common in the literature; in any case, the use
of a greek symbol has become well-entrenched by now.)
To indicate explicitly the dependence of Θs on the choice of λ, one might
write Θλs . Note that Θs seems to depend not just on λ and s but also on the
choice of µ. One can prove that any other antidominant weight satisfying
the hypotheses on µ will produce an isomorphic functor. (There seems to
7.16. Self-Dual Projectives 149

be no direct argument for this, but Example 10.8 below indicates how to
derive it from the classification of “projective functors”.)
In Theorem 7.14 we considered Θs L(w · λ) = Tµλ L(w · µ), a module with
simple head and simple socle both isomorphic to L(w · λ). What remains
mysterious is the structure of Rad Θs L(w · λ)/ Soc Θs L(w · λ). In the Lie
group context where Zuckerman’s analogous translation functors appear,
Vogan [253] conjectured that this module should be semisimple, in which
case a recursive computation of characters would be possible. This will be
explored in more detail in 8.10 in the context of the Kazhdan–Lusztig Con-
jecture. (The subquotient of Θs L(w · λ) in question is sometimes referred to
as the “Jantzen middle”; he investigated a parallel semisimplicity conjecture
in the prime characteristic setting.)
From the adjoint property of translation functors we see that for all
M ∈ O,
HomO (Tλµ M, Tλµ M ) ∼ µ
= HomO (M, Tµλ Tλ M ).
In particular, the identity map on Tλµ M corresponds to an adjunction mor-
phism M → Θs M . (See Gelfand–Manin [111, II.3.24].) This morphism will
not in general be injective (for example when M 6= 0 but Tλµ M = 0); but
there is always a cokernel to consider:
M → Θs M → Shs M → 0.
In this way we introduce a “shuffling” functor Shs = Shλs on O, to be
explored further in Chapter 12.
Exercise. Let g = sl(2, C). For a fixed regular antidominant weight λ, set
µ := −ρ and let s be the nontrivial element of W . Work out the effect of
the functors Θs and Shs on the Verma modules and simple modules in Oλ ,
using some of the previous theory.

7.16. Self-Dual Projectives

In Chapter 4 we observed that a projective module P (λ) cannot be self-dual
unless λ is antidominant (Exercise 4.8). When λ is integral and antidomi-
nant, we showed directly in Theorem 4.10 that P (λ) is self-dual and has a
standard filtration involving each M (w·λ) exactly once. The proof amounted
to showing that P (λ) ∼ = T−ρλ M (−ρ). Nothing quite so straightforward can

be done if λ fails to be integral; but the self-duality remains true. This was
first shown by Irving [134], whose proof using translation and wall-crossing
functors is given here. (See the Notes below.)
Theorem. Let λ ∈ h∗ be antidominant. Then:
(a) P (λ) ∼
= P (λ)∨ .
150 7. Translation Functors

(b) Any standard filtration of P (λ) involves each M (w·λ) with w ∈ W[λ]
precisely once as a quotient. Equivalently, BGG Reciprocity implies
that [M (w · λ) : L(λ)] = 1 for all w ∈ W[λ] .

Proof. (1) First we observe that it is enough to prove the theorem when λ
is regular, which will allow us to apply Theorem 7.14 in that case. Assume
for the moment the theorem is true for all regular antidominant λ.
If µ is an arbitrary antidominant weight, subtracting nρ from it for large
enough n will obviously produce a compatible regular antidominant weight
λ. Since P (λ) is self-dual by assumption, Tλµ P (λ) is also projective and
self-dual by Proposition 7.1. We claim that this module is nonzero. Thanks
to Corollary 4.8, Soc P (λ) is a direct sum of copies of L(λ). If L(λ) is such a
submodule, exactness of Tλµ insures that Tλµ L(λ) is a submodule of Tλµ P (λ).
Since µ\ lies in the upper closure of the chamber containing λ\ , Theorem 7.9
shows that Tλµ L(λ) ∼ = L(µ) and is therefore nonzero.
µ
Now Tλ P (λ) is isomorphic to a direct sum of n > 0 indecomposable
projectives P (w · µ) with w ∈ W[µ] = W[λ] . Each P (w · µ) has head L(w · µ)
and contains L(µ) in its socle. Accordingly Hd Tλµ P (λ) is the direct sum
of the n modules L(w · µ) while Soc Tλµ P (λ) contains the direct sum of n
copies of L(µ). But Tλµ P (λ) is self-dual, so its head and socle are isomorphic.
It follows that Tλµ P (λ) is the direct sum of n copies of P (µ). Self-duality
further implies that Tλµ P (λ) is the direct sum of n copies of P (µ)∨ . Since
these modules are indecomposable, they must all be isomorphic to P (µ).
Part (b) of the theorem for P (µ) then follows easily from the correspond-
ing fact for P (λ).
(2) From now on λ is both regular and antidominant. In particular,
L(λ) = M (λ) is self-dual and a Verma module. It also occurs as the socle
of each M (w · λ). If wλ is the longest element in W[λ] , it follows from the
definition that the weight λ◦ := wλ · λ is dominant. So M (λ◦ ) = P (λ◦ ) is
projective (by Proposition 3.8(a)) as well as a Verma module.
The objective now is to pass from M (λ◦ ) = P (λ◦ ) to P (λ) using wall-
crossing functors Θs . Fix a reduced expression wλ = s1 . . . sn with si ∈ ∆[λ]
and set Θ := Θs1 ◦· · ·◦Θsn . Note that the exactness of Θ and the embedding
L(λ) ⊂ M (λ◦ ) yield an embedding ΘL(λ) ⊂ ΘM (λ◦ ).
(3) Since λ is regular, we can take advantage of the precise way in which
a wall-crossing functor Θs applies to a Verma module M (w · λ) in Oλ such
as M (λ◦ ) or M (λ) = L(λ). Taking into account Theorem 7.6, we see from
Theorem 7.14(a) that Θs M (w · λ) is a nonsplit extension of the two Verma
modules M (w · λ) and M (ws · λ) when ws > w. Here L(w · λ) occurs just
once as a composition factor of each Verma module.
7.16. Self-Dual Projectives 151

(4) From parts (d) and (f) of Theorem 7.14, we see that Θs L(w · λ)
is either 0 or has just one composition factor isomorphic to L(λ) when
w 6= 1, while Θs L(λ) has two composition factors of this type (one in each
of the Verma modules produced by wall-crossing). In particular, the only
composition factors of type L(λ) in ΘM (λ◦ ) are produced by applying wall-
crossing functors to L(λ). In view of step (3), it follows that

(∗) [ΘM (λ◦ ) : L(λ)] = [ΘL(λ) : L(λ)].

(5) Iteration of wall-crossings, starting at either M (λ) or M (λ◦ ), shows

that each resulting Verma module M (w · λ) has the composition factor L(λ)
only once (and then it is the socle). Moreover, it is easy to check that any
w ∈ W[λ] can occur here: write wλ = ww0 with `(wλ ) = `(w) + `(w0 ) and
choose a corresponding reduced expression for wλ .
(6) Since λ◦ is dominant, M (λ◦ ) is projective. So ΘM (λ◦ ) is also projec-
tive (Proposition 7.1). By Theorem 7.14(a), each wall-crossing functor Θs
takes a Verma module M (w · λ) to a nonsplit extension involving the two
Verma modules M (w · λ) and M (ws · λ). Therefore the projective module
ΘM (λ◦ ) acquires in n steps a standard filtration involving the final Verma
module M (λ) = L(λ) just once. Thanks to BGG Reciprocity, P (λ) occurs
once as a direct summand of ΘM (λ◦ ).
(7) On the other hand, ΘL(λ) is self-dual (because L(λ) is). Since
L(λ) ⊂ Soc ΘL(λ), we get L(λ) ⊂ Hd ΘL(λ). From (∗) it follows that
the image of ΘL(λ) ⊂ ΘM (λ◦ ) under the projection ΘM (λ◦ ) → P (λ) →
L(λ) cannot be zero. So the image of ΘL(λ) in P (λ) cannot be a proper
submodule (lest it lie in the radical). Using the projective property, we get
P (λ) ⊂ ΘL(λ).
(8) In turn, the self-duality of ΘL(λ) implies that the injective module
Q(λ) = P (λ)∨ ⊂ ΘL(λ) ⊂ ΘM (λ◦ ). Being injective, Q(λ) must be a
direct summand of ΘM (λ◦ ); the latter being projective, Q(λ) must also
be projective (and indecomposable). In particular, its head is simple and
therefore isomorphic to the simple submodule L(λ) of its dual P (λ). In
other words, P (λ)∨ ∼= P (λ), which proves (a).
(9) For part (b) of the theorem, we use the fact developed in step 5):
(each M (w · λ) has L(λ) as a composition factor exactly (once. This implies,
by BGG reciprocity, that all M (w · λ) with w ∈ W[λ] occur just once in each
standard filtration of (P (λ). 

Exercise. It is instructive to follow the details of the proof in the case g =

sl(3, C), with λ integral. There are two reduced expressions w◦ = s1 s2 s1 =
s2 s1 s2 , each leading to different standard filtrations of ΘL(λ) and ΘL(λ◦ ).
The latter module is a direct sum of P (λ) and one other indecomposable
152 7. Translation Functors

projective. Both ΘL(λ) and ΘL(λ◦ ) have standard filtrations involving 8

Verma modules, with 6 distinct ones involved in P (λ) in each case.

Notes
(7.1–7.4) Translation functors were extensively developed by Jantzen, in
both the category O setting and the parallel characteristic p theory (where
W is replaced by an affine Weyl group). See [148, Kap. 2] as well as later
refinements for algebraic groups in [153, II.7]. (Independently, Zuckerman
[262] worked out similar ideas for the category of Harish-Chandra modules.)
Around the same time a broader view of some of the ideas involved in trans-
lation was developed by Bernstein–Gelfand [24] under the rubric “projective
functors”; this will be discussed in Chapter 10.
The argument in Lemma 7.5 originates in Jantzen’s paper [146, Satz 5].
For the equivalence of blocks in 7.8 we follow Jantzen [148, 2.15, Bem. 2)];
he refers to similar arguments used by Zuckerman [262] in the setting of
Harish-Chandra modules. The special case involving regular integral blocks
is formulated by BGG in [27, Thm. 4].
The arguments in 7.9 are analogues of those in Jantzen [153, II.7.15].
(For his earlier approach in O, see [148, 2.11–2.13].)
(7.11) The proof of the theorem differs from Jantzen’s [148, 2.24], which
is somewhat longer and relies on a formal character comparison, together
with BGG Reciprocity.
(7.12– 7.15). These results are drawn from Jantzen [148, Chap. 2] and
[153, II.7]. The terminology “wall-crossing functor” does not appear ex-
plicitly in Jantzen [148] but was soon in common use and appears in [153,
II.7.21] as well as [153, II.C], where Lusztig’s algebraic group analogue of
the Kazhdan–Lusztig Conjecture is explained in detail.
The proof of Proposition 7.13 was suggested by Jantzen.
In unpublished notes [133, 5.3], Irving first worked out a lengthy proof
of Theorem 7.16, which he later streamlined in [134, §3] in the context of
a more general conjecture for parabolic category O. In other unpublished
notes, Joseph [161, 3.13] then shortened the argument further by taking
advantage of the work of Bernstein–Gelfand [24] on projective functors.
 Chapter 8

Kazhdan–Lusztig
Theory

Following Verma’s 1966 thesis and the realization that the composition fac-
tor multiplicities of Verma modules could not as readily be pinned down as
he expected at first, a considerable amount of effort went into finding an
algebraic solution of the problem: work of BGG, Deodhar, Jantzen, Joseph,
and others. In the previous chapters we have seen how to refine the problem,
but not how to solve it completely except in isolated cases.
The landmark 1979 paper of Kazhdan–Lusztig [171] proposed a precise
conjecture about the multiplicities for the principal block of O, in terms of
the values at 1 of certain polynomials attached to the Iwahori–Hecke algebra
of W . Following Deodhar’s suggestion in [81], we refer to the developments
inspired by [171] as “Kazhdan–Lusztig theory” (or KL theory, for short).
Their approach was motivated in part by geometric questions involving the
flag variety and unipotent variety associated with g, as we shall recall below.
New ideas coming from geometry, topology, and analysis were already in the
air at the time and pointed quickly to a proof of the conjecture.
After reviewing briefly in 8.1 what was known by 1979, we outline the
construction of Kazhdan–Lusztig polynomials (8.2–8.3) before stating the
Kazhdan–Lusztig Conjecture formally (8.4). Its relationship with singu-
larities of Schubert varieties and intersection cohomology is then explored
in 8.5–8.6, followed by a comparison with Jantzen’s earlier criterion for a
composition factor of a Verma module to have multiplicity 1 (8.7).
Independent proofs of the conjecture were published in 1981 by Beilinson–
Bernstein [16] and Brylinski–Kashiwara [55], based on similar techniques

153
154 8. Kazhdan–Lusztig Theory

involving D-modules and intersection cohomology. In 8.8 we can only intro-

duce this literature, which requires ideas far beyond the scope of the earlier
algebraic study of category O.
The KL Conjecture leads to structural information about the modules
involved, not just multiplicity data, as suggested by an equivalent formula-
tion due to Vogan in terms of Extn functors together with a deeper study of
wall-crossing functors (8.10–8.11). It turns out that the coefficients of KL
polynomials yield information about extensions of simple modules and the
layer structure of naturally occurring Loewy filtrations (8.14–8.15). A re-
finement due to Beilinson–Bernstein of the geometric methods used to prove
the KL Conjecture ultimately settles the Jantzen Conjecture as well (8.13).
This chapter marks a shift from “textbook” mode to “survey” mode, in
which we no longer attempt to provide full details of proofs but rather guide
the reader into the more advanced literature. In line with this goal, we often
focus just on integral weights, starting with the principal block.
Notational convention: When x ≤ w in W , the difference `(w) − `(x)
occurs frequently; it will be abbreviated to `(x, w).

8.1. The Multiplicity Problem for Verma Modules

To put things in perspective, recall the discussion in 7.10, where we saw
that translation functors can be used for integral weights to reduce the
computation of formal characters of simple modules (or composition factor
multiplicies of Verma modules) to a single regular block such as the prin-
cipal block O0 . Together with the BGG–Verma Theorem (5.1), this shifts
attention to the role of W and its Bruhat ordering. But the substantial
algebraic progress made on the problem in the 1970s fell short of reaching
a complete solution. The Kazhdan–Lusztig Conjecture [171, Conj. 1.5] of-
fered a conceptual (and computable) interpretation of multiplicities for O0 ,
based on a new set of ideas.
On the other hand, the treatment of arbitrary weights λ ∈ h∗ requires
further work, even though Jantzen correctly foresaw the leading role to be
played by the reflection group W[λ] . A basic reduction, due to Soergel [237,
Thm. 11], will be explained in 13.13.
Before taking up Kazhdan–Lusztig theory we review briefly the progress
actually made before 1979. Some low rank computations were done inde-
pendently by several people. But the most definitive results are due to
Jantzen in [148] and earlier papers, based on algebraic tools he developed:
contravariant forms, Jantzen filtration and sum formula, translation func-
tors, together with other methods we have not discussed related to primitive
ideals of U (g) and associated varieties of modules. The results are of two
8.1. The Multiplicity Problem for Verma Modules 155

types: (1) Computation of all multiplicities when Φ (or sometimes Φ[λ] ) in-
volves just irreducible components of rank ≤ 3 or types A4 , D4 . These can
now be reproduced more routinely by computations of the type shown in
Goresky’s tables [115]. (2) More general qualitative or recursive formula-
tions concerning the multiplicities. Here is a quick summary of both types:
• In the background are the early examples found by BGG [25] and
Conze–Dixmier [72], exhibiting weights for which [M (λ) : L(µ)] >
1; here the weights are dominant but singular. Deodhar–Lepowsky
[83] then showed that in rank ≥ 3 there always exist similar exam-
ples involving regular dominant weights.
• To work out all multiplicities for types A2 , B2 , G2 is already some-
what challenging: for A2 , Exercise 5.3 indicates how to fill in details
for integral weights by using the Jantzen sum formula. Jantzen
himself treated the other rank 2 cases and found in general that
for blocks involving a group W[λ] of rank 2, all composition factor
multiplicities of Verma modules are 1.
• Jantzen [148, 5.24] was able to determine all multiplicities in ranks
up to 3 as well as in type A4 . He remarked also that he could use
recent methods of Vogan to complete the case D4 . This and his
more general results prompted him to observe that in all examples
treated so far, multiplicities depend essentially just on the group
W[λ] (and the root system Φ[λ] ) and to ask whether this is true in
general.
• The role of Weyl groups was exhibited most generally by Jantzen
[148, 4.13–4.14] as follows, for weights which are not necessarily
integral: Suppose λ, µ are regular, antidominant, compatible, while
all components of Φ[λ] are of classical type A` –D` ; then λ, µ have
the “same multiplicities”: for all w, w0 ∈ W ,

[M (w · λ) : L(w0 · λ)] = [M (w · µ) : L(w0 · µ)].

• In a similar spirit he used his full arsenal of methods to find, for any
g and arbitrary λ, µ ∈ h∗ satisfying µ ↑ λ, necessary and sufficient
conditions for [M (λ) : L(µ)] = 1 [148, 5.20]. (This will be revisited
in 8.7.) The conditions here are essentially recursive in nature: for
all ν ∈ h∗ such that µ ↑ ν ↑ λ but ν 6= λ, we have [M (ν) : L(µ)] = 1
together with

|Φ+ + +
λ | − |Φµ | = |{α ∈ Φλ | µ ↑ sα · λ}|.

From this one recovers by an easy induction: [M (λ) : L(µ)] = 1

whenever µ ↑ λ and µ is antidominant.
156 8. Kazhdan–Lusztig Theory

From the early work emerged a strong sense that a full determination
of composition factor multiplicies in Verma modules would depend mostly
(perhaps entirely) on the reflection group W[λ] and would require a recursive
approach rather than a closed formula for each pair of weights. The correct
framework is provided by Kazhdan–Lusztig theory.

8.2. Hecke Algebras and Kazhdan–Lusztig Polynomials

Here is a concise outline of the elementary algebraic development due to
Kazhdan–Lusztig [171], which leads to the introduction of their special
polynomials. Our notation, following Humphreys [130, Chap. 7], differs
slightly from theirs. Everything here can be done uniformly for an arbitrary
Coxeter group (without reference to Lie theory), but our immediate concern
is just with the Weyl group W attached to g.
Inspired by earlier work of Iwahori on representations of finite Chevalley
groups, one starts with a deformation H of the integral group ring of W
using an indeterminate q as parameter. Setting q = 1 recovers ZW . Call H
the Hecke algebra (or Iwahori–Hecke algebra) of W . More precisely, H
is defined by generators and relations over the ring of Laurent polynomials
A := Z[q, q −1 ], the generators Tw being in bijection with elements of W . For
example, when s is a simple reflection, Ts2 = (q − 1)Ts + qT1 , with T1 acting
as the identity in the ring structure on H.
The elements Tw turn out to be an A-basis of H. Moreover, these are
units in the ring H:
X
(Tw−1 )−1 = (−1)`(w) q −`(w) (−1)`(x) Rx,w (q)Tx .
x≤w

Here we use the Bruhat ordering of W (which is seen to be unavoidable!),

while Rx,w ∈ Z[q] is a recursively computable monic polynomial of degree
`(x, w) = `(w) − `(x), with Rw,w (q) = 1. For example, the relation quoted
above leads to
Ts−1 = q −1 Ts − (1 − q −1 )T1 .

The next ingredient is an involution of H, denoted by bar, which inter-

changes q with q −1 in the ring A and sends Tw to (Tw−1 )−1 . The objective
is to locate a new basis of H indexed by W consisting of elements Cw which
satisfy Cw = Cw . This requires working with a square root of q, as seen in
the formula
1
Cs := q − 2 (Ts − qT1 ).
1
So the base ring A is replaced by the ring of Laurent polynomials in q 2 and
1
q − 2 . Now one can prove:
8.3. Examples 157

Theorem (Kazhdan–Lusztig). There are unique elements Cw (w ∈ W )

fixed by the involution on H and satisfying the conditions:
X
Cw = (−1)`(w) q `(w)/2 (−1)`(x) q −`(x) Px,w (q)Tx ,
x≤w

where Pw,w (q) = 1, Px,w (q) ∈ Z[q], and deg Px,w (q) ≤ (`(x, w) − 1)/2 when-
ever x < w.

The “change of basis” polynomials Px,w (q) are the Kazhdan–Lusztig

polynomials, or KL polynomials for short. The cases when x < w and
deg Px,w (q) = (`(x, w)−1)/2 are especially interesting, as illustrated for type
C3 in 8.6 below; write x ≺ w when this happens. One can set Px,w (q) = 0
in case x 6≤ w. It is conjectured in [171] (for arbitrary Coxeter groups) that
all coefficients of these polynomials are nonnegative; this is still an open
question in general. In our case the discussion in 8.5 will interpret these
coefficients as dimensions of certain vector spaces.
We remark that Soergel [239] provides a streamlined development of
these ideas, avoiding the R-polynomials entirely; his approach is motivated
in part by the failure so far to find a straightforward interpretation of Rx,w (q)
in the framework of category O (Remark 8.11).

8.3. Examples
First we quote several elementary facts which follow directly from the alge-
braic theory in [171]:
(a) For all pairs x ≤ w in W , we have Px,w (0) = 1. Combined with
the degree bound for KL polynomials, this shows that Px,w (q) = 1
whenever `(x, w) ≤ 2. (See for example [130, Exer. 7.11]). These
statements hold for arbitrary Coxeter groups.
(b) In case W is finite, as in the case of a Weyl group, we have Px,w◦ (q) =
1 for all x ∈ W . (See [130, Exer. 7.14].)
(c) If W is a dihedral group, such as the Weyl group in types A2 , B2 , G2 ,
then Px,w (q) = 1 for all x ≤ w. (See [130, 7.12(a)].)
While KL polynomials are recursively computable, a computer is soon
needed: see [130, 7.12] for further discussion and references as of 1992. The
combinatorial aspects of KL polynomials continue to be explored by many
people in the wider context of Coxeter groups: see Brenti [47], Björner–
Brenti [29, Chap. 5]. Here we sketch one small and one very large example.
(1) The first case in which KL polynomials of degree > 0 actually occur
is W = S4 , the Weyl group of the Lie algebra sl(4, C) of type A3 . Here
Px,w (q) = 1 + q for two pairs satisfying x ≺ w [171, 6.1]; in turn, Py,w (q) =
1 + q holds whenever y < x. (This is confirmed by Goresky’s machine
158 8. Kazhdan–Lusztig Theory

computations posted on his Web page [115].) Using the standard numbering
1, 2, 3 of vertices in the Coxeter–Dynkin diagram, the two pairs are

s2 < s2 s1 s3 s2 and s1 s3 < s1 s3 s2 s3 s1 .

It is interesting to compare the diagram of the Bruhat ordering of S4 given

by Björner–Brenti [29, Fig. 2.4], where permutation labelling (abcd) is used.
It is not difficult to locate the two special pairs. For example, the first pair
is represented by the permutations (1324) and (3412).
(2) As |W | increases, the difficulty of carrying out explicit computations
of KL polynomials by a naive application of the original algorithm becomes
evident. Among the exceptional types, E8 illustrates well the challenge
involved. Here |W | = 214 35 52 7 = 696, 729, 600.
A smaller, but still daunting, version of the problem for E8 arises in
the study of unitary representations of real semisimple Lie groups. This is
explained informally by Vogan [256], following the successful computation
for the split real group of type E8 of related “Kazhdan–Lusztig–Vogan poly-
nomials” attached to a geometrically defined H-module. These polynomials
encode essential character information and can be computed by an algo-
rithm parallel to the original algorithm in [171]: see Lusztig–Vogan [204],
Vogan [255]. The E8 computation involved a team of Lie group specialists,
including F. du Cloux and others with special programming skills. (Some
algebraic background for Vogan’s program is described below in the complex
case: see 8.10–8.11.)
Whereas the KL polynomials are correlated with the geometry of the
complex flag variety (8.5), notably the |W | orbits of a Borel subgroup, the
KLV polynomials involve a smaller number of orbits on this variety under
the action of a (complexified) maximal compact subgroup of the real group.
In the case at hand, this number is still impressively large: 320, 206. In fact,
nontrivial local systems have to be paired with many orbits, for a total of
453, 060 pairs. Many reductions of the KLV algorithm are needed in order
to make the computation feasible. Specific results include:

• The actual number of distinct KLV polynomials for the R-split

group of type E8 is 1, 181, 642, 979.
• The largest coefficient of any KLV polynomial is 11, 808, 808, while
the value of this polynomial at 1 is 60, 779, 787.
• So far the largest value at 1 found for any KLV polynomial for E8
is 62, 098, 473.
8.4. Kazhdan–Lusztig Conjecture 159

8.4. Kazhdan–Lusztig Conjecture

In their 1979 paper, Kazhdan and Lusztig formulated a precise conjecture
[171, Conj. 1.5] for the formal characters of simple modules in the principal
block O0 of O. Here there are |W | simple modules. Since the antidominant
Verma module M (−2ρ) is simple, it is natural to start with its highest weight
λ := −2ρ and parametrize modules as Mw := M (w · λ), Lw := L(w · λ).
With this notation, we know that [Mw : Lx ] 6= 0 precisely when x ≤ w in
the Bruhat ordering (5.2).
Conjecture (Kazhdan–Lusztig). Consider the principal block O0 of O,
with simple modules Lw and Verma modules Mw indexed as above by w ∈ W .
Then the composition factor multiplicities of Verma modules are determined
in K(O0 ) by the equations
X
(1) [Lw ] = (−1)`(x,w) Px,w (1) [Mx ].
x≤w
In particular, the multiplicities depend only on W .

Thanks to a sort of “inversion” formula for the KL polynomials attached

to finite Coxeter groups ([171, §3]), the KL conjecture is easily seen to be
equivalent to:
X
(2) [Mw ] = Pw◦ w,w◦ x (1) [Lx ], or [Mw : Lx ] = Pw◦ w,w◦ x (1).
x≤w

Notice that two extreme cases are consistent with the conjecture, in
view of the facts noted above in 8.3: On one hand, the antidominant Verma
module M (λ) is already simple, in agreement with P1,1 (q) = 1. On the
other hand, the classical Weyl–Kostant formula developed in Chapter 2
corresponds to the case of L(w◦ · λ). Here all Px,w◦ (q) = 1 and these occur
with alternating signs in the conjectured formula above as they do in the
classical formula.
One striking consequence of the conjecture is that the numerical data
involved must be the same for the Lie algebras of types B` and C` : these
have the same Weyl group, even though the Lie algebras are nonisomorphic.
Exercise. Let g = sl(4, C). Deduce from the inverse formulation (2) of the
conjecture together with 8.3(1) that a Verma module such as M (0) whose
highest weight is regular, dominant, and integral must have precisely 22
composition factors of multiplicity 1 and two of multiplicity 2.

Though the KL Conjecture looks at first somewhat opaque, the formu-

lation does bring W to the forefront. Kazhdan and Lusztig were actually
motivated at first by the deep work of Springer on Weyl group represen-
tations in connection with the unipotent variety of a semisimple algebraic
160 8. Kazhdan–Lusztig Theory

group G over C having Lie algebra g. The individual unipotent classes are
smooth (as group orbits) but their closures sometimes have singularities.
A convenient desingularization involves the flag variety G/B, where B has
Lie algebra b. This variety in turn has a nice cell decomposition, but here
too the closures of cells may be singular. The geometry of the flag variety,
and indirectly the Springer representations, turn out to be intimately re-
lated with KL polynomials, as we discuss next. It is somewhat miraculous
that these geometric ideas lead to a solution of the multiplicity problem for
Verma modules.

8.5. Schubert Varieties and KL Polynomials

As promised, we outline briefly one geometric interpretation of KL polyno-
mials; see Kazhdan–Lusztig [172] (as well as the 1980 account by Gelfand–
MacPherson [110], written just before the proof of the KL Conjecture).
Let G be a simply connected semisimple algebraic (or Lie) group over
C whose Lie algebra is isomorphic to g. Fix a Borel subgroup B relative to
Φ+ , with a maximal torus T . Then the Weyl group W of Φ is isomorphic to
NG (T )/T . Corresponding to this set-up is the Bruhat decomposition G =
S
w∈W BwB (disjoint union), where the double coset BwB is independent
of the choice of coset representative for w in NG (T ).
Write X := G/B, the flag variety. This is a smooth projective variety
of dimension m = |Φ+ |, which inherits from the Bruhat decomposition a de-
composition into Bruhat cells Xw := BwB/B. Here Xw is isomorphic to an
affine space of dimension `(w). The closure Xw is called a Schubert variety.
It is the union of Xw and some smaller Bruhat cells, giving a cell decompo-
sition. Indeed, Chevalley showed that Xx ⊂ Xw if and only if x ≤ w in the
Chevalley–Bruhat ordering; this is where the ordering originated. Schubert
varieties include many special cases of interest, including flag varieties, but
are not always smooth.
The cell decomposition of Xw makes its classical cohomology easy to
compute: Since the real dimension of a cell Xx is 2`(x), all cohomology
groups in odd degrees vanish while the even Betti number dim H 2i (Xw ) is
the number of x ≤ w for which `(x) = i. These data combine to give the
Poincaré polynomial
`(w)
X
Pw (q) := dim H 2i (Xw )q i .
i=0

In [171] (and its Appendix) it was already observed that there is a close
connection between the failure of local Poincaré duality for Xw and the fact
that certain associated Px,w (q) with x < w fail to equal 1. This occurs in
8.6. Example: W of Type C3 161

case Xw is not “rationally smooth” along some cells. Examples are seen, for
example, in two Schubert varieties when g = sl(4, C) [171, 1.6]; as noted
earlier, there are two w ∈ W for which not all Px,w (q) are trivial.
In [172] Kazhdan and Lusztig went on to interpret their polynomials in
terms of the newly perfected Deligne–Goresky–MacPherson theory of inter-
section cohomology and the Beilinson–Bernstein–Deligne theory of perverse
sheaves on singular complex varieties (to which Gabber made important
contributions). We use the notation IH for the associated intersection coho-
mology groups. (In [172] they actually work with varieties over an algebraic
closure of Fp rather than over C, but the cohomology calculations are done
in characteristic 0.)
In this framework the KL polynomials turn out to be the Poincaré poly-
nomials for local intersection cohomology of Schubert varieties. If x ≤ w
the polynomial is 1 precisely when Xw is rationally smooth along Xy for
all x ≤ y ≤ w. For Schubert varieties in general, intersection cohomology
vanishes in odd degrees; this is consistent with the fact that KL polynomials
1
involve the variable q rather than q 2 .
Theorem (Kazhdan–Lusztig). For each x ≤ w in W , we have
`(w)
X
Px,w (q) = q i dim IHx2i (Xw ).
i=0

In particular, all coefficients of Px,w (q) are nonnegative.

The global Poincaré polynomial for intersection cohomology is given

explicitly by:
`(w)
X X
dim IH 2i (Xw )q i = q `(x) Px,w (q).
i=0 x≤w
Unlike the usual Poincaré polynomial, this one always exhibits duality (orig-
inating in Verdier duality): the coefficients of q i and q `(w)−i agree.

8.6. Example: W of Type C3

It is instructive to look at the case when W is of type C3 , using Goresky’s
tables [115] to track a couple of cases where singularities occur in Schubert
varieties. (In this algebraic context, we use “singularity” as a shorthand for
the failure of rational smoothness, ignoring other topological singularities.)
Note that for C3 and some other cases, his numbering of vertices in the
Coxeter–Dynkin graph of W differs from that in Bourbaki [45]. Here his
labels 1, 2, 3 are the reverse of the standard ones, so s1 s2 has order 4 while
s2 s3 has order 3 and s1 s3 = s3 s1 .
162 8. Kazhdan–Lusztig Theory

For type C3 (or B3 , where W is the same), we have |W | = 48. The

elements are listed by indices 1–48, compatible with increasing length; the
omitted index 1 refers to 1 ∈ W . The length of each element is given, to-
gether with its “code”: for example, 213 denotes the reduced decomposition
w = s2 s1 s3 . (When many reduced decompositions of a single w occur, one
gets into a typical computational problem in constructing such tables.) The
indices of elements of length `(w) − 1 which lie below w in the Bruhat order-
ing (corresponding to cells of codimension 1 in Xw ) are listed; this permits
a recursive determination of all x < w. From this in turn one can compute
the Betti numbers, which are given explicitly in the table.
In many cases the Schubert varieties are (rationally) smooth. But when
singularities occur along certain cells, this is indicated together with the
associated KL polynomials. Take for example w with index 47; its length
is 8 and its code is 21232123. The Betti numbers are 1, 3, 5, 7, 8, 8, 6, 3, 1,
corresponding to 42 cells in Xw belonging to 42 elements x ≤ w.
There is a singularity of minimal codimension along the 3-dimensional
cell represented by x ∈ W with index 15 and code 232 (the longest element
in a subgroup of type A2 ). Its closure is smooth and is in fact isomorphic
to the flag variety of SL(3, C). The KL polynomial is Px,w (q) = 1 + q 2 ,
abbreviated in the table by 1 0 1. Moreover, Xw has singularities along
the 6 cells Xy lying in Xx , with corresponding KL polynomials Py,w = 1+q 2 .
(These are not listed individually but are easily recovered from the Bruhat
graph.) Adding up all KL polynomials Pz,w with z ≤ w, one gets the
coefficients of the intersection cohomology Poincaré polynomial, called IH
in the tables: 1, 3, 6, 9, 10, 9, 6, 3, 1. Note that the total difference between
these dimensions and the Betti numbers is 6, agreeing with the number of
z < w for which Pz,w = 1 + q 2 .
Things get more complicated when unrelated singularities occur. Take
for example w with index 43 and Betti numbers 1, 3, 5, 7, 8, 7, 4, 1. There is
a singularity along each of two 4-dimensional cells with indices 17 and 22; in
both cases the KL polynomial is 1+q, which propagates to lower elements in
the Bruhat ordering until one reaches a different singularity at the element
of index 3 (the simple reflection s2 ). Here and also at the identity element
of W the KL polynomial changes to 1 + q + q 2 . The IH polynomial then
has coefficients 1, 4, 9, 13, 13, 9, 4, 1.

8.7. Jantzen’s Multiplicity One Criterion

It is instructive to compare the KL Conjecture with the known multiplicity
results summarized in 8.1. We already indicated in 8.3 that small rank
cases computed earlier agree with results predicted by the conjecture. A
more interesting example is the general criterion developed by Jantzen for
8.7. Jantzen’s Multiplicity One Criterion 163

[M (λ) : L(µ)] = 1. Its proof required a considerable amount of technical

work. But the criterion itself (for the principal block) can be recovered
from the KL Conjecture: this requires elementary combinatorial arguments
together with the fact that KL polynomials have nonnegative coefficients
when W is finite (8.5). On the other hand, the proof of the conjecture
involves deep methods drawn from other parts of mathematics.
Here is an outline of steps involved, based on an unpublished letter from
Lusztig to Jantzen early in 1980 as well as the paper of Gabber–Joseph [107]
(which appeared in preprint form later in 1980). The reader is encouraged
to fill in the omitted details.
(1) The key lemma involves further development of the combinatorics
of the auxiliary polynomials Rx,w (q) for x ≤ w (8.2). Here Rw,w (q) = 1,
while the recursive recipe by which the polynomials are computed shows
that Rx,w (q) = q − 1 in case `(w) − `(x) = 1 (which means w = rx for some
0 (1) = 1.
reflection r). Using prime to denote derivative, this case gives Rx,w
Lemma. Let x < w in W . Then
(
0 1 if `(w) − `(x) = 1
Rx,w (1) =
0 otherwise.

The proof, given in Gabber–Joseph [107, 2.2], uses induction on `(w)

together with the recursive definition of R-polynomials in Kazhdan–Lusztig
[171, 2.0.b–c]. Separate arguments are needed when s is a simple reflection
such that sw < w, to deal with the cases sx < x and sx > x.
(2) Now combine the lemma with a fundamental identity relating R-
polynomials and KL polynomials for fixed x ≤ w (see [171, §2], [172,
(2.1.5)], [107, 2.3], or [130, 7.10(20)]):
X
Rx,y (q)Py,w (q) = q `(x,w) Px,w (q −1 ).
x≤y≤w

Just differentiate formally, set q = 1, and apply the lemma, to conclude:

Corollary. If x < w in the Weyl group W , then
X
0
2Px,w (1) = `(x, w)Px,w (1) − Prx,w (1),
x<rx≤w

where r runs over all reflections in W .

(3) So far the arguments use only elementary combinatorics of the poly-
nomials, but now we have to know that all coefficients of KL polynomials
are nonnegative. For our Weyl group W , this follows from the geometric
interpretation in 8.5. In particular, the statement Px,w (1) = 1 is equivalent
to Px,w (q) = 1, since all KL polynomials have constant term 1.
164 8. Kazhdan–Lusztig Theory

If x < w, write L := `(x, w) and let R be the number of reflections r for

which x < rx ≤ w. Then we claim: Let x < w, and suppose that Py,w (q) =
1 whenever x < y ≤ w. Then Px,w (q) = 1 if and only if R = L.
By assumption, Prx,w (q) = 1 for each of the R reflections r such that
x < rx ≤ w, so step (2) yields:
0
2Px,w (1) = LPx,w (1) − R.
Write Px,w (q) = i ci q i , where 0 ≤ i ≤ (L − 1)/2 and all ci ≥ 0. Then
P
differentiate and substitute to get
X X
2 ici = L ci − R,
i>0 i
or X
(L − 2i)ci = R − L.
i
Each term on the left side is nonnegative, so R − L ≥ 0, while R = L if and
only if ci = 0 for all i ≥ 1, or if and only if Px,w (q) = 1.
With this formalism in hand and assuming the truth of the KL Con-
jecture, we can recover Jantzen’s multiplicity one theorem [148, 5.20] for
the principal block O0 . The general statement given at the end of 8.1 has
to be adapted first, using the notation in 8.4: write λ = w · (−2ρ) and
µ = x · (−2ρ), where x < w. Thanks to Exercise 5.3, |Φ+ λ | = `(w) and
|Φµ | = `(x), so the difference is `(x, w) = L. Recall that α ∈ Φ+
+
λ means
sα · λ < λ, which translates into rw < w if r = sα . The number of such
reflections for which µ ↑ sα ·λ is then the number of r for which x ≤ rw < w.
Theorem (Jantzen). In O0 , we have [Mw : Lx ] = 1 for x < w if and only
if the following two conditions are satisfied:
(a) For all y ∈ W such that x ≤ y < w, we have [My : Lx ] = 1.
(b) `(x, w) is the number of reflections r for which x ≤ rw < w.
For the proof we assume the truth of the KL Conjecture, in the inverse
format 8.4(2): [Mw : Lx ] = Pw◦ w,w◦ x (1). Equivalently, since all coefficients
of the polynomial are nonnegative in this case, Pw◦ w,w◦ x (q) = 1.
Given (a) and (b), we want to deduce from the steps above that this
value is 1. For this, set w0 := w◦ w and x0 := w◦ x. Using the general fact
`(w◦ z) = `(w◦ ) − `(z) (0.3), we get `(w0 , x0 ) = `(x, w). Now r0 := w◦ rw◦
runs over all reflections as r does. Combined with the fact that y < z if
and only if w◦ z < w◦ y, this allows us to translate x ≤ rw < w in (b) into
w0 < r0 w0 ≤ x0 . From step (3) above, we obtain Pw0 ,x0 (q) = 1, which by the
KL Conjecture proves that [Mw : Lx ] = 1.
For the reverse implication, we have to invoke Verma’s Theorem on
embeddings of Verma modules. Given [Mw : Lx ] = 1, the conditions x ≤
8.8. Proof of the KL Conjecture 165

y < w in (a) insure that My ,→ Mw and thus [My : Lx ] = 1 as required.

Then part (3) above and the KL Conjecture imply (b).

8.8. Proof of the KL Conjecture

Once the KL Conjecture was announced, a number of independent attempts
were made to understand it better and find a method of proof. Most of
these attempts (especially those relying on algebraic methods alone) ran into
insuperable barriers. But along the way some consequences of the conjecture
as well as its relationship with the earlier Jantzen Conjecture were worked
out. This work also suggested how one might interpret the coefficients of KL
polynomials in terms of the module structure of Verma modules (including
the Jantzen filtration). Some ideas due to Deodhar [78], Gabber–Joseph
[107], and Gelfand–MacPherson [110] will be discussed below in the context
of later refinements by Irving.
The first—and to date only—proofs of the KL Conjecture were published
independently in 1981 by Beilinson–Bernstein [16] (in a detailed research
announcement) and Brylinski–Kashiwara [55] (in a longer paper); see also
Beilinson’s 1983 ICM talk [15]. In each case the methods used, drawn from
recent work on D-modules and perverse sheaves, are essentially the same.
In a Bourbaki seminar soon after these papers appeared, Springer [242]
explained the ideas involved together with the conjectured applications to
his own work on Weyl group representations. This seminar talk is still
useful for its detailed overview of the deep mathematics here. (An extensive
historical account is provided by Kleiman [188].) In the following section
we outline the main steps in the proof of the KL Conjecture; filling in the
details requires a considerable amount of background in modern sheaf theory
and related matters.
For nonintegral (but compatible) weights, the analogue of the KL Con-
jecture does not follow straightforwardly from the machinery developed to
prove the original conjecture. These methods only go as far as the case of
weights with coordinates (relative to fundamental dominant weights) in Q.
Instead, one can transfer the problem for a block Oλ to an equivalent mod-
ule category involving integral weights for a semisimple Lie algebra having
Weyl group W[λ] . This is worked out by Soergel [237, 2.5]: see Theorem
11 and the related Bemerkung, discussed below in 13.13. (A later approach
by Kashiwara–Tanisaki [170] treats more generally the case of nonintegral
weights for Kac–Moody algebras.) Now that the original KL Conjecture
along with its analogue for arbitrary blocks of O have been established,
it becomes important in further work on O to identify clearly where the
statement—or the proof—of the conjecture is used.
166 8. Kazhdan–Lusztig Theory

Over the years the “localization” method articulated in the proof of

the conjecture has continued to reverberate in neighboring parts of repre-
sentation theory. (The method itself has also been revisited by Joseph–
Perets–Polo [164].) This success in translating an intractable combinatorial
problem in algebra into a problem in algebraic geometry has inspired many
further developments in “geometric representation theory”.

8.9. Outline of the Proof

Following the geometric interpretation of KL polynomials in the setting of
Schubert varieties (8.5), it becomes clear that one wants to find a pathway
from the representation theory of category O to the local geometry of the
flag variety X = G/B. At the risk of oversimplifying matters, we indicate
some of the main steps involved. The full proof requires far more background
and details; see the recent book by Hotta–Takeuchi–Tanisaki [125].
(A) Rather than looking for an explicit category equivalence between
the block O0 and a geometric category, it is best to start with the U (g)-
modules on which Z(g) acts trivially, i.e., by the character χ0 . These form
a subcategory of Mod U (g) whose intersection with O0 contains the Mw
and Lw which figure in the original formulation of the KL Conjecture. If
we denote by J0 the two-sided ideal of U (g) generated by Ker χ0 and set
U0 := U (g)/J0 , we are therefore looking at the category of left U0 -modules.
If G is a semisimple group with Lie algebra g and B a Borel subgroup
with Lie algebra b, consider the flag variety X = G/B (as in 8.5). The idea
is to work with differential operators, not just functions, on X: locally, these
arise from partial differentiation operators combined with multiplication by
functions. Besides the sheaf OX of regular functions on X (a notation unre-
lated to category O), we have a larger sheaf DX of differential operators on
X. Here DX is a quasi-coherent OX -module. The algebra DX := Γ(X, DX )
of global sections is the algebra of (global) differential operators on X.
On the other hand, g is viewed classically as the space of right-invariant
tangent vector fields on G. This leads to an epimorphism U (g) → DX ,
whose kernel turns out to be J0 . So U0 ∼ = DX . This is derived indirectly
using a well-known theorem of Kostant in the commutative setting: the
ideal of polynomial functions on g vanishing on the nilpotent variety N ⊂ g
is generated by the G-invariant nonconstant polynomials. (The link with X
is that the cotangent bundle T ∗ X projects naturally onto N and provides a
resolution of singularities.)
(B) Working just in the category of DX -modules for U (g) conceals too
much of the local behavior relative to the stratification of X by Bruhat
cells. To remedy this, one moves the problem into a category of sheaves on
X, namely the category C of all DX -modules which are quasi-coherent over
8.9. Outline of the Proof 167

the structure sheaf OX . This has to be compared with the category C 0 of

DX -modules.
There are natural functors between C and C 0 : in one direction, F 7→
Γ(X, F), a module over DX . In the other direction M 7→ DX ⊗DX M . The
fact that these define an equivalence of categories between C and C 0 is highly
nontrivial: it rests on the proof that X is “D-affine” (a subtle sheaf-theoretic
analogue of the classical theorem of Borel–Weil), requiring that DX -modules
be generated by their global sections and have zero sheaf cohomology in
degrees > 0. This equivalence is a kind of noncommutative localization,
inspired by the connection between the sheaf of regular functions on an
affine variety and the local rings realized as global sections of the structure
sheaf over affine open sets.
Our main concern is with the subcategory C00 consisting of finitely gener-
ated, U (b)-finite modules (which includes all Mw and Lw ). Its counterpart
C0 consists of coherent sheaves M killed by an ideal of finite codimension
in U (b) and having a U (b)-stable “good” filtration. Under the category
equivalence, one sees that C0 is equivalent to C00 .
In this situation, it is shown that the D-modules in C0 are holonomic
and have regular singularities.
(C) Now it requires some close study to identify the specific sheaves Mw
and Lw corresponding to Mw and Lw . Here one encounters local cohomology
and complexes of sheaves.
To make contact with the local description of KL polynomials in 8.5,
one needs to go beyond the techniques of classical sheaf theory such as
derived functors. The (bounded) derived category Dcb (X) of constructible
sheaves on X provides the correct setting. Here the objects are complexes
of sheaves, but with a more subtle notion of morphism than usual in sheaf
theory. As in the theory of D-modules, one has to study delicate features
of various functors associated to maps: surjections G/B → G/P with P1 -
fibers) (where P is “minimal”) or embeddings Xw ,→ X.
There is a sophisticated modern version of classical ideas about systems
of differential equations, known as the Riemann–Hilbert Correspondence,
which allows one to pass from D-modules to objects in the derived category.
Unfortunately Dcb (X) is not itself an abelian category. But the crucial ob-
jects corresponding to Verma modules and simple modules turn out to be
perverse sheaves, which do form an abelian category. Here the objects have
finite filtrations with simple quotients, much as in O0 .
To a Verma module corresponds an easily constructed complex, which
is nonzero in only one degree and is supported on a single Bruhat cell,
whereas the simple objects are more elusive. But in the Grothendieck group
168 8. Kazhdan–Lusztig Theory

one has (as in category O) a unipotent change of basis matrix relating the
two types of objects. Here one wants to express the simple objects in terms
of the Verma objects. As indicated in 8.5, the local intersection cohomology
groups recover the KL polynomials as Poincaré polynomials. Finally these
yield the multiplicities in K(O0 ) conjectured by Kazhdan–Lusztig.

8.10. Ext Functors and Vogan’s Conjecture

Work of Vogan and others in the 1970s involving the broader study of uni-
tary representations of real reductive Lie groups led him to a conjecture
about wall-crossing functors which turned out to be equivalent to the KL
Conjecture: see [254, §3] as well as [253, 255]. In particular, his viewpoint
allows one to interpret the KL polynomials in terms of n-cohomology. This
translates formally, as indicated in Theorem 6.15, into the language of the
functors Extn .
Following fundamental work of Harish-Chandra and others, the descrip-
tion of unitary representations can to some extent be reduced to the alge-
braic study of “Harish-Chandra modules” for the complexified Lie algebra
of the Lie group relative to a maximal compact subgroup. Assuming that
G is semisimple (and linear), the Lie algebra g is identified with the one we
have been studying. If K is a maximal compact subgroup of G, the resulting
Harish-Chandra modules are usually called (g, K)-modules. A prototype for
the entire theory is the case of principal series representations for complex
Lie groups, which translates into the language of category O (see Chapter
10 below). In the papers cited above, Vogan starts here. The idea is to com-
pute characters of irreducibles in an algorithmic style; this depends heavily
on getting information about n-cohomology at each step.
Using Zuckerman’s version of translation and wall-crossing functors for
Harish-Chandra modules, joint work of Speh and Vogan led to a general
conjecture formulated by Vogan in [253, 3.15]. This encapsulates what is
needed to carry out an inductive step in the algorithm. It become known as
Vogan’s Conjecture and can be described briefly in the general setting,
using his notation. Given a simple (g, K)-module X and a simple root α for
which translation to the α-wall fails to annihilate X, the associated wall-
crossing functor ψα ϕα yields a complex 0 → X → ψα ϕα (X) → X → 0 with
cohomology Uα (X).
Conjecture. In the above notation, the (g, K)-module Uα (X) is semisimple.

In the special case corresponding to the principal block O0 , with modules

Mw , Lw (w ∈ W ) indexed by weights w · (−2ρ) as before, this translates as
follows. Recall the discussion in 7.14–7.15. Setting s = sα and assuming
that ws > w, we know that Θs Lw has simple head and simple socle both
8.11. KLV Polynomials 169

isomorphic to Lw . This yields a complex:

0 → Lw → Θs Lw → Lw → 0.
Then the conjecture is that Rad Θs Lw / Soc Θs Lw is semisimple.

8.11. KLV Polynomials

How does Vogan’s Conjecture relate to the KL Conjecture? Here is a brief
heuristic explanation (which does not correlate precisely with the history).
Recall Proposition 6.14, which was proved in a fairly elementary way (using
BGG Reciprocity): in K(O0 ) we have
X X
[Lw ] = (−1)i dim ExtiO (Mx , Lw ) [Mx ].
x≤w i≥0

Comparing this with the KL Conjecture and using (−1)i = (−1)−i , we see
that the truth of the conjecture would imply
X
Px,w (1) = (−1)`(x,w)−i dim ExtiO (Mx , Lw ).
i≥0

This makes it tempting to look at a sort of generating function in the variable

1
q 2 : the choice of variable insures that we get a polynomial of degree at most
`(x, w) thanks to Theorem 6.11(b). Define for x ≤ w in W ,
X
Pex,w := (−1)`(x,w)−i q (`(x,w)−i)/2 dim ExtiO (Mx , Lw ).
i≥0

In the axiomatic setting of “BGG algebras”, Irving [138, §4] calls this a
KLV polynomial. If we assume the truth of the KL Conjecture, setting
q = 1 recovers Px,w (1) in view of the remarks above.
Now Vogan’s conclusions about the KL Conjecture can be summarized
as follows:

Theorem. With modules in the block O0 labelled as above, the following

statements are equivalent:
(a) The KL Conjecture is true.
(b) For all x ≤ w, the KLV polynomial Pex,w is a polynomial in q and
coincides with Px,w (q). Thus
`(x,w)−2i
X
Px,w (q) = q i dim ExtO (Mx , Lw ).
i≥0

(c) For all x ≤ w, ExtiO (Mx , Lw ) = 0 unless i ≡ `(x, w) (mod 2).

170 8. Kazhdan–Lusztig Theory

Most of this is made explicit in [254, §3], notably the fact that (b)
implies (a). Condition (c) mirrors in a way the vanishing of cohomology
in odd degrees which appears in the geometric setting (8.5) and indirectly
accounts for the fact that Px,w is a polynomial in q rather than just q 1/2 .

Remark. In their study of the homological algebra associated with the KL

Conjecture, a possible interpretation of the auxiliary polynomials Rx,w (q)
was proposed by Gabber–Joseph [107, §5] (just before the conjecture it-
self was settled). Somewhat in parallel with the above description of the
coefficients of KL polynomials in terms of various dim ExtiO (Mx , Lw ), they
defined (using different notation) polynomials of the form
X
R
ex,w (q) := (−1)`(x,w)−i q i dim ExtiO (Mx , Mw ).
i

The hypothesis that R ex,w (q) = Rx,w (q) became known as the Gabber–
Joseph Conjecture and remained open for a decade. Its plausibility came
from the fact that the R-polynomials
e satisfy some of the same recursion
relations used to compute R-polynomials. Some supporting evidence for the
absolute values of coefficients predicted by the conjecture emerged from the
methods developed by Carlin [56] to compute Extn between Verma modules
with regular integral highest weights. This involves construction of a spectral
sequence for derived functors of adjoint functors (here translation functors).
In particular, Carlin proves that the highest nonvanishing Ext group has
dimension 1:
`(x,w)
dim ExtO (Mx , Mw ) = 1 if x ≤ w.

He also recovers known results for related ExtiO (Mx , Lw ), in the spirit of
Vogan’s work.
But eventually Boe [31] was able to use a computer to locate a spe-
cific counterexample to the Gabber–Joseph conjecture; his results were then
replicated by hand. The idea behind this is fairly simple: since the coeffi-
cients of the R-polynomials
e alternate in sign (by definition), it is enough to
locate an R-polynomial having two successive nonzero coefficients with the
same sign. The “smallest” possible example (relative to rank and `(x, w))
turns out to occur when W has type B4 = C4 . Further examples were found
in types D4 and A5 , and presumably exist commonly in higher ranks.
In spite of this negative outcome, the work of Beilinson–Ginzburg–
Soergel [22] on Koszul duality (discussed in 13.15 below) might lead to
a more subtle version of the conjecture.
8.12. The Jantzen Conjecture and the KL Conjecture 171

8.12. The Jantzen Conjecture and the KL Conjecture

As mentioned earlier, the KL Conjecture immediately raised questions about
its possible relationship with the Jantzen filtration of a Verma module (5.3).
The Jantzen filtration with its associated sum formula was introduced in
Chapter 5 primarily as a tool in the proof of the BGG Theorem; but it was
also expected to be helpful at least indirectly in the quest for composition
factor multiplicities.
Recall the set-up, for an arbitrary Verma module M (λ), λ ∈ h∗ . The
filtration M (λ) = M (λ)0 ⊃ M (λ)1 ⊃ . . . involves proper inclusions, with
M (λ)1 = N (λ) = Rad M (λ), and ends at 0. As observed in 5.3, the last
nonzero term M (λ)n has index n = |Φ+ +
λ |, where Φλ = {α > 0 | sα · λ < λ}.
Now the sum formula reads:
X X
ch M (λ)i = ch M (sα · λ).
i>0 α∈Φ+
λ

The question raised by Jantzen in [148] about whether this filtration is

compatible in a natural way with embeddings M (µ) ,→ M (λ) became known
as the Jantzen Conjecture:
(∗) Let µ ↑ λ in h∗ , with µ < λ, and consider the associated embedding
M (µ) ,→ M (λ). Set r := |Φ+ + i
λ | − |Φµ |. Then M (µ) ∩ M (λ) = M (µ)
i−r for

all i ≥ r. In particular, M (µ) ⊂ M (λ)r .

Thanks to the BGG Theorem 5.1, it would be enough to prove this when
µ = sα ·λ < λ with α > 0. The crux of the problem is how to characterize the
filtration and sum formula intrinsically: the original construction required
passage to C[T ] and then specialization T 7→ 0.
Various special cases have also been studied under the rubric “Jantzen
Conjecture”:
• In (∗), consider only the case µ = sα · λ < λ with α simple; here
r = 1.
• Let λ and µ be regular, with λ = w · ν and µ = x · ν for some
antidominant weight ν. By Exercise 5.3, r = `λ (x, w) (using the
length function in W[λ] = W[µ] ), so condition (∗) becomes:
M (x · ν) ∩ M (w · ν)i = M (x · ν)i−`(x,w) for all i ≥ r.
• Let λ and µ be integral ; for example, consider just the principal
block O0 or other regular integral block.
In [107] Gabber and Joseph showed by algebraic methods that the truth
of a weak version of the Jantzen conjecture would imply the truth of the KL
Conjecture (not yet proved at the time). More precisely, they formulated
(as did Deodhar and Gelfand–MacPherson, independently) a “generalized
172 8. Kazhdan–Lusztig Theory

KL Conjecture” based on the Jantzen filtration: the multiplicities of simple

modules in the filtration layers should be given by the coefficients of a corre-
sponding KL polynomial. By setting q = 1 one obtains the KL Conjecture.
Here is a precise statement of their main theorem:
Theorem (Gabber–Joseph). Let λ ∈ h∗ be regular and antidominant. If
α ∈ ∆[λ] , w ∈ W[λ] , and wsα > w, assume that
M (w · λ)i ⊂ M (wsα · λ)i+1 ∩ M (w · λ) for all i ≥ 0.
Then
(a) All layers M (w · λ)i are semisimple.
(b) The (generalized) KL Conjecture holds for all weights linked to λ.
Therefore the KL Conjecture is true for these weights.

Notice here that the hypothesis is a weaker version of the Jantzen Con-
jecture, involving only embeddings associated with simple roots. The fact
that this is enough to imply the KL Conjecture suggests that the Jantzen
Conjecture may be significantly stronger and is therefore unlikely to provide
a pathway to proving the KL Conjecture. This is reinforced by the work of
Beilinson–Bernstein discussed in the following section. But in another di-
rection, Barbasch [13] was able to draw further consequences from the work
of Gabber–Joseph for the Loewy structure of Verma modules: see 8.16.
The proof of the theorem above relies on moderately complicated induc-
tions involving indices in the Jantzen filtration, coupled with study of wall-
crossing functors and exact sequences. This yields a proof that Uα L(w · λ)
is semisimple (using Vogan’s notation as in 8.10) and thus that all layers in
the filtration are semisimple. Induction comes heavily into play when com-
puting the multiplicities of simple modules in layers and then comparing
the KL polynomials. Along the way, it is shown that the multiplicity data
are compatible with the Jantzen sum formula. It is also shown (without
invoking the hypothesis of the theorem) that the methods of proof recover
some known consequences of the KL Conjecture.
Remark. For each λ ∈ h∗ , the Jantzen Conjecture locates a maximal vector
of weight sα · λ in a definite submodule M (λ)i . On the other hand, a Shapo-
valov element as in 4.12 produces such embeddings uniformly across a root
hyperplane. Carlin [58] investigates (especially in type A` ) the connection
of Shapovalov elements with Jantzen filtrations as λ varies.

8.13. Weight Filtrations and Jantzen Filtrations

The proof of the Jantzen Conjecture by Beilinson–Bernstein required a sig-
nificant extension of the geometric techniques used in the proof of the KL
8.14. Review of Loewy Filtrations 173

Conjecture. While the main ideas were worked out early in 1981 and outlined
in Beilinson’s 1983 ICM talk [15], a detailed proof only appeared in 1993 [18,
5.3]. The basic idea is to endow the related category of perverse sheaves with
enriched structure, allowing the Jantzen filtration to be interpreted there as
a weight filtration in the sense of Beilinson–Bernstein–Deligne and Gabber
[17, §5].
The notion of “weight” here is unrelated to the Lie-theoretic notion
arising in representation theory. It comes instead from the study of sheaves
over Q` (the algebraic closure of an `-adic field) on a scheme defined over a
finite field Fq of characteristic p 6= `. The ideas here arose in Deligne’s proof
of the Weil Conjectures. Without attempting to explain the details, we can
suggest the flavor of the language employed. For each power q n there is a
Frobenius operator acting on points of the scheme over Fq . Given a fixed
point x for this action, there is an induced action of the Frobenius on the
stalk at x of the given sheaf F. If w ∈ Z and the eigenvalues for this induced
action are algebraic numbers whose complex conjugates all have absolute
value (q n )w/2 for any fixed point x, then F is said to be pointwise pure of
weight w. A sheaf is then said to be mixed if it has a finite filtration with
pointwise pure quotients. These ideas carry over to complexes of sheaves as
well.
It is shown in [17, 5.3] that every simple mixed perverse sheaf is pure
and that any mixed perverse sheaf admits a unique finite increasing filtration
(the “filtration par le poids”) such that the associated ith graded piece is
pure of weight i. Morphisms of sheaves then preserve such filtrations. Based
on these results, the correspondence between Verma modules and perverse
sheaves developed in the proof of the KL Conjecture can be strengthened:
Theorem (Beilinson–Bernstein). If λ ∈ h∗Q , the perverse sheaf corre-
sponding to the Verma module M (λ) is mixed and the counterpart of the
Jantzen filtration is a weight filtration. As a result, the Jantzen Conjecture
is true.

As in the case of the KL Conjecture, the passage to arbitrary weights

can be made by using the later work of Soergel [237, Thm. 11].
Remark. As pointed out in Remark 7.8, Jantzen raised in [148, 5.17] the
question of how Jantzen filtrations behave relative to translation functors.
This seems to be unresolved, but one may expect the interpretation in terms
of weight filtrations to play a key role.

8.14. Review of Loewy Filtrations

As observed in 8.5 above, the coefficients of KL polynomials have a natural
interpretation in geometry as dimensions of certain cohomology groups. It
174 8. Kazhdan–Lusztig Theory

is reasonable to ask whether there is also a representation-theoretic inter-

pretation. The answer involves the submodule structure of Verma modules.
While this structure can become extremely complicated, some aspects of it
turn out to correlate well with KL theory (in Vogan’s formulation).
First we have to review briefly some standard module theory, which
applies quite generally to modules satisfying both chain conditions. Since
notation and terminology vary somewhat in the literature, we have to make
our own conventions precise. Consider all (ascending or descending) filtra-
tions of a nonzero module M having successive quotients which are nonzero
and semisimple. A composition series is the longest possibility, but we seek
instead the shortest and call it a Loewy filtration (or Loewy series); the
semisimple quotients are then called Loewy layers and the number of these
the Loewy length of M , written ``M . (The length turns out to depend
only on M .)
There are two extreme examples:
(1) The radical filtration is defined inductively starting with Rad0 M :=
M and Rad1 := Rad M (the radical of M ), the smallest proper submodule
N such that M/N is semisimple (equal to the intersection of all maximal
submodules). In general, Radk+1 M is the smallest proper submodule of
Radk M for which the quotient is semisimple. The chain conditions in-
sure that this filtration is well-defined and ends at 0. Write Radk M :=
Radk M/ Radk+1 M . Sometimes Rad0 M is called the head of M , denoted
Hd M .
(2) The socle filtration is defined inductively starting with Soc0 M := 0
and Soc1 := Soc M (the socle of M ), the largest semisimple submodule of
M (equal to the sum of all simple submodules). In general, Sock M is the
(unique) submodule of M satisfying Soc(M/ Sock−1 M ) = Sock M/ Sock−1 M .
The chain conditions insure that this filtration is well-defined and ends at
M . Write Sock M := Sock M/ Sock−1 M .
These Loewy filtrations have a natural hereditary property relative to a
submodule N ⊂ M : for example, Sock N = N ∩ Sock M . (In the case of the
radical filtration, one has to shift indices to take account of the difference in
Loewy lengths.)
Given any descending Loewy filtration M = M 0 ⊃ M 1 ⊃ . . . , one
shows inductively that Radr−k M ⊂ M r−k ⊂ Sock M if r = ``M . In case
the radical and socle filtrations coincide, giving the unique Loewy filtration,
we say that M is rigid. Thus the interesting questions about a module M
include: What is its Loewy length? Is M rigid? What are the composition
factors of each Loewy layer?
8.15. Loewy Filtrations and KL Polynomials 175

8.15. Loewy Filtrations and KL Polynomials

As a consequence of the KL Conjecture and Jantzen Conjecture (or their
proofs), one gets considerable insight into the Loewy structure of Verma
modules and associated extensions. This has implications for the structure
of projective covers as well. For an arbitrary λ ∈ h∗ , we know initially
that Hd M (λ) ∼= L(λ) and Rad M (λ) = N (λ), while Soc M (λ) ∼= L(µ), the
simple module with antidominant highest weight in the block determined
by λ.
To keep the formulations simple, we look first at the principal block
O0 , whose simple modules Lw , Verma modules Mw , and indecomposable
projectives Pw correspond to the weights w · (−2ρ), w ∈ W . As we discuss
in the following section, most of the results adapt to all integral weights
via translation functors and also (sometimes in modified form) to arbitrary
weights. Adaptations to parabolic categories are more complicated, however
9.17).
Recall from 8.2 that x ≺ w means x < w, with deg Px,w (q) as large as
possible: (`(x, w) − 1)/2. In this case, denote by µ(x, w) the leading coeffi-
cient of Px,w (q); in all other cases, set µ(x, w) = 0. The inversion formula
for KL polynomials insures that µ(x, w) = µ(w◦ w, w◦ x) [171, Cor. 3.2].
Now we can summarize in a portmanteau theorem the main structural
results following from the KL Conjecture:
Theorem. In the principal block O0 , denote modules as above by Lw , Mw , Pw
with w ∈ W . Assume the truth of the KL Conjecture for O0 .
(a) Each Verma module Mw is rigid and has Loewy length `(w) + 1.
(b) Loewy layers of Verma modules are determined by coefficients of
KL polynomials: If x < w, then
X
Pw◦ w,w◦ x (q) = [Soc`(x)+1+2k Mw : Lx ] q k
k
X
= [Rad`(x,w)−2k Mw : Lx ] q k .
k
(c) If x < w, then dim Ext1O (Lx , Lw ) = µ(x, w).
(d) The antidominant projective P1 is rigid and has Loewy length equal
to 2`(w◦ ) + 1. Moreover,

Sock P1 ∼
M
= Sock−2`(w◦ w) Mw .
w∈W

(e) In general, Pw is rigid if and only if [Mw◦ : Lw ] = 1. In that case

its Loewy length is 2`(w◦ ) − `(w) + 1.
176 8. Kazhdan–Lusztig Theory

We defer to the following section a discussion of the history and proofs

of these statements. First we comment on some parts of the theorem and
illustrate them in special cases.
(1) Given part (a), the equivalence of the two sums in part (b) is a
simple bookkeeping exercise: here Sock = Rad`(w)−k+1 . Notice that the
Loewy lengths given in parts (a), (d), and (e) are consistent with what we
know about the extreme cases w = 1 and w = w◦ .
(2) In (b) the placement of simple modules on Loewy layers is most
readily visualized in terms of the radical filtration. Assume x < w. The
default situation occurs when Pw◦ w,w◦ x (q) = 1, in which case [Mw : Lx ] = 1
(by the KL Conjecture). Here the index k = 0 in the summation corresponds
to the unique radical layer on which Lx lies; this is layer `(x, w). In general,
whenever x < w one copy of Lx always occurs on this layer; but other copies
may occur on other layers.
Consider for example the 8-dimensional Schubert variety Xw in type
C3 treated in 8.6 above. Here Mw has 42 distinct composition factors Lx
with x ≤ w (corresponding to Bruhat cells), but 6 of them occur twice when
Px,w (q) = 1+q 2 . The elements x in question have lengths 3, 2, 2, 1, 1, 0 in W .
So their “default” radical layers are numbered 5, 6, 6, 7, 7, 8. The respective
extra composition factors lie on radical layers 1, 2, 2, 3, 3, 4.
(3) We claim that (c) follows readily from (b). To see this, we just have
to interpret Ext1 in terms of radical layers. Recall from Proposition 3.1(c)
and Theorem 3.2(e) that whenever µ < λ in h∗ , we have
HomO N (λ), L(µ) ∼ = ExtO L(λ), L(µ) ∼




= ExtO L(µ), L(λ) .
Here N (λ) = Rad M (λ). On the other hand,
HomO Rad M (λ), L(µ) ∼



= HomO Rad1 M (λ), L(µ) ,
whose dimension is [Rad1 M (λ) : L(µ)] since radical layers are semisimple.
In our situation, it follows from (b) that
dim Ext1O (Lx , Lw ) = [Rad1 Mw : Lx ] = [Rad`(x,w)−2k Mw : Lx ],
the coefficient of q k in Pw◦ w,w◦ x (q). This in turn forces `(x, w) − 2k = 1, or
k = (`(x, w) − 1)/2, the largest possible degree of the polynomial. So the
coefficient in question is µ(w◦ w, w◦ x) = µ(x, w).
In the framework of part (c), the “default” simple modules Lx for which
dim ExtO (Lw , Lx ) = 1 correspond to elements x < w with `(x, w) = 1.
In the C3 example just discussed, there are three such x. But there is
also a nonzero Ext group involving a single Lx for which `(x, w) = 5; here
Px,w (q) = 1 + q 2 , so x ≺ w. All Ext groups here have dimension 1, but
in general very large coefficients can occur in KL polynomials (as in type
8.16. Some Details 177

E8 , discussed in 8.3) and might yield high dimensional Ext groups for large
values of µ(x, w).
(4) In part (d), the reader can translate the result into the language
of radical layers in the spirit of (b). The layer picture is readily visualized
here and is consistent with the “inverted” arrangement of Verma modules
in a standard filtration of a projective module suggested formally by BGG
Reciprocity. First place Mw◦ , with its head in the middle of the picture and
its radical layers numbered from top to bottom as 0 to `(w◦ ) + 1. Then
place Mw◦ w exactly `(w) layers higher. In this way the total number of
layers becomes 2`(w◦ ) + 1, with L1 at the top and bottom of the picture
and Lw◦ in the middle. Moreover, the self-dual feature of P1 becomes visible
when all the pieces are assembled.

Exercise. When g = sl(3, C), use parts (d) and (e) of the theorem to work
out the Loewy layers of all Mw and Pw .

8.16. Some Details

Obviously much more needs to be said about the history and proof of The-
orem 8.15, as well as about the case of an arbitrary block Oλ with λ ∈ h∗ .
We conclude this chapter with an overview of these matters and references
to the primary sources.
(1) The history is somewhat complicated, since some questions about
the possible relationship between coefficients of KL polynomials and the
structure of Verma modules were already raised in the immediate aftermath
of the KL Conjecture. As already mentioned, independent work of Deodhar
[78], Gelfand–MacPherson [110], and Gabber–Joseph [107] explored the
idea that the coefficients might predict natural filtration layers.
Deodhar formulated this combinatorially in the Hecke algebra setting by
introducing polynomials in q whose sum would be Px,w (q) and whose values
at 1 should have this interpretation. The viewpoint of Gelfand–Macpherson
was more topological: they set up a sort of dictionary between category
O data and Schubert varieties, which turned out not to lead to a proof of
the KL Conjecture. Even so, their dictionary also suggested connections
between KL polynomials and layer structure in Verma modules.
As stated above in 8.12, Gabber–Joseph [107] were able to show that the
truth of the Jantzen Conjecture would imply the KL Conjecture. Along the
way they showed that the Jantzen Conjecture already implies the semisim-
plicity of layers in the Jantzen filtration. This was strengthened by Barbasch
[13] to conclude that the Jantzen filtration of any Verma module must then
coincide with the socle filtration. This implies as in Theorem 8.15(c) that
178 8. Kazhdan–Lusztig Theory

dim ExtO (Lx , Lw ) is equal to the leading coefficient µ(x, w) of Px,w (q) when
its degree is (`(x, w) − 1)/2 (and is 0 otherwise).
(2) Once it became clear from the work of Beilinson–Bernstein that the
Jantzen Conjecture would be even harder to prove than the KL Conjecture,
systematic study of Loewy filtrations by Irving led to a proof of other parts of
Theorem 8.15, along with generalizations: see [135, 136, 137, 141]. While
many details are required, the essential methods are inductive in nature: a
natural starting point is the case of an antidominant Verma module.
Working with Vogan’s version of the KL Conjecture, Irving made a cru-
cial observation about its relationship with Loewy lengths:
Proposition. The KL Conjecture for O0 is equivalent to the statement: For
all M ∈ O0 and any simple reflection s, the functor Θs satisfies ``(Θs M ) ≤
``M + 2.

What can be said about the Loewy structure of Verma modules and
projectives when λ ∈ h is arbitrary? Thanks to the category equivalences
induced by translation functors, which obviously respect socle and radical
series, Theorem 8.15 carries over at once to blocks involving regular integral
weights. In fact, the proofs work equally well for arbitrary regular weights.
The case of singular blocks is more delicate: see the papers by Irving
[137, 141]. It turns out that the results for regular blocks can be adapted to
the singular case by careful use of translation functors. In particular, Verma
modules are again rigid and the multiplicities of composition factors in their
Loewy layers are again given by coefficients of certain KL polynomials. The
polynomials are of the type Px,w (q), where x and w are minimal length coset
representatives for cosets W/Wµ◦ when µ is singular. Moreover, the radical
series of projectives are described much as in the regular case. (In his first
paper, Irving had to invoke some of the deeper geometric ideas involving
weight filtrations. But the second paper shows how to proceed in a more
self-contained manner.)
(3) One concluding remark may be appropriate: While the basic facts
about Loewy structure of Verma modules and projective modules in O have
by now been established, the proofs depend on knowing the truth of the KL
Conjecture (in a formulation covering all weights). Serious attempts by Irv-
ing and others to find more elementary algebraic proofs have all foundered.
Similarly, it seems unlikely at this point that the KL Conjecture or Jantzen
Conjecture will be proved by purely algebraic methods. But in mathematics
it is risky to make predictions.
Part II. Further
Developments

179
 Chapter 9

Parabolic Versions of
Category O

For applications to Lie group representations a relative version Op of cate-

gory O is often required. This is a subcategory determined by a parabolic
subalgebra p ⊃ b of g. Study of Op is roughly parallel, in the Lie group
setting, to the study of manifolds G/P generalizing the flag manifold G/B.
We have deferred this topic until now partly in order to avoid an overlay
of extra notation, but also due to the greater complexity of some of the
results. While a combined treatment might be more efficient, we now have
the option of applying some of the earlier theory for O to Op rather than
starting again from first principles. In any case, the simple modules in Op are
already known from the study of O; so the emphasis shifts to understanding
the analogues of Verma modules.
We start with a quick review in 9.1 of parabolic subalgebras, along with
related roots and weights. Then the main ideas in this chapter fall under a
number of headings:

• Basic features of Op : definition and first properties (9.3), parabolic

Verma modules (9.4).

• Formal characters and Grothendieck group (9.6): in principle the

characters of parabolic Verma modules can be deduced from the
known results about O. A more intrinsic approach is based on a
relative version of Kazhdan–Lusztig theory (9.7).

• Projective modules and BGG Reciprocity in Op (9.8)

181
182 9. Parabolic Versions of Category O

• Module structure: maps between parabolic Verma modules (9.10),

simplicity criterion (9.12–9.13), socles, self-dual projectives (9.14),
blocks (9.15), analogue of the BGG resolution (9.16), filtrations
and rigidity (9.17).

Along the way, a number of special cases are examined involving most
often “large” parabolics. To avoid too many technical detours, we often
emphasize integral weights; side remarks will indicate what is true in general.

9.1. Standard Parabolic Subalgebras

First we recall from 0.1 the relevant subalgebra structure in g. Each subset
I ⊂ ∆ defines a root system ΦI ⊂ Φ, with positive roots Φ+ I , negative roots
−
ΦI , and Weyl group WI ⊂ W generated by all sα with α ∈ I. Associated
with the root system ΦI are a number of subalgebras of g:

pI standard
P parabolic subalgebra
lI h ⊕ α∈ΦI gα (Levi subalgebra of pI )
L
uI + + gα (nilradical of pI )
Lα∈Φ \ΦI
u−I α∈Φ− \Φ− gα
I
gI [lI , lI ] (semisimple subalgebra of g)
L
hI Lα∈I Chα (Cartan subalgebra of gI )
nI + gα
Lα∈ΦI
n−I − gα
T α∈ΦI
zI α∈I Ker α (= center of lI )

In the extreme case I = ∅, we have pI = b; at the other extreme, p∆ = g. In

all cases we have direct sum decompositions:
pI = lI ⊕ uI
lI = gI ⊕ zI
gI = n−I ⊕ hI ⊕ nI
h = hI ⊕ zI
n = nI ⊕ uI
n− = n−I ⊕ uI
−

g = u− −
I ⊕ lI ⊕ uI = uI ⊕ pI

Here WI is the Weyl group of gI (or of the reductive algebra lI ).

As before the reader has to be aware of divergent notation and terminol-
ogy in the literature; for example, S is commonly used rather than I, while
the letter m often denotes a Levi subalgebra or its derived algebra in the Lie
group context.
9.2. Modules for Levi Subalgebras 183

9.2. Modules for Levi Subalgebras

To deal with representations of g arising from a parabolic subalgebra p = pI ,
we also need to supplement the notation of the previous section by consid-
ering weights.
From h = hI ⊕ zI we get a corresponding decomposition h∗ = ∼ h∗ ⊕ z∗ ,
I I
where for example λ ∈ h∗I is 0 on zI . It is easy to check that the α ∈ I form
a basis of h∗I , while the fundamental weights $β with β ∈ ∆ \ I form a basis
of z∗I . For each λ ∈ h∗ , denote by λI its h∗I component, identifiable with its
restriction to hI . Then the set of λ for which λI is dominant integral on hI
may be denoted by
∗ ∨
Λ+ +
I := {λ ∈ h | hλ, α i ∈ Z for all α ∈ I}.

The condition on λ is equivalent to hλ + ρ, α∨ i ∈ Z>0 for all α ∈ I (or all

α ∈ Φ+ + +
I ). Obviously Λ ⊂ ΛI .

Exercise. Let λ ∈ Λ+ +
I . If w ∈ WI and w 6= 1, prove that w · λ 6∈ ΛI .

The λI with λ ∈ Λ+I are precisely the highest weights of finite dimensional
gI -modules. Such a module has a natural structure of simple lI -module,
where zI acts by the restriction of λ to zI . Denote this lI -module (or its
restriction to gI ) by LI (λ). Conversely, every finite dimensional simple lI -
module has this form: by Schur’s Lemma, the elements of zI must act by
scalars. But when zI 6= 0, many simple lI -modules have the same restric-
tion to gI . Observe also that Weyl’s complete reducibility theorem for finite
dimensional gI -modules (0.8) extends immediately to modules for the reduc-
tive Lie algebra lI on which h acts semisimply.
Besides the simple modules LI (λ), we need the corresponding Verma
modules for gI ; these too become U (lI )-modules in a natural way. Since
they will play only an auxiliary role below, we denote them in an ad hoc
way by
VI (λ) := U (lI ) ⊗U (h⊕nI ) Cλ with λ ∈ h∗ .
As in 1.16, it is easy to write down the formal character, as a function on h:
letting pI be the partition function for the root system ΦI analogous to the
Kostant function p, we get ch VI (λ) = pI ∗ e(λ).
When λ ∈ Λ+ I , the description of the maximal submodule of VI (λ) given
in Theorem 2.6 extends readily from gI to lI , yielding an exact sequence of
U (lI )-modules:
M
(1) VI (sα · λ) → VI (λ) → LI (λ) → 0.
α∈I
184 9. Parabolic Versions of Category O

9.3. The Category Op

Fix a standard parabolic subalgebra p = pI ⊃ b, corresponding to I ⊂ ∆.
One can introduce the parabolic category Op (or OI ) from first principles
(as for example in Rocha [226]) by axioms similar to those used for O in 1.1.
In this spirit, Op is defined to be the full subcategory of Mod U (g) whose
objects M satisfy:
(Op 1) M is a finitely generated U (g)-module.
(Op 2) Viewed as a U (lI )-module, M is a direct sum of finite dimensional
simple modules.
(Op 3) M is locally uI -finite.
Obviously Op is a full subcategory of O and contains all finite dimen-
sional modules. To simplify notation, we write HomO (M, N ) rather than
HomOp (M, N ) when M, N ∈ Op . There are two extreme examples: the case
I = ∅ just recovers O, while I = ∆ yields the (semisimple) category of finite
dimensional U (g)-modules.
Since we have already studied the basic properties of O, it is more effi-
cient at this point to rely instead on a single property to identify Op as a
subcategory of O.
Lemma. Let M ∈ O have set of weights Π(M ). The following conditions
on M are equivalent:
(a) M is locally n−
I -finite.
(b) For all α ∈ I and µ ∈ Π(M ), we have dim Mµ = dim Msα µ .
(c) For all w ∈ WI and µ ∈ Π(M ), we have dim Mµ = dim Mwµ .
(d) The set Π(M ) is stable under WI .

Proof. Assuming (a), the action on Mµ of the copy of sl(2, C) generated by

xα and yα for a fixed α ∈ I produces a finite dimensional submodule of M ;
this is clearly stable under h. Then the standard theory in 0.9 yields (b).
Since WI is generated by the reflections sα for α ∈ I, (c) follows at once and
trivially implies (d).
Now assume (d). Given a vector v ∈ M of weight µ, only finitely many
standard PBW monomials in U (nI ) can take v to a nonzero vector. Therefore
only finitely many weights of the form µ + ν with ν a Z+ -linear combination
of Φ+
I can lie in Π(M ). Now the longest element wI of WI interchanges ΦI
+

and Φ− +
I . Thanks to (d), wI stabilizes Π(M ), so only finitely many Z -linear
− 0
combinations of ΦI can be added to µ := wI µ to yield a weight of M . Again
by (d), µ0 is a typical weight of M , forcing M to be locally n− I -finite. 

With this in hand, it is easy to verify some elementary facts about Op :

9.3. The Category Op 185

Proposition. Fix p = pI as above.

(a) M ∈ O lies in Op if and only if M satisfies the equivalent conditions
of the lemma.
(b) Op is closed under duality in O.
(c) Op is closed under direct sums, submodules, quotients, and ex-
tensions in O, as well as tensoring with finite dimensional U (g)-
modules.
(d) If M ∈ Op decomposes as M =
L χ
M with M χ in Oχ , then each
p
χ
M lies in O ; this gives a decomposition Op =
p
L
χ Oχ . As a
result, translation functors preserve Op .
(e) If the simple module L(λ) lies in Op , then λ ∈ Λ+
I .

Proof. (a) Any M ∈ Op satisfies (Op 2), therefore also satisfies the condi-
tions in the lemma.
In the other direction, suppose M ∈ O satisfies those conditions. Since
any module in O satisfies (Op 1) and (Op 3), we just have to verify (Op 2).
Thanks to part (a) of the lemma, any weight vector v ∈ M generates a finite
dimensional U (gI )-module on which h acts semisimply since h normalizes nI
and n− I . It follows that every element of M lies in such a finite dimensional
U (lI )-module. Using complete reducibility, a standard argument shows that
M is a direct sum of simple U (lI )-modules, as required.
(b) Since M ∨ has the same formal character as M , part (d) of the lemma
is satisfied by M ∨ in place of M . Thus M ∨ lies in Op , by (a).
(c) and (d) These assertions follow readily from the criteria in the lemma,
coupled with (a).
(e) Suppose L(λ) ∈ Op . In particular, if v + is a maximal vector of weight
λ, then for each α ∈ I some power of yα kills v + . As in 1.6, the study of
sl(2, C)-modules shows that hλ, α∨ i ∈ Z+ ; thus λ ∈ Λ+ I . 

It is natural to ask whether the converse of (e) also holds; but it is com-

plicated to apply the criteria in the lemma directly to L(λ), since Π L(λ)
is hard to describe. Instead the proof will rely on an analogue of Verma
modules constructed in 9.4 below.
There are a couple of ways to define interesting functors from O to
Op . Given M ∈ O, define M to be the set of all n− I -finite vectors in M .
This is clearly a submodule and is the unique maximal element among the
submodules of M lying in Op . Taking advantage of the fact that duality
preserves Op , we can then set M := (M ∨ )∨ . This is the largest quotient of
M lying in Op and may be dubbed the truncation of M relative to p.
186 9. Parabolic Versions of Category O

Exercise. (1) Given an exact sequence M → N → 0 in O, prove that the

resulting homomorphism M → N is surjective.
(2) Given an exact sequence 0 → N → M in O, must the resulting map
N → M be injective?

9.4. Parabolic Verma Modules

What is the appropriate generalization of Verma modules in this setting?
The basic idea is to induce finite dimensional modules from pI to g as we
did with 1-dimensional b-modules. Start with the finite dimensional simple
lI -module LI (λ) for λ ∈ Λ+
I . Then inflate LI (λ) to a pI -module via the
∼
projection pI → pI /uI = lI , so that uI acts trivially. The resulting pI -
module is generated by a maximal vector v + of weight λ. Finally, define the
parabolic Verma module

MI (λ) := U (g) ⊗U (pI ) LI (λ).

Obviously u− −
I acts freely on LI (λ), just as n acted freely on a (simple!)
1-dimensional b-module of weight λ in the construction of Verma modules.
Thus MI (λ) is generated as a U (g)-module by the maximal vector 1 ⊗ v + of
weight λ. It follows that MI (λ) is a quotient of M (λ); in particular, L(λ) is
its unique simple quotient.
The construction of MI (λ) makes it easy to compute its formal character.
By the PBW Theorem, the module MI (λ) when viewed as a U (u− I )-module
is isomorphic to U (u−
I ) ⊗ L I (λ). Accordingly,

(1) ch MI (λ) = pI ∗ ch LI (λ),

where pI is defined like the function p in 1.16 but restricted to the sums of
positive roots not in Φ+ I
I . Note that p = p ∗ pI with pI as in 9.2.
While it is not straightforward here to write down explicitly a basis of
MI (λ) in terms of PBW monomials applied to v + , we can use the factoriza-
tion U (n− ) ∼
= U (u− −
I ) ⊗ U (nI ) to obtain a basis of the form

y1i1 . . . yrir u · v + ,

where the yi correspond to an enumeration of Φ+ \ ΦI and u · v + runs over

a basis of the finite dimensional space U (n− +
I )·v .
The exact sequence in the following theorem originates in work of Lep-
owsky [202, Prop. 2.1]:

Theorem. Let λ ∈ Λ+
I .

(a) The module MI (λ) lies in Op ; so its quotient L(λ) also does.
9.4. Parabolic Verma Modules 187

(b) There is an exact sequence

M
(2) M (sα · λ) → M (λ) → MI (λ) → 0.
α∈I

(c) MI (λ) coincides with the universal highest weight module M (λ) in
Op defined earlier in terms of truncation.

Proof. (a) To show that MI (λ) lies in Op , it will be enough (by Proposition
9.3(a)) to check condition (d) in Lemma 9.3: the set of weights of MI (λ) is
stable under WI . From the character formula (1) we know that these

weights
are characterized as the differences µ − ν, where µ ∈ Π LI (λ) and
ν is a
+ +
Z -linear combination of Φ \ ΦI . Clearly WI permutes Π LI (λ) . On the
other hand, each simple reflection sα with α ∈ I permutes the positive roots
other than α in both Φ+ and Φ+ I (0.3(1)). So for all w ∈ WI the weight
wµ − wν is again a weight of MI (λ) as required.
(b) As observed in 9.2(1), Theorem 2.6 extends to lI and gives an exact
sequence of U (lI )-modules:
M
(3) VI (sα · λ) → VI (λ) → LI (λ) → 0.
α∈I

As in the case of LI (λ), we can inflate VI (λ) to a U (pI )-module by letting

the nilradical uI act trivially. Then M (λ) ∼
= U (g) ⊗U (pI ) VI (λ). This inflate-
induce functor is exact and transforms the exact sequence into the desired
exact sequence (2) in O.
(c) It is clear that sα · λ 6∈ Λ+ / Op . It follows
I for all α ∈ I, so L(sα · λ) ∈
that M (sα · λ) lies in the kernel of the natural map ϕ : M (λ) → M (λ). So
the exact sequence in (b) implies that ϕ induces a surjection MI (λ) → M (λ).
Thanks to (a), MI (λ) ∈ Op . Since M (λ) is a universal highest weight module
in Op , this surjection must be an isomorphism. 
Exercise. On the level of formal characters, show that for λ ∈ Λ+
I :

ch M (λ) = pI ∗ ch VI (λ) = pI ∗ pI ∗ e(λ),

consistent with ch M (λ) = p ∗ e(λ) (1.16).

In view of Proposition 9.3, part (a) of the theorem shows that M ∈ O

lies in Op if and only if all its composition factors L(λ) satisfy λ ∈ Λ+
I .
It is often convenient to parametrize weights in Λ+ I by Weyl group ele-
ments. When λ ∈ Λ is dominant and regular, the weights w ·λ are in natural
bijection with the elements of W . If we then let W I be the set of minimal
length right coset representatives in WI \W , it is easy to check that
w · λ ∈ Λ+
I if and only if w ∈ W
I
188 9. Parabolic Versions of Category O

(as noted by Boe [30, 2.4]). This uses an observation which follows readily
from 0.3(4):
W I = {w ∈ W | w−1 Φ+ +
I ⊂ ΦI }.
Question: What can be said if λ is no longer required to be regular?
Remarks. (1) Jantzen [148, 2.25] discusses the analogue for Op of Theorem
7.6 on translation functors and their effect on Verma modules. As one might
expect, things get more complicated than before: translating a parabolic
Verma module to the closure of a facet may yield either the corresponding
parabolic Verma module or else 0.
(2) The name generalized Verma module is more widely used; but
its meaning has also been broadened (for example in work of Futorny, Ma-
zorchuk, and others) to include modules induced from infinite dimensional
simple U (p)-modules. Another variant is the label relative Verma mod-
ule. The reader needs to be aware of such differences in terminology, as well
as notation, throughout the literature.

9.5. Example: sl(3, C)

It is easy to illustrate some features of category Op when g = sl(3, C), as
in 4.11 and 5.4. (However, here as in other rank 2 cases the situation is
artificially simplified by the fact that multiplicities in Verma modules are all
1.) We look only at integral weights.
Denoting the simple roots by α, β, set I := {α}. Then the weights
λ ∈ Λ+ I are those lying on or above the hyperplane orthogonal to α; the
regular ones lie in the three Weyl chambers corresponding to 1, sβ , sβ sα
(which are minimal length right coset representatives in WI \W ). Label the
linked ones as w · λ for each fixed λ ∈ Λ+ .
What can be said about the parabolic Verma modules for regular weights
in Λ+I ? Since |WI | = 2, it is clear from the Weyl–Kostant character formula
that the exact sequence 9.4(3) is actually a short exact sequence and sim-
ilarly for the induced sequence 9.4(2). This makes the formal characters
(hence composition factor multiplicities) easy to work out.
Consider µ := sβ sα · λ, which is minimal among linked weights lying in
Λ+
I ;thus MI (µ) = L(µ). This is a proper quotient of M (µ) by the submodule
M (sα · µ) = M (w◦ · λ).
Next consider µ := sβ · λ. In this case the kernel of M (µ) → MI (µ) is
M (sα · µ) = M (sα sβ · λ), which has two composition factors. Thus MI (µ)
has the remaining two composition factors L(µ) = L(sβ · λ) and L(sβ sα · λ).
Finally, consider MI (λ), which is the quotient of M (λ) by M (sα · λ).
The latter has four composition factors, leaving only two for MI (λ) with
9.6. Formal Characters and Composition Factors 189

highest weights λ and sβ · λ. In particular, this gives an example in which

MI (λ) = M (λ) fails to capture all composition factors of M (λ) which lie in
Op . (Compare part (b) of Exercise 9.3.)
Exercise. Work out the nonzero module homomorphisms among the vari-
ous parabolic Verma modules in this example.

9.6. Formal Characters and Composition Factors

A central concern in our study of category O was the determination of for-
mal characters of the modules L(λ) for arbitrary λ ∈ h∗ . This turned out to
be a deep problem, leading to a solution (in KL theory) which is essentially
algorithmic in nature. An equivalent formulation of the problem involves
the two bases {[L(λ)]} and {[M (λ)]} of the Grothendieck group K(O): ex-
pressing one basis in terms of the other involves a unipotent integral matrix,
whose entries are to be calculated.
The problem for Op is formally similar. Here it is clear that K(Op ) may
be viewed as a subgroup of K(O) with two natural bases: one consisting of
symbols [L(λ)] and the other consisting of symbols [MI (λ)] with λ running
over Λ+
I . Again the problem is to express one basis in terms of the other,
which again involves a unipotent integral matrix.
Given what is already known about O, the formal character problem for
Op is already solved in principle. We observed in 9.4(1) that ch MI (λ) =
pI ∗ ch LI (λ) for all λ ∈ Λ+ +
I . This in turn is a Z -linear combination of
ch L(λ) (counted once) and various ch L(µ) with µ ∈ Λ+ I linked to λ by W
and µ < λ. To compute the coefficients one can proceed indirectly, starting
with the Weyl–Kostant formula for gI (extended to lI ):
X
(1) ch LI (λ) = (−1)`(w) ch VI (w · λ).
w∈WI

Recall that the length function on WI is just the restriction of the length
function on W , so there is no conflict here in the use of `(w): see 0.3. Also,
Exercise 9.4 shows that pI ∗ ch VI (µ) = ch M (µ). To put the pieces together,
multiply both sides of (1) (in the convolution sense) by pI to get (as in
Jantzen [147, Lemma 1]):
Proposition. If λ ∈ Λ+
I , then
X
ch MI (λ) = (−1)`(w) ch M (w · λ).
w∈WI

Note that if we had developed the full BGG resolution of LI (λ), this
would correspond to an exact sequence completing the one in Theorem
9.4(b).
190 9. Parabolic Versions of Category O

Now the idea is to compute (by the methods of KL theory) the multiplic-
ities of composition factors of all M (w · λ) and add them up with alternating
signs to get the list of composition factors (with multiplicity) of MI (λ).
Example. In the case g = sl(3, C) considered in 9.5 above, we were essen-
tially following the method outlined above. Here it was reasonably straight-
forward, since the composition factor multiplicities of Verma modules are
easy to work out and are all 1. In the inverse format, all coefficients are
then ±1. Even so, the end results in Op are simpler yet. For example, for
λ ∈ Λ+ we arrive at
[L(λ)] = [MI (λ)] − [MI (sβ · λ)] + [MI (sβ sα · λ)].
In higher rank cases (especially when p is taken to be a maximal parabolic
subalgebra), the method has an obvious drawback: it requires one to work
out a large number of composition factor multiplicties when the highest
weights involved lie outside Λ+
I , even though these ultimately cancel.

Exercise. In case λ ∈ Λ+ , combine Proposition 9.6 above with the Weyl–

Kostant character formula (2.4) to express ch L(λ) as an alternating sum
of the ch MI (w · λ), with w ranging over the set W I of minimal right coset
representatives for WI \W .

9.7. Relative Kazhdan–Lusztig Theory

A more intrinsic approach to computing formal characters within the frame-
work of category Op was developed by Deodhar [79] (with refinements in
[80]) and (independently, but slightly later) by Casian–Collingwood [61].
Their goals and methods were different, but in each case the same kind of
basic combinatorial framework is introduced: a relative Hecke module. This
is a module with involution over the Hecke algebra H of W (8.2) whose basis
is in bijection with the left (or right) coset space of WI in W and can be
labelled with minimal length elements of these cosets. Since the notational
conventions in the two papers differ from ours and from each other, we just
paraphrase briefly what is done. The common theme is a construction of
relative KL polynomials.
Deodhar was looking for a treatment that would work for a general
Kac–Moody group of type G/P , not relying on a fibration G/B → G/P .
His goal was to compute the dimension of intersection cohomology groups
of Schubert varieties in these analogues of parabolic flag varieties, using
a version of KL theory involving mainly combinatorial techniques. Here
the Hecke module and relative KL polynomials are essential. He does not
treat representations in our category Op explicitly but does point out that
the relative KL polynomials can be expressed in terms of the usual KL
polynomials for a finite Weyl group W .
9.8. Projectives and BGG Reciprocity in Op 191

On the other hand, Casian and Collingwood wanted a direct approach to

representations in Op which would permit the computation of intersection
cohomology for P -orbits on the flag variety without embedding the problem
in O. This fed into the study of homomorphisms between parabolic Verma
modules, as discussed below in 9.10. As a basic tool they formulate and
prove a relative KL Conjective for Op involving regular integral weights
(using the deep geometric methods employed in the proof of the original KL
Conjecture). This is where the Hecke module and relative KL polynomials
come in.

9.8. Projectives and BGG Reciprocity in Op

To round out the picture of category Op for a fixed I ⊂ ∆ and p = pI , we have
to look for projective (or injective) objects. By imitating closely the original
arguments of BGG [27], Rocha [226] showed how to carry over all the
essential ideas from O such as BGG Reciprocity; in this treatment the results
for O are recovered as special cases when I is empty. To generalize the notion
of standard filtration, one says that a module M ∈ Op has a standard
filtration (relative to I) if it has a filtration with subquotients isomorphic
to various MI (λ). As before, the multiplicity of each parabolic Verma
module
as a subquotient is well determined by M and written M : MI (λ) .
We can formulate as follows the main conclusions:
Theorem. Fix I ⊂ ∆.
(a) The category Op has enough projectives (and by duality enough
injectives).
(b) Among the projectives are the parabolic Verma modules MI (λ) with
λ dominant.
(c) The tensor product of a projective module in Op and a finite di-
mensional module is again projective.
(d) Every projective in Op is a direct sum of indecomposables PI (λ)
indexed by Λ+
I , where PI (λ) is a projective cover of L(λ).
(e) If P ∈ Op is projective, it has a standard filtration (relative
to I).
In particular, PI (λ) has such a filtration with PI (λ) : MI (λ) = 1
and µ > λ for all other subquotients MI (µ).
(f) The analogue of BGG Reciprocity holds in Op : if λ, µ ∈ Λ+
I , then


PI (λ) : MI (µ) = [MI (µ) : L(λ)].

Since the treatment of O in Chapter 3 differs significantly from that of

BGG, we invite the reader at this point to revisit that chapter and verify
that all the steps can indeed be generalized to Op . Here is a list of the main
points to be checked:
192 9. Parabolic Versions of Category O

(1) The observations about duality in Theorem 3.3 adapt to Op . In

particular,
dim HomO (MI (µ), MI (λ)∨ ) = δλµ ,
ExtO (MI (µ), MI (λ)∨ ) = 0 for all λ, µ ∈ Λ+
I .

(2) Either method of proof of Theorem 3.6 can be modified to show

that the tensor product of MI (λ) and a finite dimensional module M has a
standard filtration with subquotients MI (λ + µ) as µ runs over the weights
of M (counting multiplicities).
(3) The properties of standard filtrations in Proposition 3.7 carry over to
Op . If M has a standard filtration here, it is U (u−
I )-free. Similarly, Theorem
3.7 adapts to Op .
(4) The analogue of Proposition 3.8 is true in Op , leading to the existence
of enough projectives as in Theorem 3.8.
(5) Theorem 3.9 adapts to Op . In particular, for all M ∈ Op we have
dim HomO (PI (λ), M ) = [M : L(λ)].

(6) Any projective module in Op has a standard filtration, as in Theorem

3.10.
(7) The analogue of BGG Reciprocity (3.11) in (f) above holds in Op .
Remark. Using the fact that O has enough projectives, one could apply
the notion of “truncation” discussed at the end of 9.3 to prove the analogous
result for Op : start by showing that P is projective in Op whenever P is
projective in O. In particular, P (λ) ∼
= PI (λ) when λ ∈ Λ+ I .

Exercise. In the case g = sl(3, C) discussed in 9.5, show that PI (sβ · λ) is

self-dual, as is PI (sβ sα · λ).

9.9. Structure of Parabolic Verma Modules

In Chapters 4–5 we were able to answer definitively several related questions
about the structure of Verma modules, using elementary methods:
• For all λ, µ ∈ h∗ , we have dim HomO M (µ), M (λ) ≤ 1, while

every nonzero homomorphism is injective (4.2).

• Each M (λ) has a unique simple submodule (4.1), which is in fact
isomorphic to a Verma module (4.2).
• The Verma module M (λ) is simple if and only if λ is antidominant:
hλ + ρ, α∨ i ∈
/ Z>0 for all α > 0 (4.8). To prove this in the general
case we had to show first the existence of embeddings M (sα · λ) ,→
M (λ) when α > 0 and sα · λ < λ (4.6).
9.10. Maps Between Parabolic Verma Modules 193

• As a consequence, we deduced that the blocks of O are parametrized

naturally by antidominant weights (4.9).
• [M (λ) : L(µ)] ≥ 1 if and only if µ ↑ λ (that is, µ is strongly

linked
to λ), which happens if and only if HomO M (µ), M (λ) 6= 0 (5.1).
Similar questions arise in the category Op but are much less straightfor-
ward to resolve; some remain open except in special cases. Where definitive
answers are lacking, we will survey the current state of the literature. The
questions often turn out to be interrelated in subtle ways, so the section
divisions below are somewhat artificial.
• What can be said about homomorphisms MI (µ) → MI (λ), includ-
ing the dimension of the Hom space and the injectivity of nonzero
maps?
• Which modules MI (λ) are simple?
• Which simple modules L(µ) occur in Soc MI (λ), and with what
multiplicity?
• When is [MI (λ) : L(µ)] 6= 0 for λ, µ ∈ Λ+
I ?
• Is there an analogue of the Jantzen filtration for MI (λ)?
• Can the radical and socle series layers of MI (λ) be determined
explicitly? Is MI (λ) rigid?

9.10. Maps Between Parabolic Verma Modules

To study the possible module homomorphisms MI (µ) → MI (λ) with λ, µ in
Λ+I , it is natural to begin

with the associated Verma modules. We know
that HomO M (µ), M (λ) 6= 0 just when µ ↑ λ, in which case this space
has dimension 1. Moreover, any nonzero homomorphism of this type is
injective. The case g = sl(3, C) with I = {α} studied in 9.5 shows already
some complications in the parabolic setting: there is no nonzero map from
MI (sβ sα · λ) to MI (λ), while the nonzero map MI (sβ · λ) → M (λ) has a
nontrivial kernel.
In spite of these potential problems, something can always be done when
λ, µ ∈ Λ+I and µ ↑ λ. Start with an embedding ϕ : M (µ) → M (λ), which is
unique up to a nonzero scalar. Then compose ϕ with the natural projection
π : M (λ) → MI (λ). If the image of π ◦ϕ is nonzero, the universal property of
MI (µ) in Op permits us to factor this map through a nonzero homomorphism
ϕI : MI (µ) → MI (λ). Otherwise set ϕI = 0. The map ϕI is called the
standard map associated with ϕ. But it may well be 0, as illustrated
above when g = sl(3, C): here the embedding ϕ : M (sβ sα · λ) → M (λ)
yields ϕI = 0.
194 9. Parabolic Versions of Category O

This notion was introduced by Lepowsky in [203, §3]. He observed that

even when a standard map is 0 there may exist a nonzero homomorphism
between the parabolic Verma modules in question. A little thought shows
that this can happen only when M (λ) has a repeated composition factor
of type L(µ): indeed, when ϕI = 0 the canonical submodule of type M (µ)
in M (λ) must be sent to 0 by the projection M (λ) → MI (λ). In his 1982
Yale thesis (see [30]) Boe refined the methods of Lepowsky. In particular,
he worked out many further examples and gave precise conditions for a
standard map to be 0. Combining their results, one gets:

Theorem. Let λ, µ ∈ Λ+ I and consider a standard map ϕI : MI (µ) → MI (λ)

in Op associated with a nonzero map ϕ : M (µ) → M (λ).

(a) The map ϕI may be 0 or may fail to be injective even if it is nonzero.

(b) If ϕI = 0 but [M (λ) : L(µ)] > 1, there might be a nonzero homo-
morphism between these parabolic Verma modules.
(c) The map ϕI = 0 if and only if µ is strongly linked to λ by a chain
of higher weights with at least one of these not lying in Λ+
I .

Here are a few comments on the proof.

(a) This was already observed in connection with Example 9.5.
(b) Many examples emerge from the discussion in Lepowsky [199]: see
pp. 225–226.
(c) Again this is illustrated in Example 9.5. Lepowsky’s Proposition 3.9
in [203] goes in this direction, but was made more precise by Boe’s Theorem
3.3 in [30].
As a corollary of (c), one recovers an earlier result of Lepowsky [203,
3.7]: if λ ∈ Λ+ and w0 < w in the Bruhat ordering with `(w0 ) − `(w) = 1,
then a standard map MI (w0 · λ) → MI (w · λ) must be nonzero. (This can be
observed in Example 9.5.)
Having seen some of the obstacles that can arise in the study of maps
between parabolic Verma modules, one has to approach with caution the
further question:


What is the dimension of HomO MI (µ), MI (λ) ?

For Verma modules this space always has dimension ≤ 1 (Theorem 4.2(b)),
which is closely related to the fact that Soc M (λ) is always simple. As
discussed further in 9.14 below, the simple socle property breaks down for
parabolic Verma modules. This was first seen when λ is singular : an ex-
ample in which Soc MI (λ) has a repeated summand L(µ) was found by
9.11. Parabolic Verma Modules of Scalar Type 195

Irving [134, 9.6] in type D4 ; here I = {α} with α corresponding to the ver-
tex of the Dynkin diagram
connected to all other vertices. It follows that
dim HomO MI (µ), MI (λ) > 1.
All of the early examples worked out for regular integral weights pro-
duced Hom spaces of dimension ≤ 1, leading Casian–Collingwood to suggest
tentatively at the end of [61, (4.18)] that this might be a general fact. The
main result of their paper

gives a computable lower bound for the dimension
of HomO MI (µ), MI (λ) in case λ, µ are regular and integral. But a note
added in proof points out an example found by Irving–Shelton [143, 5.3]
which satisfies these conditions and has a Hom space of dimension 2: this
involves the same D4 set-up as above.

9.11. Parabolic Verma Modules of Scalar Type

The determination of Hom spaces between parabolic Verma modules is one
of the open-ended problems arising in the study of Op . Typically the best
results have been made obtained under one or more restrictive hypotheses
on the weights permitted (which may for example be assumed to be integral
or regular or both) or on the type of parabolic subalgebra involved (for
example, maximal). There has also been a lot of case-by-case study of the
simple types.
Here we look at the best-studied special case. In case dim LI (λ) = 1, Boe
[30, 4.1] defines MI (λ) to be of scalar type. In other words, hλ, α∨ i = 0 for
all α ∈ I. In particular, LI (λ) is a trivial U (n−
I )-module. Therefore MI (λ)
− −
is isomorphic to U (uI ) as a U (uI )-module.
An obvious example is MI (0), but there are many others. This special
class of modules provides a good laboratory for the closer study of parabolic
Verma modules and maps between them. Initial observations were already
made by Lepowsky [201, Thm. 1.1], [203, Prop. 3.12]:

Proposition. Let λ, µ ∈ Λ+
I , and suppose MI (µ) is of scalar type. Then

(a) MI (µ) has a unique simple submodule.

(b) Any nonzero homomorphism from MI (µ) to MI (λ) is injective.
This applies in particular to any nonzero standard map.


(c) If MI (λ) is also of scalar type, then dim HomO MI (µ), MI (λ) ≤ 1.

Proof. As observed above, the assumption that dim LI (µ) = 1 shows that
MI (µ) is free of rank one as a U (u−
I )-module. The absence of zero-divisors
in U (u−
I ) then allows one to argue just as in 4.1 and 4.2. 
196 9. Parabolic Versions of Category O

It is quite possible in (b) to have ϕI = 0, as seen in the example of

9.5. Part (c) is recovered by Boe [30, 4.5] in the special case of hermitian
symmetric spaces: see 9.18 below.
Matumoto [208, 209] has explored in a more comprehensive way the
question answered in special cases earlier: When is there a nonzero homo-
morphism between two parabolic Verma modules of scalar type? His 2006
paper includes a helpful survey of previous work on the problem, then pro-
ceeds in a mostly case-by-case fashion to classify homomorphisms. Trans-
lation functors, Jantzen’s simplicity criterion, and Kazhdan–Lusztig theory
all play important roles here. A unifying theme is the idea of building ho-
momorphisms for arbitrary p from “elementary” ones based on maximal
parabolics.
Remark. Other work related to parabolic Verma modules of scalar type
has been done by Gyoja [119, 120] and Marastoni [206]. These modules
have been studied from other viewpoints as well. In particular, Oda–Oshima
[224] have worked out good generating sets for their annihilators.

9.12. Simplicity of Parabolic Verma Modules

In category O there is a clearcut criterion for a Verma module M (λ) to
be simple: the highest weight λ must be antidominant. The proof is fairly
direct for integral weights (4.4) but otherwise depends on a closer study of
embeddings of Verma modules (4.8).
The question turns out to be far more complicated in the parabolic
setting. But it is easy to get started by formulating a natural substitute for
antidominance. Consider the following condition on a weight λ ∈ Λ+ I , which
requires in effect that λ be “as antidominant as possible”:
(∗) hλ + ρ, β ∨ i 6∈ Z>0 for all β ∈ Φ+ \ ΦI .
Theorem. Assume λ ∈ Λ+
I .
(a) If λ satisfies (∗), then MI (λ) is simple.
(b) If λ is regular and MI (λ) is simple, then λ satisfies (∗).

Although one might prefer to omit the regularity condition in (b), The-
orem 9.13 below shows (as do Jantzen’s examples) that a general simplicity
criterion requires a more delicate formulation.

Proof. (a) Suppose on the contrary that MI (λ) has a composition factor
L(µ) with µ < λ and µ ∈ Λ+ I . Thanks to the character formula Proposition
9.6, L(µ) must occur as a composition factor of some M (w · λ), w ∈ WI .
If w 6= 1, it is easy to see (Exercise 9.2) that w · λ 6∈ Φ+ I . In particular,
µ < w · λ; this is also true if w = 1.
9.12. Simplicity of Parabolic Verma Modules 197

The BGG Theorem 5.1, applied to M (w · λ), implies that µ is obtained

from w · λ by a sequence of downward reflections. Let sγ with γ > 0 be the
first of these for which γ 6∈ Φ+ +
I (it must exist since µ ∈ ΛI ), while x ∈ WI
is the product of the previous reflections with w. Now µ ≤ sγ x · λ < x · λ,
forcing hx(λ + ρ), γ ∨ i ∈ Z>0 . In turn, hλ + ρ, (x−1 γ)∨ i ∈ Z>0 .
Now x is a product of simple reflections sα with α ∈ I. Since each sα
takes positive roots other than α to positive roots (0.3(1)), it follows that
x−1 takes Φ+ \ ΦI to itself. So the root β := x−1 γ violates assumption (∗).
(b) To simplify the notation, we assume here that all weights are inte-
gral. The proof again involves the BGG Theorem, but now the regularity
condition on λ allows us to shift the problem to the Bruhat ordering of W as
in 5.2. Parametrize elements in the linkage class of λ in terms of a dominant
weight λ0 , setting λ = w · λ0 for a (unique) w ∈ W . In turn, µ < λ will
translate into µ = w1 · λ0 with w1 > w (hence `(w1 ) > `(w)). By the BGG
Theorem, this is equivalent to M (µ) ,→ M (λ).
Consider such a proper embedding. In view of Proposition 9.6, the
simplicity of MI (λ) will then force L(µ) to be a composition factor of some
M (x · λ) with x ∈ WI but x 6= 1. This translates into w < xw ≤ w1 ; unless
w1 ∈ WI w, the last inequality is strict and `(w1 ) − `(w) ≥ 2.
Now the strategy for proving (b) is simple: Suppose there exists β in
Φ+ \ ΦI for which (∗) fails. Clearly sβ 6∈ WI , where all reflections are with
respect to roots in ΦI . Set w1 := sβ w. We now have w1 > w with w1 6∈ WI w
as above. Take w1 of minimal length with this property (and forget about β).
If we can show that `(w1 ) = `(w) + 1, we have contradicted the conclusion
in the previous paragraph and thereby proved (b).
We just have to rule out `(w1 ) ≥ `(w) + 2. If this inequality holds, the
fact that all adjacent elements in the Bruhat ordering differ in length by
1 (0.4(b)) yields some w2 with `(w1 ) = `(w2 ) + 2 and w1 > w2 ≥ w. In
turn, 0.4(d) says that precisely two elements of W lie strictly between w
and w2 . Since the length difference is 1, these have the form sα w2 and sβ w2
for distinct positive roots α, β (unrelated to the discarded β above).
By the minimality of w1 , we get sα , sβ ∈ WI and thus α, β ∈ Φ+ I . On
the other hand, the definition of the ordering shows the existence of γ, δ ∈
Φ+ with w1 = sγ sα w2 = sδ sβ w2 . In turn, sγ sα = sδ sβ , or sα sβ = sγ sδ .
Applying both sides to a typical element in the root lattice QΦ, we get
Qα + Qβ = Qγ + Qδ. This forces γ, δ ∈ Φ ∩ QΦI = ΦI . Finally we reach the
contradiction w1 = sγ sα w2 ∈ WI w. 
198 9. Parabolic Versions of Category O

Part (a) was found independently by Wallach [258], Conze-Berline and

Duflo [73], and Jantzen [147, Lemma 2]. (A less elegant but more elemen-
tary proof, not requiring the BGG Theorem, was later devised by Jantzen
[148, 1.17].) The proof of part (b) follows suggestions by Jantzen.

9.13. Jantzen’s Simplicity Criterion

Jantzen’s criterion for simplicity of a parabolic Verma module relies heavily
on a generalization of the determinant formula for a contravariant form
worked out for Verma modules in 5.8. Recall from 3.15 that any highest
weight module of weight λ, such as MI (λ), has an essentially unique nonzero
contravariant form. The radical of the form is the maximal submodule and
is therefore 0 precisely when the module is simple. Using this criterion for
simplicity in the case of Verma modules was unnecessary, but here there
seems to be no obvious substitute (without using higher-powered methods
based on Kazhdan–Lusztig theory).
The generalized determinant formula is a more elaborate version of the
earlier one, with a correspondingly more elaborate proof. Once this is in
hand, specializing to a concrete weight λ ∈ Λ+
I leads quickly to the simplicity
criterion. In order to state the theorem and its corollaries precisely, we have
to introduce more notation.
Start with the character formula in Proposition 9.6 for parabolic Verma
modules:
X
(1) ch MI (λ) = (−1)`(w) ch M (w · λ).
w∈WI

It is convenient to denote by θ(λ) the right side of this equation when λ ∈ h∗

is arbitrary. It is easy to check that θ(w · λ) = (−1)`(w) θ(λ) for all w ∈ WI ;
in turn, θ(λ) = 0 if w · λ = λ for some w 6= 1 in WI .
Next we single out certain sets of roots:
Ψ := Φ \ ΦI
Ψ+
λ := {β ∈ Ψ+ | hλ + ρ, β ∨ i ∈ Z>0 }
Φβ := (QΦI + Qβ) ∩ Φ if β ∈ Ψ
Φ+
β := Φβ ∩ Φ+

We also need to formulate two conditions on λ:

∨
(M) For all β ∈ Ψ+
λ , there is a γ ∈ Φβ with hλ + ρ, γ i = 0.
∨
(M+) For all β ∈ Ψ+
λ , there is a γ ∈ Φβ with hλ + ρ, γ i = 0, sβ γ ∈ ΦI .

With all this notation in place it is easy to state (but not prove!)
Jantzen’s theorem [147, Satz 4]:
9.14. Socles and Self-Dual Projectives 199

Theorem. Let λ ∈ Λ+ +
I . Then MI (λ) is simple if and only if for all β ∈ Ψλ ,
X
θ(sγ · λ) = 0,
γ

where the sum is over all γ ∈ Φ+ ∨ >0

β \ ΦI for which hλ + ρ, γ i ∈ Z .

Here are four corollaries which then follow from the theorem, the last of
which is contained in Theorem 9.12:
Corollary. Let λ ∈ Λ+
I . Then:
(a) MI (λ) is simple if and only if the sum of all θ(sβ · λ) with β ∈ Ψ+
λ
is 0.
(b) If MI (λ) is simple, then condition (M) holds.
(c) If condition (M+) holds, then MI (λ) is simple.
(d) Let λ be regular. Then MI (λ) is simple if and only if Ψ+
λ = ∅.

In [147, Satz 4] Jantzen obtains a streamlined criterion when all com-

ponents of Φ are of type A: MI (λ) is simple if and only if condition (M+)
holds for λ.
The reader might find it an interesting project to apply these criteria to
a few concrete low rank cases in which λ is singular.

9.14. Socles and Self-Dual Projectives

In category O we found that Soc M (λ) is simple for all λ ∈ h∗ . Moreover,
the highest weight of this socle is the unique antidominant weight in the
W[λ] -orbit of λ (under the dot action). Nothing quite so straightforward can
be said in the parabolic case. As remarked in 9.10 above, examples found
in the 1980s showed that socles of parabolic Verma modules need not be
simple even when the highest weights involved are well-behaved. In general
the structure of Soc MI (λ) seems to be quite difficult to work out. But in
one case it is easy to prove simplicity (following Irving [134, 4.1]):
Proposition. Let λ ∈ h∗ be dominant and regular. If λ ∈ Λ+
I , then
Soc MI (λ) is simple.

Proof. To simplify notation we assume λ ∈ Λ+ (but the result is general).

Look first at the special case when λ = 0 on hI : that is, hλ, α∨ i = 0 for
all α ∈ I. Then MI (λ) is of scalar type, so its socle is simple by Proposition
9.11(a).
P
In general, define ν = α∈∆ cα $α to be the unique weight for which
cα = hλ, α∨ i when α ∈ I and cα = 0 when α 6∈ I. Clearly µ := λ − ν ∈ Λ+
and thus lies in Λ+ I . It vanishes on hI , so Soc MI (µ) is simple. On the
200 9. Parabolic Versions of Category O

other hand, λ and µ both lie inside the dominant W -chamber. This implies
that Tλµ and Tµλ are inverse category equivalences (7.8); they interchange
MI (λ) and MI (µ). Such an equivalence preserves socles, so Soc MI (λ) is also
simple. 

It remains a problem here to specify the highest weight of the socle. The
example for g = sl(3, C) in 9.5 already illustrates the potential difficulty of
identifying this weight in advance when λ ∈ Λ+ I is specified.
Irving went on to raise a further question: Which projective modules
PI (λ) in Op are self-dual ? In category O the fact that only the antidominant
simple module in a block can occur as the socle of a Verma module singled
out this class of simple modules for special attention. It was shown in 4.10 for
integral weights and in 7.16 for more general weights that the antidominant
simples L(λ) are precisely the heads (and socles) of self-dual projectives
P (λ).
In a similar spirit, Irving [134] formulated a reasonable conjecture and
proved it in special cases. He defines λ ∈ h∗ to be a socular weight
if it occurs as a summand in the socle of some parabolic Verma module.
The conjecture is that PI (λ) is self-dual if and only if λ is socular. In an
Addendum to that paper, following suggestions of D. Garfinkle, he filled in
the outline of a general proof of this conjecture:
Theorem. Let λ ∈ Λ+
I . Then PI (λ) is self-dual if and only if λ is socular.

The proof requires a lot of machinery, including wall-crossing functors

(7.15) as well as one notion not yet introduced here which plays a crucial
role in the study of primitive ideals in U (g): the Gelfand–Kirillov dimen-
sion of a module (13.3 below). In the proof of the theorem a key step is
the characterization of socular simple modules as those of largest possible
Gelfand–Kirillov dimension.
Exercise. In the case g = sl(3, C) treated in 9.5, verify the criterion in the
theorem.

9.15. Blocks of Op
Earlier we determined the blocks of O: those involving just integral weights
are the subcategories Oχ (1.13), while those involving nonintegral weights
require a more refinemed partition of the simple modules in Oχ (4.9). Recall
that we place two simple modules in the same block if they extend nontriv-
ially; the smallest equivalence relation generated by this relation partitions
the simple modules into blocks. In turn, an indecomposable module lies in
a block provided all its composition factors do.
9.16. Analogue of the BGG Resolution 201

What can be said about Op ? Since Op = χ Oχp , the subcategories Oχp

L
again provide at least a first approximation to blocks. For arbitrary p the
precise block decomposition of Oχp has not yet been worked out completely,
so we focus mainly on integral weights to keep the notation simple. (As in
1.13, the reader is cautioned to be aware of differing meanings assigned to
the term “block” in the literature: in some papers “block” simply means a
subcategory Oχp of Op .)
It is easy to get started in the search for blocks lying in Oχp :
Proposition. Given p and a central character χ, each block B in Oχp must
contain at least one simple parabolic Verma module.

Proof. Consider the finite partially ordered set of weights λ ∈ Λ+

I for which
χ = χλ and L(λ) lies in B. If λ is minimal in this set, then no composition
factor L(µ) of M (λ) with µ < λ can satisfy µ ∈ Λ+ I . Thus MI (λ) is simple
and lies in B. 

As a result, the simplicity criteria discussed in 9.12–9.13 limit the pos-

sible blocks. Here is a summary of special cases where the blocks have been
determined.
Examples. (1) There is one easy case: when χ = χλ with λ regular integral,
then Oχp contains only one simple parabolic Verma module (Theorem 9.12)
and is therefore already a block.
(2) Enright–Shelton [93, Cor. 12.14] explain in a particular (singular!)
case how a category Oχp splits into two blocks. (Their terminology is entirely
different from ours, however.) This occurs for a Hermitian symmetric pair
(see 9.18 below).
(3) In [51] Brundan studies the case g = gl(n, C), whose representation
theory is close to that of sl(n, C) but is enriched by the presence of a center.
His techniques are mostly special to this case, where partitions of n provide
a natural parametrization of weights, and rely especially on the degenerate
cyclotomic Hecke algebra attached to a related complex reflection group. In
the introductory discussion following the statement of his Theorem 2, he
explains how it implies for arbitrary integral weights that the blocks of Op
for gl(n, C) coincide with the categories Oχp .
(4) Boe and his student K. Platt have worked out further special cases,
in the context of the determination of the representation type of blocks:
finite, tame, or wild. This builds on recent work discussed in 13.8 below.

9.16. Analogue of the BGG Resolution

In the special case when λ ∈ Λ+ ⊂ Λ+I , it is easy to derive an analogue in O
p

of the classical Weyl–Kostant character formula (2.4) for L(λ), as indicated

202 9. Parabolic Versions of Category O

in Exercise 9.6. Start with the alternating sum formula

X
ch L(λ) = (−1)`(w) ch M (w · λ).
w∈W

Then group the terms in the sum according to the right cosets WI \W ; an
alternating sum over WI occurs in each partial sum. Finally use Proposition
9.6 to obtain the rewritten formula
X
(1) ch L(λ) = (−1)`(w) ch MI (w · λ).
w∈W I

Here the sum is over the set W I of minimal length right coset representatives
in WI \W . For example, when g = sl(3, C) and I = {α} as in 9.5, we get

ch L(λ) = ch MI (λ) − ch MI (sβ · λ) + ch MI (sβ sα · λ).

As in Chapter 6 it is natural to raise the question of realizing (1) as the

Euler character of a resolution of L(λ) by parabolic Verma modules. The
analogue of the BGG resolution is given as follows:

Theorem. Let λ ∈ Λ+ and let r be the greatest length of any element in

W I . Then there is an exact sequence

0 → CrI → . . . → C1I → C0I = MI (λ) → L(λ) → 0,

where
M
CkI := MI (w · λ).
w∈W I , `(w)=k

Moreover, the map CkI → Ck−1I is nonzero on each summand MI (w · λ),

where it is a standard map.

In the example g = sl(3, C) with I = {α}, the resolution looks like

0 → L(sβ sα · λ) = MI (sβ sα · λ) → MI (sβ · λ) → MI (λ) → L(λ) → 0.

The proof follows essentially the same lines as the proof for category O.
Inspired by an earlier joint paper with Garland, which led to a weak type
of BGG resolution in a more general Kac–Moody setting, Lepowsky [203]
was able to construct both weak and strong versions of the above “relative”
resolution by adapting the original BGG arguments. Soon afterwards Rocha
[226] streamlined the proof, showing the coincidence of weak and strong
resolutions. This was done by first investigating extensions of parabolic
Verma modules. As in the
case of Verma modules, the key point is that
ExtO MI (w · λ), MI (w0 · λ) = 0 if `(w) = `(w0 ).
9.17. Filtrations and Rigidity 203

9.17. Filtrations and Rigidity

As in category O, it is natural to ask for more detailed structural information
about parabolic Verma modules as well as projectives in Op . But the results
on O summarized in 8.14–8.15 are strongly dependent on the KL Conjecture
in Vogan’s form (or even Jantzen’s Conjecture). So one cannot expect easy
answers for Op , where even the socles of Verma modules are elusive.
In spite of this, there are a number of positive results, especially on
Loewy length, which generalize in a reasonable way the results for O. The
proofs tend to follow a similar path, relying especially on the way Loewy
length changes under wall-crossing functors. While some steps are elemen-
tary, the most definitive statements rely on the deeper ideas outlined in
Chapter 8. Most of the relevant work is due to Irving and his collaborators
Collingwood and Shelton, in papers published 1985–1990. Here is a brief
overview of some of the main points:

• Irving’s papers [135] and [136] deal mainly with category O, as

discussed in 8.15. The emphasis is on computing Loewy lengths
for Verma modules and their projective covers, using wall-crossing
functors and Vogan’s form of the KL Conjecture. Special attention
is paid to working out Loewy layers for the self-dual projective in a
block. This suggests a similar description for self-dual projectives
in Op , which are classified by “socular” weights (see 9.14).

• Irving–Shelton [143] focus on the use of simple projective modules

in Op , in tandem with wall-crossing functors, to pass from singular
weights to regular weights. This process in O is effective (say for
integral weights) since L(−ρ) = P (−ρ). In particular, simple pro-
jectives always exist in type A` , leading to explicit Loewy length
formulas. But the D4 example which leads to a non-simple socle for
a parabolic Verma module also leads to non-existence of a simple
projective module.
In a related paper, Collingwood–Irving–Shelton [70] work out
the ideas further for a special class of maximal parabolic subalge-
bras (discussed below in 9.18).

• In a longer paper (modified in the separate Errata), Irving [138]

studies more broadly a filtered version of category Op . This re-
quires further development of the earlier Hecke module formalism
introduced by Deodhar and Casian–Collingwood (9.7). Here there
are four natural bases, corresponding to four types of modules in
Op : simple, parabolic Verma, projective covers, and self-dual mod-
ules filtered by parabolic Vermas (which reappear below in Chapter
204 9. Parabolic Versions of Category O

11). This leads to “filtered” characters, which correspond to mod-

ule filtrations described in §7–§9. The end results here are typically
less definitive than those in O. For example, it is shown in Propo-
sition 7.4.1 that a parabolic Verma module (with regular integral
highest weight) is rigid provided its socle is simple.
At first Irving had hoped to avoid invoking the deeper geomet-
ric ideas in KL theory, but in the Errata he notes some problems
with the details leading up to his main results and suggests how
to recover most of them in the framework of geometric methods
introduced by Beilinson–Ginzburg.
As this quick summary indicates, there are still many uncertainties about
Loewy structure for an arbitary category Op as well as about the tools
required.
Remark. Unlike the study of filtrations in O, there is no parallel role here
for a “Jantzen filtration” in parabolic Verma modules. But Jantzen does
introduce such a filtration in [147] while developing his determinant formula
for the contravariant form. This implicitly produces a “sum formula” as well.
On the other hand, it is possible to construct a filtration in MI (λ) simply
by applying the canonical map M (λ) → MI (λ) to the submodules M (λ)i .
But it is unclear how to understand this filtration intrinsically or relate it
to Jantzen’s construction.

9.18. Special Case: Maximal Parabolic Subalgebras

To round out the chapter we take a brief look at the special case when p
is a maximal (proper) subalgebra of g. The results here are often obtained
in case-by-case fashion and are both simpler and richer in detail than one
can expect for arbitrary categories Op . For example, in some categories Oχp
with p maximal the parabolic Verma modules turn out to have no repeated
composition factors.
The literature cited here is fairly representative, but not exhaustive. The
work of Boe, Casian, Collingwood, Enright, Irving, Lepowsky, and Shelton
has been especially influential.
(A) One case which has been well-studied (starting with papers of Lep-
owsky) involves a simple matrix Lie group G of real rank one: here g is
the complexified Lie algebra of G, while p is the complexified Lie algebra of
a suitable maximal parabolic subgroup. There are only a few cases of this
type; in a standard classification, with G simply connected, these are labelled
Spin(n, 1), SU(n, 1), Sp(n, 1), F4(−20) . In his lecture notes Collingwood [66]
works out carefully the connections between Lie group representations and
categories of U (g)-modules.
9.18. Special Case: Maximal Parabolic Subalgebras 205

A starting point for the papers of Boe–Collingwood [32, 33] is the ques-
tion: Can one generalize to the parabolic setting the BGG Theorem 5.1,
relating maps between parabolic Verma modules to the existence of compo-
sition factors? As already discussed in 9.10, there are significant obstacles
when p is arbitrary. But in real rank one the authors are able to describe all
Hom spaces between parabolic Verma modules; this generalizes Lepowsky’s
results for scalar type. (Gyoja [121] provides some follow-up to [32].)
In their second paper, which is largely independent of the first and re-
covers some of its results, they show that in fact all composition factors of
parabolic Verma modules have multiplicity one; but these do not correlate
precisely with homomorphisms. Using KL theory they list explicitly the
composition factors for each real rank one type. An interesting consequence
is that the number of factors in each case is bounded independently of the
rank of g. (In all of this work the weights involved are assumed to be regular
and integral.)
(B) Maximal parabolic subalgebras of another special type occur in the
theory of Hermitian symmetric pairs (g, p). These arise from the study of
Hermitian symmetric spaces G/P , which are interesting manifolds of rel-
atively low dimension. There are five families of indecomposable classical
Hermitian symmetric pairs, which are listed in [93, Table 8.1] according to
the root systems of g and p:

(A` , Ak−1 × A`−k ), (B` , B`−1 ), (C` , A`−1 ), (D` , D`−1 ), (D` , A`−1 ).

There are also two exceptional pairs (E7 , E6 ) and (E6 , D5 ). The latter are
studied in detail by Collingwood [67].
Enright–Shelton [93] go on to consider the indecomposable classical Her-
mitian symmetric pairs. After laying some general groundwork in Part I,
they compute case-by-case for Hermitian symmetric pairs the precise com-
position factors of parabolic Verma modules in the “principal block” (where
highest weights have the form w · 0). All composition factors here occur
with multiplicity one (see also Boe–Collingwood [34]). The computation is
done by concrete methods, avoiding KL theory, which permits the recovery
of the relevant KL polynomials here. From this data they also get explicit
combinatorial formulas for dimensions of the ExtiO MI (x · 0), L(w · 0) (in

our notation) for minimal coset representatives in W I .

Shelton [232] follows up this work by studying the Exti spaces between
pairs of parabolic Verma modules, again for regular integral weights. He
obtains recursion relations for their dimensions, which are explored fur-
ther by Biagioli [28] in connection with Deodhar’s parabolic version of R-
polynomials.
206 9. Parabolic Versions of Category O

Collingwood–Irving–Shelton [70] go further with the principal block,

studying filtrations on parabolic Verma modules. In this special situation
they can prove a rigidity theorem, specify the composition factors which oc-
cur on the Loewy layers, and compute the Loewy lengths. They show more-
over that the single Loewy filtration here coincides with Gabber’s weight
filtration (see 8.13).
(C) In a two-part paper, Boe–Collingwood [35] look more broadly at
pairs (g, p) for which p is a maximal parabolic. The Hecke module approach
(9.7) figures here, along with a study of P -orbit closures in the flag variety
G/B when G and P have respective Lie algebras g, p.
They first classify completely those pairs (g, p) whose regular integral
blocks are multiplicity free in the sense that the composition factor multi-
plicities of all parabolic Verma modules are 1. The list comprises Hermitian
symmetric pairs (and close relatives), real rank one pairs (and their close
relatives), several further infinite families, and a handful of special cases in
low ranks.
All multiplicity free cases turn out to share nice properties, observed
earlier in some of the special cases: (1) All parabolic Verma modules are rigid
and admit closed formulas for the simples occurring in Loewy layers. (2)
All Hom spaces between parabolic Verma modules have dimension ≤ 1; the
dimension one cases can then be specified in terms of the “Gabber lattice”
introduced by Casian–Collingwood [61]; so their Conjecture 4.3 is verified
here.

Notes
Category Op was first studied systematically by Rocha [226].
Foundations for the study of parabolic Verma modules are discussed in
detail by Jantzen [148, 1.14–1.20], who works in somewhat greater generality
than we do. The widely-used terminology “generalized Verma module” arose
in a number of early papers such as the one by Conze-Berline and Duflo [73].
Khomenko, Mazorchuk, and others study “generalized Verma modules”
in a wider sense: here one induces from simple (possibly infinite dimensional)
lI -modules after inflating to p. (Many of their papers can be found in the
references.)
The truncation functor O → Op defined in 9.3 was introduced by Irving
[138, §5] under the name “residue”. He states that it is an exact functor; in
the Errata he corrects this but points out that exactness does hold for those
short exact sequences involved in his arguments.
 Chapter 10

Projective Functors
and Principal Series

This chapter deals with the projective functors defined by Joseph Bernstein
and Sergei Gelfand [24], together with the related category of principal
series Harish-Chandra modules for a complex semisimple Lie group having
g as Lie algebra. We do not attempt to reproduce all of the somewhat
intricate arguments in [24], where the setting is more general than category
O; instead we try to motivate and explain their main theorems, indicating
connections to translation functors.
The following chapters will introduce further endofunctors of O and
related new objects. To put all of this in perspective, we start off in 10.1
with an overview of questions arising in the study of such functors.
One main theme of [24] is the fine structure of functors related to ten-
soring with finite dimensional modules. Some essential ideas are gathered in
10.2–10.3. After defining relevant module categories in 10.4, we introduce
the formal properties of projective functors in 10.5. Examples include the
translation functors studied in Chapter 7. The classification of indecompos-
able projective functors is then explained in 10.7–10.8.
Underlying much of this work is the broader goal of understanding the
category of Harish–Chandra modules for g, which carry essential information
about Lie group representations. In 10.9 we provide a brief overview of
the relevant ideas here. It turns out that category O is closely related to
the subcategory of “principal series” modules (10.10). (This will play a
further role in Chapter 12.) The work of Bernstein–Gelfand predates the
Kazhdan–Lusztig Conjecture; indeed, their paper was motivated in part by

207
208 10. Projective Functors and Principal Series

the attempt to understand the multiplicity problem for Verma modules in

the wider context of Harish–Chandra modules.
Notational conventions. In earlier chapters we usually emphasized an-
tidominant weights: the corresponding Verma modules are simple and thus
form a natural starting point for the study of other simple modules. But
here the emphasis is placed instead on dominant weights (relative to the
dot-action of W ), which can equally well be used to parametrize blocks Oλ
(1.13–4.9). This shift is motivated by the fact that dominant Verma mod-
ules are projective. In particular, Bernstein–Gelfand [24] take this as their
starting point.
To be consistent with our own previous conventions, we depart in many
ways from the notation of Bernstein–Gelfand. For example, they parame-
trize a module of highest weight λ by λ + ρ and use the usual rather than
the dot-action of W .

10.1. Functors on Category O

Though we took a few detours along the way, the unifying goal of Chapters
1–8 has been the determination of formal characters of simple modules in O.
Since 1980 a number of other problems and applications involving O have
attracted attention; but there is less unity to the resulting literature, which
reflects a variety of agendas.
Most of the early work in O focused on special objects: highest weight
modules or others (such as projectives) closely related to these. Recent work
has called attention to other families of modules, sometimes motivated by
ideas originating elsewhere in representation theory. Reinforcing the study
of specific modules in O is the study of numerous endofunctors, which in
some cases preserve blocks but in other cases relate one block to another.
We have already seen a number of examples: duality, tensoring with fi-
nite dimensional modules, translation functors, and wall-crossing functors.
Endofunctors are often related closely to the way families of modules are
constructed and filtered (or graded). Some of the later work builds on the
earlier study of Jantzen filtrations of Verma modules or standard filtrations
in other modules.
Most of the ideas introduced here can be developed up to a point in the
algebraic framework used in earlier chapters. But gradually the methods of
“modern” homological algebra (as codified in Gelfand–Manin [111]) become
indispensable, especially the notion of derived category. Connections with
algebraic geometry and with the representation theory of Lie groups also
become more prominent. Here are some typical questions about a functor
F on O.
10.1. Functors on Category O 209

• What basic properties does F have: left or right exactness? preser-

vation of blocks Oλ ? interaction with other functors such as duality
or tensoring with finite dimensional modules?
• Is there an axiomatic characterization of F?
• If a functor Fs is initially defined for each simple reflection s, do
the composites relative to different reduced expressions for w ∈ W
agree (braid relations)?
• If F is exact, what can be said about the homomorphism induced
by F on K(O) or the induced functor on the bounded derived
category Db (O)?
• How is F constructed (or its existence proved)?
• What does F do to “standard” modules: Verma modules and their
duals, simple modules, projectives, etc.? This is an essential test
question for any functor introduced.
• What applications does F have to problems in representation the-
ory, geometry or topology, etc.?
When F is exact, some questions are simpler to deal with. In particular,
part (c) of the following proposition will be helpful later on in the chapter.
Proposition. Let F be an exact functor on O and consider its restriction
to a single block B.
(a) If F vanishes on all simple modules in B, then FM = 0 for all
M ∈ B.
(b) If FM = 0 for a module M ∈ B, then F = 0 on every composition
factor of M .
(c) If FM (λ) = 0 when M (λ) ∈ B with λ dominant, then F = 0 on B.

Proof. (a) This is proved by a straightforward induction on the length of

M . The hypothesis takes care of length 1. For the induction step, let L(λ)
be any simple quotient of M and consider the short exact sequence
0 → N → M → L(λ) → 0.
By induction, FN = 0 = FL(λ), so exactness forces FM = 0.
(b) Use induction on the length of M . If the length exceeds 1, proceed
as in step (a), using a short exact sequence
0 → N → M → L(λ) → 0.
By assumption FM = 0, so exactness again forces FN = 0 = FL(λ). All
composition factors of M other than L(λ) occur in N , so we can appeal to
the induction hypothesis.
210 10. Projective Functors and Principal Series

(c) Thanks to part (a), it is enough to show that F vanishes at all simple
modules in B. The hypothesis coupled with part (b) shows that F vanishes
at all composition factors of M (λ). Finally, we appeal to Proposition 4.3
(extended to nonintegral weights in Example 5.1): because λ is dominant,
we have [M (λ) : L(µ)] > 0 for all simple modules L(µ) ∈ B. 

10.2. Tensoring With a Dominant Verma Module

In preparation for the study of “projective functors”, we consider an im-
portant special case of tensoring with a finite dimensional module. The
viewpoint is somewhat different from that in Chapter 7.
Here we shall make essential use of Remark 3.9: Let λ, µ ∈ h∗ , with λ
dominant, and suppose dim L < ∞. Then


(∗) dim HomO M (λ) ⊗ L, M (µ) = dim Lµ−λ .
The proof was based on the fact that M (λ) is projective when λ is dominant,
which in turn implies that M (λ) ⊗ L is projective (Proposition 3.8). In this
situation, it is reasonable to ask (for fixed λ but variable L) which inde-
composable projectives P (µ) can occur as direct summands of such tensor
products. An obvious necessary condition is that λ and µ be compatible:
µ − λ ∈ Λ (which is automatic when all weights are integral).

Theorem. Let λ, µ ∈ h∗ be compatible, with λ dominant. Then the following

conditions are equivalent:
(a) There exists a finite dimensional module L for which P (µ) is iso-
morphic to a direct summand of M (λ) ⊗ L.
(b) There exists a finite dimensional module L for which


HomO M (λ) ⊗ L, L(µ) 6= 0.
(c) hµ + ρ, α∨ i ≤ 0 whenever α > 0 and hλ + ρ, α∨ i = 0.
When these conditions are satisfied, L can be chosen so that P (µ) occurs
just once as a summand of the tensor product.

Observe first that (a) and (b) are equivalent, thanks to Theorem 3.9(b).
Note that (c) makes sense, since α ∈ Φ[λ] = Φ[µ] . The condition on α
means that sα ∈ Wλ◦ . So (c) is automatically true in case λ is regular, that
is, Wλ◦ = 1. At the other extreme is the case λ = −ρ, which was treated
earlier as an isolated example (4.10): here µ must be antidominant. The
technique used there will resurface here in the proof that (c) implies (a): by
choosing L = L(ν) with ν ∈ Λ+ in the W -orbit of µ − λ, we will see that
P (µ) occurs precisely once as a summand of the tensor product in (a).
10.3. Proof of the Theorem 211

Example. To understand what the theorem says, it is useful to consider

the integral weights for sl(3, C). Here there are three possibilities: the two
extreme cases just indicated along with the intermediate case when λ lies in
only one of the two hyperplanes bounding the dominant chamber (shifted as
usual by −ρ). The condition on µ is that it lies on or below this hyperplane.

10.3. Proof of the Theorem

The proof of Theorem 10.2 involves a number of steps, starting with the
easier implication.
(1) We want to show that (b) implies (c), or equivalently, that the failure
of (c) implies the failure of (b). Accordingly, we are given α > 0 for which
hλ + ρ, α∨ i = 0 while hµ + ρ, α∨ i > 0. This condition on µ and α insures (by
Verma’s Theorem 4.6) that M (sα · µ) embeds properly in M (µ). In turn,
whenever dim L < ∞ we get an embedding



HomO M (λ) ⊗ L, M (sα · µ) ,→ HomO M (λ) ⊗ L, M (µ) .
We claim this is a bijection, for which it suffices to show that the dimensions
of the Hom spaces agree. This will follow from (∗) above and the assumption
that sα · λ = λ: the Hom space on the left side has dimension equal to
dim Lsα ·µ−λ = dim Lsα ·µ−sα ·λ = dim Lsα (µ−λ) = dim Lµ−λ ,
since W -conjugates of a weight of a finite dimensional module have the same
multiplicity. But by (∗) again, this last dimension is the dimension of the
Hom space on the right side.
(2) Because P := M (λ) ⊗ L is projective, HomO (P, ?) is exact. If we set
Q := M (µ)/M (sα · µ) and consider the short exact sequence
0 → M (sα · µ) → M (µ) → Q → 0,


it follows that HomO M (λ) ⊗
L, Q = 0. But L(µ) is a quotient of Q,
forcing HomO M (λ) ⊗ L, L(µ) = 0 (again by the projective property of
P ). Thus (b) fails, as required.
(3) It remains to prove that (c) implies (b). The choice of L is easy, as
indicated in remarks before the proof: take ν ∈ Λ+ in the W -orbit of µ − λ
and set L := L(ν). Then set P := M (λ) ⊗ L(ν). Since µ − λ is an extremal
weight of L(ν), we get from (∗):


dim HomO P, M (µ) = dim L(ν)µ−λ = 1.
The goal is to show that M (µ) can be replaced here by L(µ). (Thanks to
(a), P (µ) will then occur as a summand of P with multiplicity 1.)
(4) The existence of a nonzero homomorphism P → M (µ) implies that
HomO P, L(µ0 ) 6= 0 for some composition factor L(µ0 ) of M (µ). If µ0 = µ,

we are done. Otherwise µ0 < µ. Now Theorem 5.1 insures that µ0 is strongly
212 10. Projective Functors and Principal Series

linked to µ. In particular, there exists β > 0 such that µ0 ≤ sβ · µ < µ.

Moreover, there are corresponding embeddings M (µ0 ) ,→ M (sβ ·µ) ,→ M (µ).
Since the last embedding is proper, we have hµ + ρ, β ∨ i ∈ Z>0 . In turn,
the assumption in (c) implies that we cannot have hλ + ρ, β ∨ i = 0. The
remainder of the proof is devoted to contradicting this statement, from which
it will follow that µ0 = µ as desired.
(5) Imitate the previous argument based on (∗) by taking ν 0 ∈ Λ+ in the
W -orbit of sβ · µ − λ, so that
HomO M (λ) ⊗ L(ν 0 ), M (sβ · µ) 6= 0.

Thanks to the embedding M (sβ · µ) ,→ M (µ), this also gives

HomO M (λ) ⊗ L(ν 0 ), M (µ) 6= 0.

But the nonzero dimension of this Hom space is equal to dim L(ν 0 )µ−λ =
dim L(ν 0 )ν . Thus ν ≤ ν 0 .
(6) From step (4) we have HomO
P, L(µ0 ) 6= 0. The projective property

of P implies that HomO P, M (µ0 ) 6= 0. The 0

embedding M (µ ) ,→ M (sβ ·
µ) then implies that HomO P, M (sβ · µ) 6= 0. By (∗) this Hom space
has dimension equal to dim L(ν)sβ ·µ−λ = dim L(ν)ν 0 . As a result, ν 0 ≤ ν.
Combined with step (5), we get ν 0 = ν.
(7) Now we can derive the contradiction promised in step (4) by some
computations with roots and weights. From ν 0 = ν we deduce sβ · µ − λ ∈
W (µ − λ); being an integral weight, its euclidean inner product with itself
equals that of µ − λ. Moreover, we know that hµ + ρ, β ∨ i ∈ Z>0 ; call it m.
Since sβ · µ = µ − mβ, a quick calculation therefore yields m = hµ − λ, β ∨ i.
Comparing the two formulas for m, we then conclude that hλ + ρ, β ∨ i = 0,
contradicting the assumption in (c). 

10.4. Module Categories

Now we turn to the the paper by Bernstein–Gelfand [24], in which a deeper
study is made of tensoring U (g)-modules with finite dimensional modules.
(To be consistent with our earlier notational conventions, we have to de-
part considerably from their notation in what follows.) Although our main
emphasis is still on category O, their broader viewpoint places the prob-
lems studied in a more natural setting. In particular, they focus on the role
played by the center Z := Z(g) of U (g) and the associated central characters
χ : Z → C.
Tensoring with a finite dimensional U (g)-module L makes sense in a
number of full subcategories of M := Mod U (g), for example the category
Mf whose objects are the finitely generated U (g)-modules and of course its
subcategory O. Less obvious is the case of the category MZf whose objects
10.5. Projective Functors 213

are the Z-finite U (g)-modules. Here a standard result of Kostant has to be

invoked [24, 2.5] to show that MZf is stable under this kind of tensoring.
Thanks to Theorem 1.1(e), O lies in both Mf and MZf .
In Chapter 7 we used the projection prλ from O to its direct summand
Oχ when χ = χλ . Such projections also play a useful role here, but with
emphasis on the central character χ rather than an individual weight; so we
write prχ . It makes sense to work in the larger category MZf , which also
decomposes naturally into a direct sum of subcategories indexed by central
characters where Z acts as the sum of χ and a nilpotent operator.
In more detail, let Jχ be the two-sided ideal of U = U (g) generated by
Ker χ and set Uχ := U/Jχ . Define a full subcategory of MZf by M(χ) :=
Mod Uχ . The objects here are modules on which Z acts by the character
χ: for example, Verma modules or simple modules in Oχ . More generally,
when Jχ is replaced by its nth power, the quotient algebra is written Uχn
and its module category is written Mn (χ). Then
M(χ) ⊂ M2 (χ) ⊂ · · · ⊂ M∞ (χ),
where
M∞ (χ) := {M ∈ M | for each v ∈ M, Jχn · v = 0 for n  0}.
With this set-up, it follows that each M ∈ MZf decomposes uniquely into
a direct sum of submodules M χ ∈ M∞ (χ) (with χ running over all central
characters). In turn, we get a projection operator prχ : MZf → M∞ (χ)
extending the operator on O.

10.5. Projective Functors

For a fixed finite dimensional U (g)-module L, denote by FL the functor on
M which sends M to M ⊗L and takes a U (g)-homomorphism ϕ : M → N to
the map ϕ⊗id. These functors preserve various subcategories of M, notably
O and Mf . As remarked above, Bernstein–Gelfand verify that MZf is also
stable under FL .
A number of properties of the functors follow readily from the definition:
• FL is exact and commutes with arbitrary direct sums and products.
• FL0 ◦ FL = FL⊗L0 .
• If L∗ is the usual dual module, then FL∗ is a left and right adjoint
to FL .
• FL takes projectives to projectives in the categories mentioned
above.
In a category whose objects are functors (say on M or one of its sub-
categories) and whose morphisms are natural transformations of functors,
214 10. Projective Functors and Principal Series

it makes sense to talk about direct sums of functors. (If the sum is infinite,
the term “coproduct” may be used in place of “direct sum” to contrast with
“product”.) In turn, there is a notion of indecomposable functor. Define a
projective functor on MZf or O to be a direct summand of some FL .
The rationale for this label will soon emerge. It is clear from the properties
of the FL that direct sums, or direct summands, of projective functors, are
again projective functors. Similarly, composites of projective functors are
projective. Projective functors take projective modules to other projective
modules in MZf (or O).
A central issue in the paper [24] is the classification of all indecomposable
projective functors, where as usual the zero functor does not qualify as
indecomposable. It will be seen below that Theorem 10.2 embodies much
of the necessary work for category O. Some care has to be taken to verify
that new functors arising as direct summands are in fact nonzero; for this
Proposition 10.1 is a useful tool in O.
The search for indecomposables can be started easily enough by com-
posing a given projective functor F on MZf (resp. O) in either order with
the projection to a subcategory M∞ (χ) (resp. a block Oλ ). This decom-
poses F into a direct sum of projective functors, each vanishing outside the
chosen subcategory and/or taking nonzero values only in that subcategory.
For example, when L is the trivial 1-dimensional U (g)-module, the identity
functor FL decomposes into a direct sum of projections in MZf .

Example. In the case of category O, the translation functors studied in

Chapter 7 provide natural examples of projective functors, though they do
not exhaust all possibilities. The initial definition of Tλµ in 7.1 required only
that λ and µ be compatible; then a finite dimensional module L = L(ν)
was specially chosen to have highest weight in the W -orbit of ν = µ − λ.
Letting χ := χλ and χ0 := χµ , we defined Tλµ to be prχ0 ◦FL ◦ prχ followed
by inclusion into O. In particular, Tλµ is a direct summand of FL .
In this construction the choice of L depended on the given pair λ, µ. If
instead L is given at the outset, the resulting functor FL might or might not
be a direct sum of translation functors. In any case, it is usually difficult
to analyze the effect of Tλµ on a Verma module or simple module without
imposing further constraints on λ, µ (as we did in Chapter 7). The reader
might find it useful at this point to consider which functors Tλµ are indecom-
posable, keeping in mind Proposition 10.1 and Theorem 10.2.

To get started toward the classification theorem, we have to ask how far
a projective functor can be decomposed into a direct sum (possibly infinite);
in particular, is there always a decomposition into indecomposable projec-
tive functors? The answer turns out to be affirmative for the category MZf
10.6. Annihilator of a Verma Module 215

[24, 3.3(i)], though the proof is somewhat indirect. Here we give a straight-
forward argument in the case of O, where such a decomposition must be
finite:
Proposition. Given a projective functor F on category O, its nonzero re-
striction to a block Oλ decomposes into a finite direct sum of indecomposable
functors.

Proof. In the block Oλ , where λ can be assumed to be dominant, there is a

unique projective Verma module M (λ). Thanks to Proposition 10.1, FM (λ)
must be nonzero. Since FM (λ) is again projective, it decomposes into a fi-
nite direct sum of indecomposable projectives in an essentially unique way;
say there are n summands. If F fails to be indecomposable on Oλ , it can
be written as F1 ⊕ F2 . Neither functor can kill M (λ) (again by Proposition
10.1); so each Fi M (λ) is a direct sum of fewer than n indecomposable pro-
jectives. This makes it clear that F can be decomposed into a direct sum
of at most n indecomposable projective functors. 

It is not obvious that a decomposition of the sort obtained here is unique;

for this the detailed classification in Theorem 10.8 seems to be needed.

10.6. Annihilator of a Verma Module

The proof of Theorem 10.7 below requires one fundamental fact about U (g)
which we have not so far introduced. For any U (g)-module M , it is rea-
sonable to ask for a description of the annihilator Ann M in U (g); this is
the two-sided ideal consisting of all elements which act as 0 on M (in other
words, the kernel of the representation U (g) → End M ).
Theorem (Duflo). Let λ ∈ h∗ , with associated central character χ = χλ .
Then Ann M (λ) is the two-sided ideal Jχ of U (g) generated by Ker χ. In
particular, this ideal depends only on χ. Therefore Ann M (µ) = Jχ for all
weights µ linked to λ.

One inclusion is obvious: Jχ ⊂ Ann M (λ). This was already used in

10.4. But the other inclusion is far more subtle. Duflo’s theorem originates
in his 1971 lecture in a Budapest summer school, published in [87] (along
with the BGG paper [26] and others). It relies in an essential way on earlier
work of Kostant (“separation of variables” theorem), which involves not just
U (g) but also the nilpotent variety in g. For a thorough account, see Dixmier
[84, 8.4].
Taking λ here to be antidominant, the fact (4.8) that M (λ) = L(λ)
then implies that Jχ is in fact a primitive ideal: the kernel of an irreducible
representation of U (g). Moreover, it can be shown in this setting (since we
216 10. Projective Functors and Principal Series

are working over an algebraically closed field) that the ideals Jχ are precisely
the minimal primitive ideals of U (g). This opens the way to a deeper study
of the primitive ideal spectrum and its relationship with category O: see
13.1–13.3 below.
Remark. In his 1997 Montreal lectures, Joseph [163] provides an algebraic
(rather than geometric) pathway to the proof, motivated by analogous issues
for Kac–Moody algebras and quantum groups.

10.7. Comparison of Hom Spaces

A fundamental obstacle to classifying indecomposable projective functors is
the difficulty in recognizing when a given functor is actually indecomposable.
In the case of a module, the algebra of endomorphisms is a key tool: for
example, M ∈ O is indecomposable if and only if EndO M has precisely two
involutions 0 and 1 (in other words, no nontrivial projections onto proper
direct summands). The same test applies to a functor F and its algebra
End F of endomorphisms.
Our goal is to reduce the study of morphisms of functors to the study of
morphisms of certain modules. In general, a morphism of functors ϕ : F → G
on a module category C induces for each module M ∈ C a homomor-
phism ϕM
: FM → GM . In turn,
ϕ 7→ ϕM defines a natural linear map
Hom F, G → HomC FM, GM . When F = G, the functorial properties
insure that this becomes an algebra homomorphism End F → EndC FM .
To get more detailed information in this direction, we focus at first on
the subcategory M(χ) attached to a central character χ; its intersection
with Oχ contains for example all Verma modules M (λ) with χ = χλ as
well as their simple quotients. Define a projective χ-functor to be a
direct summand of some FL (χ), the restriction of FL to the category M(χ).
The main technical tool for the classification of indecomposable projective
functors is the following theorem, drawn from [24, 3.5, 3.7]. This reduces
matters largely to the study of projectives in category O.
Theorem. Fix a central character χ.
(a) Let F, G be projective χ-functors. If χ = χλ , then the natural map



iλ : Hom F, G → HomO FM (λ), GM (λ)
is a monomorphism.
(b) The map iλ is an isomorphism in case λ is either dominant or
antidominant.
(c) The restriction F ∞ (χ) of a projective functor F to M∞ (χ) is com-
pletely determined by its restriction F(χ) to M(χ). Any morphism
ϕ : F(χ) → G(χ) extends to a morphism ϕ b : F ∞ (χ) → G ∞ (χ); this
10.7. Comparison of Hom Spaces 217

is an isomorphism if ϕ is and can be chosen to be idempotent if ϕ

is idempotent.

Here we outline the methods used in the proof, referring to [24] for full
details. Duflo’s Theorem 10.6 plays an essential role in the proof, together
with some classical results of Kostant. For the proofs of (a) and (b), one
first makes an easy reduction to the case when F = FL and G = FL0 for
some finite dimensional modules L, L0 .
(a) To show that iλ is injective, we exploit the U (g)-module Uχ = U/Jχ ,
where as in 10.4 Jχ is the ideal in U generated by Ker χ. The category of
Uχ -modules is just M(χ), with Uχ itself playing a sort of universal role in
this category (though it does not lie in O). Now a morphism ϕ : F → G
induces a module homomorphism
ϕUχ : FUχ → GUχ .
By choosing ordered bases
P (vj ) in L and (vi0 ) in L0 , we can express this map
0
concretely as 1⊗vj 7→ i uij ⊗vi for some uij ∈ Uχ . In turn, for an arbitrary
M ∈ M(χ), the U -map Uχ → M defined by u 7→ u · v P allows us to write
the map ϕM : M ⊗ L → M ⊗ L explicitly as v ⊗ vj 7→ i uij · v ⊗ vi0 . In
0

the special case M = M (λ), the assumption iλ (ϕ) = 0 = ϕM (λ) translates

into uij · v = 0 for all v ∈ M (λ). In other words, uij annihilates M (λ), so
uij = 0 in Uχ by Duflo’s Theorem. Now ϕM = 0 for all M , forcing ϕ = 0.
(b) Thanks to (a), we only have to show that



dim HomO FM (λ), GM (λ) ≤ dim Hom F, G .

The first step is to prove, using a standard result of Kostant, that

dim Hom F, G = dim(L ⊗ (L0 )∗ )0 , the dimension of the 0-weight space.

This involves some study of the (Uχ , U )-bimodules Uχ ⊗ L and Uχ ⊗ L0 ,

using the natural adjoint representation of g on U . (Again the argument
departs from category O.)
What can be said about the other dimension? Consider the case when
λ is dominant. Using the fact that dim L < ∞, a standard vector space
identification yields
HomO M (λ) ⊗ L, M (λ) ⊗ L0 ∼ = HomO M (λ), M (λ) ⊗ (L∗ ⊗ L0 ) .

By Theorem 3.6, the tensor product with M (λ) has a standard filtration
with quotients M (λ + ν); here ν ranges over the weights of L∗ ⊗ L0 , taken
with multiplicity. Since λ is dominant

and therefore maximal in its linkage
class, HomO M (λ), M (λ + ν) = 0 unless ν = 0 (and then the dimension is
1). So the Hom space has dimension at most dim(L∗ ⊗ L0 )0 , which is readily
seen to agree with dim(L ⊗ (L0 )∗ )0 .
218 10. Projective Functors and Principal Series

The argument when λ is antidominant is essentially the same, but based

instead on the identification
HomO M (λ) ⊗ L, M (λ) ⊗ L0 ∼ = HomO M (λ) ⊗ (L0 )∗ ⊗ L , M (λ) .

(c) The argument here is in the spirit of other lifting
arguments in alge-
bra, using the inverse limit of spaces Hom F n (χ), G n (χ) for the restrictions
of functors to Mn (χ), together with the “Jχ -adic topology”.
By combining the parts of the theorem, one sees when λ is dominant
that the indecomposability of a projective functor on M∞ (χ) is equivalent
to indecomposability of the module FM (λ). In general, a decomposition of
FM (λ) into indecomposable summands will yield a corresponding decom-
position of F.

10.8. Classification Theorem

A central result of [24] is the classification of all indecomposable projective
functors on MZf . To formulate the theorem we need a little more notation.
The group W acts on the set of all ordered pairs of compatible weights,
with w sending (λ, µ) to (w · λ, w · µ). Denote the set of orbits for this action
by Ξ. For each ξ ∈ Ξ, there clearly exists at least one pair (λ, µ) in this orbit
for which λ is dominant and µ is minimal in its dot-orbit under the isotropy
group Wλ◦ . Call such a pair proper for ξ. There may be many such pairs, one
for each block if λ 6∈ Λ. These blocks Oλ are indexed by dominant weights
λ lying in different cosets modulo Λr . The extreme case involves singleton
blocks; it is in a sense the most “generic” (but least interesting) situation.
Theorem 3.3 in [24] is formulated in the wider context of MZf but
then adapts readily to O if we take blocks into account when dealing with
nonintegral weights.
Theorem. There is a natural bijection ξ 7→ Fξ between the set Ξ and the
set of (isomorphism classes of ) indecomposable projective functors on MZf ,
satisfying the condition: If (λ, µ) is a proper pair for ξ ∈ Ξ, then Fξ sends
M (λ) to P (µ).

Proof. (1) Given an indecomposable projective functor F on O, it follows

from the discussion in 10.5 that it must be zero outside a single block Oλ
(with λ dominant). By Proposition 10.1, the projective module FM (λ) must
be nonzero. Applying Theorem 10.7, we see that FM (λ) is indecomposable,
thus isomorphic to some P (µ). By Theorem 10.2, the pair (λ, µ) properly
represents its W -orbit ξ ∈ Ξ. So we can set Fξ := F.
(2) In the other direction, let (λ, µ) be a proper pair for ξ ∈ Ξ. By
Theorem 10.2, P (µ) is a direct summand of FL M (λ) for some finite dimen-
sional module L; indeed, L can be chosen so that P (µ) occurs just once
10.9. Harish-Chandra Modules 219

as a summand. Use Proposition 10.5 (generalized to MZf ) to decompose

FL into a direct sum of indecomposable functors. One (but only one!) of
these, call it Fξ , must take M (λ) to P (µ) ⊕ P with P projective. Since Fξ
is indecomposable, Theorem 10.7 forces Fξ M (λ) to be indecomposable as
well. Thus P = 0 and Fξ M (λ) = P (µ).
Combining these steps yields the desired bijection. 
Example. Recall from 7.15 the discussion of wall-crossing functors Θs at-
tached to simple reflections in W (or W[λ] ). Given antidominant weights
λ, µ ∈ h∗ with λ regular and µ\ lying just in the s-wall of the Weyl chamber
containing λ\ , one defines Θs to be the composite Tµλ Tλµ . This is a projective
functor, 0 outside the block of λ. Combining Theorem 7.6 with Proposition
7.13, we get Θs M (wλ · λ) ∼= P (wλ s · λ); here wλ · λ is the dominant weight
in the block. In particular, this result is independent of µ. It follows from
the above theorem that the projective functor Θs is also independent of µ.

10.9. Harish-Chandra Modules

As indicated at the outset, an underlying motive in [24] for the study of
projective functors is the theory of Harish–Chandra modules. Indeed, the
main external motivation for the study of category O comes from the rep-
resentation theory of semisimple (or more generally, reductive) Lie groups.
Start with a semisimple Lie group G having finite center and a maximal
compact subgroup K, then complexify their respective Lie algebras to get
Lie algebras g and k over C. For example, if G = SL(2, R) and we take
K = SO(2, R), then g ∼= sl(2, C) while k is a Cartan subalgebra h (in general,
we would get a reductive Lie algebra).
Now if (π, V ) is a well-behaved irreducible representation of G on a Ba-
nach space V , such as a unitary representation on a Hilbert space, then the
subspace V ◦ of K-finite vectors is in a natural way an irreducible g-module.
Moreover, V ◦ is the direct sum of finite dimensional simple k-modules, only
finitely many from each isomorphism class. These are the “k-types”, whose
multiplicities go a long way toward determining π (or its “character”, in the
distribution sense). Here V ◦ is a Harish-Chandra module. To develop the
theory axiomatically in the Lie algebra setting, some treatments are more
general in scope.
This theory is rich and well-organized but still incomplete in the original
context of representations of real Lie groups. Here we outline the basic con-
structions in the best understood special case: complex semisimple groups.
Such a group is actually an algebraic group, whose Lie algebra g over C is
of the type we have been studying. The construction leads in 10.10 below
to a comparison of category O with the “principal series” Harish-Chandra
220 10. Projective Functors and Principal Series

modules. The foundations are explained in a number of sources, notably

Duflo [88], Dixmier [84, Chap. 9], Jantzen [149, Kap. 6]. But the reader
has to be aware of considerable variation in notation in the literature; we
follow mainly the version of Bernstein–Gelfand [24] treated by Jantzen.
When a complex group G with Lie algebra g over C is viewed as real, the
complexified Lie algebra becomes g × g. Its enveloping algebra is isomorphic
to U ⊗ U , where U := U (g). in turn, the center of U ⊗ U is Z ⊗ Z, with Z :=
Z(g). Central characters here correspond naturally to pairs of characters
of Z. The anti-involution τ of g (see 0.5) extends to an anti-automorphism
of U and leaves Z pointwise fixed. Now k can be realized as a copy of g
embedded into g × g by x 7→ (x, −τ (x)).
Using τ , we get an isomorphism between U and its opposite algebra,
which allows us to convert left U -modules into right U -modules. Thus a
g × g-module can be identified with a (U, U )-bimodule. Given this set-up,
the linear space Hom(M, N ) with M, N ∈ Mod U (g) acquires a natural
(U, U )-bimodule structure, as does M ⊗ N . Setting Uχ := U/Jχ as before,
we can also specialize to (U, Uχ )-bimodules where part of the central action
is fixed in advance.
For our limited purposes, a Harish-Chandra module for g × g is
defined to be a finitely generated module whose restriction to k decomposes
into a direct sum of finite dimensional simple modules, each repeated only
finitely many times. Write HC for the category of all such modules. Basic
examples can be constructed by assigning to arbitrary modules M, N ∈ O
the subspace of the linear space Hom(M, N ) spanned by all k-finite vectors.
The resulting (U, U )-bimodule L(M, N ) then lies in HC. A sort of dual
construction is based on the tensor product M ⊗ N : let D(M, N ) be the
span of all k-finite vectors in (M ⊗ N )∗ . The two constructions are related
by

D(M, N ) ∼
= L(N, M ∨ ) for M, N ∈ O.

Remark. Observe that U or its quotient U/I by any two-sided ideal is

naturally a (U, U )-bimodule. When g = k is embedded as above in g × g,
it is easy to check that the resulting action of x ∈ g is the adjoint action
u 7→ xu − ux. Thanks to the PBW filtration of U , all vectors are k-finite. In
this way we get a large family of Harish-Chandra modules in the above sense.
In case M is a U -module and I is the kernel of the associated representation,
we conclude moreover that U/I embeds in L(M, M ).

One major objective is the classification and description of simple Harish-

Chandra modules. In the special case of complex groups, this rests on the
study of category O. Following work of Duflo and others, the connections
10.10. Principal Series Modules and Category O 221

between O and Harish-Chandra modules in this case were worked out in-
dependently by Bernstein–Gelfand [24], Enright [91], Joseph [156]. We
continue to follow the version of [24], which is also developed in detail by
Jantzen [149, Kap. 6].
For a fixed central character χ, the main interest lies in the category
of (U, Uχ )-bimodules. Here the (isomorphism classes of) indecomposable
projective objects turn out to be in natural 1–1 correspondence with the
elements of the set Ξ used in 10.8 to classify indecomposable projective
functors on O. If ξ 7→ Pξ and (λ, µ) is a proper pair for ξ, then
Pξ ⊗U M (λ) ∼
= P (µ) as left U (g)-modules,
illustrating the intimate connection with category O. In turn, one can show
that Pξ has a unique simple quotient, call it Lξ . This leads to a precise
classification:
Theorem. With notation as above, assign to each ξ ∈ Ξ the unique simple
quotient Lξ of the projective module Pξ . This induces a bijection between Ξ
and the isomorphism classes of simple Harish-Chandra modules for g × g.
Moreover, every Harish-Chandra module has finite length.

To describe Lξ in more detail would require a determination of multi-

plicities of the k-types occurring in its restriction to k. This is by no means
an easy problem.
In general, a Harish-Chandra module X has a filtration of Jordan–Hölder
type with quotients of the type Lξ which are uniquely determined up to
isomorphism and order of occurrence. Write [X : Lξ ] for the resulting mul-
tiplicity. It is another basic problem to determine these multiplicities. We
turn now to a special case where a reduction can be made to the analogous
multiplicity problem for category O.

10.10. Principal Series Modules and Category O

In the Lie group theory which underlies the general formalism of Harish-
Chandra modules, some group representations are easy to construct by in-
duction. Though these representations are typically not irreducible, they
provide a natural starting point. In the category HC a related class of mod-
ules can be studied under the rubric principal series modules. In the
framework of [24], these are easily specified: associate to any compatible
pair λ, µ ∈ h∗ the Harish-Chandra module X(λ, µ) := D M (λ), M (µ) .

Here are the main features, the first result being due to Duflo:
• If λ, µ ∈ h∗ are compatible, then all modules X(w · λ, w · µ) with
w ∈ W[λ] = W[µ] have the same composition factors, counting mul-
tiplicities.
222 10. Projective Functors and Principal Series

• Let λ, ν ∈ h∗ be compatible, with λ dominant. If ξ ∈ Ξ, then

(
[M (ν) : L(µ)] if (λ, µ) is a proper pair for ξ,
[X(λ, ν) : Lξ ] =
0 otherwise.
Taken together, these properties reduce the problem of computing com-
position factor multiplicities for principal series modules to the analogous
problem for Verma modules; there the KL Conjecture in its extended form
provides an answer.
Underlying the multiplicity comparison is a category equivalence [24,
5.9]. Suppose λ is dominant, with χ = χλ . In case λ is also regular, the
functor sending X to X ⊗U M (λ) takes a principal series module X to
a module in category O. This gives an equivalence between the finitely
generated (U, Uχ )-bimodules and the subcategory of O consisting of modules
whose weights are compatible with λ. In case λ is irregular, the subcategory
of O is more complicated to characterize. It consists of the modules M
which are “projectively presentable” by direct sums of projectives P (µ) with
ξ = (λ, µ) a proper pair for some ξ ∈ Ξ. If P is the collection of such
projectives, the condition on M is that there exists a short exact sequence
P 0 → P → M → 0 with P, P 0 ∈ P. (This generalizes the case when λ
is regular, since P is then the collection of all projectives having weights
compatible with λ.)
We remark that principal series modules and the category equivalence
just described also play a significant role in Chapter 12 below.

Notes
The material in this chapter is drawn primarily from Bernstein–Gelfand
[24], but with many differences in notation and emphasis. Theorem 10.2
reformulates some of the ideas in [24] along the lines of Jantzen’s treatment
in [149, 6.18–6.26]. The proof given here follows his suggestions.
Some later developments involving projective functors can be found in
papers by Khoroshkin [187], Backelin [12], Khomenko [175].
 Chapter 11

Tilting Modules

Next we introduce some objects in O called tilting modules. Such a mod-

ule is characterized by the condition that both the module and its dual
have standard filtrations. Motivation arises indirectly from the study of fu-
sion rules in parallel module categories for semisimple algebraic groups in
characteristic p, quantum enveloping algebras at a root of unity, or affine
Kac–Moody algebras. In our setting, tilting modules play a role which is in
a subtle sense dual to that of projectives.
After recalling known examples and working out some elementary prop-
erties of these modules in 11.1, we show how to parametrize the indecompos-
able ones by their highest weights (11.2). The effect of translation functors
on certain tilting modules is explained in 11.3, leading to the development of
“fusion rules” in 11.6. Then we discuss the deeper question of determining
the formal characters of tilting modules (11.7): this cannot be understood
properly without appeal to Kazhdan–Lusztig theory. Adaptations to para-
bolic categories Op are outlined in 11.8.
Early work in this direction is due to Enright–Shelton [92], Collingwood–
Irving [69], and Irving [139, §9]. Our treatment follows mainly that of
Andersen–Paradowski [6, §1]. (However, they use dominant rather than
antidominant weights to index linkage classes as well as translation functors.)
As in [6], we simplify notation by limiting the discussion to integral weights
(from 11.2 on). This amounts to working in the full subcategory Oint of
O whose objects are modules having weights in Λ. But the ideas can be
generalized to arbitrary weights.

223
224 11. Tilting Modules

11.1. Tilting Modules

We have already encountered most of the “interesting” indecomposable ob-
jects in O, but here we introduce a further class of modules which are char-
acterized in terms of standard filtrations. The indecomposable ones turn out
to be parametrized by highest weights but are rarely highest weight modules.
While they do not appear to play as central a role in O as Verma modules
and projectives modules, their counterparts in related module categories
(notably for affine Lie algebras, algebraic groups in prime characteristic,
and quantum groups at a root of unity) have been influential.
We say that M is a tilting module if both M and M ∨ have standard
filtrations in the sense of 3.7. Therefore a self-dual module with a standard
filtration is automatically a tilting module; whether all tilting modules are
self-dual is less obvious at this point but will be proved as a consequence
of the construction in 11.2. Following Donkin, the label “tilting module” is
applied here by analogy with the partial tilting modules which are important
in the study of representations of finite dimensional algebras by Brenner–
Butler, Auslander–Reiten, Ringel, and others.
Among the modules in O already studied, obvious examples of tilting
modules are the self-dual Verma modules M (λ) = L(λ) with λ antidominant.
Thanks to Theorem 4.8, there are no other self-dual Verma modules. At the
other extreme, the projective cover of an antidominant Verma module is self-
dual and has a standard filtration (7.16); it is therefore a tilting module.

Exercise. In the case g = sl(2, C), show that every indecomposable tilting
module is of one of these two types. (It is enough to consider integral
weights.)

Given a tilting module, here are some ways to obtain others:

Proposition. Let M be a tilting module. Then:

(a) M ∨ is also a tilting module.
(b) If N is a tilting module, so is M ⊕ N .
(c) Any direct summand of M is a tilting module.
(d) If dim L < ∞, then L ⊗ M is a tilting module.
(e) Applying any translation functor to M produces another tilting
module.
(f) If N is also a tilting module, then ExtnO (M, N ) = 0 for all n > 0.

Proof. Parts (a) and (b) are obvious, while (c) follows from Proposition
3.7(b).
11.2. Indecomposable Tilting Modules 225

Part (d) follows from Theorem 3.6, coupled with Exercise 3.2. Part (e)
is a consequence of Theorem and Corollary 7.6. For (f), apply Exercise 6.12,
using the fact that N ∨ has a standard filtration. 

It turns out that there is a hidden duality between tilting and projec-
tive modules in O, which will be made explicit in a number of the results
proved below. Roughly speaking, the role of the dominant and antidominant
chambers is being reversed. This is already seen in the fact that dominant
Verma modules are projective, while antidominant ones are tilting. (For a
geometric perspective, see Beilinson–Bezrukavnikov–Mirković [19].)

Remark. As we remarked earlier, the origins of “tilting” theory lie in the

representation theory of finite dimensional algebras. While that theory can
sometimes be applied to blocks of O by taking advantage of Proposition
3.13, it is easy in the current setting to work out the initial steps directly.
Background on “tilting theory” in general, with special attention to
modular representations of algebraic groups, is provided by Jantzen [153,
II.E]; this builds on the work of Ringel and Donkin. In the case of category
O, the search for self-dual modules with standard filtrations began in the
work of Collingwood–Irving [69]. The theory was later reformulated and
extended by Andersen–Paradowski [6]. Analogues were then treated by
Soergel [239, 240], inspired in part by Arkhipov’s ideas.

11.2. Indecomposable Tilting Modules

From now on we work in Oint , whose Grothendieck group K := K(Oint ) is a
subgroup of K(O). In order to classify all tilting modules, it is enough (by
Proposition 11.1(c)) to determine the indecomposable ones. It turns out that
these are neatly parametrized (up to isomorphism) by their highest weights.
Existence is proved using translation functors (Proposition 11.1(e)). For the
isomorphism statement we have to invoke Proposition 3.7(a):
(∗) If M has a standard filtration and λ is maximal among the weights
of M , then M has a submodule isomorphic to M (λ), while the quotient
M/M (λ) has a standard filtration.

Theorem. Let λ ∈ Λ.
(a) There exists an indecomposable tilting module D(λ) in Oint such
that dim D(λ)λ = 1 and all weights µ of D(λ) satisfy µ ≤ λ. (In
particular, M (λ) occurs just once in any standard filtration of D(λ)
and then occurs as a submodule.)
(b) Every indecomposable tilting module is isomorphic to some D(λ).
226 11. Tilting Modules

(c) Up to isomorphism, D(λ) is the only tilting module in Oint hav-

ing the properties in (a). It is uniquely determined by its formal
character.
(d) The symbols [D(λ)] with λ ∈ Λ form a basis of the group K.

Proof. (a) Existence is shown by an induction on length in W , using trans-

lation functors to and from walls of Weyl chambers. So we fix a regular
antidominant weight λ and consider the linked weights w · λ as well as
translated weights in walls of Weyl chambers. To get started, simply define
D(λ) := M (λ). Translating to any weight µ in the closure of the antidom-
inant Weyl chamber then produces another simple Verma module M (µ)
which is again a tilting module and can be labelled D(µ).
Now use induction on `(w), assuming that D(w · µ) exists for all weights
w · µ in the closure of the chamber containing w · λ. Thanks to Proposition
11.1(e), translating any tilting module produces another.
Having defined D(w · λ), let s be any simple reflection for which ws >
w and choose an antidominant weight µ for which Wµ◦ = {1, s}. Then
D(w · µ) = D(ws · µ) is also defined. By Theorem 7.14, the tilting module
Tµλ D(w·µ) involves in a standard filtration a nonsplit extension of two Verma
modules, a submodule M (ws · λ) and a quotient M (w · λ); this module is
indecomposable, with head L(w · λ). Moreover, ws · λ is the unique highest
weight of Tµλ D(w · λ), occurring with multiplicity 1. So the indecomposable
summand involving this weight has the required properties of D(ws · λ).
Once D(ws · λ) is defined, it can be translated to the closure of the ws-
chamber. For weights ν in walls of this chamber which are not walls of the
w-chamber already considered, an application of Theorem 7.6 shows that
the translated module has a unique occurrence of ν. So we can extract the
desired indecomposable tilting module D(ν).
(b) Let D be an arbitrary indecomposable tilting module having λ as
one of its maximal weights. By (∗) above, there is an embedding M (λ) ,→ D
as part of a standard filtration, while D/M (λ) also has a standard

filtration.
Then Proposition 11.1(f) implies that ExtO D/M (λ), D(λ) = 0. Thus the
natural map



HomO D, D(λ) → HomO M (λ), D(λ)

is surjective. This implies that the embedding M (λ) ,→ D(λ) lifts to a

homomorphism ϕ : D → D(λ). By the same kind of reasoning, there is a
homomorphism ψ : D(λ) → D which is an isomorphism on the embedded
copies of M (λ). It follows that the endomorphism ψ ◦ ϕ is an isomorphism
on Dλ , while ϕ ◦ ψ is an isomorphism on D(λ)λ . Because the modules are
11.3. Translation Functors and Tilting Modules 227

indecomposable (and of finite length), it is a standard result that any endo-

morphism must be either nilpotent (which is impossible here) or invertible.
In turn, it follows that both ϕ and ψ must be isomorphisms.
(c) This is an immediate consequence of part (b), when D is taken to be
a module with the same properties as D(λ) in (a).
(d) This follows from part (a), coupled with the easy Exercise 1.12. 
Corollary. Every tilting module is self-dual.

Proof. It clearly suffices to consider indecomposable tilting modules. Thanks

to the theorem, such a module is isomorphic to D(λ) for some λ. But D(λ)∨
is also an indecomposable tilting module. Since it has the same formal char-
acter as D(λ), its unique maximal weight λ has multiplicity 1. Then the
uniqueness assertion in (b) insures that D(λ)∨ ∼= D(λ). 
Remarks. (1) Starting with our definition of tilting module, there seems to
be no shorter way to obtain the corollary. The earliest work in this direction,
by Collingwood–Irving [69], proceeds in a contrary direction but leads to
essentially the same results: their goal was to determine self-dual modules
having standard filtrations.
(2) The direct sum of all D(λ) in a given block corresponds to a “tilting
module” in the original sense for a finite dimensional algebra whose module
category is equivalent to the block (see 3.13).

At this point the initial examples in 11.1 fit easily into the classifica-
tion. At one extreme, D(λ) = M (λ) = L(λ) whenever λ is antidominant.
w◦ ·λ
Moreover, L(−ρ) is also projective and injective. As a result, T−ρ M (−ρ)
is both projective and tilting; its highest weight is w◦ · λ, so D(w◦ · λ) must
be a direct summand. It is also projective, with simple quotient L(λ), hence
coincides with P (λ) (and is self-dual). More precisely, our direct construc-
tion in 4.10 shows that D(w◦ · λ) ∼ λ L(−ρ). In these cases the formal
= T−ρ
character of the tilting module is therefore known. But in general it becomes
more complicated to describe: see 11.7 below.

11.3. Translation Functors and Tilting Modules

As one might expect from the case of projective modules, it is nontriv-
ial to work out the effect of translation functors on arbitrary tilting mod-
ules. Under the standard assumptions on λ, µ, Proposition 11.1 insures that
Tλµ D(w · λ) is isomorphic to a direct sum of various D(w0 · µ); but the result
could be complicated, especially if we are translating out from a wall. The-
orem 7.11 does have a sort of mirror image here, which we work out only in
the easiest case (following Andersen–Paradowski [6, Prop. 1.11]). This will
be essential for the discussion of “fusion rules” in 11.6 below.
228 11. Tilting Modules

Fix antidominant weights λ, µ ∈ Λ, with λ regular and Wµ◦ = {1, s} for

some simple reflection s. We want to analyze Tµλ D(w · µ) when ws > w,
so ws · λ > w · λ. In this situation (or more generally, when µ is just
assumed to lie in the closure of the facet of λ), Theorem 7.12 insures that
ch Tµλ M (w · µ) = ch M (w · λ) + ch M (ws · λ). Since D(w · µ) has a standard
filtration, this guarantees at least that Tµλ D(w · µ) 6= 0. More precisely,
Theorem 7.14 gives a nonsplit short exact sequence:

(1) 0 → M (ws · λ) → Tµλ M (w · µ) → M (w · λ) → 0.

Theorem. Let λ, µ ∈ Λ be antidominant, with λ regular and Wµ◦ = {1, s} for

a simple reflection s. Assume that w ∈ W satisfies ws > w, so ws·λ > w ·λ.
Then Tµλ D(w · µ) ∼= D(ws · λ).

Proof. Since w · µ is the unique maximal weight of D(w · µ) and occurs with
multiplicity 1, applying (1) above to the various Verma modules including
M (w · µ) in a standard filtration of D(w · µ) shows that Tµλ D(w · µ) has ws · λ
as its unique maximal weight (with multiplicity 1). Thus Tµλ D(w · µ) ∼ =
D(ws · λ) ⊕ D for some tilting module D having all weights < ws · λ. We
just need to show that D = 0. This is equivalent to Tλµ D = 0, in view of
Theorem 7.6.
Now the strategy is to apply Tλµ to Tµλ D(w · µ). According to Corollary
7.12, the formal character of the resulting module is 2 ch D(w · µ). But since
it is a tilting module, we deduce

Tλµ Tµλ D(w · µ) ∼

= D(w · µ) ⊕ D(w · µ).
In particular, this module has w · µ as its unique maximal weight, with
multiplicity 2. If we can show that Tλµ D(ws · λ) already has w · µ as a weight
with multiplicity 2, it will follow that Tλµ D = 0 as desired.
Start with an embedding ϕ : M (ws·λ) → D(ws·λ) (Proposition

3.7(a)).
Dualize Theorem 6.13(c) to get
ExtO M (w · λ), D(ws · λ) = 0. Applying
the functor HomO ?, D(ws · λ) to (1) above, this allows us to extend ϕ
to a homomorphism ϕ : Tµλ M (ws · µ) → D(ws · λ). Here Ker ϕ cannot
involve the weight w · λ, lest the sequence (1) split. So ϕ induces a nonzero
homomorphism M (w · λ) → D(ws · λ)/M (ws · λ). In particular, w · λ is
a weight (obviously maximal) of the quotient module. This module has a
standard filtration, so M (w · λ) must occur in such a filtration of D(ws · λ)
along with M (ws · λ). It follows from Theorem 7.6 that w · µ is a weight of
Tλµ D(ws · λ) with multiplicity (at least) 2, as claimed. 


Corollary. Under the
hypotheses of the theorem, D(ws · λ) : M (xs · λ) =
D(ws · λ) : M (x · λ) for all x ∈ W .
11.4. Grothendieck Groups 229

Proof. Thanks to the theorem, D(ws · λ) = Tµλ D(w · µ). Since tilting
modules (hence their filtration multiplicities) are determined by their formal
characters, the corollary follows from the analogue of (1) above when w is
replaced by x if xs > x (or by xs if x > xs). 

Notice how the theorem and corollary fail when ws < w, by taking the
extreme case w = s: here D(ws · λ) = M (λ).
Remark. The classification by Collingwood–Irving of self-dual modules in
O having a standard filtration [69, Thm. 3.1] leads to precisely the tilting
modules as we have defined them. Their inductive construction relies heavily
on translation functors, refined by the Bernstein–Gelfand theory of projec-
tive functors: the idea is to start with an antidominant Verma module, cross
a wall, then extract the desired indecomposable summand, etc.

11.4. Grothendieck Groups

There is a natural action of W on K(O) or its subgroup K: Using the
Z-basis consisting of all symbols [M (λ)] with λ ∈ h∗ , an element w ∈ W
sends [M (λ)] to [M (w · λ)]. This action obviously respects the blocks. On
the other hand, the Grothendieck group R of the subcategory Ofd of Oint
whose objects are finite dimensional modules has a natural ring structure.
Since tensoring with a finite dimensional module preserves Oint , the additive
group K is then a module over R.
Proposition. The natural action of W on K commutes with the R-module
structure.

Proof. Theorem 3.6 allows us to write for all µ ∈ Λ+ and λ ∈ h∗ :

X
(1) [L(µ)] · [M (λ)] = dim L(µ)π [M (λ + π)].
π∈Λ
To prove the proposition we have to manipulate this formula, using the fact
that dim L(µ)π = dim L(µ)wπ for all w ∈ W . Here each sum is taken over
π ∈ Λ:

X
w · [L(µ)] · [M (λ)] = dim L(µ)π [M (w · (λ + π)]
π
X
= dim L(µ)π [M (w · λ + wπ)]
π
X
= dim L(µ)π [M (w · λ + π)]
π


= [L(µ)] · w · [M (λ)] .
Here the formula (1) was applied in the first and last equality. 
230 11. Tilting Modules

11.5. Subgroups of K
Given a finite dimensional module L(µ) with µ ∈ Λ+ , the tensor product
D(λ) ⊗ L(µ) decomposes into the direct sum of tilting modules D(ν) with
various multiplicities. To compute these multiplicities explicitly is a difficult
problem, tied to the computation of all formal characters ch D(ν). (See 11.7
below.) But the task becomes reasonable when λ is regular and antidomi-
nant and we look only for the contributions of other regular antidominant
weights ν to the direct sum. Indeed, the multiplicities here turn out to
be given in terms of classical formulas for decomposing finite dimensional
tensor products.

Exercise. Let g = sl(2, C), with Λ ∼

= Z. Work out the tensor products of
the form M (λ) ⊗ L(µ) with λ ≤ −2 and µ ≥ 0, as direct sums of tilting
modules. Compare the occurrence of regular antidominant summands with
the classical Clebsch–Gordan formula for L(w◦ λ) ⊗ L(µ).

We can formulate the problem conveniently in the Grothendieck group

language. It was already observed in Theorem 11.2(d) that the symbols
[D(λ)] with λ ∈ Λ form a Z-basis of K. Denote by K◦ the subgroup of K
spanned by all [D(λ)] = [M (λ)] with λ regular and antidominant; such a
weight can be expressed uniquely as w◦ ·λ for some λ ∈ Λ+ . All other [D(λ)]
then span a complementary subgroup K0 of K; thus K = K◦ ⊕ K0 . How do
these Grothendieck groups interact?

Proposition. Consider the decomposition K = K◦ ⊕ K0 .

(a) The subgroup K0 is equal to the Z-span K00 of all elements of K
which are fixed by at least one reflection sα with α > 0.
(b) The subgroup K0 of K is an R-submodule; so K◦ ∼ = K/K0 inherits
a natural R-structure.

Proof. (a) First consider λ ∈ λ which is irregular : Wλ◦ 6= 1. In the Verma

basis of K, we can write [D(λ)] as a Z+ -linear combination of symbols [M (µ)]
with µ linked to λ (thus also irregular). Since Wµ◦ 6= 1 is generated by
reflections, for each µ occurring there is some sα fixing [M (µ)]. It follows
that these symbols and therefore also [D(λ)] belong to K00 .
Suppose on the other hand that λ is regular but not antidominant. In
this case the reflection s in at least one wall of the chamber containing λ
satisfies s · λ > λ. It then follows from Corollary 11.3 that we can write
X

[D(λ)] = (D(λ) : M (w · λ) [M (w · λ) + M (ws · λ)] ,
w·λ<ws·λ
11.6. Fusion Rules 231

where the sum is taken over w ∈ W and s a simple reflection. Since the
reflection wsw−1 fixes the expression in parentheses, we are again in K00 .
This shows that K0 ⊂ K00 , as required.
In the other direction, let λ be regular and antidominant; so s · λ > λ for
all reflections s. In particular, s cannot fix [D(λ)]. The same is then true of
any element of K◦ , so this subgroup of K intersects K00 trivially. Conclusion:
K00 = K0 .
(b) Combine Proposition 11.4 with part (a). 

Part (a) shows in particular that for any λ ∈ Λ and α > 0, we have
[M (λ)] + [M (sα · λ)] ≡ 0 mod K0 .
By iteration, we get:
(1) [M (λ)] ≡ (−1)`(w) [M (w · λ)] mod K0 for all λ ∈ Λ, w ∈ W.
Note in particular that [M (λ)] ≡ 0 mod K0 whenever λ fails to be regular.

11.6. Fusion Rules

For the proof of the main result, we need to recall a classical rule for decom-
posing the tensor product of two finite dimensional simple modules when
the weights of one are explicitly known (see [6, 1.15]):
X X
(−1)`(w) dim L(µ)w·ν−λ ch L(ν).


(1) ch L(µ) ⊗ L(λ) =
ν∈Λ+ w∈W

It will be convenient here to denote by λ a typical element of Λ+ , so

that w◦ · λ denotes a typical regular antidominant weight in Λ.
Theorem. The additive isomorphism K◦ → R given by [M (w◦ ·λ)] 7→ [L(λ)]
for λ ∈ Λ+ is an isomorphism of R-modules.

Proof. If λ, µ ∈ Λ+ , so w◦ · λ is regular and antidominant, we have to

rewrite the corresponding product in K (modulo K0 ) in terms of the symbols
[M (w◦ · ν)] with ν ∈ Λ+ . Start with 11.4(1) above:
X
[L(µ)] · [M (w◦ · λ)] = dim L(µ)π [M (w◦ · λ + π)].
π∈Λ

Keeping in mind that irregular weights disappear modulo K0 , we claim that

this sum is equivalent to:
X X
(−1)`(w) dim L(µ)w·ν−λ [M (w◦ · ν)].
ν∈Λ+ w∈W

which will prove the theorem in view of (1) above.

232 11. Tilting Modules

This just requires a little bookkeeping. We may assume that w◦ · λ + π in

the first double sum is regular, so there is a unique x ∈ W with x·(w◦ ·λ+π) =
w◦ · ν, where ν ∈ Λ+ . Solving, we get π = x−1 w◦ · ν − w◦ · λ. Thus
w◦ π = (w◦ x−1 w◦ ) · ν − λ. Since w◦ π has the same multiplicity in L(µ) as π,
while w := w◦ x−1 w◦ has the same parity as x−1 or x, we get a term (modulo
K0 ) of the desired form by applying 11.5(1). 

With the theorem in hand, we can see explicitly how to answer the ques-
tion raised earlier about extracting from the tensor product of a (regular) an-
tidominant Verma module and a finite dimensional module the tilting sum-
mands which belong to other regular antidominant weights. If λ and µ lie in
Λ+ , define the reduced tensor product M (w◦ · λ) ⊗ L(µ) to be the direct
sum of those tilting summands (counting multiplicity) D(w◦ · ν) = M (w◦ · ν)
with ν ∈ Λ+ which occur in the ordinary tensor product. According to
the theorem, the multiplicities can be recovered from the classical ones for
L(λ) ⊗ L(µ) given by (1). (Note that in the proof of the theorem, we have
added some details omitted in [6], where there is also a sign error in the
subscript attached to L(µ).)
Remark. There is a formally parallel development for projective modules,
which we sketch briefly. The group K becomes the Z-span in K(O) of
the symbols [P (λ)] with λ ∈ Λ. Here K0 and K◦ are the complementary
subgroups spanned respectively by the [P (λ)] with λ ∈ / Λ+ and λ ∈ Λ+ .
In the latter case, P (λ) = M (λ). As above, K0 is stable under the natural
action of R, inducing an R-module structure on the quotient K/K0 ∼ = K◦ .
+
There is again a reduced tensor product: If µ ∈ Λ , then
P (λ) ⊗ L(µ) ∼
M
= nνλµ P (ν).
ν∈Λ

If λ ∈ Λ+ , define
M
P (λ) ⊗ L(µ) := nνλµ P (ν) and [P (λ)] · [L(µ)] := [P (λ) ⊗ L(µ)].
ν∈Λ+
Again the resulting structure constants nνλµ mirror the classical decomposi-
tion of L(λ) ⊗ L(µ) into simple summands.

11.7. Formal Characters

So far we have sidestepped one of the most natural questions about the
modules D(λ): What are their formal characters?
Equivalently, what are
the standard filtration multiplicities D(λ) : M (µ) ?
This question was not addressed directly by Collingwood–Irving [69]
(or by Andersen–Paradowski [6]). But the comments in [69, §3] are quite
suggestive. They point out a sort of duality between tilting modules and
11.7. Formal Characters 233

projectives in O, manifested in the similar methods of construction based

on projective functors (in concrete terms, extracting appropriate direct sum-
mands after applying wall-crossing functors). In the case of tilting modules,
however, one starts with an antidominant Verma module; in the case of
projective modules, one starts with a dominant Verma module. This kind
of duality between tilting and projective objects is also observed in parallel
module categories, notably in the case of Kac–Moody algebras.
To make these ideas precise, recall the proof of BGG Reciprocity in
3.11. The strategy was to relate a filtration multiplicity and a composition
factor multiplicity by equating each to the dimension of a Hom space. Here
the goal can be restated as follows for integral weights (the general case is
similar). Fix an antidominant weight λ ∈ Λ. Then for all x ≤ w in W , the
question becomes:

?
(1) D(w · λ) : M (x · λ) = [M (w◦ x · λ) : L(w◦ w · λ)]
By Theorem 3.9, the left side of (1) is equal to
dim HomO D(w · λ), M (x · λ)∨ = dim HomO M (x · λ), D(w · λ) ,

using the fact that tilting modules are self-dual.

Does this agree with the right side of (1)? In the proof of BGG Reci-
procity, the analogous argument was elementary. Here that no longer seems
to be the case (but see the remarks below). The work of Soergel [241]
implies the desired result but uses deeper geometric methods related to KL
theory; this is inspired by earlier work of Andersen in the setting of quantum
groups and characteristic p representations. Andersen’s idea was to intro-
duce a filtration in the indicated Hom space, somewhat analogous to the
Jantzen filtration. In our situation, Soergel is able to analyze the resulting
filtration layers and describe their dimensions in terms of coefficients of KL
polynomials [241, Thm. 4.4]. Specializing the parameter to 1 then yields:
Theorem (Soergel). With the above notation, for all x ≤ w in W we have:


dim HomO M (x · λ), D(w · λ) = Px,w (1).

Comparing this with the inverse form of the KL Conjecture 8.4(2), we

see that the Hom space has dimension equal to the right side of (1). Thus
the KL Conjecture coupled with BGG Reciprocity shows:
Corollary. Equality holds in (1) above. Thus


D(w · λ) : M (x · λ) = Px,w (1).



Moreover, D(w · λ) : M (x · λ) = P (w◦ w · λ) : M (w◦ x · λ) .

Soergel has pointed out that (1) can be proved without use of the KL
Conjecture. The idea is to set up (using more homological algebra than we
234 11. Tilting Modules

have developed) a self-equivalence of the subcategory of O whose objects

are modules with a standard filtration. This interchanges projective mod-
ules and tilting modules while preserving filtration multiplicities as in the
corollary. (In an unpublished later German version of [240], the Kac–Moody
case in §2 is supplemented in an added §3 by an explicit discussion of the
finite dimensional case.)

11.8. The Parabolic Case

The treatment of tilting modules in this chapter extends for the most part
to a parabolic subcategory Op , defined in Chapter 9 relative to a subset
I ⊂ ∆; but some proofs become more complicated. We continue to focus
just on integral weights and outline briefly what can be done, referring to
Andersen–Paradowski [6] for details.
By tilting module in Op we mean a module M for which both M and
M ∨ have filtrations whose quotients are parabolic Verma modules MI (λ).
For each integral weight λ ∈ Λ+
I , the earlier arguments can be adapted to
show the existence of a unique (up to isomorphism) tilting module DI (λ)
having λ as highest weight.
Which parabolic Verma modules are also tilting modules in Op ? In O
the antidominant Verma modules play this role (being simple as well). Here
the answer is similar: MI (λ) is a tilting module if and only if MI (λ) = L(λ).
The weights λ ∈ Λ+ I satisfying this condition are somewhat complicated to
characterize, but are specified by Jantzen’s Theorem 9.13. In case λ is reg-
ular, the condition for simplicity of MI (λ) is just that λ be as antidominant
as possible:
hλ + ρ, β ∨ i < 0 for all β ∈ Φ+ \ ΦI .
At the other extreme, we can ask which tilting modules in Op are also
projective: that is, when is DI (λ) ∼= PI (λ)? For this it is necessary that
PI (λ) be self-dual. On the other hand, all projectives in Op have “standard
filtrations” by parabolic Verma modules (9.8). So the self-dual ones are
tilting modules. Whereas the self-dual projectives P (λ) in O are just those
with λ antidominant, the criterion for self-duality in Op is more complicated
(Theorem 9.14): The projective tilting modules PI (λ) in Op are those for
which L(λ) is a summand of the socle of some parabolic Verma module.
The fusion rules in O also have a close analogue in Op and lead back in
a similar way to classical multiplicity formulas; however, the proof is more
intricate than before: see [6, 1.20–1.23].
To work out the formal characters of the DI (λ) in terms of parabolic
KL data, one can again apply Soergel’s methods [239, 240, 241] to get a
duality between projective and tilting modules.
 Chapter 12

Twisting and
Completion Functors

Here we introduce briefly some further objects of O and associated functors.

There is a continuing flow of research papers involving these ideas, which
we cite as we go along. As in Chapters 10 and 11, much of the original
motivation comes from other areas of representation theory or neighboring
subjects.
The sources we draw on vary considerably in notation as well as in the
generality of weights considered. To streamline our account we assume that
all weights are integral and often regular as well; the principal block O0
is convenient for illustrative purposes. For consistency we label blocks by
antidominant weights, though some papers cited opt for dominant weights.
When using translation functors Tλµ , we usually take λ to be regular and
antidominant, while µ lies just in a single wall of the antidominant Weyl
chamber. (But some results in the literature are valid more generally.)
Here we survey three main topics, which have originated independently
over several decades but ultimately turn out to be intimately related:

• We start with shuffled Verma modules (later called twisted Verma

modules): see 12.2–12.5. Such a module has the same formal char-
acter as a Verma module, but with the composition factors rear-
ranged. There are several approaches to the construction of these
modules, including shuffling functors derived from wall-crossing
functors.
• Next we define (following Arkhipov) twisting functors (12.6). These
endofunctors on O originate quite differently from the shuffling

235
236 12. Twisting and Completion Functors

functors, but when applied systematically to Verma modules pro-

duce the same families.
• Earlier work of Enright and others involves completion functors
(12.9). These were first developed in a limited framework as a
tool for constructing “discrete series” modules for Lie groups alge-
braically. But a broader approach eventually leads to interesting
interactions with twisting functors (12.12).
As in the case of wall-crossing functors, some constructions begin with
a functor attached to a simple reflection (or simple root), followed by iter-
ation relative to a given reduced expression for each w ∈ W . Then it is a
problem to decide whether the resulting functor is independent of the choice
of reduced expression: if so, we say the braid relations are satisfied. This in
turn may permit applications to the study of knot and link invariants.

12.1. Shuffling Functors

Irving [142] studied certain modules in O which have the same formal char-
acters as Verma modules but whose composition factors occur in other ar-
rangements. He coined the term shuffled Verma modules. (The related
name twisted Verma modules emerged in some of the later literature.)
The prototype is a dual Verma module M (λ)∨ (3.3), but in general many new
modules occur. On both philosophical and practical grounds, one should not
expect to focus attention on just one Verma module but rather on all those
lying in a block; this shows up immediately in our use of wall-crossing func-
tors below.
A parallel family of modules occurs naturally in the setting of principal
series Harish-Chandra modules for a complex semisimple Lie group whose
Lie algebra is isomorphic to g: see 10.10. Without first establishing a cate-
gory equivalence relating principal series modules to O, one can attempt a
direct functorial construction in each regular block Oλ . This starts with the
wall-crossing functor Θs = Θλs (for a simple reflection s) defined in 7.14 for
a fixed block Oλ with λ antidominant and regular. Here Θs = Tµλ Tλµ for a
suitable weight µ in the s-wall of the antidominant Weyl chamber. Recall
that the adjoint property of translation functors yields for each M ∈ Oλ a
natural isomorphism

(1) HomO (Tλµ M, Tλµ M ) ∼ µ

= HomO (M, Tµλ Tλ M ).

The identity map on Tλµ M then corresponds to an adjunction morphism

κs : M → Θs M , which may or may not be injective. (For the general
notion, see Gelfand–Manin [111, II.3.24].)
12.2. Shuffled Verma Modules 237

The shuffling functor Shs = Shλs takes M to the cokernel of κs :

M → Θs M → Shs M → 0.
(Irving writes Cs rather than Shs , perhaps to suggest “cokernel”; however,
we are reserving the letter C for completion functors.) It is routine to check
that maps M → N induce maps Shs M → Shs N having the usual functorial
properties.
Before considering the effect of shuffling functors on Verma modules, we
observe that these functors are somewhat problematic in general. For one
thing, they do not commute with arbitrary translation functors. Moreover,
it is usually hard to determine whether or not κs is injective for a given M ;
this certainly fails if M 6= 0 but Tλµ M = 0. One special case can be handled
more easily. From (1) above we get immediately:
Lemma. Suppose dim EndO Tλµ M = 1. Then κs is nonzero and is up to
scalars the unique nonzero homomorphism M → Θs M . 

12.2. Shuffled Verma Modules

Fix an antidominant weight λ ∈ Λ, which for convenience we assume to be
regular in order to simplify the use of translation functors. Thus λ lies inside
the antidominant Weyl chamber for W . As above, we use an arbitrary µ
lying in just one s-wall of that chamber to define a wall-crossing functor Θs
and shuffling functor Shs . Here we consider the effect of Shs on a typical
Verma module M (w · λ) in Oλ .
Theorem. Fix a regular antidominant weight λ ∈ Λ. Let w ∈ W and let s
be a simple reflection.
(a) The adjunction morphism κs : M (w · λ) → Θs M (w · λ) is injective.
(b) If ws > w, then Shs M (ws · λ) ∼
= M (w · λ). On the other hand,
Shs M (w · λ) is an indecomposable module with simple head L(w · λ)
and formal character equal to ch M (ws · λ).
(c) For all w, the formal character of Shs M (w · λ) equals ch M (ws · λ).

Proof. (a) Recall from 7.6 that Tλµ M (w ·λ) ∼

= M (w ·µ). On the other hand,
dim EndO M (w · µ) = 1. It follows from Lemma 12.1 that κs is nonzero and
is up to scalars the unique nonzero homomorphism M (w · λ) → Θs M (w · λ).
(b) Since T µ M (w · λ) ∼
λ = M (w · µ), Theorem 7.14(a) provides a nonsplit
short exact sequence:
(1) 0 → M (ws · λ) → Θs M (w · λ) → M (w · λ) → 0.
In particular, ch Θs M (w · λ) = ch M (ws · λ) + ch M (w · λ).
238 12. Twisting and Completion Functors

The obvious fact that Θs M (w · λ) ∼= Θs M (ws · λ) allows us to deduce

from (1) coupled with part (a) that κs : M (ws · λ) → Θs M (ws · λ) is
proportional to the injection in this short exact sequence. In turn, we get
Shs M (ws · λ) ∼
= M (w · λ).
On the other hand, the fact that κs is nonzero on M (w · λ) implies that
κs M (w · λ) is a highest weight submodule in Θs M (w · λ) of weight w · λ.
If its projection to M (w · λ) in (1) were nonzero, the projection would be
surjective, contrary to the fact that the short exact sequence is nonsplit.
Thus κs maps M (w · λ) into the submodule M (ws · λ) of Θs M (w · λ). As a
nonzero map between Verma modules, κs is therefore injective.
By definition there is another short exact sequence:
(2) 0 → M (w · λ) → Θs M (w · λ) → Shs M (w · λ) → 0.
Character comparison with (1) shows that ch Shs M (w · λ) = ch M (ws · λ).
By Theorem 7.14(b), Θs M (w·λ) has a simple head isomorphic to L(w·λ).
So its nonzero quotient Shs M (w · λ) must have the same simple head and
therefore be indecomposable.
(c) In case ws > w, apply the second statement in (b). But if ws < w,
apply the first statement in (b) to w0 := ws. 

The ultimate objective here is to start with the family {M (w · λ)} of

Verma modules in Oλ indexed by W and obtain by some iteration process a
larger family of shuffled modules indexed by pairs (w, w0 ) ∈ W ×W which can
be described intrinsically. Although we can get started with a single functor
Shs , there is immediately a problem in composing such functors relative to a
reduced expression of a Weyl group element. How does one maintain control
of Sht Shs M (w · λ) and further iterations in the spirit of the above theorem?
And will the reduced expression chosen affect the outcome? On the level of
formal characters, part (c) of the theorem can be generalized to some extent:
Corollary. With λ and s as in the theorem, suppose the adjunction mor-
phism κs for M ∈ Oλ is injective. If ch M = ch M (w · λ) for some w ∈ W ,
then ch Shs M = ch M (ws · λ).

Proof. By assumption we have a short exact sequence

0 → M → Θs M → Shs M → 0.
Since Θs is an exact functor, inducing an endomorphism of K(O), we have
ch Shs M = ch Shs M (w · λ), which equals ch M (ws · λ) thanks to part (c) of
the theorem. 

In what follows we look first for a convenient characterization of a fam-

ily of shuffled Verma modules and then discuss several approaches to the
12.3. Families of Twisted Verma Modules 239

construction which will permit verification of the injectivity of the relevant

adjunction morphisms.

Example. Let g = sl(3, C). Consider the block O0 , where we write Mw :=

M (w · (−2ρ)) and Lw := L(w · (−2ρ)). Take s to be one of the simple
reflections and t the other. If w = st one can describe much of the struc-
ture of Shs Mw by comparing (1) and (2) above. Here the dominant Verma
module Mw◦ has six distinct composition factors, while Mw has just four:
Lw , Ls , Lt , L1 . The shuffled module has a quotient isomorphic to the embed-
ded submodule Mw , together with a submodule isomorphic to the quotient
Mw◦ /Mw . Moreover, Soc Shs Mw ∼ = Lts .
As observed above, it is less easy to say what Sht does to the modules
produced by Shs .

12.3. Families of Twisted Verma Modules

Andersen–Lauritzen [5] axiomatize the formal properties required for a fam-
ily {Mλ (x, y)} ⊂ Oλ to be called a family of twisted Verma modules.
Here λ may be any antidominant weight (but they use the convention that
λ is dominant). In the axioms, x and y run over W .

(S1) Mλ (1, 1) ∼
= M (λ).
(S2) Mλ (x, y) ∼
= Mλ (xs, sy) if xs > x, sy > y.
(S3) If λ is regular and µ antidominant, then Tλµ Mλ (x, y) ∼
= Mµ (x, y).
(S4) If λ is regular and ys > y for a simple reflection s, then for all
x ∈ W the adjunction morphism is injective on Mλ (x, y) and yields
a short exact sequence

0 → Mλ (x, y) → Θs Mλ (x, y) → Mλ (x, ys) → 0.

Thus Shs Mλ (x, y) ∼

= Mλ (x, ys).
(S5) If λ is regular and ys > y, then Θs Mλ (x, y) ∼
= Θs Mλ (x, ys).
In case y = 1, we will see below that (S1) generalizes to Mλ (x, 1) ∼ =
M (x · λ) for all x ∈ W . Thus (S4) implies that Mλ (x, s) is isomorphic to the
shuffled Verma module Shs M (x · λ), which by Theorem 12.2(b) has simple
head L(x · λ) and formal character equal to ch M (xs · λ).

Exercise. Let g = sl(2, C), with W = {1, s}. Identifying Λ with Z as usual,
assume that λ ≤ −1. Using (S1) − (S5), show how to describe all Mλ (x, y)
in terms of Verma modules and their duals.
240 12. Twisting and Completion Functors

12.4. Uniqueness of a Family of Twisted Verma Modules

Assuming the existence of a family of twisted Verma modules, we want
to prove that the family is uniquely determined, up to isomorphism, while
exhibiting the formal characters involved.
Theorem. Fix an antidominant weight λ ∈ Λ. If {Mλ (x, y)} with x, y ∈ W
is a collection of modules in Oλ satisfying (S1) − (S5), then:
(a) Mλ (x, 1) ∼
= M (x · λ) for all w ∈ W .
(b) If y = s1 . . . sn is a reduced expression, then for all x ∈ W we have
Mλ (x, y) ∼
= Shsn · · · Shs1 M (x · λ).
(c) ch Mλ (x, y) = ch M (xy · λ) for all x, y ∈ W .

Proof. (a) First we use (S3) to reduce the proof to the case when λ is
regular. Suppose we have already shown that Mλ (x, 1) ∼
= M (x · λ) in this
case, and let µ be antidominant. Combining (S3) with Theorem 7.6, we get
Mµ (x, 1) ∼
= Tλ Mλ (x, 1) ∼
µ
= Tλ M (x · λ) ∼
µ
= M (x · µ).

Fixing the regular weight λ, we now use induction on `(x) to complete the
proof. The case x = 1 is settled by (S1). Assume that Mλ (x, 1) ∼ = M (x · λ).
If there exists a simple reflection s with xs > x, we get from (S2) and the
induction hypothesis:
(1) Mλ (xs, s) ∼
= Mλ (x, 1) ∼
= M (x · λ).
According to (S4), Mλ (xs, 1) is isomorphic to the kernel of the morphism
Θs Mλ (xs, 1) → Mλ (xs, s). Replace the first module by Θs Mλ (xs, s) using
(S5) and then replace Mλ (xs, s) throughout by M (x · λ) using (1). The
upshot is that Mλ (xs, 1) is isomorphic to the kernel of the epimorphism
Θs M (x · λ) → M (x · λ). Comparison with 12.2(1) shows that this kernel is
isomorphic to M (xs · λ).
(b) As in part (a), use (S3) and Theorem 7.6 to reduce to the case when
λ is regular.
The proof now goes by induction on `(y). In case y = s1 , use (S4) to
get
0 → Mλ (x, 1) → Θs1 Mλ (x, 1) → Mλ (x, s1 ) → 0.
After using (a) to replace Mλ (x, 1) by M (x · λ), comparison with 12.2(2)
shows that Mλ (x, s1 ) ∼
= Shs1 M (x · λ). Now proceed inductively.
(c) If the equality of characters holds when λ is regular, apply (S2) on
the left and Theorem 7.6 on the right to get equality in general.
Now assume that λ is regular. Using induction on `(y), combine part
(b) and (S4) with Theorem 12.2(c) and Corollary 12.2. 
12.4. Uniqueness of a Family of Twisted Verma Modules 241

When λ is regular, the |W |2 modules Mλ (x, y) form an array whose rows

and columns are indexed by the elements of W taken in some fixed order
compatible with the Bruhat ordering (with the first row and column indexed
by 1). In particular, the first column consists of Verma modules.

Exercise. Using Theorem 12.2, show that for any w ∈ W , a module

Mλ (x, y) with character equal to ch M (w · λ) occurs in each row and each
column of the array. Exhibit the array of formal characters explicitly when
g = sl(3, C).

Remark. It is natural to ask which Mλ (x, y) are isomorphic to dual Verma

modules. Experimentation suggests a systematic pattern: for all x, y ∈ W ,
we expect

(2) Mλ (x, y)∨ ∼

= Mλ (xw◦ , w◦ y)

In particular, all dual Verma modules M (x · λ)∨ should belong to the family.
Here M (x · λ) ∼
= Mλ (x, 1), so (2) would force M (x · λ)∨ ∼
= Mλ (xw◦ , w◦ ). All
of this does turn out to be true, after one has constructed a family of twisted
Verma modules based on principal series Harish-Chandra modules.
It is tempting to base a proof of (2) instead on the theorem just proved.
(In [5] this is stated as a corollary of the theorem, but with a significant step
in the proof left unfinished.) The idea is to set Mλ0 (x, y) := Mλ (xw◦ , w◦ y)∨
for x, y ∈ W , then verify that these modules satisfy axioms (S1) − (S5).
The uniqueness of a family of twisted Verma modules in Oλ will yield the
asserted isomorphisms. One has to use the fact that translation functors
commute with duality (Proposition 7.1). Most of the details are routine,
but the verification of (S4) is tricky. Assuming that λ is regular and that
ys > y for a simple reflection s, we want to obtain the following short exact
sequence involving the adjunction morphism:

0 → Mλ0 (x, y) → Θs Mλ0 (x, y) → Mλ0 (x, ys) → 0.

Since `(ys) > `(y), we get `(w◦ ys) < `(w◦ y) from 0.3(3); so (w◦ ys)s > w◦ ys.
Therefore (S4) gives

0 → Mλ (xw◦ , w◦ ys) → Θs Mλ (xw◦ , w◦ ys) → Mλ (xw◦ , w◦ y) → 0.

Use (S5) to replace the middle term by Θs M (xw◦ , w◦ y). Then apply duality,
which commutes with Θs and reverses the order:

0 → Mλ (xw◦ , w◦ y)∨ → Θs Mλ (xw◦ , w◦ y)∨ → Mλ (xw◦ , w◦ ys)∨ → 0.

The question remains, however: Is the map here identifiable with the ad-
junction morphism for Mλ (xw◦ , w◦ y)∨ ?
242 12. Twisting and Completion Functors

12.5. Existence of Twisted Verma Modules

Andersen–Lauritzen [5] describe three approaches to the construction of a
family of twisted Verma modules, then prove that these all give the same
results:
(1) They reformulate Irving’s results in [142] (discussed above) based
on the category equivalence between a block of O and a category of Harish-
Chandra modules (10.10).
(2) A quite different approach identifies twisted Verma modules with
local cohomology modules for the action of an algebraic group G with Lie
algebra g on the flag variety (8.5). This only works directly for integral
weights.
(3) A third construction (more intrinsic to category O) is based on
Arkhipov’s general theory of twisting functors, which we discuss in 12.6
below. This applies to arbitrary weights.

12.6. Twisting Functors

As part of his study of “semi-infinite cohomology”, Arkhipov [8, 9] was led
to introduce twisting functors. These turn out to be closely related to
shuffling functors, but are in some ways better behaved. On the other hand,
Arkhipov’s construction is rather subtle; it will be outlined in 12.7 below.
The program is to define for each w ∈ W a functor Tw on the entire
category of U (g)-modules, then see how it restricts to O and its blocks. Some
of the main findings can be summarized as follows, in the later formulations
by Andersen–Stroppel [7].
Theorem. The twisting functors Tw with w ∈ W satisfy:
(a) Tw is right exact, but not left exact.
(b) If s is a simple reflection and ws > w, then Tws ∼
= Tw Ts .
(c) Tw preserves O as well as each block Oλ .
(d) Up to isomorphism, Tw commutes with the functor M 7→ M ⊗ L
when dim L < ∞.
(e) Tw commutes with translation functors.
(f) ch Tw M (λ) = ch M (w · λ) for all λ ∈ h∗ .

In their further development of Arkhipov’s ideas, Andersen and Strop-

pel show for example that the left derived functor of Tw defines an auto-
equivalence of the bounded derived category Db (O0 ) [7, Cor. 4.2]. They
also study how Ts acts on simple modules in O. Moreover, they prove that
twisting functors stabilize the collection of all dual Verma modules. It is
12.7. Arkhipov’s Construction of Twisting Functors 243

shown that all twisted Verma modules are indecomposable; in fact their
endomorphism algebras are one dimensional.
Remarks. (1) In [5, §7], Andersen and Lauritzen explain how twisting
functors provide an appropriate setting for the deformation theory needed
to construct precise analogues of Jantzen filtrations and sum formulas for
twisted Verma modules. This is made explicit (for regular integral weights)
when g has type B2 . Here all composition factor multiplicities are 1, so
that the sum formulas give unambiguous information about filtration layers.
Already in this small case one sees that twisted Verma modules need not
have a simple head or simple socle. (It is interesting to note that the same
layer structures appeared earlier in the context of “weight filtrations” for
related Lie group representations: see Table I in Casian–Collingwood [62].)
(2) Recently Abe [1] has worked out precise conditions for nonzero ho-
momorphisms to exist between twisted Verma modules. (This generalizes
the BGG criterion for Verma modules.) The setting here is the category
equivalence involving category O and principal series Harish-Chandra mod-
ules described in 10.10.

12.7. Arkhipov’s Construction of Twisting Functors

As promised, we explain now how Arkhipov’s functors are defined, following
the accounts in Andersen–Lauritzen [5, §6] and Andersen–Stroppel [5, §6].
Here the universal enveloping algebra U = U (g) plays an essential role.
The idea is to construct for each w ∈ W a U -bimodule denoted Sw , then
assign to a module M ∈ O the left U -module Sw ⊗U M . This is not likely to
lie in O, but if we “twist” the action of U by w we get the desired module
Tw M ∈ O. Here we use the fact that w induces an automorphism of Φ and
hence an automorphism of g or U . The resulting twisted action of U on a
module N takes a weight space Nµ to Nwµ .
The definition of Sw is somewhat opaque looking at first sight. Start
with the subalgebra n− − −1 −
w := n ∩ w n of n and call its universal enveloping
algebra Nw for short. For example, if w = sα is a simple reflection, n−
w = g−α
and Nw is the polynomial algebra in yα . The algebra Nw has a graded dual
Nw∗ , which is Z-graded in nonpositive degrees. This grading starts with the
standard grading of U by Λr , which defines a Z-grading on L U if all simple
∗
root vectors are placed in U1 . Then Nw is taken to be ∗
n≥0 (Nw )n (the
usual dual space of (Nw )n ). Somewhat miraculously, it can then be shown
that Sw is not just a left U -module but has a natural bimodule structure.
Theorem 12.6(b) indicates that the main focus of attention will be the
special case when w = sα for some α ∈ ∆ (call it s). Here the definition of Tw
given above becomes much more down-to-earth. In the study of enveloping
244 12. Twisting and Completion Functors

algebras (or other noncommutative noetherian rings), there is a kind of

noncommutative (Ore) localization relative to certain multiplicatively closed
subsets: see for example Dixmier [84, 3.6], Jantzen [149, Kap. 11]. In our
situation, use the set {yαn | n ∈ Z+ } to form an algebra U(s) containing U as
a proper subalgebra. In effect, the negative powers of yα are admitted (say
as left factors). Then U(s) /U has a natural structure of bimodule for U . It
turns out to be isomorphic to Ss . Thus Ts M is obtained by making U act
on Ss ⊗U M with a twist by s.
A basic conclusion that emerges from this construction and the proof of
Theorem 12.6 is
The collection of modules produced by applying all Tw to all
Verma modules in a given block Oλ is a family of twisted
Verma modules in the sense of 12.3 above.
The precise correlation with the earlier notation Mλ (x, y) is a little tricky to
work out. In any case, Theorem 12.6(f) shows that the resulting formal char-
acter data can be specified quite generally apart from the parametrization
of a block by a dominant or antidominant weight.
Remark. Recent work has provided an alternative definition of twisting
functors in terms of the completion functors defined earlier by Enright and
Joseph; this will be explained in 12.12.

12.8. Twisted Versions of Standard Filtrations

Having introduced twisted versions of Verma modules, it is natural to ask
whether these can be used to filter modules (for example, projective or tilt-
ing) which are known to have standard filtrations. In the case of projective
modules, this was investigated by Irving [142, §4] (though with too few
details given in the proof of his Theorem 4.1).
In this section, let λ ∈ Λ again be regular and antidominant. For projec-
tives there are two extremes: In the case P (w◦ ·λ) = M (w◦ ·λ), the standard
filtration by itself is the only way to filter this module using twisted Verma
modules. At the other extreme, P (λ) is self-dual and involves all M (w · λ) in
a standard filtration, each with multiplicity one. So its formal character is at
least compatible with the existence of many filtrations whose subquotients
are twisted Verma modules.
With these examples in mind, we can formulate Irving’s result as follows
(under our restrictive hypotheses on λ). It says in effect that the “larger” the
projective, the more ways there are to filter it by twisted Verma modules.
Fix x ∈ W . If y ranges over the parabolic subgroup of
W generated by all simple reflections s with xs > x, then
12.9. Complete Modules 245

Shy P (x · λ) ∼
= P (x · λ) and P (x · λ) has a filtration with
quotients of the form Mλ (x, y). In particular, P (λ) has
such a filtration involving all Mλ (1, y) for y ∈ W .
Mazorchuk [213] goes on to consider how shuffling (or twisting) functors
interact with the tilting modules introduced in Chapter 11. Since the self-
dual projective P (λ) is isomorphic to the tilting module with highest weight
w◦ · λ, Irving’s criterion applies at least here. In fact, Mazorchuk obtains a
uniform result. If x ∈ W , we say that a module M has an Fx -filtration if M
has a filtration with subquotients of the form Mλ (x, y) with y ∈ W . Define
an F y -filtration similarly for a given y.

Theorem. Let Oλ be a block parametrized by an antidominant weight λ ∈ Λ.

Then each tilting module in Oλ has an Fx -filtration and an F y -filtration, for
all x, y ∈ W .

The proof uses wall-crossing functors, induction on length in W , and

some complicated diagram-chasing.

12.9. Complete Modules

Here we change the subject, at least temporarily. Enright [90] introduced
a notion of “complete module” in a category of integral weight modules
having substantial overlap with O. The goal was an algebraic construction
of discrete series (or more generally, “fundamental series”) representations
for Lie groups, permitting a natural computation of related multiplicities.
Later Joseph [159] and Mathieu [207] independently broadened the con-
struction to encompass all of O. With the benefit of hindsight, the resulting
completion functors are seen to be intimately related to the twisting func-
tors discussed earlier; this permits an alternative approach to the definition
(12.12).
To provide some insight into Enright’s idea, we first sketch his original
construction [90, §3] while sidestepping his somewhat complicated proofs
(which have been superseded by the later work). Since notation varies a lot
in the papers cited below, we impose our own compromise choices.
The setting is the category I = I(g) of U (g)-modules M satisfying the
conditions:
(I1) M is a weight module with weights in Λ.
(I2) M is U (n)-finite, so each xα with α > 0 acts locally nilpotently on
M.
(I3) M is torsion-free as a U (n− )-module.
246 12. Twisting and Completion Functors

First we take g to be s = sl(2, C) with its standard basis (h, x, y) and

write I(s). Similarly, denote by O(s) the version of category O for s. (Later
s will be identified with the subalgebra sα ⊂ g for a root α > 0.)
The axioms for I(s) amount to the requirements: h acts on M by integral
weights (identified as usual with elements of Z); x acts locally nilpotently
on M ; y acts injectively on M . For Enright’s application it is important
that M is not required to be finitely generated. But if it is finitely gener-
ated, the third requirement translates into the statement that M is a free
U (Cy)-module, since this polynomial algebra is a PID. Evidently the finitely
generated modules in I(s) all lie in O(s).
For example, consider a regular (integral) block of O(s): here the only
highest weights are n and −n − 2 with n ∈ Z+ . The indecomposables in
such a block were classified in 3.12:
L(n), M (n), M (−n − 2) = L(−n − 2), M (n)∨ , P (−n − 2).
They all satisfy (I1) and (I2), but L(n) and M (n)∨ fail to satisfy (I3).
Now write M x for the subspace of M annihilated by x: this is spanned
by all maximal vectors. Call M complete if for each n ∈ Z+ , the action of
y n+1 maps Mnx isomorphically onto M−n−2 x . The reader can check that this
is equivalent to requiring that for any s-homomorphism ϕ : M (−n−2) → M ,
there is a unique s-homomorphism ϕ : M (n) → M with ϕ ◦ i = ϕ, where
i : M (−n − 2) → M (n) is the natural embedding.
When M ∈ I(s) is arbitrary, call a module M ∈ I(s) a completion of
M if M is complete and there is an injective map i : M → M for which
M /M is a locally finite U (s)-module. For example, M (n) is a completion
of M (−n − 2) when n ∈ Z+ , whereas P (−n − 2) is already complete. The
first nontrivial facts to be established are these:
Proposition. Let s ∼
= sl(2, C) and define I(s) as above.
(a) Each M in I(s) has, up to isomorphism, a unique completion C(M )
in I(s).
(b) If dim L < ∞ and M ∈ I(s), then L⊗M belongs to I(s). Moreover,
C(L ⊗ M ) ∼= L ⊗ C(M ).
(c) The assignment M 7→ C(M ) defines a covariant functor on I(s).

To illustrate the method of proof, consider first the modules M in I(s)

which lie in O(s). We may assume that M lies in a regular block of O(s)
corresponding to some fixed n ∈ Z+ . As recalled above, the indecomposables
in this block which lie in I(s) are quite limited. So the given M decomposes
as the direct sum of various M (n) with n ≥ −1 and various P (−n−2) (these
being complete) along with various M (−n − 2) (these not being complete).
12.10. Enright’s Completions 247

To construct C(M ), one just has to replace each M (−n − 2) by M (n) in

this direct sum.

12.10. Enright’s Completions

Now the problem solved by Enright is to define a similar notion of completion
for any semisimple Lie algebra g, relative to its canonical subalgebra s =
sα ∼
= sl(2, C)) for a fixed α > 0. Let Iα (g) consist of U (g)-modules M whose
restriction Ms to s lies in I(s). Obviously I(g) ⊂ Iα (g). Then:
Theorem. Fix α > 0 and s = sα as above. If M ∈ Iα (g), then C(Ms )
admits a unique g-module structure so that the embedding Ms ,→ C(Ms )
becomes an embedding of g-modules.

Enright’s proof involves a complicated “amalgamation” construction

along the lines of the one described above for sl(2, C). His theorem then
makes it reasonable to write Cα M in place of C(Ms ). This may be called
the α-completion of M : clearly Cα M is a completion of M in the above sense
for the action of s. If M happens to lie in O, the construction also makes it
clear that Cα M lies in O. In general, M 7→ Cα M defines a covariant, left ex-
act functor on Iα (g) which is idempotent and takes Verma modules to Verma
modules. Moreover, if L is finite dimensional, then Cα (L ⊗ M ) ∼
= L ⊗ Cα M ;
so these functors commute with translation functors. In case M has a con-
travariant form, this form can be extended to Cα M .
Whenever M ∈ I(g), it makes sense to iterate the application of the
functors Cα : taking a fixed reduced expression w = s1 . . . s
n in W with all
reflections si = sαi simple, just set Cw M := Cα1 ◦ · · · ◦ Cαn (M ). But then
it remains a problem to decide whether a different reduced expression will
yield the same functor Cw . For his application, Enright is able to bypass
this question; but it is taken up in later constructions (12.11).
Remark. Motivation for the introduction of completion functors came from
the search for an algebraic construction of discrete series representations of
a real semisimple Lie group G (with finite center) relative to a maximal
compact subgroup K. The representations in question occur in L2 (G) and
are not accessible by standard induction methods, though their “characters”
were determined by Harish-Chandra. Denoting by k and g the complexifi-
cations of the respective Lie algebras, the discrete series exists when g has a
“compact” Cartan subalgebra h (one which lies in k). A further problem is to
derive algebraically the known multiplicities of (finite dimensional!) simple
k-modules in a discrete series module, as specified by Blattner’s Conjecture.
The basic idea in the algebraic construction is to start with a known
simple module M (such as an antidominant Verma module in O) and use a
sequence of completions to build up a “lattice” of more complicated modules
248 12. Twisting and Completion Functors

above M indexed by the Weyl group of k; the elusive discrete series module
D is then the simple module at the “top” (analogous to L(n) if the process
for sl(2, C) begins with M (−n − 2) = L(−n − 2)). As a byproduct of this
construction, Enright obtains a resolution of D involving the modules in the
lattice, using the maps in the BGG resolutions (6.1) of finite dimensional
simple k-modules. Then the k-multiplicities in the discrete series module can
easily be read off.
The paper by Enright [90], which covers more generally the “funda-
mental series”, followed his joint work with Varadarajan (see also Wallach
[258]). His lecture notes [91] provide further exposition, while a brief sketch
is given by Humphreys [129, §4].

12.11. Completion Functors

Apart from the laborious and somewhat ad hoc construction employed by
Enright, some natural questions remained open: Do his functors satisfy
braid relations? Can the functors be extended in a reasonable way to all of
O, and if so, which features will generalize? Two rather different approaches
to these questions have emerged:
(1) Deodhar [77] uses a localization procedure to streamline the con-
struction of Cα , where we always take α to be a simple root. This leads to
a proof of the braid relations and other refinements of Enright’s results, but
still at the cost of doing complicated direct calculations. (An independent
proof of the braid relations is due to Bouaziz [44].) Moreover, the functors
are still limited to the category I.
Briefly put, Deodhar realizes Cα as a subfunctor of a functor Dα defined
concretely as follows. Setting y = yα , one defines for M ∈ I an equivalence
relation on the set of symbols {y −n · v | n ∈ Z+ , v ∈ M } by y −n · v ∼ y −m · v 0
if and only if y m · v = y n · v 0 . The set Dα M of equivalence classes admits a
natural vector space structure as well as a natural action of U (g) (via the
commutation relations); moreover, M embeds in Dα M . Then Cα M can be
defined to consist of all v ∈ Dα M on which every root vector xβ with β > 0
acts nilpotently. (In effect, Deodhar is using an Ore localization process;
this also occurs in Arkhipov’s construction of twisting functors outlined in
12.7.) The intuition behind the use of localization here can be seen in the
rank one case, where the process of completing M (−n−2) to M (n) amounts
to applying negative powers of y to a maximal vector.
Deodhar’s functors can be identified with Enright’s. Besides yielding the
braid relations, his construction shows conceptually that completions exist
and are unique up to isomorphism.
12.12. Comparison of Functors 249

In later work on the classification of simple weight modules, Mathieu

[207, Appendix] generalizes this localization technique in such a way as to
extend Deodhar’s construction to a wider range of modules including all of
O, while recovering the earlier results such as braid relations. The meth-
ods are then refined by König–Mazorchuk [194], using a totally different
approach: they translate the problem into the category of modules over a
finite dimensional algebra attached to a block of O. This streamlines the
proof of the braid relations, in particular.
(2) Joseph [159, 160] provides a much less computational approach,
defining completion functors on all of O which satisfy the main properties
found earlier including the braid relations. His construction shows in addi-
tion that a contravariant form on a module M lifts to its completion. This
motivates the work, since it suggests an algebraic approach to understanding
the Jantzen Conjecture.
Joseph’s method relies heavily on the category equivalence relating O
and a category of Harish-Chandra modules, which we outlined briefly at the
end of Chapter 10. In particular, the Harish-Chandra bimodules denoted
L(M, N ) there come into play. Take λ ∈ Λ to be dominant and regular.
Then for a fixed simple root α and any M ∈ O, set


Cα M := L M (sα · λ), M ⊗U (g) M (λ).
From this it is possible to compute all Cα L(µ) and to show that Cα M (w · λ)
is isomorphic to either M (w·λ) or M (sα w·λ). Moreover, this Cα agrees with
the Enright/Deodhar functor on I ∩ O. But Cα is no longer idempotent in
general. And it sometimes differs from Mathieu’s functor on O, as illustrated
in the table below.

12.12. Comparison of Functors

While the twisting and completion functors introduced in this chapter have
diverse motivations and histories, they turn out to be close relatives. Ma-
zorchuk and Stroppel draw together the relevant literature in [219], starting
with a broad overview of endofunctors on O defined relative to a simple root
(or reflection): for example, wall-crossing functors, shuffling functors, twist-
ing functors, or completion functors as defined by Enright, Joseph, Matthieu.
In tables and diagrams they summarize concisely the relationships among all
of these. Among the issues addressed are: left or right exactness, existence
of left or right adjoints, derived functors, morphisms between functors.
The rank one case is easy to illustrate: for a regular integral block of
sl(2, C) with n ∈ Z+ , the table below shows the effect on standard modules
of Arkhipov’s twisting functor as well as the completion functors in variants
defined by Joseph and Mathieu.
250 12. Twisting and Completion Functors

Arkhipov Joseph Mathieu

L(n) 0 0 0
M (−n − 2) M (n)∨ M (n) M (n)
M (n) M (−n − 2) M (n) M (n)
M (n)∨ M (n)∨ M (−n − 2) M (n)
P (−n − 2) P (−n − 2) P (−n − 2) P (−n − 2)
Table 1. Functors acting on a regular integral block, g = sl(2, C)

In their study of Arkhipov’s twisting functor Tw , Andersen–Stroppel

[7, §4] show that its right adjoint functor is isomorphic to D ◦ Tw−1 ◦ D,
where D is the duality functor on O; see also their Remark 5.5. Then in
[186, §5] Khomenko–Mazorchuk identify this adjoint functor with Joseph’s
completion functor:
Theorem. For w ∈ W , denote by Tw Arkhipov’s twisting functor and by
Gw Joseph’s completion functor. On a regular integral block of O, Gw is
right adjoint to Tw . If D denotes the duality on O, then moreover
(∗) Gw ∼
= D ◦ Tw−1 ◦ D.

The reader can easily check the isomorphism (∗) on standard modules for
sl(2, C) using the table; but the full isomorphism of functors requires more
work. In general, w−1 occurs here because a reduced expression w = s1 . . . sn
yields Tw = Ts1 . . . Tsn (Theorem 12.6(b)), whereas Gw = Gsn · · · Gs1 .
The methods of Khomenko–Mazorchuk [186] were inspired by the repre-
sentation theory of finite dimensional algebras, enabling them to go further
in several directions. In particular, they derive the braid relations for twist-
ing functors from the known braid relations satisfied by completion functors.
Remark. Mazorchuk and Stroppel [219, Thm. 1] develop an interesting
application to the study of knot invariants: when g = sl(n, C), well-chosen
endofunctors of O0 yield a categorification of the Baez–Birman “singular
braid monoid”. This results from assigning generators of the braid monoid
to well-chosen functors.
 Chapter 13

Complements

This final chapter is structured more loosely than the previous ones, but has
two main themes:
(1) First we sketch some of the interactions between category O and
other parts of representation theory. Here we can cite only a small sample
of the extensive literature. We have already pointed out close ties between
category O and Lie group representations. To illustrate the widespread
influence of O in other areas, we discuss both older and newer work, under
several headings:

• Universal enveloping algebras: classification of primitive ideals (13.1–

13.3); Kostant’s problem (13.4)

• Parallels to category O: Kac–Moody algebras (13.5–13.6); ana-

logues of highest weight modules in various categories, leading to
axiomatic approaches (13.7)

• Blocks: their quivers and representation types (13.8–13.10)

(2) To round out the chapter, we then look at more advanced view-
points on O which have been in the forefront of recent research: Soergel’s
theorem on endomorphism algebras of self-dual projective modules and the
coinvariant algebra of W (13.11–13.13); endomorphism algebras of arbitrary
projectives, (13.14); Koszul duality (13.15).
Concerning notation, the basic conventions used earlier remain in effect
except as indicated. But when the work discussed ranges more broadly we
sometimes follow the mildly conflicting notation used in the cited papers.

251
252 13. Complements

13.1. Primitive Ideals in U (g)

We start by looking back at a subject which reached maturity during the
decade 1972–1982 but still has unsolved problems. The universal envelop-
ing algebra U (g) is an interesting object of study in its own right. It is a
noncommutative noetherian ring, generated by any basis x1 , . . . , xn of g. A
resulting PBW basis induces a filtration of U (g), whose associated graded
algebra is isomorphic to the polynomial algebra S(g) in n indeterminates
(as seen in the proof of the PBW Theorem). The center Z(g) of U (g) is
itself a polynomial algebra in ` variables, isomorphic to the W -invariants
for the dot-action in U (h) ∼
= S(h); its characters or their kernels naturally
parametrize the W -linkage classes in h∗ .
The general theory of noetherian rings provides a range of tools for
the study of U (g), while at the same time U (g) serves as an illuminating
example for this theory. More to the point for us, the study of U (g) has
a symbiotic relationship with the representation theory of g and especially
with category O. Here we focus on primitive ideals of U (g): the kernels
of irreducible representations. Although there is no reasonable algebraic
classification of simple U (g)-modules, we can still ask for a classification of
primitive ideals. These form a partially ordered set Prim U (g), which has a
Jacobson topology (analogous to the Zariski topology in the commutative
case). While the topology on Prim U (g) is still incompletely understood, its
structure as an ordered set has been well studied.
The results quoted below can be found (with references to primary
sources) in the books by Dixmier [84] and Jantzen [149] together with the
surveys by Borho [40, 41], Jantzen [150, 151], Joseph [161]. Jantzen’s book
documents especially the extensive contributions of Joseph. We have also
listed in the references a small sample of related papers, including Borho–
Jantzen [42], Conze–Dixmier [72], Conze–Duflo [73], Duflo [87, 88, 89],
Joseph [155, 156, 158, 163]. The subject is unexpectedly rich, with links
to Kazhdan–Lusztig theory, Springer’s Weyl group representations, and the
geometry of nilpotent orbits in g.
As pointed out in the Notes to Chapter 1, every simple U (g)-module M
has a central character χM : Z(g) → C. If J = Ann M , then J ∩ Z(g) =
Ker χM . This yields a natural map π : Prim U (g) → Max Z(g), which plays a
fundamental role in the study of primitive ideals. Since Z(g) is a polynomial
ring, its maximal ideal spectrum may be identified with the affine space C` .
13.2. Classification of Primitive Ideals 253

13.2. Classification of Primitive Ideals

Even though O contains “few” of the simple U (g)-modules, the classical
work of Dixmier, Duflo, and others shows that O does encode indirectly a
lot of information about Prim U (g):
Theorem. Let g be a semisimple Lie algebra over C. Then:
(a) If λ ∈ h∗ , the annihilator of M (λ) is the two-sided ideal of U (g)
generated by Ker χλ ; thus it is the same for all weights linked to λ
and may be denoted Jχ with χ = χλ .
(b) For any central character χ : Z(g) → C, the ideal Jχ is primitive.
(c) Every primitive ideal of U (g) is of the form Ann L(λ) for some
simple highest weight module L(λ), λ ∈ h∗ ; call this ideal Jλ .
(d) As χ ranges over the central characters, the ideals Jχ are precisely
the minimal primitive ideals of U (g).

We already quoted parts (a) and (b) in connection with projective func-
tors in 10.6. While (a) is a fairly deep result, part (b) is then an easy
consequence: take χ = χλ with λ ∈ h∗ antidominant, so that M (λ) = L(λ).
Part (c) is a later theorem of Duflo [89].
To deduce (d) from the other statements, observe that Ann M (µ) ⊂
Ann L(µ) for all µ ∈ h∗ . In particular, when λ is antidominant and µ ∈ W ·λ,
it follows from (a) and (b) that Jλ = Jχ = Ann M (µ) ⊂ Jµ . Combined with
(c), this shows that such ideals Jχ are the minimal ones in Prim U (g).
The emerging picture can be summarized in a commutative diagram:
h∗ −−−−→ Prim U (g)
 
 π
y y
χ
h∗ /W −−−−→ Max Z(g)
The top horizontal map sends λ ∈ h∗ to Jλ , which π then sends to Jλ ∩Z(g) =
Ker χλ . Thanks to the theorem, both of these maps are surjective. The left
vertical map sends λ to the W -linkage class {w · λ | w ∈ W }; this too is
surjective. Using the fact that Jχ = Jλ when χ = χλ and λ is antidominant,
the lower horizontal map sends the linkage class to Ker χ. This last map is
actually bijective, since the natural correspondence between linkage classes
and central characters (or their kernels) is bijective.
It is customary to write briefly X := Prim U (g). The fibers Xχ of the
map π are the finite sets of primitive ideals Jλ with λ running over a single
linkage class (where χ = χλ ). Moreover, no inclusions can exist between
primitive ideals lying in distinct fibers. Part (d) of the above theorem shows
that Xχ has a unique minimal element Jχ . Conze–Dixmier [72, Lemme 2]
254 13. Complements

show that the fiber also has a unique maximal element: any ideal Jλ with
λ dominant and χ = χλ . To describe X as a partially ordered set reduces
now to describing the order structure of each fiber. Along the way one also
has to determine the sizes of the fibers.
Borho and Jantzen [42] apply translation functors systematically to re-
duce the problem further to the case when λ is regular and integral. They
show that all fibers of this type are isomorphic as partially ordered sets, while
the singular cases exhibit predictable degeneracies. Understanding a single
case such as λ = 0 is then crucial.

13.3. Structure of a Fiber

In view of the Borho–Jantzen results, we can focus now on the fiber X0 ,
writing Jw in place of Jw·0 . The fiber has at most |W | elements, including the
unique minimal element Jw◦ and the unique maximal element J1 . Beyond
this the structure of X0 is mysterious. But the work of Borho–Jantzen
suggests strongly that it should somehow depend just on W , a situation
reminiscent of the multiplicity problem for Verma modules. Part of the
problem here is to describe the equivalence relation on W which corresponds
to Jw = Jw0 .
For the finer classification of primitive ideals it is useful to consider two
invariants, which can be defined for an arbitrary primitive ideal J in terms
of the ring-theoretic properties of U (g)/J.
(1) In the classical setting, a simple U (g)-module M with dim M < ∞
corresponds to a matrix representation U (g) → Mn (C) having kernel J;
moreover, this map is surjective (Burnside). If dim M = ∞, we still have
an injective map U (g)/J → End M ; but it need not be surjective (see 13.4
below). In general one can assign to a U (g)-module M a measure of “growth”
called the Gelfand–Kirillov dimension (which was referred to earlier in the
discussion of Theorem 9.14). This is denoted Dim M and is 0 precisely when
dim M < ∞. In our situation it takes values in Z+ , because it has a well-
behaved commutative counterpart when one passes from U (g) to S(g) and
replaces J by a graded ideal in the polynomial ring.
(2) By Goldie’s Theorem, the ring U (g)/J embeds into its complete ring
of fractions (Ore) which in turn is isomorphic to the ring of r × r matrices
over a division algebra. Call r the Goldie rank , denoted rk(U (g)/J). Now
Joseph goes on to define Goldie rank polynomials: Fix Jw ∈ X0 and let λ be
a regular integral weight in the w-chamber; then set pw (λ) := rk(U (g)/Jλ ).
The Zariski density in h∗ of the regular elements of Λ then allows one to
extend pw to a polynomial function on h∗ . Its degree is given explicitly by
1
deg pw = |Φ+ | − Dim(U (g)/Jw ).
2
13.4. Kostant’s Problem 255

For example, when w = 1, this is |Φ+ | and the polynomial p1 is just the
classical Weyl dimension polynomial. By combining these and related ideas,
one concludes that Jw = Jw0 if and only if pw = pw0 . Similarly there is a
precise algorithm for the inclusion of one ideal in another.
The various Jw in X0 can be further organized into “clans”. In a clan,
all the Goldie rank polynomials have the same degree; moreover, they form
a basis for a simple CW -module.
Another essential ingredient is supplied by Borho. He associates to a
primitive ideal J a nilpotent orbit in g under the action of the adjoint group
G and proves that Dim(U (g)/J) is just the dimension of this orbit. In par-
ticular, the orbit dimensions are all even and range from 0 to |Φ|, consistent
with the above formula for deg pw .
Kazhdan–Lusztig theory [171] draws these diverse strands together and
relates them to Springer’s Weyl group representations as well. The equiv-
alence classes described above in W are the left cells in W and the CW -
modules are those involved in Springer’s theory, where the correspondence
with nilpotent orbits also plays a natural role. In type A` one gets all irre-
ducible representations of W = S`+1 , which are parametrized by partitions
of ` + 1 (as are the nilpotent orbits, in terms of sizes of Jordan blocks). But
in general the picture is much more subtle.

Example. Borho and Jantzen [42, 4.17] work out a number of examples to
illustrate the ideas in the classification of primitive ideals, including the full
order diagram for a regular integral fiber when g = sl(4, C) or sl(5, C). The
data for the first of these can be summarized as follows. The fiber X0 has
10 elements, corresponding to 10 left cells in W = S4 . The Gelfand–Kirillov
dimensions are 0, 6, 8, 10, 12 (the dimensions of the five nilpotent orbits),
while the Goldie ranks are respectively 6, 3, 2, 1, 0. The corresponding clans
have (respectively) 1, 3, 2, 3, 1 elements, these being the degrees of irreducible
representations of S4 (whose squares sum to 24 = |W |). Here 10 is the sum
of these degrees and in this case counts the number of involutions in S4 .

13.4. Kostant’s Problem

Techniques developed in category O have a bearing on a classical (still un-
solved) problem arising in Kostant’s early work on Harish-Chandra modules.
This concerns a U (g)-module M and the related bimodule L(M, M ). Recall
from 10.9 the construction, which starts with the linear space Hom(M, M )
viewed as a (U, U )-bimodule (where U = U (g)). When g = k is embedded
suitably in g × g, the space of k-finite vectors in Hom(M, M ) is a Harish-
Chandra module denoted L(M, M ). As observed in Remark 10.9, the bi-
module U and its quotient U/ Ann M are themselves locally finite under this
256 13. Complements

action of g. So the natural homomorphism U → End M induces an injection

U/ Ann M → L(M, M ). This raises the natural question:

Problem (Kostant). For which simple U (g)-modules M is the natural in-

jective homomorphism U (g)/ Ann M → L(M, M ) surjective?

While the main interest is focused on simple modules, one can of course
pose the question more generally. For example, it is known for a dominant
weight λ that M (λ) and all its quotients such as L(λ) yield positive answers:
see Jantzen [149, 6.9]. (In case dim L(λ) < ∞, this already follows from the
classical theorem of Burnside as noted in 13.3(1) above.)
The problem is still unsettled even for the simple highest weight modules
L(λ), where both positive and negative answers have been found in special
cases. Joseph [157] studied the problem systematically in terms of Goldie
rank, finding a negative answer for type B2 but a positive answer for some
L(λ). (See [157, 9.5] and [220, 11.5].) Further cases where the answer is
positive also emerged from the work of Gabber–Joseph [106].
In recent years the problem has been revisited by Mazorchuk and his
collaborators. Here is a quick overview. In [214] Kostant’s Problem is
studied from the standpoint of the Arkhipov twisting functors discussed in
Chapter 12. Their close connection with completion functors (an important
ingredient in Joseph’s work) makes twisting functors a natural tool here.
On the one hand, Mazorchuk solves the problem affirmatively in some new
cases, while on the other hand he recovers known results in a more unified
framework. For example, let λ ∈ Λ be dominant and regular. Take a subset
I ⊂ ∆, with corresponding parabolic subgroup WI of W and longest element
wI . For a fixed α ∈ I, set w := sα wI w◦ . Then Kostant’s problem has a
positive answer for L(w · λ).
In the special case g = sl(n, C), where W is the symmetric group
Sn , Mazorchuk–Stroppel [220, §11] approach the problem in terms of the
Kazhdan–Lusztig (one-sided) cells of W and find positive answers in new
as well as known cases. While this encourages optimism about finding only
positive answers in type A, counterexamples soon emerge even in this case
[221]. Definitive results for sl(n, C) when n ≤ 5 are then worked out by
Kåhrström and Mazorchuk [169] for regular integral blocks.

13.5. Kac–Moody Algebras

We turn now to the role played by category O ideas in diverse module
categories related to Lie theory. To avoid excessive and often conflicting
notation, we temporarily suspend in this and the next section our standard
conventions about the symbols g, h, W, . . .
13.5. Kac–Moody Algebras 257

The nearest relatives of a finite dimensional semisimple Lie algebra are

the infinite dimensional Kac–Moody algebras, which arose independently in
the mid-1960s (with different motivations) in the thesis work of Victor Kac
and Robert Moody. The starting point for their construction is the presen-
tation of a finite dimensional algebra by generators and relations perfected
by Serre: see Bourbaki [46, VIII, §4], Carter [60, 7.5], Humphreys [126,
18.3]. The generators are indexed by simple roots and would be denoted
xi , yi , hi in our set-up (but are more usually labelled ei , fi , hi in the Kac–
Moody literature). The relations depend only on the Cartan matrix with
integer entries hαi , αj∨ i, equivalent to the data in the Dynkin diagram.
By relaxing the conditions on matrix entries one gets a generalized Car-
tan matrix A = (Aij ): its diagonal entries are all 2, its off-diagonal entries
are nonpositive integers, with Aij = 0 just when Aji = 0. The analogue of
Serre’s presentation then typically yields an infinite dimensional Lie algebra
which we again call g. Somewhat surprisingly, g retains many features of the
“finite type” g such as a triangular decomposition of the form g = n− ⊕h⊕n,
with h abelian and finite dimensional. There is also an infinite “root sys-
tem” and “Weyl group” W . Here W is a “crystallographic” Coxeter group
in the sense that the orders of products of distinct involutive generators
lie in {2, 3, 4, 6, ∞}. One novelty is that the root spaces gα spanning n, n−
(other than those permuted by W ) need not be one dimensional; write
mα := dim gα .
The closest analogue of a finite dimensional g is the related affine Lie
algebra, which we denote by b g. Here the matrix A builds on the Cartan ma-
trix for g by adding an extra row and column, corresponding to the extended
Dynkin diagram obtained by using as an extra root the negative of the high-
est root. Unlike general Kac–Moody algebras, b g admits an explicit model:
first construct the loop algebra C[t, t−1 ] ⊗ g using the Laurent polynomial
ring and define an obvious Lie bracket, then pass to a one dimensional cen-
tral extension (this is the subtle but essential part). (For technical reasons
one then often extends the algebra by a derivation.) In this situation there
are two kinds of roots: real (conjugate under W to the given “simple” roots)
and imaginary (with root spaces all of dimension `). The latter are nonzero
integral multiples of a single root.
Affine Lie algebras have been well studied and have many links to other
parts of mathematics and physics. Beyond these the best behaved Kac–
Moody algebras are the symmetrizable ones: here A can be made symmetric
by multiplying it by a diagonal matrix with positive integer entries on the
diagonal. This includes both finite and affine types.
There is by now a vast literature devoted to Kac–Moody algebras as
well as their generalizations: vertex operator algebras, Borcherd algebras,
258 13. Complements

etc. Three books which cover the basic material, though from different per-
spectives, are Kac [166], Moody–Pianzola [223] (emphasizing the triangular
decomposition), Carter [60] (integrating the finite and affine types). To con-
vey the flavor of the published research dealing with topics parallel to those
in the original category O, we mention here just a tiny sample of papers
published over three decades: Kac [165], Garland–Lepowsky [109], Kac–
Kazhdan [167], Deodhar–Gabber–Kac [82], Rocha–Wallach [227, 228], Ku-
mar [198], Kashiwara–Tanisaki [170], Tanisaki [249], Fiebig [95, 96].

13.6. Category O for Kac–Moody Algebras

Any Kac–Moody algebra admits a category of “highest weight” modules
analogous to O. As before a weight λ is a linear function on h. But one
has to relax the finite generation assumption, allowing for infinitely many
“composition factors”. Moreover, the center of the universal enveloping
algebra plays no significant organizing role. As a result it is highly nontrivial
to generalize proofs even though a surprising number of theorems for the
finite dimensional case have good analogues here.
One starts with natural analogues of Verma modules, again written
M (λ); these have easily computed formal characters and unique simple quo-
tients L(λ). There is a reasonable analogue here of finite dimensional mod-
ules: a module M in O is called integrable if all root vectors ei , fi act locally
nilpotently on it; then dim Mµ = dim Mwµ for all µ and all w ∈ W . For a
simple module L(λ) this is equivalent to requiring that λ be dominant in-
tegral : all λ(hi ) ∈ Z+ . Now ρ is chosen (non-uniquely) to satisfy ρ(hi ) = 1
for all i. In 1974 Kac [165] was able to prove:
Theorem (Kac). Let g be a symmetrizable Kac–Moody algebra. If λ is
dominant and integral, then L(λ) is integrable, with
X
(−1)`(w) e w(λ + ρ) .


Q ∗ ch L(λ) =
w∈W

Here Y
m X
Q = e(ρ) ∗ 1 − e(−α) α = (−1)`(w) e(wρ).
α>0 w∈W

All of this closely resembles the classical Weyl formula (Theorem 2.4)
except for the root multiplicities mα ≥ 1 appearing in the denominator for-
mula. But in general the sums and products here are all infinite. The proof
by Kac is similar to the BGG proof we followed in the finite dimensional
case, relying essentially on a Casimir operator (a formal infinite sum which
acts on a module just as a finite sum) as a substitute for the full center of
the enveloping algebra.
13.7. Highest Weight Categories 259

While at first the theorem appears to be a purely formal generalization

of classical results, it was motivated in large part by a spectacular inter-
pretation of the above formula for Q. This expresses an infinite product
over positive roots in terms of an infinite sum over the Weyl group. Special
cases of this were discovered earlier in an ad hoc way by Macdonald; this
recaptures in particular the classical Jacobi Triple Product Identity related
to Dedekind’s η-function. The Lie algebra approach reveals a lot of new
“sum equals product” identities, within a conceptually unified framework.
As in the finite dimensional case, the character formula can also be
realized by a BGG-type resolution. Associated with this are Lie algebra
homology computations. (See for example Kumar [198].)
Generalizing the Weyl–Kac character formula when L(λ) is integrable,
one can formulate a precise analogue of the KL Conjecture for any L(λ):
since W is a Coxeter group, all of the combinatorial machinery is already in
place. But the proof in the Kac–Moody case requires substantial new ideas,
supplied by the work of Casian and Kashiwara–Tanisaki [170, 249].
Besides the study of multiplicities in highest weight modules, some of
the literature cited above delves into other more categorical aspects of O
for a Kac–Moody algebra. But nothing is quite so straightforward as in the
case when g is finite dimensional.

13.7. Highest Weight Categories

Beyond the case of Kac–Moody algebras, there are quite a few module cat-
egories which resemble O in some ways but differ significantly in others.
Each has its own extensive literature (and local notation), defying even the
type of brief sketch we gave in the case of Kac–Moody algebras. Here we
just list some of the active areas of study in recent decades, with reference
to books, surveys, and a somewhat arbitary sample of research papers.

• simple weight modules: Mathieu [207]

• Virasoro algebra: Rocha–Wallach [229], Boe–Nakano–Wiesner [38]
• Lie superalgebras: Brundan [49, 50]
• finite W -algebras: Brundan–Goodwin–Kleshchev [53]
• quantum enveloping algebras (and quantized function algebras):
Andersen [4], Jantzen [152], Joseph [162]
• symplectic reflection algebras, including as a special case the ratio-
nal Cherednik algebras: Gordon [114], Ginzburg–Guay–Opdam–
Rouquier [112], Guay [117], Khare [173]
260 13. Complements

• representations in prime characteristic of algebraic groups, their

Lie algebras, finite groups of Lie type: Jantzen [153], Humphreys
[132]
• Schur algebras and q-Schur algebras: Donkin [85]
• BGG reciprocity for perverse sheaves: Mirollo–Vilonen [222]
Is it possible to extract from the diversity here some common core in ax-
iomatic form? This could be modelled for example on the features of a single
block of the original category O. A number of axiomatic approaches have
been proposed: see Cline–Parshall–Scott [63, 64], Irving [138] Futorny–
Mazorchuk [104], Gomez–Mazorchuk [113], Khare [174].
The set-up formulated by Cline–Parshall–Scott is rather general and has
been especially influential. In [63, §3] an abelian category C is dubbed a
highest weight category if it is “locally artinian”, with enough injectives, and
satisfies a number of axioms. One can define the composition factors of an
object A ∈ C to be composition factors of a subobject of finite length, leading
to multiplicities [A : S] which may be infinite. Now the simple objects S(λ)
are assumed to be indexed by an interval-finite partially ordered set (thought
of as “weights”), as are objects A(λ) modelled on dual Verma modules, so
that S(λ) ⊂ A(λ) and all composition factors S(µ) of A(λ)/S(λ) satisfy
µ < λ. Moreover, S(λ) has an injective envelope I(λ) in C which is filtered
by various A(µ) (each of finite multiplicity).
This framework can be seen to encompass many of the above examples
both in characteristic 0 and in characteristic p. Moreover, there turns out
to be a close connection with the quasi-hereditary algebras introduced by
Scott: for a finite dimensional algebra A over a field, its module category is a
highest weight category precisely when A is quasi-hereditary [63, Thm. 3.6].
Irving [138] studies a narrower class of highest weight categories at-
tached to BGG algebras. Here the category is required to have a duality
functor, leading to BGG reciprocity.

13.8. Blocks and Finite Dimensional Algebras

By decomposing O (or Op ) into a direct sum of subcategories, each having
only finitely many simple objects, one is able (in principle) to invoke the
well-developed techniques of the representation theory of finite dimensional
algebras. This is usually not easy to do in practice, but some new approaches
and insights outside the usual framework of Lie theory are suggested. Here
we look briefly at two related topics:
(1) Quiver attached to a block.
(2) Representation type of a block.
13.9. Quiver Attached to a Block 261

The easiest decomposition of O involves the subcategories Oχ attached

to central characters χ (or W -linkage classes of weights). It was shown in
Proposition 3.13 that each Oχ is equivalent to the category of (right) A-
modules for some finite dimensional algebra A; a convenient choice is the
endomorphism algebra of a projective generator.
If the weights involved are required to be integral, the subcategories Oχ
are actually the blocks of O: see 1.13. (As before the reader is cautioned
to be aware of different usages for the term “block” in the literature cited.)
Otherwise Oχ has a refined block decomposition, using the nonempty in-
tersections of a linkage class with cosets of Λr in h∗ . Here the blocks Oλ
are indexed by their unique antidominant weights λ: see 4.9 and Remark
3.5. The blocks in Oχ are all isomorphic as partially ordered sets (even
equivalent as categories, thanks to Mathieu [207, A.3]).
We remark that Soergel’s work, discussed in 13.13 below, shows when
λ 6∈ Λ that Oλ is the subcategory corresponding to a linkage class of W[λ] in
the analogue of category O for a semisimple Lie algebra having Weyl group
W[λ] . This makes it possible to limit attention just to integral weights, which
we now do. The main ideas can be illustrated conveniently for
the principal
block, using the previous convention that Lw = L w · (−2ρ) , etc.

13.9. Quiver Attached to a Block

The use of quivers in Lie theory goes back to early work of I.M. Gelfand and
his collaborators. But the first systematic study of quivers for O seems to
have been carried out by Irving, starting with his unpublished notes [133]
(see [134]). Recent work by Stroppel [246] and Vybornov [257] goes much
further. The quiver viewpoint has the advantage of stripping away all but
the most essential features of the module category in question.
Here is a quick review of the construction. Given a finite dimensional
algebra A, its quiver is a directed graph having vertices in bijection with the
simple modules Si . There are nij arrows i → j if dim ExtA (Sj , Si ) = nij (the
multiplicity of Si as a summand of the top radical layer of the projective
cover Pj of Sj ). Each arrow represents a nonsplit extension having length
2 (equivalently, a nonzero homomorphism Pi → Pj whose image meets the
top radical layer of Pj in a copy of Si ).
Up to Morita equivalence, A may be assumed here to be a basic algebra:
all dim Si = 1; so A ∼
L
= i Pi . Then A is isomorphic to the quotient of its
path algebra by the ideal of relations, where the path algebra has a basis
consisting of all directed paths in the quiver and an obvious product (= 0
when two paths cannot be composed). Thus one wants to describe the
quiver with relations in order to recover the full category of A-modules:
262 13. Complements

these correspond to representations of the quiver, defined by assigning a

vector space to each vertex and a linear map to each arrow.
In the case of O, Theorem 3.2(e) insures that



dim ExtO L(λ), L(µ) = dim ExtO L(µ), L(λ) for all λ, µ.

Rather than use pairs of arrows, we can just connect vertices with a double
arrow when Ext 6= 0. The simplest example is the principal block for sl(2, C)
[246, 5.1.1]. Labelling the vertices (left to right) by the weights 0 and −2,
we get the quiver

• ←→ •

Observe how to identify the five nonisomorphic indecomposable modules in

O0 specified in 3.12 with five (indecomposable) representations of the quiver.
In this case one needs only the single relation

(• ←− •) ◦ (• −→ •) = 0.

But beyond rank one the number of relations required grows very fast.
In general the KL Conjecture implies for the block O0 that if x < w in
W , then dim ExtO (Lx , Lw ) = µ(x, w); this is the lead coefficient of the KL
polynomial Px,w (q) if its degree is as large as possible, or 0 otherwise. (See
Theorem 8.15(c).) But it is not obvious how to describe all the relations
involved in the quiver.
Going beyond Irving’s direct methods, Stroppel [246] takes advantage
of Soergel’s identification of a regular block of O with a category of modules
over the coinvariant algebra of W (discussed in 13.12 below). She formulates
an algorithm for computing the quiver with relations of any such block and
carries out the details for examples in rank 2 as well as for sl(4, C). In
this last case the resulting quiver diagram and list of relations (given in the
Appendix) are impressive; as she observes, there are “a lot of arrows and very
many relations”. Indecomposable modules can then be pictured in detail as
“representations” of the quiver, for example the dominant Verma module in
a regular integral block of sl(4, C). Here some multiple composition factors
occur, leading to submodules which are not sums of Verma submodules and
which can be located precisely using the quiver.
Vybornov [257] refines Stroppel’s quiver algorithm for an arbitrary sim-
ple g, leading to a computer executable algorithm. His approach is en-
tirely different, based on modeling categories of perverse sheaves on various
spaces including the flag variety by “IC-modules” (without use of derived
categories).
13.10. Representation Type of a Block 263

13.10. Representation Type of a Block

In studying the module category of a finite dimensional algebra A, one has
to look at all possible indecomposable modules, including the finitely many
simple modules and their projective covers. There are three possibilities for
the “representation type” of A: finite type, meaning that up to isomorphism
A has only finitely many indecomposables; tame type, meaning that A has
infinitely many nonisomorphic indecomposables, which can nonetheless be
parametrized in some reasonable way; wild type, when neither of these alter-
natives is true. In the literature, there are precise definitions of tame type
(differing somewhat in the details) which we won’t pursue here.
When A is the group algebra of a finite group, much work has been done
to sort out the possibilities here in terms of the structure of the group. In
the case of a finite group of Lie type, the outcome has a somewhat negative
flavor: in only a few small rank cases does one have finite or tame represen-
tation type. (These are summarized in Humphreys [132, 8.9].) Recent work
on blocks of category O or Op leads to a somewhat similar outcome. Here
we have already seen in 3.12 a small rank case of finite representation type:
a regular integral block of g = sl(2, C), which contains just five indecompos-
able modules. Of course, the singular block involving just the weight −ρ is
also of finite type (and is moreover semisimple).
Roughly speaking, the classification of blocks by representation type is
done indirectly. Typically one shows (using quivers) that a known wild type
of module category embeds in the given block, insuring that it too is of wild
type. Then a lot of case-by-case work is needed to sort out the relatively few
remaining cases. We can summarize the end result for O in the case when
g is simple. Consider a block Oλ with λ ∈ Λ antidominant. The isotropy
group Wλ◦ of λ is then the Weyl group of a root subsystem Ψ ⊂ Φ.
Theorem. Fix an antidominant weight λ ∈ Λ, with isotropy group Wλ◦ and
root system Ψ ⊂ Φ. Then the representation type of Oλ depends just on the
pair (Φ, Ψ), as follows.
(a) Finite type: (Φ, Φ), (A1 , ∅), (A2 , A1 ).
(b) Tame type: (A3 , A2 ), (B2 , A1 ).
(c) Wild type: all others.

This theorem is the main result of Futorny–Nakano–Pollack [105], who

rely mainly on techniques such as quivers from the representation theory
of finite dimensional algebras together with a comparison of blocks for Levi
subalgebras of g. The theorem is recovered a bit later in a completely
different way by Brüstle–König–Mazorchuk [54], building in part on a paper
by the last two authors [194]. They exploit heavily the work of Soergel on
264 13. Complements

projectives in O and the coinvariant algebra of W (13.12 below) along with

Schubert calculus for the cohomology algebra of the related flag variety.
The story for blocks in parabolic categories Op is considerably more
complicated and less complete, the blocks themselves not yet being fully
determined. Boe–Nakano [37] have worked out the representation type of
certain blocks, while further cases have been treated by Boe’s student K.
Platt. As above, most blocks are wild. When the weights involved are
regular (and integral), the blocks of each representation type are known.
Very few of them are of finite type: for the root systems of types A` , B` , C` ,
the subset I ⊂ ∆ must be respectively of type A`−1 , B`−1 , C`−1 ; the only
other case is G2 with I of type A1 . Among the blocks involving the linkage
class of a singular weight λ, detailed results are obtained for those which
depend on two disjoint subsets of ∆: one is the given set I with p = pI
while the other defines the subsystem consisting of all roots α for which
hλ + ρ, α∨ i = 0.

13.11. Soergel’s Functor V

Now we change gears once again and introduce some influential work of
Soergel [237] and Beilinson–Ginzburg–Soergel [22], which was stimulated
by earlier unpublished conjectures of Bernstein and these authors. They
look at category O from a more advanced homological viewpoint, making
further connections to perverse sheaves and the geometry of flag varieties.
One striking theme that emerges is the role of graded (not just filtered)
categories. Eventually the geometric side of the subject is captured under
the rubric “geometric representation theory”, as in Gaitsgory’s lecture notes
[108].
The first goal is to formulate precisely some of the main results in [237,
§2], where the methods are still algebraic; but we follow our own notational
conventions when conflicts occur. Here it is important to allow arbitrary
weights λ ∈ h∗ and blocks Oλ , with the corresponding reflection subgroups
W[λ] of W . In particular, let λ be antidominant and consider the projective
module P = P (λ). When λ is integral, we saw in 4.10 that P is self-
dual, hence also injective, while each Verma module in its block Oλ occurs
precisely once as a quotient in any standard filtration. With more effort, the
same result was later proved in general using translation functors (7.16).
Write E = EndO P and denote by M the category of finite dimensional
E-modules. For any M ∈ O, observe that the vector space HomO (P, M ) has
a natural structure of E-module: simply compose an endomorphism P → P
with a given homomorphism P → M . Now Soergel defines a functor V = VP
from O to M, which sends M 7→ HomO (P, M ) and behaves in the obvious
13.12. Coinvariant Algebra of W 265

way on morphisms. For example, when M = P we just get VP = E. With

this set-up, a fundamental structure theorem states:
Struktursatz. Let λ ∈ h∗ be antidominant and write P = P (λ), E =
EndO P . If Q is any projective module in the block Oλ and M ∈ O, then
the functor V defined above induces a vector space isomorphism
∼
HomO (M, Q) → HomE (VM, VQ).
Remarks. (1) For an easy example, take M = Q = P : then the the-
orem just expresses the natural vector space isomorphism from E onto
HomE (E, E).
(2) More recently Beilinson–Bezrukavnikov–Mirković [19] have recovered
the Struktursatz in a geometric setting.

To make the functor V effective, one of course has to understand the

structure of the finite dimensional algebra E. Indeed, the proof of the Struk-
tursatz given in [237, 2.3] relies on a concrete description of E. This turns
out to have a natural connection with W . The computation of dim E is
already suggestive. According to Theorem 3.9(c), dim E = [P : L(λ)]. This
equals the index in W[λ] of the isotropy group Wλ◦ thanks to the way P is
filtered by Verma modules in Oλ and the fact that L(λ) occurs just once as
a composition factor of each.

13.12. Coinvariant Algebra of W

In general it is not obvious how to determine the endomorphism algebra of an
indecomposable projective in O. There is one obvious step in this direction:
left multiplication by an element of Z(g) is always a U (g)-endomorphism,
so there is a natural map from Z(g) to the endomorphism algebra. When
E = End P with P = P (λ) for an antidominant λ ∈ h∗ , it follows from
Soergel’s description of E in the theorem below that this natural map is
surjective. In particular, E is commutative. The key result is that E depends
only on W[λ] and Wλ◦ , and is in fact isomorphic to the coinvariant algebra
C of W[λ] if λ is regular.
We have to recall how C is defined, referring for details to Bourbaki [45,
V, §5], especially 5.2, Thm. 1 and Thm. 2 (see also Humphreys [130, 3.6]).
The definition makes sense for an arbitrary finite reflection group, which we
just call W for convenience.
Viewing W as a subgroup of a suitable GL(V ) with dim V = n, its action
on W extends naturally to the (graded) symmetric algebra S(V ). By Cheval-
ley’s theorem, the subalgebra S(V )W of invariants is itself a polynomial
algebra in n indeterminates and is generated by homogeneous elements of
S(V ); their degrees are uniquely determined and have product |W |. (When
266 13. Complements

W is a symmetric group, this is just the theorem on elementary symmetric

polynomials.) In turn, the ideal I generated by homogeneous invariants of
positive degree is homogeneous and the quotient S(V )/I is a graded finite
dimensional algebra called the coinvariant algebra of W . Denote it by C. As
a W -module it affords the regular representation of W ; thus dim C = |W |.
It is well known from the classical work of Borel that the coinvariant
algebra of the Weyl group W of g is isomorphic as a graded algebra to the
cohomology algebra of the flag variety G/B (as in 8.5). This fact is used
in the proof of the theorem below; it also suggests links between O and the
geometry of Schubert varieties implicated in the KL Conjecture.
Endomorphismensatz. Let λ and P = P (λ) be as in the Struktursatz,
and recall that Wλ◦ denotes the isotropy group of λ in W[λ] . Let C be the
◦
coinvariant algebra of W[λ] as above, with subalgebra C Wλ of invariants un-
◦
der Wλ◦ . Then there is a surjective homomorphism Z(g) → C Wλ , having as
kernel AnnU (g) P . As a result,
◦
EndO P ∼
= C Wλ ,
which is just C if λ is regular.

For the proof of the theorem in [237, 2.2], Soergel relies heavily on
translation functors (especially wall-crossing functors), together with a close
study of the interaction between C and Z(g).
Once the Endomorphismensatz is in place, the proof of the Struktursatz
is relatively short but requires a number of techniques. In one step, the
copy of the dominant Verma module embedded in P is characterized as the
subspace of P annihilated by the unique maximal ideal of C. In another
step, the modules in O killed by V are characterized as those having Gelfand–
Kirillov dimension < dim n; then projective functors come into play.

13.13. Application: Category Equivalence

The results of Soergel discussed above make more precise the earlier ob-
servation that key features of O depend solely on the Weyl group. For
example, the multiplicities of composition factors of Verma modules with
integral weights are given by the values at 1 of KL polynomials attached to
W and therefore coincide for Lie algebras of types B` and C` .
Here we get insight into the way in which a block of Oλ associated with a
nonintegral weight λ, together with the subgroup W[λ] of W , can be related
to an integral block of another Lie algebra having W[λ] as its Weyl group.
(As mentioned in earlier chapters, this reduces many problems such as the
proof of the KL Conjecture to the integral case.) Soergel’s theorem [237,
2.5, Thm. 11] can be stated formally as follows.
13.14. Endomorphisms and Socles of Projectives 267

Theorem. Let g and g0 be semisimple Lie algebras, with respective Weyl

groups W and W 0 . Fix weights λ for g and λ0 for g0 , with corresponding
blocks Oλ and Oλ0 0 and reflection subgroups W[λ] and W[λ 0 . If there is an
0]

isomorphism between these Coxeter groups which takes the isotropy group of
λ to the isotropy group of λ0 , then Oλ is equivalent to Oλ0 0 , with L(λ) sent
to L(λ0 ) and M (λ) sent to M (λ0 ).

For the application just indicated, take λ to be nonintegral and then

use the fact that W[λ] is the Weyl group of the root system Φ[λ] (Theorem
3.4) to find a semisimple Lie algebra g0 whose Weyl group is isomorphic to
W[λ] . Then λ0 can be taken to be any integral weight whose isotropy group
agrees with that of λ. (Note a subtle point in this category equivalence:
since there is typically a different weight ρ for each Lie algebra, one cannot
directly compare a linkage class of weights for g with one for g0 . So the proof
of the theorem is necessarily somewhat indirect.)
To prove the theorem, a preliminary reduction is made (using translation
functors) to the case of regular weights. Now the two blocks in question are
determined by the endomorphism algebras of projective generators, which
under the functor V can be compared in terms of C-modules for the shared
reflection group. Here one builds up the images under V of indecomposable
projectives inductively in the C-module setting using the counterpart of
wall-crossing functors.
The simple modules have categorical meaning and therefore correspond
naturally. This also happens for the Verma modules: first characterize them
categorically as projective objects in appropriate subcategories (this was
remarked after the proof of Proposition 3.8).

13.14. Endomorphisms and Socles of Projectives

Building on Soergel’s work, Stroppel [246] studies the endomorphism alge-
bras of arbitrary projective modules in an integral block of O (along with
related quivers, discussed in 13.9 above). For convenience we state the
main result [246, Thm. 7.1] just for the block O0 , with standard modules
Lw , Mw , Pw labelled by the weight w · (−2ρ).
Theorem. If x ∈ W , the following three statements are equivalent:
(a) EndO Px is commutative.
(b) The filtration multiplicity (Px : Mw◦ ) = 1.
(c) There exists an epimorphism Z(g) → EndO Px .

Although this is stated without explicit reference to Soergel’s functor

V, the proof uses his results in an essential way; indeed, a different proof
268 13. Complements

of the implication (a) ⇒ (b) was given by Soergel. Stroppel observes that
(b) implies the more precise conclusion that the endomorphism ring is a
quotient of the coinvariant algebra C.
By BGG Reciprocity, (b) is the same as [Mw◦ : Lx ] = 1. This can be
checked using Jantzen’s multiplicity one criterion (8.7), which predates the
KL Conjecture.
Soergel’s work already covers the extreme case when x = 1; here Px is
the self-dual projective module in the block. At the other extreme, Pw◦ =
Mw◦ has the trivial endomorphism ring C (4.2). In case g = sl(3, C), the
multiplicity in (b) is always 1. Here Stroppel exhibits concretely the way
each endomorphism ring can be written as a quotient of C.
Under the hypothesis of the theorem, it is also shown in [246, Thm. 8.1]
that Soc Px is the direct sum of (Px : Mw◦ ) copies of the simple Verma
module L1 . This number is one just when the endomorphism ring is com-
mutative, by the theorem. The result is of course consistent with the two
extreme cases. Again its statement does not involve Soergel’s functor V. The
proof uses the fact that V kills all simple modules not of maximal Gelfand–
Kirillov dimension: this leaves only copies of the simple Verma module in
the socle.
Even when the conditions of the theorem are satisfied, it is not obvious in
general how to describe the structure of EndO Px precisely. However, in all
cases the dimension of the endomorphism ring is computable (in principle)
from Kazhdan–Lusztig theory: thanks to Theorem 3.9(c), it is equal to
[Px : Lx ]. After filtering Px by Verma modules Mw with x ≤ w and using
BGG Reciprocity, version 8.4(2) of the KL Conjecture yields
X X
dim EndO Px = [Mw : Lx ] = Pw◦ w,w◦ x (1).
x≤w x≤w

This raises the question of whether the structure of each EndO Px depends
only on W and its Bruhat ordering.

13.15. Koszul Duality

By exploiting further the geometric methods used in the proof of the KL
Conjecture (especially perverse sheaves), Soergel [237] and then Beilinson–
Ginzburg–Soergel [22] refine significantly the algebraic results described in
13.11 and 13.12. In particular, this work brings graded modules into the
picture, by comparing a graded Ext algebra with a not obviously graded End
algebra. Important precursors include the preprints of Beilinson–Ginzburg
[20] and Soergel [236].
In [237] the coinvariant algebra C plays a major role in this transition to
geometric thinking. In one of his main theorems (Zerlegungssatz), Soergel
13.15. Koszul Duality 269

essentially categorifies the Hecke algebra H of W (8.2) using the category

of finitely generated graded C-modules. This leads to certain graded C-
modules (in bijection with the Kazhdan–Lusztig self-dual basis of H) which
are graded versions of the VPw (w ∈ W ). The price paid for the use of deeper
geometric methods here is the limitation to Weyl groups, even though the
C-module framework makes sense for arbitrary finite reflection groups.
A high point in [22] is a surprising Koszul duality theorem, which ex-
presses a hidden “parabolic–singular
L duality”. In general, a ring A is called
a Koszul ring if A = n≥0 An is a graded ring, where A0 is a semisimple
ring and the A-module A/A+ ∼
L
= A0 (with A+ := n>0 An ) has a projective
resolution by graded modules Pn with Pn generated by its nth graded com-
ponent. Under a mild finiteness condition, there is a well-defined Koszul
dual ring A! : the (opposite ring of the) graded ring Ext•A (A0 , A0 ). Then
(A! )! ∼
= A. For example, the symmetric algebra of a finite dimensional vec-
tor space and the exterior algebra of its dual space are Koszul algebras and
in fact Koszul duals.
These ideas come into play in O when it is viewed as the module category
over the endomorphism ring of a projective generator. (Here the duality
functor can be used to pass to the opposite algebra.) In what follows all
weights are assumed to be integral but possibly singular. Each antidominant
weight λ ∈ Λ determines a block Oλ as well as a subset I = Iλ of ∆:

Iλ := {α ∈ ∆ | hλ + ρ, α∨ i = 0}.

As before we denote by WI the resulting subgroup of W and by W I the set

of minimal length right coset representatives in WI \W . Here WI is just the
isotropy group Wλ◦ ; thus W I parametrizes the simple modules in the block
Oλ to which L(λ) belongs. Now set
M M
Lλ := L(w · λ) and P λ := P (w · λ).
w∈W I w∈W I

Here P λ is a projective generator for the block.

On the other hand, I determines a parabolic subalgebra p = pI and the
regular block O0p of Op containing the trivial module. The simple modules
L(x · 0) lying in this block are parametrized by left coset representatives in
W/WI , which can be chosen to be those of maximal length. Here we denote
by Lp the direct sum of simples in the block and by P p the direct sum of
their projective covers in Op .
In these two blocks, the respective simple modules L(w·λ) and L(x·0) are
in natural bijection, if we send w 7→ w◦ x−1 . Although the simple modules
in the two situations are otherwise unrelated, a duality emerges (somewhat
miraculously) in [22]:
270 13. Complements

Koszul Duality. With the above notation, there are algebra isomorphisms
EndO P λ ∼ = Ext• p (Lp , Lp ),
O
EndO P p ∼
= Ext•O (Lλ , Lλ ).
The two Ext algebras are Koszul algebras and are in fact Koszul duals.

When λ is regular (essentially the case of O0 ) this restates the earlier

result of Soergel [237, 3.5]; here the Ext algebra is self-dual in the Koszul
sense. In general the isomorphisms in the theorem reveal an underlying
graded structure in each of the End algebras, which does not appear to be
accessible by algebraic means alone. In particular, category O is seen to be
a “graded semisimple” category. All of this lies fairly deep: for example, the
fact that the algebra defining O0 is Koszul and self-dual is actually enough
to imply the KL Conjecture for O0 , though at present its proof depends on
the truth of that conjecture.
Remark. Further work has brought out more strongly the ubiquity of
Koszul duality as a theme in the study of O and related geometry. For
example, Backelin [11] applies the methods of [22] to the parabolic cate-
gory Op . He shows that the algebra which defines an integral (but possibly
singular) block of Op is also Koszul and works out the Koszul dual. In an-
other direction, Ryom–Hansen [230] shows that the category equivalence of
[22] interchanges the derived functors of translation functors for the singular
category and Zuckerman functors for the parabolic category. This result is
recovered by Mazorchuk, Ovsienko, and Stroppel [216], where the theme
of Koszul duality for pairs of functors is carried much further. In [263]
Mazorchuk extends and reformulates more algebraically the treatment of
Koszul dual functors in [11].
Bibliography

1. N. Abe, On the existence of homomorphisms between principal series of complex

semisimple Lie groups, arXiv:0712.2122 [math.RT]
2. K. Akin, On complexes relating the Jacobi–Trudy identity with the Bernstein–
Gelfand–Gelfand resolution, J. Algebra 117 (1988), 494–503.
3. H.H. Andersen, Schubert varieties and Demazure’s character formula, Invent. Math.
79 (1985), 611–618.
4. , Twisted Verma modules and their quantized analogues, pp. 1–10, Combina-
torial and Geometric Representation Theory (Seoul, 2001), Contemp. Math., 325,
Amer. Math. Soc., Providence, RI, 2003.
5. H.H. Andersen and N. Lauritzen, Twisted Verma modules, pp. 1–26, Studies in Mem-
ory of Issai Schur, Progr. Math., 210, Birkhäuser, Boston, 2003.
6. H.H. Andersen and J. Paradowski, Fusion categories arising from semisimple Lie
algebras, Comm. Math. Phys. 169 (1995), 563–588.
7. H.H. Andersen and C. Stroppel, Twisting functors on O0 , Represent. Theory 7
(2003), 681–699.
8. S. Arkhipov, A new construction of the semi-infinite BGG resolution, unpublished
preprint, arXiv:q-alg/9605043
9. , Algebraic construction of contragredient quasi-Verma modules in positive
characteristic, pp. 27–68, Representation Theory of Algebraic Groups and Quantum
Groups, Adv. Stud. Pure Math., 40, Math. Soc. Japan, Tokyo, 2004.
10. E. Backelin, Representation of the category O in Whittaker categories, Internat.
Math. Res. Notices 1997, no. 4, 153–172.
11. , Koszul duality for parabolic and singular category O, Represent. Theory 3
(1999), 139–152.
12. , The Hom-spaces between projective functors, Represent. Theory 5 (2001),
267–283.
13. D. Barbasch, Filtrations on Verma modules, Ann. Sci. École Norm. Sup. (4) 16
(1983), 489–494.
14. H. Bass, Algebraic K-Theory, W.A. Benjamin, New York, 1968.

271
272 Bibliography

15. A. Beilinson, Localization of representations of reductive Lie algebras, pp. 699–710,

Proc. Intern. Congr. Math. (Warsaw, 1983), PWN, Warwaw, 1984.
16. A. Beilinson and J. Bernstein, Localisation de g-modules, C.R. Math. Acad. Sci. Paris
292 (1981), 15–18.
17. A. Beilinson, J. Bernstein, and P. Deligne, Faisceaux pervers, Analyse et topologie sur
les espaces singuliers, I (Luminy, 1981), 5–171, Astérisque, 100, Soc. Math. France,
Paris, 1982.
18. , A proof of Jantzen conjectures, I.M. Gelfand Seminar, 1–50, Adv. Soviet
Math., 16, Part 1, Amer. Math. Soc., Providence, RI, 1993.
19. A. Beilinson, R. Bezrukavnikov, and I. Mirković, Tilting exercises, Mosc. Math. J. 4
(2004), 547–557.
20. A. Beilinson and V. Ginzburg, Mixed categories, Ext-duality and representations
(results and conjectures), preprint, 1986.
21. , Wall-crossing functors and D-modules, Represent. Theory 3 (1999), 1–31.
22. A. Beilinson, V. Ginzburg, and W. Soergel, Koszul duality patterns in representation
theory, J. Amer. Math. Soc. 9 (1996), 473–527.
23. J. Bernstein, I. Frenkel, and M. Khovanov, A categorification of the Temperley–Lieb
algebra and Schur quotients of U (sl2 ) via projective and Zuckerman functors, Selecta
Math. (N.S.) 5 (1999), 199–241.
24. J. Bernstein and S. Gelfand, Tensor products of finite and infinite dimensional rep-
resentations of semisimple Lie algebras, Compositio Math. 41 (1980), 245–285.
25. I.N. Bernstein, I.M. Gelfand, and S.I. Gelfand, Structure of representations generated
by highest weights, Funktsional. Anal. i Prilozhen. 5 (1971), no. 1. 1–9; English transl.,
Funct. Anal. Appl. 5 (1971), 1–8.
26. , Differential operators on the base affine space and a study of g-modules,
pp. 21–64, Lie Groups and their Representations (Proc. Summer School, Bolyai János
Math. Soc., Budapest, 1971), Halsted, New York, 1975.
27. , On a category of g-modules, Funktsional. Anal. i Prilozhen. 10 (1976), no. 2,
1–8; English transl., Funct. Anal. Appl. 10 (1976), 87–92.
28. R. Biagioli, Closed product formulas for extensions of generalized Verma modules,
Trans. Amer. Math. Soc. 356 (2004), 159–184.
29. A. Björner and F. Brenti, Combinatorics of Coxeter Groups, Springer-Verlag, New
York, 2005.
30. B.D. Boe, Homomorphisms between generalized Verma modules, Trans. Amer. Math.
Soc. 288 (1985), 791–799.
31. , A counterexample to the Gabber–Joseph conjecture, pp. 1–3, Kazhdan–
Lusztig Theory and Related Topics (Chicago, IL, 1989), Contemp. Math., 139, Amer.
Math. Soc., Providence, RI, 1992.
32. B.D. Boe and D.H. Collingwood, A comparison theory for the structure of induced
representations, J. Algebra 94 (1985), 511–545.
33. , A comparison theory for the structure of induced representations II, Math.
Z. 190 (1985), 1–11.
34. , A multiplicity one theorem for holomorphically induced representations,
Math. Z. 192 (1986), 265–282.
35. , Multiplicity free categories of highest weight representations, Comm. Algebra
18 (1990), 947–1032; part II, 1033–1070.
Bibliography 273

36. B.D. Boe and M. Hunziker, Kostant modules in blocks of category OS ,

arXiv:math.RT/0604336
37. B.D. Boe and D.K. Nakano, Representation type of the blocks of category OS , Adv.
Math. 196 (2005), 193–256.
38. B.D. Boe, D.K. Nakano, and Emilie Wiesner, Category O for the Virasoro algebra:
cohomology and Koszulity, Pacific J. Math. 234 (2007), 1–22.
39. A. Borel and N. Wallach, Continuous Cohomology, Discrete Subgroups, and Repre-
sentations of Reductive Groups, 2nd ed., Math. Surveys Monogr., 67, Amer. Math.
Soc., Providence, RI, 2000.
40. W. Borho, A survey on enveloping algebras of semisimple Lie algebras. I, pp. 19–50,
Lie algebras and related topics (Windsor, Ont., 1984), CMS Conf. Proc., 5, Amer.
Math. Soc., Providence, RI, 1986.
41. , Nilpotent orbits, primitive ideals, and characteristic classes (a survey),
pp. 350–359, Proc. Intern. Congr. Math. (Berkeley, Calif., 1986), Amer. Math. Soc.,
Providence, RI, 1987.
42. W. Borho and J.C. Jantzen, Über primitive Ideale in der Einhüllenden einer halb-
einfachen Lie-Algebra, Invent. Math. 39 (1977), 1–53.
43. R. Bott, Homogeneous vector bundles, Ann. of Math. 66 (1957), 203–248.
44. A. Bouaziz, Sur les représentations des algèbres de Lie semi-simples construites par
T. Enright, pp. 57–68, Noncommutative harmonic analysis and Lie groups (Marseille,
1980), Lecture Notes in Math., 880, Springer-Verlag, Berlin, 1981.
45. N. Bourbaki, Groupes et algèbres de Lie, Chapters 4–6, Hermann, Paris, 1968; 2nd
ed., Masson, Paris, 1981; English transl., Springer-Verlag, Berlin, 2002.
46. , Groupes et algèbres de Lie, Chapters 7–8, Hermann, Paris, 1975; English
transl. (with Chapter 9), Springer-Verlag, Berlin, 2005.
47. F. Brenti, Kazhdan–Lusztig polynomials: history, problems, and combinatorial in-
variance, Sém. Lothar. Combin., 49 (2002/04), Art. B49b, 30 pp.
48. K.S. Brown, Extensions of “thickened” modules of the Virasoro algebra, J. Algebra
269 (2003), 160–188.
49. J. Brundan, Kazhdan–Lusztig polynomials and character formulae for the Lie super-
algebra gl(m | n), J. Amer. Math. Soc. 16 (2003), 185–231.
50. , Tilting modules for Lie superalgebras, Comm. Algebra 32 (2004), 2251–2268.
51. , Centers of degenerate cyclotomic Hecke algebras and parabolic category O,
arXiv:math.RT/0607717
52. , Symmetric functions, parabolic category O and the Springer fiber,
arXiv:math/0608235 [math.RT], to appear in Duke Math. J.
53. J. Brundan, S.M. Goodwin, and A. Kleshchev, Highest weight theory for finite W -
algebras, arXiv:0801.1337 [math.RT]
54. Th. Brüstle, S. König, and V. Mazorchuk, The coinvariant algebra and representation
types of blocks of category O, Bull. London Math. Soc. 33 (2001), 669–681.
55. J.L. Brylinski and M. Kashiwara, Kazhdan–Lusztig conjecture and holonomic sys-
tems, Invent. Math 64 (1981), 387–410.
56. K. Carlin, Extensions of Verma modules, Trans. Amer. Math. Soc. 294 (1986), 29–43.
57. , Completion and translation in O, Comm. Algebra 16 (1988), 1921–1932.
58. , Local systems of Shapovalov elements, Comm. Algebra 23 (1995), 3039–3049.
274 Bibliography

59. H. Cartan and S. Eilenberg, Homological Algebra, Princeton Univ. Press, Princeton,
1956.
60. R.W. Carter, Lie Algebras of Finite and Affine Type, Cambridge Univ. Press, Cam-
bridge, 2005.
61. L. Casian and D.H. Collingwood, The Kazhdan–Lusztig conjecture for generalized
Verma modules, Math. Z. 195 (1987), 581–600.
62. , Weight filtrations for induced representations of real reductive Lie groups,
Adv. in Math. 73 (1989), 79–146.
63. E. Cline, B. Parshall, and L. Scott, Finite dimensional algebras and highest weight
modules, J. Reine Angew. Math. 391 (1988), 85–99.
64. , Abstract Kazhdan–Lusztig theories, Tohoku Math. J. 45 (1993), 511–534.
65. D.H. Collingwood, Category O0 , n-homology and the reducibility of generalized prin-
cipal series representations, Duke Math. J. 50 (1983), 1201–1224.
66. , Representations of rank one Lie groups, Research Notes in Mathematics,
137, Pitman (Advanced Publishing Program), Boston, MA, 1985.
67. , The n-homology of Harish-Chandra modules: generalizing a theorem of
Kostant, Math. Ann. 272 (1985), 161–187.
68. , Jacquet modules for semisimple Lie groups having Verma module filtrations,
J. Algebra 136 (1991), 353–375.
69. D.H. Collingwood and R.S. Irving, A decomposition theorem for certain self-dual
modules in the category O, Duke Math. J. 58 (1989), 89–102.
70. D.H. Collingwood, R.S. Irving, and B. Shelton, Filtrations on generalized Verma
modules for Hermitian symmetric pairs, J. Reine Angew. Math. 383 (1988), 54–86.
71. D.H. Collingwood and B. Shelton, A duality theorem for extensions of induced highest
weight modules, Pacific J. Math. 146 (1990), 227–237.
72. N. Conze and J. Dixmier, Idéaux primitifs de l’algèbre enveloppante d’une algèbre de
Lie semi-simple, Bull. Sci. Math. 96 (1972), 339–351.
73. N. Conze–Berline and M. Duflo, Sur les représentations induites des groupes semi-
simples complexes, Compositio Math. 34 (1977), 307–336.
74. C.W. Curtis and I. Reiner, Methods of Representation Theory, I, Wiley, New York,
1981.
75. P. Delorme, Extensions dans la categorie O de Bernstein–Gelfand–Gelfand. Applica-
tions, preprint, 1978.
76. , Extensions in the Bernstein–Gelfand–Gelfand category O. Applications,
Funktsional. Anal. i Prilozhen 14 (1980), no 3, 77–78; English transl., Funct. Anal.
Appl. 14 (1980), 228–229.
77. V.V. Deodhar, On a construction of representations and a problem of Enright, Invent.
Math. 57 (1980), 101–118.
78. , On the Kazhdan–Lusztig conjectures, Indag. Math. 44 (1982), 1–17.
79. , On some geometric aspects of Bruhat orderings, II. The parabolic analogue
of Kazhdan–Lusztig polynomials, J. Algebra 111 (1987), 483–506.
80. , Duality in parabolic set up for questions in Kazhdan–Lusztig theory, J. Al-
gebra 142 (1991), 201–209.
81. , ed., Kazhdan–Lusztig Theory and Related Topics (Chicago, IL, 1989), Con-
temp. Math., 139, Amer. Math. Soc., Providence, RI, 1992.
Bibliography 275

82. V.V. Deodhar, O. Gabber, and V. Kac, Structure of some categories of representa-
tions of infinite-dimensional Lie algebra, Adv. in Math. 45 (1982), 92–116.
83. V.V. Deodhar and J. Lepowsky, On multiplicity in the Jordan–Hölder series of Verma
modules, J. Algebra 49 (1977), 512–524.
84. J. Dixmier, Algèbres enveloppantes, Gauthier–Villars, Paris, 1974; reprint of English
translation, Enveloping Algebras, Amer. Math. Soc., Providence, RI, 1996.
85. S. Donkin, The q-Schur Algebra, London Math. Soc. Lecture Note Series, 253, Cam-
bridge Univ. Press, Cambridge, 1998.
86. Yu.A. Drozd and V. Mazorchuk, Representation type of ∞ 1
λ Hµ , Q.J. Math. 57 (2006),
319–338.
87. M. Duflo, Construction of primitive ideals in an enveloping algebra, pp. 77–93, Lie
Groups and their Representations (Proc. Summer School, Bolyai János Math. Soc.,
Budapest, 1971), Halsted, New York, 1975.
88. , Représentations irréductibles des groupes semi-simples complexes, pp. 26–88,
Analyse harmonique sur les groupes de Lie, Lecture Notes in Math., 497, Springer-
Verlag, Berlin, 1975.
89. , Sur las classification des idéaux primitifs dans l’algèbre enveloppante d’une
algèbre de Lie semi-simple, Ann. of Math. 105 (1977), 107–120.
90. T.J. Enright, On the fundamental series of a real semisimple Lie algebra: their irre-
ducibility, resolutions and multiplicity formulae, Ann. of Math. 110 (1979), 1–82.
91. , Lectures on representations of complex semisimple Lie groups, Tata Inst.
Lectures on Mathematics and Physics, 66, Springer-Verlag, Berlin, 1981.
92. T.J. Enright and B. Shelton, Decompositions in categories of highest weight modules,
J. Algebra 100 (1986), 380–402.
93. , Categories of highest weight modules: Applications to Hermitian symmetric
spaces, Mem. Amer. Math. Soc. 67 (1987), no. 367.
94. T.J. Enright and N.R. Wallach, Notes on homological algebra and representations of
Lie algebras, Duke Math. J. 47 (1980), 1–15.
95. P. Fiebig, Centers and translation functors for the category O over Kac–Moody al-
gebras, Math. Z. 243 (2003), 689–717.
96. , The combinatorics of category O over symmetrizable Kac–Moody algebras,
Transform. Groups 11 (2006), 29–49.
97. J. Franklin, Homomorphisms between Verma modules in characteristic p, J. Algebra
112 (1988), 58–85.
98. W. Fulton and J. Harris, Representation Theory, Springer-Verlag, New York, 1991.
99. V. Futorny, S. König, and V. Mazorchuk, S-subcategories in O, Manuscripta Math.
102 (2000), 487–503.
100. , A combinatorial description of blocks in O(P, Λ) associated with sl(2)-
induction, J. Algebra 231 (2000), 86–103.
101. , Categories of induced modules for Lie al;gebras with triangular decomposi-
tion, Forum Math. 13 (2001), 641–661.
102. V. Futorny and V. Mazorchuk, Structure of α-stratified modules for finite-dimensional
Lie algebras. I, J. Algebra 183 (1996), 456–482.
103. , BGG-resolution for α-stratified modules over simply-laced finite-dimensional
Lie algebra, J. Math. Kyoto Univ. 38 (1998), 229–240.
104. , Highest weight categories of Lie algebra modules, J. Pure Appl. Math. 138
(1999), 107–118.
276 Bibliography

105. V. Futorny, D.K. Nakano, and R.D. Pollack, Representation type of the blocks of
category O, Q.J. Math. 52 (2001), 285–305.
106. O. Gabber and A. Joseph, On the Bernstein–Gelfand–Gelfand resolution and the
Duflo sum formula, Compositio Math. 43 (1981), 107–131.
107. , Towards the Kazhdan–Lusztig conjecture, Ann. Sci. École Norm. Sup. (4)
14 (1981), 261–302.
108. D. Gaitsgory, Geometric representation theory, lecture notes, Harvard Univ., 2005.
109. H. Garland and J. Lepowsky, Lie algebra homology and the Macdonald–Kac formulas,
Invent. Math. 34 (1976), 37–76.
110. S. Gelfand and R. MacPherson, Verma modules and Schubert cells: a dictionary,
pp. 1–50, P. Dubreil and M.-P. Malliavin Algebra Seminar (Paris, 1981), Lecture
Notes in Math., 924, Springer-Verlag, Berlin, 1982.
111. S.I. Gelfand and Yu.I. Manin, Methods of Homological Algebra, 2nd ed., Springer-
Verlag, Berlin, 2003.
112. V. Ginzburg, N. Guay, E. Opdam, and R. Rouquier, On the category O for rational
Cherednik algebras, Invent. Math. 154 (2003), 617–651.
113. X. Gomez and V. Mazorchuk, On an analogue of BGG-reciprocity, Comm. Algebra
29 (2001), 5329–5334.
114. I.G. Gordon, Symplectic reflection algebras, arXiv:0712.1568 [math.RT]
115. M. Goresky, Tables of Kazhdan–Lusztig polynomials, currently available online at
http://www.math.ias.edu/˜goresky/tables.html
116. W.A. de Graaf, Constructing homomorphisms between Verma modules, J. Lie Theory
15 (2005), 415–428.
117. N. Guay, Projective modules in the category O for the Cherednik algebra, J. Pure
Appl. Algebra 182 (2003), 209–221.
118. K. Günzl, The fine structure of translation functors, Represent. Theory 3 (1999),
223–249.
119. A. Gyoja, Further generalization of generalized Verma modules, Publ. Res. Inst.
Math. Sci. 29 (1993), 349–395.
120. , A remark on homomorphisms between generalized Verma modules, J. Math.
Kyoto Univ. 34 (1994), 695–697.
121. , A duality theorem for homomorphisms between generalized Verma modules,
J. Math. Kyoto Univ. 40 (2000), 437–450.
122. Harish-Chandra, On some applications of the universal enveloping algebra of a
semisimple Lie algebra, Trans. Amer. Math. Soc. 70 (1951), 28–96.
123. P.J. Hilton and U. Stammbach, A Course in Homological Algebra, Springer-Verlag,
New York, 1971.
124. A. van den Hombergh, Note on a paper by Bernstein, Gelfand and Gelfand on Verma
modules, Indag. Math. 36 (1974), 352–356.
125. R. Hotta, K. Takeuchi, and T. Tanisaki, D-Modules, Perverse Sheaves, and Repre-
sentation Theory, Progr. Math., 236, Birkhäuser, Boston, 2008.
126. J.E. Humphreys, Introduction to Lie Algebras and Representation Theory, Springer-
Verlag, New York, 1972.
127. , A construction of projective modules in the category O of Bernstein–
Gelfand–Gelfand, Indag. Math. 39 (1977), 301–303.
Bibliography 277

128. , Finite and infinite dimensional modules for semisimple Lie algebras, pp. 1–
64, Lie Theories and their Applications, Queen’s Papers in Pure and Appl. Math.
No. 48, Kingston, Ont., 1978.
129. , Highest weight modules for semisimple Lie algebras, pp. 72–103, Represen-
tation Theory I, Lecture Notes in Math., vol. 831, Springer-Verlag, Berlin, 1980.
130. , Reflection Groups and Coxeter Groups (Cambridge Studies in Advanced
Mathematics, 29), Cambridge Univ. Press, Cambridge, 1990.
131. , Conjugacy Classes in Semisimple Algebraic Groups, Math. Surveys Mono-
graphs, vol. 43, Amer. Math. Soc., Providence, RI, 1995.
132. , Modular Representations of Finite Groups of Lie Type, London Math. Soc.
Lecture Note Series, 326, Cambridge Univ. Press, Cambridge, 2006.
133. R.S. Irving, Projective modules in the category O, unpublished manuscript, 1982.
134. , Projective modules in the category OS : self-duality, Trans. Amer. Math. Soc.
291 (1985), 701–732.
135. , Projective modules in the category OS : Loewy series, Trans. Amer. Math.
Soc. 291 (1985), 733–754.
136. , The socle filtration of a Verma module, Ann. Sci. École Norm. Sup. (4) 21
(1988), 47–65.
137. , Singular blocks of the category O, Math. Z. 204 (1990), 209–224.
138. , BGG algebras and the BGG reciprocity principle, J. Algebra 135 (1990),
363–380.
139. , A filtered category OS and applications, Mem. Amer. Math. Soc. 83 (1990),
no. 419.
140. , Graded BGG algebras, pp. 181–200, Abelian Groups and Nonabelian Rings,
Contemp. Math., 130, Amer. Math. Soc., Providence, RI, 1992.
141. , Singular blocks of the category O, II, pp. 237–248, Kazhdan–Lusztig Theory
and Related Topics (Chicago, IL, 1989), Contemp. Math., 139, Amer. Math. Soc.,
Providence, RI, 1992.
142. , Shuffled Verma modules and principal series modules over complex semisim-
ple Lie algebras, J. London Math. Soc. 48 (1993), 263–277.
143. R.S. Irving and B. Shelton, Loewy series and simple projective modules in the category
OS , Pacific J. Math. 132 (1988), 319–342; correction, ibid. 135 (1988), 395–396.
144. N. Jacobson, Lie Algebras, Wiley Interscience, New York, 1962; Dover reprint, 1979.
145. J.C. Jantzen, Darstellungen halbeinfacher algebraischer Gruppen und zugeordnete
kontravariante Formen, Bonner Math. Schriften, No. 67, 1973.
146. , Zur Charakterformel gewisser Darstellungen halbeinfacher Gruppen und Lie-
Algebren, Math. Z. 140 (1974), 127–149.
147. , Kontravariante Formen auf induzierten Darstellungen halbeinfacher Lie-
Algebren, Math. Ann. 226 (1977), 53–65.
148. , Moduln mit einem höchsten Gewicht, Lecture Notes in Math., 750, Springer-
Verlag, Berlin, 1979.
149. , Einhüllende Algebren halbeinfacher Lie-Algebren, Springer-Verlag, Berlin,
1983 (reviewed by D.A. Vogan, Jr. in Bull. Amer. Math. Soc. 12 (1985), 279–283.
150. , Einhüllende Algebren halbeinfacher Lie-Algebren, pp. 393–401, Proc. Intern.
Congr. Math. (Warsaw, 1983), PWN, Warsaw, 1984.
278 Bibliography

151. , Primitive ideals in the enveloping algebra of a semisimple Lie algebra, pp. 29–
36, Noetherian Rings and their Applications, Math. Surv. Monogr., 24, Amer. Math.
Soc., Providence, RI, 1987.
152. , Lectures on Quantum Groups, Amer. Math. Soc., Providence, RI, 1996.
153. , Representations of Algebraic Groups, Academic Press, Orlando, 1987; 2nd
ed., Amer. Math. Soc., Providence, RI, 2003.
154. , Character formulae from Hermann Weyl to the present, to appear.
155. A. Joseph, Gelfand–Kirillov dimension for the annihilators of simple quotients of
Verma modules, J. London Math. Soc. 18 (1978), 50–60.
156. , Dixmier’s problem for Verma and principal series submodules, J. London
Math. Soc. 20 (1979), 193–204.
157. , Kostant’s problem, Goldie rank and the Gelfand–Kirillov conjecture, Invent.
Math. 56 (1980), 191–213.
158. , Towards the Jantzen conjecture, Compositio Math. 40 (1980), 35–67; II 40,
69–78; III 41 (1981), 23–30.
159. , The Enright functor and the Bernstein–Gelfand–Gelfand category O, Invent.
Math. 67 (1982), 423–445.
160. , Completion functors in the O category, pp. 80–105, Noncommutative har-
monic analysis and Lie groups (Marseille, 1982), Lecture Notes in Math., 1020,
Springer-Verlag, Berlin, 1983.
161. , Primitive ideals in enveloping algebras, pp. 403–414, Proc. Intern. Congr.
Math. (Warsaw, 1983), PWN, Warsaw, 1984.
162. , Quantum Groups and Their Primitive Ideals, Springer-Verlag, Berlin, 1995.
163. , Sur l’annulateur d’un module de Verma, pp. 237–300, Representation The-
ories and Algebraic Geometry (Montreal, PQ, 1997), NATO Adv. Sci. Inst. Ser. C
Math. Phys. Sci., 514, Kluwer Acad. Publ., Dordrecht, 1998.
164. A. Joseph, G.S. Perets, and P. Polo, Sur l’équivalence de catégories de Beilinson et
Bernstein, C.R. Math. Acad. Sci. Paris 313 (1991), 705–709.
165. V.G. Kac, Infinite-dimensional Lie algebras and Dedekind’s η-function, Functional
Anal. Appl. 8 (1974), 68–70.
166. , Infinite Dimensional Lie Algebras, 3rd ed., Cambridge Univ. Press, Cam-
bridge, 1990.
167. V.G. Kac and D.A. Kazhdan, Structure of representations with highest weight of
infinite-dimensional Lie algebras, Adv. Math. 34 (1979), 97–108.
168. J. Kåhrström, Tensoring with infinite-dimensional modules in O0 , arXiv:0708/2218
169. J. Kåhrström and V. Mazorchuk, A new approach to Kostant’s problem,
arXiv:0712.3117 [math.RT]
170. M. Kashiwara and T. Tanisaki, Characters of irreducible modules with non-critical
highest weights over affine Lie algebras, pp. 275–296, Representations and Quantiza-
tions (Shanghai, 1998), China High. Educ. Press, Shanghai, 2000.
171. D. Kazhdan and G. Lusztig, Representations of Coxeter groups and Hecke algebras,
Invent. Math. 53 (1979), 165–184.
172. , Schubert varieties and Poincaré duality, pp. 185–203, Geometry of the
Laplace operator, Proc. Sympos. Pure Math., XXXVI, Amer. Math. Soc., Provi-
dence, R.I., 1980.
173. A. Khare, Category O over a deformation of the symplectic oscillator algebra, J. Pure
Appl. Algebra 195 (2005), 131–166; erratum 199 (2005), 319–320.
Bibliography 279

174. , Axiomatic framework for the BGG category O, arXiv:math.RT/0502227

175. O. Khomenko, Categories with projective functors, Proc. London Math. Soc. 90
(2005), 711–737.
176. O. Khomenko and V. Mazorchuk, The Shapovalov from for generalized Verma mod-
ules, Dopov. Nats. Akad. Nauk Ukr. Mat. Prirodozn. Tekh. Nauki 1999, no. 6, 28–32.
177. , A note on simplicity of generalized Verma modules, Comment. Math. Univ.
St. Paul. 48 (1999), 145–148.
178. , On the determinant of Shapovalov form for generalized Verma modules, J.
Algebra 215 (1999), 318–329.
179. , Generalized Verma modules over the Lie algebra of type G2 , Comm. Algebra
27 (1999), 777–783.
180. , Schubert filtration for simple quotients of generalized Verma modules, Ark.
Mat. 38 (2000), 319–326.
181. , An irreducibility criterion for generalized Verma modules, Dopov. Nats.
Akad. Nauk Ukr. Mat. Prirodozn. Tekh. Nauki 2001, no. 5, 22–24.
182. , Generalized Verma modules induced from sl(2, C) and associated Verma mod-
ules, J. Algebra 242 (2001), 561–576.
183. , On multiplicities of simple subquotients in generalized Verma modules,
Czechoslovak Math. J. 52 (127) (2002), 337–343.
184. , Rigidity of generalized Verma modules, Colloq. Math. 92 (2002), 45–57.
185. , Structure of modules induced from simple modules with minimal annihilator,
Canad. J. Math. 56 (2004), 293–309.
186. , On Arkhipov’s and Enright’s functors, Math. Z. 249 (2005), 357–386.
187. S. Khoroshkin, Projective functors and the restriction of a Verma module to a sub-
algebra of Levi type, Ann. Global Anal. Geom. 10 (1992), 81–86.
188. S.L. Kleiman, The development of intersection homology theory,
arXiv:math.HO/0701462, to appear in Pure Appl. Math. Q.
189. A.W. Knapp, Lie Groups, Lie Algebras, and Cohomology, Math. Notes 34, Princeton
Univ. Press, Princeton, NJ, 1988.
190. S. König, The global dimension of the regular blocks of the BGG-category O of a
semisimple complex Lie algebra, unpublished preprint.
191. , Cartan decompositions and BGG-resolutions, Manuscripta Math. 86 (1995),
103–111.
192. , Blocks of category O, double centralizer properties, and Enright’s comple-
tions, pp. 113–134 in Algebra—Representation Theory, ed. K.W. Roggenkamp and
M. Ştefănescu, Kluwer, Dordrecht, 2001.
193. , Ringel duality and Kazhdan–Lusztig theory, Pacific J. Math. 203 (2002),
415–428.
194. S. König and V. Mazorchuk, Enright’s completions and injectively copresented mod-
ules, Trans. Amer. Math. Soc. 354 (2002), 2725–2743.
195. , An equivalence of two categories of sl(n, C)-modules, Algebr. Represent.
Theory 5 (2002), 319–329.
196. B. Kostant, A formula for the multiplicity of a weight, Trans. Amer. Math. Soc. 93
(1959), 53–73.
197. , Lie algebra cohomology and the generalized Borel–Weil–Bott theorem, Ann.
of Math. 74 (1961), 329–387.
280 Bibliography

198. S. Kumar, Bernstein–Gelfand–Gelfand resolution for arbitrary Kac–Moody algebras,

Math. Ann. 286 (1990), 709–729.
199. J. Lepowsky, Conical vectors in induced modules, Trans. Amer. Math. Soc. 208
(1975), 219–272.
200. , Existence of conical vectors in induced modules, Ann. of Math. 102 (1975),
17–40.
201. , Uniqueness of embeddings of certain induced modules, Proc. Amer. Math.
Soc. 56 (1976), 55–58.
202. , Generalized Verma modules, the Cartan–Helgason theorem, and the Harish-
Chandra homomorphism, J. Algebra 49 (1977), 470–495.
203. , A generalization of the Bernstein–Gelfand–Gelfand resolution, J. Algebra
49 (1977), 496–511.
204. G. Lusztig and D.A. Vogan, Jr., Singularities of closures of K-orbits on flag mani-
folds, Invent. Math. 71 (1983), 365–379.
205. A.V. Lutsyuk, Homomorphisms of the modules Mχ , Funktsional. Anal. i Prilozhen.
8 (1974), no. 4, 91–92; English transl., Funct. Anal. Appl. 8 (1975), 351–352.
206. C. Marastoni, Generalized Verma modules, b-functions of semi-invariants and duality
for twisted D-modules on generalized flag manifolds, C.R. Math. Acad. Sci. Paris 335
(2002), 111–116.
207. O. Mathieu, Classification of irreducible weight modules, Ann. Inst. Fourier (Greno-
ble) 50 (2000), 537–592.
208. H. Matumoto, On the existence of homomorphisms between scalar generalized Verma
modules, pp. 259–274, Representation Theory of Groups and Algebras, Contemp.
Math., 145, Amer. Math. Soc., Providence, RI, 1993.
209. , The homomorphisms between scalar generalized Verma modules associated
to maximal parabolic subalgebras, Duke Math. J. 131 (2006), 75–118.
210. V. Mazorchuk, α-stratified modules over the Lie algebra sl(n, C), Ukrain. Mat. Zh.
45 (1993), no. 9, 1215–1224; English transl., Ukrainian Math. J. 45 (1993), no. 9,
1360–1371.
211. , On the structure of an α-stratified generalized Verma module over Lie algebra
sl(n, C), Manuscripta Math. 88 (1995), 59–72.
212. , Generalized Verma Modules, Mathematical Studies Monograph Series, 8.
VNTL Publishers, L’viv, 2000.
213. , Twisted and shuffled filtrations on tilting modules, C.R. Math. Acad. Sci.
Soc. R. Can. 25 (2003), no. 1, 26–32.
214. , A twisted approach to Kostant’s problem, Glasg. Math. J. 47 (2005), 549–
561.
215. , Some homological properties of the category O, Pacific J. Math. 232 (2007),
313–341.
216. V. Mazorchuk, S. Ovsienko, and C. Stroppel, Quadratic duals, Koszul dual functors,
and applications, arXiv:math.RT/0603475, to appear in Trans. Amer. Math. Soc.
217. V. Mazorchuk and C. Stroppel, Translation and shuffling of projectively presentable
modules and a categorification of a parabolic Hecke module, Trans. Amer. Math. Soc.
357 (2005), 2939–2973.
218. , Projective-injective modules, Serre functors and symmetric algebras,
arXiv:math.RT/0508119, to appear in J. Reine Angew. Math.
Bibliography 281

219. , On functors associated to a simple root, J. Algebra 314 (2007), 97–128.

220. , Categorification of (induced) cell modules and the rough structure of gener-
alized Verma modules, arXiv:math.RT/070281
221. , Categorification of Wedderburn’s basis for C[Sn ], arXiv:0708.3949
[math.RT], to appear in Arch. Math. (Basel).
222. R. Mirollo and K. Vilonen, Bernstein–Gelfand–Gelfand reciprocity on perverse
sheaves, Ann. Sci. École Norm. Sup. (4) 20 (1987), 311–324.
223. R.V. Moody and A. Pianzola, Lie Algebras with Triangular Decompositions, Wiley
Interscience, New York–Toronto, 1995.
224. H. Oda and T. Oshima, Minimal polynomials and annihilators of generalized Verma
modules of the scalar type, J. Lie Theory 16 (2006), 155–219.
225. P. Ostapenko, Inverting the Shapovalov form, J. Algebra 147 (1992), 90–95.
226. A. Rocha–Caridi, Splitting criteria for g-modules induced from a parabolic and the
Bernstein–Gelfand–Gelfand resolution of a finite-dimensional, irreducible g-module,
Trans. Amer. Math. Soc. 262 (1980), 335–366.
227. A. Rocha–Caridi and N.R. Wallach, Projective modules over graded Lie algebras. I,
Math. Z. 180 (1982), 151–177.
228. , Highest weight modules over graded Lie algebras: resolutions, filtrations and
character formulas, Trans. Amer. Math. Soc. 277 (1983), 133–162.
229. , Characters of irreducible representations of the Virasoro algebra, Math. Z.
185 (1984), 1–21.
230. S. Ryom-Hansen, Koszul duality of translation and Zuckerman functors, J. Lie The-
ory 14 (2004), 151–163.
231. N.N. Shapovalov, On a bilinear form on the universal enveloping algebra of a complex
semisimple Lie algebra, Funktsional. Anal. i Prilozhen 6 (1972), no. 4, 65–70; English
transl., Funct. Anal. Appl. 6 (1972), 307–312.
232. B. Shelton, Extensions between generalized Verma modules: the Hermitian symmetric
cases, Math. Z. 197 (1988), 305–318.
233. W. Soergel, D-modules et équivalence de Enright–Shelton, C.R. Math. Acad. Sci.
Paris 307 (1988), 19–22.
234. , Universelle versus relative Einhüllende: Eine geometrische Untersuchung
von Quotienten von universellen Einhüllenden halbeinfacher Lie-Algebren, Math.
Ann. 284 (1989), 177–198.
235. , n-cohomology of simple highest weight modules on walls and purity, Invent.
Math. 98 (1989), 565–680.
236. , Parabolisch-singuläre Dualität für Kategorie O, Max-Planck-Institut, Bonn,
MPI/89–68.
237. , Kategorie O, perverse Garben und Moduln über den Koinvarianten zur Weyl-
gruppe, J. Amer. Math. Soc. 3 (1990), 421–445.
238. , Gradings on representation categories, pp. 800–806, Proc. Intern. Congr.
Math. (Zürich, 1994), Birkhäuser Verlag, Basel, 1995.
239. , Kazhdan–Lusztig polynomials and a combinatoric for tilting modules, Rep-
resent. Theory 1 (1997), 83–114.
240. , Character formulas for tilting modules over Kac–Moody algebras, Represent.
Theory 2 (1998), 432–448.
241. , Andersen filtration and Hard Lefschetz, to appear in Geom. Funct. Anal.
(2008)
282 Bibliography

242. T.A. Springer, Quelques applications de la cohomologie d’intersection, Bourbaki Sem-

inar, Vol. 1981/1982, Exp. 589, Astérisque, 92–93, Soc. Math. France, Paris, 1982.
243. R. Steinberg, Lectures on Chevalley Groups, Yale Univ. Math. Dept., 1967–68.
244. C. Stroppel, Homomorphisms and extensions of principal series, J. Lie Theory 13
(2003), 193–212.
245. , Category O: gradings and translation functors, J. Algebra 268 (2003), 301–
326.
246. , Category O: quivers and endomorphism rings of projectives, Represent.
Theory 7 (2003), 322–345.
247. , Composition factors of quotients of the universal enveloping algebra by prim-
itive ideals, J. London Math. Soc. 70 (2004), 643–658.
248. , Categorification of the Temperley–Lieb category, tangles, and cobordisms via
projective functors, Duke Math. J. 126 (2005), 547–596.
249. T. Tanisaki, Character formulas of Kazhdan–Lusztig type, pp. 261–276, Representa-
tions of finite dimensional algebras and related topics in Lie theory and geometry,
Fields Inst. Commun., 40, Amer. Math. Soc., Providence, RI, 2004.
250. V.S. Varadarajan, Lie Groups, Lie Algebras, and their Representations, Prentice–
Hall, Englewood Cliffs, NJ, 1974; reprinted by Springer-Verlag, New York, 1984.
251. D.N. Verma, Structure of certain induced representations of complex semisimple Lie
algebras, Ph.D. thesis, Yale Univ., 1966.
252. , Structure of certain induced representations of complex semisimple Lie al-
gebras, Bull. Amer. Math. Soc. 74 (1968), 160–166; errata, 628.
253. D.A. Vogan, Jr., Irreducible characters of semisimple Lie groups I, Duke Math. J.
46 (1979), 61–108.
254. , Irreducible characters of semisimple Lie groups II. The Kazhdan–Lusztig
conjectures, Duke Math. J. 46 (1979), 805–859.
255. , Irreducible characters of semisimple Lie groups III. Proof of the Kazhdan–
Lusztig conjecture in the integral case, Invent. Math. 71 (1983), 381–417.
256. , The character table for E8 , Notices Amer. Math. Soc. 54 (2007), 1122–1134.
257. M. Vybornov, Perverse sheaves, Koszul IC-modules, and the quiver for the category
O, Invent. Math. 167 (2007), 19–46.
258. N.R. Wallach, On the Enright–Varadarajan modules: a construction of the discrete
series, Ann. Sci. École Norm. Sup. (4) 9 (1976), 81–101.
259. C.A. Weibel, An Introduction to Homological Algebra, Cambridge Univ. Press, Cam-
bridge, 1994.
260. Wai Ling Yee, The signature of the Shapovalov form on irreducible Verma modules,
Represent. Theory 9 (2005), 638–677.
261. A.V. Zelevinsky, Resolutions, dual pairs, and character formulas, Funktsional. Anal.
i Prilozhen. 21 (1987), no. 2, 74–75; English transl., Funct. Anal. Appl. 21 (1987),
152–154.
262. G. Zuckerman, Tensor products of finite and infinite dimensional representations of
semisimple Lie groups, Ann. of Math. 106 (1977), 295–308.
263. V. Mazorchuk, Applications of the category of linear complexes of tilting modules
associated with the category O, to appear in Algebr. Represent. Theory
Frequently Used
Symbols

Symbol Description Section

g semisimple Lie algebra 0.1

h Cartan subalgebra 0.1
` rank of g (= dim h) 0.1
Φ root system of g relative to h 0.1
∆ simple system in Φ 0.1
Φ+ positive roots relative to ∆ 0.1
gα root space 0.1
n sum of positive root spaces 0.1
n− sum of negative root spaces 0.1
b standard Borel subalgebra h ⊕ n 0.1
p = pI parabolic subalgebra, I ⊂ ∆ 0.1
l = lI Levi subalgebra 0.1
u = uI nilradical of p 0.1
(hα , xα , yα ) standard basis elements, α ∈ Φ+ 0.1
E euclidean space spanned by Φ 0.2
α∨ coroot 2α/(α, α) 0.2
sα reflection relative to α ∈ Φ 0.2
Λr root lattice in E 0.2

283
284 Frequently Used Symbols

Symbol Description Section

W Weyl group 0.3

`(w) length of w ∈ W 0.3
w◦ longest element of W 0.3
WI parabolic subgroup of W (I ⊂ ∆) 0.3
w0 ≤ w Bruhat ordering of W 0.4
Z(g) center of U (g) 0.5
τ transpose map on U (g) 0.5
Λ integral weight lattice in E 0.6
Λ+ dominant integral weights 0.6
µ≤λ partial ordering of weights 0.6
Γ Z+ -linear combinations of ∆ 0.6
$α fundamental weight, α ∈ ∆ 0.6
ρ sum of fundamental weights 0.6
C Weyl chamber in E 0.6
Mod U (g) category of all U (g)-modules 0.7
Mλ weight space in M for λ ∈ h∗ 0.7
Π(M ) set of weights of M 0.7
O subcategory of Mod U (g) 1.1
Oint modules in O with integral weights 1.1
M (λ) Verma module 1.3
L(λ) simple quotient of M (λ) 1.3
N (λ) maximal submodule of M (λ) 1.3
χλ central character 1.7
w·λ w(λ + ρ) − ρ (w ∈ W, λ ∈ h∗ ) 1.8
[M : L(λ)] composition factor multiplicity 1.11
K(O) Grothendieck group of O 1.11
Rad M radical of module M 1.11
Soc M socle of module M 1.11
Oχ subcategory of O 1.12
Mχ summand of M in Oχ 1.12
O0 principal block of O 1.13
X0 additive group of formal characters 1.14
ch M formal character of M 1.14
p Kostant function in X0 1.16
P original Kostant partition function 1.16
Frequently Used Symbols 285

Symbol Description Section

M∨ dual of module M ∈ O 3.2

Φ[λ] integral root system of λ ∈ h∗ 3.4
∆[λ] simple system in Φ[λ] 3.4
W[λ] integral Weyl group of λ ∈ h∗ 3.4
Oλ
block in O 3.5
M : M (λ) standard filtration multiplicity 3.7
P (λ) projective cover of L(λ) 3.9
Q(λ) injective hull of L(λ) 3.9
µ↑λ strong linkage of weights 5.1
M (λ)i ith submodule in Jantzen filtration 5.3
M (λ)i ith layer in Jantzen filtration 5.3
Φ+λ index set in Jantzen sum formula 5.3
Tλµ translation functor 7.1
Hα α-hyperplane shifted by −ρ in E 7.3
Fb upper closure of facet F 7.3
Cb upper closure of Weyl chamber C 7.3
E(λ) span of Φ[λ] in E 7.4
λ\ integral part of λ ∈ h∗ 7.4
Wλ◦ stabilizer of λ 7.4
wλ longest element in W[λ] 7.13
Θs wall-crossing functor 7.14
`(x, w) `(w) − `(x) when x ≤ w 8.0
Px,w (q) Kazhdan–Lusztig polynomial, x ≤ w 8.2
G/B flag variety 8.5
Xw Bruhat cell in G/B 8.5
Op parabolic subcategory of O 9.3
MI (λ) parabolic Verma module 9.4
WI minimal coset representatives in WI \W 9.4
Index

adjoint group, 2, 27 composition series, 29

adjoint representation, 2 contravariant form, 68
adjunction morphism, 149, 236 convolution, 33
affine Lie algebra, 257 coroot, 3
affine reflection, 53 Coxeter group, 4
affine Weyl group, 53
alternating function, 39 direct sum of functors, 214
annihilator of a module, 215 discrete series, 247
antidominant weight, 54 dominant chamber, 133
artinian category, 28 dominant integral weight, 7
dominant weight, 54
BGG algebra, 260 dot action, 24
BGG category, 13 dot-regular weight, 24
BGG reciprocity, 65 dual root system, 3
BGG resolution, 108 duality in O, 50
block, 30
Borel subalgebra, 2 facet, 132
braid relations, 236 family of twisted Verma modules, 239
Bruhat cell, 160 filtration, 17
Bruhat decomposition, 160 filtration length, 58
Bruhat ordering, 5 flag variety, 122, 160
formal character, 32
fundamental weight, 7
Cartan decomposition, 2
Cartan invariant, 3
Cartan matrix, 257 Gabber–Joseph Conjecture, 170
Cartan subalgebra, 1 Gelfand–Kirillov dimension, 254
category O, 13 generalized Cartan matrix, 257
center of U (g), 6 generalized KL Conjecture, 172
central character, 23 generalized Verma module, 188
chamber, 133 global dimension, 120
Chevalley Restriction Theorem, 27 Goldie rank, 254
Chevalley–Bruhat ordering, 5 Goldie rank polynomial, 254
coinvariant algebra, 265, 266 Grothendieck group, 29
compatible weights, 130
complete module, 246 Harish-Chandra homomorphism, 23
completion, 246 Harish-Chandra module, 168, 220

287
288 Index

head, 29 parabolic category Op , 184

head of a module, 174 parabolic subalgebra, 2
Hecke algebra, 156 parabolic subgroup of W , 4
height of positive root, 3 parabolic Verma module, 186
Hermitian symmetric pair, 205 parabolic Verma module of scalar type, 195
highest weight category, 260 parabolic–singular duality, 269
highest weight module, 16 partial ordering of weights, 9
homological dimension, 120 PBW basis of U (g), 6
Poincaré polynomial, 160
indecomposable functor, 214 primitive ideal, 215, 252
induced module, 17 primitive vector, 16
induction, 17 principal block, 31
injective hull, 63 principal series module, 221
injective object, 60 projective χ-functor, 216
integral weight lattice, 7 projective cover, 62
Iwahori–Hecke algebra, 156 projective dimension, 120
projective functor, 207, 214
Jantzen Conjecture, 97, 171 projective generator, 68
Jantzen middle, 149 projective object, 60
Jantzen sum formula, 96
Jordan–Hölder length, 29 quiver, 261

Kac–Moody algebra, 257 radical, 29

Kazhdan–Lusztig Conjecture, 159 radical filtration, 174
Kazhdan–Lusztig polynomial, 157 radical of a module, 174
Killing form, 2 real rank one, 204
KL Conjecture, 159 reduced expression in W , 4
KL polynomial, 157 reduced tensor product, 232
KLV polynomial, 158, 169 reductive Lie algebra, 2
Kostant function, 34 regular weight, 24
Kostant partition function, 34 relative Hecke module, 190
Kostant’s Problem, 255 relative KL polynomial, 190
Koszul duality, 269 relative Verma module, 188
Koszul ring, 269 representation type, 263
rigid module, 174
length function, 4 root lattice, 3
length of a module, 29 root space, 2
Levi subalgebra, 2 root system, 2, 3
linkage class, 24
linked weights, 24 scalar type, 195
Loewy filtration, 174 Schubert variety, 160
Loewy layer, 174 semisimple Lie algebra, 1
Loewy length, 174 Shapovalov matrix, 102
Loewy series, 174 shuffled Verma module, 236
long exact sequence, 48 shuffling functor, 149, 237
longest element of W , 4 simple reflection, 4
loop algebra, 257 simple system, 2, 3
lower closure, 133 singular weight, 24
socle, 29
maximal vector, 15 socle filtration, 174
multiplicity free block, 206 socle of a module, 174
multiplicity of composition factor, 29 socular weight, 200
multiplicity of weight, 9 square, 117
standard basis, 2
noetherian category, 14 standard filtration, 58, 191
standard map, 193
Ore localization, 244 standard PBW ordering, 6
Index 289

strong linkage, 93
strongly linked weights, 93
sum formula, 96
symmetrizable Kac–Moody algebra, 257

Tensor Identity, 56
tilting module, 223, 224
tilting module in Op , 234
translation functor, 130
transpose map, 7
truncation, 185
twisted Harish-Chandra homomorphism, 26
twisted Verma module, 236, 239
twisting functor, 242

universal enveloping algebra, 6

universal highest weight module, 18
upper closure, 133

Verma flag, 58
Verma module, 18
Vogan’s Conjecture, 168

wall, 133
wall-crossing functor, 148, 236
weak BGG resolution, 109
weight, 9
weight filtration, 173
weight module, 9
weight space, 9
Weyl chamber, 8, 132
Weyl Denominator Formula, 41
Weyl group, 4
Weyl’s Complete Reducibility Theorem, 9
 Representations of Semisimple Lie Algebras
in the BGG Category O
(Revisions)
xvi In last paragraph of Preface, read “Stroppel” in place of “Stropple”.

35 The last symbol in the Exercise should be [M (λ)].

124 In line 8 of (1) in the proof, the first Ext term in the exact sequence
should be ExtO .

146 In line 1, replace P (wλ · λ) by P (wλ wµ◦ · λ).

150 In line −1, replace “bonce” by “once”.

200 In the third line of 9.15, read “more refined partition”.

243 Revise the third paragraph of 12.7, starting with line 4:

“and Nw = C[yα ]. The algebra Nw is Z-graded: the standard grading
of U by Λr induces a Z-grading on U if all simple root vectors are
placed in U1 . The graded dual Nw∗ , with nth graded piece the dual
space (Nw )∗−n , then becomes a Z-graded Nw -bimodule. Define Sw :=
U ⊗Nw Nw∗ . Somewhat miraculously, . . . ”

265 In line −4, replace “on W ” by “on V ”.

269 In line −15, replace WI by WI .