Patrick Lei

All notes

Category O

Talks by Kevin Chang, Fan Zhou, Che Shen and Patrick Lei, Columbia, Fall 2021.

Category O Learning Seminar
Fall 2021
Notes by Patrick Lei
Lectures by Various

Columbia University

Disclaimer

These notes were taken during the seminar using the vimtex package of the editor neovim. Any errors are mine and not the speakers’. In addition, my notes are picture-free (but will include commutative diagrams) and are a mix of my mathematical style and that of the lecturers. If you find any errors, please contact me at plei@math.columbia.edu.

Also, we should note that Fan used different notation than the rest of us because he likes to be special.

1 Kevin (Sep 29): Review of semisimple Lie algebras and introduction to category 𝒪

1.1 Review of semisimple Lie algebras

Throughout this lecture, we will work over ℂ.

Definition 1.1.1.

A Lie algebra 𝔤 is semisimple if any of the following equivalent conditions hold:

  1. 1.

    𝔤 is a direct sum of simple Lie algebras (those with no nonzero proper ideals).

  2. 2.

    The Killing form κ⁡(x,y)≔tr⁡(ad⁡(x)⁢ad⁡(y)) is nondegenerate.

  3. 3.

    The radical (maximal solvable ideal) of 𝔤 is zero.

Some examples of semisimple Lie algebras include 𝔰⁢𝔩n,𝔰⁢𝔬n,𝔰⁢𝔭2⁢n, and in some sense (the classification of simple Lie algebras), these are essentially all semisimple Lie algebras.

Now given a semisimple Lie algebra 𝔤, we will fix a Cartan subalgebra 𝔥⊂𝔤, which is just a maximal abelian subalgebra of semisimple elements. This gives us a root decomposition

where 𝔤α is the subspace of 𝔤 where 𝔥 acts with weight α. Some important facts about these root systems are the following:

  • •

    For all α, we have dim𝔤α=1.

  • •

    For all roots α,β, we have [𝔤α,𝔤β]⊂𝔤α+β.

  • •

    If α is a root, so is −α.

In addition, the α are required to form a (reduced) root system (denoted Φ), the precise definition of which is deliberately omitted. Given a choice of Borel subalgebra containing 𝔥, we obtain a set Φ+ of positive roots and a set Δ of simple roots. In addition, given a root system Φ, there is a dual root system Φ∨, whose roots are

Now suppose that 𝔤 is a semisimple Lie algebra with root system Φ. For every α∈Φ+, we may choose xα∈𝔤α and yα∈𝔤−α, and these determine some hα=[xα,yα]∈𝔥. This choice can be made such that α⁡(hα)=2.

Recall that the Lie algebra 𝔰⁢𝔩2 is spanned by the matrices

Then the choice of xα,yα,hα gives an embedding 𝔰⁢𝔩2→𝔤. These maps, ranging over all α, cover all of 𝔤. Now a basis of 𝔤 is given by xα,yα,α∈Φ and hαi for the simple roots αi. Therefore, to specify 𝔤, we only need to give commutation relations for the basis elements.

Now suppose that Φ is some root system. We would like to construct a semisimple Lie algebra 𝔤 with root system Φ. We want to build a semisimple Lie algebra. To do this, choose a set of simple roots αi, and consider the Lie algebra

where the relations are as follows:

  • •

    [hαi,hαj]=0.

  • •

    We have [xαi,yαj]=hαi if i=j and this commutator vanishes otherwise.

  • •

    [hαi,xαj]=⟨αj,αi∨⟩⁢xαj.

  • •

    [hαi,yαj]=−⟨αj,αi∨⟩⁢yαj.

  • •

    ad⁡(xαi)1−⟨αj,αi∨⟩⁢(xαj)=0 if i≠j.

  • •

    ad⁡(yαi)1−⟨αj,αi∨⟩⁢(yαj)=0 if i≠j.

The first four relations are called the Weyl relations and the last two are called the Serre relations. Given this data, we end up with a semisimple Lie algebra 𝔤Φ with root system Φ. In addition, if 𝔤 is any other semisimple Lie algebra with root system Φ, there is an isomorphism 𝔤Φ→∼𝔤. Moreover, we have a bijection between semisimple Lie algebras and reduced root systems, which restricts to a bijection between simple Lie algebras and irreducible root systems.

Table 1.1: Root systems and Lie algebras
Irreducible root systems simple Lie algebras
An 𝔰⁢𝔩n+1
Bn 𝔰⁢𝔬2⁢n+1
Cn 𝔰⁢𝔭2⁢n
Dn 𝔰⁢𝔬2⁢n
E6,E7,E8,F4,G2 exceptional Lie algebras

We will now discuss the finite-dimensional representation theory of semisimple Lie algebras 𝔤.

Theorem 1.1.2 (Weyl’s complete reducibility theorem).

Any finite-dimensional representation of 𝔤 is decomposes as a direct sum of simple representations.

Now suppose that M is a finite-dimensional 𝔤-representation. Then M has a weight decomposition

These λ are integral weights, which simply means that ⟨λ,α∨⟩∈ℤ for all roots α. For any root α, xα⁢(Mλ)⊂Mλ+α and yα⁢(Mλ)⊂Mλ−α. We would like to think that the xα raise the weights and yα lower the weights, so we introduce a partial order. We say that λ≥μ if λ−μ∈ℤ≥0⁢Φ+.

By Weyl’s complete reducibility theorem, it remains to classify the irreducible representations of 𝔤. These are in bijection with the dominant integral weights, which in particular means that ⟨λ,α∨⟩≥0 for all α∈Φ+. For any dominant weight λ, there is a unique highest-weight representation L⁡(λ). Here, L⁡(λ) is generated by a single maximal vector v of weight λ. This means that for all positive roots α, xα⁢v=0.

1.2 Introduction to category 𝒪

We would now like to study infinite dimensional representations of 𝔤. Of course, this is impossibly complicated in general, so we will impose some finiteness conditions on our representations.

Definition 1.2.1.

The category 𝒪 is the full subcategory of U⁡(𝔤)-modules M satisfying:

  1. 1.

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

  2. 2.

    M is 𝔥-semisimple and has a weight decomposition M=⨁λ∈𝔥∗Mλ.

  3. 3.

    M is locally 𝔫-finite, where 𝔫=⨁α∈Φ+𝔤α. Precisely, this means that the U⁡(𝔫) generated by any v∈M is finite-dimensional.

Here are some facts about category 𝒪, which are stated without proof.

  • •

    For all M in our category and weights λ, the weight space Mλ is finite-dimensional.

  • •

    𝒪 is a Noetherian (everything satisfies the descending chain condition) abelian category.

We will now describe some infinite-dimensional objects in category 𝒪.

Definition 1.2.2.

For any weight λ, the Verma module M⁡(λ) associated to λ is the module

where 𝔟=𝔥+𝔫 is the Borel subalgebra associated to our choice of positive roots and ℂλ is the 𝔟-module associated to the 1-dimensional representation of 𝔥 with weight λ and the identification 𝔟/𝔫=𝔥.

2 Fan (Oct 06): Beginnings in category 𝒪: Vermas, central characters, and blocks

Recall the following fact ahout semisimple Lie algebras. If we have a decomposition

where

  1. 1.

    dim𝔤α=1;

  2. 2.

    ℤ⁢Φ⊂𝔥∗ is a lattice of maximal rank;

  3. 3.

    α∈Φ implies −α∈Φ;

  4. 4.

    [[𝔤α,𝔤−α],𝔤α]≠0,

then 𝔤 is a semisimple Lie algebra.

Now recall that the Weyl group W is the group generated by the reflections sα in the roots. For any w∈W, we define the length

Next, there is the Bruhat order, where if w2=s⁢w1 and ℓ⁡(w2)>ℓ⁡(w1), we say w1<w2.

Finally, throught this lecture, we will denote the weight lattice by Λ and the root lattice by Q. In addition, we will denote the λ-weight space of a module M by Mλ, and Mλ will be the Verma associated to λ. Also, we will need the notion of the universal enveloping algebra, which we will not write down here.

2.1 Definitions

Recall that 𝒪 is the full subcategory of 𝖬𝗈𝖽⁡(𝒰⁢𝔤) of modules M such that

  1. 1.

    M is finitely generated over 𝒰⁢𝔤.

  2. 2.

    M is 𝔥-semisimple.

  3. 3.

    M is locally 𝔫-finite.

  4. 4.

    dimMλ<∞.

  5. 5.

    The set of weights of M is contained in some finite union of cones λ−Q+.

Theorem 2.1.1.

The following properties hold for 𝒪:

  1. 1.

    𝒪 is Noetherian.

  2. 2.

    𝒪 is closed under submodules, quotients, finite direct sums, and is abelian.

  3. 3.

    𝒪 is closed under tensoring with finite-dimensional representations (in fact, if tensoring with 𝒪 is exact and lands in 𝒪, then N must be finite-dimensional).

  4. 4.

    ℳ is locally Z⁢𝔤-finite.

  5. 5.

    All M∈𝒪 are finitely generated over 𝒰⁢𝔫−.

2.2 Highest weight modules

Definition 2.2.1.

A vector v+ is a maximal vector if 𝔫+⁢v+=0.

Definition 2.2.2.

A module M is a highest weight module if there exists a maximal v+∈M generateing M.

Definition 2.2.3.

Let λ∈𝔥∗ and consider the 𝔟+-module ℂλ. Then the Verma module for λ is the module

Note that we have the standard adjunction Hom𝔤⁡(Mλ,−)=Hom𝔟⁡(ℂλ,−).

Theorem 2.2.4.

For any highest weight module M with highest weight λ,

  1. 1.

    M=⟨f1n1⁢⋯⁢f|Φ+|n|Φ|⟩, and in particular M is 𝔥-semisimple.

  2. 2.

    All weights of M are at most λ.

  3. 3.

    For any μ<λ, dimMμ<∞, and dimMλ=1. In addition, M∈𝒪.

  4. 4.

    Any quotient of M is also a highest weight module with highest weight λ.

  5. 5.

    Any submodule of a highest weight module with weight μ<λ is a proper submodule. if M is simple, all maximal vectors have weight v+λ.

  6. 6.

    There exists a unique maximal submodule, and thus M has a unique simple quotient and thus is indecomposable.

  7. 7.

    All simple highest weight modules with highest weight λ are isomorphic, so dimEnd⁡M=1.

Corollary 2.2.5.

Let M∈𝒪. Then M admits a filtration whose successive quotients are highest weight modules.

2.3 Verma modules

Let Mλ be a verma module and Lλ be the unique simple quotient, and Nλ be the unique maximal submodule.

Theorem 2.3.1.

Any simple L∈𝒪 is isomorphic to Lλ for some λ.

Proposition 2.3.2.

Let Σ be the set of simple roots and σ∈Σ. Let λ∈𝔥∗ such that σ∗⁢(λ)∈ℕ. Choose v+λ∈Mλ a maximal vector. Then

In particular, there exists a nonzero morphism Mλ−(σ∗⁢λ+1)⁢σ↪Mλ.

Lemma 2.3.3.

We have the commutation relations

2.4 Examples

We will discuss the example of 𝔰⁢𝔩2. Let ϕi be the operator that outputs the i-th diagonal of a matrix. Then let α=2⁢ϕ1=ϕ1−ϕ2 be the root. Let α† be the matrix such that κ⁡(α†,−)=α⁡(−). In particular, we have α†=14⁢h.

Then note that if we choose units so that ϕ1=1, then α=2. If λ=n, then the Verma module for λ has weights n,n−2,…, and the simple module has weights n,n−2,…,−n. For non-integral weights, we just have an infinite-dimensional representation. To see this, note that hitting any non-integral weight with 𝔫− will not reach another maximal vector.

2.5 Finite-dimensional modules

Theorem 2.5.1.

For any weight λ, dimLλ<∞ if and only if λ∈Λ+ is a dominant integral weight. This is equivalent to dimLλμ=dimLλw⁡(μ) for all w∈W.

This result tells us that weights of Lλ are actually symmetric under the Weyl group.

Proof.

First, if the span of v∈Mλ is finite-dimensional for 𝔰⁢𝔩2, then all of 𝔥 stabilizes Span𝔰⁢𝔩2⁢v. This is because if vμ∈N≔⟨v⟩, then

and thus h⁡(fi⁢vμ)∈ℂ⁢fi⁢vμ.

Next, if dimLλ<∞, then after restriction to 𝔰⁢𝔩2, we have λ⁡(hi)=αi∗⁢(λ)∈ℕ, and thus λ∈Λ+.

Now suppose that λ∈Λ+. Then after restricting Lλ to 𝔰⁢𝔩2i, the span of v+λ is isomorphic to Lαi∗⁢(λ)⋅ϕ1, and in particular it is finite-dimensional. Next, we show that Lλ is a sum of finitely many 𝔰⁢𝔩2i-summands. To see this, consider the sum M of all 𝔰⁢𝔩2i-submodules of Lλ. But then if we denote a summand by N, we note that 𝔤⊗N is a finite-dimensional representation of 𝔰⁢𝔩2, so the natural morphism

lands inside M. But then M=Lλ because M is a nonzero submodule.

Next, recall that for M∈Rep⁡(𝔰⁢𝔩2), then the sets of weights are invariant under reflection across the origin, so we have isomorphisms

Because Lλ is a sum of finite-dimensional representations of 𝔰⁢𝔩2, for any v∈Lλμ, consider the finite-dimensional 𝔰⁢𝔩2-module containing it. If we add them up, we know that Lλμ is some finite-dimensional 𝔰⁢𝔩2-module Nμ. But then the si generate W, and thus for all w∈W, Lλμ≅Lλw⁡(μ).

Finally, for any orbit of W, there exists exactly one representative in the dominant weight lattice, and because there are only finitely many dominant integral weights less than λ, there must only be finitely many orbits, so Lλ is finite-dimensional. □

2.6 Central actions

Here, we will consider the action of Z⁢𝔤 on a module M. Suppose that M is a highest weight module for weight λ. Then we note that

and therefore z⁢v+λ=ϑλ⁢(z)⁢v+λ. Therefore z acts by 𝒱λ⁢(z) on any highest weight module of weight λ, and we call the function ϑλ:Z⁢𝔤→ℂ a central character. In general, all algebra morphisms Z⁢𝔤→ℂ arise in this way. Then we have the decomposition

and write π𝔥:𝒰⁢𝔤→𝒰⁢𝔥 for the morphism killing 𝔫±. Then ϑλ⁢(z)=λ⁡(π𝔥⁢(z)), so π𝔥:ℤ⁢𝔤→𝒰⁢𝔥 is an algebra homomorphism, and we will call this φHC≕ϖ, the Harish-Chandra morphism. In particular, we obtain a morphism

Also, we will consider ϖ∘w⋅, where w∘λ=w⁡(λ+ϱ)−ρ, where

If we can identify the two morphisms on a Zariski-dense subset, they must agree in general. First, note that ϖ⁡(λ)=ϑλ, and now it suffices to show that ϑλ=ϑw∘λ for λ∈Λ.

To prove this, if there exists σ∈Σ such that σ∗⁢(λ)∈N, then Msσ∘λ⊂Mλ, and thus θλ=θsσ∘λ. In addition, if σ∗⁢(λ)=−1, we have sσ∘λ=λ, and if σ∗⁢λ≤−2, we can reverse the roles of λ,σ∘λ because

2.7 More on Harish-Chandra

Consider the twisted Harish-Chandra morphism

This gives us a morphism ψHC. In particular, we have

Theorem 2.7.1.

The image of ψHC is contained in (S⁢𝔥)W.

To see this, note that

Theorem 2.7.2.
  1. 1.

    ψHC is an isomorphism Z⁢𝔤→S⁢𝔥W;

  2. 2.

    If λ,μ are linked, then ϑλ=ϑμ;

  3. 3.

    Every element of Hom𝖠𝗅𝗀⁡(Z⁢𝔤,ℂ) arises in this way.

There is a simple way to see the last two parts of the theorem if we assume some algebraic geometry.

For a central character ϑ:Z⁢𝔤→ℂ, consider the module

where n depends on z. We have a decomposition

which gives us

Now we may define subcategories of 𝒪 given by

Some examples are that all highest weight modules of weight λ are contained in 𝒪ϑλ.

Proposition 2.7.3.

We have a decomposition

We will now consider blocks of category 𝒪. We say that simple modules S1,S2 are in the same block if there is a nontrivial extension of S2 by S1. For general M, we know that M has a finite Jordan-Hölder decomposition because 𝒪 is Artinian, so M is in some block if all of its Jordan-Hölder quotients are.

Proposition 2.7.4.

If λ∈Λ, then 𝒪ϑλ is a block of 𝒪.

To prove this, if μ<λ are linked, then we have the diagram

giving us an exact sequence

3 Che (Oct 13): Formal characters and applications to finite-dimensional modules

3.1 Weyl character and dimension formulas

Today we will see what Category 𝒪 tells us about finite-dimensional modules. We will fix a semisimple Lie algebra 𝔤=𝔫+⊕𝔥⊕𝔫−.

Definition 3.1.1.

Given M∈𝒪, define the function ch⁡M:𝔥∗→ℤ≥0, where λ↦dimMλ.

Also let eλ be the characteristic function of λ. Now given f,g:𝔥∗→ℤ≥0, define the convolution product

For example, eλ∗eμ=eλ+μ. Here, we assume that f,g are supported on a finite union of things of the form λ−Γ, where Γ are the non-negative weights. We will call the set of such functions 𝒳.

Proposition 3.1.2.
  1. 1.

    If 0→M′→M→M′′→0 is an exact sequence, then ch⁡(M)=ch⁡(M′)+ch⁡(M′′).

  2. 2.

    If M∈𝒪 and L is finite-dimensional, then ch⁡(L⊗M)=ch⁡(M)∗ch⁡(L).

Proof.
  1. 1.

    Note that dimMμ=dimMμ′+dimMμ′′ for any such exact sequence.

  2. 2.

    Note that dim(L⊗M)λ=∑μ+ν=λdimLμ⁢dimMν. □

Last time we considered central characters for a highest weight module of weight λ and highest weight vector v+

If L⁡(μ) is a subquotient of the Verma module M⁡(λ), then, then χμ=χλ. Equivalently, μ and λ are linked by some element w of the Weyl group. Because 𝒪 is Artinian, for all M∈𝒪 we have a finite filtration

where all Mi+1/Mi is simple, so they must be isomorphic to some L⁡(λi). But then we have

Our goal is now to compute the a⁡(λ,w).

Let λ1,λ2,…,λn be weights linked to lambda arranged such that if λi≥λj, then i≥j. Then we should have some identity of the form

This implies that ch⁡L⁡(λ)=∑w∈Wb⁡(λ,w)⁢ch⁡M⁡(w∘λ). Note that b⁡(λ,1)=1.

Definition 3.1.3.

Define the Kostant function

Proposition 3.1.4.

p=ch⁡M⁡(0). More generally, eλ∗0=ch⁡M⁡(λ).

Proof.

By the PBW theorem, we know that M⁡(0) is spanned by 𝒰⁡(𝔫−). This is apparently equivalent to the definition of p. □

Definition 3.1.5.

Define the function

Also define

Note that fα∗(1−e−α)=1. Also note that

Next, p=∏α>0fα and if α is a simple root, then sα⋅q=−q. The reason for this is that sα⁢(α)=−α but sα fixes the other positive roots. This implies that w⋅q=(−1)ℓ⁡(w)⁢q.

Theorem 3.1.6 (Weyl character formula).

If λ∈Λ+, then

Proof.

If we apply q∗− to the formula

we obtain

Because λ∈Λ+, we know that L⁡(λ) is finite-dimensional and all weight spaces are symmetric. If α is a simple root, then we apply sα to both sides, and we obtain

because sα⁢(w∘λ+ρ)=sα⁢w⁢(λ+ρ)=sα⁢w∘λ+ρ. Therefore we see that b⁡(λ,sα⁢w)=−b⁡(λ,w), so b⁡(λ,w)=(−1)ℓ⁡(w). □

We would now like to compute dimL⁡(λ) for dominant integral weights λ. We want something like

except that both sides here vanish, so this is too naïve. For example, if we consider 𝔰⁢𝔩2, we have

If we divide, we actually obtain ch⁡L⁡(λ)=eλ+eλ−2+⋯+e−λ.

In the general case, let μ∈𝔥∗ and t∈ℝ. Define Fμ,t:𝒳→ℝ by extending eλ↦et⁡(λ,μ) linearly. Applying Fρ,t to the Weyl charaacter formula, we obtain

Note here that Fρ,t⁢(eλ∗eμ)=Fρ,t⁢(eλ)⋅Fρ,t⁢(eμ). In the t→0 limit, we have Fρ,t⁢ch⁡L⁡(λ)→dimL⁡(λ) and et⁡(ρ,ρ)→1. Therefore we obtain

This is called the Weyl dimension formula.

3.2 Maximal submodules of Verma modules

Theorem 3.2.1.

Let λ∈Λ+ and α1,…,αk be simple roots of 𝔤. Then

is the maximal submodule of M⁡(λ).

Remark 3.2.2.

Last time we saw that M⁡(sαi∘λ)⊂M⁡(λ).

Lemma 3.2.3.

Let a,b∈𝒰⁢𝔤. Then

This is proved by induction, so like a certain Fields medalist, we omit the proof.

If xα,xβ correspond to roots α,β, then note that eventually we will have [xα,⋯⁣[xα,[xα,xβ]]]=0. In fact, four times is enough.

Lemma 3.2.4.

Let α be a simple root. Then for any v∈M⁡(λ), there exists N≫0 such that yαN⋅v=0 in M⁡(λ)/∑M⁡(sαi∘λ).

Proof.

We proceed by induction. Suppose that v=yi1⁢yi2⁢⋯⁢yit⁢v+. When t=0, then yα(α,λ)+1⁢v+=0. For t>0, we have

The first term on the right hand side vanishes by the inductive hypothesis, and the second term becomes (−)⋅yαN−3⁢yi2⁢⋯⁢yi+=0 by the inductive hypothesis. □

Proof of Theorem.

By the discussion last time and the second lemma, we know that M⁡(λ)/∑M⁡(sαi∘λ) is finite-dimensional. This implies that M⁡(λ)/∑M⁡(sλi∘λ)=L⁡(λ)⊕M′, but we know that the left hand side is a highest weight module, so M′=0. □

Remark 3.2.5.

We have a resolution

This is called the BGG resolution.1

4 Kevin (Oct 20): Duality and projectives in category 𝒪

4.1 Duality

Recall from the finite-dimensional story that 𝔤-representation M have duals M∨ with 𝔤-action

for x∈𝔤,f∈M∗,v∈M. This is not well-behaved for infinite-dimensional representations (for example, M∗⁣∗≇M), so in this case we would like to construct a better-behaved duality functor.

Note that every semisimple Lie algebra 𝔤 has a transpose τ:𝔤→𝔤 (if 𝔤 is a matrix Lie algebra, this is literally the transpose) which is an anti-automorphism. Here, we have

This allows us to define1

Definition 4.1.1.

Let M=⨁λMλ∈𝒪. Then the dual of M is defined by

Proposition 4.1.2.

M∨∈𝒪.

Proof.

To prove finite generation, note that M∨ has finite length (here, L⁢(λ)∨=L⁡(λ) because duality preserves formal characters and exchanges quotients and submodules). Clearly the weight spaces are finite-dimensional by assumption, and the weights lie in some union ⋃λ−Λ because formal characters are preserved, so we have local 𝔫-finiteness. □

Here are some more facts about duality.

  • •

    Duality is a contravariant functor. This is obvious because everything is defined on the level of weight spaces.

  • •

    There is a natural isomorphism M∨⁣∨≅M. This is clear because we are taking double duals of finite-dimensional things and adding them up, so in particular duality is an anti-equivalence of categories.

  • •

    We have L⁢(λ)∨≅L⁡(λ). On the other hand, duality for M⁡(λ) is complicated.

  • •

    τ fixes Z⁡(𝔤) by an exercise in Humphreys.2 In particular, this means that (Mχ)∨=(M∨)χ.

4.2 Projectives

Recall that P is projective if Hom⁡(P,−) is right exact.3 Our goal is to prove that 𝒪 has enough projectives (which will mean that we can do homological algebra). The first thing we will do is introduce dominance and antidominance.

Recall that for λ∈Λ, W⁢λ contains one dominant weight and one antidominant weight. This gives us two(!) good choices for representatives of W⁢λ. Unfortunately, we care about nonintegral weights,4 and we cannot choose representatives of W⁢λ for general λ∈𝔥∗.

Remark 4.2.1.

From now on we will use the w∘− action (because this is all we care about), and therefore our new notion of (anti)dominance will not restrict to the old notion of dominance.5

Definition 4.2.2.

A weight λ∈𝔥∗ is dominant if ⟨λ+ρ,α∨⟩∉ℤ<0 for all α∈Φ+. A weight λ∈𝔥∗ is antidominant if ⟨λ+ρ,α∨⟩∉ℤ>0 for all α∈Φ+.

Note that this is not the same as the undotted definition. For example, −ρ is dominant. Also, the set W∘λ can have multiple dominant and/or antidominant weights.

Definition 4.2.3.

We define the subgroup

where Λr is the root lattice. We also define

In fact, W[λ] is the Weyl group of Φ[λ]. We may similarly define Δ[λ].

Proposition 4.2.4.

The following are equivalent:

  1. 1.

    λ is dominant.

  2. 2.

    ⟨λ+ρ,α∨⟩≥0 for all α∈Δ[λ].

  3. 3.

    λ≥sα∘λ for all α∈Δ[λ].

  4. 4.

    λ≥w∘λ for all w∈W[λ].

Proof.

Clearly 1 implies 2, and 2 implies 1 because positive roots are sums of simple roots with nonnegative coefficients. To prove that 2 is equivalent to 3, note that

Finally, to see that 3 is equivalent to 4, note that 4 implies 3 automatically. To prove that 3 implies 4, we induct on ℓ⁡(w). If w=w′⁢sα with ℓ⁡(w′)=ℓ⁡(w)−1, we see that

It is clear that λ−w′∘λ≥0 while λ−sα∘λ is a nonnegative multiple of α. Because of the length condition, we see that w′∘α is positive. □

Corollary 4.2.5.

The orbit W[λ]∘λ has a unique (anti)dominant weight.

Proof.

This is because 1 is equivalent to 4 in the proposition. □

Theorem 4.2.6.
  1. 1.

    If λ is dominant, then M⁡(λ) is projective.

  2. 2.

    If P∈𝒪 is projective and L∈𝒪 is finite-dimensional, then P⊗L is projective.

  3. 3.

    𝒪 has enough projectives.

Proof.
  1. 1.

    Consider M↠N and suppose v∈N is a maximal weight vector with weight λ (coming from a map M⁡(λ)→N). Assume that M=Mχ,N=Nχ. Our goal is to lift v to a maximal weight vector in M, but because M→N is surjective, we can lift v to v′∈Mλ. If v′ is maximal, then we are done, so suppose that v′ is not maximal.

    In this case, there exists x∈𝒰⁢𝔫 such that x⁢v′ is a maximal vector with weight greater than λ. However, this weight must be linked to λ, so by dominance of λ, it cannot exist.

  2. 2.

    Here, we use the tensor-Hom adjunction

    Because L∗⊗− is exact and Hom𝒪⁡(P,−) is exact, the functor Hom𝒪⁡(P⊗L,−) is exact and thus P⊗L is projective.6

  3. 3.

    The first thing we want to do is to find projectives mapping onto L⁡(λ). For large n, λ+n⁢ρ is dominant. This implies that M⁡(λ+n⁢ρ) is projective, but then M⁡(λ+r⁢ρ)⊗L⁡(n⁢ρ) is projective.

    In fact, there exists a surjection M⁡(λ+n⁢ρ)↠L⁡(n⁢ρ). To see this, if M is a 𝒰⁢𝔤-module and L is a 𝒰⁢𝔟-module, then

    This is known as the tensor identity and is apparently not obvious unless you have the arrogance level of a certain Chinese mathematician. Because M⁡(λ+n⁢ρ)=𝒰⁢𝔤⊗𝒰⁢𝔟ℂλ+n⁢ρ, we obtain

    The surjection comes from the fact that the lowest weight of L⁡(n⁢ρ) is −n⁢ρ, so we can kill all of the higher weights.

    The rest of the proof is simply homological algebra. For a general M∈𝒪, because 𝒪 is Artinian, we can induct on the length of M. First consider a short exact sequence

    By assumption, there exists a surjection P↠N, and this morphism lifts to P→M. If P does not surject onto M, then Im(P→M) cannot intersect L⁡(λ) (otherwise it would contain all of L⁡(λ) and thus surject onto M). This implies that Im(P→M)≅N, which splits the exact sequence. □

By standard homological algebra, because 𝒪 is Artinian and has enough projectives, then 𝒪 has projective covers (i.e. unique minimal projectives surjecting onto M). If we define P⁡(λ) to be the projective cover of L⁡(λ), the P⁡(λ) are precisely the indecomposable projectives. Therefore every projective is a direct sum of P⁡(λ).

Theorem 4.2.7.
  1. 1.

    P⁡(λ) has a standard filtration, which is a filtration with subquotients that are Verma modules.

  2. 2.

    (BGG reciprocity) The multiplicity of M⁡(μ) in the composition series for P⁡(λ) is given by

5 Fan (Oct 27): more on tHe strUcture, simplicity criterioN, ANd embedding of verma modules

We will be discussing the following results during this lecture:

Theorem 5.0.1.

Let λ∈𝔥∗. Then Mλ=Lλ if and only if λ is ρ-antidominant (this is the notion of antidominance that we discussed last time).

Theorem 5.0.2.

For any λ∈𝔥∗, if α∈Φ+ has sα∘λ≤λ, then Msα∘λ↪Mλ.

5.1 Basic facts

Recall that the socle of a module M is defined to be the direct sum of its simple submodules. A fact from ring theory1 is that if R is left noetherian and has no right zero-divisors, then any two left ideals intersect nontrivially.

Proposition 5.1.1.

for all λ∈𝔥∗, Mλ has a unique simple submodule.

Proof.

Recall that Mλ≅𝒰⁢𝔫−, but then simple submodules of Mλ are ideals in 𝒰⁢𝔫−, which intersect nontrivially. □

Theorem 5.1.2.

For any λ,μ∈𝔥∗,

  1. 1.

    Any nonzero φ:Mμ→Mλ is an injection.

  2. 2.

    dimHom𝒪⁡(Mμ,Mλ)≤1.

  3. 3.

    If Lμ⊆Mλ is the unique simple submodule, then Lμ=Mμ.

Proof.
  1. 1.

    φ is determined by the image φ⁡(v+μ)=y⋅v+λ for some y∈𝒰⁢𝔫− of the highest weight vector. But then φ⁡(y′⁢v+μ)=y′⋅y⋅v+λ, and this follows from the fact that 𝒰⁢𝔫− has no zero divisors.

  2. 2.

    If φ1,φ2 are morphisms and L is the unique simple submodule of Mμ, then φ1⁢(L),φ2⁢(L) are both simple and thus isomorphic, so there exists c such that (φ1−c⁢φ2)⁢(L)=0, and so φ−c⁢φ2=0 because it is not injective.

  3. 3.

    Suppose Lμ is the unique simple submodule of Mλ. Then we have the sequence

    and so Mμ must inject into Mλ and thus Mμ=Lμ. □

Proposition 5.1.3.

For λ∈Λ+−ρ, if w=sn⁢⋯⁢s1 is a reduced word expression, then there exists a sequence of embeddings

Proof.

We will induct on ℓ⁡(w). Choose some 0<k<n. Recall that ℓ⁡(sα⁢w)>ℓ⁡(w) if and only if w−1⁢α∈Φ+. Note that ℓ⁡(sk+1⁢⋯⁢s1)>ℓ⁡(sk⁢⋯⁢s1). Thus if w′=sk⁢⋯⁢s1, then w′−1⁢σk+1>0. We now compute

Then we know Msk+1⁢⋯⁢s1∘λ⊆Msk⁢⋯⁢s1∘λ. □

5.2 Proof of first theorem

If λ is ρ-antidominant, we will prove that Mλ=Lλ. Recall that antidominance is equivalent to λ≤w∘λ for all w∈Wλ. But the Jordan-Holder factors of Mλ look like Mμ for μ≤λ with μ is linked to λ. This is the same as μ∈Wλ∘λ. But then the only weight that can appear is λ, so Mλ=Lλ.

Now we will prove that if Mλ=Lλ, then λ is antidominant. This is hard, so we will assume for now that λ∈Λ. Because Mλ is simple, then Nλ=0. Now assume that for some σ∈Σ, σ∗⁢(λ+ρ)>0. Then

and so Msσ∘λ↪Mλ. This gives a contradiction.

We now finish the proof in the general case. Assume there exists α∈Φ+ such that α∗⁢(λ+ρ)∈ℤ+. Then sα∘λ=λ−α∗⁢(λ+ρ)⁢α<λ. By the second theorem, we have an injection Msα∘λ↪Mλ, and thus Mλ is simple.

5.3 More basic facts

Proposition 5.3.1.

Let λ,μ∈𝔥∗ and σ∈Σ. If Msσ∘μ⊆Mμ⊆Mλ, then:

  • •

    If σ∗⁢(λ+ρ)≤0, then Mλ⊆Msσ∘λ.

  • •

    If σ∗⁢(λ+ρ)>0, then Msσ∘μ⊆Msσ∘λ⊊Mλ.

Proof.

In the first case, then we know that

and so Mλ↪Msσ∘λ.

In the second case, let v+λ,v+μ be maximal vectors of Mλ,Mμ. Then we know that

and because Msσ∘μ⊆Mμ, fσσ∗⁢(μ+ρ)⁢v+μ∈Msσ∘μ is a maximal vector. Also because Mμ⊆Mλ, there exists y∈𝒰⁢𝔫− such that y⋅v+λ=v+μ.

It is a fact that if 𝔫 is a nilpotent Lie algebra and ξ∈𝔫,x∈𝒰⁢𝔫, for any n∈ℤ+ there exists k∈ℤ such that ξk⁢x∈𝒰⁢𝔫⁢⟨ξn⟩. Applying this fact to fσ∈𝔫−, y∈𝒰⁢𝔫−, and n=σλ+ρ∈ℤ+, there exists k such that

But then

If k<σ∗⁢(μ+ρ), then we know Msσ∘μ⊆Msσ∘λ. On the other hand, if k>σ∗⁢(μ+ρ), we have

but this is equal to eσ⁢fσk⁢v+μ, so we are done. □

5.4 Proof of second theorem

We will prove this result in the integral case. Assume that λ∈Λ and μ≔sα∘λ≤λ. If μ∈Λ, then there exists w such that μ′≔w−1∘μ∈Λ+−ρ. Thus there exists w=sn⁢⋯⁢s1 such that if we write μn=μ,μn−1=sn−1⁢⋯⁢s1∘μ′,…, then

If we define λ′=w−1∘λ and λn,…,λ1 analogously and assume that μ<λ and thus μk≠λk, then we have

where βk=sk+1⁢⋯⁢sn⁢(α) so that

Note that this is the difference of a multiple of βk and a sum of positive roots.

Continuing, note that λk,μk are linked to λ. These satisfy μ′=μ0≥⋯≥μn=μ, μ<λ, and μ′>λ′. Thus there exists k such that λk<μk and λk>μk+1. But then

and in fact we obtain a multiple of βk. Now there exists Mμk+1⊆Mλk+1 and Mμk2⊆Mλk+2, so Mμn=Mμ⊆Mλn=Mλ. To construct this, note that

This implies that sk+1 flips βk and thus βk=σk+1. Also, βk+1=−σk+1. Therefore,

But then the coefficient of σk+1 in λk+1−sk+1∘λ is positive, and thus Mμk+1≤Mλk+1.

We then have Mμk+2⊆Mμk+1⊆Mλk+1, and therefore either Mλk+1⊆Msk+2∘λk+1=Mλk+2 or Mμk+2⊆Mλk+2⊆Mλk+1, and so we are done.

We sketch the rest of the proof. Fix α∈Φ+ and n>0 is an integer. Then define

Then define H={λ∣α∗⁢(λ+ρ)=n}. We know that Λ∩H⊆X~⊆H. But now Λ∩H is Zariski-dense in H, so we need to prove that X~ is a closed subscheme of H. But now Mλ−α↪M exactly when there exists y∈𝒰⁢𝔫− such that y∈(𝒰⁢𝔫−)−n⁢α. But now 𝔫+⋅y⋅v+λ=0, and this is equivalent to

and this is equivalent to

Finally we have [ei,y]=yi+yi′⁢hi. If we write y=f1−⁢⋯⁢fℓ−, we have a map

by an argument about passing from ei⁢f1⁢⋯⁢fℓ to f1⁢⋯⁢fℓ⁢ei. But now our chain of equivalences continues to (yi+λ⁡(hi)⁢yi′)=0 for all i. But now X~ is given by the rank of the map being lower than usual, so it is a closed subscheme.

6 Patrick (Nov 03): CAts, bOndage, and why you can’T do representAtioN theory without Geometry

Note: these are the speaker’s notes.

6.1 Introduction

Our goal is to prove the following theorem, due to Bernstein-Gelfand-Gelfand.

Theorem 6.1.1 (BGG).

Let λ,μ∈𝔥∗.

  1. 1.

    If μ is strongly linked to λ, then M⁡(μ)↪M⁡(λ). This implies [M⁡(λ):L⁡(μ)]≠0.

  2. 2.

    If [M⁡(λ):L⁡(μ)]≠0, then μ is strongly linked to λ.

We first need to define what it means for weights to be strongly linked. Then we will state a result of Jantzen that we will use to prove this theorem. Next, we will do some linear algebra before proving the result of Jantzen. Finally, we will use Jantzen to prove BGG. Note that this is a different order than in which Humphreys does things.

Example 6.1.2 (Cursed example).

Consider 𝔤=𝔰⁢𝔩4. Number the simple roots and fundamental weights as usual. This means that the simple roots and fundamental weights are

Write c1⁢ϖ1+c2⁢ϖ2+c3⁢ϖ3 as (c1,c2,c3). This means that

Consider λ=(1,−2,−1). Then we note that λ+ρ=α1 but if w=s2⁢s3⁢s2⁢s1⁢s2, then w⁡(λ+ρ)=−α3. But this implies that

and therefore w∘λ<λ. However, apparently there is no embedding of M⁡(w∘λ) inside M⁡(λ).

The reason for this is actually that if you inspect the Bruhat order for W=S4. The weight in Λ+−ρ linked to λ is μ=(0,−1,0). Here, we note that λ=s2⁢s3∘μ while w∘λ=s3⁢s2⁢s1∘μ. However, s3⁢s2⁢s1 and s2⁢s3 are incomparable.

6.2 Strong linkage

The first thing we need to do is to define what it means for two weights μ,λ to be strongly linked.

Definition 6.2.1.

Let λ,μ∈𝔥∗. Write μ↑λ if μ=λ of there exists a root α>0 such that μ=sα∘λ<λ. Note that this is equivalent to ⟨λ+ρ,α∨⟩∈ℤ>0. More generally, if μ=λ or there exists α1,…,αr∈Φ+ such that

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

Note that if we repeatedly apply Theorem 4.6 (the second big theorem in Fan’s lecture), we immediately obtain the first part of the theorem. The second part of the proof is more involved.

Before we proceed, we need to discuss the Bruhat order on the Weyl group (in fact, this can be defined for any Coxeter group, but we will not need this). Let λ be a regular integral weight which is antidominant. Then suppose that α>0 is a positive root and that sα∘(w∘λ)<w∘λ for some w∈W. By the second main theorem of the previous lecture, we know that

But we know that this is equivalent to

Because λ is antidominant, we know w−1⁢α<0. This tells us that if w′=sα⁢w, then

All of this is reversible, so by definition we see that w′∘λ<w∘λ if and only if w′<w. This implies, after we prove the main theorem, that

Corollary 6.2.2.

Let λ be a regular antidominant weight and let w,w′∈W. Then [M⁡(w∘λ):L⁡(w′∘λ)]≠0 if and only if w′≤w.

6.3 Jantzen filtration

Proposition 6.3.1 (Jantzen).

Let λ∈𝔥∗. Then M⁡(λ) has a filtration by submodules

such that M⁢(λ)i≫0 for i large enough and the following conditions hold:

  1. 1.

    Each nonzero quotient M⁢(λ)i/M⁢(λ)i+1 has a nondegenerate contravariant form in the sense of 3.14.

  2. 2.

    M⁢(λ)1=N⁡(λ) is the unique maximal submodule of M⁡(λ).

  3. 3.

    At the level of formal characters, we have

This filtration is called the Jantzen filtration and the formal character formula is called the Jantzen sum formula. We will write

Also note that the summation on the right side of the sum formula is over a set of positive roots which we will call Φλ+. In fact, if M⁢(λ)n≠0 but M⁢(λ)n+1=0, the sum formula tells us that n=|Φλ+|.

Example 6.3.2.

Let λ be regular, antidominant, and integral. Then we know that Φw∘λ+ is the set of all α>0 satisfying sα∘(w∘λ)<w∘λ. But this is the same as

but of course this is equivalent to w−1⁢α∨<0 by defintion of λ being integral and antidominant. Therefore in fact here n=ℓ⁡(w).

Here are some reasonable questions about these things:

  1. 1.

    Is the Jantzen filtration unique relative to properties such as the sum formula and existence of nondegenerate contravariant forms on the quotients?

  2. 2.

    What are the composition factor multiplicities in each filtration layer M⁢(λ)i?

  3. 3.

    Are the filtration layers semisimple? If so, does the filtration coincide with one of the standard module filtrations with semisimple quotients?

  4. 4.

    How does the Jantzen filtration behave with respect to M⁡(μ)↪M⁡(λ)?

A more precise formulation of the last question is called the Jantzen conjecture, and this can only be proved (as of whenever the book was published) using geometric tools like Beilinson-Bernstein localization. For completeness, it is stated here:

Conjecture 6.3.3 (Jantzen).

Suppose that μ↑λ and set r≔|Φλ+|−|Φμ+|. Then M⁡(μ)⊂M⁢(λ)i if i≥r while M⁡(μ)∩M⁢(λ)i=M⁢(μ)i−r if i≥r.

Example 6.3.4.

Let 𝔤=𝔰⁢𝔩3. Suppose that λ is regular, antidominant, and integral. We will write down the Jantzen filtration for M⁡(w∘λ) and then deduce that all composition factors have multiplicity 1. Let α,β be the simple roots and write w=sα⁢sβ. The sum formula tells us that

But this implies that the Jantzen filtration is

This implies that there are four composition factors each with multiplicity 1.

6.4 Proofs of results

We will now prove Jantzen’s theorem and apply it to prove the main theorem. I assume that everyone is familiar with the basic theory of finitely generated modules over a principal ideal domain.1

Let A be a principal ideal domain and p∈A be a prime element. Suppose that M is a free A-module of rank r with a nondegenerate symmetric bilinear form (−,−). Let D be the determinant of the bilinear form – this is well-defined up to a unit. Let M¯=M/p⁢M, which is a vector space over A¯≔A/p⁢A. For n∈ℤ, define

For notation, we will write M∗=Hom⁡(M,A) and M∨⊆M∗ for the image of M under the bilinear form. We will write e1,…,er for a basis of M and f1,…,fr for the dual basis.

Lemma 6.4.1.
  1. 1.

    We have the identity

  2. 2.

    For all n, the modified bilinear form (−,−)n=p−n⁢(−,−) induces a nondegenerate form on M⁡(n)¯/M⁡(n+1)¯.

Proof.

Write f=∑aj⁢fj for some f∈M. Then writing (ei,fj)=fj⁢δi⁢j, we know that (ei,f)=ai⁢di. On the other hand, we know that f∈M⁡(n) if and only if

for all i. But if we write ni≔vp⁢(di), then we know that M⁡(n) is spanned by fi for all i where n≤ni together with pn−ii⁢fi for the i such that n>ni. But this implies that

and in particular M⁡(n)¯=0 for n≫0.2 This means the sum is defined, and therefore we have

For the second part, it ic clear that this form takes values in A, but we need to check that there is an induced form on M⁡(n)¯. But here, if e∈p⁢M∩M⁡(n), we see that

But if f∈M⁡(n+1), the same argument gives us (f,M⁡(n))n⊂p⁢A. This gives us a form on M⁡(n)¯/M⁡(n+1)¯. Finally we know that if n≤nj, then (ei,fj)=δi⁢j⁢p−n⁢dj≠0, and thus we conclude that the form is nondegenerate. □

Before we continue, we will state some facts about contravariant forms that were not covered by Kevin.

Definition 6.4.2.

Let (−,−) be a symmetric bilinear form on a module M. Then (−,−) is contravariant if (u⋅v,w)=(v,τ⁡(u)⋅w) for all u∈𝒰⁢𝔤 and v,w∈M.

It is not even clear that these exist, but here are some properties:

  1. 1.

    If M has a contravariant form, then Mλ⟂Mμ for λ≠μ.

  2. 2.

    If M is a highest weight module, then it has a unique (up to scalars) nonzero contravariant form.

  3. 3.

    If N⊂M is a submodule, then N⟂ is also a submodule.

  4. 4.

    Any contravariant form on a highest weight module M must induce the zero form on its maximal submodule. Moreover, the maximal submodule is the radical of the form and the form is nondegenerate if and only if M=L⁡(λ) is simple.

Proof of Jantzen.

The reason we did all of that linear algebra above is because we will extend to the field K≔ℂ⁡(T). Also write A=ℂ⁡[T]. Then apparently we can adapt the entire theory of Lie algebras to 𝔤K≔𝔤⊗ℂK and 𝔤A≔𝔤⊗ℂA. I will not check this, and you should not ask me to do so. We will use the lemma to construct filtrations here and then kill T to obtain the Jantzen filtration.

Let λ∈𝔥∗ be a weight and write λT≔λ+T⁢ρ∈𝔥K∗. Clearly λT is antidominant, and thus M⁡(λT) is simple. But now we know that the contravariant form on M⁡(λT) is nondegenerate by Theorem 3.15 in Humphreys. Here, recall that the contravariant form satisfies for all v,w∈M and y∈𝒰⁢𝔤 the identity

Now recall that 𝒰⁢𝔤A=A⊗ℂ𝒰⁢𝔤⊂K⊗ℂ𝒰⁢𝔤=𝒰⁢𝔤K is an A-form. This induces an A-form M⁢(λT)A⊂M⁡(λT). Each weight space is a free A-module of finite rank and the contravariant form is nondegenerate after restricting to each weight space. If Γ is the space of positive linear combinations of simple roots, then write MλT−ν for the A-form of M⁢(λT)λT−ν. Set

These form a decreasing filtration of M⁢(λT)A.

Now we can quotient by the ideal (T)⊂ℂ⁡[T] and we obtain a decreasing filtration M⁢(λ)i on M⁡(λ)≅M⁢(λT)A/T⁢M⁢(λT)A. By the lemma, the quotients M⁢(λ)i/M⁢(λ)i+1 have induced nondegenerate contravariant forms. Also, we know that the filtration of each individual weight space eventually ends at 0, and thus for large enough i we have M⁢(λ)i=0 because only finitely many weights are linked to λ.

Next, we note that M⁡(λ)/M⁢(λ)1 is a highest weight module with a nondegenerate contravariant form and therefore is simple. Because M⁡(λ) has a unique simple quotient, we see that M⁢(λ)1=N⁡(λ). First, we will express the sum

in terms of the determinants of the contravariant forms on the steps of the filtration. Here, by some magic involving Shapovalov elements that was discovered by Jantzen and Shapovalov independently while working on their PhD theses, the determinant of the contravariant form on the λT−ν weight space of M⁡(λT) is given by

where 𝒫 is the Kostant partition function. Note that the Kostant partition function gives the number of ways to write a weight as a sum of positive roots with non-negative coefficients. This is actually unique only up to a unit, but this means that vT⁢(Dν⁢(λT)) is well-defined. But now we know that

This is not a multiple of T unless ⟨λ+ρ,α∨⟩=r. But this means that α∈Φλ+ and vT for this term is 1. But this implies that for fixed ν and α, the contribution is given by

Finally, we compute formally and obtain the result

This comes from a variable change ν↦ν+⟨λ+ρ,α∨⟩⁢α in ν. But now we precisely have

and therefore if we fix α the sum becomes

as desired. □

Remark 6.4.3.

Remember that ρ disappears when we kill T. Remember we only needed the fact that ⟨ρ,α∨⟩≠0. We may consider what ahppens if we replace ρ by another weight, but a definitive answer to this is only provided by Belinson-Bernstein using a geometric approach.

Proof of BGG.

We will finally prove the BGG theorem. We will induct on the number of linked weights μ≤λ. If λ is minimal in its linkage class, then we know M⁡(λ)=L⁡(λ), so we are done. Now suppose that μ<λ and [M⁡(λ):L⁡(μ)]>0. But this implies that

But now the sum formula for the Jantzen filtration gives us

for some α∈Φλ+. By the inductive hypothesis, there exists α1,…,αr∈Φ+ such that

Because sα∘λ<λ, we are done. □

7 Fan (Nov 10): why You shoUld care about the bgg resolutioN, aNd more on homologicAl consideratioNs in category o

7.1 The BGG resolution

Let λ∈Λ+ be a dominant (in the old sense) weight. A BGG resolution is something of the form

Later, we will see that there is something called the BGG resolution, but we will probably not prove it.

Theorem 7.1.1 (Weak BGG).

There exists a resolution

such that the standard filtration of Dkλ consists of the weights {w∘λ∣ℓ⁡(w)=k}.

A fact is that For all subalgebras 𝔤⊇𝔞 there exists a resolution of the trivial resolution

where the differentials are given by

Lemma 7.1.2.

For a finite dimensional representation V of 𝔟 which is 𝔥-semisimple, the standard filtration of 𝒰⁢𝔤⊗𝒰⁢𝔞V contains precisely the weights of V, respecting multiplicity.

Recall the following theorem from your first course on Lie algebras:

Theorem 7.1.3 (Lie).

For V a finite dimensional representation of a solvable 𝔟, there exists

such that dimVi/Vi+1=1, and there exists a nonzero v∈V and λ:𝔟→k such that ξ⁢v=λ⁡(ξ)⁢v for all ξ∈𝔟.

Proof of lemma.

Using the previous theorem, we have a filtration

and the quotients are 𝒰⁢𝔤⊗Vi/Vi+1=Mλ, where λ is a weight of V. □

Lemma 7.1.4.

For ϑ a central character and M having a standard filtration, we have

Proof.

Choose a standard filtration

such that Mi/Mi−1=Mλi. Then we have

a standard filtration of Mϑ, and also we note that Miϑ/Mi−1ϑ=(Mi/Mi−−1)ϑ=Mλiϑ. □

Proof of weak BGG for λ=0.

Consider the resolution

Taking central characters with ϑ0, we have a new resolution

But we also know that

Now recall that w⁢ρ−ρ=−∑α∈Φ+w−1⁢α∈Φiα. Also we will use the fact that if S⊆Φ+ has

then S={α∈Φ+∣w−1⁢α∈Φ−}≔Φw. This tells us that

To prove the general case, we will use the translation functor Tμλ. This is given by projecting onto ϑμ, tensoring by Lλ−μ, and then including into ϑλ.

Proof in general.

Let Bk=(𝒰⁢𝔤⊗𝔟𝔤/𝔟)θ0. Then recall we have a resolution

Tensoring by Lλ, which is finite-dimensional, we obtain

Finally, projecting to ϑλ, we have

Now if we consider Std⁡(Bk⊗Lλ), we have the filtration

Also we know that Bk,i⊗Lλ/Bk,i−1⊗Lλ=Mwi∘0⊗Lλ. We will use the fact that if 0=M0⊂⋯⊂Mn=M, then Std⁡M=⨆Std⁡Mi/Mi−1. Therefore

Projecting onto the central character ϑλ, we have

Now we know that w∘0+μ=u∘λ for some u∈W and this means that

and thus λ=u−1⁢(μ+w⁢ρ−u⁢ρ). Because ρ≥u−1⁢w⁢ρ and u−1⁢μ≤λ, we know that

Using the previous equality for λ, we see that ρ=u−1⁢w⁢ρ, which means that u=w and μ=u⁢λ. □

Theorem 7.1.5.

Let λ,μ∈𝔥∗.

  1. 1.

    If Ext𝒪1⁡(Mμ,Mλ)≠0, then μ↑λ and μ≠λ.

  2. 2.

    If λ∈Λ+ and Ext𝒪⁡(Mw∘λ,Mu∘λ), then u<w.

Proof of Strong BGG.

We induct on |Std⁡Dkλ|. The base case is length 2. Here, we have

Because u,w have the same length, the Ext1 vanishes, and so the exact sequence must split. In the inductive step, we can pull the direct sum out of the Ext functor, and each individual Ext vanishes, so we are done. □

This argument in fact proves that weak BGG is strong BGG. Note that this was not available to BGG in their original paper.

Proof of theorem.

By proposition 3.1 of Humphreys, we know μ≠λ. Thus there exists a non-split extension

We know that there exists a lift φ:Pμ→M. If Imφ∩Mλ=0, then it must be Mμ and so Mμ∩Mλ=0, and so it splits. This means that Imφ∩Mλ≠0. Now take the standard filtration

We know that Std⁡Pμ contains weights at least μ with μ appearing exactly once. But then by BGG reciprocity, we know that

and this is equivalent to μ↑μi.

Because Imφ∩Mλ≠0. We can consider the images φ⁡(Pμ)⊃⋯⊃φ⁡(P0)=0. Choose the smallest i such that φ⁡(Pi)∩Mλ≠0. Thus Mλ has a submodule isomorphic to the image of φ⁡(Pi/Pi−1)=φ⁡(Mμi) If we quotient out more stuff, we get a copy of Lμi inside Mλ. But now [Mλ:Lμi]>0, and thus μi↑λ. By definition, μ↑λ.

Now we will state another fact. Let λ be integral, dot-regular, and ρ-antidominant. Then u∘λ↑w∘λ if and only if u≤w. □

7.2 Why you should care

We will now explain why you should care about the BGG resolution.

Theorem 7.2.1 (Bott, Kostant).

Let λ∈Λ+. Then

as 𝔥-modules, where w0 is the longest element of W.

Proof.

Recall that Hk⁢(𝔫−,Lλ)=Rk⁢Lλ𝔫−=Rk⁢Homn−⁡(ℂ,Lλ)=Rk⁢Hom𝔫−⁡(Lλ∗,ℂ). But we know that Lλ∗=L−w0⁢λ. We now consider the BGG resolution of L−w0⁢λ, and so we have

Note that Hom𝔫−⁡(M,ℂ)=Homℂ⁡(M/𝔫−⁢M,ℂ)=(M/𝔫−⁢M)∗. Therefore

□

Apparently there are two facts that if w1,w4 differ by length 2, then either there are two w2,w3 between them or there is nothing. Also, the BGG resolution is apparently unique.

Next, for λ a ρ-antidominant weight, we can define 𝒪λ={M∣μ∈JH⁡M⇒μ∈Wλ∘λ}.

Theorem 7.2.2.

The projective dimension of Mw∘λ is ℓ⁡(w), the projective dimension of Lw∘λ is 2⁢ℓ−ℓ⁡(w), and the homological dimension of 𝒪ϑλ is 2⁢ℓ.

We will now state some more facts. First, if M,N have standard filtrations, then Ext𝒪>0⁡(M,N∨)=0.

Theorem 7.2.3.

The following are equivalent:

  1. 1.

    M has a standard filtration;

  2. 2.

    Ext𝒪>0⁡(M,Mλ∨)=0 for all λ∈𝔥∗;

  3. 3.

    Ext𝒪1⁡(M,Mλ∨)=0 for all λ∈𝔥∗;

  4. 4.

    H0⁢(𝔫+,M∨)=0 (H1⁢(𝔫+,M∨)=0);

  5. 5.

    H0⁢(𝔫−,M)=0 (H1⁢(𝔫−,M)=0).

Proposition 7.2.4.

We have the identity

8 Fan (Nov 17): VERmas and sImpleS undEr the translation functor, facets, chambers, anD walls

The reason we should care about translation functors is the following:1 for any antidominant λ, Pλ=Pλ† is self dual. Also,

Before we continue, we will do some more homological stuff. If we consider Hom𝔟⁡(ℂλ,−):𝒪→𝖵𝖾𝖼𝗍, this is actually the same as Hom𝔤⁡(Mλ,−). Of course, this does not remain true in the derived category, but we also have Hom𝔟⁡(ℂλ,−)=((−)𝔫+)λ. If we consider

we need to be careful about how we derive things. Recall that Lie algebra cohomology is computed by the Koszul complex

Note that 𝒰⁢𝔫 is most definitely not 𝔫-finite, but somehow Ext𝒪 can be computed in the category of 𝔤-modules that are 𝔥-semisimple and locally 𝔥-finite. Now we have

8.1 The actual content of this lecture

Definition 8.1.1.

Let λ,μ∈𝔥∗ such that λ≡μ(modΛ). Then λ−μ∈Λ. Denote by (λ−μ)+∈Λ+ the representative of the orbit of λ−μ under the undotted action. Then the translation functor Tμλ is given by

Proposition 8.1.2.

Tμλ is exact, preserves projectives, commutes with taking duals, and Tμλ⁢Mμ has a standard filtration.

Proof.

Exactness is obvious. Preserving projectives and commuting with duals are both from Kevin’s talk, and Tμλ⁢Mμ having a standard filtration follows from Mμ⊗L(λ−μ)+ having a standard filtration follows from

and the only M that works is Mλ. □

Proposition 8.1.3.

For λ,μ compatible, the functors Tμλ,Tλμ are both left and right adjoints.

From now on, all chambers will be discussed with respect to −ρ.

Definition 8.1.4.

A facet F is a subset of E determined by Φ+=ΦF+⊔ΦF⁢0⊔ΦF−. F is given by

We can define F¯ by making our inequalities weak. We can also define F̂ by allowing ≤0 and  ̂ F by allowing ≥0.

We will assume the following facts without proof:

  1. 1.

    {α∣±α∈ΦF⁢0} is a root system. It has the Weyl group given by the subgroup of W fixing F pointwise.

  2. 2.

    If there exist λ,w such that w∘λ∈λ, then w∘λ<λ′ for all λ′∈F.

For any λ∈𝔥∗, let Eλ⊆E be spanned by Φλ over ℝ. Let λ♮∈Eλ be such that for all α∈Φλ, α∗⁢(λ)=α∗⁢(λ♮).

Definition 8.1.5.

Define Wλ∘=StabWλ⁡λ.

We will use the following facts without proof:

  1. 1.

    For all w∈Wλ, λ−w∘λ=λ♮−w∘λ♮ and (w∘λ)♮=w∘λ♮.

  2. 2.

    We have StabWλ⁡λ=StabWλ⁡λ♮=StabW⁡λ.

  3. 3.

    If Φλ=Φμ=Φλ+μ, then (λ+μ)♮=λ♮+μ♮.

Lemma 8.1.6.

For λ,μ∈𝔥∗ compatible with μ♮∈F and λ♮∈F¯, for all ν≠(λ−μ)+∈Wt⁡L(λ−μ)+, μ+ν∉Wλ∘λ.

Theorem 8.1.7.

Let λ,μ be antidominant and compatible. For μ♮∈F⊆Eμ and λ∈F¯,

for all w∈Wλ=Wμ.

Proof.

We consider (Mw∘μϑμ⊗L(λ−μ)+)ϑλ. The standard filtration of this is given by

Also, (Tμλ⁢Mw∘μ:Mw∘λ)=1 because w∘λ=w∘μ−w⁡(λ−μ). But (w∘μ)♮=w∘μ♮∈w∘F. Also w∘λ∈w∘F¯. By the lemma, for any ν≠(λ−μ)+∈Wt⁡L(λ−μ)+, μ+ν∉Wλ∘λ. But now only w∘λ can appear in the standard filtration, and so we are done. □

Corollary 8.1.8.

For the setting of the theorem, if M∈𝒪μ has a standard filtration, then Tμλ⁢M has a standard filtration.2

Proof.

Induct on the size of the standard filtration. Consider

Applying the translation functor, we have

By the inductive hypothesis, we are done. □

Proposition 8.1.9.

In the same setting, we have

Proof.

Consider Mw∘μ→Lw∘λ→0. Applying the translation functor, we have

If Tμλ⁢Lw∘μ is nonzero, it is a highest weight module. But now we have

and this becomes

This gives a morphism Mw∘λ→Tμλ⁢Lw∘μ→Mw∘λ†. But now recall that dimHom⁡(Mμ,Mλ†)=δμ⁢λ, and therefore, Lw∘λ≅Tμλ⁢Lw∘μ. □

Theorem 8.1.10.

Consider the same setting as before. Then Tμλ⁢Lw∘μ=Lw∘λ if and only if w∘λ♮∈w∘F̂.

Proof.

We know that Tμλ is exact, so it does not increase the length of the Jordan-Hölder filtration. Thus it brings a unique Lu⁢w∘μ to Lw∘λ. But Tμλ⁢Lu⁢w∘μ=Lu⁢w∘λ=Lw∘λ, and thus u⁢w∘λ=w∘λ.

Now let λ♮∈G be a facet in Eλ. Suppose that w∘λ♮∈w∘F̂. For all α∈w⁢ΦG⁢0, w−1⁢α∈ΦG⁢0. Thus (w−1⁢α)∗⁢(λ♮+ρ)=0 and α∗⁢(w∘λ♮+ρ)=0. But now α∈Φ(w∘F)⁢0 or Φ(w∘F)−, and we know that w∘μ♮∈F, so α∗⁢(w∘μ♮+ρ)≤0. Thus

for all α∈w⁢ΦG⁢0. Because Ww∘λ∘ is the Weyl group of {α∣±α∈Φ(w∘G)⁢0}, u⁢w∘μ♮≥w∘μ♮ for all μ∈Ww∘λ0. Thus u⁢w∘μ≥w∘μ, so u⁢w∘μ=w∘μ. Therefore Tμλ brings Lw∘μ to Lw∘λ.

If w∘λ♮∉w∘F̂, we know that w∘λ♮∈

 ̂

w∘F
. But now because w∘μ♮∈w∘F, α∈Φ(w∘F)+ has α∗⁢(w∘μ♮+ρ)>0. But now sα⁢w∘μ♮<w∘μ♮, and thus sα⁢w∘μ<w∘μ. This implies that there exists Msα⁢w∘μ↪MP⁢w∘μ, and so after the translation functor, we have

except sα⁢w∘λ=w∘λ. □

9 Kevin (Dec 01): Kazhdan-Lusztig theory

So far, we have cared about the Verma modules Mλ and the simple modules Lλ in category 𝒪. Recall that we have two bases [Mλ],[Lλ] of K⁡(𝒪). We would like to know the change of basis matrix. There is an answer when λ is regular and integral. If λ,λ′ are regular and integral, we know that 𝒪λ≅𝒪λ′ via the translation functors. Therefore, we only need to consider the principal block 𝒪0. The answer in this case is the following:

Conjecture 9.0.1 (Kazhdan-Lusztig).

For all w∈W, let Mw,Lw have highest weight −w⁡(ρ)−ρ. Then we have

where εw=(−1)ℓ⁡(w). Also, we have

where w0 is the longest element of W. The Py,w⁢(q) are called the Kazhdan-Lusztig polynomials (which arise from the study of Hecke algebras) and are only nonzero for y≤w.

This result was proved by Beilinson-Bernstein and by Brylinski-Kashiwara using the theory of D-modules.

9.1 Combinatorics

Definition 9.1.1.

Let ℋ~ be the associative ℤ⁡[q]-algebra with basis Tw for all w∈W subject to the following relations:

  • •

    Tw⁢Tw′=Tw⁢w′ if ℓ⁡(w⁢w′)=ℓ⁡(w)+ℓ⁡(w′);

  • •

    (Ts+1)⁢(Ts−q)=0 if s is a simple reflection.

The Hecke algebra is defined to be ℋ≔ℋ~⊗ℤ⁡[q]ℤ⁡[q±1/2]. This has a duality operation defined by

and q1/2¯=q−1/2.

Apparently this comes from number theory,1 and you are supposed to specialize q to a prime power so that it has number-theoretic meaning. Here are some good properties:

  • •

    If we set q=1, ℋ specializes to ℤ⁡[W].

  • •

    If we set q to be a prime power, then ℋ⊗ℤ⁡[q±1/2]ℂ is the algebra of intertwining operators on the space of functions on the flag variety of G⁡(𝔽q).

The Kazhdan-Lusztig polynomials Py,w⁢(q) arise from the following theorem as a change of basis:

Theorem 9.1.2 (Kazhdan-Lusztig).

For any w∈W, there exists a unique cw∈ℋ such that c¯w=cw and

The Py,w are polynomials in q of degree at most 12⁢(ℓ⁡(w)−ℓ⁡(y)−1) with Pw,w=1.

Proof.

First define Rx,y∈ℤ⁡[q1/2,q−1/2] for all x,y∈W by

We can check that

The relations and the fact that R1,1=1 tell us that Rx,y≠0 if and only if x≤y. Also, Rx,y is a polynomial of q of degree ℓ⁡(y)−ℓ⁡(x) when x≤y and Rx,x=1.

Now we will prove that if the cw exist, then they are unique. Because c¯w=cw, we have

Comparing coefficients of Tx, we have

After cancelling, we obtain

Note that the first term on the left has only positive powers of q−1/2 and the second term only has positive powers of q1/2. This means that the Px,w are uniquely determined by downward induction on the length of y.

We will now prove the existence of the cw inductively on the length of w. We begin with c1=T1=1 and cs=q−1/2⁢Ts−q1/2. Now assume that cw′ for all w′ with ℓ⁡(w′)<ℓ⁡(w). Write w=s⁢v, where ℓ⁡(w)=ℓ⁡(v)+1. Introduce a relation ≺, where y≺w when y<w and Py,w is a polynomial in q of degree exactly 12⁢(ℓ⁡(w)−ℓ⁡(y)−1). If y≺w, define μ⁡(y,w) to be the coefficient of highest degree of Py,w. Now define

Clearly, c¯w=cw. From the second relation, we obtain

Here, c=1 if s⁢y<y and c=0 if s⁢y>y. We simply check that all terms are polynomials of low enough degree. This is clear except when c=1 and y≺v, in which the two terms qc⁢Py,v and μ⁡(y,v)⁢qy−1/2⁢qv1/2⁢Py,y could be problematic. However, the two terms cancel out, and so we are done. □

9.2 Geometry

Let G be the simply-connected semisimple group over ℂ corresponding to 𝔤. The flag variety ℬ is defined as the set of Borel subgroups of G with the geometric structure coming from an identification with G/B for a fixed Borel B. This is a smooth projective variety. By the Bruhat decomposition, we have G=⨆w∈WB⁢w⁢B, and this gives us a decomposiion

We can check that ℬw≅𝔸ℓ⁡(w). Taking the closures B¯w, we obtain the Schubert varieties. These are singular, so we need to consider the intersection cohomology instead of the usual cohomology.

Suppose we have a variety X of dimension n. Define the intersection complex ICX∈Db⁢(X,ℚ) as follows. First, there exists a nice stratification X=Xn⊇Xn−1⊃⋯⊃X0 such that

  • •

    Each Xi is closed;

  • •

    Xi−Xi−1 is smooth of pure dimension i;

  • •

    The stratification is a Whitney stratification.

Given such a stratification, let Ui=X−Xn−i. We know that U1 is smooth and that

Denote the inclusions Ui⊂Ui+1 by ji. Then we define

This is independent of the choice of Whitney stratification. Of course, we may now define intersection cohomology I⁢H∗⁢(X) as the hypercohomology of this complex. For smooth X, ICX=ℚX, and so intersection cohomology is the same as ordinary cohomology. Also, intersection cohomology satisfies Poincaré duality and hard Lefschetz.

We can relate intersection cohomology to the Kazhdan-Lusztig stuff as follows:

Theorem 9.2.1 (Kazhdan-Lusztig).
  1. 1.

    We have

    where IHy means relative to ℬy⊆ℬ¯w.

  2. 2.

    We have the relation

Corollary 9.2.2.

Py,w has nonegative coefficients.

10 Patrick (Dec 08): Koszul duality for people who aren’t Peter May

Note: these are the speaker’s notes.

Before we begin, we will establish some notation.

Notations 10.0.1.

Let A=⨁Aj be a graded ring. We will denote by A⁢-⁢𝖬𝗈𝖽 the category of left A-modules and A⁢-⁢𝗆𝗈𝖽 the category of graded A-modules. These carry functors Hom,Ext,… and hom,ext,…, respectively. Also, the subcategories of finitely-generated modules will be called A⁢-⁢𝖬𝗈𝖿 and A⁢-⁢𝗆𝗈𝖿. Finally, we will denote the grading shift ⟨n⟩ by (M⁢⟨n⟩)i=Mi−n. Finally, write k≔A0. Here, we will assume that k is a field, but after certain adjustments we may allow k to be a noncommutative semisimple ring.

10.1 Koszul duality

Definition 10.1.1.

A graded ring A=⨁j≥0Aj is a Koszul ring if A0 is semisimple and admits a graded projective resolution

such that Pi is generated by its degree i component Pii.

Example 10.1.2.

The Koszul resolution

shows that S∙⁢V is a Koszul ring.

There is an alternative characterization of Koszul rings.

Definition 10.1.3.

Let M be a graded module. Then M is pure of weight n if and only if M=M−n.

In the cases we will consider, any simple A-module is pure and any pure module is semisimple.

Proposition 10.1.4.

Let A be a graded ring with A0 semisimple. Then the following are equivalent:

  1. 1.

    A is Koszul.

  2. 2.

    For any pure A-modules M,N of weights m,n we have extAi⁡(M,N)=0 whenever i≠m−n.

  3. 3.

    We have extAi⁡(A0,A0⁢⟨n⟩)=0 whenever i≠n.

We will list some more properties of Koszul rings:

Proposition 10.1.5.

If A is a Koszul ring, so is the opposite ring Aop.

Proof.

Define M⊙ by Mi⊙=M−i∗. Then consider a projective resolution

such that Pi=A⁢Pii. But now we note that

is an injective resolution of k in 𝗆𝗈𝖽⁢-⁢A. Now we note that

whenever i≠n. Thus extAopi⁡(k⁢⟨−n⟩,k)=0 whenever i≠n, and so Aop is Koszul. □

Definition 10.1.6.

A graded ring A is quadratic if the natural map

is surjective and the kernel R⊆T⁢A1 is generated by R∩A1⊗A1.

Proposition 10.1.7.

Any Koszul ring is quadratic.

We will now assume that all Ai are finitely generated k-modules (I guess we will call it locally k-finite). We will now define a candidate dual algebra.

Definition 10.1.8.

Let A=Tk⁢V/⟨R⟩ be a locally k-finite quadratic ring over k, where R⊆V⊗V. Then define its quadratic dual A!≔TkV∗/⟨R⟂⟩, where R⟂⊆V∗⊗V∗ is the annihilator of R.

Remark 10.1.9.

Clearly A!!=A.

Proposition 10.1.10.

Let A be a locally finite Koszul ring. Then A! is also Koszul.

Proof.

We will use the fact (without proof) that the Koszul complex of A is actually

with differential given by Hom(Ai+1!,A)→Hom(Ai+1!⊗V,A⊗V)→Hom(Ai!,A). In coordinates if idV=∑vα∗⊗vα, this is given by

Therefore, we have a bigraded space

where the differential has bidegree (1,1) and the cohomology appears in bidegree (0,0). Taking duals, we obtain the space

where the differential has degree (1,1) and the cohomology appears in degree (0,0). Except this is a graded projective resolution of k as an A!-module, so A! is Koszul. □

Theorem 10.1.11.

Let A be a locally finite Koszul ring over k. Then ExtA∙⁡(k,k) is canonically isomorphic to (A!)op.

We will write ExtA∙⁡(k,k)≕E⁡(A). It is easy to see that E⁡(E⁡(A))=A.

We will now give a numerical criterion for a ring to be Koszul. Suppose that the Ai are finite-dimensional for all i and that A0 is a product of copies of a field F. Then define the Hilbert polynomial to be the matrix with entries

Lemma 10.1.12.

Suppose that A is Koszul. Then P(A,t)P(A!,−t)T=1.

Theorem 10.1.13.

Let E=E⁡(A). Then A is Koszul if and only if P⁡(A,t)⁢P⁡(E,−t)=1.

Now we are finally able to define a derived category version of Koszul duality, which is what we really wanted. Write C⁡(A) for the homotopy category of complexes in B⁢-⁢𝗆𝗈𝖽. We will use Mi for the grading in the complex and Mji for the grading in the module Mi. Now define the category C↑⁢(A) to be the full subcategory of C⁡(A) where Mji=0 if i≫0 or i+j≪0. Similarly, define C↓⁢(A) to be the full subcategory of C⁡(A) where Mji=0 if i≪0 or i+j≫0. Define the derived categories D↑⁢(A),D↓⁢(A).

Theorem 10.1.14.

Let A be a locally finite Koszul ring. Then there exists an equivalence of triangulated categories D↓(A)≅D↑(A!).

We will not prove this result, but we will construct a functor. Let M∈C⁡(A). Then consider the bigraded vector space

with the differentials coming from the Koszul complex and from M and given by

where vα∗,vα are as before. Now we consider the total differential d=d′+d′′ and write

In fact, F sends C↓⁢(A) to A↑(A!) and takes acyclic complexes to acyclic complexes, so it induces a derived functor DF:D↓(A)→D↑(A!).

The inverse functor is given as follows. Let N∈C(A!). Then define

This will have anticommuting differentials given by

Then we consider the total differential d=d′+d′′ and write

One can check that G sends C↑(A!) to C↓⁢(A) and that G takes acyclic things to acyclic things. Thus there is a derived functor DG:D↑(A!)→D↓(A).

We will write K≔D⁢F for the Koszul duality functor. Here are some more properties of this functor.

Theorem 10.1.15.

There is a canonical isomorphism K⁡(M⁢⟨n⟩)=(K⁢M)⁢[−n]⁢⟨−n⟩.

Theorem 10.1.16.

Let A be a Koszul ring over k that is a finite-dimensional k-vector space. Also suppose that A! is left Noetherian. Then there are embeddings Db⁢(A⁢-⁢𝗆𝗈𝖿)⊆D↓⁢(A) and Db(A!-𝗆𝗈𝖿)⊆D↑(A!) such that Koszul duality induces an equivalence

10.2 Parabolic-singular duality

Let S⊂W be the set of simple reflections. Then for any subset Sι⊂S, let Wι⊂W be the subgroup generated by Sι, wι be its longest element, and Wι⊂W be the set of longest representatives of the cosets W/Wι.

Let λ∈𝔥∗ be integral and dominant (but possibly singular). Set 𝒪λ⊂𝒪 to be the full subcategory of objects with the same central character is L⁡(λ). Set

Then the simple objects in 𝒪λ are precisely the L⁡(x∘λ) for all x∈Wλ.

Let 𝔟⊂𝔮⊂𝔤 be a parabolic subalgebra. Define 𝒪𝔮⊂𝒪0 to be the subcategory of all objects that are locally 𝔮-finite. Let S𝔮⊂S be the simple reflections corresponding to 𝔮. For all x∈W𝔮 set Lx𝔮≔L⁡(x−1⁢w0∘0)∈𝒪𝔮. These represent the simple objects in 𝒪𝔮. Call their projective covers Px𝔮.

The main theorem of this part of the talk is

Theorem 10.2.1.

Suppose that Sλ=S𝔮. Then there are isomorphisms of finite-dimensional algebras

where the summations are over x∈Wλ=W𝔮. The algebras on the right are both Koszul rings and are Koszul dual to each other.

Note that 𝒪𝔟=𝒪0. In particular, in this case, we obtain the following:

Corollary 10.2.2 (Soergel ’90).

Let L∈𝒪 be the direct sum of all simple modules with the same central character as ℂ=L⁡(0). Let P∈𝒪 be the direct sum of their projective covers. Then there exists an isomorphism

Moreover, Ext𝒪∙⁡(L,L) is a self-dual Koszul ring.

The proof of the theorem is geometric and developing the machinery would take to long, so we will prove the equality of dimensions for the first isomorphism. Using BGG reciprocity and the Kazhdan-Lusztig conjectures, we have

By an argument involving localization (translating the problem to geometry), this is precisely the dimennsion of Ext𝒪𝔮⁡(⨁Lx𝔮,⨁Lx𝔮).

Now write AQ=Ext𝒪∙⁡(⨁Lx𝔮,⨁Lx𝔮) and AQ=Ext𝒪∙⁡(⨁L⁡(x∘λ),⨁L⁡(x∘λ)). Also, write A𝔮=End𝒪⁡(⨁Px𝔮).

Corollary 10.2.3.

There is a ring isomorphism AQ=E⁡(AQ).

Proof.

By the theorem, there is an equivalence 𝒪λ≅AQ⁢-⁢𝖬𝗈𝖿. This is because if 𝒜 is an abelian category whose objects have finite length and P∈𝒜 is a projective generator, then there is an equivalence Hom𝒜⁡(P,−):𝒜→𝖬𝗈𝖿⁢-⁢E, where E=End𝒜⁡P. But now this equivalence identifies L⁡(x∘λ) with AQ0⁢1x, and from this we deduce the desired isomorphism. □

Proposition 10.2.4.

The rings AQ and AQ are Koszul.

Proof.

First, note that 𝒪𝔮=𝖬𝗈𝖿⁢-⁢A𝔮. But now the simples correspond under this equivalence, and therefore

This tells us that AQ=E⁡(A𝔮,op). Once we prove that A𝔮 is Koszul, we are done.

To prove this, consider that 𝒪𝔟≅𝖬𝗈𝖿⁢-⁢AB. But now we know that 𝒪𝔮 consists of objects in 𝒪𝔟 such that if [M:Lx𝔟]≠0, then x∈W𝔮, and therefore 𝒪𝔮≅𝖬𝗈𝖿⁢-⁢(AB/IQ), where IQ is generated by everything in W∖W𝔮. Thus A𝔮≅𝒪𝔮 because they are both endomorphism algebras of the same progenerator. Now restricting modules over AB/IQ to AB induces injections on Ext groups, so we now only need to prove that AB is Koszul.

To prove that AB=AB is Koszul, we will use the numerical criterion. We will simply use the fact (proved geometrically) that the Hilbert polynomial of AQ is given essentially by the intersection cohomology matrix

In fact, we have P⁡(AQ,t)=(PQ)T⁢PQ. But now we know that 𝒪𝔟=𝒪0 and that Lx𝔟=L⁡(x−1⁢w0∘0), and thus we have E⁡(AB)=AB=AB with 1x corresponding to 1x−1⁢w0. Therefore, we may define

and we have P⁡(E⁡(AB),t)=PBT⁢PB. We now have

by some magic of Kazhdan-Lusztig, which says that PB⁢(t)⁢(PB⁢(−t))T=1. □

Remark 10.2.5.

You could also prove that AB is Koszul by using a result of Bezrukavnikov that if A0,A1 are finite-dimensional and A≅E⁡(A), then A is Koszul.