Patrick Lei

All notes

Geometric rep theory and universal centralizers

Lectures by Xin Jin, Boston College, Spring 2026.

Abstract.

The goal of this course is to introduce many geometric objects and tools that are in the heart of GRT. Below is the tentative plan of the course:

  • •

    Part I. I will talk about the geometry of complex semisimple Lie algebras and groups. Specific topics include

    1. (1)

      The geometry of flag variety, nilpotent cone, Grothendieck-Springer resolution, Steinberg variety and an introduction to Springer theory

    2. (2)

      The wonderful compactification of semisimple groups and more general toroidal compactifications of reductive groups.

    We will follow some standard text/notes (and some additional references):

    • –

      Chriss-Ginzburg, Representation theory and complex geometry, mostly chapter 3 and part of chapter 2

    • –

      Evens-Jones, On the wonderful compactification

  • •

    Part 2. We will focus on a class of completely integrable systems, called the universal centralizers JG (or under other names: Toda systems, regular centralizer group schemes, or bi-Whittaker reductions), associated with a semisimple (or reductive) group G. This class of completely integrable systems lies at the crossroads of many different subjects in differential geometry, GRT (including geometric Langlands), (homological and enumerative) mirror symmetry, and mathematical physics (e.g. the class of Coulomb branches). Selected topics include:

    1. (1)

      A partial log-compactification of JG, by Balibanu (as an application of (2) in part 1)

    2. (2)

      Symplectic geometry of JG and sketch of its homological mirror symmetry (based on some of my recent work)

    3. (3)

      Depending on interest, I will sketch the proof of a hyperKahler metric on it constructed from gauge theory (following Donaldson, Kronheimer and Bielawski).

  • •

    Prerequisite: Basic knowledge of differential topology and complex geometry is assumed. Familiarity with semisimple Lie algebras and algebraic groups, as well as their representations, will be helpful but is not strictly required. We will follow the approach of Chriss–Ginzburg in recalling and supplementing the necessary background as needed.

1 Introduction

Representation theory studies some group G with some structure by considering representations (V,ρ:G→GL(V)) where V is a vector space and ρ is a group homomorphism with suitable properties. This seems like a purely algebraic question, but it is already very hard even when G is a finite group (in the nonabelian case). We may now bring in various geometric methods to construct representations of G:

  1. (1)

    If G acts on an algebraic variety X, then there is an action of G on the space of functions on X;

  2. (2)

    If G acts on an algebraic variety X, then we may consider a G-equivariant line bundle ℒ on X (which is a line bundle on [X/G]), we may consider the cohomology groups Hi⁢(X,ℒ) which are representations of G. If X is projective, then these cohomology groups are finite-dimensional;

  3. (3)

    If G acts on X, then it also acts on the cohomology H∙⁢(X) for any generalized cohomology theory;

  4. (4)

    There are also more hidden and subtle symmetries, where instead of actual group actions, we consider correspondences acting on cohomology groups. We will see examples of this later when we discuss Springer theory. In addition, we can lift this to the level of derived categories and consider Fourier-Mukai transforms;

1.1 Overview of the course.

The first part of the course will focus on the basics of geometric representation theory, largely focusing on Springer theory, the Borel-Weil-Bott theorem, and the geometry of the wonderful compactification G¯.

The second part of the course will focus on a completely integrable system called the Toda system and its homological mirror symmetry. One motivation for this is the notion of categorification, for example going from the Hecke algebra to the Hecke category, which will turn out to be a category of constructible sheaves.

One origin of categorification in algebraic geometry is Grothendieck’s function-sheaf correspondence, which we will adapt to our setting (in particular, we will not need ℓ-adic sheaves). One particularly simple example of this is upgrading the integers ℤ to the derived category of vector spaces over a field k, where the original integer is recovered as the Euler characteristic.

For a more sophisticated example, we may consider a smooth manifold X with a stratification 𝒮={Sα} by locally closed submanifolds. We may consider the space 𝖥𝗎𝗇𝒮⁢(X) of constructible functions (taking values in integers) with respect to 𝒮, which are locally constant on each stratum Sα. The natural categorification of this is the derived category 𝖲𝗁𝒮⁢(X) of constructible sheaves on X, where all cohomology sheaves are locally constant on each stratum Sα. The original constructible function is recovered by taking the Euler characteristic of the stalk at each point. In addition, we may recover the functions as

1.2 Homological mirror symmetry.

We will now discuss a connection to symplectic geometry. If we have a stratification 𝒮, then we may consider the union

of conormal bundles to the strata and consider the microlocal sheaf theory. In particular, each constructible sheaf ℱ has a microlocal stalk at each (x,ξ)∈T𝒮∗⁢X as in Figure 1. For example, if we have a restriction map ℱ⁡(U)→ℱ⁡(V), then the microlocal stalk is

where the shift is not canonical. One upshot of this approach is that

T∗⁢S1≅S1×ℝZero Section (ξ=0)x1x2Vx1Vx2VU1VU2Tx1∗⁢S1ξ>0p=(x1,ξ)Microlocal Stalk μp⁢ℱDepends on Vx1→VU1and Vx1→VU2
Figure 1. A microlocal stalk

The following theorem inspired many subsequent works.

Theorem 1.1 (Nadler-Zaslow).

The dg-category 𝖲𝗁c⁢(X) of constructible sheaves on X equivalent the Fukaya category 𝖥𝗎𝗄⁡(T∗⁢X) of the cotangent bundle T∗⁢X of X.

We will now discuss a geometric interpretation. We may consider the characteristic cycle CC⁡(ℱ) of a constructible sheaf, which is a conic Lagrangian cycle in T∗⁢X. However, the objects of the Fukaya category are smooth Lagrangian submanifolds with the cochain complexes of morphisms being given by counting interesection points and holomorphic polygons. See Figure 2 for a visualization. It is usually very difficult to calculate the Fukaya category directly, but the theorem allows us to translate these questions into questions about constructible sheaves, which are more manageable.

uL0L1pq𝒲⁢𝖥𝗎𝗄⁢(T∗⁢S1)
Figure 2. Two objects in the Fukaya category
Sketch of proof.

We will begin by discussing why there is a fully faithful functor Shc⁡(X)→𝖥𝗎𝗄⁡(T∗⁢X). The first step is to find some nice generators of the constructible sheaf category. For this, we consider open embeddings j:U↪X and extension by zero j!ℂU of the constant sheaf on U. These have characteristic cycles given by the conormal bundle to the boundary ∂U (together with the zero section at U). We will send this sheaf j!ℂU to a smoothing of its characteristic cycle, as depicted in Figure 3.

x (Base)ξ (Fiber)U=(a,b)CC(j!ℂU)LagrangianSmoothing
Figure 3. Characteristic cycle and its smoothing

To prove essential surjectivity, all of the Hom complexes in the Fukaya category can be calculated using Morse theory. For more details, see the original paper by Nadler-Zaslow. □

One reason to consider the Fukaya category is homological mirror symmetry, which was originally proposed by Kontsevich. For certain symplectic manifolds M, there is a mirror partner Mˇ, which is a complex algebraic variety, and Kontsevich conjectured that there is an equivalence of categories

The simplest example is when M=T∗⁢S1, where we need to consider the fully wrapped Fukaya category. We have an equivalence

(modulo technical details), but the upshot is that all characteristic cycles are supported inside the zero section. Local systems are on S1 are simply representations of π1⁢(S1)≅ℤ, which are simply modules over the group algebra ℂ⁡[ℤ]≅ℂ⁡[t,t−1]=𝒪⁡(ℂ×). Ignoring finiteness conditions, these are simply quasicoherent sheaves on ℂ×.

Remark 1.2.

For the purposes of representation theory, we will treat π1⁢(S1)=𝕏∙⁢(S1) as cocharacters, which will be identified with the characters 𝕏∙⁢(ℂ×) on the mirror side. This picture generalizes if we consider algebraic tori of any rank and is a version of Langlands duality in the abelian case.

Remark 1.3.

Homological mirror symemtry is closely related to the geometric Langlands correspondence, where we send the Hitchin system for G to the character variety for the Langlands dual group G∨. There are some technical issues, but the Betti geometric Langlands correspondence is essentially homological mirror symmetry here.

One modification of this is to complexify the Fukaya side into T∗⁢T, and in this case the (appropriate version of the) wrapped Fukaya category still reduces to local systems on T. Traditionally, we considered T∗⁢Tc (the cotangent bundle of the compact form) because of the SYZ mirror symmetry philosophy, which predicts that a manifold M and its mirror Mˇ have dual Lagrangian torus fibrations. In general, it is very hard to use SYZ to construct mirrors because there are usually singular fibers, which are hard to control.

However, one way to produce a mirror is to believe in homological mirror symmetry and then try to find a variety Mˇ such that

If the fibration has a Lagrangian section, then each smooth fiber becomes a Lie group, so 𝖥𝗎𝗄⁡(M) is expected to pick up a “convolution” symmetric monoidal structure, which should correspond to the tensor product of coherent sheaves in some kind of non-linear Fourier transform.

In the second part of the course, we will generalize the duality T∗⁢T↔T∨ to the universal centralizer (or Toda system) JG, where G is a semisimple complex Lie group. This is related to representation theory for two reasons.

  • •

    JG is also called the bi-Whittaker reduction and represents something like the generic part of the representation of G, namely the “commutative part” of the Hecke category;

  • •

    If we add a loop to G, then we obtain the affine Toda system;

  • •

    There is a partial (log) compactification J¯Glog over the base 𝔤∗⫽G constucted by Balibanu whose fibers are a kind of regular Hessenberg variety (a generalization of Springer fibers). The central fiber is the Peterson variety YP, which is a closed subvariety of the flag variety G/B and is related to the quantum cohomology of G∨/P∨ via the isomorphism

1 Geometry of complex semisimple Lie algebras and groups

2 Review of semisimple Lie algebras/groups

Eventually, the groups we are interested in are SLn⁡(ℂ), SOn⁡(ℂ), and Sp2⁢n⁡(ℂ) with Lie algebras 𝔰⁢𝔩n⁢(ℂ), 𝔰⁢𝔬n⁢(ℂ), and 𝔰⁢𝔭2⁢n⁢(ℂ). However, along the way we will encounter groups like Borels and unipotent groups.

2.1 Generalities.

Every Lie group has both a left and right action of G on itself given by the multiplication map

We may consider the tangent space 𝔤=Te⁢G at the identity, and left translation of vector fields gives an isomorphism

The right action gives another trivialization

These two are related by the diagram

where ι is the map sending an element of G to its inverse. If we fix ξ∈𝔤, then applying L gives us a left-invariant vector field ξℓ on G, and similarly applying R gives a right-invariant vector field ξr.

Lemma 2.1.
  1. (1)

    Left (resp. right) invariant vector fields on G are closed under the Lie bracket;

  2. (2)

    The left and right-invariant vector fields induce the same Lie bracket on 𝔤.

The main idea of the proof is that the Lie bracket is preserved by diffeomorphisms and in particular left (or right) translation.

Remark 2.2.

If we have any action a:G×X→X, then taking the derivative

gives an infinitesimal action of G on X (as in a map from 𝔤 to vector fields on X). For example, L is simply

and a similar statement is true for R.

One important action of G on itself is the adjoint action, where G acts on itself by conjugation. On the tangent space, we have (Rg−1)∗⁢Lg⁣∗=Lg⁣∗⁢(Rg−1)∗. By definition, we now have

which induces a map

Therefore, another equivalent definition of the Lie bracket is to simply differentiate the adjoint action. There is also an exponential map exp:𝔤→G, and the Lie bracket is defined as

Remark 2.3.

Any finite-dimensional abstract Lie algebra is isomorphic to a subalgebra of 𝔤⁢𝔩n by Ado’s theorem, which says that every finite-dimensional Lie algebra has a finite-dimensional faithful representation. In particular, every finite-dimensional Lie group over ℝ is locally an immersed subgroup of GLn⁢(ℝ) for some n.

We will give some principles of Lie theory:

  • •

    The Lie group is (up to taking its universal cover) determined by its Lie algebra. Given a connected Lie group (H,e), then any discrete covering (H′,e′) inherits a unique group structure whose kernel lies in the center of H′ (this works because the fundamental group of H is abelian due to the group structure). Conversely, any discrete central quotient of H is a Lie group with the same Lie algebra. If we work analytically, H has a universal cover, which is the unique simply connected Lie group with the same Lie algebra as H.

  • •

    If G and H are Lie groups such that G is connected, then any Lie group homomorphism

    is uniquely determined by its derivative dρ:𝔤→𝔥, which is a Lie algebra homomorphism.

  • •

    For 𝔤=𝖫𝗂𝖾⁡G and any embedding 𝔥↪𝔤, the subgroup of G generated by exp⁡(𝔥) is an immersed Lie subgroup of G with Lie algebra 𝔥. For example, if G=S1×S1 and 𝔥 is a line with irrational slope, then H is immersed rather than embedded.

  • •

    Let G and H be finite dimensional Lie groups with G connected and simply connected. Then a map f:Te⁢G→Te⁢H is the differential of a Lie group homomorphism ρ:G→H if and only if it is a Lie algebra homomorphism. The standard trick to prove this is to consider the graph

    which is a Lie algebra embedding. Then we get an immersed subgroup of G×H (this receives a map from G because G is simply connected), which we project down to H.

By Ado’s theorem, we see that the embedding 𝔤↪𝔤⁢𝔩n⁢(ℝ) gives an immersed subgroup G→GLn⁢(ℝ). In addition, the third item actually implies that there is a unique simply connected Lie group for any Lie algebra 𝔤.

Example 2.4.

Let X be a projective variety. Then studying the identity component Aut⁡(X)0 is similar to studying

which is a standard result in deformation theory. For example, if X=G/P is a partial flag variety, then H0⁢(X,T⁢X)=𝔤 in almost all cases, but if P=B (the full flag variety), then we really do have

In particular, we have

and when G is semisimple, the center is discrete.

We will call a Lie algebra homomorphism

a representation of 𝔤. For example, we previously considered the adjoint representation

We can also take the direct sums V⊕W and tensor products V⊗W of representations. It should be noted that the action on V⊗W should take the form

in analogy with differentiating the action of G on V⊗W, which takes the form

2.2 Rough classification of Lie algebras.

Given a Lie algebra 𝔤, there is a lower central series

and the smaller derived series

Definition 2.5.

A Lie algebra is nilpotent if the lower central series becomes zero at some point.

One example of a nilpotent Lie algbera is if we consider upper-triangular matrices with zeroes on the diagonal (usually denoted 𝔫).

Definition 2.6.

A Lie algebra is solvable if the derived series eventually vanishes.

An example of a solvable Lie algebra is the algebra 𝔟 of all upper-triangular matrices. The reason for this is that [𝔟,𝔟]=𝔫. If we write

then we see that [𝔥,𝔥]=0, [𝔥,𝔫]=𝔫, and [𝔫,𝔫]⊂𝔫.

Theorem 2.7 (Engel).

Let 𝔤⊂𝔤⁢𝔩⁢(V) be a Lie subalgebra such that every X∈𝔤 is a nilpotent endomorphism of V. Then there exists a nonzero v∈V such that X⁡(v)=0 for all X∈𝔤.

Engel’s theorem implies that all X∈𝔤 can be made (strictly) upper-triangular simultaneously.

Theorem 2.8 (Lie).

Let 𝔤⊂𝔤⁢𝔩⁢(V) be a solvable Lie subalgebra over ℂ. Then there is a common eigenvector for all X∈𝔤.

In particular, Lie’s theorem implies that all X∈𝔤 can be simultaneously made upper-triangular.

We will now consider all solvable ideals in 𝔤. In fact, if we have a finite collection of solvable ideals, then the sum is still solvable, so there exists a maximal solvable ideal of 𝔤, called the radical of 𝔤 and denoted Rad⁡(𝔤). This fits into the exact sequence

2.3 Semisimple and reductive Lie algebras.

Definition 2.9.

A Lie algebra 𝔤 is semisimple if Rad⁡(𝔤)=0.

Example 2.10.

The Lie algebra 𝔤/Rad⁡(𝔤) is semisimple.

Definition 2.11.

A Lie algebra is reductive if Rad⁡(𝔤)=Z⁡(𝔤).

Remark 2.12.

There are other characterizations of reductive Lie algebras, for example that the nilradical (maximal nilpotent ideal) Rnil⁢(𝔤), which fits into the exact sequence

vanishes.

We now give some important properties of semisimple Lie algebras.

Theorem 2.13.

Let 𝔤 be a semisimple Lie algebra. Then every finite-dimensional representation of 𝔤 decomposes uniquely as a direct sum of irreducible representations.

Sketch of proof.

There are two ways to do this, either algebraically (which is tedious) or by using Weyl’s unitary trick. The key fact using the unitary trick is that every complex semisimple Lie algebra 𝔤 has a unique compact real form 𝔤c satisfying 𝔤c⊗ℝℂ=𝔤 (being a real form) and being the Lie algebra of a simply connected compact Lie group.

For example, the compact real form of 𝔰⁢𝔩n⁢(ℂ) is 𝔰⁢𝔲n (traceles anti-Hermitian matrices), which corresponds to SUn⊂SLn⁡(ℂ). The unitary trick is to restrict to the action of the compact form Gc on V and then try to find a Gc-invariant Hermitian metric on V, which will split W⊂V for free. The way to find the inner product is to pick any Hermitian metric ⟨−,−⟩ and a Haar measure μ on Gc such that

and define a new inner product by

(which is the same trick we use for finite groups). Finally, the Gc-representations lift to G, so we are done. □

Theorem 2.14.

In fact, a Lie algebra is semisimple if and only if all of its finite-dimensional representations are semisimple.

Proof.

Consider the adjoint representation ad:𝔤→𝔤⁢𝔩⁢(𝔤). Because it is semisimple, we have a decomposition

In particular, we must have [Rad⁡(𝔤),𝔤ss]=0, where 𝔤ss=𝔤/Rad⁡(𝔤).

If Rad⁡(g)≠0, then by Lie’s theorem says that any representation must have an eigenvector, so all irreducible representations must be 1-dimensional. This implies that

An abelian Lie algebra has indecomposable representations which are not semisimple, for example representing a line by matrices of the form [0∗00], so the radical must have vanished to begin with. □

Remark 2.15.

Levi’s theorem actually states that 𝔤=Rad⁡(𝔤)⋊𝔤ss.

Theorem 2.16 (Preservation of Jordan decomposition).

Let 𝔤 be semisimple. For any X∈𝔤, there exist Xs,Xn∈𝔤 such that Xs+Xn=X and for any finite-dimensional representation

the Jordan normal form of ρ⁡(X) has ρ⁢(X)s=ρ⁡(Xs) and ρ⁢(X)n=ρ⁡(Xn).

Remark 2.17.

If we consider the adjoint representation of 𝔤 (which is faithful), then in fact the condition on the Jordan decomposition in any representation actually implies that X=Xs+Xn.

Example 2.18.

If we consider 𝔤=ℂ, then the representation

has ρ⁢(X)s and ρ⁢(X)n not in the image of ρ in general.

Definition 2.19.

If 𝔤 is semisimple, then we say X∈𝔤 is semisimple if X=Xs and nilpotent if X=Xn.

We will now consider reductive Lie algebras. In fact, by Levi’s theorem and the fact that [Z⁡(𝔤),𝔤ss]=0, 𝔤 is reductive if and only if it is the product of a semisimple and an abelian Lie algebra, which is equivalent to 𝔤=Z⁡(𝔤)⊕[𝔤,𝔤].

Definition 2.20.

A Lie group is semisimple if its Lie algebra is semisimple.

We will now consider all complex semisimple Lie groups, which are actually affine algebraic groups. By Borel’s theorem, they are actually all linear, meaning closed subgroups of some GLn. If 𝔤 is semisimple, then there is a simply connected form Gsc and a tower down to the adjoint form Gsc=Gsr/Z⁡(Gsc).

Example 2.21.

For the Lie algebra 𝔰⁢𝔩n, then the simply connected form is SLn and the adjoint form is PGLn.

Example 2.22.

If we consider 𝔰⁢𝔭2⁢n, then the simply connected form is Sp2⁢n⁢(ℂ) and the adjoint form is PSp2⁢n⁢(ℂ).

Example 2.23.

For SOn, there is a form SOn, which for n≥3 has fundamental group ℤ/2, so the simply connected form is the universal cover Spinn and the adjoint form is PSOn.

Remark 2.24.

We may also consider simple Lie algebras, which are semisimple Lie algebras with no nontrivial ideals. Then the classical simple Lie algebras are the ones we have already listed and are the only infinite families.

There is also the Cartan-Killing criterion of semisimple Lie algebras.

Definition 2.25.

An invariant bilinear form on 𝔤 is given by

satisfying the condition that

Definition 2.26.

The Killing form is the form

Because the trace has cyclic invariance and the trace is invariant under conjugation, this is invariant and symmetric.

Theorem 2.27 (Cartan-Killing criterion).

A Lie algebra 𝔤 is semisimple if and only if the Killing form is nondegenerate.

In particular, the means that the adjoint and coadjoint representations are isomorphic.

Remark 2.28.

If Z⁡(𝔤)≠0, then adx=0 for all x∈Z⁡(𝔤), so B⁡(x,−)=0 for B the Killing form.

We now briefly pause to consider issues of analytic vs algebraic categories. If we are considering semisimple Lie groups over ℂ, then they are in fact also (linear) algebraic groups. However, if we leave the semisimple setting, then there are differences. For example, if we consider 𝔤=ℂ, then the algebraic options are (ℂ,+)=𝔾a, (ℂ×,×)=𝔾m, and compact 1-dimensional tori, namely elliptic curves. Usually, in algebraic groups theory, we will only consider affine algebraic groups.

Theorem 2.29 (Borel).

Every affine algebraic group G is linear, and has a closed immersion to some GL⁡(V).

In the analytic world, there is the map

which is not algebraic. In particular, 𝔾a and 𝔾m are totally different and have no nontrivial morphisms between them (and we have to consider their formal completions to recover the exponential map). In fact, the category 𝖱𝖾𝗉⁡(𝔾m) is semisimple (prove this using the unitary trick for S1), while the same is not true for 𝔾a (consider the representation by matrices of the form [1∗01]). However, any analytic Lie group over ℂ not containing a compact torus is algebraic.

We now proceed to defining reductive groups. Recall that we essentially need to figure out what to do with the center Z⁡(𝔤) of a reductive Lie algebra 𝔤. In the algebraic group world, we will consider tori, which are groups isomorphic to products of copies of 𝔾m.

Definition 2.30.

A reductive group G is a finite quotient of the product of an (affine) algebraic torus and a semisimple algebraic group.

In particular, we have

where Gder=[G,G] is semisimple and Z⁡(G) might be disconnected. For example, Z⁡(SLn)=μn, which is clearly not connected. Alternatively, there is a quotient morphism

where again Gad=G/Z⁡(G). Here, we see that Z⁡(G)/Z⁡(Gder) is always connected. Using Langlands duality, we can produce a cover by something with connected center. A third alternative is to rewrite the product as

where Z⁢(Gder)0=Z⁢(G)0∩Gder.

Example 2.31.

In the case where G=GLn, we have Z⁡(GLn)=ℂ× and (GLn)der=SLn, so we have

Reductive groups are characterized by complete reducibility of their finite-dimensional representations (in characteristic zero). In addition, algebraic and analytic morphisms between reductive groups agree, so we no longer need to worry about the difference between algebraic and analytic categories. Representations of reductive groups also satisfy the preservation of Jordan decompositions, namely that for any g∈G, there exist gs,gu∈G such that g=gs⁢gu=gu⁢gs and for any finite-dimensional representation

the Jordan decomposition of ρ⁡(g) satisfies ρ⁢(g)s=ρ⁡(gs) and ρ⁢(g)u=ρ⁡(gu).

Lemma 2.32.

If 𝔤 is semisimple, then any invariant bilinear form is symmetric. In addition, any invariant non-degenerate bilinear form is a scalar multiple of the Killing form.

Proof.

Any invariant bilinear form B induces a map

as morphisms of G≔Gad-representations. Recall that 𝔤 decomposes into a product

of simple Lie algebras, so we have a morphism

Using the Killing form, we have an isomorphism 𝔤i≅𝔤i∗, so by Schur’s lemma, the map 𝔤i→𝔤j∗ is zero if i≠j and a scalar multiple of the identity if i=j. Therefore, any invariant bilinear form is a sum of scalar multiples of Killing forms on each simple factor, which are all symmetric. The second statement follows similarly. □

2.4 Cartan subalgebras.

We will begin by considering general 𝔤 and specialize to semisimple ones later.

Definition 2.33.

If 𝔞⊂𝔤 is a Lie subalgebra, then the normalizer of 𝔞 in 𝔤 is given by

Definition 2.34.

A subalgebra 𝔥⊂𝔤 is a Cartan subalgebra if

  1. (1)

    𝔥 is nilpotent;

  2. (2)

    𝔥=𝔫𝔤⁢(𝔥) (as in it is equal to its own normalizer).

We will see later that every Lie algebra has a Cartan subalgebra.

For any x∈𝔤, let

denote the characteristic polynomial of adx. If we choose coordinates x1,…,xn on 𝔤, then each ai⁢(x) is a homogeneous polynomial in x1,…,xn of degree n−i.

Definition 2.35.

The rank of 𝔤 is the least i such that ai⁢(x) is not identically zero.

Example 2.36.

If 𝔤≅ℂn is abelian, then Px⁢(T)=Tn for all x, so 𝔤 has rank n.

Remark 2.37.

Note that i<n in the nonabelian case because an⁢(x)=1 is never zero and i>0 because x∈ker⁡(adx) for all x and therefore a0⁢(x)=0 for all x.

From now on, we will denote the rank of 𝔤 by r.

Definition 2.38.

We say that x∈𝔤 is strongly regular if ar⁢(x)≠0.

Remark 2.39.

We will see another notion of regularity later, so we call this strongly regular to distinguish it.

We will denote the set of strongly regular elements by 𝔤sreg. The definition of a strongly regular element implies the following.

Proposition 2.40.

The set 𝔤sreg is a connected open dense subset of 𝔤 (in the Zariski topology and hence also the analytic topology).

Theorem 2.41.

If x∈𝔤sreg, then the generalized eigenspace of 0 for adx, denoted 𝔤x0, is a Cartan subalgebra of 𝔤.

Note that by definition, we have dim𝔤x0=r, so the dimension of this Cartan equals the rank of 𝔤. In fact, we have already described all Cartan subalgebras.

Theorem 2.42.

Let 𝔤 be a Lie algebra over ℂ. Then all Cartan subalgebras of 𝔤 are of the form 𝔤x0 for some x∈𝔤sreg.

These Cartan subalgebras are all conjugate under the action of

so in particular Ginn acts transitively on the set of Cartan subalgebras.

Remark 2.43.

If G is reductive, then Ginn=Gad=G/Z⁡(G).

2.5 Root systems.

We will now specialize to the semisimple case, which is much more explicit.

Theorem 2.44.

Let 𝔥 be a Cartan subalgebra of a semisimple Lie algebra 𝔤. Then

  1. (1)

    𝔥 is abelian;

  2. (2)

    𝔥 is its own centralizer, i.e. Z𝔤⁢(𝔥)=𝔥;

  3. (3)

    Every element of 𝔥 is semisimple.

  4. (4)

    The restriction of the Killing form to 𝔥 is nondegenerate.

Example 2.45.

If 𝔤=𝔰⁢𝔩n, then the Cartan subalgebras are precisely the conjugates of the diagonal matrices. This has dimension n−1 (given the constraint that the trace vanishes), so the rank of 𝔰⁢𝔩n is n−1.

For a semisimple Lie algebra, if x∈𝔥=𝔤x0, then in fact 𝔥=ker⁡adx for a general choice of x. The necessary condition on x is that it is regular semisimple, which we can take to simply mean that ker⁡adx is as small as possible. Because we can simultaneously diagonalize all adx for x∈𝔥, we have a decomposition

where

and Φ is the set of nonzero α∈𝔥∗ such that 𝔤α≠0.

Example 2.46.

Returning to 𝔰⁢𝔩n, if we let

then twe see that

If we refer to the coordinates on ℂn as e1,…,en, then the roots are given by

where αi⁢j⁢([x1000…000xn])=xi−xj. We refer to the set Φ as the roots of 𝔤 with respect to 𝔥 and 𝔤α as the root space corresponding to α.

Because the Killing form is nondegenerate when restricted to 𝔥, we may identify ei∗−ej∗∈𝔥∗ with ei−ej∈𝔥. In particular, we may consider Φ as a subset of 𝔥 itself.

Corollary 2.47.
  1. (1)

    𝔥 is a maximal abelian subalgebra of 𝔤;

  2. (2)

    Every strongly regular element is semisimple;

  3. (3)

    Gad acts transitively on the set of all Cartan subalgebras of 𝔤.

Warning 2.48.

Not all maximal abelian subalgebras are Cartan subalgebras, for example the nilpotent matrices in 𝔰⁢𝔩2.

Lemma 2.49.

Assume that G is semisimple. Let T=NG⁢(𝔥)0 (usually, it is not connected). Then T is a maximal torus in G with T≅(ℂ×)r.

Sketch of proof.

Choose a compact real form 𝔤c⊂𝔤. We now consider the set

Now let ξc∈𝔤c be strongly regular. Then Z𝔤c⁢(ξc)=𝔥c is a Cartan subalgebra of 𝔤c. Next, the normalizer

is a maximal torus of Gc of real dimension r, so we can conclude the proof by complexifying everything. □

Example 2.50.

In the case of 𝔰⁢𝔩n, if ξ is strongly regular, then we see that ar⁢(ξ)≠0, and so dimker⁡adξ=r=n−1, so ξ must be semisimple. In particular, the root space decomposition is

Lemma 2.51.

Choose a root decomposition

Then ξ∈𝔥 is strongly regular if and only if there exists α∈Φ such that α⁡(ξ)≠0 for all α∈Φ.

Proof.

Note that ξ∈𝔥 is strongly regular if and only if ker⁡adξ=𝔥, which holds if and only if adξ|𝔤α is nonzero for all α∈Φ, which holds if and only if α⁡(ξ)≠0 for all α∈Φ. □

Remark 2.52.

From now on, we will refer to strongly regular elements as regular semisimple elements in the semisimple case and refer to the set of such elements as 𝔤rs. We will also denote

which is given by the equation

Example 2.53.

The Lie algebra 𝔤=𝔰⁢𝔩2 is generated by the matrices

with the relations

Then we see that the Cartan is 𝔥=ℂ⁢h and the root spaces are 𝔤α=ℂ⁢e and 𝔤−α=ℂ⁢f, where α⁡(h)=2. Therefore, the root system is Φ={α,−α} and is shown in Figure 4. In this case, a choice of α is the same as choosing a set Φ+⊂Φ of positive roots. In this case, α is also a simple root, but in general positive and simple roots are not the same.

α−α
Figure 4. Root system of 𝔰⁢𝔩2

Now set α∨=h∈𝔥, which is a simple coroot. Then we have ⟨α∨,α⟩=2. We may now consider the maximal torus

The adjoint action of T on 𝔤 is given by

Note that α⁡(h)=2, which matches the exponent. We may now consider the cocharacter lattice

If we project down to the adjoint form PGL2=SL2/±I we see in fact that

Note that this cocharacter lattice is dual to the root lattice and is called the coweight lattice. We can now identify the character lattice of T=Tsc with the weight lattice ℤ⋅12⁢α, whereas 𝕏∙⁢(Tad)=ℤ⋅α is the root lattice.

Now consider the exact sequence

This induces the exact sequence

In particular, we see once again that π1⁢(Gad)≅Z⁡(Gsc). If we instead consider maximal tori, we have

which induces the exact sequence

on fundamental groups. This gives us the identification

and relates the coroot and coweight lattices.

Example 2.54.

In the case of 𝔰⁢𝔩3, the root system is drawn in Figure 5. In particular, the roots are given by

𝒞ω23ω12α13−α23−α13−α12α23α12
Figure 5. Root system of 𝔰⁢𝔩3

In order to treat the coroots, note that

for any root α. Therefore, we may choose eα∈𝔤α and fα∈𝔤−α such that

For example, if we consider α=α12, then we may choose

We now have the coroot αi⁢j∨=hαi⁢j such that ⟨αi⁢j∨,αi⁢j⟩=2.

To choose a basis, we cut 𝔥∗ by a general linear hyperplane and choose a positive direction. This gives a set Φ+ of positive roots. For example, we will choose Φ+={αi⁢j|i<j} for 𝔰⁢𝔩3. Inside this, there is a unique set of simple roots Π⊂Φ+ satisfying

  1. (1)

    It is a basis of 𝔥ℝ;

  2. (2)

    Any β∈Φ+ is a nonnegative integer combination of elements of Π.

For example, in this case, we have Π={α12,α23}⊂Φ+. The simple coroots are then simply Π∨={α12∨,α23∨}. Then the Cartan matrix is given by

We should also explain why the picture has the angles as shown. If we recall that the root lattice Q is identified with the character lattice X∙⁢(Tad), this is proved by considering the kernel ker⁡Q⊂Tad. In particular, anything killed by Q must actually act trivially on the root spaces and acts trivially on 𝔥 by definition, so it lives in the center of Gad, which is trivial. Continuing as before, we consider the coweight lattice, which is spanned by the matrices

𝒞ω23∨ω12∨α13∨−α23∨−α13∨−α12∨α23∨α12∨
Figure 6. Coroot system of 𝔰⁢𝔩3
Remark 2.55.

The Cartan matrix is symmetric if and only if the Dynkin diagram is simply laced, i.e. has no multiple edges.

Unfortunately, we will not be able to discuss root systems in detail for time reasons. In type A (where 𝔤=𝔰⁢𝔩n+1), the picture generalizes that of 𝔰⁢𝔩3, where we can choose Φ+={αi⁢j|i<j} and Π={αi,i+1|1≤i≤n}. As a remark, we will also refer to the weight lattice as P and the coweight lattice as P∨. In general, we will interpret the fundamental weights ωi as the dual basis to the simple coroots αi∨.

In the group picture, if we have an intermediate form G with maximal torus T, then we have the inclusions

induced by

To find the center of G, we simply have the quotient Z⁡(G)=X∙⁢(Tad)/X∙⁢(T), which comes from the long exact sequence on homotopy groups coming from

To find the fundamental group of G, we apply this to the exact sequence

which gives us π1⁢(G)=X∙⁢(T)/X∙⁢(Tsc). In particular, this specializes to

For example, if Gsc=SL3, then we have

Because this group has prime order, there are no intermediate forms between SL3 and PGL3.

Using the Killing form, there is the formula

and this actually implies that

Because this needs to be an integer, we see that for all roots, the angle θ between them must satisfy 4⁢cos2⁡θ∈ℤ. This restricts the possible angles to π2,2⁢π3,π3,π4,3⁢π4,π6, or 5⁢π6 unless α=±β.

Remark 2.56.

Root systems are classified by Dynkin diagrams, which is equivalent to classifying Cartan matrices. The entries of the Cartan matrix satisfy Ci⁢i=2 and Ci⁢j≤0. Non-symmetry can be explained by the fact that in some root systems, αi and αj may have different lengths. This gives arise to the notions of long roots and short roots. In fact, there are only two possible lengths for all roots (under the Weyl group action, any root can become simple).

For an example of a root system with two lengths, consider the B2 root system, whose root and coroot systems are shown in Figure 7.

𝒞e1−e1e2−e2e1+e2−e1−e2e1−e2−e1+e2ω1ω2α1α2
(a) Root system of B2
𝒞2⁢e1−2⁢e12⁢e2−2⁢e2e1+e2−e1−e2e1−e2−e1+e2ω1∨ω2∨α1∨α2∨
(b) Coroot system of B2
Figure 7. Root and coroot systems of type B2

2.6 Weyl groups.

Recall that the maximal torus T has Lie algebra 𝔥. Then we define the Weyl group

Here, we know that 𝔫𝔤⁢(𝔥)=𝔥, so T⊂NG⁢(T).

Example 2.57.

If g=𝔰⁢𝔩n, then WT≅Sn because we can permute the diagonal entries.

The Weyl group acts on the set Φ of roots. Each root α corresponds to a reflection sα∈WT given by

This satisfies sα2=id and fixes the hyperplane α⟂⊂𝔥∗ pointwise.

In the 𝔰⁢𝔩n example, the root αi⁢j corresponds to the reflection that swaps the i-th and j-th coordinates. In particular, its kernel is the set of all matrices satisfying xi=xj.

The Weyl group is generated by the simple reflections, which correspond to the simple roots Π. We will refer to the simple roots as αi∈Π and the corresponding reflections as si=sαi. Then, the Weyl group has the presentation

and is a finite reflection group.

The reflection hyperplanes α⟂ cut the space 𝔥ℝ∗ into a set of chambers, which are fundamental domains for the Weyl action and are in bijection with the choices of positive roots Φ+. In particular, the set Φ+ corresponds to the chamber

Note also that the choice of a set of positive roots corresponds to a choice of Borel 𝔟⊃𝔥. In particular, there is the direct sum decomposition

In addition, for any α,β∈Φ we have

2.7 Regular elements.

Let ξ∈𝔥reg≔𝔥∖⋃α∈Φ+ker⁡α. Note that WT acts freely on 𝔥reg. Finally, let r=rank⁡𝔤=dim𝔥.

Proposition 2.58.

Let x∈𝔤. Then the centralizaer Z𝔤⁢(x) contains an r-dimensional abelian subalgebra.

Proof.

Since the regular semisimple locus in dense, there exists a sequence xn→x of regular semisimple elements. For each xn, its centralizer 𝔥n=Zg⁢(xn) is r-dimensional ane abelian.

Finally, the compactness of the Grassmannian Gr⁡(r,𝔤) implies that there is some subsequence 𝔥n=Z𝔤⁢(xn) which converges to some limit 𝔲. We need to check that 𝔲⊂Z𝔤⁢(x), but we know that if y∈𝔲, there is some sequence yn→y such that yn∈𝔥n. Because the Lie bracket is continuous on both arguments, we see that

We now need to check that 𝔲 is abelian. But this follows similarly: if y,z∈𝔲, there are sequences yn→y and zn→z such that yn,zn∈𝔥n. Then

Corollary 2.59.

For all x∈𝔤, then dimZ𝔤⁢(x)≥r and dimZG⁢(x)≥r. In addition, every maximal abelian subalgebra has dimension at least r.

Definition 2.60.

An element x∈𝔤 is called regular if dimZ𝔤⁢(x)=r. Equivalently, the adjoint orbit G⋅x has maximal dimension.

Example 2.61.

An example of a regular element which is not semisimple is

Its centralizer is given by

and is 2-dimensional.

We will now discuss a general criterion for regularity. Recall the Jordan decomposition

where xs is semisimple and xn is nilpotent. Because ad⁡x=ad⁡xs+ad⁡xn=(ad⁡x)s+(ad⁡x)n, we have

Proposition 2.62.

The element x∈𝔤 is regular if and only if xn is regular in the derived algebra of ℓxs≔Z𝔤⁢(xs).

Remark 2.63.

The centralizer ℓxs is a reductive subalgebra of 𝔤, called a Levi subalgebra. Here, if xs∈𝔥 for some Cartan 𝔥 and

is the root decomposition, then we have

We now consider the root system

This is the root system for ℓxsder (note we need to take the derived subalgebra to eliminate the center). Finally, we note that

so the rank of the root system increases as we move to smaller strata of 𝔥 (with respect to the stratification by root hyperplanes). The structure of the Levi subalgebra is in fact

where the center is (Φ′)⟂ and the semisimple part has root system Φ′.

Example 2.64.

In 𝔰⁢𝔩3, the matrix

lies in the root hyperplane α12=0 and its Levi subalgebra is given by traceless block-diagonal matrices of the form

Taking the derived subalgebra, we obtain a copy of 𝔰⁢𝔩2 in the upper-left block. The center is simply matrices of the form

In 𝔰⁢𝔩n, we can group the blocks by shared eigenvalues and proceed similarly. If the blocks have sizes m1,…,mk, then the derived subalgebra of the Levi is given by 𝔰⁢𝔩m1×⋯×𝔰⁢𝔩mk and the center has dimension k−1.

Proof of Proposition 2.62.

Note first that

Set-theoretically, we have

If dimZ⁡(ℓ)=k, then rank⁡ℓder=dimZ⁢(ℓ)⟂=r−k, where the orthogonal complement is taken in 𝔥∗. Therefore, we know that x is regular if and only if dimZℓder⁢(xn)=r−k=rank⁡ℓder, which is equivalent to xn being regular in ℓder. □

Example 2.65.

We may assume that y∈𝔰⁢𝔩n is already in Jordan normal form. Then in fact it is regular if and only if all Jordan blocks have distinct eigenvalues. Here, the Levi subalgebra is given by block-diagonal matrices where the blocks are grouped by shared eigenvalues. However, any regular nilpotent element in 𝔰⁢𝔩m has a single Jordan block, so we see that each eigenvalue of y must correspond to a single Jordan block.

2.8 Flag varieties and the Bruhat decomposition.

We will now discuss Borel subalgebras 𝔟⊂𝔤.

Definition 2.66.

A Borel subalgebra in 𝔤 is a maximal solvable subalgebra. A Borel subgroup in G is a maximal connected solvable subgroup.

Note that if we have a Borel subalgebra, we can integrate it to get a Borel subgroup B⊂G. Here are some properties:

  1. (1)

    All Borel subalgebras are conjugate under the adjoint representation;

  2. (2)

    If 𝔟 is a Borel subalgebra, then 𝔫𝔤⁢(𝔟)=𝔟 and NG⁢(𝔟)=B;

  3. (3)

    Every Borel subalgebra contains a Cartan subalgebra. On the other hand, a fixed Cartan is contained in |WT|-many Borel subalgebras, which form a WT-torsor. In fact, they are in bijection with the set of choices of positive roots via

    For example, we can let 𝔥 be the diagonal matrices in 𝔰⁢𝔩n, and we can generate all Borels by taking the upper-triangular matrices as a base point and conjugating by permutation matrices.

  4. (4)

    If h∈𝔟 is regular semisimple, then in fact 𝔥≔Z𝔤⁢(h)⊂𝔟.

Now let ℬ be the set of Borel subalgebras of 𝔤. We can give it the structure of an algebraic variety as follows. First, consider the Grassmannian

which is naturally an algebraic variety which carries a G-action induced from the adjoint representation. Then ℬ sits inside as a subset which is in fact the orbit of any Borel 𝔟 under this action. This gives

It is easy to check that ℬ is in fact closed, so it is a projective variety called the flag variety of G (and the quotient description means it is a homogeneous space).

Remark 2.67.

Because the Borel subgroup B always contains the center Z⁡(G), the flag variety is insensitive to the choice of intermediate form of G. If G is reductive, then this construction kills the center, so we only see the semisimple part of G in the flag variety.

Example 2.68.

If G=SLn (or PGLn), then in fact

is the usual flag variety of ℂn. This clearly admits a GLn-equivariant surjection from the variety of frames

If we consider the standard flag

then the stabilizer of this flag is the standard Borel subgroup B+ of upper-triangular matrices, so we see that ℬ=G/B+.

Now we may restrict the G-action on ℬ to the Borel B, where we choose a particular Borel B as a base point. Then the set of B-orbits is denoted by

We may also consider the G×G-action on ℬ×ℬ and restrict to the diagonal subgroup GΔ. Then the set of GΔ-orbits is given by

To check that this is well-defined, suppose that 𝔟′′=b⋅𝔟′. This follows from the fact that

To check surjectivity, we look at the orbit of (𝔟1,𝔟2), but we can bring 𝔟1 to 𝔟, and surjectivity follows. To check injectivity, we see that if GΔ⋅(𝔟,𝔟′)=GΔ⁢(𝔟,𝔟′′), then (𝔟,𝔟′′)=g⋅(𝔟,𝔟′) for some g∈G. But if g⋅𝔟=𝔟, then g∈B and thus 𝔟′′∈B⋅𝔟′.

We will now consider a more refined bijection which will give the Bruhat decomposition. Fix a maximal torus T⊂B, or equivalently a Cartan subalgebra 𝔥⊂𝔟. Then we have a map

where we choose a representative w˙∈w⁢T of a coset in WT=NG⁢(T)/T. Because T⋅B=B, this map is well-defined. Taking stock, we now have maps

Our goal is to show that the final arrow is a bijection.

Theorem 2.69 (Bruhat decomposition).

The map WT→B\G/B is a bijection.

Remark 2.70.

This implies the decomposition

Example 2.71.

If G=SLn and we choose B is the upper-triangular matrices with T the diagonal matrices, then we can present w˙ as a permutation matrix, and the Bruhat decomposition says that any invertible matrix can be written as a product of an upper-triangular matrix, a permutation matrix, and another upper-triangular matrix.

The proof of this theorem is based on the Bialynicki-Birula decomposition, which is an analog of Morse theory. If we have an action of ℂ× on a projective variety X, then we may consider the fixed point locus 𝒲=Xℂ×. We will assume for now that 𝒲 is discrete (hence finite).

Example 2.72.

If X=ℙ1, then we consider the action λ⋅[a:b]=[a:λ⁢b] with fixed points 0=[1:0] and ∞=[0:1]. For any other point, we consider the limit

In this case, the attracting locus of 0 is 𝔸1=ℂ, while the attracting locus of ∞ is just the point ∞.

For each fixed point w∈𝒲, we consider the tangent space Tw⁢X, which is a representation of ℂ×. We can decompose it into weight spaces

where Tw⁢X⁢[n] is the subspace on which λ∈ℂ× acts by multiplication by λn. By assumption, there are no zero weights, so we have the decomposition

where the first summand is the sum of positive weight spaces and the second summand is the sum of negative weight spaces.

Theorem 2.73 (Bialynicki-Birula decomposition).

The variety X decomposes into a disjoint union

where the attracting manifold Xw is the locally closed subvariety

and is also given by

Proof of Bruhat decomposition.

We will establish bijections

Note that the third bijection follows directly from the Bialynicki-Birula decomposition, so we need to establish the other bijections.

First note that there is a bijection between T-fixed points in ℬ and the set of all Borels 𝔟′⊃𝔥. This is simply the infinitesimal version of the fact that if T⋅B′=B′, then T⊂NG⁢(B′)=B′. Now this is in bijection with the set of w˙⋅𝔟 for w∈WT, which establishes the first bijection.

Now choose a cocharacter h:ℂ×↪T which is regular and a primitive nonzero cocharacter of T. For this ℂ×-action, the fixed points are in bijection to the set of all Borel subgroups B′ containing ℂ×. Taking the derivative at the identity, this is in bijection with

Because h is regular, then 𝔥=Z𝔤⁢(h)⊂𝔟′, so we have recovered the set of all 𝔟′⊃𝔥.

It remains to identify B-orbits with attracting sets of the ℂ×-fixed points. If 𝔟w∈ℬℂ× is a Borel, then we consider its attracting set

The attracting sets are the positive parts of the tangent spaces at the fixed points, and so we must have T𝔟w⁢ℬ≅𝔤/𝔟w because ℬ≅G/Bw. Note that the action of ℂ× on 𝔤 decomposes into eigenspaces

where 𝔥 will disappear when we quotient by 𝔟w. Our goal is to have

Considering the root decomposition

considering ℂ×-weights gives us

There is a Weyl chamber (see Figure 6 for the SL3 example) which pairs with the positive roots Φ+ positively, so we are free to choose 𝔥 in this chamber. Once we choose 𝔥, then the inclusion 𝔫+→𝔤/𝔟w factors through T𝔟w⁢ℬ.

Remark 2.74.

One convenient choice for a regular cocharacter h is h=2⁢ρ∨, where

This implies that

which lies in the coroot lattice Q∨=𝕏∙⁢(Tsc). By definition, we have ⟨ρ∨,αj⟩=1 for all simple roots, so it pairs positively with all positive roots.

Now let U⊂[B,B] such that 𝖫𝗂𝖾⁡U=𝔫+. Then if we consider the orbit U⋅𝔟w, we have

Dimension counting shows that U⋅𝔟w is open and dense in ℬ𝔟w, so our goal is to show that it is also closed, i.e. that U⋅𝔟w=ℬ𝔟w.

Lemma 2.75.

Suppose that U is a unipotent group acting on an affine variety X=Spec⁡A. Then any U-orbit in X is closed.

Proof.

Let 𝒪 be a U-orbit and 𝒪¯ be its closure. Suppose that C≔𝒪¯∖𝒪 is nonempty (and hence is a closed subvariety of X) and has defining ideal 0≠I⊂ℂ⁡[𝒪¯]. Because C is U-stable, so is I.

Lemma 2.76.

For all f∈I, the span Spanℂ⁡U⋅f is finite-dimensional.

Proof.

The action of U on X gives a coaction

Then we see that

and thus

which is finite-dimensional. □

Now consider IU⊃(Spana≠0⁡U⋅a)U≠0. Choose 0≠f∈IU. Then f is U-invariant, so it must be constant on 𝒪, and hence constant on 𝒪¯. But we know that f vanishes on C, so it must vanish on 𝒪¯, which contradicts the fact that f is nonzero. Therefore, C must be empty and 𝒪 is closed. □

The two lemmas together emply that U⋅𝔟w=B⋅𝔟w=ℬ𝔟w. □

Example 2.77.

In the case when G=SL2, then the flag variety if ℙ1, and if we choose 𝔟 to be the upper-triangular matrices with 𝔥 the diagonal matrices, then the Weyl group is

and the nontrivial element w swaps the upper-triangular and lower-triangular Borel subgroups. See Figure 8.

w⋅𝔟𝔟wℙ1≅S⁢L2/B
Figure 8. Action of Weyl group on Borels for SL2.

Note that the Bruhat decomposition even refines to

and gives a cell decomposition of ℬ into affine spaces (which are even dimensional). This allows us to easily compute the homology and cohomology of ℬ, which is given by

We can compute the dimension of each cell as follows. Note that

where U− is the unipotent radical of the opposite Borel. Now we compute

which is called the length of w relative to Φ+. For example, the identity has length 0 and the longest element w0 has length #⁢Φ+.

Example 2.78.

In the SLn case, the flag variety has dimension

The cohomology has dimension

Later, we will see that this is actually a representation of the Weyl group W and describe its algebra structure.

Remark 2.79.

In fact, a Borel subgroup B⊂G is a minimal subgroup of G such that the quotient G/B is projective.

Theorem 2.80 (Fixed point theorem of Borel).

A connected solvable linear algebraic group acting algebraically on a non-empty complete variety has a fixed point.

2.9 Parabolic subgroups.

If X=G/H is a homogeneous space which is compact, we can consider the action of a Borel subgroup B⊂G on X. By Borel’s fixed point theorem, this has a fixed point, say g⁢H. But then the stabilizer g⁢H⁢g−1 of this point contains B, and therefore

so H contains a Borel. Conversely, if H contains a Borel, then G/H admits a surjection from the flag variety G/B and is hence projective.

Definition 2.81.

A parabolic subgroup of G is a (necessarily connected) subgroup P⊂G which contains a Borel subgroup.

Remark 2.82.

We can drop connectedness from the definition because P is in fact always connected.

Proposition 2.83.

A subgroup P⊂G is parabolic if and only if G/P is projective.

We can in fact “classify” parabolic subgroups. If we fix a Borel B, then we can ask for all parabolic subgroups containing B. Because P is connected, we can reduce to the corresponding problem for Lie algebras, namely finding 𝔭⊃𝔟 such that [𝔭,𝔭]⊂𝔭. We may consider the adjoint action of 𝔥 on 𝔭, which decomposes as

The first two factors give 𝔟, so our goal is to determine the possible Γ. We will not discuss the details and only state the end result.

Proposition 2.84.

There is a one-to-one correspondence between 𝔭⊃𝔟 and subsets S⊂Π of the simple roots. It is given by

If we restrict the Cartan matrix to S, we obtain a root system

which decomposes into irreducible root systems. We can then decompose

where 𝔭u is the nilradical and ℓ is the Levi factor, then ℓss=ℓ/Z⁡(ℓ) has root system determined by S. The center is given by

In particular, we have

For example, in the SL3 example, if we choose S to be one simple root, then we see only roots living in the span of that simple root, which gives us a copy of 𝔰⁢𝔩2.

Example 2.85.

If 𝔤=𝔰⁢𝔩n and 𝔟=𝔟n is the upper-triangular matrices, then the parabolic subalgebras are simply block upper-triangular matrices and the Levi factor is given by block-diagonal matrices. The root system of the Levi is determined by the sizes of the blocks, and the center is given by scalar matrices in each block.

Remark 2.86.

The varieties G/P are known as partial flag varieties. In the G=SLn example, they are given by flags

where we require that dimVi/Vi−1=ni is the size of the i-th block.

2.10 Representation theory.

We will give a classification of irreducible representations of 𝔤.

Theorem 2.87.

Fixing 𝔥⊂𝔟⊂𝔤, the (isomorphism classes of) irreducible representations of 𝔤 are in one-to-one correspondence with dominant weights λ∈Λ, where

is the nonnegative integral span of the fundamental weights.

In the 𝔰⁢𝔩3 example, this corresponds to the shaded region in Figure 5.

Remark 2.88.

The case of groups is more subtle because of fundamental group issues. However, representations of 𝔤 are “the same” as representations of the simply connected form Gsc. This happens because we can integrate 𝔤→𝔤⁢𝔩⁢(V) to Gsc→GL⁡(V). In particular, this means that irreducible representations of Gsc are the same as positive coweights 𝕏∙⁢(Tsc)+=Λ+.

The way we do this is by considering the restriction to 𝔥 for any 𝔤-representation. These are all semisimple by preservation of Jordan decomposition. This gives a decomposition

so because we can integrate to Gsc, the 𝔥-action upgrades to a Tsc-action, and we have a decomposition

We can now consider the action of the unipotent radical U⊂B on V, which will not preserve the weight decomposition.

U has a nonzero fixed point because the B-action on ℙ⁡(V) has a fixed point by Borel’s fixed point theorem. If we consider the action of U on Vμ, then

where Q+ is the positive part of the root lattice. We will say that μ≤ν if ν−μ∈Q+. In this language, we have

Definition 2.89.

A highest weight vector vλ of V is a nonzero vector such that U⋅vλ=vλ.

Theorem 2.90.

Every irreducible representation V has a unique highest weight vector vλ up to scaling, and this λ is called the highest weight. Conversely, V is uniquely determined by its highest weight.

For an arbitrary finite-dimensional representation V, we can compute the decomposition

by considering the action of U on V. To decompose, we look for the fixed subspace VU, which is the span of the highest weight vectors. Finite-dimensional representations of B are classified by 𝕏∙⁢(Bsc/Bscu)=𝕏∙⁢(Tsc)=Λ. For all λ∈Λ, let ℂλ be the corresponding irreducible representation of B. Then the associated line bundle

on G/B is G-equivariant.

Theorem 2.91 (Borel-Weil).

The global sections Γ⁡(ℬ=G/B,ℒλ) are isomorphic to Vλ∗ as Gsc-representations.

Remark 2.92.

There is a generalization of this result, called the Borel-Weil-Bott theorem, which calculates the higher cohomology of these line bundles.

We will not give a detailed proof of the Borel-Weil theorem. However, the idea is that a section of ℒλ is a pair (g,f⁡(g)⋅1) where f:Gsc→ℂ (here, we always use the simply connected form and drop the subscripts), and we must be invariant under the B-action

and therefore we must have

for all g∈G and b∈B. Equivalently, we can ask for

Then we are looking for f∈𝒪⁢(G)(B,λ)≅Γ⁡(ℬ,ℒλ). To identify this representation, we need to find the highest weight with respect to B or B−. It is a bit easier to work with B−, so we can pass to U−. Then we consider

Using the Bruhat decomposition

we will use the longest element w0 (for SLn this is the permutation which exchanges 1 and n, 2 and n−1, etc.) to write

Then B−⁢B⊂G is open, so we can consider the behavior of functions on this open. This decomposes as the Cartesian product

so we are looking for functions in 𝒪⁢(U−⁢T⁢U)U−×U,(T,λ)=𝒪⁢(T)(T,λ). This gives us the 1-dimensional space ℂ⋅yλ, so the highest weight space we are trying to calculate is at most 1-dimensional. It remains to show that yλ has no poles outside of B−⁢B. Using the Bruhat decomposition again, the codimension 1 orbits with respect to B−×B are given by B−⁢s˙i⁢B for simple reflections si.

Lemma 2.93.

For all simple coroots αi∨, let ωi be the dual fundamental weight. Then div⁡(yωi)=B−⁢s˙i⁢B¯.

Now if λ∈Λ+, we see that yλ is globally defined and that

If λ∉Λ+, then

with some of the ai being negative, so yλ has a pole along Di of order |ai|. Thus, B− acts on 𝒪⁢(G)(B,λ) with weight −λ, and the end result is that the representation must be Vλ∗.

Remark 2.94.

Note that Pic⁡(G)=0 because of the Bruhat decomposition

and therefore the the exact sequence

has the first object given entirely by principal divisors Di=B−⁢s˙i⁢B¯. Equivalently, we see that 𝒪⁡(G) is a UFD.

However, if we move away from the simply-connected form, then there are nontrivial line bundles, and

Theorem 2.95 (Peter-Weyl).

Consider the action of Gsc×Gsc on 𝒪⁡(Gsc). Then we have a decomposition

We may obtain this decomposition by considering U−×U-fixed vectors and obtaining

which has weight −λ for B− in the first factor and weight λ for B in the second factor. In the other direction, we associate the tensor v1∗⊗v2 with the function

3 Springer theory

3.1 The Grothendieck-Springer resolution.

Proposition 3.1.

Let 𝔟 and 𝔟′ be two Borel subalgebras of 𝔤. Then there is a canonical isomorphism 𝔟/[𝔟,𝔟]→∼𝔟′/[𝔟′,𝔟′].

Proof.

Because all Borels are conjugate, we have

This isomorphism preserves the nilradical, so it induces an isomorphism on the quotients. For a different g′ that conjugates 𝔟 to 𝔟′, we have g′=g⁢b for some b∈B, and it remains to check that

is the same as

Passing to the infinitesimal version, the action of 𝔟 on 𝔟/[𝔟,𝔟] is zero because the adjoint action of 𝔟 on itself sends everything to the nilradical, so it integrates to the trivial action of B. □

Definition 3.2.

The quotient 𝔟/[𝔟,𝔟] is called the universal Cartan of 𝔤 and is denoted by ℌ.

Remark 3.3.

We must note that ℌ is not a subalgebra of 𝔤. Instead, we should view it as a pair (𝔥,𝔟⊂𝔥). Indeed, for all such pairs, we can identify 𝔥 and ℌ by

If 𝔥⊂𝔟′ is a different Borel, then 𝔟′=w⁡(𝔟)=w˙⋅𝔟, and therefore relating the sequences via

requires conjugation by w˙ and thus requires the choice of a Borel. However, we do end up with a pair (ℌ,Φℌ+), which gives a universal set of positive roots containing a universal set Πℌ of simple roots. We will call this a “pinned” version of the usual Cartan. This allows us to discuss the Weyl group 𝕎 and the universal simple reflections. Here, we note that any choice of 𝔥⊂𝔟 gives a canonical isomorphism

Lemma 3.4.

For any Borel subalgebras 𝔟1 and 𝔟2 there exists a Cartan subalgebra 𝔥⊂𝔟1∩𝔟2.

Proof.

Using the Bruhat decomposition relative to B1⊃T1, we write

There exists some u1∈U1 and w∈WT1 such that

We consider u1−1⋅𝔟2=w˙⋅𝔟1 and u1−1⋅𝔟1=𝔟1. In particular, we see that

Applying u1 to this inclusion gives us the desired result. □

Remark 3.5.

If we consider the set of pairs {(𝔥,𝔟)}, this is actually isomorphic to G/T. Forgetting 𝔥 corresponds to the map G/T→G/B, and the fiber over the set

Definition 3.6.

Define the relative position of two Borel subalgebras 𝔟1 and 𝔟2 as follows. Choose a Cartan 𝔥⊂𝔟1∩𝔟2. The pair (𝔥,𝔟1) determines a choice of positive roots, and we also know that 𝔟2=w˙⋅𝔟1. We then have canonical identifications (𝔥,𝔟1)≅(ℌ,Φℌ+) and WT≃𝕎. Therefore, we have

for a unique w∈𝕎.

The reason this is well defined is because for any two 𝔥1,𝔥2⊂𝔟1∩𝔟2, we have 𝔥2=b⋅𝔥1 for some b∈B1∩B2. This implies that in the diagram

we actually have w˙2=Adb⁡w˙1, so they are sent to the same w∈𝕎.

In practice, if we fix a pair (𝔥,𝔟), then given 𝔟1 and 𝔟2, we first write 𝔟1=u⁢w˙⋅𝔟, then apply

and then compute which Bruhat cell (u⁢w˙)−1⋅𝔟 lies in.

Remark 3.7.
  1. (1)

    If we consider the action of GΔ on ℬ×ℬ, then the orbits are canonically indexed by 𝕎.

  2. (2)

    If we have 𝔟1→w1𝔟2→w2𝔟3, the relative position of 𝔟3 with respect to 𝔟1 is not determined by w1 and w2. However, if ℓ⁡(w1)+ℓ⁡(w2)=ℓ⁡(w1⁢w2), then the relative position is indeed w1⁢w2.

Example 3.8.

If G=SL2, then ℬ×ℬ=ℙ1×ℙ1. The GΔ-action decomposes this into

and the diagonal corresponds to the identity element (so 𝔟1=𝔟2), while the other element corresponds to the large open set. If we consider

then we can get either 1 or s depending on whether 𝔟1=𝔟3.

In general, suppose 𝔟1,𝔟2,𝔟3⊃𝔥 and

Suppose that 𝔟1=𝔟. Applying w˙−1 to the second arrow, we see that

which tells us that w˙1−1⁢𝔟3=w˙2⁢𝔟 and hence that 𝔟3=w˙1⁢w˙2⁢𝔟.

Definition 3.9.

The Grothendieck-Springer resolution is the incidence variety

This variety admits two maps

which are both G-equivariant. Over ℬ, the fibers are simply given by 𝔟, so we have the tautological vector bundle over 𝔟. Therefore, we have a morphism

over G/B≅ℬ.

The fiber

which is the zero locus of the vector field on ℬ corresponding to x is called the Springer fiber. Later, we will restrict to x in the nilpotent cone. The first extreme case is if x=0, in which case μ−1⁢(0)=ℬ. On the other hand, if x∈𝔤ssreg, then x∈𝔥reg for some Cartan, so 𝔥⊂𝔟 for any 𝔟∋x, and thus the Springer fiber is simply

which is a torsor for the Weyl group. In particular, 𝔤~rs is actually a 𝕎-torsor over 𝔤rs.

Proposition 3.10.

The map μ is proper.

Proof.

Consider the commutative diagram

and use the fact that the horizontal arrow is a closed immersion and the fact that proper morphisms are closed under composition. □

Now consider the morphism

Note that if we consider 𝔤 only, then the target must be 𝔥⫽W because there is no preferred order of the eigenvalues (in the 𝔰⁢𝔩n case, the others are similar). We now have the correspondence

Theorem 3.11 (Chevalley restriction theorem).

Fix a Cartan 𝔥⊂𝔤. Then there is the restriction of functions 𝒪⁡(𝔤)→𝒪⁡(𝔥), which induces

This φ is an isomorphism.

Proof.

We begin with injectivity. If P∈𝒪⁢(𝔤)G satisfies P|𝔥=0, then P|𝔤s=0. This is dense, so P=0.

Now let P∈𝒪⁢(𝔥)W. Our goal is to find P~∈𝒪⁢(𝔤)G such that P~|𝔥=P. Note that P determines P~ on 𝔤rs, but we need to ensure that P~ extends to all of 𝔤 (since the boundary has codimension 1).

First, using the correspondence Equation 1, we have Pℌ∈𝒪⁡(ℌ) after we choose some 𝔟⊃𝔥. Because P is W-invariant, this in fact does not depend on the choice of 𝔟. We now have a function P~=ν∗⁢Pℌ on 𝔤~, which is G-invariant because

Our goal now is to descend it to 𝔤. Let x∈𝔤rs. Note that 𝔤~rs→𝔤rs is a 𝕎-torsor and thus 𝔤rs=𝔤~rs/𝕎. In particular, we see that

Note that P~|𝔤~rs is 𝕎-invariant, so it descends to R∈𝒪⁢(𝔤rs)G. Moving to the analytic topology, we want to show that for all x∈𝔤∖𝔤rs and any small ball D∋x, then R|D∩𝔤rs is bounded. Note that μ−1⁢(D¯) is compact because μ is proper. Note that P~|μ−1⁢(D¯)∩𝔤~rs is bounded, so R|D¯∩𝔤rs is bounded and thus extends to a holomorphic function on D.

Finally, we need to check that R|𝔥=P. For any x∈𝔥 and any choice of 𝔟∋x, we see that

Example 3.12.

For 𝔤=𝔰⁢𝔩n, we can write

and 𝒪⁢(𝔤)G=ℂ⁡[a2,…,an]. Transported to 𝔥, these are simply the elementary symmetric polynomials in the eigenvalues.

Note also that the Chevalley restriction theorem gives a canonical isomorphism 𝒪⁢(𝔤)G≃𝒪⁢(ℌ)𝕎. Later, we will see that 𝒪⁢(ℌ)𝕎 is a polynomial ring in r variables, so ℌ⫽𝕎 is an affine space of dimension r. The embedding 𝒪⁢(ℌ)𝕎↪𝒪⁡(𝔤) gives the adjoint quotient map

Remark 3.13.

We can in fact choose homogeneous generators of degrees d1≤d2≤⋯≤dr, and we will have d1=2 and dr being the Coxeter number h.

Proposition 3.14.

The diagram

commutes.

Proof.

It suffices to check that

because ℌ⫽𝕎 is affine. This is true because

essentially by definition. □

Proposition 3.15.

Let 𝔟∈ℬ be a Borel, 𝔫=[𝔟,𝔟], and x∈𝔟. Then for all P∈𝒪⁢(𝔤)G, the restriction P|x+𝔫 is constant.

Proof.

For all n∈𝔫, we have

Remark 3.16.

If we use the Jordan decomposition, then in fact P⁡(x)=P⁡(xs) for any P∈𝒪⁢(𝔤)G because P⁡(x)=Pℌ⁢(xs).

3.2 Nilpotent cone.

Definition 3.17.

The set of nilpotent elements

is closed, G-stable, and ℂ×-stable and is called the nilpotent cone.

Recalling the Grothendieck-Springer resolution μ:𝔤~→𝔤, we define 𝒩~=μ−1⁢(𝒩). Now we can consider the diagram

We see here that

Note in fact that the condition of being nilpotent is equivalent to P⁡(x)=0 for all P∈𝒪⁢(𝔤)G. This identifies 𝒩~ as a G-equivariant vector bundle over ℬ with fiber 𝔫, and in particular that

Under this presentation, we see that μ⁡(g,x)=g⋅x.

Theorem 3.18.

There is a natural identification 𝒩~≅T∗⁢ℬ.

Proof.

Note that if 𝔟∈ℬ, we have T𝔟⁢ℬ=𝔤/𝔟, so the cotangent space T𝔟⁢ℬ=𝔟⟂⊂𝔤∗. Using the Killing form, we can identify 𝔤∗≅𝔤, and under this identification, 𝔟⟂=𝔫. Therefore, we have a canonical isomorphism T∗⁢ℬ≅G×B𝔟⟂≅G×B𝔫≅𝒩~. □

Later, we will see that μ𝒩 is the moment map for the G-action on T∗⁢ℬ induced by the G-action on ℬ. Let 𝒪⁢(𝔤)+G⊂𝒪⁢(𝔤)G be the functions which vanish at the origin.

Proposition 3.19.

x∈𝒩 if and only if P⁡(x)=0 for all P∈𝒪⁢(𝔤)+G.

Corollary 3.20.

The nilpotent cone 𝒩 is an irreducible variety of dimension 2⁢dim𝔫.

Proof.

The map μ𝒩 is proper and surjective, and 𝒩~≅T∗⁢ℬ is irreducible. We also see that dim𝒩≤dim𝒩~=2⁢dim𝔫. On the other hand, we know 𝒪⁢(𝔤)+G=𝒪⁢(ℌ)+𝕎=⟨f1,…,fr⟩, and the morphism χ:𝔤→ℌ⫽𝕎 expresses 𝒩=χ−1⁢(0) and shows that dim𝒩≥dim𝔤−r=2⁢dim𝔫. □

Proposition 3.21.

There are finitely many G-orbits in 𝒩.

Idea of proof.

This is easy to see for 𝔤=𝔤⁢𝔩n via the Jordan normal form. In the general case, we choose a faithful representation 𝔤↪𝔤⁢𝔩n, which embeds nilpotent cones by preservation of Jordan decomposition. The following lemma finishes the proof using the fact that 𝔤 is semisimple (so it has a 𝔤-stable complement in 𝔤⁢𝔩n). □

Lemma 3.22.

Let 𝔤⊃𝔞 such that 𝔞 has an 𝔞-stable complement W⊂𝔤. Let G and A be connected algebraic groups with Lie algebras 𝔤 and 𝔞. For all G-orbits Ω in 𝔤, any irreducible component of Ω∩𝔞 is a single A-orbit.

Proof.

Let Z⊂Ω∩𝔞 be any irreducible component. For any x∈Z, we have

This implies that Tx⁢Z=[x,𝔞]=Tx⁢(A⋅x), so A⋅x is open and dense in Z. This implies that for any x1,x2∈Z, A⋅x1∩A⋅x2≠∅, so the two orbits coincide. □

Corollary 3.23.

The nilpotent cone 𝒩 contains an open dense orbit, which is exactly

Proof.

For all x∈𝒩reg, we want to check that dimG⋅x=dim𝒩. This follows from the fact that

Example 3.24.

In the case of 𝔰⁢𝔩2, we see that

See Figure 9 for a visualization of this variety. We can interpret 𝒩~=T∗⁢ℙ1 as

and therefore anything in the open part is similar to [0100] and clearly has a one-dimensional kernel, while the origin has a two-dimensional kernel and thus its fiber is all of ℙ1.

0
Figure 9. Nilpotent cone of 𝔰⁢𝔩2

Now fix a B⊃T. This gives an action of T on 𝔫/[𝔫,𝔫]≅⨁αi∈Π𝔤αi. Choose a basis ei∈𝔤αi for 1≤i≤r. Define the subset

Proposition 3.25.

The set 𝔫r is a single B-orbit consisting of regular nilpotent elements.

Proof.

Note that 𝔫r is open and dense in 𝔫 (here, 𝒩reg∩𝔫 must be open and dense in 𝔫), so 𝔫r must contain a regular element x. The image of 𝔫r in 𝔫/[𝔫,𝔫] is a single T-orbit, so it remains to show that x+[𝔫,𝔫] is a single U-orbit. On the level of tangent spaces, we see that

and that this holds on all of x+[𝔫,𝔫]. This tangent space has dimension

and therefore U⋅x⊂x+[𝔫,𝔫] is open and dense. However, the fact that U is unipotent and x+[𝔫,𝔫] is affine implies that it is also closed, so U⋅x=x+[𝔫,𝔫]. □

Proposition 3.26.

Every regular nilpotent element is contained in a single Borel.

In particular, this means that the morphism μ:𝒩~→𝒩 is an isomorphism over the regular locus.

Proof.

Fix a regular nilpotent element

where each 0≠ei∈𝔤αi. Because this consists entirely of elements from the simple root spaces, there exists h∈𝔥 such that [h,e]=2⁢e. This h is determined uniquely by the condition αi⁢(h)=2 for every simple root αi∈Π. This implies that for all positive roots α∈Φ+, α⁡(h)≠0 and thus h∈hreg∩𝕏∙⁢(T).

Note that this h determines a one-parameter subgroup i:ℂ×→T. The adjoint action is given by Adt⁡(e)=t2⁢e and lifts to an action of ℂ× on μ−1⁢(e). For all 𝔟∈μ−1⁢(e), we see that

and because μ−1⁢(e) is finite, the action is trivial. In particular, this implies that 𝔟 is a fixed point of the action of ℂ× on ℬ induced by i. This happens if and only if h∈𝔟, and because h is regular, we see that 𝔥⊂𝔟. The set of Borels satisfying this form a W-torsor, but any 𝔟 containing e must contain all simple root spaces 𝔤αi, and there exists a unique Borel satisfying this condition. □

In particular, we see that 𝒩~→𝒩 is a resolution of singularities, called the Springer resolution.

Remark 3.27.

If x∈𝒩reg, then μ−1⁢(x) is a singleton. On the other hand, if x∈𝔤rs, then μ−1⁢(x) is a 𝕎-torsor. In general if x∈𝔤reg, then μ−1⁢(x) is finite. More generally, the map 𝔤→𝔥⫽W sends x↦[xs] (after conjugating x to be in 𝔥). In particular, the set μ−1⁢(x) can be identified with the set 𝕎/Stab𝕎⁡(x).

3.3 Jacobson-Morozov theorem.

The idea is that for every nonzero e∈𝒩, we can find an 𝔰⁢𝔩2-triple (e,h,f).

Theorem 3.28.

Let 0≠e∈𝒩. Then there exist h,f∈𝔤 such that h is semisimple, f is nilpotent, [h,e]=2⁢e, [h,f]=−2⁢f, and [e,f]=h.

Note that in 𝔰⁢𝔩2, the matrices

satisfy these relations. For a general 𝔤, these give an embedding 𝔰⁢𝔩2↪𝔤 sending the standard basis to (e,h,f). Also, in the general case, e+ker⁡adf=e+𝔤f is a transverse slice to G⋅e, and if e is regular, it is called the principal slice or Kostant slice and satisfies e+𝔤f=𝔥⫽W.

Corollary 3.29.

For all e∈𝒩, there exists a rational group homomorphism SL2→G sending [0100] to e.

Proof of Theorem 3.28.

Recall the Killing form

The fact that this is G-invariant means that

To simplify things, we will write Gx=ZG⁢(x) and 𝔤x=ℨ𝔤⁢(x)=ker⁡adx for x∈𝔤.

Lemma 3.30.

Assume that e∈𝒩 and 𝔤e consists entirely of nilpotent elements. Then (e,𝔤e)=0.

Proof.

We need to show that for any x∈𝔤e, the trace tr⁡(ade⁡adx)=0. However, both ade and adx are nilpotent, and they commute, so their product is also nilpotent. □

Our goal now is to reduce to the case where 𝔤e⊂𝒩 so that we can induct on the rank of 𝔤. Assume that if the rank is strictly less than r, then the theorem holds. For all e∈𝒩, we consider the set 𝔤e. If 𝔤e⊄𝒩, then there exists x∈𝔤e which is not nilpotent. Writing x=xs+xn, we see that [e,x]=0 implies that both [e,xs]=0 and [e,xn]=0, so e∈𝔤xs∩𝔤xn. Because xs is semisimple and nonzero, 𝔤xs⊂𝔤 is a proper Levi subalgebra and is hence reductive. Because e is nilpotent, we have

which is semisimple and has strictly smaller rank than 𝔤.

In the case where 𝔤e⊂𝒩, we use the ade-invariance of the Killing form to obtain

and therefore that Im⁢ade⊂(ker⁡ade)⟂=(𝔤e)⟂. Because the Killing form is nondegenerate, the two are actually equal. This implies that e∈Im⁢ade and thus that there exists h∈𝔤 such that [h,e]=2⁢e. Decomposing h=s+n into a semisimple and nilpotent part, we see that [s,e]=2⁢e. Therefore, we can replace h by s.

Our final goal is to find f. Consider the weight decomposition of adh:𝔤→𝔤. This decomposes

Here, note that h∈𝔤0 and e∈𝔤2, and because [h,e]=2⁢e, we see

Here, the calculation is

We want to find f∈𝔤−2 such that [e,f]=h, or equivalently, show that h∈Im⁢ade=(𝔤e)⟂. Because [h,e]=2⁢e and [h,𝔤e]⊂𝔤e, we see that the subspace ℂ⁢h+𝔤e is a solvable subalgebra in 𝔤. By the theorem of Lie and Engel, there exists a basis of 𝔤 such that adh is upper-triangular and for any x∈𝔤e, adx is strictly upper-triangular. This implies that

for any x∈𝔤e because adh⁡adx is strictly upper-triangular. This implies that h∈(𝔤e)⟂=Im⁢ade, so there exists f such that [e,f]=h. Because h∈𝔤0, we can choose f∈𝔤−2. □

Proposition 3.31.

All 𝔰⁢𝔩2-triples containing e∈𝒩 are conjugate under Ge≔ZG⁢(e).

For a fixed triple (e,h,f), we can consider the induced representation of 𝔰⁢𝔩2 from any representation of 𝔤. Recall that representations of 𝔰⁢𝔩2 are classified by their highest weight, which is a positive integer n. See Figure 10 for a visualization. For non-irreducible representations, we can draw a diagram with all of the irreducible components.

−n−n+2−n+4⋯n−4n−2neeeeeeffffff
Figure 10. Irreducible 𝔰⁢𝔩2-representation Vn with highest weight n.

Specifying to V=𝔤, we see that 𝔤=𝔤f⊕Imade. If we decompose

note that [gk,𝔤ℓ]⊂𝔤k+ℓ.

Remark 3.32.

Note that dim𝔤e=dim𝔤f≥dim𝔤h=dim𝔤0 and equality holds if and only if all weights are even.

If e is regular, then dim𝔤e=r, and therefore f is regular and dim𝔤h=dim𝔤e=r, and therefore e,f,h are all regular. In this case, we have a principal 𝔰⁢𝔩2-triple. Note here that all weights are even. Also, note that

but is not necessarily contained in the strictly positive part. By definition, set

By definition, we have Vtriv=ℨ⁡(e,f,h)⊂𝔤, and if there is no trivial part under 𝔰⁢𝔩2, then 𝔤e is contained in the strictly positive part. Now lift 𝔲 to a unipotent subgroup U⊂Ge. Note that Ge may be disconnected, but U is connected.

Lemma 3.33.

The affine space h+𝔲 is a single U-orbit.

Proof.

First, note that if x∈𝔲, then [h+x,𝔲]⊂𝔲 because 𝔲 consists of weight spaces. In fact, they are actually equal because [h,𝔲]=𝔲. By a standard argument, the orbit U⋅h⊂h+𝔲 is open, but U is unipotent, so it is also closed. □

Proof of Proposition 3.31.

Fix (e,h,f) and (e,h′,f′). Note that [h′−h,e]=0, so h′−h∈𝔤e. In particular, we see that h′−h∈𝔤e∩[𝔤,e] because we also have [e,f′−f]=h′−h. Therefore, h′∈h+𝔲=U⋅h, so there exists u∈U such that u⋅h=h′.

Now we have u⋅(e,h,f)=(e,h′,u⋅f). Note that

and therefore [e,u⁢f−f′]=0. This implies that u⁢f−f′∈𝔤e, but u⁢f−f′∈𝔤−2 while 𝔤e is concentrated in nonnegative weights, so u⁢f=f′. Therefore, u⋅(e,h,f)=(e,h′,f′). □

Remark 3.34.

We have in fact also shown that if we fix e and h, then f is unique.

We will now consider the structure of Ge.

Proposition 3.35.

The unipotent subgroup U is the unipotent radical of Ge and H≔ZG⁢(e,h,f) is the maximal reductive subgroup of Ge.

Example 3.36.

If e is regular, then so are h and f. When we described 𝒩reg, recall that 𝔫reg consisted of all elements which are nonzero on all simple root spaces. If we let

thhen h is actually unique is given by

while f is lower-triangular with nonzero entries on the subdiagonal. In this case, we have ZG⁢(e,h,f)=Z⁡(G). For all g∈ZG⁢(e,h,f), we see that g∈ZG⁢(h)=T, and so under some choice of Borel 𝔟 containing h∈𝔥reg, we see that

and therefore if g∈ZG⁢(e), this is equivalent to α⁡(g)=1 for all simple roots α, which is equivalent to g∈Z⁡(G).

Because e is regular, we see that dimGe=r. Using the proposition, we see that dimU=r, and therefore that 𝔤e=𝔲. Here, we see that

Looking at the decomposition of 𝔤 with respect to this h, then in fact

or namely all weights are even. There is no trivial 𝔰⁢𝔩2 representation appearing as a summand of 𝔤 because ZG⁢(e,h,f) is discrete.

Proof of Proposition 3.35.

Our goal is to show that Ge=H⁢U and that H∩U is trivial. We already know that U⊲Ge because 𝔲 is an ideal in 𝔤e.

We will use the fact that the centralizer of a reductive subgroup in G is always reductive. Because the image of SL2 in G given by e,h,f is reductive, so is H. We also know that H∩U⊲H, and because H is reductive, the intersection is trivial.

Finally, we need to prove that Ge=H⁢U. For any g∈Ge, recall that

for some u∈U. Therefore, we see that u−1⁢(g)∈ZG⁢(e,h,f) and therefore g∈U⁢H=H⁢U (here, we use normality of U). □

3.4 Regular 𝔰⁢𝔩2-triples.

Let e∈𝒩reg and consider the Kostant slice e+𝔤f. Recall that the map χ:𝔤→𝔥⫽W restricts to an isomorphism

Note that 𝒪⁢(𝔤)G≃𝒪⁢(𝔥)W is an isomorphism of graded rings.

Theorem 3.37 (Chevalley-Shephard-Todd).

We have 𝒪⁢(𝔥)W≃ℂ⁡[P1,…,Pr], where P1,…,Pr are homogeneous elements of degrees d1≤⋯≤dr which are independent of the choice of generators. These d1,…,dr are called the exponents of 𝔤.

Theorem 3.38 (Kostant).

Let (e,h,f) be a regular 𝔰⁢𝔩2-triple. Recall that the decomposition

into irreducible 𝔰⁢𝔩2-representations Vj satisfies dimVj=2⁢λj+1, where λj>0. Order Vj such that 1=λ1≤⋯≤λk. Then k=r and dj=λj+1.

Proof.

Consider the decomposition

For each Vj, the lowest weight is −2⁢λj. Because 𝔤f∩Vj is the lowest weight space of Vj, we can let the set of 0≠ξj∈𝔤f∩Vj be a basis of 𝔤f. Define a map

We will show that dψ is surjective everywhere, or namely that ψ is a submersion. Because this morphism is G-equivariant, it suffices to check when g=1. Let x=(1,a1,…,ar). Then

Our goal is to prove that this is surjective. We compute

Now consider the ℂ×-action on e+𝔤f, which is given by

Note that the map ψ is equivariant with respect to this action, where ℂ× acts on ℂr by ai↦t−2⁢λi−2⁢ai. Note that the t→∞ limit contracts everything to the origin, so it was enough to check that we had a surjection there. Because the differential is surjective, G⋅(e+𝔤f)=Im⁢ψ is open and dense in 𝔤, and therefore the restriction

is injective.

For a generator Pj∈ℂ⁢[𝔤]G, let Rj∈ℂ⁡[a1,…,ar] be given by

Because P1,…,Pr are algebraically independent, so are R1,…,Rr. We now use the Euler identity, which says that for all homogeneous P∈ℂ⁡[x1,…,xr], we have

Denote this vector field by Z. Applying this to the Pj on 𝔤, let u=Zu denote the tangent vector at Tu⁢𝔤. Clearly, we have

Now let

Then we see that

The polynomial Rj is a sum of monomials of the form a1m1⁢j⁢⋯⁢akmk⁢j, and we have the identity

for every monomial in Rj.

We now want to show that each dj≥1+λj. Suppose there exists j0 such that dj0<1+λj0. Then for every i≥j0, we must have mi⁢j0=0. On the other hand, for all j<j0, we have dj≤1+λj0 while mi⁢j=0 for j≤j0. This implies that R1,…,Rj0 depend only on a1,…,aj0−1, so they cannot be algebraically independent, which is a contradiction.

Finally, we need to show that dj≤1+λj. We will follow Kostant’s proof.

Proposition 3.39.

Let 𝔤 be semisimple of rank r and d1,…,dr be the exponents of 𝔤. Then

Assuming the proposition, we see that

so we must actually have an equality everywhere. □

Proposition 3.40.

The morphism

is an isomorphism.

2 Universal centralizers

4 Defining the universal centralizer

Definition 4.1.

Define the universal centralizer to be

Note that this has a morphism to ℂr given by forgetting g and applying ρ. This map is flat, and the general fiber is a maximal torus.

This universal centralizer admits a symplectic structure. To see this, we will realize it as a Hamiltonian reduction.

Definition 4.2.

A symplectic manifold (M2⁢n,ω) is a smooth manifold of dimension 2⁢n equipped with a closed non-degenerate 2-form ω.

Note that ω induces an isomorphism T⁢M≃T∗⁢M.

Example 4.3.

The most basic example is the cotangent bundle

where (q1,…,qn) are local coordinates on X and p1,…,pn are the corresponding covectors in the fiber.

Example 4.4.

A Kähler manifold M has a symplectic form ω given by the imaginary part of the Kähler metric. For example, we can consider ℙℂn with the Fubini-Study metric.

We will consider the action of G on T∗⁢X. If there is an action of G on X, then there is an induced action of G on T∗⁢X given by

Exercise 4.5.

Check that this action preserves the symplectic form.

If we consider the infinitesimal action, then for every η∈𝔤 there is a vector field vη on T∗⁢X. We then consider the 1-form

for some Hamiltonian function Hη. The moment map μ:T∗⁢X→𝔤∗ is defined by the relation

Note that this Hη is only defined up to a constant, but can be fixed by requiring that 𝒪⁡(𝔤∗→C∞⁢(T∗⁢X)) is a Poisson map. In particular, the formula is given by setting

Example 4.6.

Consider the left and right actions of G on T∗⁢G. The right action gives an isomorphism T∗⁢G≃G×𝔤∗ given by

We can now try to compute the moment maps. The moment map for the left action is given by

and the moment map for the right action is given by

To see this, note that

for the left action.

Proposition 4.7.

The universal centralizer JG is isomorphic to a Hamiltonian reduction of T∗⁢G.

Here, we can consider the left and right actions of the unipotent N⊂G. We then consider

We can now compose this to 𝔫∗×𝔫∗ by using the Killing form to identify 𝔫∗≅(𝔟⟂)∗=𝔤/𝔟. Now consider the inclusion

and recall that 𝔫/[𝔫,𝔫] is the sum of the simple root spaces. Dualizing, this is an injection

Now choose f∈⨁α∈Π𝔤−α. This is invariant under the adjoint action of N, and therefore

In general, we can consider a coadjoint orbit 𝒪 and the Hamiltonian reduction μ−1⁢(𝒪)/G. In this case, we are considering

Proposition 4.8.
  1. (1)

    f+𝔟⊂𝔤reg;

  2. (2)

    If 𝒮=f+𝔤e is the Kostant slice, then Ad:N×𝒮≃f+𝔟.

Proof.
  1. (1)

    We will prove that e+𝔟−⊂𝔤reg. Write the decomposition

    with respect to the 𝔰⁢𝔩2-triple (e,h,f). Now consider the ascending filtration

    Let x=e+v∈e+𝔟−. We see that v∈𝔟−=𝔤≤0 preserves the filtration, so 𝔤x has an induced filtration

    We will show that the image of 𝔤≤jx/𝔤≤j−2x⊂𝔤j lies in 𝔤e. Choose any element y∈𝔤≤jx. We can write y=yj+y′ where yj∈𝔤j and y′∈𝔤≤j−2. Then we see that

    This implies that [yj,e]=0. Now we compute

    and therefore x is regular.

  2. (2)

    We will not give a full proof of this. The idea is that the tangent map d(1,e)⁢α of the opposite version α:N−×(e+𝔤f)→e+𝔟− lands inside 𝔟− and is given by (n,x)↦[n,e]+x. The rest of the proof is to define a ℂ×-action on both sides which contracts everything to a point and such that α is equivariant. On the right, we define

    where ξ=e+x. Note that x∈𝔟−, so t−1 acts on it with non-negative weights and thus the action has positive weights. On the left, note that N− is generated by the negative root groups. Then Adγ⁡(t−1) has positive weight on it, so clearly both factors are contracted to the identity. The proof concludes with the following lemma. □

Lemma 4.9.

Let X1 and X2 be smooth irreducible affine varieties with contracting ℂ×-actions. Let α:X1→X2 be ℂ×-equivariant such that the tangent map at the fixed point is an isomorphism. Then α is an isomorphism.

Theorem 4.10.

The composition e+𝔤f↪𝔤reg→𝔤reg/G is a bijection.

Proof.

Choose a set of homogeneous generators P1,…,Pr of 𝒪⁢(𝔤)G. Then we have a diagram

We know that ρ is an isomorphism. We need to prove that ρ¯ is injective. To do this, we will show that P1,…,Pr separate G-orbits on 𝔤reg. Suppose that x,y∈𝔤reg have the same semisimple part. We claim that they lie in the same G-orbit. This is because their nilpotent parts n and n′ must lie in (𝔤s)reg=(𝔤s,der)reg, so they are in the same nilpotent orbit in 𝔤s,der. Because of this, we can see that P⁡(s+n)=P⁡(s). This happens because for any nonzero constant, s+c⁢n is conjugate to s+n. Therefore, we see that P⁡(s+c⁢n)=P⁡(s+n), and taking the c→0 limit gives us P⁡(s)=P⁡(s+n).

It now suffices to show that 𝒪⁢(𝔥)W separates W-orbits. Let s,t∈𝔥 and W⁢s∩W⁢t=∅. We claim that there exists P∈𝒪⁢(𝔥)W such that P|W⋅s=0 and P|W⋅t=1. To see this, we start with R∈𝒪⁡(𝔥) such that R⁡(w⋅s)=0 and R⁡(w⋅t)=1 for all w∈W and the apply the averaging trick to obtain

We will now return to the problem of considering the moment map

Recall here that fmod𝔟∈𝔤/𝔟≅𝔫∗ descends to f:𝔫/[𝔫,𝔫]→ℂ and that we want to consider μN×N−1⁢(f,f)/N×N, which is called the bi-Whittaker reduction N\fT∗⁢G/fN. Here, we have

Because f+𝔟≃N×(f+𝔤e), we can kill the N factors and make the identification

4.1 Bruhat Decomposition.

Recall that JG=N\fT∗⁢G/fN and thus has a map to N\G/N, which has the Bruhat decomposition

It is natural to ask what the preimages are.

Proposition 4.11.
  1. (1)

    w appears in the image of 𝒷:JG→N\G/N if and only if w=w0⁢wS for some S⊂Π, where w0 is the longest element and ws∈wS=⟨si|i∈S⟩.

    Denote the Levi subgroup associated to S by LS. This can be obtained by integrating

    Alternatively, choose a generic element t∈S⟂=ℨS, and then LS=ZG⁢(t).

  2. (2)

    Let ℬS=𝒷−1⁢(w˙0⁢w˙S⁢T). Then

Note that the second factor is an affine space 𝔸|S| and Z⁡(LS) is a torus which may not be connected.

Example 4.12.

Let G=SL2 and make the standard choice of e,h,f. Then let

Then the general fiber over e+𝔤f≅𝔸1 is ℂ× but the central fiber has two components given by

Away from the central fiber (or in general over the regular locus), we have

There is a unique simple root α, and so we see that

which is the union of Kostant sections. Note that this holds generally for any semisimple G.

In general, the complement of the Kostant sections is ℬ∅≅T∗⁢T.

Remark 4.13.

If G=GLn, then JG⊂Hilbn⁡(ℂ××ℂ) is open. Note that this contains (T×𝔥reg)/W as another open subset.

We will now begin the proof of the proposition. Let w=w0⁢ws. To simplify the proof, consider a torsor of JG over 𝔠=f+𝔤e given by

Now our Bruhat decomposition is with respect to N−\G/N, and we can get rid of the w0 everywhere.

Up to the action of N−×N, we can always reduce g=w˙⁢h for some representatives w˙∈NG⁢(T) and h∈T. We need to determine for which w˙⁢h there exists ξ∈f+𝔟 such that Adw˙⁢h⁡(ξ)∈e+𝔟−. This requires that w⁡(−Π)⊂Π∪Φ−, so let S=w⁡(−Π)∩Π. Then w must send −Π∖w−1⁢(S) to Φ−. Also, we see that wS⁢w sends w−1⁢(S) to wS⁢(S)=−S and thus we have

for all w∈WS. This implies that wS⁢w⁢(−Π)⊂Φ− and therefore wS⁢w=1. Because wS2=1, we see that w=wS. In summary, if S=w⁡(−Π)∩Π, then w=wS, so we have proved that if w∈Im⁡(𝒷), then w=wS.

Now fix some S⊂Π and let g=w˙S⁢h for some h∈T. Write f+𝔟∋ξ=fS+fΠ∖S+t+ξ+, where fS=∑α∈Sfα, fΠ∖S=∑β∉Sfβ, t∈𝔥, and ξ+∈𝔫. We need to show that

Note here that fS↦Adw˙S⁢h⁡(fS), which we need to equal eS, fΠ∖S is sent to an element of 𝔟−, Adw˙S⁢h⁡(t)∈𝔟−, and ξS is sent to 𝔤ΦS−⊂𝔟−, so we need ξ+′↦eΠ∖S. This implies that

In summary, we need h∈T to satisfy

Note that there are some “free variables” like t∈𝔥 and ξS∈𝔫S=𝔤ΦS+, so we can choose h such that Adh⁡eα=α⁡(h)⁢eα, and thus the choices of h form a torsor for Z⁡(LS)={h∈T|α⁡(h)=1⁢ for all ⁢α∈S}. In particular, we see that fS+t+ξS∈ℓS. Taking the quotient by conjugation of NS, we see that the choices can be identified with fS+𝔟S/NS≅ℓS⫽LS. Therefore, we have proved that

If we denote the previously defined JG-torsor by 𝒥G, we see that this preimage is actually an isotropic submanifold in 𝒥G.

Note that there is a canonical open symplectic embedding

which is given by the formula

If we consider the composition b:𝒥G→N−\G/N→Spec⁡𝒪⁢(G/N)N−=Spec⁡ℂ⁡[𝕏+⁢(T)] (another integrable system), then the map

is proper.

4.2 Wonderful compactification.

Note that the horizontal integrable system has target

By the Peter-Weyl theorem, we see that this is actually

The monoid in question is a free monoid generated by the fundamental weights if G is simply connected. Our goal is to produce a partial log compactification

where YΣ^ is a complete toric variety and the map

is finite. Roughly, it will be something like

For a reference, read Brion-Kumar’s book on Frobenius splitting methods.

More generally, we can consider equivariant embeddings of G, where we are looking for a normal variety X with an action of G×G and an open equivariant embedding G↪X.

  1. (1)

    When G=T is a torus, then these are just toric varieties;

  2. (2)

    If G=Gad, then there is the wonderful compactification 𝕏=G¯.

The wonderful compactification satisfies the following properties:

  1. (1)

    𝕏 is smooth and complete;

  2. (2)

    The boundary 𝕏∖G is a union of smooth divisors X1,…,Xr with simple normal crossings;

  3. (3)

    The closures of the smaller G×G-orbits are given by ⋂i∈IXi where I⊂{1,…,r} is nonempty.

In general, equivariant embeddings will combine the features of toric varieties and the wonderful compactification. The idea is that G⁢Ln⊂Mn×n, while PGLn⊂ℙ⁡(Mn×n).

Let G=Gad and G~=Gsc. Choose a regular dominant weight λ∈Λ+,reg. Note that there is a map

which induces a map

Lemma 4.14.

For all regular dominant weights λ, Vλ satisfies the following:

  1. (1)

    The T-eigenspace Mλ of weight λ is a B-invariant line. All the other weights are strictly smaller than λ in the dominance order;

  2. (2)

    For all positive roots α, we have g−α⁢Mλ≠0 and the morphism G/B→ℙ⁡(Vl⁢a⁢m⁢b⁢d⁢a) sending g⁢B to g⋅Mλ is a closed embedding;

  3. (3)

    Dually, M−λ∗⊂Vλ∗ and the morphism G/B− to ℙ⁡(Vλ∗) sending g⁢B− to g⋅M−λ∗ is a closed embedding;

  4. (4)

    The action of G (resp 𝔤) on ℙ⁡(Vλ) factors through a faithful action of Gad (resp 𝔤ad).

Note that Z=Z⁡(G) acts on Vλ by a character λ:Z→ℂ×. If we consider the other weights

all α⁡(Z)=1, so the action factors through Gad. The same applies to the Lie algebra action. On the other hand, if g∈G fixes every line in Vλ, then g is in every Borel, so g∈Z. This implies that the action of Gad is faithful, so there is an embedding Gad↪PGL⁡(Vλ).

Now fix λ as above and let M=Vλ. Then consider the G~×G~-module End⁡(M)=M∗⊗M and let h∈End⁡(M) be the identity matrix.

Lemma 4.15.

The orbit (G×G)⋅[h] is isomorphic to

Proof.

Suppose that (g1,g2)⋅[h]=[h]. This is equivalent to g1⁢g2−1⁢v=c⁢v for a fixed scalar c∈ℂ× for all v∈M. This happens if and only if g1⁢g2−1∈Z. Therefore, the stabilizer of [h] is given by (Z×Z)⁢Δ⁢G~. □

Now let ℙ=ℙ⁡(End⁡(M)) and 𝕏=(G~×G~)⋅[h]¯. We need to prove that 𝕏 is smooth and that the boundary is snc. To study 𝕏, one key ingredient is to analyze the closure of (T×T)⋅[h].

Lemma 4.16.

Let G be a reductive group and X be a G×G-variety. Assume that x∈X is such that (G×G)⋅x is open in X and that the isotropy group (G×G)x contains Δ⁢G. Set X′≔(T×T)⋅x¯. Then:

  1. (1)

    All G×G-orbits in X intersect X′;

  2. (2)

    X contains finitely many G×G-orbits;

  3. (3)

    X is complete if and only if X′ is complete.

Remark 4.17.

If G is not adjoint, then we need to consider toroidal compactifications, which depend on a choice of W-equivariant toric compactification of T.

Let mi be a basis of T-eigenvectors of M and mi∗ be the dual basis. Then h=∑imi∗⊗mi. There are some summands like hλ=mλ∗⊗mλ and hλ−αi=mλ−αi∗⊗mλ−αi where λ is the highest weight. The vector hλ is a B−×B-eigenvector. Let ℙ0={mλ⊗mλ∗≠0}. Now let

and T¯ad,0=T¯ad∩ℙ0. Then ℙ0 is B×B−-stable because applying the action to

does not kill the leading term.

Lemma 4.18.
  1. (1)

    The inclusion Tad↪T¯ad,0 extends to an isomorphism γ:𝔸r≃T¯ad,0.

  2. (2)

    The action of W×W on T¯ad induces an action of Δ⁢W, and

    is a smooth complete toric variety whose fan consists of the Weyl chambers and their faces.

Proof.
  1. (1)

    Because Δ⁢T fixes [h], we can compute that

    where μi=λ−∑j∈Πnj(i)⁢αj. Then the embedding

    is given by

    It is easy to see that ıT⁢(T)¯≅𝔸r.

  2. (2)

    Note that

    By definition, the fan of T¯ad,0 is the negative Weyl chamber because the functions are −αi for i∈Π, but when we apply w, this corresponds to moving to a different chamber. Gluing these together, T¯ad will contain the toric variety with the complete fan given by all Weyl chambers and their faces. Clearly, these two things must be equal. □

Proposition 4.19.

The map

is an N×N−-equivariant isomorphism.

Proof.

The restriction of Γ to N×N−×Tad identifies the source with the open Bruhat cell in G. To show that Γ is an isomorphism, we need to construct an N×N−-equivariant morphism β:𝕏0→N×N− and then show that β−1⁢(1,1)≅T¯ad,0. Note that Tad⊂β−1⁢(1,1), but it is easy to see that

so clearly we see that β−1⁢(1,1)=T¯ad,0.

We will now construct β. Regard ℙ=ℙ⁡(End⁡(M)) as a space of rational maps ℙ⁡(M)⇢ℙ⁡(M). By definition, ℙ0 consists exactly of those rational maps defined at mλ and send this to ℙ⁡(M)0={mλ∗≠0}. For all φ∈G, it gives a regular self-map of ℙ⁡(M) preserving G⋅[mλ]=G/B. By the Bruhat decomposition, we see that

so we can consider the action of each Bruhat cell N−⁢w⁢N on [mλ] and see that

There is a map G∩𝕏0→ℙ⁡(M)0 and we want to factor it through a rational map 𝕏0→ℙ⁡(M)0. Because the image of G∩𝕏0 is already closed, the map extends to some partial compactification. This gives a map 𝕏0→N− which is clearly equivariant. The map 𝕏0→N is defined dually. □

Theorem 4.20.
  1. (1)

    𝕏 is covered by the G×G-translates of 𝕏0 (and in particular is smooth);

  2. (2)

    The boundary ∂𝕏=𝕏∖G is the union of r smooth prime divisors 𝕏1,…,𝕏r with normal crossings;

  3. (3)

    For all subset I⊂{1,…,r}, the stratum

    is the closure of a unique G×G-orbit 𝒪I. There is a unique closed orbit 𝒪{1,…,r}=G⋅[hλ]≅G/B−×G/B.

4.3 Log compactification of the universal centralizer.

Recall that JG=N\fT∗⁢G/fN. Here, we will need G to be adjoint, but the first step will be to replace the cotangent bundle with the log cotangent bundle TD∗⁢G of the wonderful compactification. To understand D, we will consider T¯ad,0=T¯∩𝕏0. This is a toric variety whose fan is the single cone given by the negative Weyl chamber, so the functions are simply αi⁢(t) for the simple roots αi. The T-orbits are indexed by I⊂{1,…,r} with standard points

Everything in T¯ can be translated to T¯ad,0 by the Weyl action. Every G×G-orbit in 𝕏 intersects T¯ad,0, so they are also indexed by T-orbits and will be called 𝒪I≔(G×G)⋅[zI]. These are stabilized by

where PIc is the parabolic subgroup associated to Ic. In particular, there is a fiber sequence

The orbit closures are given by

and the fibers of 𝒪¯I→G/PIc−×G/PIc are the wonderful compactification of LIc/Z⁡(LIc). In fact, there is an isomorphism

and from Di=𝒪¯{i} we see that the boundary is normal crossings.

Now consider the log tangent bundle TD⁢G¯ where local sections are vector fields tangent to the stratification by D. Explicitly, if we have local coordinates x1,…,xn such that D is defined by x1⁢⋯⁢xk=0, then the local sections are generated by x1⁢∂x1,…,xk⁢∂xk,∂xk+1,…,∂xn. The log cotangent bundle TD∗⁢G¯ is the dual of this, and its local sections are generated by

Now TD∗⁢G¯ inherits an action of G×G and in particular (G¯,D) is log homogeneous for the action of G×G. In particular, there is

We will denote the kernel by RG¯, which is called the bundle of isotropy subalgebras. We can show that RG¯ is the restriction of the tautological bundle of

where n=dim𝔤. It is in fact a Lagrangian subbundle of G¯×𝔤×𝔤 with respect to (β,−β) on 𝔤×g, where β is the Killing form. Now note that

Lemma 4.21.

There is an identification

Now consider the compactified moment map

where we view (a,x,y)∈RG¯⊂G¯×𝔤×𝔤.

Lemma 4.22.
  1. (1)

    The image of μ¯ is 𝔤×𝔤⫽G𝔤;

  2. (2)

    The variety 𝔤×𝔤⫽G𝔤 is normal;

  3. (3)

    μ¯ has connected fibers.

Sketch of proof.

For the first part, μ¯ is clearly proper, so the image is closed. Note that the image contains 𝔤reg×𝔤reg/G𝔤reg because the uncompactified moment map μ:T∗⁢G→𝔤×𝔤 is contains this locus. The image of μ¯ must be its closure, so it must be 𝔤×𝔤⫽G𝔤.

For the second part, 𝔤×𝔤⫽G𝔤 is a complete intersection defined by

where Pi are generators of 𝒪⁢(𝔤)G. Now 𝔤reg×𝔤reg/G𝔤reg is open and smooth, and the complement has codimension at least 2 because the complement of 𝔤reg has codimension at least 2, so 𝔤×𝔤⫽G𝔤 is smooth in codimension 1 and is therefore normal.

For the third part, we know that μ¯ is proper and its image is normal. We only need to show that a general fiber is connected, in which case all fibers are connected by Stein factorization. For this, we will use the general fact that μ¯−1⁢(xrs) is a connected toric variety. □

Now define a fiberwise partial compactification of JG by

Theorem 4.23 (Balibanu).

J¯ is smooth and the boundary J¯G∖JG is a divisor with simple normal crossings. Moreover, J¯ has a natural log symplectic structure extending the symplectic structure on JG, which is a closed nondegenerate 2-form with logarithmic poles along the boundary.

Sketch of proof.

Note that

Now consider μ¯−1⁢(f,f)/(N×N). Here, N×N acts freely on (f+𝔟)×(f+𝔟), so we can use a logarithmic version of Hamiltonian reduction. Here, we need D∩μ¯−1⁢(f,f) to be a divisor, and it certainly avoids μ−1⁢(f,f), and μ¯−1⁢(f,f) is connected and smooth, and hence irreducible. This also proves that μ−1⁢(f,f)/N×N⊂μ¯−1⁢(f,f)/N×N is open and dense. □

Theorem 4.24 (Jin, to appear).

Let 𝒥G→N−\eG/fN≅Spec⁡ℂ⁡[𝕏+⁢(T)]. This compactifies to

where YΣ^ is the complete toric variety associated to the fan Σ^⊂𝕏∙⁢(T)⊗ℤℝ whose rays are ℝ≥0⁢αi∨ for i∈Π and R≥0⁢(−λj∨) for j∈Π and whose dual polytope is homeomorphic to a cube. For each maximal cone σS in the fan, there is the identification

where Z⁡(LS)¯0 is the closure of Z⁡(LS) in T¯ad,0. Finally, we have