Patrick Lei

All notes

Lie groups and representations II

Lectures by Andrei Okounkov, Columbia, Spring 2021.

Lie Groups and Representations
Spring 2021
Notes by Patrick Lei
Lectures by Andrei Okounkov

Columbia University

Disclaimer

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

1 Lie algebras and algebraic groups

If G is a compact Lie group, then 𝔀=𝖫𝗂𝖾⁑(G) has an invariant metric (βˆ’,βˆ’). If 𝔀1βŠ‚π”€ is an ideal, then 𝔀1βŸ‚ is also an ideal and we have 𝔀=𝔀1βŠ•π”€2. In particular, we have

The simple nonabelian Lie algebras are very interesting and very special, and there are only countably many of them. Recall that they are classified by root systems.

1.1 Solvable and Nilpotent Lie Algebras

These are built out of abelian Lie algebras. They are not very interesting, but it is easy to find them. In some sense, if we consider the moduli space of Lie algebras, most Lie algebras will be nilpotent.

Definition 1.1.1.

Define the commutant 𝔀′ of a Lie algebra to be the span of [𝔀,𝔀]. Here, we have an exact sequence

By analogy, we may define Gβ€² to be the commutator subgroup of a Lie group G.

Theorem 1.1.2.

If G is simply connected, then Gβ€² is a Lie subgroup.

Example 1.1.3.

The commutant of the group of all matrices of the form

is the set of all matrices of the form

Now if we take

where Ξ› is a lattice in the center ℝ2, then Gβ€² is the image of

which is not necessarily a Lie subgroup.

Proof.

Consider the exact sequence 0→𝔀′→𝔀→ℝkβ†’0. Because G is simply connected, we can lift to an exact sequence

We know that Gβ€²βŠ†Gβ€²β€² but then 𝖫𝗂𝖾⁑(Gβ€²β€²)=𝔀′ and therefore we must have Gβ€²β€²=Gβ€². β–‘

Definition 1.1.4.

A Lie algebra is called solvable if (𝔀′)′⁒⋯=0. In other words, repeatedly taking the commutant eventually reaches 0. Alternatively, one should think about 𝔀 as an iterated extension by abelian Lie algebras.

Similarly, a group G is called solvable if (Gβ€²)′⁒…=1.

Corollary 1.1.5.

A connected Lie group G is solvable if and only if 𝖫𝗂𝖾⁑(G) is solvable.

Example 1.1.6.

The group BβŠ‚G⁒Ln of upper-triangular matrices is solvable. In some sense, this is an universal example.

Note that if G1βŠ‚G and G is solvable, then G1 is solvable. Conversely, if Gβ†ͺG2 and G is solvable, then so is G2. Next, if

is an exact sequence and G1,G2 are solvable, then so is G.

A stronger condition than being solvable is being nilpotent.

Definition 1.1.7.

A Lie algebra 𝔀 is nilpotent if [[[𝔀,𝔀],𝔀],…]=0.

There is a similar definition for Lie groups, and we have

Corollary 1.1.8.

A connected Lie group is nilpotent if and only if 𝖫𝗂𝖾⁑(G) is nilpotent.

Example 1.1.9.

The group of unitriangular matrices (equivalently the Lie algebra of strictly upper-triangular matrices) is nilpotent. Again, this is in some sense a universal example.

Theorem 1.1.10 (Lie).

If 𝔀 is a solvable Lie algebra over β„‚ and 𝔀→𝔀⁒𝔩⁒(V) is a representation, then 𝔀 maps into the set of upper-triangular matrices in some basis.

Remark 1.1.11.

β„‚ or any algebraically closed field of characteristic 0 is important because we need to ensure that every operator actually has an eigenvalue and therefore can be upper-triangularized. Having characteristic 0 is also important because 𝔰⁒𝔩2⁒(β„€/2⁒℀) is solvable. Here, we have [h,e]=[h,f]=0, and in particular, the defining representation cannot be upper-triangularized.

Proof.

The usual proof is induction. We simply need to find one common eigenvector, and first we find an eigenvector in 𝔀′ and then extend to 𝔀. We will prove the result more globally. If G has a common eigenvector v1, then the line ℂ⁒v1βˆˆβ„™β‘(V) is fixed by G. Therefore, if G is triangular in the basis v1,…,vn, this is equivalent to fixing a flag ℂ⁒v1βŠ‚β„‚β’v1+ℂ⁒v2βŠ‚β‹―βŠ‚V. Then by Borel-Morozov, a fixed flag exists because the flag variety is projective. β–‘

Theorem 1.1.12 (Borel-Morozov).

If G is a connected solvable affine algebraic group over an arbitrary algebraically closed field acting on a proper variety X, then XG is nonempty.

If we apply this discussion to the adjoint representation, we see that over a field of characteristic 0, 𝔀 is solvable if and only if 𝔀′ is nilpotent.

Theorem 1.1.13 (Engel).

Suppose π”€βŠ‚π”€β’π”©β’(V) consists of nilpotent operators. Then 𝔀 is contained in the set of strictly upper-triangular matrices for some basis.

The usual proof of this is by induction, so we skip it.

Corollary 1.1.14.

If we apply this to the adjoint representation, then 𝔀 is nilpotent is nilpotent if and only if ad⁑x=[x,βˆ’] is nilpotent for every x.

This result has a global analog, due to Kolchin (who incidentally was once a professor at Columbia).

Theorem 1.1.15 (Kolchin).

Let GβŠ‚G⁒L⁒(V) be any group consistent of unipotent operators. Then G is contained in the set of unitriangular matrices for some basis of V.

Proof.

By induction, it is enough to find one common fixed vector v1. We will assume that V is irreducible. Then we consider Span⁑(G)βŠ‚End⁑(V). On End⁑(V), we have a nondegenerate pairing (a,b)=tr⁑(a⁒b), and thus if we consider

we see that for all g, g=1. Therefore dimV=1 and every element is fixed. β–‘

Proof of Borel-Morozov.

Consider the exact sequence 1→G′→G→abelian→1. By induction on the dimension of G′ we see that XG′≠ is closed and thus proper. Now rename X=XG′, so we only need to prove the result for abelian G.

Any algebraic group action on an algebraic variety has a closed orbit π’ͺ, which in this case is proper. On the other hand, π’ͺ=G/stabilizer is an affine algebraic group and therefore must be a point.

For another proof, every affine abelian group is built out of 𝔾a,𝔾m, so we simply prove the result for these two groups. For 𝔾m, let x∈X. Then we simply consider the limit as tβ†’0 of tβ‹…x, which exists by the valuative criterion of properness. This must be a fixed point. For 𝔾a=𝔸1, we run the same argument except we consider the limit at ∞. β–‘

In the Lie theorem, GβŠ‚G⁒L⁒(n,β„‚) is an arbitrary connected solvable Lie group. We need to see that the Zariski closure of G is still solvable. If we write 𝔀¯=𝖫𝗂𝖾⁑(GΒ―), then we have

Lemma 1.1.16.

[𝔀¯,𝔀¯]=[𝔀,𝔀].

Proof.

We show that [𝔀¯,𝔀]=[𝔀,𝔀]. Consider

In particular, we have GβŠ‚GΒ―βŠ‚G~ because G~ is closed, and therefore [𝔀¯,𝔀]βŠ‚[𝔀,𝔀].

Now consider the same construction but with 𝔀 replaced by 𝔀¯. This implies that [𝔀¯,𝔀¯]βŠ‚[𝔀,𝔀]. β–‘

Corollary 1.1.17 (Borel).

All maximal connected solvable subgroups BβŠ‚G are conjugate for any connected linear algebraic group G.

Proof.

The idea is that if B0βŠ†G is connected solvable of maximal dimension, then X=G/B0 is projective. Then any other B will have a fixed point g⁒B0∈X, and so gβˆ’1⁒B⁒g fixes B0. This implies that gβˆ’1⁒B⁒gβŠ‚B0 and so they must be equal (by maximality).

Really, we will prove that GβŠ†G⁒L⁒(n,k). Therefore we will consider the action of G on the flag variety Fl⁑(n), and the stabilizer of any point is solvable. Then stabilizers of maximal dimension correspond to orbits of smallest dimension, which are closed and thus projective. Choose some maximal B0, which stabilizes a point in a closed orbit. Then B0 is solvable and X=G/B0 is projective, so the argument above works. β–‘

Now consider the variety X=G/B. This is called the flag variety for G.

Example 1.1.18.
  1. 1.

    If G=G⁒L⁒(n,k), then G/B is the usual flag variety. Here, B is all upper-triangular matrices.

  2. 2.

    Let G be one of the classical groups. Suppose g preserves a flag 0βŠ‚V1βŠ‚V2βŠ‚β‹―βŠ‚Vn=kn. Then g also preserves ViβŸ‚ and intersections of the Vi and their orthogonal complements, so we impose ViβŸ‚=Vnβˆ’i. Thus we take the space of flags to be

    We need to check that G acts on X transitively, so we check it up to V⌊n/2βŒ‹, which is a maximal isotropic subspace.

Theorem 1.1.19.

For all v1,…,vm∈kn, invariants of G=S⁒O or G=S⁒p are generated by (vi,vj) and minors like vi1βˆ§β‹―βˆ§vim. But all of these vanish because the vi are all orthogonal, so there are no invariants.

Definition 1.1.20.

A linear algebraic group G is an affine algebraic variety over k which is also a group.

Theorem 1.1.21 (Chevalley).

Over any field of characteristic 0, any group scheme is reduced and hence smooth.

Example 1.1.22.

Consider the group 𝔸1=𝔾a, the additive group of k. Then k⁑[G]=k⁑[t], and so the addition map (t1,t2)↦t1+t2 corresponds to the map f⁑(t)↦f⁑(t1+t2).

If char⁑k=p, then t↦tp is a group homomorphism. This gives us an exact sequence

Here, the first term is an affine group scheme because Δ⁒tp=tpβŠ—1+1βŠ—tp and therefore k⁑[t]/tp has a well-defined coproduct.

Therefore in characteristic 0, we can simply consider algebraic varieties. Then G is smooth, and we note that the maps m:GΓ—Gβ†’G,i:Gβ†’G,1β†’G induce maps Ξ”:Aβ†’AβŠ—A (comultiplication), S:Aβ†’A (antipode), and Ξ΅:Aβ†’k (counit). Here, the comultiplication is required to be coassociative, and the antipode is required to satisfy the identity

In other words, the diagram

commutes.

Now A has two sets of tensors:

  1. 1.

    As a commutative algebra over k, it has ΞΌ:AβŠ—Aβ†’A, which is dual to Ξ”:Gβ†’GΓ—G and the unit ΞΉ:kβ†’A. In principle, the multiplication does not need to be commutative.

  2. 2.

    As functions on a group, it gets Ξ”:Aβ†’AβŠ—A,Ξ΅:Aβ†’k,S:Aβ†’A. Note that the comultiplication may not be cocommutative.

Definition 1.1.23.

A Hopf algebra A over a field k is a bialgebra over k such that the axioms listed above for k⁑[G] are satisfied.

Note that there is no need for A to be commutative and that the set of axioms is symmetric. Therefore we can consider the dual of a Hopf algebra, where all vector spaces are replaced with their duals and all maps are replaced by the dual maps. Now we see that linear algebraic group schemes over k are equivalent to finitely generated commutative Hopf algebras over k.

Now let f∈A. We see that (Δ⁒f)⁒(g,h)=f⁑(g⁒h)=βˆ‘ci⁒(g)⁒fi⁒(h), so if g=1, then f is in the span of the fi. Also,

so every f∈A belongs to a finite-dimensional subspace that is invariant under the left regular representation. This implies that every affine algebraic group G is contained in G⁒L⁒(N,k)

Now note that if Gβ†’πœ‘H is a homomorphism of algebraic groups, then ImΟ†βŠ‚H is closed.

Theorem 1.1.24.

For all subgroups HβŠ†G, there exists a morphism Gβ†’G⁒L⁒(V) such that Im(H) is contained in the stabilizer of a line.

Proof.

Let IH be the ideal of H in k⁑[G]. Then H is the stabilizer of IHβŠ‚k⁑[G] under the natural G-action. Here, note that Lhβˆ’1⁒f⁒(g)=f⁑(g⁒h), so if we set g=1, then f⁑(h)=0 for all h∈stab⁑(IH).

Note that a tangent vector to an algebraic variety is an map in Hom⁑(Spec⁑k⁑[Ρ]/Ρ2,X) that sends the closed point of Spec⁑k⁑[Ρ]/Ρ2 to x∈X. Therefore, we have

Next, from last time, we know that dimSpan⁒{f⁑(gβˆ’1)}<∞, s IH=(f1,…,fk), where fi∈L, a finite-dimensional G-invariant subspace. Let LH=IH∩L. Then H is the stabilizer of LH, so H stabilizes a point in G⁑(dimLH,dimL). Now we note that β‹€kLHβŠ†β‹€kL is a line, as desired. β–‘

Definition 1.1.25.

We define G/H to be the orbit of the line that is stabilized in ℙ⁑(V).

Note that this definition is not necessarily independent of V and we also need to know what properties it satisfies. From now, we will assume H is smooth and dim𝖫𝗂𝖾⁑(H)=dimH. Then we have an exact sequence

and therefore G→Orbit is separable.

Theorem 1.1.26.

If X→Y is dominant, separable, and generically one-to-one, then it is birational.

Proposition 1.1.27.

Let x∈P⁑(V) be as above and let y∈Y, where Y is any variety with a G-action such that HβŠ‚Gy. Then there exists a unique G-invariant map Gβ‹…xβ†’Gβ‹…y such that x↦y.

Proof.

Consider the map g↦(gβ‹…x,gβ‹…y) that sends Gβ†’Gβ‹…xΓ—Gβ‹…y. Then the map p1:Gβ‹…xΓ—Gβ‹…yβ†’Gβ‹…x must be separable. But then GyβŠ‚Gx=H implies p1 is one-to-one. This implies that p1 is birational when restricted to the image of g↦(gβ‹…x,gβ‹…y). But this means that p1 is an isomorphism, so we can take p2∘p1βˆ’1 as the required map. β–‘

Now we will study what the space G/H looks like. In the case where G is connected and B is a maximal solvable groups, then the flag variety G/B is projective. The space G/H could also be an affine variety.

Definition 1.1.28.

A group G is reductive if for any GβŠ‚G⁒L⁒(V), we can write V=⨁Vi, where the Vi are irreducible G-modules.

Remark 1.1.29.

Over β„‚, this definition is equivalent to being the complexificaiton of a compact group.

Theorem 1.1.30 (Matsushita-Onishchik).

If G is reductive, then G/H is affine if and only if H is reductive.

Note that for most H, G/H is neither projective nor affine. For example, if we consider G⁒L⁒(2) and let

then G/H is the orbit of a single vector, which is 𝔸2βˆ–0. However, G/H is always quasiprojective, so it can be embedded in projective space.

Proposition 1.1.31.

The space G/H is quasiaffine if and only if H is the stabilizer of a point in an affine algebraic variety with a G-action. Such subgroups are called observable.

Proposition 1.1.32.

Ths space G/H is projective if and only if BβŠ‚H. Such subgroups are called parabolic.

Proof.

First, if G/H is projective, then (G/H)Bβ‰ βˆ… and thus B is conjugate to a subgroup of H. On the other hand, β–‘

1.2 Invariant Theory

Now note that if G acts on X, then G×X→X is a morphism of algebraic varieties. Now we want to study the space X/G. We can consider this in some world more general than algebraic varieties (namely stacks), but this is beyond the scope of this course. Instead, we will consider the best possible approximation in the category of schemes. Here, we will consider Y=X/G if for all Z (with the trivial action), G-equivariant maps X→Z factor uniquely through Y.

Our goal is to show that the GIT quotient X/G exists if X is affine and G is reductive.

Example 1.2.1.

Consider the action of G⁒L⁒(V) on V,VβŠ—V,Vβˆ—,Vβˆ—βŠ—V,…. Then the first (second?) fundamental theorem of invariant theory says that all invariants of these actions come from contracting tensors. For example, if we consider Vβˆ—βŠ—V=Hom⁑(V,V), the invariants are generated by the coefficients of the characteristic polynomial. This means that Hom⁑(V,V)/G⁒L⁒(V)=𝔸dimV.

Remark 1.2.2.

Note there are several notions of being reductive. The first is structural. The second is being linearly reductive, which means that we need something like kβ†’k⁑[G]β†’k, where the last map is some sort of invariant integration. Finally, there is the notion of being geometrically reductive. If k has characteristic 0, then all of these notions are equivalent.

Lemma 1.2.3.

Suppose we can split every module with an invariant element as k→M→k. Then all representations are linearly reductive.

Proof.

Let M1βŠ‚M be some submodule. We want a G-invariant map Mβ†’M1, which requires a G-equivariant map Hom⁑(M,M1)β†’Hom⁑(M1,M1) that maps onto 1M1. But this problem is resolved by taking the transpose of a matrix acting on M that preserves M1. β–‘

Note that G⁒L⁒(V) is not linearly reductive if char⁑k=p. In this case, consider the action of G⁒L⁒(2) on the polynomials of degree d in x1,x2. Then the span of x1p,x2p does not split off.

Definition 1.2.4.

A group G is reductive if the radical of G is a torus. Equivalently, the unipotent radical of G is trivial. Here, the radical of G is defined to be

and is the largest normal connected solvable subgroup. The unipotent radical is defined to be the largest normal unipotent connected subgroup, and is

Definition 1.2.5.

A group G is geometrically reductive if for any G-module M and line of G-fixed points, there exists a complementary divisor given by a G-invariant polynomial f⁑(m) that does not vanish on the line.

Example 1.2.6.

Finite groups fail to be linearly reductive in positive characteristic. For example, the representation of β„€/p⁒℀ given by

is not completely reducible. On the other hand, they are geometrically reductive.

Proof.

Take any f0 such that f0⁒(0)=0 and f0⁒(x)=1, where x is a fixed point. Then take

Now we can choose the Taylor series of f⁑(x) to be homogeneous of degree p. β–‘

Theorem 1.2.7 (Haboush).

Group-theoretic reductivity is equivalent to geometric reductivity.

Corollary 1.2.8.

Let A be an algebra with a G-action and suppose Aβ† B, which also has a G-action. Then linear reductivity means the natural map AGβ† BG is surjective. Geometric reductivity means that for all f∈BG, there exists m=pk such that bm∈Im(AG). In particular, BG is integral over AG.

Theorem 1.2.9 (Nagata, Popov,…).

A group G is reductive if and only if for all finitely generated (commutative) algebras A, the algebra AG of invariants is finitely generated.

This result is extremely hard. Instead, we will prove

Theorem 1.2.10 (Hilbert).

Let X be an affine variety over a field k of characteristic 0 and G a reductive group. Then k⁒[X]G is finitely generated.

Proof.

Let XβŠ†V and Gβ†ͺG⁒L⁒(V). Now k⁒[V]Gβ† k⁒[X]G by linear reductivity. Consider the ideal I=(k⁒[V]+G). This is finitely generated by another theorem of Hilbert (from the same paper). If f1,…,fk are generators, then we will show that they also generate k⁒[V]G.

Let F∈k⁒[V]dG for some d>0. Then F∈I and we can write F=βˆ‘ci⁒fi. Now we will take the average over G, which is linear over invariants. Now we obtain F=βˆ‘cΒ―i⁒fi, where the cΒ―i are all invariants of degree less than d. By induction on d, we are done. β–‘

For the proof in arbitrary characteristic, there is a book on invariant theory by T. Springer.

Now consider a map Xβ†’πœ‹(f1,…,fk)𝔸k, where X is an affine variety with an action of a reductive group G. Then we will show that

  1. 1.

    The map Ο€ takes G-invariant Xβ€²βŠ†X to closed subsets.

  2. 2.

    If Xβ€²,Xβ€²β€² are disjoint G-invariant closed subsets, then π⁑(Xβ€²)βˆ©Ο€β‘(Xβ€²β€²)=βˆ….

  3. 3.

    For any open UβŠ†Ο€β‘(X), Ο€βˆ—β’π’ͺU=π’ͺΟ€βˆ’1⁒(U)G.

In particular, we will show that if G is reductive and Xβ€²,Xβ€²β€²βŠ†X are closed and disjoint, then there exists f∈k⁒[X]G such that f⁑(Xβ€²)=0,f⁑(Xβ€²β€²)=1. To see this, we know that IXβ€²+IXβ€²β€²=k⁑[X], so we can find f0∈IXβ€²,f1∈IXβ€²β€² such that f0+f1=1. Thus f0⁒(Xβ€²)=0,f0⁒(Xβ€²β€²)=1. Then if f0,…,fm span f0⁒(gβˆ’1βˆ’), then the map Xβ†’(f0,…,fm)𝔸m+1 sends Xβ€² to (0,…,0) and Xβ€²β€² to (1,…,1).By geometric reductivity, there exists a polynomial P⁑(f0,…,fm) which is invariant and takes values 0 on Xβ€² and 1 on Xβ€²β€².

Note that if G is not reductive, then closed subsets cannot be separated by invariants. For an example, consider the action of 𝔾a on 𝔸2 by translating the second coordinate. Then (x,0),(y,0) cannot be separated by invariants.

Now we need to show that π⁑(X) is closed. If not, then if pβˆˆΟ€β‘(Xβ€²)Β―βˆ–Ο€β‘(Xβ€²), then Ο€βˆ’1⁒(p) is closed and disjoint from Xβ€². But this implies there exists f such that f⁑(Xβ€²β€²)=f⁑(p)=1 and f⁑(Xβ€²)=f⁑(π⁑(Xβ€²))=0, a contradiction.

Finally, let UβŠ†Ο€β‘(X) be given by {F1β‰ 0,F2β‰ 0}. Then

This all implies that π⁑(X)=X/G is the categorical quotient of X under the action of G. To see this, observe that if Ui is an affine open cover of Z, then pβˆ’1⁒(Ui) cover X, so Xi=Xβˆ–pβˆ’1⁒(Ui) is closed and β‹‚Xi=βˆ…. Now let Vi=Yβˆ–Ο€β‘(Xi). These form an open cover of Y, so now write pΒ―:Viβ†’Ui. Then we have

and this must be unique, so Ο€βˆ—β’pΒ―βˆ—=pβˆ—, so pΒ―:Viβ†’Ui.

Therefore we have proved that if X is affine and G is reductive, then Y=Specβ‘π•œβ’[X]G is the categorical quotient. Note that this is surjective, and for p∈Y, Ο€βˆ’1⁒(p) is nonempty and contains a unique closed orbit.

Now we will discuss quotients of general varieties by algebraic groups. This is very complicated because x∈X may not have a G-invariant affine neighborhood (consider the example of Hironaka). Now if we consider XβŠ‚β„™β‘(V) for a G-module V with Vβˆ—=π’ͺ⁑(1), then π’ͺ⁑(1) is a very ample line bundle on X with a linearzation by G. Similarly to Y=Specβ‘π•œβ’[X]G, we may consider the affine cone XΜ‚ over X and then take Y=Projβ‘π•œβ’[XΜ‚]G. This is covered by open sets where {Fi⁒(x)β‰ 0}, and then ℙ⁑(V)βˆ–{F⁑(x)=0} is an affine G-invariant set.

Not all points have an invariant polynomial Fi such that Fi≠0. The points that do are called semistable.

Definition 1.2.11.

The GIT quotient Xβ«½G is defined to be Projβ‘π•œβ’[XΜ‚]G=Xs/G, where Xs is the stable locus.

The unstable points are those such that there is no invariable Fd such that Fd⁒(x)β‰ 0. But this implies that the closure of the orbit of x in V contains 0∈V.

Note that if Ο‡:G→𝔾m is a character, then V↦VβŠ—Ο‡ does not change the action on ℙ⁑(V) because Sd⁒Vβˆ—β†¦Sd⁒Vβˆ—βŠ—Ο‡βˆ’d sends Ο‡-covariants to invariants. Therefore, even in the affine situation, it makes sense to consider Xβ«½G=Proj⁑covariants. For the most basic example, consider ℙ⁑(V)=Proj⁑⨁Sd⁒Vβˆ—. Then V/𝔾m is a point, and V⫽𝔾m=ℙ⁑(V). On the other hand, we have Vβ«½Ο‡=t𝔾m=Proj⁑ℂ=βˆ…, so in both cases the map Xβ«½Gβ†’X/G is uninteresting.

Now we want to find generators of the algebra π•œβ’[X]G=π•œβ‘[f1,…,fN]. Then the affine scheme X/G is cut out by the relations among the fi. Finding the relations is incredibly hard, so we can try to find the generators. Results of this form go under the form of the first fundamental theorem of invariant theory. Here, we will assume G=G⁒L⁒(n),S⁒L⁒(n). These fit into the exact sequence

Therefore S⁒L⁒(V)-invariants are the same as G⁒L⁒(V)-covariants with respect to the determinant character. We know that π•œβ‘[X] contains a finite-dimensional G-invariant module M, which can be included in π•œβ’[G]βŠ•m. This implies that any X can be emdedded in some VβŠ•m1βŠ•(Vβˆ—)βŠ•m2≕Mm1,m2 because there is a natural map π•œβ‘[End⁑(V)]β†’π•œβ‘[G]. This gives us a map π•œβ‘[Mm1,m2]β†’π•œβ‘[X] that restricts to invariants, so we have reduced the problem of finding invariants to vector spaces.

Theorem 1.2.12 (First fundamental theorem of invariant theory).

The invariants of S⁒L⁒(V) acting on VβŠ•m1βŠ•(Vβˆ—)βŠ•m2 are generated by

  1. 1.

    Contracting tensors: (v1,…,vm1,β„“1,…,β„“m2)β†¦βŸ¨vi,β„“j⟩;

  2. 2.

    Determinants of the form det⁑(vi1…vin) with weight det and dually for the β„“j with weight detβˆ’1 (weights are under the action of G⁒L).

Proof.

Note that Mm1,m2=Hom⁑(π•œm1,V)βŠ•Hom⁑(V,π•œm2). Now the two parts parts have actions by the groups G⁒L⁒(m1),G⁒L⁒(m2) and maximal tori Tm1,Tm2. THen the weights record how many times we use a particular vector or covector. Now it suffices to consider functions of weight (1β„“,0k).

To see this, we use a polarization trick. If degvi⁑f⁑(v1,…)=d, then we can write vi=βˆ‘Ξ»i⁒ui and now we have a function of m1+dβˆ’1 vectors u1,…,ud,v2,…,vm1. Expanding this, we obtain a new polynomial f~ that is linear in each of the u1,…,ud. Then considering the polynomial f~⁒(v1,…,v1,v2,…,vm) gives us the desired reduction.

But now functions on Mm1,m2 linear in each of the v1,…,vm,β„“1,…,β„“m2 are just the space (Vβˆ—)βŠ—m1βŠ—VβŠ—m2. We will show that

The scalars tβ‹…I act with weights tβˆ’m1+m2 so there are no invariants unless m1=m2. Now if m1=m2, we are looking for

This result is known as Schur-Weyl duality. If we consider the natural map G⁒L⁒(V)Γ—Smβ†’End⁑(VβŠ—m). In fact, each piece of the product generates the commutant of the other. We see that both images are semisimple subalgebras in End⁑(VβŠ—m). Now the desired result is equivalent to proving that End⁑(VβŠ—m)Sm is the image of G⁒L⁒(V). Then polynomials on End⁑(V) of degree m are the same as Sm⁒End⁑(V)βˆ—=End⁑(VβŠ—m)Sm. Suppose that End⁑(VβŠ—m)SmβŠ‹G⁒L⁒(V). But then we can consider G⁒L⁒(V)βŸ‚ in the set of polynomials of degree m. Let P be such a polynomial. Then P⁑(g,…,g)=0 for all g∈G⁒L⁒(V). But then by Zariski density of G⁒L⁒(V) in End⁑(V), we see that P=0. This now tells us that

Now we need to compute the additional S⁒L⁒(V)-invariants, which are given by

We will introduce new covectors β„“Β―1,…,β„“Β―n and consider the functions

which is an invariant and thus contained in π•œβ‘[⟨vi,β„“j⟩,⟨vi,β„“Β―j⟩]. Now det is multilinear and skew-symmetric, so each β„“Β―j has to be used exactly once. But now f is a product f1,f2 where f1βˆˆπ•œβ‘[⟨vi,β„“j⟩] and f2 is contained in the antisymmetrization of ∏⟨vik,β„“Β―k⟩, so f2=det⁑(vi⁒k)β‹…det⁑(β„“Β―k). β–‘

1.2.1 Finite Subgroups of S⁒L⁒(2,β„‚)

Now consider a finite group GβŠ‚S⁒L⁒(2,β„‚). For example, G is cyclic, dihedral, etc. Now we have an exact sequence

and now we can find in S⁒O⁒(3) symmetries of the Platonic solids A4,D4,A5 corresponding to tetrahedron, cube, and dodecahedron. Now if γ∈S⁒O⁒(3) has order 3 with eigenvalues 1,ΞΆ3,ΞΆ32, then we know Ξ³~∈S⁒L⁒(2) has eigenvalues ΞΆ6,ΞΆ6βˆ’1. Now for G∈A~4,S~4,A~5 and V=β„‚2, we know that S⁒L⁒(2)=S⁒p⁒(2) preserves the skew pairing. We know

so now consider the Hilbert/PoincarΓ© series

By an observation of Hilbert, this is a rational function for any finitely generated graded module over a finitely generated algebra. But now we know that ℂ⁑[a1,…,am]β† A. If ai has degree di, then the free module has Hilbert series

In general, a module M has a finite gree resolution

This gives us

Theorem 1.2.13 (Molien).

Let G be a reductive group over β„‚ acting on a vector space V. Then

To do the actual computation, we can use the Weyl character formula. This is simply

and this can be computed using residues. Of course, if G is finite, then we just sum over conjugacy classes. For example, if G=A4, then these are cycles of signature either (3,1) or (2,2), and therefore

is a group of order 24. Now the conjugacy classes are given by

Remark 1.2.14.

Andrei admires the mathematicians of the past who were able to compute things by hand. Now he cannot imagine performing these computations without a computer. It is important to note that we should always use a free and open-source program to perform such computations and not something proprietary like some programs that shall not be named.12

This tells us that

This suggests generators of degree 6,8,12 and a relation in degree 24.

Theorem 1.2.15 (E. Noether).

The ring (Sβˆ™β’V)G is generated in degree at most |G|.

Proof.

By polarization, we know Sd⁒V is spanned by vd. But then we know that (Sd⁒V)G is spanned by polynomials of the form

and this can be expressed in elementary symmetric functions of degree at most |G|. β–‘

Now we can rewrite

Let x,y,z be the generators of degree 6,8,12, and then in degree 24, we have some relation

There are no further relations because dimV/A4~=2, so we have a map

Remark 1.2.16.

This classification of finite subgroups of S⁒L⁒(2) also gives us Du Val singularities, the classification of simple Lie algebras, the McKay correspondence, and many other interesting objects in mathematics.

Now consider the action of S⁒L⁒(2,β„€) on the upper half-plane β„‹. Then we have an exact sequence

and we have an action of S⁒L⁒(2,β„€/m) on β„‹/Γ⁑(m).

But now Γ⁑(m) has no torsion, so we have finitely many cusps, corresponding to the action of Γ⁑(m) on β„š, and β„‹/Γ⁑(m) is a curve of genus g=g⁑(Γ⁑(m)). For m=3,4,5, we have g=0 and thus S⁒L⁒(2,β„€/m) acts on β„™1.

Now the tetrahedron corresponds to the standard fundamental domain for S⁒L⁒(2,β„€). The cube corresponds to the below:

Refer to caption
Figure 1.1: Fundamental domain subdivided

The dodecahedron and icosahedron correspond to the following:

Refer to caption
Figure 1.2: Fundamental domain for icosahedron

The cusps correspond to points of 5-fold symmetry, and correspond to 0,1,2,∞,Ο†,1/Ο†,…, and the points converging to 1/Ο† are given by ratios Fn/Fnβˆ’1 of Fibonacci numbers.

Remark 1.2.17.

Instead of just considering the icosahedron, we should consider an infinite strip with the given pattern in the picture.

1.3 Jordan Decomposition

Let π•œ=π•œΒ―. Then for all g∈G=G⁒L⁒(V), we can decompose g into Jordan blocks. In particular, we can write g=gs⁒gn, where gs is semisimple and gn is strictly upper triangular (in some basis). The analogous decomposition for ΞΎβˆˆπ”€ is ΞΎ=ΞΎs+ΞΎn. Now if VβŠ†π•œn is invariant under g, it is invariant under both gs,gn.

Consider the regular representation ρ of G=G⁒L⁒(n,π•œ) on π•œβ‘[G].

Lemma 1.3.1.

For any g∈G, we have ρ⁒(g)s=ρ⁑(gs).

Corollary 1.3.2.

If G is an algebraic group, then gs,gn∈G for all g∈G. Moreover, if Ο†:G1β†’G2 is a homomorphism, then φ⁒(g)s=φ⁑(gs) and φ⁒(g)n=φ⁑(gn).

Proof.

By Chevalley, G is the stabilizer of a subspace on π•œβ‘[G⁒L⁒(n)] and therefore because g stabilizes the subspace, so do gs,gn.

For the second part, we can pull back Ο†βˆ—:π•œβ‘[G2]β†’π•œβ‘[G1] and then the desired result is obvious. β–‘

1.4 More Solvable Lie Algebras

We will return to solvable Lie algebras. Assume charβ‘π•œ=0.

Theorem 1.4.1.

Let π”€βŠ‚π”€β’π”©β’(V) be a Lie subalgebra. Then 𝔀 is solvable if and only if tr⁑x⁑[y,z]=0 for all x,y,zβˆˆπ”€.

Proof.

One direction is clear by Borel. Here, π”€βŠ‚π”Ÿ is contained in the subalgebra of upper-triangular matrices. In the other direction, the form tr⁑x⁑[y,z]∈(Ξ©3⁒𝔀)𝔀. In particular, we see that if 𝔀=𝖫𝗂𝖾⁑(G), then this becomes a bi-invariant 3-form on G. This gives a class in HdR3⁒(G)GΓ—G. Now we have an exact sequence

and here the unipotent radical is homeomorphic to ℝn, while Greductive is a product of simple nonabelian Gi and a torus up to a finite cover. Then rk⁑π3 is the number of simple nonabelians, and so the morphisms S⁒U⁒(2)≃S3β†ͺ(Gi)compact generate Ο€3βŠ—β„š.

Now suppose that tr⁑x⁑[y,z]=0. Now we will show that 𝔀 is solvable. It suffices to show that 𝔀′=[𝔀,𝔀] is nilpotent. By Engel, it suffices to show that any xβˆˆπ”€β€² is nilpotent. Consider the subalgebra

This is the Lie algebra of an algebraic group

But then tr⁑x⁒ξ=0 for all xβˆˆπ”€β€²,ΞΎβˆˆπ”€~. Now if x=βˆ‘[yi,zi], then we have

so if xβˆˆπ”€βŠ‚π”€~, then xsβˆˆπ”€~. Now considering f⁑(xs)βˆˆπ”€~, we will obtain some condition on ad⁑f⁑(xs). These will have the same eigenvectors as ad⁑xs. If Ei⁒j is an eigenvector of ad⁑xs with eigenvalue Ξ»iβˆ’Ξ»j, then it has eigenvalue f⁑(Ξ»i)βˆ’f⁑(Ξ»j) under ad⁑f⁑(xs). If there exists ψ such that f⁑(Ξ»i)βˆ’f⁑(Ξ»j), then ad⁑f⁑(xs)=ψ⁑(ad⁑xs) and thus f⁑(xs)βˆˆπ”€~.

Now if f is linear over β„š, then f⁑(xs)βˆˆπ”€~ and ad⁑f⁑(xs)=f⁑(ad⁑xs). Next we see that tr⁑xs⁒f⁑(xs)=0 because we can embed {Ξ»i}βŠ‚β„‚ and then take f⁑(Ξ»i)=λ¯i. Then we see that tr⁑xs⁒f⁑(xs)=βˆ‘|Ξ»i|2=0, and then we see that all Ξ»i=0. Alternatively, if dimβ„šβ¨β„šβ’Ξ»i>0, then there exists a nonzero f:β¨β„šβ’Ξ»iβ†’β„š, but then f⁒(βˆ‘Ξ»i⁒f⁒(Ξ»i))=βˆ‘fi⁒(Ξ»)2. β–‘

Definition 1.4.2.

Define the Killing form by

Remark 1.4.3.

Killing apparently lived a very sad life and did not get the recognition he deserved. Unfortunately, Andrei (and I) do not know more about him.

Theorem 1.4.4.

A Lie algebra 𝔀 is solvable if and only if (x,[y,z])=0.

Proof.

Consider the exact sequence

Then solvability of 𝔀 is equivalent to solvability of ad⁑𝔀. β–‘

Theorem 1.4.5 (Cartan Criterion).

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

Proof.

Suppose (βˆ’,βˆ’) is degenerate. Then π”€βŸ‚ is a solvable ideal in 𝔀. But then if ℐ is a solvable ideal with ℐn+1=0, then π”ž=ℐn is an abelian ideal. Therefore, for all x,y, we see that

and therefore for all aβˆˆπ”ž,yβˆˆπ”€, we have ad⁑(a)⁒ad⁑(y)⁒ad⁑(a)=0, so (ad⁑(a)⁒ad⁑(y))2=0, and thus tr⁑ad⁑(a)⁒ad⁑(y)=0. But then aβˆˆπ”€βŸ‚. β–‘

Corollary 1.4.6.

If 𝔀 is semisimple, then 𝔀=⨁𝔀i is a sum of simple nonabelians.

Proof.

Suppose π”₯βŠ‚π”€ is an ideal. Then π”₯∩π”₯βŸ‚=0 (because it is a solvable ideal). Note that if π”₯βŠ‚π”€ is an ideal, then (h1,h2)π”₯=(h1,h2)𝔀. β–‘

2 Geometric and topological aspects

2.1 Lie algebra cohomology

There are three kinds of properties:

  • β€’

    General abstract properties;

  • β€’

    Properties derived from the structure theory 𝔀=π”₯βŠ•β¨Ξ±π”€Ξ±;

  • β€’

    Properties derived from the calssification of root systems.

We will begin with the first. If 𝔀 is semisimple, then:

  1. 1.

    The category π–¬π—ˆπ–½fd⁑𝔀 is semisimple. In particular, every finite-dimension 𝔀-module has the form M=⨁Mi, where the Mi are simple.

  2. 2.

    The algebra 𝔀 has no deformations.

  3. 3.

    All derivations of 𝔀 are inner derivations. In particular, we see that 𝔀=𝖫𝗂𝖾⁑(Aut⁑(𝔀)). In addition, in the exact sequence

    the two outside terms vanish.

  4. 4.

    For any Lie algebra 𝔀any, we have

    and this exact sequence splits into 𝔀any=𝔀ss⋉radical.

All of these pheonomena fit under the umbrella of the vanishing of some cohomology groups.

Definition 2.1.1.

Let 𝔀 be a Lie algebra over a field π•œ and M be a 𝔀-module. We may consider the complex

and define

In the homework, we will show that d2=0, so we may define the Lie algebra cohomology Hn⁒(𝔀,M).

For an example in low dimension, we see that C0=Mβ†’C1 is given by dω⁒(ΞΎ1)=ΞΎ1⁒(Ο‰), so H0⁒(M)=M𝔀. For a more modern definition, we see that Hi⁒(𝔀,M) are the derived functors of Mβ†’M𝔀.

We may motivate this formula in the following way from the de Rham differential. Suppose G acts on a manifold X. This gives a morphism 𝔀→Γ⁑(X,T⁒X) into the vector fields. Then if ΞΎ1,…,ΞΎn+1 are vector fields on X and Ο‰ is an n-form on X, we have

Proposition 2.1.2.

The formula for dω⁒(ΞΎ1,…,ΞΎn) is the same formula as in Lie algebra cohomology, where ΞΎi⁒ω is the Lie derivative of Ο‰ along ΞΎi.

Proof.

Andrei’s proof is way too confusing. This is also Proposition 12.19 in Lee’s smooth manifolds book. The proof there is the same, but is done in a much easier-to-digest way. β–‘

Theorem 2.1.3.

Let 𝔀 be a semisimple Lie algebra over a field of characteristic 0.

  1. 1.

    If M is irreducible and nontrivial, then Hβˆ™β’(𝔀,M)=0.

  2. 2.

    Hβˆ™β’(𝔀,π•œ) is the free anticommutative algebra on finitely many generators of degree contained in {3,5,7,…}. When π•œ=β„‚, this is the same as Hβˆ™β’(G,β„‚). This is also the same as the cohomology of the maximal compact subgroup. For example,

    In particular, if 𝔀 is semisimple, then H1,H2 vanish for all M. Also, we have an isomorphism H3⁒(𝔀,π•œ)β‰…k#⁒simple factors, and this is just the space of invariant bilinar forms.

Proof.
  1. 1.

    Suppose M is nontrivial and irreducible. Without loss of generality, assume 𝔀 is simple and consider π”€βŠ†π”€β’π”©β’(M). We will show that multiplication by dim𝔀 is homotopic to 0. Consider the form B⁑(x,y)=trM⁑x⁒y, which is nondegenerate. Then if {e1,…,ed},{f1,…,fd} are dual bases of 𝔀, the Casimir element βˆ‘ei⁒fi commutes with 𝔀 and is nonzero because trβ‘βˆ‘ei⁒fi=βˆ‘tr⁑ei⁒fi=dim𝔀. Now we will define our homotopy by

    Now, we compute d∘h+h∘d. We have

    Then writing

    we have

    Now we may verify that all of the relevant terms cancel, so

    The second term is given by inserting the tensor [eβˆ—βŠ—fβˆ—,ΞΎiβŠ—1+1βŠ—ΞΎ]. But then eβˆ—βŠ—fβˆ— is an invariant tensor, so ad⁑(ΞΎ)⁒eβˆ—βŠ—fβˆ—=[ΞΎβŠ—1+1βŠ—ΞΎ,eβˆ—βŠ—fβˆ—]=0.

  2. 2.

    We may assume that π•œ=ℝ or π•œ=β„‚, so let G be a connected compact Lie group. We will see that Hβˆ™β’(𝔀)=(β‹€βˆ™π”€)G≃Hβˆ™β’(G). We will compute ordinary cohomology using the de Rham complex. First, we will note that g∈G acts trivially on Hβˆ™β’(G). To see this, observe that ΞΎβˆˆπ”€ acts on forms by the Lie derivative LΞΎ, and [Lξ⁒ω]=0 if dΟ‰=0 by the Cartan formula.

    Now for any compact group G acting on a manifold X, the inclusion (Ξ©i⁒X)Gβ†’Ξ©i⁒X induces an isomorphism on cohomology. To see this, we simply note that ∫gβˆ—β’Ο‰β’dg is cohomologous to Ο‰. Therefore, we may consider right-invariant forms in Ξ©i⁒G. But then

    and the differential on β‹€iπ”€βˆ— is the differential from Lie algebra cohomology. Now if we consider (Ξ©i)GΓ—G≃(β‹€iπ”€βˆ—)G, the differential vanishes. To see this, the map g↦gβˆ’1 perserves the biinvariants but acts by (βˆ’1)i on β‹€iπ”€βˆ—, and so dβ†¦βˆ’d, so d=0. Because 𝔀 acts trivially on cohomology, it is also possible to see that

    is a quasi-isomorphism. β–‘

Example 2.1.4.
  1. 1.

    Let 𝔀 be abelian. Then

  2. 2.

    The condition that G is compact is important. Note that

Next, we will actually prove that Hβˆ™β’(𝔀,π•œ)=π•œβ’βŸ¨Ο‰2⁒diβˆ’1⟩i=1,…,rank⁑𝔀. In particular, by a theorem of Hopf, this is a Hopf algebra. Similarly, if G is a compact Lie group then Hβˆ™β’(G,ℝ) is a Hopf algebra. Here are some properties of the cohomology:

  • β€’

    It is graded and supercommutative.

  • β€’

    Under the map Gβ†’GΓ—Gβ†’πœ‡G given by g↦(g,1)β†’g, if we write

    then (1βŠ—Ξ·)⁒Δ⁒ω=Ο‰ because Δ⁒ω=Ο‰βŠ—1+1βŠ—Ο‰+H>0βŠ—H>0.

Theorem 2.1.5 (Hopf).

Any finitely generated graded supercommutative Hopf algebra and the second property has the form π•œβ’βŸ¨Ο‰mi⟩.

Corollary 2.1.6.

If, additionally, our Hopf algebra is assumed to be finite-dimensional, then all of the mi are odd.

Proof.

Let Ο‰mi be generators with m1≀m2≀⋯. Then let β„‹k be the algebra generated by Ο‰m1,…,Ο‰mk. Then we know that Δ⁒ωmk=Ο‰mkβŠ—1+1βŠ—Ο‰mk+β‹―, so each β„‹k is a sub-bialgebra. Now it suffices to show that β„‹k=β„‹kβˆ’1β’βŸ¨Ο‰mk⟩.

Suppose there is a relation R=βˆ‘i=0dci⁒xi=0 of degree d. But now if we consider the ideal Δ⁒R modulo 1βŠ—βŸ¨β„‹kβˆ’1,x2⟩, which does not contain x for grading reasons, then we have

and this must be a relation of smaller degree. Therefore we have no relations beyond supercommutativity. β–‘

Now an interesting problem is to compute the degrees of the generators. For example, we have

Theorem 2.1.7 (Cartier-Kostant-Gabriel-…).

If β„‹ is a supercommutative Hopf algebra over a field π•œ of characteristic 0, then

where G is a (typically finite) group and 𝔀 is a Lie superalgebra over π•œ.

The elements with Δ⁒g=gβŠ—g are called grouplike, and the elements with Δ⁒ξ=ΞΎβŠ—1Β±1βŠ—ΞΎ are called primitive. The grouplike elements give us G, and the primitive elements give us 𝔀.

Now we may take the dual Hopf algebra, and this gives us another graded supercommutative Hopf algebra. These give us algebraic supergroups over π•œ. But this algebraic supergroup must be an odd vector space. Another consequence of the theorem is that a commutative Hopf algebra over π•œ has no nilpotent elements. This implies that all group schemes over π•œ are reduced and therefore smooth.

Recall that if G is a compact connected Lie group, then Hβˆ™β’(G,ℝ)=β„β’βŸ¨Ο‰2⁒diβˆ’1⟩i=1,…,rank⁑G.

Theorem 2.1.8.

We have dimℝHβˆ™β’(G,ℝ)=2rank⁑G, where rank⁑G is the dimension of the maximal torus.

We would also like to compute the di. In fact, di are the degrees of the generators of (Sβˆ™β’π”€βˆ—)G. We know that every element of 𝔀 is conjugate to an element of 𝔱=𝖫𝗂𝖾⁑T. The normalizer of this is the Weyl group W. Now, by definition, we have

and this is free on some generators pdi of degree |pdi|=di for i=1,…,dim𝔱=rk⁑G.

Theorem 2.1.9.

The di defined in the various ways are the same numbers and are called exponents of G.

Example 2.1.10.

For G=U⁑(n), we have {di}={1,2,…,n} and pd=tr⁑ξd. Alternatively, we can use the coefficients of the characteristic polynomial. In addition, we see that

and (β‹€βˆ™π”€βˆ—)G is generated by Ο‰1⁒(ΞΎ)=tr⁑ξ,Ο‰3⁒(ΞΎ1,ΞΎ2,ΞΎ3)=tr⁑ξ1⁒[ΞΎ2,ΞΎ3], and in general

In this formula, we observe that d must be odd.

Proof of Theorem 3.2.8.

Use the Molien series. We have

Setting V=β‹€βˆ™π”€, we see that

This now gives us

by change of variables. β–‘

2.2 Classifying Spaces and Flag Varieties

We already know that Hβˆ™β’(G)=β„β’βŸ¨Ο‰2⁒diβˆ’1⟩. On the other hand, we know the cohomology of the flag manifold Hβˆ™β’(G/T) is all even, and finally we have the cohomology

On the other hand, we have

For G=U⁑(n), this becomes the space of polynomials divided by symmetric polynomials of positive degree and has dimension n!=|W|. This is the fiber over 0 of the map 𝔱→𝔱/W. To define B⁒G, consider the category of spaces with a free action of G (equivalently principal G-bundles) for any group G. Given a commutative diagram

with φ⁑(gβ‹…x)=φ⁑(g)⋅φ⁑(x), we would like to consider the possibilities for Ο†X for a fixed Ο†G. If we consider the graph of Ο†X, this is just a section of XΓ—Xβ€²/X because G acts freely, this is the same as a section of (XΓ—Xβ€²)/Gβ†’X/G.

Theorem 2.2.1.

If X′ is contractible, then there exists a unique φX:X→X′ compatible with φG.

Proposition 2.2.2.

For any compact group G, there exists a contractible space E⁒G with a free G-action.

Corollary 2.2.3.
  1. 1.

    E⁒G is unique up to homotopy.

  2. 2.

    E⁒G is functorial in G.

  3. 3.

    For any free action of G on X, the map X→X/G is the pullback of

    for some map X/Gβ†’B⁒G. Therefore, we see that

Proof of proposition.

For all G compact, there exists an embedding GβŠ†U⁑(n), so it suffices to consider U⁑(n). Consider the embedding U⁑(n)β†ͺMat⁒(n,N)rank=n for some N≫0. For example, we have the action of U⁑(1) on β„‚nβˆ–0≃SNβˆ’1, so we can consider S∞, which is contractible. Therefore we have B⁒U⁒(1)=S∞/U⁑(1)=β„‚β’β„™βˆž.

In general, we can consider the action of U⁑(n) on

and this last space becomes contractible as Nβ†’βˆž. To see this, it sits inside of (SN)n, and thus as Nβ†’βˆž, we obtain a contractible subspace of (S∞)n. Finally, we have B⁒U⁒(n)=Gr⁑(n,∞). β–‘

Therefore, we have a tautological U⁑(n)-bundle on Gr⁑(n,∞) and a tautological β„‚n-bundle where the fiber above a subspace is the subspace itself. The vector bundle is the associated bundle of the U⁑(n)-bundle and the U⁑(n)-bundle is the bundle of unitary operators on the vector bundle. Also, we have proved that

The same statement holds for O⁑(n), real vector bundles, and the real Grassmannian. Explicitly over B, consider the exact sequence

Now the kernel defines a map Bβ†’Gr⁑(Nβˆ’n,N)≃Gr⁑(n,N). Taking Nβ†’βˆž, we obtain the desired result.

Remark 2.2.4.

The spaces E⁒G and B⁒G are naturally approximated by algebraic varieties such as Gr⁑(n,N) and are therefore ind-schemes.

We want to show that β„βˆžβˆ–0 is contractible. We can consider the function T on β„βˆžβˆ–0 given by T⁑(x1,x2,…,0,…)=(0,x1,x2,…). Then x,T⁒x are never collinear for xβ‰ 0 and thus T is homotopic to the identity. But then T⁒x is never collinear to e1=(1,0,…,) and therefore T is nullhomotopic. We also see that Cβˆžβˆ–0 is contractible. Also, SβˆžβˆΌβ„βˆžβˆ–0 is contractible and so are Stiefel manifolds

The Stiefel manifold has a free action of G⁒L⁒(n,β„‚), so this is E⁒G⁒L⁒(n,β„‚). We also see that B⁒G⁒L⁒(n,β„‚)=Gr⁑(n,∞,β„‚). We know that the maximal compact subgroup of G is homotopy equivalent to G, so B⁒G⁒L⁒(n,β„‚)=B⁒U⁒(n).

Remark 2.2.5.

The high-brow way that we prove that all of these spaces are contractible is by proving that they are weakly contractible, and then smashing the remaining part that weakly contractible implies contractible for spaces with the homotopy type of a CW complex.

Now the cohomology Hβˆ™β’(B,G) is intimately connected to characteristic classes for principal G-bundles. If G acts on a manifold M, we have a map 𝔀=𝖫𝗂𝖾⁑(G)β†’Vect⁑(M), so we may consider the Lie derivative Lx:Ξ©k⁒Mβ†’Ξ©k⁒M for xβˆˆπ”€. For the Lie derivatives, recall the identity [Lx,ΞΉy]=ΞΉ[x,y]. This gives us a super Lie algebra 𝔀̂ with 𝔀̂1=β„‚d, 𝔀̂0=𝔀, and π”€Μ‚βˆ’1β‰…Ad⁑𝔀. This gives us the super-Lie bracket

Now we see that [𝔀̂1,𝔀̂0]=0, [𝔀̂1,π”€Μ‚βˆ’1] is given by the Cartan formula, and [𝔀̂0,π”€Μ‚βˆ’1] is the adjoint action of 𝔀 on itself. Therefore, if M is a manifold with a G-action, then Ξ©βˆ™β’M is a supercommutative DG algebra with an action of 𝔀̂.

If the action of G is free, we may choose a G-invariant metric on M. For every v∈Tm⁒M, we have its projection onto Tm⁒G⁒m≃𝔀. This gives us a connection, which is a G-invariant 1-form with values in 𝔀 and thus gives us a map

Therefore, if a map π”€Μ‚β†’π’œβˆ™, where π’œ is a supercommutative DG algebra means that G acts on M, we would like to give an interpretation of a connection Ξ±:π”€βˆ—β†’π’œ1 such that [α⁑(ΞΎ)]⁒(ΞΉx)=⟨ξ,x⟩.

Theorem 2.2.6.

There exists a unique acyclic supercommutative DG algebra with H0⁒(π’œβˆ™)=β„‚,Hi⁒(π’œβˆ™)=0,i>0. We will denote this universal algebra by 𝔼.

Proof.

Set π’œ0=β„‚ and π’œ1=α⁑(π”€βˆ—) with Lx acting by the coadjoint action. We set d:π’œ0β†’π’œ1 to be the zero map. We also set ΞΉx⁒(α⁑(ΞΎ))=⟨ξ,x⟩. Because d:π’œ1β†’π’œ2 must be an isomorphism, we see that π’œ2=β⁑(π”€βˆ—)βŠ•β‹€2π’œ1 with the coadjoint action of 𝔀. Then we define ix by

Note that 𝔼 looks like Ξ©βˆ™β’π”€, which are polynomials in x multiplied by β‹€dxi. For Ξ²βˆˆπ’œ2 and Ξ±βˆˆπ’œ1, we can define dβˆ—Ξ²β‘(ΞΎ)=Ξ±,dβˆ—Ξ±β‘(ΞΎ)=0. Therefore, dβˆ— is a derivation, and the Laplacian

is the identity on π’œ1 and π’œ2. This implies that multiplication by (k+β„“) on (𝔼1)k⁒(𝔼2)β„“ is homotopic to the identity and thus Hβˆ™β’(𝔼)=β„‚.

Now we return to our manifold M and base B=M/G. Then the image of Hβˆ™β’(B)β†ͺHβˆ™β’(M) are the so-called basic forms, which vanish on ΞΉx (horizontal) and are G-invariant. In particular, they are killed by both ΞΉx,Lx. Then the map Hβˆ™β’(M)β† Hβˆ™β’(G) has kernel the horizontal forms, so we need to consider the horizontal forms.

Proposition 2.2.7.

Horizontal forms are generated by curvatures, which have the form

where Ξ΄ is the map π”€βˆ—β†’β‹€2𝔼1 given by the transpose of the Lie bracket.

We now have

because α⁑(ΞΎ)⁒[x,βˆ’]=βˆ’Ξ±β‘(adx⁑ξ). Therefore we can write

Therefore, we have 𝔼horizontal=Sβˆ™β’π”€βˆ— and 𝔼basic=(Sβˆ™β’π”€βˆ—)G with zero differential, so we have H2β£βˆ™β’(𝔼basic)=(Sβˆ™β’π”€βˆ—)G. Now we have a transgression1 map H2⁒m⁒(𝔼basic)β†’H2⁒mβˆ’1⁒(G), which is the β€œinverse” Ο€βˆ˜dβˆ—m:𝔼Gβ†’(β‹€βˆ™π”€βˆ—)G of the differential. This vanishes on ((Sβˆ™β’π”€)>02)2.

Corollary 2.2.8.

𝔼horizontal=Sβˆ™β’(curvatures). This is because ΞΉΞΎ:π”€βŠ—π”€βˆ—β†’β„‚ is a perfect pairing.

In some sense, we have proven that

Theorem 2.2.9.

Hβˆ™(B⁒G,β„‚)≃(Sβˆ™β’π”€βˆ—)G.

However, we would like to discuss Hβˆ™β’(G). Now we want to describe the transgression. If p∈Hβˆ™β’(B⁒G), then want to compute dβˆ’1(p)∈Hβˆ™βˆ’1⁒(E⁒G), and then we can restrict this to Hβˆ™β’(G) by killing the horizontal forms. Therefore, polynomial functions on 𝔀 map to polynomial differential forms on 𝔀 by dβˆ—. This gives us something like

and this takes polynomials of degree m to elements of degree 2⁒mβˆ’1. Now in the chain

the choice of dβˆ’1 does not matter. Also, this map is zero on ((S>0β’π”€βˆ—)G)2 because dβˆ’1(p1⁒p2)=p1⁒dβˆ’1p2 if deg⁑p1,deg⁑p2>0, which is killed by transgression. Therefore, primitive elements of (Sβˆ™β’π”€βˆ—)G map isomorphically onto primitive elements of (β‹€βˆ™π”€βˆ—)G. Now we have explained the relationship between Hβˆ™β’(G) and Hβˆ™β’(B⁒G).

We now want to add the third vertex of the triangle, which is Hβˆ™β’(G/T). Recall that any element of 𝔀 is conjugate to some element of 𝔱 and that (Sβˆ™β’π”€βˆ—)G=(Sβˆ™β’π”±βˆ—)W, which is a free algebra. Therefore,

We will see that this is free over (Sβˆ™β’π”±βˆ—)W.

Example 2.2.10.

For G=S⁒U⁒(2) and G/T=S2, we have 𝖫𝗂𝖾⁑𝔱=ℝ and W={Β±1}, and also H2⁒(S2)=ℂ⁑[x]/(x2).

Now we have a map Sβˆ™β’π”±βˆ—/(S>0β’π”±βˆ—)Wβ†’H2β£βˆ™β’(G/T) because the image of the map G/Tβ†’B⁒Tβ†’B⁒G is a point, so we have to kill all positive degree elements in Hβˆ™β’(B⁒G).

Theorem 2.2.11 (Chevalley-Sheppard-Tod).

Let Ξ“βŠ‚G⁒L⁒(V) be finite. Then the following are equivalent:

  1. 1.

    Ξ“ is generated by complex reflections r such that rk⁑(rβˆ’1)=1.

  2. 2.

    ℂ⁑[V]=R is free over the invariants S=ℂ⁒[V]Ξ“.

  3. 3.

    S=ℂ⁑[p1,…,pdimV].

There is a classification of all such groups generated by complex reflections, but Andrei does not remember what it is.

The key to the proof of this theorem is divided difference operators: If sΞ± is a reflection, then Vs is a hyperplane, so for all f∈R, we may consider fβˆ’sΞ±β‹…fΞ±. This vanishes on Vs, lowers the degree by 1, and commutes with x. For example, we could choose

Proof.

Let e1,e2,… be a basis of R/S>0β‹…R. We will show that it is also a basis of R over S. Suppose there exists a relation βˆ‘gi⁒ei=0, where gi∈S. Then either g1βˆˆβˆ‘i>0S⁒gi or e1∈R⁒S>0. Inducting on the degree of e1, if deg⁑e1=0, then

Averaging over Ξ“, we obtain the desired expression. If deg⁑e1>0, then we can apply divided difference operators to reduce the degree. The divided differences of some function f all vanish only when sα⁒f=f for all Ξ±, which is equivalent to f∈RG=S. Therefore we can assume f∈S>0, so if g1=βˆ‘ci⁒gi, then

form another basis of R/S>0. But now we obtain a shorter relation

Therefore we have proved freeness. Finally, we see that the rank of R over S is deg⁑(Vβ†’V/Ξ“)=|Ξ“|.

Now note that Vβ†’V/Ξ“ is flat. Therefore V/Ξ“ is smooth. Recall that if X is a scheme, a point p∈X is smooth if and only if a minimal resolution of β‹―β†’π’ͺXβ†’π’ͺp is finite.2. The most important point to check is 0. Here, have a finite resolution

so after tensoring with R, flatness gives us finiteness for Rβ†’R/R⁒S>0.

To prove the final leg, we will use Molien series. Recall that V/Ξ“ is a cone with vertex 0 and is also smooth, so it must be affine space. Recall that S=⨁iβ‰₯0Si, where S0=β„‚. Then

This implies that the number of generators is dimV/Ξ“=dimV. This implies that |Ξ“|=∏mi. For example, if G=S⁒U⁒(n), then |W|=n!=n⁒(nβˆ’1)⁒(nβˆ’2)⁒⋯⁒2, and the exponents are 1,…,n. The next term in the expansion gives us the subgroup of all reflections fixing Ξ±=0 in Ξ“, and in fact it is cyclic and contained in the roots of unity. Denote a generator by ΞΆk. Then we have

This implies that #⁒reflections=βˆ‘(miβˆ’1). For example, for Ξ“=Sn, we have (n2)=(1βˆ’1)+(2βˆ’1)+β‹―+(nβˆ’1). Now set Ξ“β€²βŠ‚Ξ“ be the subgroup generated by reflections. Then ℂ⁑[pi]=SβŠ‚Sβ€²=RΞ“β€²=ℂ⁑[piβ€²], where piβ€² has degree miβ€² and pi has degree mi. If we order m1≀⋯≀mr and m1′≀⋯≀mrβ€², then miβ‰₯miβ€² because otherwise the p1,…,pi would be algebraically dependent. Thus βˆ‘(miβˆ’1)β‰₯βˆ‘(miβ€²βˆ’1) and in in fact the inequality is strict unless mi=miβ€² for all i. However, both sums count reflections, which is the same for Ξ³,Ξ³β€², so in fact mi=miβ€². This implies that ∏mi=∏miβ€², and thus |Ξ“|=|Ξ“β€²|, so Ξ“=Ξ“β€². β–‘

Now we have a diagram

which commutes. In particular, Hβˆ™β’(G/T,β„‚) is the regular representation of the Weyl group, which follows from the Lefschetz fixed point formula, which states that if f:Mβ†’M, then we have

which says that the number of fixed points with multiplicity is equal to the intersection product of Ξ” and Γ⁑(f). In our case, wβ‰ 1 has no fixed points and thus its trace vanishes. For w=1, all points are fixed, so we obtain χ⁑(G/T)=|W|.

Now recall that H2⁒k⁒(G/T)←ciSkβ’π”±βˆ—. Then Ξ»βˆˆπ”±βˆ— maps to a G-invarint 2-form that restricts to λ⁑([x,y]) at the origin and that Hβˆ™β’(G/T) is the regular representation of W. But then we see that Hβˆ™β’(G/T)W=H0 and thus (S>0β’π”±βˆ—)Wβ†’0. This also follows from the fact that the composition G/Tβ†’B⁒Gβ†’B⁒G maps G/T to a point. Therefore we have a map

Both spaces have dimension W, but we want to know if this is an isomorphism. Equivalently, we want to know whether the cohomology of G/T is generated by c1⁒(taut). Both of these are 0-dimensional Gorenstein rings. Of course, we have HdimM=β„‚β‹…[M] for any closed manifold M, and we call this the socle. This has dimension 1 as a vector space. In addition, we have a perfect pairing

and in particular dimHk=dimHdimMβˆ’k.

Another important example is a 0-dimensional complete intersection ZβŠ‚π”Έd. This has

where f1,…,fd is a regular sequence. For example, a dimension 0 subscheme of the plane generated by a monomial ideal is Gorenstein if and only if it is generated by two elements. Equivalently, its Young diagram is a rectangle. Then the socle of π’ͺZ is spanned by π•œβ’det⁑(βˆ‚fiβˆ‚xj). In the case of a monomial ideal (x1d1,x2d2), this is spanned by x1d1βˆ’1⁒x2d2βˆ’1. Now a map between Gorenstein rings is injective if and only if it preserves the socle. Then we know that Sβˆ™β’π”±W=ℂ⁑[p1,…,pr], where r is the rank, so I=(p1,…,pr). Therefore Sβˆ™β’π”±βˆ—/(S>0β’π”±βˆ—)W is a complete intersection. It has socle given by

and is the first anti-invariant J. This means sα⁒J=βˆ’J and thus the socle is the sign representation of W. But then any f with sα⁒f=βˆ’f has to vanish along Ξ±=0, so α∣f. In particular, ∏α∣J. We conclude that the socle of Sβˆ™β’π”±βˆ—/(S>0β’π”±βˆ—)W is β„‚β‹…βˆΞ± with

Now it suffices to check where ∏α goes in Hβˆ™β’(G/T). We know that ∏α is the volume form on 𝔀/𝔱, so it goes to Htop⁒(G/T)G, as desired.

Remark 2.2.12.

There are other approaches to proving this result. Let M=Gr⁑(k,n) and consider the locus

This is the zero locus of a section of Hom⁑(L1,β„‚n/L2), where L1,L2 are the tautological bundles. This has rank k⁒(nβˆ’k) and thus G⁒r⁒(k,n)=U⁑(n)/U⁑(k)Γ—U⁑(nβˆ’k). Now in Hβˆ™β’(MΓ—M), the class [diag] is given by the characteristic classes of L1,L2. Like in Lefschetz, this acts by the identity operator on Hβˆ™β’(M) and thus Hβˆ™β’(M) is spanned by the characteristic classes of L1.

It remains to discuss the relationship between Hβˆ™β’(G) and Hβˆ™β’(G/T). In principle, we can consider Tβ†ͺGβ†’G/T, but studying this requires spectral sequences. Alternatively, we will study this using the Weyl integration formula. Recall that the map

is generically |W|-to-1. Therefore we have a map

and because the action (g,t)↦(g⁒wβˆ’1,w⁒t⁒wβˆ’1) is free, the middle term is a smooth manifold. This gives us a map Hβˆ™β’(G)β†’H⁒(G/TΓ—T)W. This is a map between Gorenstein rings and is injective on the socle by inspection. Thus it remains to compute the dimension. By the KΓΌnneth formula, we have

The first term is the regular representation of W and the second term is β‹€βˆ™π”±βˆ— with the natural W action. But then we see that H⁒(G/TΓ—T)W has dimension dimβ‹€βˆ™π”±βˆ—=2r. Chasing equivalences, we have

Here, all elements of Sβˆ™β’π”±βˆ— have their degrees doubled. In particular, deg⁑dpi=2⁒miβˆ’1.

Remarks 2.2.13.

Recall that Lie algebra cohomology give the derived functors of Mβ†’M𝔀=Hom⁑(π•œ,M). This is computed by resolving, taking invariants, and then taking the cohomology. This is a general principle in homological algebra, and in fact we can apply this to topological spaces. If G acts freely on M, then M/G is nice (a smooth manifold) and therefore is nice. Otherwise, it is better to consider (MΓ—E⁒G)/G and the fibration Mβ†ͺ(MΓ—E⁒G)/Gβ†’B⁒G. Then we can define the equivariant cohomology

and this is a module over HGβˆ™β’(pt)=Hβˆ™β’(B⁒G). Then we can view Spec⁑HGβˆ™β’(M) (here the Spec is taken as a superscheme) as a sheaf over π”±βˆ—/W. If G acts on M freely, then (MΓ—E⁒G)/G is homotopy equivalent to M/G. The module structure over Hβˆ™β’(E⁒G) via Hβˆ™β’(B⁒G)β†’Hβˆ™β’(E⁒G)=H0. Now we obtain the skyscraper sheaf over π”±βˆ—/W with stalk Hβˆ™β’(M) at the origin.

In this language, we have HTβˆ™β’(G/T)=Sβˆ™β’π”±βˆ—βŠ—(Sβˆ™β’π”±βˆ—)WSβˆ™β’π”±βˆ—, and this has a map to Hβˆ™β’(G/T) killing the positive degree part of the first factor of the tensor product. This recovers the cohomology of Hβˆ™β’(G/T) that we computed before.

Now let V be a rank r vector bundle on X. Then V is given by a map Xβ†’B⁒U⁒(r)=Gr⁑(r,∞). This induces a map Hβˆ—β’(B⁒U⁒(r),β„€)β†’Hβˆ—β’(X,β„€). Now there is a cell decomposition Gr⁑(r,N) into Schubert cells, which is given by the row reduced echelon form of a matrix in Gr⁑(r,N). Now this gives a basis

where Ξ£ ranges over the Schubert cells. Then the character

is the character of an irreducible representation of U⁑(r). Now in terms of the characteristic classes, pulling back ℀⁑[e1,…,er] gives us classes ck⁒(V)∈H2⁒k⁒(X), where ei are the elementary symmetric polynomials.

Now recall that Gr⁑(N,r) parameterizes surjections β„‚Nβ†’Lr, and if Xβ†’πœ‘[𝔖], then we know 𝔖 is a locus where a section of L vanishes, and V=Ο†βˆ—β’L. Then Cr⁒(V)=Ctop⁒(V). Thus if X is a complex manifold, T⁒X has rank dimX, so ctop⁒(T⁒X) is the locus where a vector field vanishes, so

Now c1 parameterizes the locus where r generic sections have rank rβˆ’1, or equivalently det⁑V=0. Therefore c1⁒(V)=c1⁒(β‹€rV). Then ck describes the locus where rβˆ’k+1 sections have rank rβˆ’k. Then the splitting principle tells us that we can write βˆ‘ck⁒(V)=∏(1+xi), where xi are the Chern roots. This is because if 0β†’V1β†’Vβ†’V2β†’0 is an exact sequence, then c⁑(V)=c⁑(V1)⁒c⁑(V2).

Now we will consider the case of a line bundle V=β„’. Then the curvature of β„’ is a class in H2⁒(X,ℝ). Then we can trivialize away from the zero locus of a section s=0, If we draw a loop in the total space of β„’ over Xβˆ–x, then this may not be trivial. If the loop around D2 trivial, then ∫D2curv=0, but if there is a zero in the loop, then the integral is the total angle of rotation 2⁒π times the order of vanishing. Thus c1=curv2⁒π and x1=t12⁒π.

Now if Ο€:XΜ‚β†’X is the flag bundle of V, then Ο€βˆ—β’V=β„’1βŠ•β‹―βŠ•β„’r. Then

Now we can use this to prove that

To see this, note that every symmetric polynomial of degree k is determined uniquely by its values on x=(x1,…,xk,0,…,0). It is enough to consider vector bundles V=Vβ€²βŠ•β„‚rβˆ’k, where Vβ€²=β„’1βŠ•β‹―βŠ•β„’k is a sum of line bundles. Then if s1,…,sk are sections of β„’1,…,β„’k, then all s1,…,sk must vanish, so ek⁒(x)=x1⁒x2⁒⋯⁒xk.

Now we want to see that [𝔖λ]=sλ⁒(x) using equivariant cohomology. Consider the action of G⁒L⁒(N) on Gr⁑(r,N). It is easy to see that the Schubert classes are preserved by the Borel subgroup. Now there is no difference if we take the Borel or the maximal torus A, so we have a map

This is flat of length (Nr) because we are taking the Chern roots x1,…,xr up to permutation. If β„‚nβ† L, then c⁑(L)∣c⁑(β„‚n). Because c⁑(L)=∏(1+t⁒xi) while c⁑(β„‚n)=∏(1+t⁒ai). Now if we consider the vector bundle

we see that ai∈Hβˆ™β’(B⁒A). But then we must have xi=aj for all i and some j.

By flatness, we obtain a result called equivariant formality. If aβˆˆπ–«π—‚π–Ύβ‘(A), the fiber above a is simply Hβˆ™β’(Gr⁒(r,N)a). If a=0, then we obtain Hβˆ™β’(Gr⁑(r,N)), and if a is generic, then Gr⁒(r,N)a is a set of coordinate subspaces and is thus a disjoint union of (Nr) points.

Now if [𝔖λ]βˆˆβ„€[a1,…,aN][x1,…,xr]Sr/∼, this has degree |Ξ»| in x1,…,xr and misses many fixed points. For example, on β„™N, the class (1,βˆ—,…,βˆ—) hits everything, while the class (0,…,0⏟k,1,βˆ—,…) misses anything with a 1 in the first k components. Thus we have a polynomial of degree k in x that vanishes on x=a1,…,ak, so we obtain (xβˆ’a1)⁒⋯⁒(xβˆ’ak)≕pk⁒(x). Now we want a symmetric version of this, which is a Schur function in the

By an interpolation argument, this is exactly [𝔖λ]∈HAβˆ™β’(G). In ordinary cohomology, we set a=0 and obtain the Schur function.

2.3 Minimizing norms

Let G be a compact Lie group. This is associated to a complex Lie algebra 𝔀. Then there exists a simply connected G~ such that 𝖫𝗂𝖾⁑G~=𝔀. By the Lie theorem, we have a central extension

Therefore complex representations of G are the same as complex representations of K. In particular, G is linearly reductive.

On the other hand, by Peter-Weyl, L2⁒(K) has a dense subset given by ⨁VEnd⁑(V), where V ranges over all irreducible representations. This is a finitely generated commutative Hopf algebra generated by matrix elements of a faithful representation. Therefore it is ℂ⁑[G] for some linear algebraic group with 𝖱𝖾𝗉⁑G=𝖱𝖾𝗉⁑K. This also works when K is not connected. For example, if K=(S1)n is abelian, then G=(β„‚Γ—)n.

Now we will prove that if G is a complex reductive group, there exists a compact KβŠ‚G such that 𝖫𝗂𝖾⁑(K)βŠ—β„β„‚=𝖫𝗂𝖾⁑(G). In other words, every complex reductive group is a complexification of a compact Lie group. Now if G is a complex reductive group, then there exists an exact sequence

Then Ad⁑(G) is open in Aut⁑(𝔀), and thus everything is algebraic. Therefore we have an embedding Gβ†ͺG⁒L⁒(β„‚N)βŠƒU⁑(N). Now all of the maximal compact subgroups are conjugate because any compact group preserves some Hermitian metric, so we hope for K=G∩U⁑(n).

What we want to do is to minimize some norm. Consider X=G⁒L/G=Spec⁑ℂ⁒[G⁒L]G. This is a finitely generated algebra with a linear action of G⁒L, so if we take a G⁒L-invariant subspace that contains the generators, then there is a closed embedding Xβ†ͺV into a finite-dimensional G⁒L-module V as a closed orbit. Now choose a U-invariant norm β€–βˆ’β€–2 on V and minimize it on X. Set

Proposition 2.3.1.

dimℝXmin≀dimβ„‚X.

Assuming this, choose x=g⁒G∈Xmin, where g∈G⁒L. Then

and therefore dimβ„‚G≀dimℝgβˆ’1⁒U⁒g∩G. In fact, the real dimension of a compact subgroup is at most the complex dimension of G, and in fact any compact subgroup is totally real in G. Here a submanifold Y of a complex manifold X is totally real if Ty⁒T∩i⁒Ty⁒T=0 for all y∈Y. In particular, we have dimℝXmin=dimβ„‚X and dimℝG∩gβˆ’1⁒U⁒g=dimβ„‚G. Therefore it remains to prove the inequality.

First, note that β€–βˆ’β€–2 is a plurisubharmonic or J-convex function.

Definition 2.3.2.

Let X be a complex manifold and f:X→ℝ be a real function. Then f is plurisubharmonic if (βˆ‚Β―iβ‘βˆ‚jf) is positive semidefinite and strictly plurisubharmonic if (βˆ‚Β―iβ‘βˆ‚jf) is positive definite.

This is equivalent to the function Uβ†’X→𝑓ℝ being subharmonic for all open UβŠ‚β„‚, which is equivalent to the Laplacian being nonnegative (or strictly positive). To see this, note that

and so we have the product of the Hessian and the norm of βˆ‚ziβˆ‚w. Therefore the restriction of a plurisubharmonic function to a complex submanifold remains plurisubharmonic. Thus β€–βˆ’β€–2 is clearly plurisubharmonic because its Hessian is the standard Hermitian metric.

Proposition 2.3.3.

The minima of any strictly plurisubharmonic function are totally real.

Proof.

Note that Tx⁒Xmin is in the kernel of the Hessian of f. Therefore it cannot contain any complex lines and is thus totally real. β–‘

There are some variations. Let XβŠ‚β„‚N be an affine variety (or a Stein manifold) and consider f⁑(x)=β€–xβˆ’pβ€–2 for some fixed p. This is a Morse function for generic p. Then the negative index of f⁑(x) at any critical point x0 is at most dimβ„‚X, so the tangent space contains no complex lines. This implies that X has the homotopy type of a CW complex of real dimension at most dimβ„‚X. This can be found in a 1959 paper of Andreoti and Frankel about the Lefschetz hyperplane theorem.

Theorem 2.3.4 (Lefschetz hyperplane theorem).

Let Z be a smooth projective manifold of dimension n. Let D=π’ͺ⁑(1) be a hyperplane section. Then the restriction map Hi⁒(Z)β†’Hi⁒(D) is an isomorphism for i≀nβˆ’2 and injective for i=nβˆ’1.

Now let G be a complex reductive group and let Gβ†’G⁒L⁒(V) be a representation. Then we know that Spec⁑ℂ⁒[V]G parameterizes closed G-orbits. For example, under the action of β„‚Γ— on 𝔸2 by (t,tβˆ’1), the closed orbits have the form x1⁒x2=cβ‰ 0 and the origin. Here, the only closed orbit in x1⁒x2=0 is the origin. Therefore, we have Spec⁑ℂ⁒[x1,x2]G=Spec⁑ℂ⁑[x1⁒x2]. For any closed orbit we can look for minima of a K-invariant Hermitian metric β€–βˆ’β€–2, where K is a compact real form of G. We may assume that β€–βˆ’β€–2 has the form c1⁒‖x1β€–2+c2⁒‖x2β€–2. We would like to prove the following result:

Theorem 2.3.5 (Kempf-Ness; Matsushita-Onishchik).

The orbit Gβ‹…v is closed if and only β€–βˆ’β€–2 attains a minimum. When this is the case, the minima form a single K-orbit and there are no other critical points of β€–βˆ’β€–2. Finally, the stabilizer of v is reductive.3

Corollary 2.3.6.

We can identify V/G with the quotient of the critical loci of β€–βˆ’β€–2 by K. Equivalently, the moment map ΞΌ:V→𝖫𝗂𝖾⁑(K)βˆ— vanishes.

Returning to our example, the critical locus of β€–βˆ’β€–2 is the set {c1⁒|x1|2βˆ’c2⁒|x2|2=0}, which is the union of two lines. This intersects each orbit exactly once, as desired.

It is clear that if the orbit is closed, the minimum of the norm function is attained. Now we will assume Kempf-Ness and prove Matsushita-Onishchik. Write Xmin for the locus where the minimum of the norm is attained. Then we know that dimℝXmin≀dimβ„‚X=dimβ„‚Gβˆ’dimβ„‚H, where H is the stabilizer of a point v∈Xmin. However, we can extend this to the inequality

This implies that dimβ„‚H≀dimℝK∩H. Of course K∩H is totally real, so the reverse inequality holds and thus H is the complexification of K∩H, so it is reductive. In fact, all inequalities are equalities, so dimℝXmin is the dimension of any K-orbit contained in Xmin. This implies that Xmin is smooth. We will also see later that Xmin is connected, so it forms a single orbit.

Now for g∈G⁒L⁒(n), write g⁒gβˆ—βˆˆG⁒L⁒(n)/U⁑(n). This is some Hermitian metric, so G⁒L⁒(n)/U⁑(n) can be identified with the space of Hermitian metrics on β„‚n. Then we know that

Recall that for GβŠ‚G⁒L⁒(n) reductive, there exists some metric β€–βˆ’β€–2 such that G∩U⁑(n) is a compact real form K. Then we have a morphism G/Kβ†’G⁒L⁒(n)/U⁑(n), and the image is exp⁑(i⁒𝖫𝗂𝖾⁑K). Then we know every element of 𝖫𝗂𝖾⁑(K) is conjugate to an element of 𝖫𝗂𝖾⁑(T) for a maximal torus TβŠ‚K, Therefore we can represent G/K as k⁒A⁒kβˆ’1 for some K, where A=exp⁑(i⁒𝖫𝗂𝖾⁑T)βŠ‚G. In particular, we have the decomposition G=K⁒A⁒K. In our previous example, G=β„‚βˆ—,K=U⁑(1),A=ℝ>0.

This means we can write gβ‹…v=k1⁒a⁒k2⁒v, and only a changes β€–βˆ’β€–2. Now if Ξ±βˆˆπ–«π—‚π–Ύβ‘A, we know

and therefore we can write

This is a convex nonnegative function. When the Ξ±i are not all of the same sign, we have a unique maximum, but not necessarily when all Ξ±i>0. We can consider the weights of the T-action on V in the characters of T. In the first case, the Newton polytope contains the origin, so our function is strictly convex and bounded below. In the second case, the Newton polytope does not contain the origin, so the closure of A⁒v contains some other vector vβ€² and thus is not closed. This implies K⁒A⁒K⁒v is not closed. There is a third case in which 0 is contained on the boundary on the Newton polytope. In this case, we have a semistable point.

In conclusion, of β€–βˆ’β€–2 has a local minimum, then the orbit is closed and there are no other critical points. In addition, for any maximal torus TβŠ‚K, the Newton polytope of the weights of the T-action on V contains the origin. This is called the Hilbert-Mumford criterion.4

Now consider the action of K on coadjoint orbits in 𝖫𝗂𝖾⁑(K)βˆ—. Then every ΞΎβˆˆπ–«π—‚π–Ύβ‘(K) defines a function which is the Hamiltonian for the vector field adβˆ—β‘(ΞΎ). Recall that the coadjoint orbits are Poisson manifolds. For a basic example of a Hamiltonian, the function H=12⁒(p2+q2) gives us the flow which is rotation with velocity 1. Changing coordinatse, we see that H=12⁒‖zβ€–2 generates ei⁒t and H=m2⁒‖zβ€–2 generates ei⁒m⁒t. If we write i⁒m=ΞΎβˆˆπ–«π—‚π–Ύβ‘(U⁑(1)), then Ξ±=ΞΎiβˆˆπ–«π—‚π–Ύβ‘A. Of course, we have β€–et⁒α⁒zβ€–2=22⁒m⁒t⁒‖zβ€–2 and thus

We obtain the same formula when ΞΎ is a larger diagonal matrix, where H=12β’βˆ‘mi⁒|zi|2. In this case, we have

This implies that critical points of β€–βˆ’β€–2 are the same as the zeroes of the moment map.

Now we can generalize this in several directions:

  1. 1.

    We can consider GIT quotients. For example, if we have a character ΞΎ:Gβ†’β„‚βˆ—, then we can replace V/G by

    Instead of μ=0, we can consider μ=±dχ.

  2. 2.

    Many moduli problems are fomally quotients by infinite-dimensional groups. Then the moment map equations (or minimization of β€–βˆ’β€–2) are very useful and important PDEs. A very classical example of this is the work of Hitchin-Kobayashi-Donaldson-Uhlenbeck-Yau-…who studied stable holomorphic vector bundles. In fact, stable holomorphic bundles are precisely those with Hermitian Yang-Mills connection. These minimize β€–curvatureβ€–L22.

    For example, if we consider a curve C and a line bundle of degree 0, this line bundle lives in Jac0⁒(C). Then we are looking for flat unitary line bundles, which have the form (C~Γ—β„‚)/Ο€1⁒(C) under a map Ο€1⁒(C)β†’U⁑(1). Therefore our line bundles are parameterized by

    In fact, this is isomorphic to the Jacobian as a smooth manifold. More recently, there is the work of Chen-Donaldson-Sun in higher dimension.

2.4 Symmetric spaces

Now let GβŠ‚G⁒L⁒(V) be a complex semisimple group and K be a maximal compact. Then K is fixed under σ⁑(g)=(gβˆ—)βˆ’1, which is an antiholomorphic automorphism of G. Now consider a G orbit of β€–βˆ’β€–2, which is thet set {g⁒gβˆ—βˆ£g∈G} of positive self-adjoint elements of G. Now T1⁒(G)=𝔨+i⁒𝔨. Also, G/K has a G-invariant metric, and this is unique up to multiple if K is simple. Also, Οƒ preserves the metric, so it acts as an isometry of G/K, which is a symmetry about 1∈G/K.

Definition 2.4.1.

Let M be a Riemannian manifold. We say M is a Riemannian symmetric space if for all m, there exists some isometry Οƒm fixing m such that Οƒm acts by βˆ’1 on Tm⁒M.

In particular, G/K is an example of a Riemannian symmetric space. There is a structure theory of Riemannian symmetric spaces, and the classification is essentially the classification of all real Lie groups. Here are some properties of symmetric spaces. If m,mβ€² lie on a geodesic, then Οƒm′⁒σm is translation by the distance between mβ€²,m. In particular, M is complete. In addition, the isometry group G of M acts transitively on M.

This implies that M=G/H where H is the stabilizer of some point m∈M. In particular, H is compact because it is a closed subgroup of O⁑(Tm⁒M). We also know that H commutes with Οƒm, so Οƒm∈Aut⁑(G) is an inner automorphism. Now we have

where 𝖫𝗂𝖾⁑(H) are transformations in O⁑(Tm⁒M) and Tm⁒M moves m. We see that 𝖫𝗂𝖾⁑(H) is the +1 eigenspace and Tm⁒M is the βˆ’1 eigenspace. Therefore a Riemannian symmetric space M is the same as a Lie group G with a choice of automorphism Οƒ and compact subgroup (GΟƒ)1βŠ‚HβŠ‚GΟƒ, where M=G/H.

Now consider the symmetric space Sn. If we choose (x1,…,xn,xn+1)=(0,…,0,1), the involution fixing the north pole is the map

which has two fixed points. This means that Οƒm does not have to have a unique fixed point. Now if we conjugate O⁑(n+1) by Οƒ=(11β‹±1), then the fixed locus is

Now if we take H=O⁑(n), we can remove the Β±1 if we take Sn/Β±1=ℝℙn. Now the classification of Riemannian symmetric spaces is as follows:

  1. 1.

    Spaces of the form G/K, where G is a real Lie group and K is a maximal compact group. These have negative curvature.

  2. 2.

    Spaces of the form KΓ—K/Ξ”, where K is a compact Lie group. Thus we have M=K with the action of K by conjugation.

  3. 3.

    There are compact summetric spaces with positive curvature, and these are dual to the negative curvature case.

An important example of a symmetric space is the hyperbolic space ℍn=S⁒O⁒(n,1)+/S⁒O⁒(n). This is dual to Sn=S⁒O⁒(n+1)/S⁒O⁒(n). The classification reduces again to the classification of root systems, which sit inside the class of finite reflection groups. Of course, finite reflection groups sit inside discrete reflection groups. In fact, the maximal symmetric spaces have H=O⁑(Tm⁒M), and these are classified by

We also must have [𝔰⁒𝔬⁒(n),ℝn]βŠ‚β„n and [ℝn,ℝn]βŠ‚π”°β’π”¬β’(n) because 𝔰⁒𝔬⁒(n) is the positive eigenspace of some involution and ℝn is the negative eigenspace. This gives us an 𝔰⁒𝔬⁒(n)-equivariant map β‹€2ℝn→𝔰⁒𝔬⁒(n). These are both irreducible and isomorphic, so these are multiplication by a real constant c. If c>0, then 𝔀=𝔰⁒𝔬⁒(n+1) and G/H=Sn. If c=0, then 𝔀=𝔰⁒𝔬⁒(n)⋉ℝn and G/H=ℝn, and if c<0, then 𝔀=𝔰⁒𝔬⁒(n,1) and G/H=ℍn.

3 Semisimple Lie algebras

3.1 Reflection groups and root systems

Now we want to consider discrete isometry groups Ξ“βŠ‚π–¨π—Œπ—ˆβ‘(M) where M is one of Sn,ℝn,ℍn. This is in fact equivalent to the fact that for all x∈M, Γ⁒xβŠ‚M is discrete. Therefore the subsets {MΞ³}Ξ³βˆˆΞ“ are also discrete in M. To see this, consider any point m. Then the distance d⁑(m,γ⁒m)≀2⁒d⁒(m,MΞ³). Thus if MΞ³ is not discrete, the distances d⁑(m,γ⁒m) accumulate. If Ξ“ is generated by reflections, then the set of reflecting hyperplanes is discrete.

Now we will consider the Dirichlet fundamental domain

This is a locally finite intersection of half-spaces. We are mainly interested in the case where M is compact, so we will have a globally finite intersection. In the compact case, we may have to subdivide the boundaries βˆ‚π’Ÿ such that each facet F corresponds to a particular Ξ³FβˆˆΞ“:

π’ŸΞ³Fβ’π’ŸFmay have tosubdivide
Figure 3.1: Subdivision of facet

In the globally finite case, we obtain Ξ³FiβˆˆΞ“. These generate Ξ“ and the relations are as follows: Note that F′≔γFβˆ’1⁒(F) is a facet of π’Ÿ. It is also easy to see that Ξ³Fβ€²=Ξ³Fβˆ’1. In the case of a reflection group, then F=Fβ€² and Ξ“F2=1. In other words, {Ξ³F} is a symmetric set of generators. We also have more complicated relations coming from codimension 2 strata. If the fixed hyperplanes of r1,r2 have angle p⁒iP⁒m, then r1⁒r2 is rotation by 2⁒θ, so we obtain the relation (r1⁒r2)m=1. In fact, if we have the square lattice, then Ξ³1⁒γ2=Ξ³2⁒γ1.

Theorem 3.1.1.

The relations above are all relations between the Ξ³F. This means that

Proof.

Consider the exact sequence 1β†’kerβ†’Ξ“~=βŸ¨Ξ“F∣relationsβŸ©β†’Ξ“β†’1. Now consider

But now we have a covering

However, Mβˆ–codimΒ β‰₯3 is 1-connected, so this covering is an isomorphism. β–‘

Therefore it remains to classify possible π’Ÿ. First, however, we will consider a very classical example.

Example 3.1.2.

Consider a triangle with angles Ο€m12,Ο€m13,Ο€m23. Then we can consider the discrete reflection group Ξ“=⟨s1,s2,s3∣si2=1,(si⁒sj)mi⁒j=1⟩. We have the following cases for where the triangle lives:

Thus in the first case Ξ“ is finite, in the second case it has polynomial growth, and in the final case it has exponential growth. To define the growth, we need to consider the length function on Ξ“. Here, we count the number of hyperplanes separating π’Ÿ,Ξ³β’π’Ÿ, and this is the length ℓ⁑(Ξ³), which is also the length of the shortest Ξ³=si1⁒⋯⁒siβ„“. Some nice properties are that ℓ⁑(Ξ³)=ℓ⁑(Ξ³βˆ’1) and that ℓ⁑(γ⁒si)=ℓ⁑(si⁒γ)=ℓ⁑(Ξ³)Β±1. Then we consider the function that takes x to the number of words with length less than x, and consider the growth of this function.

There are finitely many solutions to 1m12+1m23+1m12β‰₯1, and these correspond to platonic solids (spherical) or regular tesselations of the plane (Euclidean).

Now we are ready to classify discrete reflection groups acting on Sn or ℝn. We can embed SnβŠ‚β„n+1, so the two cases are really the same. We want to impose that the normal vectors to the bounding hyperplanes span (ℝn)βˆ—. We also assume that there is no partition of the normal vectors into two mutually perpendicular sets. We will carry out the classification using the connection to complex semisimple Lie algebras (or compact semisimple real Lie algebras). Consider the outer normals ei and consider the Coxeter matrix (ei,ej)i⁒j. Now we know

Because this is a Gram matrix, it is positive semidefinite. Also, the diagonal emenets are all positive and all off-diagonal elements are nonpositive.

Lemma 3.1.3.

Together with indecomposability, the two properties imply that either the ei are linearly independent (in which case π’Ÿ is a cone over a tetrahedron, which is the image of ℝβ‰₯0n under a linear map) or that ker⁑Q=ℝ⋅v, where v has all positive coordinates.

In the first case, π’Ÿ is the cone over a simplex and Ξ“ is finite. In the second case, π’Ÿ is a simplex, so it is the image of {xiβ‰₯0βˆ£βˆ‘xi=1}. In this case, we obtain an irreducible reflection group in Iso⁑(ℝn).

Proof.

Consider the matrix Q as a quadratic form. Then the kernal ker⁑Q={v∣Q⁑(v,v)=0}. For v∈(v1,…,vn), denote |v|=(|v1|,…,|vn|). Thus Q⁑(|v|,|v|)≀Q⁑(v,v). Thus if Q⁑(v,v)=0, we also have Q⁑(|v|,|v|)=0. Now suppose that some vi=0 and Q⁒v=0. Then (Q⁒v)i=βˆ‘jβ‰ iqi⁒j⁒vj. If there exist iβ‰ j such that qi⁒jβ‰ 0, then Q⁑(|v|,|v|)<Q⁑(v,v). This is impossible if Q is indecomposable, so no vi can vanish for v∈ker⁑Q. Thus dimℝker≀1. Moreover, the kernel is closed under vβ†’|v|, so it must be spanned by a vector with vi>0. β–‘

This implies that the classification of reflection groups in Sn,ℝn is the same as the classification of indecomposable positive-semidefinite Coxeter matrices. This is explicit and fun, and involves solving inequalities.

Now let 𝔨=𝖫𝗂𝖾⁑(K) for a compact group K with maximal torus T. Then write 𝔀=π”¨βŠ—β„β„‚ and 𝔱=𝖫𝗂𝖾⁑(T)βŠ—β„β„‚. Then we have a decomposition

and we know 𝔀α is 1-dimensional if and only if Ξ± is a root. Next, consider eΞ±βˆˆπ”€Ξ±,fΞ±βˆˆπ”€βˆ’Ξ±. Then we write hΞ±=[eΞ±,fΞ±]βˆˆπ”±, and this is unique. Then we have (hΞ±,h)=α⁑(h) and [hΞ±,eΞ±]=2⁒eΞ± and [hΞ±,fΞ±]=βˆ’2⁒eΞ±. Therefore we have an embedding 𝔰⁒𝔩⁒(2)Ξ±βŠ‚π”€, so an S⁒L⁒(2)Ξ±βŠ‚G. Next, consider the map

We need 2⁒(Ξ±,Ξ²)(Ξ±,Ξ±)βˆˆβ„€ so that sα⁒(Ξ²) is in the lattice spanned by the roots. The reflections in the roots Ξ± generate a finite group W of reflections, which also preserves the root lattice. Now we may consider the Cartan matrix (2⁒(Ξ±i,Ξ±j)(Ξ±i,Ξ±i)). This is an integer matrix with 2 on the diagonal and nonpositive off diagonal entries. The Cartan matrix must be positive definite.

Now each reflecting hyperplane is the zero set of a linear function (which is a root). Now we can choose a fundamental domain, and this gives us a partition of roots into positive and negative roots. After reflection by some si, if Ξ²β‰ Ξ±i is a positive root, then Ξ²βˆ’Ξ±i is still positive after Ξ±i. Therefore every positive root has a sum of the form Ξ²=βˆ‘mi⁒αi, where the Ξ±i are the simple positive roots (corresponding the the faces). Moreover, this expression is unique (because the simple roots are independent as affine linear functions).

Now recall that for a semisimple Lie algebra 𝔀, we have

where Ξ± ranges over the roots. We know that [h,eΞ±]=α⁑(h)⁒eΞ±, that [𝔀α,𝔀β], that dimℂ𝔀α=1, and that π”€Ξ±βŸ‚π”€Ξ² if Ξ±+Ξ²β‰ 0. Also, there exist eΒ±Ξ±βˆˆπ”€Β±Ξ± such that (eΞ±,eβˆ’Ξ±)=1. If we write hα≔[eΞ±,eβˆ’Ξ±], then (hΞ±,h)=α⁑(h). Now our goal is go from root systems to Lie algebras.

Because every root is either positive or negative, and every positive root can be written uniquely as a sum of simple roots, we can write

Note that if 𝔀 is finite-dimensional, then 𝔫+ is nilpotent.

Proposition 3.1.4.

The nilpotent Lie algebra 𝔫+ is generated by ei:eΞ±i, where {Ξ±1,…,Ξ±r} are the simple positive roots.

Proof.

We show that if Ξ±+Ξ² is a root, then [𝔀α,𝔀β]=𝔀α+Ξ². Consider the action of (𝔰⁒𝔩2)Ξ± on 𝔀n⁒α+Ξ². Now each 𝔀nΞ±+Ξ² is one-dimensional. Now recall that the weights of irreducible representations with respect to hΞ± are Ξ»,Ξ»βˆ’2,…,2βˆ’Ξ»,βˆ’Ξ». But now recall that fΞ± decreases the weight by 2 and eΞ± increases the weight by 2, and so the action of eΞ±:𝔀β→𝔀β+Ξ± is nonzero, as desired. β–‘

Corollary 3.1.5.

The Lie algebra 𝔀=⟨hi,ei,fi⟩/relations, where ei=eΞ±i,fi=eβˆ’Ξ±i,hi=hΞ±i, where the Ξ±i are the simple roots.

The relations are given by

We will write ai⁒j≔αi⁒(hi). We also have [hi,fj]=βˆ’ai⁒j⁒fj and [ei,fj]=Ξ΄i⁒j⁒hi.

Lemma 3.1.6 (Chevalley-Serre).

For all i,j, we have ad⁑(ei)1βˆ’ai⁒j⁒ej=0.

For example, for the A2 root system with Cartan matrix (2βˆ’1βˆ’12), we have ad⁑(e1)2⁒e2=ad⁑(e2)2⁒e1=0.

Proof.

Write β–‘=ad⁑(ei)1βˆ’ai⁒j⁒ej. We will prove that if [fk,β–‘]=0, then ad⁑(𝔀)⁒░ is invariant under 𝔀. Then β–‘ is a lowest weight vector, so ad⁑(𝔀)β’β–‘βŠ‚π”«+ would be a proper submodule, contracticting the simplicity of 𝔀.

Now we need to prove that [fk,β–‘]=0. This is clear for kβ‰ i,j, so we need to check it for k=i and k=j. For k=i, we have weights ai⁒j,ai⁒j+2,…,βˆ’ai⁒j for hi (by the action of (𝔰⁒𝔩2)i on ej), so β–‘ corresponds to the weight βˆ’ai⁒j+2, so we must have β–‘=0 (because otherwise it would generate an infinite-dimensional module). Now we consider the case when k=j. Here, we have

Now if ai⁒j<0, we see [ei,ei]=0 and if ai⁒j=0, we also have vanishing. β–‘

Theorem 3.1.7.

For a simple Lie algebra 𝔀, we have 𝔀=⟨hi,fi,ei⟩/relations with the Cartan matrix C=(ai⁒j)=2⁒(Ξ±i,Ξ±j)(Ξ±j,Ξ±j). Moreover, for any matrix C such that ai⁒i=2, ai⁒j≀0, and C is symmetrizable, then the matrix 𝔀 is simple and finite-dimensional if and only if C corresponds to a finite root system.

Remark 3.1.8.

Note that the conditions that ai⁒i=2,ai⁒j≀0, and C is symmetrizable can be relaxed in principle.

3.2 Kac-Moody Lie algebras

Let C=(ai⁒j) be a matrix such that ai⁒i=2,ai⁒j≀0, and ai⁒j=0 if and only if ai⁒j=0. Note this is weaker than being symmetrizable. Now we define the Lie algebra

Now we have the decomposition 𝔀~=π”₯βŠ•π”«~+βŠ•π”«~βˆ’, where 𝔫~+ is the free Lie algebra generalized by the ei, whose universal enveloping algebra is a free associative algebra. Note if C=(2), then 𝔀~=𝔰⁒𝔩2. Also, we have

so in fact all roots are already either positive or negative. Here, note that

and we also know that

The previous argument for β–‘=ad⁑(ei)1βˆ’ai⁒j⁒ej shows that it is a lowest weight vector in 𝔫+. Now we know that 𝔫+βŠ‡ad⁑(𝔀~)⁒░. More generally, we can consider the maximal submodule of ad⁑(𝔀~) (which cannot contain the ei).

Definition 3.2.1.

We define the Kac-Moody algebra 𝔀KM to be the quotient of 𝔀~ by the maximal submodules in 𝔫+ and π”«βˆ’.

Theorem 3.2.2 (Gabber-Kac).

If C is symmetrizable, then the maximal submodules in 𝔫+,π”«βˆ’ are generated by the Chevalley-Serre relations.

Theorem 3.2.3.

If dim𝔀KM<∞, then C corresponds to a finite reflection group.

Remark 3.2.4.

There are even larger generalizations of this, for example Borcherds-Kac-Moody Lie algebras and other classes of Lie algebras.

Now Chevalley-Serre tells us that for large enough M, ad⁑(ei)M applied to any generator vanishes. Thus for any xβˆˆπ”€KM, there exists M=M⁑(x) such that ad⁑(ei)M⁒x=ad⁑(fi)M⁒x=0. I particular, any xβˆˆπ”€KM is contained in a finite-dimensional 𝔰⁒𝔩⁒(2)i=⟨ei,fi,hi⟩-module.

By definition, a 𝔀KM-module is integrable if for all 𝔰⁒𝔩⁒(2)i, the module decomposes as a sum of finite-dimensional modules. Thus, in this language, the adjoint representation of 𝔀KM is integrable. Now we can integrate each 𝔰⁒𝔩⁒(2)i-module to S⁒L⁒(2)i, and the matrix (0βˆ’110) acts by reflection ri. Then we see that

and the ri generate the Weyl group of 𝔀KM. In particular, it permutes the set of roots.

Recall that when we had an invariant quadratic form, we had the identity

Therefore there exists an invariant quadratic form if and only if there exists a diagonal D such that C⁒D is symmetric. Now note that 𝔀~=π”₯βŠ•π”«+βŠ•π”«βˆ’, and 𝔫+,π”«βˆ’ are dual with respect to the quadratic form. In fact, we obtain expressions of the form

Now if we consider the ideal of relations 𝔯+βŠ‚π”«+, then we note that

Now note that any invariant bilinear form on 𝔀 is nondegenerate because π”€βŸ‚ is an ideal which does not intersect π”₯, and thus vanishes. If ai⁒iβ‰ 0, then we can assume ai⁒i=2, and therefore we have a subalgebra 𝔰⁒𝔩⁒(2)i=⟨ei,fi,hi⟩, then for jβ‰ i, ej is the lowest weight vector for 𝔰⁒𝔩⁒(2)i with weight ai⁒j. If ai⁒j∈0,βˆ’1,βˆ’2,…, then it is possible that the 𝔰⁒𝔩⁒(2)i-module is finite-dimensional. Kac-Moody theory is about integrable modules. Also, recall that ad⁑(ei)1βˆ’ai⁒j⁒ej=0βˆˆπ”€, and by Gabber-Kac this is the complete list of relations. Also recall the adjoint representation is integrable.

Example 3.2.5.

Consider C=(2a12a212). Note that if we want det⁑C>0, we can only have (a12,a21)=(βˆ’1,βˆ’1),(βˆ’2,βˆ’1),(βˆ’3,βˆ’1). Now the three cases correspond to the Lie algebras 𝔰⁒𝔩3,𝔰⁒𝔭⁒(4)=𝔰⁒𝔬⁒(5),G2. In the case where det⁑C=0, or (a12,a21)=(βˆ’2,βˆ’2),(βˆ’4,βˆ’1), then we need to enlarge π”₯ to have Ξ±j⁒(hi)=ai⁒j. The extra elements in π”₯ will be central, and we will need to consider a central extension of 𝔰⁒𝔩⁒(2)βŠ—β„‚β‘[tΒ±1]. Here, we will have

Now we have

These are called affine Lie algebras, and these correspond to reflection groups in ℝn. Now consider the case when det⁑C<0. This is much more complex, and suppose a12=a21=βˆ’m, where m>4. Then the quadratic form β€–(x,y)β€–2=2⁒x2βˆ’2⁒m⁒x⁒y+2⁒y2 has signature (1,βˆ’1). Now after we kill the Serre relations, we have our Weyl group W=⟨r1,r2⟩, where

It is easy to see that r2⁒(1,0)=(1,m),r2⁒(0,1)=(0,βˆ’1),r1⁒(0,1)=(m,1),r1⁒(1,0)=(βˆ’1,0). Now the vectors with β€–βˆ’β€–2=2 form a hyperbola, and it is easy to see that r1,r2 are translations on the hyperbola. Note that these are solutions to x2βˆ’m⁒x⁒y+y2=1 and are units in β„šβ‘(m2βˆ’4). Also, note that

is periodic and the coefficients are 1,mβˆ’2.

The fact that there are no roots of 𝔀~ in β„€β‰₯0×℀≀0βˆ–{(0,1),(1,0),(0,0)} and the W-action implies that there are no roots with β€–βˆ’β€–2>2. Now for Ξ²=m1⁒α1+m2⁒α2, we would like to minimize the height m1+m2. Now if (ri,Ξ²)=Ξ²βˆ’(Ξ±i,Ξ²)⁒αi, then we can decrease the height unless Ξ²=Ξ±i. Now in every W-orbit of a root Ξ², either there is a simple root Ξ±i, so β€–Ξ²β€–2=2, or there is a root with (Ξ²,Ξ±i)>0 for all i. Now roots are either real (which means dim𝔀β=1) or imaginary (which implies exponential growth in the cone bounded by lines of slope m/2,2/m). In general, if the signature of our matrix (nβˆ’1,1), there is a similar picture, where we have a light cone (notation stolen from physics, all hyperboloids will asymptotically converge to this), and inside we have the funramental domain. Unfortunately, there is only a compact fundamental domain in low dimension, so the theory fails in high dimension. If the signature is worse, then the theory gets worse. However, every root is either real or belongs to the orbit of a fundamental domain (which is the set of vectors such that (Ξ²,Ξ±i)≀0 for all but finitely many roots).

It should be clear that dim𝔀<∞ if and only if C>0. This explains why we end up with infinite-dimensional Lie algebras when we attempt classification.

3.3 Integrable representations of Kac-Moody Lie algebras

Let 𝔀=π”₯βŠ•π”«+βŠ•π”«βˆ’ be a symmetrizable Kac-Moody Lie algebra and C=(ai⁒j) be the Cartan matrix. Then the Weyl group W=⟨ri⟩ acts on π”₯,π”₯βˆ—.

Now if M is a module and v∈M, by definition v is a highest weight vector if hi⁒v=λ⁑(hi)⁒v for some λ∈π”₯βˆ— and ei⁒v=0 for all i. Some examples of highest weight modules are Verma modules MΞ», which is the free module generated by a highest weight vector vΞ». Equivalently, we can write

This is not even close to being integrable (just note we can act infinitely many times by fi). To make MΞ» integrable, we need λ⁑(hi)=0,1,… to be a nonnegative integer and fiλ⁑(hi)+1⁒v=0. This gives us a candidate for an irreducible integrable module LΞ» with highest weight1 Ξ». We will see later that this is indeed irreducible by the classification of irreducible finite-dimensional modules. We will also compute the character using the Weyl-Kac formula.

We will now consider the Casimir element. First suppose that dim𝔀<∞ and let xi,yi be dual bases for the Killing form (βˆ’,βˆ’). Then the Casimir element is

Now because 𝔀α is dual to π”€βˆ’Ξ±, we have

This acts by a scalar in MΞ» because it commutes with all fi, and so

For any h∈π”₯, we have (h,hΞ±)=α⁑(h)⁒(eΞ±,fΞ±). Now consider

Lemma 3.3.1.

For all r, ri⁒(ρ)=Οβˆ’Ξ±i. In other words, ρ⁑(hi)=1.

Proof.

Note that ri permutes the set of positive roots distinct from Ξ±i. Therefore ri⁒(Ξ±i)=βˆ’Ξ±i, and the desired result follows by splitting the expression for ρ. β–‘

Remark 3.3.2.

Some analog of this also makes sense if dim𝔀=∞.

We can take any other function of the set {Ξ±>0βˆ£Ξ±β‰ Ξ±i}, such as x↦1βˆ’eβˆ’x. Then the product

is a Laurent series in eβˆ’Ξ±i which is often analytically convergent. Under the action of ri, we obtain

Therefore this product is W-anti-invariant. Note that the product ∏α>011βˆ’eβˆ’Ξ± is the character of hπ”₯ acting on 𝒰⁑(π”«βˆ’). Here, the operators 1,fΞ±,… have weight 1,eβˆ’Ξ±.eβˆ’2⁒α,…, and taking the sum, we obtain the expression. Therefore, the character of a Verma module is

Theorem 3.3.3 (Weyl-Kac).

The character of the integrable representation LΞ» is given by

Returning to the Casimir element, we now write

For infinite-dimensional modules with highest weight, this expression makes sense because the final term in the sum is locally finite. Here, hi is the dual basis to the hi. This still commutes with the fi. Therefore C acts on MΞ» by (Ξ»,Ξ»)+2⁒(Ξ»,ρ)=(Ξ»+ρ,Ξ»+ρ)βˆ’(ρ,ρ).

To prove this, consider the exact sequence

Here, M is some highest-weight module, so it has a surjection from ⨁Mμi. Thus there exists a resolution of Lλ by Verma modules of the form

By the action of the Casimir element, we have β€–ΞΌi+ρ‖2=β€–Ξ»+ρ‖2. In addition, the character of LΞ» is W-invariant by integrability. Together, we obtain that the character of LΞ» is a sum of the form

where Ξ»βˆ’ΞΌ is either a sum of positive roots or zero, cΞ»=1, β€–ΞΌ+ρ‖2=β€–Ξ»+ρ‖2, cΞΌ is anti-invariant under the action of W by wβ‹…ΞΌ=w⁑(ΞΌ+ρ)βˆ’Ο.

Lemma 3.3.4.

Every ΞΌ such that Ξ»βˆ’ΞΌ is a sum of roots can be brought by the Wβ‹… action to the positive cone (where (ΞΌ,Ξ±i)β‰₯0). The positive cone is also called the dominant cone.

Proof.

Choose w such that in the expression Ξ»+Οβˆ’w⁑(ΞΌ+ρ)=βˆ‘mi⁒αi, βˆ‘mi is minimal. If (ΞΌ,Ξ±i)<0, then we can apply ri and decrease mi. β–‘

Lemma 3.3.5.

If (Ξ»,Ξ±i)>0,(ΞΌ,Ξ±i)β‰₯0, Ξ»β‰₯ΞΌ, and (Ξ»,Ξ»)=(ΞΌ,m⁒u), then Ξ»=ΞΌ.

Proof of this result is left as an exercise to the reader.

As a consequence, we have:

  1. 1.

    If cΞΌβ‰ 0, then the intersection of Wβ‹…ΞΌ with the dominant cone is simply Ξ», and thus cμ⁒(βˆ’1)w⁒δWβ‹…Ξ».

  2. 2.

    In the exact sequence ⨁𝒰⁑(π”«βˆ’)⁒fiλ⁑(hi)+1⁒vβ†’MΞ»β†’cokerβ†’0, the first term is simply ⨁riMriβ‹…Ξ». Now the cokernel is irreducible because it is integrable and therefore there is no room for other singular vectors. In fact, this can be continued to a Bernstein-Gelfand-Gelfand resolution of LΞ» by

As a corollary of Weyl-Kac we obtain

Corollary 3.3.6 (Denominator identity).

Apply the formula to L0=β„‚ to obtain

and rearrange to obtain

For example, when 𝔀=𝔰⁒𝔩⁒(n), the positive roots are of the form (0,1,…,βˆ’1,0), so eΞ±=xixj. Then we have ρ=((nβˆ’1)/2,…,(1βˆ’n)/2), and therefore we have

This has the more familiar form

For 𝔰⁒𝔩̂2, this gives us the Jacobi triple product identity for theta functions

3.4 Affine Lie Algebras

Let Ξ“=⟨ri⟩ be a group of isometries in ℝn. For example, if r1,r2,r3 are reflections along the sides of an equilateral triangle, we have the AΜ‚2 group. Now isometries of ℝn embed in isometries of ℍn+1 because we can embed ℝn as a horocycle. Now consider the group AΜ‚1 generated by r1,r2 with r1⁒r2 of infinite order. If we consider spheres Sn in ℍn+1, we have an action of S⁒O⁒(n) on Sn which commutes with the action of S⁒O⁒(n,1) on ℍn+1. Note that S⁒O⁒(n) stabilizes a vector of negative norm. In the limit as the radius approaches ∞, the curvature approaches 0, so we obtain ℝn.

Now the semidirect product S⁒O⁒(nβˆ’1)⋉ℝn stabilizes a null vector, so it acts on ℝnβŠ‚β„n+1. Now if we fix that S⁒O⁒(n,1) preserves the form x0⁒xn+x12+β‹―+xnβˆ’12, then if we consider the light cone where β€–xβ€²β€–2+x0⁒xn=0, then our isometry group acts in a way such that orbits are like parabolas. In the case of AΜ‚1, the orbit of a single point in ℝ1 will lie on a parabola when we embed in ℝ2,1.

Now our reflection group AΜ‚1 could have Cartan matrix (2βˆ’2βˆ’22) or (2βˆ’4βˆ’12). The first corresponds to a central extension 𝔰⁒𝔩̂⁒(2) of 𝔰⁒𝔩⁒(2)βŠ—β„‚β‘[t,tβˆ’1] while the second corresponds to a twisted loop algebra. Here, if 𝔀 is a Lie algebra and Οƒ is an automorphism of order m, then we may consider the subalgebra

For example, note that 𝔰⁒𝔩⁒(2)⁒[tΒ±1] is generated by the matrices

Then [e1,f1]=h1=βˆ’[e0,f0]. If we define h0≔[e0,f0], then for C=(2βˆ’2βˆ’22), h0+h1 is contained in the kernel of the map

Now note that if we set deg⁑e0=(1,0),deg⁑e1=(0,1), then the Lie algebra ⟨h0,e0,f0,h1,f1,e1⟩ is graded by β„€2. Because C is degenerate, only one of these two gradings is internal. To restore the grading, we introduce the element D≔t⁒ddt, where D⁒e0=e0 and D⁒f0=βˆ’f0. Now we set

where we set h2=D. Now we have a map

Now because there are three simple roots, the matrix of the invariant bilinear form is

and this has signature (2,1). In the basis of C=(h0+h1),h1,D, this takes the form

which is precisely of the form x0⁒xn+1+x12+β‹―+xn2.

Now we would like to describe the central extension in terms of matrices. Here, we have

then we can write 1=Res⁑tr⁑P′⁒(t)⁒Q⁒(t)=Res⁑(ddt⁑P⁑(t),Q⁑(t)) if we set e1=P⁑(t),f1=Q⁑(t). Using integration by parts, this is skew-symmetric. Now we have

Now the eigenvalues of (C,h1,D) on each factor are (0,2,n),(0,βˆ’2,n),(0,0,n). The values x+n⁒(nβ‰₯0),βˆ’x+n⁒(nβ‰₯1),n⁒(nβ‰₯1) are positive for x∈(0,1). Now if we define

then if we set exp⁑(0,0,βˆ’1)=q,exp⁑(0,1,0)=x, we obtain

This is simply the genus 1 theta function ϑ⁑(x), which has the property that ϑ⁑(1)=0 and that

Therefore Ο‘ is a section of a degree 1 line bundle over 𝔼≔ℂ×/qβ„€. Note that Ξ” converges analytically if |q|<1. Because βˆ’qβˆ’1⁒xβˆ’1 is invertible, we may use it as a clutching function. By Riemann-Roch, Ο‘ only has a single zero at x=1.

Another way to write solutions of the equation ϑ⁑(q⁒x)=βˆ’qβˆ’1⁒xβˆ’1⁒ϑ⁒(x) is using the form βˆ‘xm⁒qm2/2. Here we want to think of m22 as (m2), so if we apply x↦q⁒x, we have

and thus we have Ο‘~⁒(q⁒x)=xβˆ’1⁒(~⁒ϑ)⁒(x). Therefore ϑ⁑(x)β‰ˆconst⁒ϑ~⁒(βˆ’q⁒x), and thus everything is unique up to a multiple section of a line bundle on 𝔼. Classically, this is the Jacobi product formula. As a sum, we really have the sum in the Weyl-Kac formula, because

This also satisfies some q-difference equation, so is a section of a Line bundle of 𝔼rank⁑𝔀. In general, we have

and this can be rewritten as an expression in the characters of 𝔀finite is we sum over W first and as a theta function if we sum over the lattice first.