LIE THEORY AND TOPOLOGY

ALLEN KNUTSON
1. MOTIVATION OF LIE GROUPS
Dene a topological Lie group to be a group object in the category of topological man-
ifolds. Until stated otherwise G is assumed nite-dimensional.
Theorem 1.1. Let Gbe a compact, simple Hausdorff group. Then either Gis a nite simple group
or it is Lie.
The latter are much easier to classify: they are U(n), O(n), U(n, H) mod their centers, or one
of ve special cases (of dimensions 248), each more or less blamable on the octonions.
Nonexamples: pronite groups are compact and not Lie, but not simple.
Theorem 1.2. Any closed subgroup of a Lie group is Lie.
Well dene a Lie group to be a group object in the category of smooth manifolds. The
principal example is GL
n
(R).
Theorem 1.3. Any topological Lie group is uniquely (equivariantly) smoothable, and indeed,
uniquely real-analytic.
Moreover, any measurable homomorphism of Lie groups is automatically continuous smooth
real-analytic.
However, not every Lie group is real algebraic, one standard example being

SL
2
(R).
Proof: the center of an algebraic group is algebraic, and a discrete algebraic variety (of
nite type) should be nite, but Z(

SL
2
(R))

=
1
(SL
2
(R))

= Z.
Differential topology is a great place to work, in that any smooth Lie group comes
with an adjoint action on the tangent space g at the identity. Differentiating that map
G End(g) at the identity, we get the map ad : g End(g).
Theorem 1.4. Each connected normal subgroup H of G gives a subrepresentation h g.
The group structure on G induces a Lie algebra structure on g, but well have very
little use for it. In particular there are many subalgebras that dont come from closed sub-
groups, because of irrational-ow-on-a-torus problems. It is interesting to note, though,
that every Lie algebra has a faithful matrix representation, and that the map ad only de-
pends on the Lie algebra.
Hereafter the term simple group will mean one whose normal subgroups are nite
rather than trivial.
Theorem 1.5. If G is connected and simple, its normal subgroups are central, and G/Z(G) is
simple in the usual sense.
Date: Draft of December 5, 2011.
1
2 ALLEN KNUTSON
2. REDUCTION TO COMPACT GROUPS
Recall that any group has a unique maximal normal solvable subgroup. In a topological
group we can add the adjective connected.
Dene the Killing form on g to be
X, Y := Tr(ad X ad Y).
This extends to a GG-invariant pseudoRiemannian metric on G.
Theorem 2.1. (1) If G is simple, this metric is unique.
(2) If G is compact, this metric is negative semidenite, and its kernel is tangent to Z(G).
(3) If G is simple and this metric is denite, then G is compact. (Since ad only depends on the
Lie algebra, this says that if G, G

have the same Lie algebra and G is compact with nite

center, then G

is also compact.)
(4) The radical of the Killing form is the tangent space to the unique maximal normal solvable
connected subgroup.
A group is semisimple if its Killing form is nondegenerate.
Theorem 2.2. (1) (Levi 1905) There exists a semisimple Levi subgroup complementary to
the radical of the Killing form.
(2) (Mal

cev 1942) Any two such choices are conjugate by elements of the maximal normal
solvable connected subgroup.
So far weve split the group into the kernel of , and a semisimple complement. We
can go further inside the semisimple part, essentially splitting into positive and negative
parts:
Theorem 2.3. (Iwasawa 1949) Let G be connected and semisimple. Let K be a maximal com-
pact connected subgroup. Then there exists a complementary subgroup A N, where A is Ad-
diagonalizable and N is Ad-unipotent, and AN is diffeomorphic to a vector space. In particular K
is a deformation retract of G.
There is another decomposition G = KP, where P is the exponential of k

g and is
not a subgroup.
If G is real algebraic, N will be too but A need not be. We can replace A by its Zariski
closure A

, but then the map K A

N G is only onto not bijective.

3. COMPACT CONNECTED LIE GROUPS
Well usually use K for a compact connected Lie group.
Example 1. S
1
.
Example 2. T = (S
1
)
n
, called an n-torus.
Example 3. If
1
(K) is nite, then

K.
Theorem 3.1. If Z(K) is nite, then
1
(K) is too.
Example 4. K
1
K
2
. . . K
m
.
Example 5. K/, where Z(K).
By the long exact sequence, if
1
(K) = 0, then
1
(K/) = .
LIE THEORY AND TOPOLOGY 3
Theorem 3.2. Any K is a quotient of K

:= T K
1
. . . K
m
by a nite central subgroup,
where each K
i
is simple and simply-connected. We can take T = Z(K)
0
, and K
1
. . . K
m
as the
universal cover of the commutator subgroup K

.
Note that the quotient map K

K induces an isomorphism of their Lie algebras.

Each K
i
is real algebraic, so K is algebraic too. Indeed, there is a complex matrix group K
C
containing K as a subgroup, homotopy retract, and xed points of an antiholomorphic involution.
In particular, if K has a nite center, then there are nitely many connected groups with
its Lie algebra, and they are all compact. (Whereas a torus has the same Lie algebra as a
vector space.)
Example 6. U(n)

= (T
1
SU(n))/Z
n
.
Example 7. SU(2)/Z
2

= SO(3). That gives three obvious quotients of SU(2) SU(2).
The nonobvious one turns out to be SO(4), where one thinks of U(1, H) U(1, H) acting
on H by left and right multiplication. People make a big deal about this when studying
the hydrogen atom, e.g. http://math.ucr.edu/home/baez/classical/runge.pdf.
Example 8. For n > 2,
1
(SO(n)) = Z
2
. (Proof: SO(3)

= RP
3
, and we get a long exact
sequence on homotopy using the bration over the sphere SO(n)/SO(n1).) Its universal
cover is called Spin(n). Z(SO(odd)) = 1, so Z(Spin(odd))

= Z
2
, but Z(SO(even))

= Z
2
,
so |Z(Spin(even)| = 4. Strangely, that center is Z
4
for n 2 mod 4, Z
2
Z
2
for n 0 mod 4.
Example 9. Spin
c
(n) := (Spin(n)S
1
)/(Z
2
)

, so quotients to SO(n) but not to Spin(n).

Example 10. SU(n), SO(n), U(n, H). These collide at U(1, H)

= SU(2) SO(3), also at
Spin(4)

= SU(2)
2
as mentioned above, again at U(2, H)

= Spin(5), and also at Spin(6)

=
SU(4) (as the image of SU(4) U(Alt
2
C
4
) is SO(6)),
Example 11. G
2
= Aut(O) is 14-dimensional, simple, centerless, simply-connected.
Example 12. F
4
= Aut(Herm
3
(O)) is the automorphism group of the exceptional
Jordan algebra of real dimension 27, the 3 3 octonionic-Hermitian matrices, under
anticommutator. F
4
is 56-dimensional, centerless, simply-connected.
Example 13. E
6
also acts on the Jordan algebra, but only preserving the determinant
3-form, or one can see it as the collineation group of the octonionic projective plane. (Note
that collineation groups are usually semidirect products, with one factor being the auto-
morphism group of the skew-eld involved!) It is 78-dimensional, and the center of its
simply-connected form is Z
3
.
Example 14. E
7
is a 133-dimensional group, that acts on a 64-fold known as the quate-
roctonionic projective plane, and the center of its simply-connected form is Z
2
.
Example 15. E
8
is 248-dimensional, centerless, and simply-connected. Its smallest linear
representation is its adjoint representation, which is pretty obnoxious if you ask me.
4. AVERAGING
Theorem 4.1. (Haar 1933) On any locally compact group G there is a left-invariant measure,
unique up to scale. Its total volume is nite exactly if G is compact (in which case we often scale it
to be 1).
If G has no one-dimensional representations, then its left-invariant measures are also right-
invariant. (Non-example: the {ax +b} solvable group.)
4 ALLEN KNUTSON
If G is Lie, existence is really easy; we even have left-invariant volume forms (unique
up to scale).
Theorem 4.2. On a compact group K, there exists a biinvariant Riemannian metric. The expo-
nential map k K dened using its geodesics is K-equivariant.
(If K has discrete center, the negative of the Killing form will do.)
Proof. Pick any metric and average it under the K K-action.
Theorem 4.3. If K acts on a real resp. complex vector space V, then V can be given a K-invariant
orthogonal resp. Hermitian form.
Thus, any K-subrepresentation of V has a K-invariant complement.
Thus, any nite-dimensional representation of K is a direct sum of irreducible representations.
Lots more to say about that, of course! Lets start with
Lemma 4.4. Let T act linearly on V = R
n
. Then V is the direct sum of V
T
and a bunch of
2-dimensional irreps. In particular dim
R
V dim
R
V
T
mod 2.
If T acts linearly on V = C
n
, then V is a direct sum of complex-one-dimensional irreps.
Proof. Pick a T-invariant orthogonal form on V, so we can split complements U

to sub-
representations.
If v V C is an eigenvector for t, then v + v, i(v v) generate a T-subrep U

= R
2
.
If weve already split off the invariants, then U must be irreducible, since any homomor-
phism of T O(1) is trivial.
The complex case is similar but simpler.
5. MAXIMAL TORI AND WEYL GROUPS
A maximal torus T of a compact Lie group is what youd expect.
1
Let t T be a
topological generator, i.e. t = T (thats most elements). Then
K t

= K/C
K
(t)

= K/C
K
(T)
and since T is a maximal torus, the group C
K
(T)/T must be nite, so this space K t is a
nite quotient of K/T. (Later well see that if K is connected, then C
K
(T) = T.)
The xed points of t on K/T are also xed by T;
TkT/T = kT/T k N(T)
so the xed points correspond to W := N(T)/T, the Weyl group.
Lemma 5.1. (1) If K is compact positive-dimensional, then K contains a circle.
(2) If T is maximal, then C
K
(T)/T is nite. (Later we will showthat if Kis connected, C
K
(T) =
T.)
1
For noncompact groups, one generalizes the notion of torus to a connected group that acts diagonal-
izably on g, so e.g. the diagonal matrices in SL
n
(R) count as a torus.
LIE THEORY AND TOPOLOGY 5
Proof. Pick X k \0, and let J = exp(R X). Then J is connected, and abelian, so exp : j J
is a group surjection. Since J is connected and Hausdorff, ker exp is a lattice so J is a torus.
Then tori contain circles.
If C
K
(T)/T isnt nite, then we can choose a circle subgroup J

in C
K
(T), and take its
preimage J in C
K
(T). This is a compact connected group, so we know what it looks like.
But for T to be codimension one in it, J has to be a torus. In which case T wasnt maximal.

Theorem 5.2. The Weyl group is nite.

Proof. We need to show N
K
(T)
0
= T, so W is discrete, but also compact hence nite. We
already noted that C
K
(T)
0
= T.
The kernel of the conjugation action of N(T) on T is C
K
(T), so there is a faithful action
of a nite quotient of W on T. But Ts automorphism group is GL
dimT
(Z), so discrete.
If we can pick X t, such that exp(X) = t, then the isolated xed points says that Xs
vector eld has isolated zeros. To apply Poincar e-Hopf, we need to compute indices:
Lemma 5.3. If V
T
= {

0}, and X t such that t := exp(X) topologically generates T, then the

index of Xs unique xed point is 1.
Proof. By assumption, V = U
i
, where U
i
is a 2-d irrep of T. The homomorphism T
O(U
i
)

= O(2) lands inside SO(2), so gives index 1. When we take the sum of all the
irreps, we take the product of the indices and get 1.
Theorem 5.4. (1) The Euler characteristic of K/T is |W| ,= 0.
(2) The Lefschetz number of any k acting on K/T is |W| ,= 0.
(3) Any element is contained in a conjugate of T.
(4) Any two maximal tori are conjugate, in particular, all of the same dimension, called the
rank of K. (E.g. Unitary matrices are diagonalizable.)
Proof. (1) The vector eld given by X is almost complex, so all the indices are positive.
(Actually theyre 1.)
(2) Since K is connected, any ks Lefschetz number is the Euler characteristic.
(3) Since k must have a xed point, kxT = xT, we learn k xTx
1
.
(4) Let k be a topological generator of a second maximal torus, and apply the previous.

One view is that maximal tori are analogous to Sylow subgroups, and this ,= 0 argu-
ment takes the place of the counting argument that one uses to prove that all p-Sylows
are conjugate. Well see another application of this analogy later.
Factoid: if K is simply connected, then every abelian subgroup lies in a maximal torus.
Non-examples:
The diagonal matrices in SO(3) dont live in any maximal torus (SO(2), or a con-
jugate).
There exist innite-dimensional symplectomorphismgroups with nite-dimensional
maximal tori of differing dimensions.
6 ALLEN KNUTSON
The idea that we can cover a group using conjugates of an abelian subgroup will turn
out to be incredibly great. Note that we can never cover a nite group G with conjugates
of a proper subgroup H (even nonabelian):
|
gG
gH| = |
gG/N(H)
gH| < |

gG/N(H)
gH| =

gG/N(H)
|gH| = |G/N(H)||H| = |G|/|N(H)/H| |G|.
The rst can only be = if N(H) = G, and the second only if N(H) = H.)
6. MORSE THEORY ON COADJOINT ORBITS
Itd be nice if Xs vector eld was something like the gradient of a Morse function, so
we could do Morse theory on K t. The natural place for X k to induce a function is
on k

. Since exponential maps are local diffeomorphisms, we get to roughly correspond

the conjugation orbits of K on K with the orbits of K on k, and on k

(again using the

K-invariant metric).
One reason that g

is a great space to work on, in general, is that it has an interesting

Poisson structure, which is a shadow of the fact that
Fun(g

) = Sym(g) = gr Ug.
The Poisson bivector (Alt
2
Tg

) is easy to write down:

(X, Y)|

:= ([X, Y]), X, Y (T

= g
Its symplectic leaves turn out to be exactly the orbits of G, called coadjoint orbits.
Example: K = U(n). Then we can equivariantly identify k

with the space of Hermitian

matrices, and the orbits O

are isospectral sets, i.e. correspond to spectra (

1
. . .
n
).
The simplest case is = (1, 1, 1, . . . , 1, 0, . . . , 0) with k 1s, in which case the map M
Image(M) corresponds O

with the Grassmannian of k-planes.

More generally, we can take M to the nested list (

top i eigenspaces), corresponding

it to a partial ag manifold {(V
1
< V
2
< . . . < C
n
)}. It is rather amazing that the real
vector space k

has been naturally partitioned into compact complex manifolds! (Of course
symplectic manifolds are always even-real-dimensional, but they are not always complex
[Thurston 76].)
In Morse theory, one uses a Riemannian metric to build a vector eld from a function
(or really, from the 1-form obtained as its derivative). What is perhaps unsatisfying is
that the resulting vector elds do not annihilate the metric or the function. If one uses
a Poisson bivector instead, the resulting Hamiltonian vector elds do annihilate the
Poisson tensor and the function.
Theorem 6.1. If K is compact, then any coadjoint orbit can be given a K-invariant Riemannian
metric compatible with its symplectic structure, such that the result is almost K ahler.
(In fact it will be honestly complex, but we wont use that.)
Proof. Pick a compatible almost complex structure, get a metric, average it.
Lemma 6.2. Let T act on a Hermitian vector space V

=

i
C

i
, where each C

i
is an irrep with
character T

. Then for X t, the function

f
X
: V R, (z
1
, . . . , z
m
)

1
2
|z
i
|
2
X,
i

LIE THEORY AND TOPOLOGY 7

has Poisson gradient equal to the action of X, and is unique with this property up to addition of a
constant.
If no X,
i
= 0, then it is a Morse function, with index 2#{
i
: X,
i
< 0}.
Theorem 6.3. K/T has an even-dimensional cell decomposition, with cells indexed by W. In
particular, (K/T) = |W|, and K/T is simply-connected. Moreover C
K
(T) = T.
Proof. Fix a K-invariant metric on k. Pick k

such that the corresponding

k is
within the injectivity radius of exp : k K, and such that t := exp(

) topologically
generates T. Then
K

= K t

= K/C
K
(t)

= K/C
K
(T).
On k

, we have the linear functional X, which restricted to K

gives a function. We can

analyze its critical points using the Hamiltonian vector eld (which is not the Riemannian
gradient) to determine that they are isolated and even index.
Let T be a maximal torus, t T a topological generator within the injectivity radius of
exp, and X t its logarithm. Then X denes a map

X
: K X R, , X
which will turn out to give a Morse decomposition into even-dimensional cells, gener-
alizing the usual one for projective space. In the next section we gure out what the
dimensions are of those cells.
6.1. The Weyl group is a reection group. Note rst that W acts on t preserving the
Killing form, so, orthogonally.
Theorem 6.4. Let T act on g/t, a sum of 2-irreps.
(1) These irreps U are all nonisomorphic.
(2) Each t + u is the tangent space to a subgroup L
U
isomorphic to SU(2) T
n1
, possibly
mod (Z
2
)

.
(3) W(L
U
) W is isomorphic to Z
2
.
Because complex representation theory is easier than real, well often complexify k to
k
C
:= C k. Well use T

to denote the set of one-dimensional complex representations or

weights of T, which we can identify with a lattice in t

by
: T U(1)

: t iR i

.
Then W preserves T

inside t

.
Example: K = U(n), W

= S
n
, T

= Z
n
, with the usual action.
The root system T

of K is the set of weights in (g/t) C. So each U in the above

decomposition gives a pair of such roots, and well use L

or L

in place of L
U
.
If we pick a random functional X t, i.e. no X, = 0, then we can use it to dene a
positive system
+
. Let
1

+
denote the simple roots, that arent sums of other
positive roots.
Theorem 6.5. (1) The action of W(L
U
) on T

is by reections.
(2)
1
is linearly independent.
(3) The reections {r

W(L

)}

1
generate W.
8 ALLEN KNUTSON
(4)
1
spans a lattice inside T

of index |Z(K)|, called the root lattice.

Proof. (1) We already calculated that the action of W
U(n)
on T

is the standard action

of S
n
on Z
n
, which barely modied tells us that W
SU(2)
acts on its T

= Z by reec-
tion (negation). Enlarging SU(2) to SU(2) T
n1
just gives the reection a large
invariant space.
Because K preserves the denite form on k

, N(T) preserves the restriction of

that form to t

.
(2) The proof goes by a better fact: each ,

0. Anyway its ndable in any Lie

textbook.
(3) This is too. Its really a statement about reection groups, not Lie theory. (Note,
though, that it uses Kconnected; otherwise we could take the Z
3
half of N(T) inside
SU(3).)
(4) The kernel in K of the adjoint action is Z(K) C
K
(T) = T; write Ad T for T/Z(K), a
maximal torus of K/Z(K). The U(1)-dual of the exact sequence of compact groups
1 Z(K) T Ad T 1
is the exact sequence of discrete groups
0 Z(K)

root lattice 0.
http://en.wikipedia.org/wiki/Pontryagin_duality

At this point we can jack into the purely combinatorial theory of reection groups, to
learn things like
Theorem 6.6. If w W, then |
+
\ w
+
| = (w), the length of w written as a minimal
product of simple reections.
Corollary 6.7. The Morse cell of
X
, coming down from the point wT/T K/T, has dimension
2(w). In particular, if Z(K) is nite then H
2
(K/T)

= Z
dimT
.
Dene the height of a root as the sum of the coefcients when is expanded in
the simple roots
1
. (Then if we let
k
denote the set of roots of height k, it motivates the
notation

and
1
.)
7. TOPOLOGY OF K/T
The long exact sequence on homotopy for T K K/T gives us
. . .
2
(T)
2
(K)
2
(K/T)
1
(T)
1
(K)
1
(K/T)
0
(T) = 1.
We know the homotopy groups of T. But lets be more precise than just to say
1
(T)

=
Z
dimT
; we can identify it as a group with the coweight lattice , the kernel of the expo-
nential map t T. (This is the Z-dual of the weight lattice T

.)
1
2
(K)
2
(K/T)
1
(K)
1
(K/T) 1
Since K/T is simply-connected, we can kill the last one, and invoke Hurewicz:
1
2
(K) H
2
(K/T)
1
(K) 1
LIE THEORY AND TOPOLOGY 9
So in particular
1
(K) is abelian (but the Eckman-Hilton argument lets one see that for
any group).
The Z-dual of this map H
2
(K/T) is a map H
2
(K/T) T

, and turns out to be

the map c
1
(K
T
C

) taking a representation to the rst Chern class of its associated

bundle, under the associated-bundle construction for the principal bundle K K/T.
Theorem 7.1.
2
(G) = 1 for any nite-dimensional Lie group.
Proof. Retract G to K. The above sequence says
2
(K) is free abelian, so nonzero iff
H
2
(K; R) ,= 0. We can compute the latter using de Rham cohomology, as follows. In-
side the complex of forms lies the nite-dimensional subcomplex of left-invariant forms,
and we can average forms to show that the inclusion induces a homotopy equivalence
of the complexes. Then the Lie algebra cohomology H
2
(k) turns out to be measuring
nontrivial central extensions, and there are none.
So for instance, any principal bundle over S
3
with nite-dimensional structure group
must be trivial! (Of course the tangent bundle is, because S
3
is a group, but not every
Z
2
-bundle over S
1
is trivial, for instance.)
Corollary 7.2. If K is centerless, then
1
(K) is nite.
Proof. If K is centerless, than
1
spans T

, so in
1 H
2
(K/T)
1
(K) 1
H
2
(K/T) and have the same rank.
It turns out that
3
(K) is always Z for a compact simple group (e.g. SO(5) but not
SO(4)). Basically, each L

K gives the generator of

3
, remembering that

L

= S
3
.
7.1. Schubert calculus. Let S
w
H

(K/T) denote the cohomology class dual to the cycle

X
w
owing down into the point w, so deg S
w
= 2(w), and c
w
uv
Z denote the structure
constants in the multiplication S
u
S
v
=

w
S
w
.
Theorem 7.3. (Kleiman 1973) Each c
w
uv
is nonnegative.
Proof. In fact K/T is a complex manifold, and each X
w
is a subvariety, called a Schubert
variety.
We can compute

K/T
S
w
S
v
by intersecting X
v
with w
0
X
w
. Morse-theoretically, we get
the closure of the union of the lines from w
0
w down to v.
(1) If w
0
w , v, this is empty.
(2) If w
0
w > v, this is positive-dimensional.
(3) If w
0
w = v, this is just a point.
So

K/T
S
w
S
v
=
w
0
w,v
. (N.B. We need to use the fact that this Morse function is Palais-
Smale.)
Hence c
w
uv
=

S
u
S
v
S
w
0
w
= |X
u
(g X
v
) (h S
w
0
w
)| for generic g, h K, because the
complex structure guarantees that each point in that triple intersection contributes 1 to
the cohomological intersection.
10 ALLEN KNUTSON
There are many extensions of this result to other cohomology theories (e.g. equivariant
quantum K-theory), but manifestly positive formul for very few of them. The simplest
(and most important) case is when K t is a Grassmannian, and there we do have many
rules.
http://www.math.cornell.edu/
~
allenk/plenary.pdf
7.2. A non-topological applications of the Morse theory: Horns inequalities.
Lemma 7.4. Let f be a Morse function on M with C the set of critical points, and M =

C
M
c
the Morse decomposition. If m M
c
, then f(m) f(c).
Theorem 7.5. (Helmke-Rosenthal 1995) Let H
a
+ H
b
+ H
c
= 0, where each H
d
is a Hermitian
matrix with spectrum (d
1
d
2
. . . d
n
). Let (, , ) be a triple of Schubert classes on
Gr
k
(C
n
) such that

,= 0. Then
a + b + c 0
where we consider , , as vectors from {0
nk
1
k
}.
Proof. Each H
d
gives a Morse function V Tr(H
d

v
) on the k-Grassmannian, whose
critical points come when V is a sum of eigenlines of H
d
. By the integral, there exists a
V Gr
k
(C
n
) in the intersection of the three Morse strata for the three different Morse-
Schubert stratications. Hence
0 = Tr(0) = Tr((H
a
+H
b
+H
c
)
V
) = Tr(H
a

v
) +Tr(H
b

v
) +Tr(H
c

V
) a+ b+ c.

Klyachko proved that these give all the inequalities on a, b, c. Belkale proved that its
enough to consider

= 1. Tao, Woodward, and I proved that all those remaining

inequalities are indeed necessary. Much more about this is at the URL above.
8. A HINT OF REPRESENTATION THEORY
Theorem 8.1. Let V, W be reps of an arbitrary group G.
(1) Hom(V, W) and V

W are reps, and the natural map V

W Hom(V, W) is G-
equivariant.
(2) If V is nite-dimensional, then g Tr(g|
V
) is constant on conjugacy classes, and called
the character of V.
(3) If V, W are isomorphic nite-dimensional reps, they have the same character.
(4) Let Hom
G
(V, W) := Hom(V, W)
G
, the equivariant maps or intertwiners. If V, W are
irreducible, then dimHom
G
(V, W) = [V

= W] (Schurs lemma).
If G is compact:
(1) Tr(g|
V
) = Tr(g|
V
).
(2) Let
G
|
V
=

G
g|
V
. Then
G
is a projection V V
G
.
(3) Consequently, if V is nite-dimensional, Tr(g|
V
) = dimV
G
.
Theorem 8.2. The characters of the irreps of a compact group K are orthonormal elements of the
Hermitian vector space L
2
(K; C).
LIE THEORY AND TOPOLOGY 11
Proof.
Tr(g|
V
), Tr(g|
W
) :=

K
Tr(g|
V
)Tr(g|
W
) =

K
Tr(g|
V
)Tr(g|
W
)
=

K
Tr(g|
V

W
) =

K
Tr(g|
Hom(V,W)
) = dimHom(V, W)
K
= dimHom
K
(V, W)
and that is 1 or 0 by Schurs lemma.
Of course, theyre not a basis for L
2
(K), since theyre constant on conjugacy classes;
really we might hope that they be a basis for L
2
(K/) (as indeed they are). We study that
space in the next section.
Corollary 8.3. Let V, W be reps of a compact connected group K, with maximal torus T. Then
V, W are isomorphic if they isomorphic as T-representations.
The nicest way to write down a T-representation is as a function T

N, taking
dimHom
T
(C

, V) =

T
t

Tr(t|
V
).
If the representation comes from K, then this multiplicity diagram will be W-invariant.
When dimHom
T
(C

, V) > 0, call a weight of V.

The biggest theorem in the subject requires a concept early from the next section.
9. CONJUGACY CLASSES
Each denes a hyperplane in t, and this gives a decomposition of t into top-
dimensional cones called Weyl chambers. It is a wonderful result in nite reection
groups that W acts simply transitively on the set of Weyl chambers. Having picked
+
,
we have exactly broken this W-symmetry to have a positive Weyl chamber
t
+
:= {X : X, 0
1
}
which provides a system of representatives:
t
+
t t/W.
This is one of the great benets of W being a reection group. (We will see an example
later of how much more annoying moduli spaces are when quotienting by a group that
isnt one.)
Because the simple roots are linearly independent, this cone is always an orthant times
a vector space (whose dimension is that of Z(K)).
Recall that given X t not perpendicular to any , i.e. in the interior of some Weyl
chamber, we can dene
+
:= { : X, > 0. Then dening t
+
as above, we obtain
the chamber that contains X.
We can now state the big theorem in the subject of representations of compact con-
nected groups, which uses the corresponding chamber in t

.
Theorem 9.1 (of the highest weight). Fix X t, dening a positive Weyl chamber t

+
of a
connected compact Lie group K. Then the map
[V] arg max
T

X, : is a weight of V

taking an isomorphism class [V] of K-irreps to its highest weight is a bijection {[irreps]} T

+
.
12 ALLEN KNUTSON
Moreover, that dimHom
T
(C

, V) is 1, i.e. the high weight vectors are unique up to scale.

How to think about the lower-dimensional walls in the Weyl hyperplane arrangement?
If X, = 0, then C
K
(X) L

, and vice versa. So as we go to smaller faces, the centralizer

jumps dimension, and the K-orbit shrinks.
9.1. Conjugacy classes in k

(coadjoint orbits). So far weve shown that every conjugacy

class of K meets T, so using the exponential map, every orbit in k goes through t. (That
only works within the injectivity radius, but we can rescale to work in there.) And then,
we can use W to cut t down further to its positive Weyl chamber.
Using an invariant form, we can regard t

as a subspace of k

, and dene a positive Weyl

chamber there as well.
In both cases, though, we havent discussed the issue of whether the chamber repre-
sents some K-conjugacy class more than once (it will turn out it does not).
Example: if K = U(n), and we naturally identify t

with R
n
, then the usual positive
Weyl chamber is (
1

2
. . .
n
). If we restrict to SU(n), then it gets the additional
condition

i
= 0, and becomes a pointed cone.
9.2. Conjugacy classes in K. So far T/W maps onto the space K/ of conjugacy classes.
To know its an isomorphism, we need
Lemma 9.2. Two elements of T are K-conjugate iff theyre N(T)-conjugate.
Proof. Let t, gtg
1
T be the two elements, so t g
1
Tg. Let H = C
K
(t)
0
. Then H
T, g
1
Tg. Hence some h H has h
1
Th = g
1
Tg, since all tori in the compact connected
Lie group H are conjugate. So hg
1
N(T), and
gtg
1
= gh
1
thg
1
= (hg
1
)
1
t(hg
1
).

The corresponding statement for Sylow subgroups of a nite group is called Burn-
sides fusion theorem, and has very much the same proof.
Corollary 9.3. The subset t

+
k

meets each coadjoint orbit exactly once.

Call an element of K regular if it lies in a unique maximal torus, which includes the
case of topological generators but is more general. Example: a unitary matrix is regular
iff it has distinct eigenvalues (rather than their logs being incommensurable).
This property is obviously invariant under conjugacy, so it sufces to understand which
elements of T are regular. If t T lies in another torus, then dimC
G
(t) > dimT, so t acts
trivially on some part of k/t, i.e. (t) = 1 for some .
This is easiest to analyze up on t, where the condition is , X Z, giving a decompo-
sition of t into Weyl alcoves. Note that nearby 0, this is just the decomposition into Weyl
chambers.
T/W

= (t/)/W

= t/(W)
Theorem 9.4. If K is simple and simply connected, then W is an afne reection group,
whose new generator is reection through the hyperplane ,
0
1 where
0
is the lowest root
in , the one of greatest negative height.
In particular, T/W can be identied with a simplex T
+
inside t.
LIE THEORY AND TOPOLOGY 13
If t is on a face of dimension m of the positive Weyl alcove T
+
, the group C
K
(t) has
central rank m.
So t is on a vertex iff C
K
(t) is semisimple. The rank(G) + 1 many such conjugacy
classes are called special.
Z(K) acts on T
+
by rigid motions, taking the identity vertex to the other central
vertices.
Hence |Z(K)| rank(K) +1. Equality holds exactly for K = SU(n).
To compute the order (resp. adjoint order) of a conjugacy class, scale it until it lies
on a reection of the identity vertex (resp. a central vertex).
W T
+
, considered inside t, is a polytope that tesselates t. The map exp : W T
+
T
is onto, and one-to-one away from the boundary.
If K has nite center, we can still use this technology to analyze Ks conjugacy
classes by K/ = (

K/)/
1
(K). But this may not be a polytope, as in PU(3). (Or it
may be, as in SO(5).)
10. LOOP GROUPS
10.1. Loop spaces. Let M be a Riemannian manifold, and LM = Map(S
1
, M) be the
space of smooth based loops into M. This is an innite-dimensional Fr echet manifold,
with tangent spaces
T

LM

= (S
1
;

TM), LM.
We can dene a metric on the loop space:
v, w :=

S
1
v|
t
, w|
t

where the latter , occurs inside T

M
(t). We also can dene a 1-form
(v) =

S
1
v|
t
,

(t)
and take d of it to get a closed 2-form.
Finally, we can dene an action functional
A() =

S
1
1
2
|

(t)|
2
whose critical points are the geodesic loops.
Theorem 10.1. Where is nondegenerate, the rotate the loop vector eld is the Hamiltonian
vector eld of the action functional.
For a generic metric, the closed geodesics are isolated, and A is a Morse function
indeed, Morse invented Morse theory for this application.
One can do some amazing, if nonrigorous, stuff with this 2-form [At83]. But its in some
sense boring since its d of a 1-form.
If M is a group G, then TM

= G g, so each tangent space is isomorphic to Lg. In
particular, we can talk about the derivative of a tangent vector, and get another tangent
vector!
14 ALLEN KNUTSON
10.2. The based loop group. Fix a compact connected Lie group K. (So the metric will be
very nongeneric.)
The correct space to work on will be not LK, but K = Map(S
1
, K), the space of smooth
based loops into K. Both of these are groups, under pointwise multiplication, something
like a limit of K
n
as the n points become dense in S
1
, and they are related by
LK/K

= K, identifying K

= {constant loops}.
One benet of this identication is to put a circle action on K, which doesnt exist for
general M. But each turns out to be the wrong group!
K has very nice geodesics:
Theorem 10.2. : S
1
K is a basepoint-preserving geodesic iff it is a one-parameter subgroup.
We can conjugate it to lie in T, and then to get its generator to lie in the positive Weyl chamber t
+
.
So the space of critical points of A is a disjoint union of adjoint K-orbits, one for each dominant
coweight in t
+
. When the dominant coweight lies in the interior, the orbit is a K/T.
The index of a stratum is the height of the coweight, and in particular, nite.
This is the example for which Bott invented Morse-Bott theory; A turns out to be a
Morse-Bott function.
Also, there is a natural symplectic 2-form on K:
(v, w) =

S
1
v

, w
which is antisymmetric by integration-by-parts. (It is even the imaginary part of a K ahler
form on K [Pr82].) Then the circle action is Hamiltonian, and generated by the energy
functional.
Theorem 10.3. If G acts transitively and symplectically on a symplectic manifold M, then M is
a cover of a central extension of G, but not necessarily of G itself.
(Even better: if Ms 2-form is the curvature of a Hermitian line bundle /, then some central
extension of G can have its action lifted to /.)
Proof. On the Lie algebra level, g symp(M). There is an exact sequence 0 H
0
(M)
C

(M) symp(M) H
1
(M) 0 of Lie algebras making C

(M) symp(M) a
central extension. Pull it back to get a central extension ^ g of g, and the dual of this gives a
G-equivariant map M^ g

.
The group version is based on Aut(/) Symp(M), where the automorphisms of /
may move the base but must preserve parallel transport.
Let R
2n
act on itself by translation, preserving the standard symplectic form. Then
the above construction discovers the Heisenberg group.
Let Sp(R
2n
) act on R
2n
\

0. This is a double cover of the minimal coadjoint orbit
(R
2n
\

0)/. If one tries to act on the Hermitian line bundle, one discovers the
metaplectic group.
Let LK act on K. Then the above discovers that LK has a central extension

LK, of
which K is a coadjoint orbit.
LIE THEORY AND TOPOLOGY 15
Central extensions of a group are related to elements of H
2
. If the group is compact
connected, then that H
2
is also related to H
2
of its Lie algebra. So we get the weird situation
that even though H
2
(R
2n
) = 0, we have H
2
(r
2n
)

= H
2
(t
2n
)

= H
2
(T
2n
) ,= 0, so R
2n
can
have a central extension. The symplectic Lie algebra is semisimple, so has no central
extensions, but the group is homotopic to U(n) so has a double (or even Z-) cover. Finally,
H
2
(LK)

= H
3
(K)

=
3
(K)

= Z for K simple and simply-connected, which gives a hint as
to why LK should have a canonical central extension.
Morse-Bott theory on this manifold is a little weird, not so much because its innite-
dimensional but because its noncompact. Consider Morse theory on the punctured torus,
using a function that goes to at the puncture. The Morse strata then form a gure 8,
which is only a deformation retract of the punctured torus, rather than equal to it.
Theorem 10.4. K deformation-retracts to the union Gr =

t+
Gr

of the nite-dimensional
Morse-Bott strata. Each Gr

is isomorphic to a complex vector bundle over a complex ag manifold

K/K

, each closure is a projective variety, and the union is an ind-scheme.

We heard already that dominant weights control representation theory. So how can we
pull representations out of these Gr

?
Baby case: K = U(n), K has Z-many components. The minimum A-stratum on the
kth component is isomorphic to Gr
k mod n
(C
n
). The homology of that manifold is

n
k

-
dimensional, which by amazing coincidence is also the dimension of the kth fundamental
representation of U(n)!
Of course, thats also the cohomology of that manifold. But Gr

is singular in general, so
these will differ, and which should we use? In general we wont want to use either, but
the intersection homology, which we give a brief picture of.
The homology of a singular (or any) space is easy to think about geometrically, using
cycles. The cohomology is just as easy if the space is smooth. The best smoothness we
have available here is that each Gr

Gr

is stratied by smooth manifolds Gr

, so
instead of thinking about arbitrary cycles, we think about cycles that behave well with
respect to the stratication.
The cohomology of a (compact, oriented, and nicely) stratied space X is easy to de-
scribe using cycles C that are dimensionally transverse to the strata Y:
dimC (C Y) dimX dim(Y = X Y)
whereas to compute homology we dont need any condition:
dimC (C Y) 0.
For intersection homology in middle perversity, we split the difference:
dimC (C Y)
1
2

dimX dimY

.
Call such C intersection homology chains, and make a complex with them, with the usual
boundary as differential, to dene IH(X). It is naturally a module over H

(X) by a sort of
cap product, but this isnt so useful since H

(X) can be so nasty.

Theorem10.5 (Geometric Satake correspondence). There is a natural action of
L
Gon IH(Gr

),
making the latter into the irrep of
L
G with highest weight . Here
L
G is the Langlands dual
group of G, whose weight lattice is the Z-dual of Gs coweight lattice, and vice versa.
16 ALLEN KNUTSON
The proof of this (rst approximated by Ginzburg, now spread out over papers of
Ginzburg, Lusztig, and Mirkovi c-Vilonen) is rather indirect; they make a category that
looks like the representations of some group, in that it has tensor products (the hard part)
and a forgetful functor to Vec, whose simple objects are the IH(Gr

). Then the Tannaka re-

construction theorem says that this category is the representations of some group. Which
one? With not much work, they relate its weight lattice with Gs coweight lattice, nishing
the identication.
Theorem 10.6. (1) (Mirkovi c-Vilonen 99) If one does Morse theory on the singular space
Gr

, using a component X of the T moment map, the Morse strata are reducible varieties,
and their components give a basis of IH(Gr

).
(2) (Jared Anderson 03) One can use these cycles to compute weight multiplicities and tensor
products of representations. It is even enough to know just their moment polytopes, M-V
polytopes.
(3) (Kamnitzer 05) The polytopes are all distinct, and there is a simple characterization of
them, bypassing all the innite-dimensional geometry.
REFERENCES
[At83] M. Atiyah, Circular symmetry and stationary-phase approximation, Colloquium in honor of Lau-
rent Schwartz, Vol. 1 (Palaiseau, 1983), Ast erisque No. 131 (1985), 4359.
[Pr82] A. N. Pressley, The energy ow on the loop space of a compact Lie group, J. London Math. Soc. (2)
26 (1982), no. 3, 557566.