The Moment Map

J.W.R
March 21, 2007
Throughout M is a manifold, G is a Lie group and LG is the Lie algebra
of G. Denote by Di(M) the group of smooth self dieomorphisms of M,
and by A(M) the Lie algebra of smooth vector elds on M. We denote by
Flow(X) the time one map of the ow of the vector eld X. Note that
Flow : A(M) Di(M)
is formally analogous to the exponential map
exp : LG G
More precisely if
f
t
= Flow(tX)
then
f
0
= id
M
the identity map of M and
d
dt
f
t

t=t
0
= X f
t
0
Given some structure on M we denote by Di(M, ) the subgroup
of Di(M) consisting of those dieomorphisms which preserve . Similarly,
A(M, ) denotes the Lie subalgebra of A(M) consisting of those vector elds
which preserve the structure. For example, if is a dierential form then
Di(M, ) = f Di(M) : f

=
and
A(M, ) = X A(M) : (X) = 0
1
where (X) denotes the Lie derivative of in the direction X:
(X) =
d
dt
Flow(tX)

t=0
A Lie group action of G on M is a group homomorphism
G Di(M) : a a
M
for which the evaluation map is smooth. A Lie group action determines (and
in case G is connected and simply connected is determined by) a Lie algebra
action
LG A(M) : A A
M
via the formula
exp(tA)
M
= Flow(tA
M
)
for A LG and t R. We can calculate A
M
by the formula
A
M
=
d
dt

t=0
exp(tA)
M
.
1 Some formulas
The notations used in the exterior calculus vary slightly from author to au-
thor depending on the choice of multiplicative constants. In the following
three denitions one can take any value of e and the various laws (viz. is
associative, skew-commutative; () and d are skew-derivations, etc.) will
hold. Most authors take e = 1; some take e = 0.
For T
p
(M), T
q
(M), X
0
, X
1
, . . . , X
p
A(M) we dene (with
e = 1)
=

p + q
q

e
ALT( )
(X
1
)X
2
, . . . , X
p
= p
1e
(X
1
, X
2
, . . . , X
p
)
(d)(X
0
, X
1
, . . . , X
p
) = (p + 1)
e1

i
()
i
(D
X
i
)(X
0
, . . . ,

X
i
. . . , X
p
)
In the last formula D is any covariant derivative (the result is independent
of the choice)
2
Wedge product of one-forms
(X, Y ) = (X)(Y ) (Y )(X)
Wedge product of a one-form with a two-form
(X, Y, Z) = (X)(Y, Z) + (Y )(Z, X) + (Z)(X, Y )
Cartans innitesimal homotopy formula
(X) = d(X) + (X)d.
Palaiss Formula
(d)(X
0
, X
1
, . . . , X
p
) =

j
()
j
(X
j
)

(X
0
, . . . ,

X
j
, . . . , X
p

j<k
()
j+k
((X
j
)X
k
, X
0
. . . , X
j
. . . , X
k
. . . , X
p
).
Palaiss formula for zero-forms
d(X) = (X)
Palaiss formula for one-forms
d(X, Y ) = (X)((Y )) (Y )((X)) + ([X, Y ])
Palaiss formula for two-forms
d(X, Y, Z) = (X)((Y, Z)) + (Y )((Z, X)) + (Z)((X, Y ))
(X, [Y, Z]) (Y, [Z, X]) (Z, [X, Y ])
2 Symplectic Mechanics
Let (M, ) be a symplectic manifold: i.e. is a symplectic form (=closed,
non-degenerate two-form) on M. We call the dieomorphisms f and vector
elds X which preserve symplectic. As usual Di(M, ) denotes the
3
group of symplectic dieomorphisms, and A(M, ) denotes the Lie algebra
of symplectic vector elds.
Since d = 0 Cartans formula gives
(X) = d(X)
so X A(M, ) i (X) is closed. Thus any function H on M determines
a vector eld X A(M, ) via the formula
(X) = dH
(X is the symplectic gradient of H) and when H
1
(M) = 0 every vector
eld X A(M, ) has this form. We call X the Hamiltonian vectoreld
and H a Hamiltonian for X. We also write
X = H
M
for the Hamiltonian vector eld with Hamiltonian H. Note that X deter-
mines H only up to a locally constant function (an additive constant when
M is connected.)
The Poisson brackets
T(M) T(M) T(M) : (H, K) H, K
dened by
H, K = (H
M
, K
M
)
give T(M) the structure of a Lie algebra which renders the map H H
M
a
homomorphism of Lie algebras:
H, K
M
= [H
M
, K
M
].
In other words, the Poisson brackets H, K is a Hamiltonian for the Lie
brackets [H
M
, K
M
]. This Lie algebra is called the Poisson algebra and
denoted by T(M, ) so
T(M, ) = T(M)
as vector spaces.
The kernel of the homomorphismH H
M
consists of the locally constant
functions and the homomorphism is onto if the rst De Rham cohomology
4
H
1
(M) of M vanishes. Hence in case M is connected and simply connected
we have an exact sequence
0 R T(M, ) A(M, ) 0
of Lie Algebras.
According to Darbouxs theorem there exist, at any point of M, co-
ordinates q
1
, . . . , q
n
, p
1
, . . . , p
n
such that
=
n

i=1
dq
i
dp
i
.
Such co-ordinates are called symplectic co-ordinates. In symplectic co-
ordinates the Poisson brackets are given by
H, K =
n

i=1

H
p
i
K
q
i

H
q
i
K
p
i

and the trajectories of the vector eld H

M
are the solutions of Hamilto-
nians equations:
q
i
=
H
p
i
, p
i
=
H
q
i
.
3 Darboux
Darbouxs theorem is easily seen to be equivalent to the following
Theorem 3.1 Given two symplectic forms
0
and
1
dened in a neighbor-
hood of 0 in R
2n
there is a dieomorphism f, dened in a possibly smaller
neighborhood of 0 with f(0) = 0 and
f

1
=
0
.
For the proof rst make a linear change of co-ordinates so that
1
and
0
agree at 0. Then dene

t
=
0
+ t(
1

0
)
for 0 t 1. Since that value of
t
at the origin 0 R
2n
is independent
of t (and hence non-degenerate) it follows that all the
t
are symplectic in
5
a neighborhood of 0 R
2n
. We shall nd a time-dependent vector eld
X
t
dened near 0 R
2n
so that the curve of dieomorphism germs t t
determined by solving the dierential equation
(1)
d
dt
f
t
= X
t
f
t
with initial condition
2) f
0
(x) = x
satises
(3) f
t
(0) = 0
and
(4) f

t

t
=
0
.
Assume for the moment that (1)-(4) hold. Dierentiate with respect to t and
evaluate at t; we obtain
(5)
d
dt
f

t

t
= f

(X
t
)
t
+
1

.
Hence we require X
t
to satisfy
(6) (X
t
)
t
+
1

0
= 0.
We will also imose the condition
(7) X
t
(0) = 0.
Suppose we have found X
t
satisfying (6) and (7). By the existence theorem
for ordinbary dierential equations we may nd f
t
satisfying (1) and (2).
Equation (6) implies equation (3) so that the dieomorphisms f
t
x the ori-
gin. (This will assure that the change of variables is valid in a neighborhood
of the origin.) Equations (5) and (6) give that f

t

t
is independent of t.
Hence, taking f = f
1
gives f

1
=
0
as required. Thus it is enough to
construct X
t
solving (6) and (7).
By Cartans formula it suces to choose X
t
to be the unique solution of
(X
t
)
t
=
where d =
1

0
. To achieve X
t
(0) = 0 we require that vanishes at 0.
This is can be achieved by replacing by + d for some function .
6
4 Invariance
The formula
(H
M
)K = K, H
holds on any symplectic manifold (M, ). Heres the proof:
(H
M
)K = (H
M
)dK
= (H
M
)(K
M
)
= (K
M
, H
M
)
= K
M
, H
M
.
Thus if K, H = 0, the function K is constant along the trajectories of the
vector eld H
M
.
5 Volume
A symplectic manifold (M, ) admits a volume
=
n
where the dimension of M is 2n. Obviously, any symplectic dieomorphism
(or vector eld) is volume-preserving (that is, it preserves ). This is called
Liouvilles theorem. (It is easy to see that a Hamiltonian vector eld is
volume-preserving, since its divergence is zero in symplectic co-ordinates.)
When M is compact we can use to choose a canonical Hamiltonian H
for any Hamiltonian vector eld X: we simply impose the condition

M
H = 0
in addition to the condition X = H
M
. When M is connected, this condition
determines a Hamiltonian H for X uniquely.
1
For any functions H and K we have

M
H, K = 0
1
If M is not connected, impose this condition on each component.
7
The proof is by Stokes theorem and
H, K = ((K
M
)H)
= (K
M
)(H)
= d(K
M
)(H).
Hence the map X H determines a Lie algebra splitting of the exact
sequence
0 R T(M, ) A(M, ) 0.
6 Classical mechanics
When M = R
n
R
n
and (q, p) are the position and momentum of a particle
of mass m subjected to a potential V = V (q), we take
H =
1
2m
|p|
2
+ V (q)
to be the energy of the particle and Hamiltons equations reduce to Newtons
equations:
m q = p
p = grad V.
where grad V is the gradient of V .
7 Symplectic Group Actions
A symplectic Lie group action is a Lie group action where each element of
the group G is represented by a symplectic dieomorphism:
G Di(M, ) : a a
M
.
In case G is connected, the action of G is symplectic if and only if corre-
sponding action of its Lie algebra LG is symplectic:
A
M
A(M, )
for A LG.
8
We call the action Hamiltonian i each vector eld A
M
(for A LG)
is Hamiltonian. (This is always the case when M is simply connected.) The
action is Hamiltonian i the Lie algebra homomorphism
LG A(M, ) : A A
M
lifts to a linear map
LG T(M, ) : A
A
.
For each A LG the function
A
is a Hamiltonian for the vector eld A
M
:
() d
A
= (A
M
).
The map
: M LG

given by
(z), A) =
A
(z)
is called a moment map for the action. Equation () determines the mo-
ment map only up to an additive constant which depends on A; i.e. for
LG

A
=
A
+ (A)
is also a Hamiltonian for A
M
.
8 Equivariance
Let : M LG

be a moment for a Hamiltonian group action. Recall the

co-adjoint action
G Aut(LG

) : a ad(a
1
)

.
of the Lie group on the dual of its Lie algebra.
Proposition 8.1 Assume that G is connected. Then the map
: M LG

is equivariant i the map

LG T(M, ) : A
A
is a homomorphism of Lie algebras:

A
,
B
=
[A,B]
for A, B LG.
9
Proof: Since G is connected, equivariance with respect to the Lie group
actions is the same as equivariance with respect to the corresponding Lie
algebra action. In other words, the map is equivariant i
(A
M
) = Ad(A)

for A LG. Evaluate both sides at B LG. The left side becomes
(A
M
)
B
=
A
,
B

(since
A
and
B
are Hamiltonians for A
M
and B
M
) and the right hand side
becomes
[A,B]
.
Now in case M is connected, any two Hamiltonians for the same vector
eld dier by an additive constant so that each moment map determines
a skew-symmetric bilinear form
2
(LG) by

A
,
B
=
[A,B]
+ (A, B).
Now solves a co-cycle identity:
d = 0
and replacing by
= +
(where LG

) replaces by

dened by

= + d
where d :
2
(LG)
3
(LG) is dened by
d(A, B, C) = ([A, B], C) + ([B, C], A) + ([C, A], B)
and d :
1
(LG)
2
(LG) is dened by
d(A, B) = ([A, B]).
If we can solve d = (for example, if the Lie algebra cohomology H
2
(LG)
vanishes which is the case when LG is semi-simple) then we can choose so
that the map A
A
is a homomorphism of Lie algebras; i.e. so that is
equivariant.
Of course, when M is compact it is easy to choose an equivariant moment:
we impose the condition that

A
= 0.
(See section 5.)
10
9 Angular Momentum
Take M = R
n
R
n
as before, and G = SO(n) the special orthogonal group:
SO(n) = a R
nn
: det(a) = 1, a

= a
1
.
Take the action to be given by:
a
M
(q, p) = (aq, ap).
This action is symplectic. To prove this note that = d where
=
n

i=1
p
i
dq
i
= p, dq)
so
a

M
= ap, d(aq))
= ap, a(dq))
= p, dq)
=
so a

M
= a

M
d = da

M
= d = . The Lie algebra of SO(n) is
so(n) = A R
nn
: A

= A
A (equivariant) moment map is given by

A
(q, p) = p, Aq).
When n = 3 there is a bijective correspondence
so(3) R
3
: A
determined by
Av = v
for v R
3
( is the cross product) which identies
A
with the angular
momentum about the axis .
11
10 Heisenberg
The action of R
2
on itself by translations is Hamiltonian. To see this let
= dp dq and note that the innitessimal translations /p and /q are
Hamiltonian with respective Hamiltonians
H(q, p) = q + h, K(q, p) = p + k.
For any choice of the additive constants h and k the map = (H, K) is a
moment for this action, but
H, K = 1
whereas the group is abelian so there can be no equivariant moment.
The Heisenberg group is the set of all real 3 3 matrices of form
a =

1 x z
0 1 y
0 0 1

.
It acts on M = R
2
via the formula
a
M
(p, q) = (p + x, q + y)
and this action is also Hamiltonian. This time however there is a equivariant
moment
(p, q) = (q, p, 1)
where the triple = (, , ) R
3
is identied with a point in the dual of
the Heisenberg algebra via the formula
, A) = x + y + z
for
A =

0 x z
0 0 y
0 0 0

.
12
11 Co-adjoint
A Lie group G acts linearly on its Lie algebra LG via the adjoint action:
ad : G Aut(LG),
ad(a)A = aAa
1
for a G and A LG. Hence G acts linearly on the dual LG

of the Lie
algebra LG via the co-adjoint action:
G Aut(LG

) : a ad(a
1
)

.
Theorem 11.1 Let N LG

be an orbit of the co-adjoint action of G.

Denote by a a
N
the restriction of the co-adjoint action to N:
a
N
= ad(a
1
)

[N
for a G. Then there is a unique G-invariant symplectic form on N such
that the inclusion
: N LG

is an equivariant moment. It is dened by the formula

() (A
N
(), B
N
()) = , [A, B])
for N LG

and A, B LG.
Proof: Before we start the proof, note that the tangent space T

N to N at
N is the vector subspace of LG

given by
T

N = A
N
() : A LG.
This is because G acts transitively on N.
Now assume is G-invariant and that the inclusion is a moment for
the (symplectic) action of G on the orbit N. We prove uniqueness of by
proving the formula (). Evidently for A LG, a Hamiltonian
A
for A
N
is given by

A
() = , A)
13
for N. Hence
(A
N
(), B
N
()) = ((A
N
))(B
N
())
= d
A
()B
N
()
= B
N
(), A)
= Ad(B)

, A)
= , Ad(B)A)
= , [B, A])
= , [A, B])
Now we prove existence. Dene by (). If A
N
() = 0 then the right side
of () vanishes (by part of the last calculation) and similarly if B
N
() = 0.
Hence, is well-dened. The above computation veries the formula
d
A
= (A
N
).
The formula
Ta
N
A
N
a
1
N
= (ad(a)A)
N
(which holds for any action) easily implies the invariance of . To see that
is closed we use Palaiss formula:
d(A
N
, B
N
, C
N
) = (A
N
)(B
N
, C
N
) + (B
N
)(C
N
, A
N
) + (C
N
)(A
N
, B
N
)
+(A
N
, [B
N
, C
N
]) + (B
N
, [C
N
, A
N
]) + (C
N
, [A
N
, B
N
]).
Both the rst three and the last three terms on the right vanish by the Jacobi
identity.
12 Reduction
Let P be a manifold and
P
be a closed two-form on P. (Typically P will be
a submanifold of a symplectic manifold M and
P
will be the restriction to
P of the symplectic form
M
on M.) For each point p P dene a subspace
K
p
= p T
p
P : ( p)
P
= 0
Proposition 12.1 Assume that p K
p
is a sub-bundle of TW, i.e. that
dim(K
p
) is constant. Then the subundle K is integrable.
14
Proof: If X is a vector eld on P then X is a section of K i (X)
P
= 0.
When X is a section of K, we have (X)
P
= d(X)
P
+(X)d
P
= 0 since

P
is closed. Hence, any section of K leaves
P
invariant. Hence, if X and
Y are sections of K, then
([X, Y ])
P
= ((Y )X)
P
= (Y ) ((X)
P
) (X)(Y )
P
= 0
as required.
Now assume that the foliation admits a quotient manifold. This means
that there is a surjective submersion
: P B
whose bers are the leaves of the foliation. Thus
K
p
= T
p

1
((p))
for p P.
Proposition 12.2 There is a unique two-form
B
on B such that

P
=

B
.
Moreover,
B
is symplectic. The symplectic manifold (B,
B
) is called the
reduction of (P,
P
).
Proof: To dene
B
at a point b B choose a point p P with (p) = b
and dene
()
B
(v, w) =
P
( v, w)
for v, w T
b
N where v, w TzW are lifts of v, w:
(T
p
) v = v, (T
p
) w = w.
Since the right hand side of () vanishes if either v = 0 or w = 0 the
denition is independent of the choice of the lifts v, w. Since the vector elds
tangent to the ber leave
P
invariant and act transitively on the ber, the
denition is independent of the choice of p
1
(b). Since

d
B
= d
P
= 0
and is a surjective submersion, it follows that d
B
= 0, i.e.
B
is closed.
Finally,
B
is non-degenerate, for if v T
b
B is such that
B
(v, w) = 0 for
all w T
b
B, then
P
( v, w) = 0 for all w T
p
P, so v K
p
, so v = 0.
15
13 Reduction and Moment
Now assume that : M LG

is an equivariant moment for a sympletic

G-manifold and put
P =
1
()
where LG

is a regular value for . The tangent space to P at a point

p P is given by
T
p
P = p T
p
M : d
A
(p) p = 0 A LG.
This can be rewritten in the form
T
p
P = p T
p
M : (A
M
(p), p) = 0 A LG
i.e.
T
p
P = (T
p
O)

where O is the orbit of G through P and denotes sympletic orthogonal

complement. Thus
K
p
= (T
p
P) (T
p
P)

= (T
p
P) (T
p
O)
= B
M
(p) : d
A
(p)B
M
(p) = 0 A LG
= B
M
(p) : Ad(B)

= 0
= B
M
(p) : B LStab(, G)
where
Stab(, G) = a G : ad(a)

=
is the stabilizer (isotropy) group of = (p). This stabilizer group Stab(, G)
leaves invariant the manifold P =
1
() since the moment is equivariant.
If the stabilizer group is connected its orbits are the leaves of the foliation
tangent to the sub-bundle K so the reduction : P B (if it exists) is the
projection of the Stab(, G)-manifold P =
1
() onto its orbit space.
14 Reduction of the order
Classically, the reduction construction was used to study a Hamiltonian
system which admitted some symmetry. By restricting to a level surface
16
P =
1
() of the moment and passing to the reduction B one obtains a
system on a lower dimensional manifold B. One says the new system is
obtained by reduction of the order.
Heres how it works. Take a symplectic manifold (M, ) with a symplectic
group action G Di(M, ) : a a
M
and equivariant moment : M
LG

. Assume that the Hamiltonian H is invariant by G:

a

M
H = H
for a G. Let P =
1
() be a regular level of the moment ,
P
= [P
the restriction, B = P/Stab(, G) the orbit space of P by the stabilizer sub-
group, : P B the projection onto the orbit space, and
B
the unique
symplectic form on B satisfying

B
=
P
.
If we dierentiate the equation a

M
H = H we obtain that (A
M
)H = 0
for A LG. Hence
(H
M
)
A
= H,
A
= (A
M
)H = 0
for all A so (H
M
) = 0 so the vector eld H
M
is tangent to the level surfaces
of . In particular, H
M
is tangent to P =
1
() so we dene the vector
eld H
P
A(P,
P
) to be the restriction
H
P
= H
M
[P.
On the other hand the condition a

M
H = H for a G shows that H[P is
invariant by the stabilizer subgroup Stab(, G). Hence there is a function
K T(B,
B
) which lifts to H[P:

K = H[P.
This K determnes a Hamiltonian vector eld K
B
on B.
It is easy to see that the projection : P B intertwines the vector
eld H
P
on P and the vector eld K
B
on B:
T H
P
= K
B
.
Heres the proof. Choose b B and p
1
(b). We must show that
TH
P
(p) = K
B
(b). For this it is enough to show that
B
(TH
P
(p),

b) =
17

B
(K
B
(b),

b) for all

b T
b
B. Choose

b T
b
B and p T
p
P with T p =

b.
Then

B
(TH
P
(p),

b) =

B
(H
P
(p), p)
=
P
(H
P
(p), p)
= (H
M
(p), p)
= dH(p) p
= d(

K(b)) p
= (

dK)(b) p
= dK(b)

b
=
B
(K
B
(b),

b)
as required.
15 Projective space
For example, take M = C
n+1
0, G = S
1
acting diagonally
a
M
(z) = exp(i)z
for z C
n+1
, a = exp(i) S
1
, and
= i
n

j=0
dz
j
d z
j
= 2
n

j=0
dy
j
dx
j
where z
j
= x
j
+ iy
j
are the standard co-ordinates on C
n+1
. Also is the
imaginary part of the (at) Hermitean inner product on C
n+1
:
(v, w) = v, iw) w, iv)
for v, w T
z
M = C
n+1
where
v, w) =
n

j=0
v
j
w
j
for v = (v
0
, v
1
, . . . , v
n
), w = (w
0
, w
1
, . . . , w
n
).
18
A moment for the action is given by
(z) = |z|
2
.
The level surface P =
1
(1) is the unit sphere S
2n+1
and the stabalizer
subgroup Stab(1, G) is the whole group S
1
(because the group is abelian)
so the orbit space N = CP
n
is complex projective space and the projection
: S
2n+1
CP
n
is the Hopf map.
A skew-Hermitean matrix A u(n+1) determines a real valued function
H : C
n+1
R by
H(z) = Az, iz).
The corresponding Hamiltonian vector eld on M = C
n+1
0 is given by
H
M
(z) = Az.
Since H is invariant by the circle action, we obtain a symplectic vector eld
A
N
on N = CP
n
. Thus the usual action of the unitary group U(n + 1) on
CP
n
is symplectic.
16 Linear actions
If G GL(V ) is a linear representation of a compact group G on a real
vector space V there is a direct sum decomposition
V = V
0
W
1
W
q
where V
0
is the xed point set of the linear representation and W
1
, . . . , W
q
are the irreducible components of the representation. By averaging we may
nd an invariant inner product , ) on V and take the decomposition to be
orthogonal.
Now suppose is a non-degenerate skew symmetric form on V invariant
by G. Then there is a skew-symmetric automorphism of V with
(v, w) = v, w)
for v, w V . Write = PJ where P is positive denite and J is orthogonal.
Since

= , and

commute, hence P =

and J = P
1

commute. By replacing , ) by (v, w) Pv, w) we may assume = J.

19
Now J
1
= J

= J so J
2
= I: J is a complex structure on V . The
Hermitean inner product
(v, w) v, w) + i(v, w)
is invariant by G. We may decompose V into irreducible complex represen-
tations (this might be dierent from the real decomposition)
2
and obtain a
J-invariant orthogonal decomposition
V = V
0
V
1
V
p
.
Now the summands are orthogonal with respect to as well as with repect
to the inner product. Any innitessimally unitary endomorphism A of V is
Hamiltonian:
A = J grad Q
where
Q(v) =
1
2
JAv, v).
Now specialize to the case where G is the torus T = T
n
= R
n
/Z
n
. In this
case the irreducible components V
1
, . . . , V
p
are all two dimensional (over R)
and the decomposition over R agrees with the decomposition over C. Any A
in the image of LT has form A = J grad Q where
Q = a
1
(x
2
1
+ y
2
1
) + a
p
(x
2
p
+ y
2
p
)
where (x
i
, y
i
) are suitable symplectic linear co-ordinates on V
i
for i = 1, . . . , p.
Note in particular that
(1) The index of the quadratic form Q is even.
(2) The restriction of the skew-symmetric form to the space V
i
is non-
degenerate. In particular, [V
0
is non-degenerate.
2
An example where the decomposition into real irreducibles is dierent from the de-
composition into complex irreducibles is given by O(n) acting on C
n
. (The skew form
is the imaginary part of the standard Hermitean inner product on C
n
.)
20
17 Symplectic linearization
The action of a compact group G on a manifold M may be linearized near
a xed point z M: use an equivariant exponential map T
z
M M, say
the one arising from the geodesic spray of an invariant Riemannian metric.
The equivariant exponential map intertwines the linear action of G on T
z
M
with the action of G on M. Since the xed point set of the linearized action
of G on T
z
M is a vector subspace of T
z
M, this proves that the xed point
set of the action of G on M is a submanifold of M.
Now suppose that the action is symplectic. The equivariant exponential
map need not be symplectic, but we shall prove that there is a symplectic
linearization of the action. What is required is an equivariant Darbouxs
theorem: if
0
and
1
are G-invariant symplectic forms which agree at a
xed point z of the G-action, then there is an equivariant dieomorphism f,
dened in an invariant neighborhood of z, satisfying f(z) = z and f

1
=
0
.
To prove this we simply imitate the proof of Darbouxs theorem given in
section 3 above. First write
1

0
= d and choose to be invariant by
averaging over the group. Then the vector eld X
t
dened by (X
t
)
t
=
is equivariant with respect to G since
t
and are. Hence the solution
f
t
commutes with the G-action. In particular, the dieomorphism f = f
1
commutes with the G-action. As it solves f

1
=
0
it provides the required
linearization.
18 Morse inequalities
In their most general form, the Morse inequalities state that
dimH
k
(M)
r

i=1
dimH
k
(M
i
, M
i1
)
where the M
i
form a ltration of the space M:
= M
0
M
1
M
2
M
r
= M.
These inequalities are usually applied in the following situation:
M
i
= x M : f(x) a
i

21
where f : M R is a smooth function and the a
i
are regular values of f
and there is exactly one critical value c
i
with
a
i1
< c
i
< a
i
.
We denote by = (f) the set of critical points of f:
= x M : df(x) = 0
and by
i
those at level c
i
:

i
= x : f(x) = c
i

so that
=
1

2

r
and

i
int (M
i
M
i1
) .
The critical elements of f we mean the connected components of the
set (f) of critical points of f. To ease the exposition we shall assume that
each
i
connected, so that the critical elements are
1
,
2
, . . . ,
r
. This
assumption can easily be dropped.
Let x be a critical point f. The Hessian D
2
f(x) is the quadratic form on
the tangent space dened by
D
2
f(x) x
2
=
d
2
dt
2

t=0
f((t))
where x T
x
M and is any curve in M with (0) = x and (0) = x. (Since
df(x) = 0 the denition is independent of the choice of .) The dimension of
a maximal subspace of T
x
M on which the Hessian D
2
f(x) is negative-denite
is called the index of the critical point x. Thus at any critical point x we
have a (non-unique) direct sum decomposition
T
x
M = N
0
x
N
s
x
N
u
x
orthogonal with respect to the Hessian D
2
f(x) such that D
2
f(x) vanishes on
N
0
x
, is positive-denite on N
s
x
, and negative-denite on N
u
x
. The dimension
u of N
u
x
is the index of the critical point x.
The function f is called non-degenerate in the sense of Morse i each
critical point x is non-degenerate meaning that the Hessian D
2
f(x) is a
22
non-degenerate quadratic form on the tangent space T
x
M. It follows in this
case that the critical points are isolated so that is a discrete set (nite if M
is compact). In this case the critical elements are points. By our assumption
that each
i
is connected it follows that
i
is a single point and M
i
/M
i1
has the homotopy type of S
u
where u is the index. Hence
dimH
k
(M
i
, M
i1
) =

1 if k = u,
0 otherwise.
We obtain
Proposition 18.1 (Morse inequalities) For a function f which is non-
degenerate in the sense of Morse we have that
dimH
k
(M) n
k
(f)
where n
k
(f) is the number of critical points of f with index k.
The function f is called non-degenerate in the sense of Bott i
the critical set is a submanifold of M and at each x the restriction
of the Hessian D
2
f(x) to the normal space T
x
M

is non-degenerate. (This
condition is independent of the choice of Riemannian metric used to compute
the normal space.) In this case the index of a critical point x is constant for
x
i
. In this case for each x the Hessian vanishes on T
x
. In fact
the restriction of TM to each critical element
i
of admits a (non-unique)
vector bundle splitting
TM[
i
= T
i
N
s
i
N
u
i
orthogonal with respect to the Hessian D
2
f such that the Hessian vanishes
on T
x

i
, is positive denite on N
s
i
and negative denite on N
u
i
. The ber
dimension u of N
u
i
is the the common index of the critical points x
i
.
The space M
i
/M
i1
has the homotopy type of the Thom space D/D where
D is the unit disk bundle of N
u
i
. Hence, by the Thom isomorphism, we have
that
dimH
k
(M
i
, M
i1
) = dimH
ku
(
i
).
In particular, dimH
k
(M
i
, M
i1
) = 0 for k < u and dimH
u
(M
i
, M
i1
) = 1
(by our assumption that
i
is connected). We obtain
23
Proposition 18.2 (Morse-Bott inequalities) For a function f which is
non-degenerate in the sense of Bott we have that
dimH
k
(M) n
k
(f)
where n
k
(f) is the number of critical elements of f with index k.
19 Conected levels
Lemma 19.1 (Atiyah) Assume that M is compact and connected, that the
function f : M R is non-degenerate in the sense of Bott, and that neither
f nor f has a critical element of index 1. The each level f
1
(c) for c R
is connected (or empty).
Proof: By continuity we may assume w.l.o.g that c is a regular value. Let
M
+
c
= f
1
([c, )), M

c
= f
1
((, c]),
so
f
1
(c) = M
+
c
M

c
.
By Mayer-Vietoris we have an exact sequence
H
1
(M) H
0
(f
1
(c)) H
0
(M
+
c
) H
0
(M

c
) H
0
(M).
By the Morse-Bott inequalities we have
H
1
(M) = 0
so it is enough to prove that M

c
is connected. (To see that M
+
c
is also
connected replace f with f.)
Now M

c
is dieomorphic to M
k
for some k = 1, 2, . . . , r. Consider the
exact sequence
H
1
(M
i
, M
i1
) H
0
(M
i1
) H
0
(M
i
) H
0
(M
i
, M
i1
).
Now H
1
(M
i
, M
i1
) ,= 0 only if the index of the critical element
i
is 1
or 0. The former case is excluded by hypothesis and in the latter case
i
is a local minimum. In any case (no matter what the index is) the map
H
1
(M
i
, M
i1
) H
0
(M
i1
) is zero so dimH
0
(M
i1
) dimH
0
(M
i
). We
have equality for each i since by hypothesis dimH
0
(M
r
) = dimH
0
(M) = 1.
Thus dimH
0
(M
i
) = 1 for each i = 1, 2, . . . , r so each M
i
is connected as
required.
24
20 Symplectic toral actions
Let (M, ) be a compact symplectic manifold and T
n
Di(M, ) be a
Hamiltonian action of the torus T
n
= R
n
/Z
n
. Using an invariant inner
product we identify
LT
n
= LT
n
= R
n
.
Let
= (f
1
, f
2
, . . . , f
n
) : M R
n
.
be an equivariant moment map for the action and
f =
A
be a linear combination of the components of Thus the corresponding
Hamiltonian vector eld f
M
= A
M
is an innitessimal generator of the toral
action. Its trajectories have tori as closures: sub-tori of the orbits of the
action. (Atiyah calls f
M
almost periodic.)
Lemma 20.1 Under these hypotheses the function f : M R is non-
degenerate in the sense of Bott and has only crtical elements of even index.
(Hence, every level set f
1
(c) is connected.)
Proof: The critical points of f are the zeros of the corresponding vector eld
A
M
. With out loss of generality (replace T by a sub-torus) we may assume
that the one-parameter group subgroup of T generated by A is dense in T.
Now critical points of f =
A
are the xed points of the action. At any xed
point z of the action we may choose symplectic co-ordinates which linearize
the action. In such co-ordinates f is a quadratic form
f = a
1
(x
2
1
+ y
2
1
) + a
p
(x
2
p
+ y
2
p
)
as in section 16 above which shows the critical point has even index.
21 Symplectic and convex
Theorem 21.1 (Atiyah) Let M be compact and symplectic and
: M LT

= R
n
be an equivariant moment for a symplectic action of the n dimensional torus
T = T
n
on M. Let Z
1
, . . . , Z
r
be the connected components of the xed point
set of the action. Then
25
(1) Each (non-empty) ber
1
() ( LT

) is connected.
(2) The moment is constant on each Z
i
(i = 1, . . . , r).
(3) The image (M) is a convex hull of the points f(Z
1
), . . . , f(Z
r
).
Proof: We prove part (1) by induction on n. The case n = 1 is lemma 20.1.
Now assume the theorem for n 1 and write
(x) = ((x), f(x))
where : M R
n1
and f : M R. Also put
= (, c)
where R
n1
and c R. It is enough to prove (1) for a dense set of so
by Sards theorem we may assume that is a regular value of . Form the
manifold
N =
1
().
By the induction hypothesis N is connected. Since the torus is abelian, the
stabalizer is everything: Stab(, T
n1
) = T
n1
.
As a warm-up we make the (unwarranted) assumption that the quotient
Q = N/T
n1
is a manifold. Well see how to eliminate this assumption at
the end of this section. Form the reduction : N Q where Q = N/T
n1
.
The moment is invariant by the action of T
n
(since the torus is abelian)
so the function f is invariant by the action of T
n1
so there is a function
g : Q R with f[N = g . Evidentally
1
() =
1
(g
1
(c)). The ber of
is T
n1
which is connected and the level set g
1
(c) N is connected by
lemma 20.1. Hence
1
() is connected as required.
For part (2) note that the xed point set of the action is precisely the set
of points z M where d
A
(z) = 0 for every A LT: in other words it is
the set of points z where d(z) = 0. As this set is a manifold it is clear that
is constant on each of its components.
For part (3) we rst show that the image (M) is convex, i.e. that it
intersects every line L in R
n
= LT

in a segment. Choose a linear projection

: R
n
R
n1
and a point R
n1
so that L =
1
(). Apply part (1) to
. Then ( )
1
() is connected so its image

( )
1
()

= (M)
1
()
26
is also connected as required.
For almost every A R
n
= LT the one-parameter group t exp(tA) is
dense in T and so the trajectories of the corresponding vector eld A
M
are
dense in the orbits of T. This means that the zeros of A
M
are precisly the
xed points of the T-action. But the zeros of A
M
are the critical points of its
Hamiltonian
A
. Hence
A
achieves its maximum on one of the components
Z
i
of the xed point set. This means that the image (M) lies to one side
of the hyperplane of all LT

such that
, A) = (Z
i
), A).
Since this holds for almost every A it follows that the image (M) is a subset
of the convex hull of the points (Z
i
) R
n
= LT

. But we have already

seen that the image is convex so it must equal this convex hull.
Heres how to get rid of the unwaranted assumption that Q is a manifold.
We work directly with N =
1
(). Since

1
() = (f[N)
1
(c)
it is enough to show that f[N satises the hypothesis of Lemma 19.1, namely
that f[N is non-degenerate in the sense of Bott and the critical elements of
f[N have even index. For this we choose a critical point x N of f[N.
By the method of Lagrange multipliers x is a critical point of the function
= f + : M R
for some choice of the linear functional R
(n1)
. Now is a linear
combination of the components of the moment map so by Lemma 20.1
its critical set (that is, the xed point set of its Hamiltonian ow) is a
submanifold of M and D
2
f(x) : T
x
M T
x
M R has even index. We will
show
22 Eigenvalue inequalities
Theorem 22.1 (Schur-Horn) A vector x = (x
1
, x
2
, . . . , x
n
) R
n
is the
diagonal of a Hermitean matrix whose eigenvalues are the components of the
vector = (
1
,
2
, . . . ,
n
) if and only if x is a convex combination of the
vectors

for S
n
. Here S
n
is the permutation group on 1, 2, . . . n and

= (
(1)
,
(2)
, . . . ,
(n)
).
27
The theorem may be reformulated as follows. Let D
n
denote the set of
all real diagonal matrices. Let D
n
and let
O() = aa
1
: a U(n)
denote the orbit of under unitary similarity. For any square n n matrix
A, let (A) denote the n n diagonal matrix having the same diagonal as
A. Then the Schur-Horn theorem takes the form
(O()) = convex hull of O() D
n
.
Let G = U(n) be the unitary group and LG = u(n) be its Lie algebra.
Thus u(n) is the set of all matrices iA where A is Hermitean. Notice that
LT = iD
n
is the Lie algebra of the maximal torus T = exp(iD
n
) in U(n)
consisting of all diagonal unitary matrices. The Lie algebra u(n) admits an
invariant inner product
A, B) = real part of tr(AB

)
and with respect to this inner product the map
: u(n) iD
n
which assigns to each matrix A u(n) the diagonal matrix having the same
diagonal entries as A is orthogonal projection onto iD
n
. Thus the Schur-Horn
theorem is a special case of the following
Theorem 22.2 (Kostant) Let G be a compact Lie group, LG be its Lie
algebra, T be a maximal torus in G, and LT be its Lie Algebra. Let LG be
endowed with a G-invariant inner product. Then the orthogonal projection
of any orbit of G in LG onto LT is the convex hull of the intersection of that
orbit with LT.
Remark 22.3 The orthogonal projection LG LT is independent of the
choice of the invariant inner product. To see this write
G = G
1
G
2
G
k
T
0
where each G
i
is compact with simple Lie algebra and T
0
is a torus. Then
T = T
1
T
2
T
k
T
0
28
where T
i
is a maximal torus in G
i
for i = 1, 2, . . . , k. Now consider the
corresponding decompostion of the Lie algebra:
LG = LG
1
LG
2
LG
k
LT
0
.
Each G
i
acts irreducibly on LG
i
(else LG
i
would not be simple) so by Schurs
lemma this direct sum decomposition must be orthogonal with respect to any
invariant inner product , ) on LG. Thus the invariant inner product is given
by
A, B) = A
1
, B
1
)
1
+A
2
, B
2
)
2
+ A
k
, B
k
)
k
+A
0
, B
0
)
0
where , )
i
is a G
i
-invariant inner product on LG
i
for i = 1, 2, . . . , k and
, )
0
is a T
0
-invariant inner product on LT
0
. But the only invariant inner
product on a simple Lie algebra is a multiple of the Killing form, so the
orthogonal complement is given by
LT

= LT

1
LT

2
LT

k
where LT

i
is the sum of the root spaces of LG
i
. (This argument also shows
that to prove Kostants theorem it is enough to prove it when LG is simple.)
Proof of Kostants theorem: Using the G-invariant inner product we identify
LG and LG

. The orbit N is a symplectic G-manifold and the inclusion

N LG is an equivariant moment. Restrict the action to T: then the
inclusion followed by the orthogonal projection onto LT is an equivariant
moment. The xed points of the action are the points N at which
, [A, B]) = 0
for all A LT and all B LG. In other words is a xed point for the
T-action exactly when [A, ] = 0 for all A LT and this the case exactly
when LT.
23 References
[AM] R. Abraham, J. Marsden: Foundations of Mechanics, (Second edi-
tion), Benjamin-Cummings, 1978.
[A] M.F. Atiyah: Convexity and commuting Hamiltonians, Bull. London
Math. Soc. 14 (1982) 1-15.
29
[B] R. Bott: Non-degenerate critical manifolds, Ann. of Math. 60 (1954)
248-261.
[H] A. Horn: Doubly stochastic matrices and the diagonal of a rotation
matrix, Amer. J. Math., 76 (1954), 620-630.
[Ki] F.C. Kirwan: Cohomology of quotients in symplectic and algebraic ge-
ometry, Princeton Mathematical notes, 31 (1984).
[Ko] B. Kostant: On convexity, the Weyl group, and the Iwasawa decom-
positionm, Ann. Sci. Ec. Norm. Sup,, 6 (1973), 413-455.
[M] J. Milnor: Morse theory, Annals of Math Studies 51, Princeton Univer-
sity Press (1963).
[MW] J. Marsden, A. Weinstein: Reduction of symplectic manifolds with
symmetry, Rep. Math. Phys., 5 (1974) 121-130.
[S] I. Schur: Uber eine klasse von mittelbildungen mit anwendungen auf der
determinantentheorie, Sitzungsberichte der Berliner Mathematischen
Gesellschaft 22 (1923) 9-20.
30