Chapter 1

Linear groups
We begin, as we shall end, with the classical groupsthose familiar groups
of matrices encountered in every branch of mathematics. At the outset, they
serve as a library of linear groups, with which to illustrate our theory. Later
we shall find that these same groups also serve as the building-blocks for the
theory.

Definition 1.1 A linear group is a closed subgroup of GL(n, R).

Remarks:
1. We could equally well say that: A linear group is a closed subgroup of
GL(n, C). For as we shall see shortly, GL(n, C) has an isomorphic image as a closed subgroup of GL(2n, R); while conversely it is clear that
GL(n, R) can be regarded as a closed subgroup of GL(n, C).
2. By GL(n, k) (where k = R or C) we mean the group of invertible n n
matrices, ie
GL(n, k) = {T M(n, k) : det T 6= 0}
where M(n, k) denotes the space of all n n real or complex matrices
(according as k = R or C).
3. We define a norm kXk on M(n, k) by
2

kXk =

|Xij | =

i,j

tr(X 0 X)
tr(X X)

if k = R,
if k = C,

where as usual X 0 denotes the transpose and X the conjugate transpose.

This is just the usual Euclidean norm, if we identify M(n, k) with k N , where
424III

11

424III

12

N = n2 , by taking as the coordinates of the matrix X its n2 entries Xij . The

requirement that a linear group should be closed in GL(n, k) refers to this
metric topology. In other words, if T (i) is a sequence of matrices in G
tending towards the matrix T GL(n, k), ie
kT (i) T k 0,
then T must in fact lie in G.
Examples:
1. The general linear group GL(n, R)
2. The special linear group
SL(n, R) = {T GL(n, R) : det(T ) = 1}
3. The orthogonal group
O(n) = {T GL(n, R) : T 0 T = I}
In other words
O(n) = {T : Q(T v) = Q(v) v Rn },
where
Q(v) = x21 + . . . + x2n

(v = (x1 , . . . , xn ) Rn ),

ie O(n) is the subgroup of GL(n, R) leaving the quadratic form Q invariant.

4. The special orthogonal group
SO(n) = O(n) SL(n)
5. The complex general linear group GL(n, C). This calls for some explanation, since GL(n, C) is not a group of real matrices, as required by Definition 1. However, we can represent each complex matrix Z M(n, C) by a
real matrix RZ M(2n, R) in the following way. (Compare the realification of a representation discussed in Part I Chapter 11.)
If we forget scalar multiplication by non-reals, the complex vector space
V = Cn becomes a real vector space RV of twice the dimension, with basis
(1, 0, . . . , ), (i, 0, . . . , 0), (0, 1, . . . , 0), (0, i, . . . , 0), . . . , (0, 0, . . . , 1), (0, 0, . . . , i).

424III

13

Moreover each matrix Z M(n, C), ie each linear map

Z:V V
defines a linear map
RZ : RV RV,
ie a matrix RZ M(2n, R).
Concretely, in passing from Z to RZ each entry
Zj,k = Xj,k + iYj,k
is replaced by the 2 2 matrix
Xj,k Yj,k
Yj,k Xj,k

The map
Z 7 RZ : M(n, C) M(2n, R)
is injective; and it preserves the algebraic structure, ie
R(Z + W ) = RZ + RW
R(ZW ) = (RZ)(RW )
R(aZ) = a(RZ) a R
RI = I
R(Z ) = (RZ)0 .
It follows in particular that RZ is invertible if and only if Z is; so R restricts
to a map
Z 7 RZ : GL(n, C) GL(2n, R).
Whenever we speak of GL(n, C), or more generally of any group G of
complex matrices, as a linear group, it is understood that we refer to the
image RG of G under this injection R.
The matrix X GL(2n, R) belongs to GL(n, C) if is built out of 2 2
matrices of the form
!
x y
y x

424III

14

This can be expressed more neatly as follows. Let

0 1

1 0

iI 7 J =

0 1
1 0
...

Since any scalar multiple of the identity commutes with all matrices,
X M(n, C) = (iI)X = X(iI).
Applying the operator R,
X RM(n, C) = JX = XJ.
Converseley, if JX = XJ then it is readily verified that X is of the required
form. Thus
RM(n, C) = {X M(2n, R) : JX = XJ};
and in particular
GL(n, C) = {T GL(2n, R) : JX = XJ}
6. The complex special linear group
SL(n, C) = {T GL(n, C) : det T = 1}
Note that the determinant here must be computed in M(n, C), not in M(2n, R).
Thus
 
T = i
/ SL(1, C),
although
RT =

0 1
1 0

SL(2, R).

7. The unitary group

U(n) = {T GL(n, C) : T T = I}
where T denotes the complex transpose of T . In other words
U(n) = {T : H(T v) = H(v) v Cn },

424III

15

where H is the Hermitian form

H(v) = |x1 |2 + . . . + |xn |2

(v = (x1 , . . . , xn ) Cn ).

Since R(T ) = (RT )0 , a complex matrix is unitary if and only if its real
counterpart is orthogonal:
U(n) = GL(n, C) O(2n)
8. The special unitary group
SU(n) = U(n) SL(n, C)
9. The quaternionic general linear group GL(n, H). The quaternions
q = t + xi + yj + zk

(t, x, y, z R),

with multiplication defined by

i2 = j 2 = k 2 = 1, jk = kj = i, ki = ik = j, ij = ji = k
form a division algebra or skew field. For the product of any quaternion
q = t + xi + yj + zk
with its conjugate
q = t xi yj zk
gives its norm-square
kqk2 = qq = t2 + x2 + y 2 + z 2 ,
so that if q 6= 0,
q 1 =

q
.
kqk2

We can equally well regard H as a 2-dimensional algebra over C, with each

quaternion taking the form
q = z + wj

(z, w C),

and multiplication in H being defined by the rules

jz = zj, j 2 = 1

424III

16

Surprisingly perhaps, the entire apparatus of linear algebra extends almost

unchanged to the case of a non-commutative scalar field. Thus we may
speak of an n-dimensional vector space W over H, of a linear map t : W
W , etc.
A quaternionic vector space W defines a complex vector space CW by
forgetting scalar multiplication by non-complex quaternions (ie those involving j or k), in just the same way as a complex vector space V defines
a real vector space RV . If W has quaternionic dimension n, with basis
{e1 , e2 , . . . , en }, then CW has complex dimension 2n, with basis
{e1 , je1 , e2 , je2 , . . . , en , jen }.
Moreover each matrix Q M(n, H), ie each linear map
Q:W W
defines a linear map
CQ : CW CW,
ie a matrix CQ M(2n, C).
Concretely, in passing from W to CW each entry
Qr,s = Zr,s + iWr,s
is replaced by the 2 2 complex matrix
Zr,s Wr,s
Wr,s Zr,s

The map
Q 7 CQ : M(n, H) M(2n, C)
is injective; and it preserves the algebraic structure, ie
C(Q + Q0 ) = CQ + CQ0
C(QQ0 ) = (CQ)(CQ0 )
C(aQ) = a(CQ) a C
CI = I
C(Q ) = (CQ) .

424III

17

In this last relation, Q denotes the quaternionic matrix with entries

(Q )rs = Qsr .
To identify M(n, H) as a subspace of M(2n, C), consider the automorphism of H
q 7 q = jqj 1 .
In terms of its 2 complex components,
q = z + wj 7 q = z + wj.

(Note that q 6= q; indeed, the map q 7 q is not an automorphism of H, but

an anti-automorphism.)
Let J denote the diagonal matrix

j 0

0
j
J =

..

M(n, H);

and consider the map

= JQJ 1 : M(n, H) M(n, H).
Q 7 Q
We see from above that
= CQ,
C(Q)
denotes (for X M(n, C)) the matrix with entries
where X
r,s = Xr,s .
X
Now

0 1

1 0

J 7 CJ =

0 1
1 0
..

which we take the liberty of also denoting by J (as we did earlier, when
defining the embedding of GL(n, C) in GL(2n, R), although there we regarded J as a real matrix rather than a complex one).
Thus

M(n, H) {X M(2n, C) : JXJ 1 = X}

424III

18

then X is constructed
Conversely, it is readily verified that if JXJ 1 = X,
from 22 matrices of the form specified above, and so arises from a quaternionic matrix. Hence

M(n, H) = {X M(2n, C) : JXJ 1 = X}

It follows from the properties of the map C listed above that if Q is invertible
then so is CQ; so C restricts to a map
T 7 CT : GL(n, H) GL(2n, C).
In fact, our argument above gives the concrete embedding
GL(n, H) = {T GL(2n, C) : JT J 1 = T}
10. The symplectic group
Sp(n) = {T GL(n, H) : T T = I}
Since (CT ) = C(T ) it follows that
T Sp(n) = (CT ) (CT ) = I,
and so
Sp(n) = GL(n, H) U(2n)
Thus
Sp(n) = {T GL(2n, C) : JT J 1 = T & T T = I}.
Since T1 = T 0 from the second relation, the first relation can be re-written
T 0 JT = J.
This gives an alternative description of the symplectic group:
Sp(n) = {U U(2n) : U 0 JU = J}
In other words, Sp(n) consists of those unitary matrices leaving invariant
the skew-symmetric form
J(x, y) = x1 y2 x2 y1 + x3 y4 x4 y3 + . . . + x2n1 y2n x2n y2n1 .

424III

11. The Lorentz group

O(1, 3) = {T GL(4, R) : G(T v) = G(v) v R4 }
where G is the space-time metric
G(v) = t2 x2 y 2 z 2 .
In matrix terms
O(1, 3) = {T : T 0 GT = G},
where

G=

1 0
0
0
0 1 0
0
0 0 1 0
0 0
0 1

19

424III

110

Exercises
1. Prove that every linear group is locally compact. (Hint: Show that every
open subset, and every closed subset, of a locally compact space is locally
compact.)
2. Show that

SO(2)
= U(1),

Sp(1)
= SU(2).

3. Show that if G and H are linear groups then G H is linearisable, ie there

is a linear group isomorphic to G H.
4. Show that there is a discrete linear group isomorphic to any finite group G.
5. Show that there is a discrete linear group isomorphic to Zn .
6. Show that R is linearisable.
7. Show that R is linearisable. Show also that
R
= R C2
(where C2 denotes the cyclic group of order 2).
8. Show that C is linearisable.
9. Show that C is linearisable. Show also that
C
6= C G
for any topological group G.

Chapter 2
The Exponential Map
Napier introduced logarithms to convert difficult multiplication into easy
addition. Our motivation is much the same, though we are dealing with
matrices rather than numbers. As in the numerical case, it is simpler to
start with the exponential functiondefined by an everywhere convergent
matrix power-seriesand derive the logarithmic function as the inverse in a
suitable restricted zone.

Proposition 2.1 For each matrix X M(n, k) (where k = R or C) the exponential sequence
X2 X3
+
+
I +X +
2!
3!
converges.
Proof I In Chapter 1 we defined the norm kXk on M(n, k) by
kXk2 =

kXij k2 ,

i,j

In other words,
2

kXk =
Lemma 2.1

tr(X 0 X) if k = R
tr(X X) if k = C

1. kX + Y k kXk + kY k

2. kXY k kXkkY k
3. kaXk = |a|kXk(a k)
Proof of Lemma B We suppose that k = C; the real case is identical, with X 0 in
place of X .
424III

21

424III

22

1. We know that
tr Z Z = kZk2 0
for all Z M(n, k). Setting Z = X + Y (where R),
tr X X + (tr X Y + tr Y X) + 2 tr Y Y 0
for all R. Hence
| tr X Y + tr Y X|2 4kXk2 kY k2 ,
and so
| tr X Y + tr Y X| 2kXkkY k.
We note for future reference that if X and Y are hermitian, ie X =
X, Y = Y , then tr X Y = tr XY = tr Y X = tr Y X; and so
X, Y hermitian = tr XY kXkkY k.
But now (taking = 1),
kX + Y k2 = kXk2 + tr(X Y + Y X) + kY k2
kXk2 + 2kXkkY k + kY k2 ;
whence
kX + Y k kXk + kY k.
2. We have
kXY k2 =
=
=
=

tr(XY ) XY
tr Y X XY
tr X XY Y
tr P Q,

P = X X,

Q = Y Y .

where
These 2 matrices are hermitian and positive-definite; and
kXk2 = tr P,

kY k2 = tr Q.

Thus it is sufficient to show that for any 2 such matrices

tr P Q tr P tr Q.

424III

23

But as we noted in the first part of the proof,

tr P Q kP kkQk.
It is sufficient therefore to show that
kP k tr P.
for any positive hermition matrix X. Since
kP k2 = tr P P = tr(P 2 ),
this is equivalent to proving that
tr(P 2 ) (tr P )2 .
But if the eigenvalues of P are 1 , . . . , n , then those of P 2 are 21 , . . . , 2n ,
and
tr(P 2 ) = 21 + + 2n (1 + + n )2 = (tr P )2 .
since i 0 for all i.
For a tensorial proof of this result, let S denote the skew-symmetrizer
Sjlik =
under which


1 i k
j l li jk .
2

1
S(u v) = (u v v u);
2

and let
Z = S(X Y ).
Then
Z = (X Y )S,
and so, since S 2 = S,
tr Z Z = tr(X Y )S(X Y )
= tr S(X Y )(X Y )
= tr S(X X Y Y )
1
(tr X X tr Y Y tr X XY Y )
=
2
1
(tr X X tr Y Y tr(XY ) XY )
=
2

1
kXk2 kY k2 kXY k2
=
2

Since tr Z Z 0, we conclude that

kXY k kXkkY k.

424III

24

3. We have
kaXk2 = tr X a
aX
= a
a tr X X
= |a|2 kXk2 .
C

To show that the exponential series converges for any matrix X, we compare
its partial sums with those in the scalar case. By the lemma above,
kX i /i! + . . . + X j /j!k xi /i! + + xj /j!,
where x = kXk.
It follows that
kX i /i! + . . . + X j /j!k 0
as i, j 7 . Since every Cauchy sequence converges in RN , this proves the
proposition. J
Definition 2.1 For each matrix X M(n, k) (where k = R or C) we set
eX = I + X +

X2 X3
+
+
2!
3!

Examples:
1. e0 = I
2. If X is diagonal, say

1
..

X=

.
n

then eX is also diagonal:

eX =

e1
..

.
en

Proposition 2.2 If X, Y M(n, k) commute, ie

XY = Y X,
then
eX eY = eX+Y = eY eX .

424III

25

Proof I Since XY = Y X, (X + Y )m can be expanded by the binomial theorem:

(X + Y )m = X m + mX m1 Y + . . . + Y m .
The result follows on summation as in the ordinary (scalar) case
ex+y = ex ey ,
all convergences being absolute.

Corollary 2.1 For all X M(n, k),

eX GL(n, k),
ie eX is invertible, with
(eX )1 = eX .
Thus the exponential establishes a map
M(n, k) GL(n, k) : X 7 eX .
Proposition 2.3 For T GL(n, k), X M(n, k),
1

T eX T 1 = eT XT .
Proof I For each m,
(T XT 1 )m = T X m T 1 .
The result follows on summation.

Proposition 2.4 If the eigenvalues of X are

1 , . . . , n
then the eigenvalues of eX are
e1 , . . . , en .
Proof I As we saw in Chapter 1, each matrix X M(n, C) defines a matrix
RX M(2n, R). Conversely, each matrix X M(n, R) defines a matrix CX
M(n, C), namely the matrix with the same entries (now regarded as complex
numbers).
Lemma 2.2

1. C(eX ) = eCX

for all X M(n, R)

424III
2. R(eX ) = eRX

26

for all X M(n, C)

Proof of Lemma B
1. This is immediate, since the matrices arising on each side are identical, the
only difference being that in one case they are regarded as real and in the
other as complex.
2. This follows from
(RX)m = R(X m ),
on summation.
C

Proof of Proposition 4, continued: Since X and CX have the same eigenvalues,

we can suppose X complex, by Part 1 of the Lemma.
Although a complex matrix cannot in general be diagonalised, it can always
be brought to triangular form:

T XT

=Y =

1 a12
0 2
..
..
.
.
0 0

. . . a1n
. . . a2n

..
..
. .

. . . n

(This can be proved by induction on n. Taking an eigenvector of X as first

element of a new basis,

SXS 1 =

b12 . . . b1n
0
..
.
T1
0

The result follows on applying the inductive hypothesis to the (n 1)

(n 1) matrix T1 .)
But then

c12 . . . c1n
m
1
0 m
. . . c2n

2
..
..
..
. .

.
m
0
0 . . . n

424III

27

and so on summation

eY

e1 w12 . . . w1n
0 e2 . . . w2n
..
..
...
.
.
...
0
0 . . . en

The result now follows, since

eY = T eX T 1
by Proposition 3; so eX and eY have the same eigenvalues.
J

Corollary 2.2 For each X M(n, k),

det eX = etr X .
Proof I If the eigenvalues of X are 1 , . . . , n then
det X = 1 . . . n
tr X = 1 + . . . + n .
Hence
det eX = e1 en = e1 ++n = etr X .
J

Proposition 2.5

1. eX = (eX )0

2. eX = (eX )

for all X M(n, R)

for all X M(n, C)

Proof I For each m,

(X 0 )m = (X m )0 ,
(X )m = (X m ) .
The result follows on summation. J
We turn to the analytic properties of the exponential map.
Proposition 2.6 There exists an open neighbourhood U 3 0 in M(n, R) which
is mapped homeomorphically by the map X 7 eX onto an open neighbourhood
V = eU 3 I in GL(n, R).

424III

28

Lemma 2.3 Suppose X M(n, R) satisfies

kXk < 1.
Then
1. T = I + X GL(n, R).
2. The series
log T = X

X2 X3
+
...
2
3

is convergent.
3. elog T = T .
Proof of Lemma B
1. Explicitly,
(I + X)1 = I X + X 2 X 3 + . . . .
2. Convergence follows as in the scalar case, but with the matrix norm kXk in
place of the absolute value |x|.
3. We know that if |x| < 1 then
1 + x = elog(1+x) = 1 + (x x2 /2 + . . .) + (x x2 /2 + . . .)2 /2! + . . . .
Moreover the convergence on the right is absolute; and the identity therefore
holds for any matrix X satisfying kXk < 1.
C

Proof of Proposition I Let

V = {T GL(n, R) : kT Ik < 1};
and let
U = log V = {log T : T V }.
Then it follows from Part 3 of the Lemma that the maps
T 7 log T : V U and X 7 eX : U V
are mutually inverse. Since eX and log T are continuous (as in the scalar case)
and V is open, it follows that U is also open. J

424III

29

Remark: We shall call the region U (and also occasionally its image V ) the logarithmic zone, since for T U the exponential has the well-defined inverse log T :
elog T = T
log eX = X

for all T U
for all X V

Proposition 2.7 For each X M(n, R) the map

t 7 etX : R GL(n, R)
is a continuous homomorphism; and every continuous homomorphism R GL(n, R)
is of this form.
Proof I Since sX and tX commute, it follows from Proposition 3 that
esX etX = e(s+t)X .
Hence the map t etX is a homomorphism, which is clearly continuous.
Conversely, suppose
T : R GL(n, R)
is a continuous homomorphism. For sufficiently small t, say
t J = [c, c],
T (t) must lie in the logarithmic zone; and we can therefore set
X(t) = log T (t)
for t J.
Since T : R GL(n, R) is a homomorphism,
T (s)T (t) = T (s + t).
We want to convert this to the additive form
X(s) + X(t) = X(s + t)
by taking logarithms. To this end, note first that T (s) and T (t) commute:
T (s)T (t) = T (s + t) = T (t)T (s)

for all s, t.

424III

210

It follows that X(s) and X(t), as power-series in T (s) and T (t), also commute:
for all s, t J.

X(s)X(t) = X(t)X(s)

So by Proposition 2, if s, t and s + t all lie in J,

eX(s)+X(t) = eX(s) eX(t) = T (s)T (t) = T (s + t) = eX(s+t) .
Now if s and t are small enough, say
s, t J 0 = [c0 , c0 ],
then not only X(s), X(t) and X(s + t), but also X(s) + X(t), will lie in the
logarithmic zone V . In that case, on taking logarithms in the last relation,
for all s, t J 0 .

X(s) + X(t) = X(s + t)

We want to deduce from this that
X(t) = tX

for some X. Note first that, on replacing t by c0 t (and X by c0 1 X), we can

suppose that c0 = 1, ie that the last relation holds for
s, t I = [1, 1].
We have to show that
X(t) = tX(1)

for all t I.

Suppose first that s is a positive integer. By repeated application of the basic

identity,
1
1
1
1
1
sX( ) = X( ) + . . . + X( ) = X( + . . . + ) = X(1),
s
s
s
s
s
ie

1
1
X( ) = X(1).
s
s

Now suppose 0 r s. Then

r
1
1
1
r
X( ) = X( ) + . . . + X( ) = rX( ) = X(1).
s
s
s
s
s
We have thus established the result for rationals t = r/s [0, 1]. Since the
rationals lie densely among the reals, it follows by continuity that
X(t) = tX(1)

for all t [0, 1].

424III

211

The extension to the negative interval [-1,0] is immediate, since

X(0) + X(0) = X(0) = X(0) = 0,
and so
X(t) + X(t) = X(0) = X(t) = X(t).
Returning to the given homomorphism,
T (t) = eX(t) = etX

for all t J 0 .

We can extend the range by taking powers:

T (nt) = T (t)n = (etX )n = entX

for all n N, t J 0 .

Hence
T (t) = etX

for all t R.

Finally, the uniqueness of X follows on taking logarithms for sufficiently

small t. J
Summary: The exponential map sets up a 11 correspondencemore precisely, a homeomorphismbetween a neighbourhood U of 0 in M(n, R)
and a neighbourhood V of I in GL(n, R). The inverse logarithmic function projects this polar region V back into M(n, R)which we can identify with the tangent-space to GL(n, R) at I. The picture is reminiscent of
Mercators projection of the globe onto the pages of an atlas, or Riemanns
projection of the sphere onto a plane.

424III

212

Exercises
In Exercises 0115 calculate eX for the given matrix X:

1.

1 0
0 2

6.

1 1
1 1

11.

a b
b a

2.

0 1
0 0

7.

1 1
1 1

12.

0 b
b 0

!
!

3.

1 2
0 1

8.

1 1
0 0

13.

a b
b a

4.

1 0
1 1

9.

1 1
2 2

14.

a b
b a

5.

0 1
1 0

10.

1 1
1 1

15.

a b
c a

20.

1 0
0 1

25.

1 2
2 1

In Exercises 1625 determine whether or not the given matrix is of the form eX
for some X M(2, R).

16.

1 0
0 0

21.

0 1
0 0

17.

1 0
0 1

22.

1 2
2 1

!
!

18.

2 0
0 3

23.

2 1
1 2

19.

1 1
0 1

24.

2 1
1 2

!
!

Chapter 3
The Lie Algebra of a Linear Group
I: The Underlying Space
Suppose G GL(n, R) is a linear group. The rays in M(n, R) corresponding to 1-parameter subgroups trapped inside G fill a vector subspace
LG. This correspondence between closed subgroups of GL(n, R) and certain subspaces of M(n, R) is the foundation of Lie theory.

Definition 3.1 Suppose G GL(n, R) is a linear group. Then we set

LG = {X M(n, R) : etX G

for all t R}.

Remark: In the case of the classical groups GL(n, R), SO(n, R), etc, considered
in Chapter 1, it is customary to denote the corresponding space by the same letters
in lower case, eg
o(n) = LO(n),
sl(n, R) = LSL(n, R),
sp(n) = LSp(n).

Proposition 3.1 If G GL(n, R) is a linear group then LG is a vector subspace

of M(n, R).
Proof I We have to show that
1. X LG, a R = aX LG,
2. X, Y LG = X + Y LG.
424III

31

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

32

The first result follows at once from the definition of LG. The second is a
consequence of the following result.
Lemma 3.1 Suppose T (t) G for 0 < t d; and suppose
T (t) = I + tX + o(t)
for some X M(n, R), ie
(T (t) I)/t X as t 0.
Then
X LG.
Proof of Lemma B We must show that
etX G

for all t.

The argument can be divided into 5 steps.

1. Recall the formula for the scalar exponential as a limit:
(1 +

x m
) ex
m

as m .

This is most simply proved by taking the logarithm of each side. On the left
x
log (1 + )m
m


x
)
m
x
1
= m( + o( ))
m
m
= x + o(1).

= m log(1 +

In other words
x
log (1 + )m x
m


as m .

The result follows on taking the exponentials of both sides.

2. It is evident from this proof that we can replace 1 + x/m by any function
a(m) satisfying
1
x
+ o( );
a(m) = 1 +
m
m
for any such function
a(m)m ex .

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

33

3. Both this last result and its proof carry over to the matrix case. If
A=I+
with X M(n, R), then

X
1
+ o( ),
m
m

A(m)m eX

4. Applying this with

A(m) = T (
we deduce that
T(

t
)
m

t m
) etX .
m

5. Since T (t/m) G for sufficiently large m, and since G is a group,

T(

t m
) G.
m

Hence, since G is closed,

etX G.
C

Remark: In geometric language this result shows that LG can be regarded as the
tangent-space to G at T = I.
Proof of Proposition 1 (completion). Suppose X, Y LG. Then
etX etY G

for all t.

But
etX etY

= (I + tX)(I + tY ) + o(t)
= I + t(X + Y ) + o(t).

Hence by the Lemma

X + Y LG.
J

Examples:
1. The General Linear Group: Since etX GL(n, R) for all X M(n, R),
by Proposition 2.1,
gl(n, R) = M(n, R).

THE LIE ALGEBRA I: THE UNDERLYING SPACE

2. The Special Linear Group:

424III

34

sl(n, R) consists of all trace-free matrices, ie

sl(n, R) = {X M(n, R) : tr X = 0}.

By definition
X sl(n, R) = etX SL(n, R) for all t
= det(etX ) = 1 for all t.
To the first order in t,
etX = I + tX + o(t).
Hence
det(etX ) = det(I + tX) + o(t)
= 1 + t tr X + o(t);
for on expanding the determinant the off-diagonal terms in tX will only
appear in second or higher order terms. Hence
X sl(n, R) = tr X = 0.
Conversely,
tr X = 0 = det etX = et tr X = 1
= etX SL(n, R).
Remarks:
(a) A linear group G is said to be algebraic if it can be defined by polynomial conditions Fk (T ) on the matrix entries Tij , say
G = {T GL(n, R) : Fk (T ) = 0, k = 1, . . . , N }.
Not every linear group is algebraic; but all the ones we meet will be.
The technique aboveworking to the first order in tis the recommended way of determining the Lie algebra LG of an unknown linear group G. From above,
X LG = etX G
= Fk (I + tX + ) = 0
= Fk (I + tX) = O(t2 )

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

35

Thus
LG {X : Fk (I + X) = O(t2 )}.
In theory it only gives us a necesssary condition for X to lie in LG;
and it is easy to think of artificial cases where the condition is not in
fact sufficient. For example the equation
tr(I T )0 (I T ) = 0
defines the trivial group {I}; but it is satisfied by etX to the first order
in t for all X.
In practice the condition is usually sufficient. However it must be
verifiedhaving determined LG (as we hope) in this way, we must
then show that eX does in fact lie in G. (Since the condition on X is
linear this will automatically ensure that et X G for all t R.)
(b) An alternative way of describing the technique is to say that since each
defining condition is satisfied by etX identically in t, the differential of
this condition must vanish at t = 0, eg
sl(n, R) {X :

d
det etX = 0 at t = 0}.
dt

(c) To summarise: Given a linear group G,

i. Find all X for which I + tX satisfies the equations for G to the
first order in t; and then
ii. Verify that eX G for these X.
3. The Orthogonal Group:

o(n) consists of all skew-symmetric matrices, ie

o(n) = {X : X 0 + X = 0}.
For
(I + tX)0 (I + tX) = I + t(X 0 + X) + o(t).
Hence
X o(n) = X 0 + X = 0.
Conversely,
X 0 + X = 0 = X 0 = X
0
= (eX )0 = eX = eX = (eX )1
= eX O(n).

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

36

4. The Special Orthogonal Group:

so(n) = o(n).
This follows from the trivial
Lemma 3.2 If G is an intersection of linear groups, say
G = Gi ,
then
LG = LGi .
Applying this to
SO(n) = O(n) SL(n, R)
we deduce that
so(n) = o(n) sl(n, R).
But
X 0 + X = 0 = tr X = 0,
ie
o(n) sl(n, R).
Hence
so(n) = o(n) sl(n, R) = o(n).
The reason why SO(n) and O(n) have the same Lie algebra is that they
coincide in the neighbourhood of I. In effect, O(n) has an extra piece
far from I, where det T = 1. Since Lie theory only deals with what
happens in the neighbourhood of I, it cannot distinguish between groups
that coincide there.
Technically, as we shall see in Chapter 4, 2 linear groups having the same
connected component of I (like SO(n) and O(n)) will have the same Lie
algebra.
5. The Complex General Linear Group:
gl(n, C) = M(n, C).

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

37

For from Example 1.5,

X M(n, C) =
=
=
=
=

XJ = JX
X m J = JX m
etX J = JetX
etX GL(n, R)
X gl(n, R);

while conversely
etX GL(n, C)

for all t = etX J = JetX

= XJ = JX
= X M(n, C),

on equating coefficients of t.
6. The Complex Special Linear Group:
sl(n, C) = {X M(n, C) : tr X = 0}.
Note that tr X here denotes the trace of X as a complex matrix. The result
follows exactly as in Example 2.
7. The Unitary Group:

u(n) consists of all skew-hermitian matrices, ie

u(n) = {X M(n, C) : X + X = 0}.

For
(I + tX) (I + tX) = I + t(X + X) + o(t).
Hence
X u(n) = X + X = 0.
Conversely,
X + X = 0 =
=
=
=

X = X

(etX ) = etX = etX = (etX )1

etX U(n) for all t
X u(n).

8. The Special Unitary Group:

su(n) = {X M(n, C) : X + X = 0, tr X = 0}.

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

38

This follows at once from the Lemma above:

SU(n) = U(n) SL(n, C) = su(n) = u(n) sl(n, C).
Notice that su(n) is not the same as u(n); a skew-hermitian matrix is not
necessarily trace-free.
9. The Symplectic Group:
sp(n) = {X M(2n, C) : X + X = 0, XJ = JX}.
For
Sp(n) = Sp(n, C) U(2n),
where
Sp(n, C) = {T GL(n, C) : XJ = JX}.
Hence
sp(n) = sp(n, C) u(2n),
by the Lemma above. The result follows, since
sp(n, C) = {X M(2n, C) : XJ = JX}
just as in Example 5.
Definition 3.2 The dimension of a linear group G is the dimension of the real
vector space LG:
dim G = dimR LG.
Examples:
1. dim GL(2, R) = 4.
2. dim SL(2, R) = 3. For the general matrix in sl(2, R) is of the form
a b
c a

giving a vector space of dimension 3.

3. dim O(n) = 3. For the general matrix in o(3) is of the form

0 a b

0 c
a

b c
0

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

39

4. dim SO(3) = 3. For so(3) = o(3).

5. dim U(2) = 4. For the general matrix in u(2) is of the form
ia
b + ic
b + ic
id

6. dim SU(2) = 3. For the general matrix in su(2) is of the form

ia
b + ic
b + ic
ia

7. dim Sp(1) = 3. For the general complex matrix in sp(1, C) is

z w
w z

If this is also in u(2) then z must be pure imaginary, say

ia
b ic
b ic
ia

In Chapter 2 we defined the logarithmic zone in GL(n, R): an open neighbourhood of I mapped homeomorphically onto an open set in M(n, R). Our next
result shows that every linear group has a logarithmic zone.
Proposition 3.2 Suppose G is a linear group. Then there exists an open neighbourhood W 3 0 in LG which is mapped homeomorphically by eX onto an open
neighbourhood eW 3 I in G.
Proof I This result is rather remarkable. It asserts that there exists a > 0 such
that
kXk , eX G = X LG.
Suppose this is not so. Then we can find a sequence Xi M(n, R) such that
Xi 0, eXi G, Xi
/ LG.
Let us resolve each Xi into components along and perpendicular to LG (where
perpendicularity is taken with respect to the inner product associated to the quadratic
form kXk2 ):
Xi = Yi + Zi
(Yi LG, Zi LG).

THE LIE ALGEBRA I: THE UNDERLYING SPACE

Notice that kYi k, kZi k kXi k.
Consider the set of matrices

424III

310

Zi
kZi k

on the unit ball in M(n, R). Since this ball is compact, we can find a convergent
subsequence. Taking this subsequence in place of the original sequence, we may
assume that
Zi
Z
kZi k
Since Zi LG it follows that Z LG.
Now consider the sequence
Ti = eYi eXi
Y2
(Yi + Zi )2
= I Yi + i
I + (Yi + Z + i) +
+
2!
2!
= I + Zi + O(kXi kkZi k),
!

since each remaining term will contain Zi and will be of degree 2 in Yi , Zi . Let
kZi k = ti . Then
Zi = ti (Z + Ei ),
where Ei 0. Thus
Ti = I + ti Z + o(ti ).
From the Lemma to Proposition 1 above, this implies that
Z LG.
But Z LG. So our original hypothesis is untenable; we can find a > 0 such
that if kXk then
eX G X LG.
J

Corollary 3.1 If G is a linear group then

LG = 0 G is discrete.
Proposition 3.3 The connected component of I in a linear group G is a normal
open subgroup G0 , generated by the exponentials eX (X LG):
G0 = {eX1 eX2 . . . eXr : X1 , X2 , . . . , Xr LG}.

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

311

Proof I The set of all such matrices is clearly closed under multiplication and
inversion, and so forms a subgroup G0 .
The path
t 7 etX1 . . . etXr (0 t 1)
connects I to
eX1 . . . eXr 3 G0 .
Hence G0 is connected.
By Proposition 2, eLG is a neighbourhood of 0 G. Hence geLG is a neighbourhood of g for each g G. But
geLG G0 if g G0 .
Hence G0 is open. Recall that this implies G0 is also closed. (For each coset of G0
is open. Hence any union of cosets is open. Hence the complement of G0 , which
is the union of all cosets apart from G0 itself, is open, ie G0 is closed.)
Since G0 is also connected, it must be the connected component of I in G.
Finally, if g G then gG0 g 1 is also connected. Hence
gG0 g 1 G0 ,
ie G0 is normal in G.

Remarks:
1. Note that by this Proposition, if G is a linear group then
G connected = G arcwise connected.
2. This Proposition is usually applied in reverse, ie we first determine (by some
other means) the connected component G0 of I in G. The Proposition then
shows that each element of G0 is expressible as a product of exponentials
of elements of LG.
The following resultwhich really belongs to homotopy theoryis often
useful in determining G0 .
Lemma 3.3 Suppose the compact linear group G acts transitively on the compact
space X; and suppose x X. Let S = S(x) denote the stabiliser subgroup of x,
ie
s = {T G : T x = x}.
Then
S and X connected = G connected.

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

312

Remark: Those familiar with homotopy will recognise in this the 0-dimensional
part of the infinite exact homotopy sequence
0 0 (S) 0 (G) 0 (X) 1 (S) . . . ,
where i denotes the ith homotopy group) of the fibre space (G, X, S) with total
space G, base space X, fibre S and projection
g 7 gx : G X.
Although we shall assume no knowledge of homotopy theory, it is interesting to
note that the 1-dimensional part of this sequence will play a similar role in Chapter
7, in showing that
S and X simply connected = G simply connected.
Proof of Lemma B Suppose g G0 . Then gS is connected; and so gS G0 .
Hence
G0 = G0 S.
Since G0 is closed (and so compact), so is
gG0 x X.
On the other hand since G0 is open, G G0 is closed, and so therefore is
(G G0 )x = X G0 x.
Thus G0 x is both open and closed; and so
G0 x = X,
since X is connected. But since G0 = G0 S this implies that
G0 = G,
ie G is connected.

Examples:
1. SO(n) is connected for all n. For consider the action
(T, v) T v
of SO(n) on the (n 1)-sphere
S n1 = {v Rn : |v| = 1}.

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

313

This action is transitive; and the stabiliser subgroup of

en = (0, 0, . . . , 1)
can be identified with SO(n 1). Applying the Lemma,
SO(n 1) and S n1 connected = SO(n) connected.
Thus by induction, starting from SO(1) and S 1 , we deduce that SO(n) is
connected for all n.
2. SO(n) is the connected component of I in O(n). For
T O(n) = T 0 T = I = (det T )2 = 1 = det T = 1.
Thus no path in O(n) can enter or leave SO(n), since this would entail a
sudden jump in det T from +1 to 1, or vice versa.
3. U(n) is connected for all n. For consider the action of U(n) on
S 2n1 = {v C n : |v| = 1}.
The stabiliser subgroup of
en = (0, 0, . . . , 1)
can be identified with U(n 1); and so we deduce by induction, as in
Example 1, that U(n) is connected for all n.
4. SU(n) is connected for all n. Again this follows from the action of SU(n)
on S 2n1 .
5. Sp(n) is connected for all n. This follows in the same way from the action
of Sp(n) on
S 4n1 = {v Hn : |v| = 1}.
6. SL(n, R) is connected for all n. For suppose T SL(n, R). Then T 0 T is a
positive-definite matrix, and so has a positive-definite square-root, Q say:
T 0 T = Q2 .
(To construct Q, diagonalise the quadratic form v 0 T v and take the squareroot of each diagonal element.) Now set
O = T Q1 .

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

314

Then
O0 O = I,
ie O O(n). In fact, since det T = 1 and det Q > 0,
O SO(n).
Thus
T = OQ,
where O SO(n) and Q P , the space of positive-definite matrices. Thus
SL(n, R) = SO(n)P.
Now P is connected; in fact it is convex:
A, B P = tA + (1 t)B P

for all t [0, 1].

Since SO(n) is also connected, so too is

SL(n, R) = SO(n)P.
Summary: To each linear group G GL(n, R) there corresponds a vector
subspace LG M(n, R). Two linear groups G and H correspond to the
same subspace if and only if they have the same connected component of
the identity: LG = LH G0 = H0 . In other words, there is a 11
correspondence between connected linear groups G and the subspaces LG.
The exponential map and the inverse logarithmic map define a homeomorphism (but not a homomorphism) between a neighbourhood U of I in
G and a neighbourhood V of 0 in LG. With a little artistic licence we may
say that the logarithmic projection turns subgroups into subspaces.

THE LIE ALGEBRA I: THE UNDERLYING SPACE

424III

315

Exercises
In Exercises 0110 determine the dimension of the given group
1.

GL(n, R) 2. SL(n, R) 3.

6.

SL(n, C)

7.

U(n)

O(n)

4. SO(n)

8. SU(n) 9.

Sp(n)

5. GL(n, C)
10.

O(1, 3)

In Exercises 1115 determine the connected component of I, and the number

of components, in each of the following groups.
11.

GL(n, R) 12.

GL(n, C) 13.

SL(n, C) 14.

O(1, 1) 15.

O(1, 3)

Chapter 4
The Lie Algebra of a Linear Group
II: The Lie Product
The subspace LG corresponding to a linear group G is closed under the Lie
product [X, Y ] = XY Y X, and thus consitutes a Lie algebra. Algebraically, the Lie product reflects the non-commutativity of Gan abelian
group has trivial Lie algebra. Geometrically, the Lie product measures the
curvature of G.

Definition 4.1 For X, Y M(n, R) we set

[X, Y ] = XY Y X.
The matrix [X, Y ] is called the Lie product of X and Y .
Proposition 4.1 If G is a linear group then
X, Y LG = [X, Y ] LG.
Proof I Our argument is very similar to that used in the proof of Proposition 3.1
to show that
X, Y LG = X + Y LG.
Suppose
X, Y LG.
Then etX , etY G for all t R; and so
etX etY etX etY G.

424III

41

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

42

But
etX etY etX etY
t2 X 2
t2 Y 2
t2 X 2
t2 Y 2
)(I + tY +
)(I tX +
)(I tY +
) + o(t2 )
2
2
2
2
= I + t2 [X, Y ] + o(t2 )
= I + s[X, Y ] + o(s),
= (I + tX +

with s = t2 . Hence by the Lemma to Proposition 3.1,

[X, Y ] LG.
J

Proposition 4.2 The Lie product [X, Y ] defines a skew-symmetric bilinear map
LG LG LG, ie
1. [aX, Y ] = a[X, Y ]
2. [X1 + X2 , Y ] = [X1 , Y ] + [X2 , Y ]
3. [X, Y1 + Y2 ] = [X, Y1 ] + [X, Y2 ]
4. [Y, X] = [X, Y ]
In addition it satisfies Jacobis identity
[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0
Proof I All is clear except for (J), and that is a matter for straightforward verification:
[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]]
= X(Y Z ZY ) (Y Z ZY )X + Y (ZX XZ)
(ZX XZ)Y + Z(XY Y X) (XY Y X)Z
= 0,
the 12 terms cancelling in pairs.

Definition 4.2 A Lie algebra L over k (where k = R or C) is a finite-dimensional

vector space L over k, together with a skew-symmetric bilinear map L L L,
which we denote by [X, Y ], satisfying Jacobis identity (J).
If G is a linear group then the real Lie algebra defined on LG by the Lie
product is called the Lie algebra of G.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

43

Remarks:
1. Henceforth LG (and similarly gl(n, R), etc) will denote the Lie algebra of
G, ie the space LG together with the Lie product on this space.
2. Note that the Lie algebra of a linear group is always real, even if G is complex, ie G GL(n, C). The point of introducing complex Lie algebras
will only become apparent in Chapter7, when we consider complex representations of linear groups.
3. It follows at once from the skew-symmetry of the Lie product that
[X, X] = 0 X L.
4. In defining the Lie product in a Lie algebra L with basis e1 , . . . , em it is only
necessary to give the m(m-1)/2 products
[ei , ej ]

(i < j)

since these determine the general product [X, Y ] by skew-symmetry and

linearity.
5. With the same notation, we can express each Lie product [ei , ej ] of basis
elements as a linear combination of these elements:
[ei , ej ] =

ckij ek .

The m3 scalars
ckij

(1 i, j, k m)

are called the structure constants of the Lie algebra. In theory we could
define a Lie algebra by giving its structure constants; in practice this is
rarely a sensible approach.
Examples:
1. The space so(3) consists of all skew-symmetric 3 3 matrices. As basis we
might choose

0 0 0

U = 0 0 1
,
0 1 0

0 0 1

V = 0 0 0
,
1 0 0

0 1 0

W = 1 0 0

0 0 0

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

44

It is readily verified that

[U, V ] = U V V U = W, [U, W ] = U W W U = V, [V, W ] = V W W V = U.
Thus we can write
so(3) = hU, V, W : [U, V ] = W, [V, W ] = U, [W, U ] = V i.
2. The space su(2) consists of all skew-hermitian 2 2 complex matrices with
trace 0. As basis we might choose
A=

i 0
0 i

B=

0 1
1 0

C=

0 i
i 0

It is readily verified that

[A, B] = 2C, [A, C] = 2B, [B, C] = 2A.
Thus
su(2) = hA, B, C : [A, B] = 2C, [B, C] = 2A, [C, A] = 2Bi.
Notice that the Lie algebras so(3) and su(2) are isomorphic,
so(3) = su(2),
under the correspondence
2U A, 2V B, 2W C.
Intuitively, we can see how this isomorphism arises. The covering homomorphism : SU(2) SO(3) establishes a local isomorphism between
SU(2) and SO(3). If one remains close to I these 2 groups look exactly the
same. But Lie theory is a local theorythe Lie algebra LG depends only
on the structure of G near to I. (In particular, it depends only on the connected component of I in G, whence so(3) = o(3).) Thus covering groups
have the same Lie algebra. We shall returnmore rigorouslyto this very
important point in Chapter 8.
But we note now that the groups SO(3) and SU(2) are certainly not isomorphic, since one has trivial, and the other non-trivial, centre:
ZSO(3) = {I, I}, ZSU(2) = {I}.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

45

3. The space sl(2, R) consists of all real 2 2 matrices with trace 0. As basis
we might choose
H=

1 0
0 1

E=

0 1
0 0

F =

0 0
1 0

It is readily verified that

[H, E] = 2E, [H, F ] = 2F, [E, F ] = H.
Thus
sl(2, R) = hH, E, F : [H, E] = 2E, [H, F ] = 2F, [E, F ] = Hi.
This Lie algebra is not isomorphic to so(3) = su(2). For the subspace
hU, V i is closed under the Lie product. (It corresponds, as we shall see in
Chapter 6, to the subgroup of lower triangular matrices.) But it is readily
verified that no 2-dimensional subspace of so(3) is closed under the Lie
product.
4. As an exercise, let us determine all the Lie algebras of dimension 3. This
will allow us to introduce informally some concepts which we shall define
formally later.
(a) The Lie algebra LG is said to be abelian if the Lie product is trivial:
[X, Y ] = 0

for all X, Y L.

We shall look at abelian Lie algebras in the next Chapter. Evidently

there is just 1 such algebra in each dimension.
(b) The derived algebra L0 of a Lie algebra L is the subspace spanned by
all Lie products:
L0 = {[X, Y ] : X, Y L}.
This is an ideal in L:
X L, Y L0 = [X, Y ] L0 .
Evidently L0 = 0 if and only if L is abelian.
(c) The centre of the Lie algebra L,
ZL = {Z L : [X, Z] = 0 for all X L},
is also an ideal in L. Evidently L is abelian if and only if ZL = L.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

46

(d) Suppose X L. We denote by ad X : L L the linear map defined

by
ad X : Y 7 [X, Y ].
Jacobis identity can be re-stated in the form
ad[X, Y ] = ad X ad Y ad Y ad X.
To see this, apply both sides to Z. On the left we have
ad[X, Y ](Z) = [[X, Y ], Z] = [Z, [X, Y ]].
On the right we have
ad X ad Y (Z)ad Y ad X(Z) = [X, [Y, Z]][Y, [X, Z]] = [X, [Y, Z]]+[Y, [Z, X]].
The two are equal by Jacobis identity.
Dimension 1 Since [X, X] = 0, if dim L = 1 then L is abelian.
Dimension 2 Either L is abelian, or dim L0 = 1, since
[aX + bY, cX + dY ] = (ad bc)[X, Y ].
Suppose dim L0 = 1, say L0 = hXi. Then we can find a Y such that
[X, Y ] = X. Thus
L = hX, Y : [X, Y ] = Xi.
So there are just 2 Lie algebras of dimension 2.
Dimension 3 We have dim L0 = 0, 1, 2 or 3.
If dim L0 = 0 then L is abelian.
If dim L0 = 1, say L0 = hXi, then it is not hard to show that the centre
ZL must also be of dimension 1; and
L = L1 ZL,
where L1 is a 2-dimensional Lie algebra. Thus there is just 1 Lie
algebra in this category:
L = hX, Y, Z : [X, Y ] = X, [X, Z] = [Y, Z] = 0i.
If dim L0 = 2, say L0 = hX, Y i, then L0 must be abelian. For suppose
not. Then we can find X, Y L0 such that [X, Y ] = X. Suppose
[Y, Z] = aX + bY, [Z, X] = cX + dY.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

47

Then Jacobis identity gives

[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = bX cX + cX + dY.
Thus b = d = 0. But then L0 = hXi, contrary to hypothesis.
Thus L0 = hX, Y i is abelian. Take any Z L \ L0 . Since ad Z(L)
L0 , ad Z defines a map
AZ : L0 L0 .
If we take Z 0 = aX + bY + cZ in place of Z then
AZ 0 = cAZ .
Thus the map AZ is defined up to a scalar multiple by L. Conversely,
it is readily verified that any linear map AZ : L0 L0 defines a Lie
algebra, since all 3 terms in Jacobis identity for X, Y, Z vanish.
There are 2 families of solutions, according as the eigenvalues of AZ
are real, or complex conjugates. Note that AZ is surjective, since
adZ (L) = L0 . Hence the eigenvalues of AZ are not both 0 (ie AZ
is not idempotent).
If the eigenvalues are complex conjugates ei then we can make = 1
by taking 1 Z in place of Z. This gives the family of Lie algebras
L() = hX, Y, Z : [X, Y ] = 0, [Z, X] = cos Xsin Y, [Z, Y ] = sin X+cos Y i.
Similary if the eigenvalues are real we can take them as 1, . and we
obtain the family
L() = hX, Y, Z : [X, Y ] = 0, [Z, X] = X, [Z, Y ] = Y i.
We come now to the case of greatest interest to us, where
L0 = L.
A Lie algebra with this property is said to be semisimple. (We shall
give a different definition of semisimplicity later, but we shall find in
the end that it is equivalent to the above.) Perhaps the most important
Theorem in this Part is that every representation of a semisimple Lie
algebra is semisimple. But that is a long way ahead.
Since
ad[X, Y ] = ad X ad Y ad Y ad X,

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

48

it follows that
tr ad[X, Y ] = 0.
Thus
X L0 = tr ad X = 0.
In particular if L0 = L then tr ad X = 0 for all X L. Hence the
eigenvectors of ad X are 0, . On the other hand, the characteristic
equation is real. So the eigenvectors are either of the form 0, or
0, i with real.
Before going further, let us dispose of one possibility: that the eigenvalues of ad X might always be 0, 0, 0. Recall that this is the case if
and only if ad X is nilpotent. (This follows from the Cayley-Hamilton
Theorem, that a linear transformationor square matrixsatisfies its
own characteristic equation.)
A Lie algebra L is said to be nilpotent if ad X is nilpotent for all X
L. As we shall see later (Engels Theorem) a nilpotent Lie algebra L
cannot have L0 = L, ie L cannot be both nilpotent and semisimple.
This is very easy to establish in the present case, where dim L = 3.
Note first that
L = hX, Y, Zi = L0 = h[Y, Z], [Z, X], [X, Y ]i.
It follows that if L0 = L then
[X, Y ] = 0 X, Y linearly dependent.
Now suppose ad X is nilpotent for some X 6= 0. Then ad X =
6 0,
since otherwise [X, Y ] = 0 for all Y L. Thus we can find Y L
such that
Z = ad X(Y ) 6= 0 but ad X(Z) = [X, Z] = 0.
This implies, as we have seen, that Z = X. Thus
ad Y (X) = [Y, X] = [X, Y ] = X.
So ad Y has eigenvalue 6= 0, and is not nilpotent.
Thus there exists an X L with ad X not nilpotent, with the 2 possibilites outlined above.
(a) For some X L, ad X has eigenvalues 0, where > 0.
Taking 1 X in place of X, we may suppose that = 1. Taking the
eigenvectors of X as a basis for L, we get
[X, Y ] = Y, [X, Z] = Z.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

49

Suppose
[Y, Z] = aX + bY + cZ.
Jacobis identity yields
[X, [Y, Z]]+[Y, [Z, X]]+[Z, [X, Y ]] = bY cZ+[Y, Z]+[Z, Y ] = 0.
Thus b = c = 0, ie
[Y, Z] = aX.
Dividing Y by a, we get
[Y, Z] = X.
So we have just 1 Lie algebra,
L = hX, Y, Z : [X, Y ] = Y, [X, Z] = Z, [Y, Z] = Xi.
In fact
L = sl(2, R)
under the correspondence X 7 21 H, E 7 Y, F 7 Z.
(b) Alternatively, ad X has eigenvalues 0, i with some 0 for
every X L. (For otherwise we fall into the first case.) Choose
one such X. As before, on replacing X by 1 we may suppose that
= 1, ie ad X has eigenvalues 0, i. Taking the i-eigenvector of
ad X to be Z + iY ,
[X, Z + iY ] = i(Z + iY ) = Y + iZ.
Thus
[X, Z] = Y, [X, Y ] = Z.
Suppose
[Y, Z] = aX + bY + cZ.
Jacobis identity yields
[X, [Y, Z]]+[Y, [Z, X]]+[Z, [X, Y ]] = bZcY +[Y, Y ]+[Z, Z] = 0.
Thus b = c = 0, ie
[Y, Z] = aX.
Dividing Y, Z by

|a|, we may suppose that

[Y, Z] = X.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

410

If [[Y, Z] = X then

0 0 1

ad Y = 0 0 0
.
1 0 0
This has eigenvalues 0, 1, and so falls into the first case. Hence
L = hX, Y, Z : [X, Y ] = Z, [Z, X] = Y, [Y, Z] = Xi.
In other words,
L = so(3).
It remains to show that so(3) = su(2) and sl(2, R) are not isomorphic. We shall see later that this follows from the fact that so(3) and
su(2) are compact, while sl(2, R) is not; for we shall show that the
compactness of a group G is reflected in its Lie algebra LG. But we
can reach the same result by a cruder argument. From our argument
above, it is sufficient to show that ad X has eigenvalues 0, i for
every X so(3). But it is readily seen that

0 0 0
0 0 1
0 1 0

ad U =
0 0 1 , ad V = 0 0 0 , ad W = 1 0 0 .
0 1 0
1 0 0
0 0 0
Thus, with respect to this basis, ad X is always represented by a skewsymmetric matrix. The result follows since the eigenvalues of such a
matrix are either purely imaginary or 0.
Summary: To each linear group G there corresponds a Lie algebra LG.
Most of the information about G is encoded in LG; and since LG is far
easier to analyseusing standard techniques of linear algebrait provides
a powerful tool in the study of linear groups.

THE LIE ALGEBRA II: THE LIE PRODUCT

424III

411

Exercises
In Exercises 0110 determine the Lie algebra of the given group
1. GL(2, R)

2.

SL(3, R)

3.

8. SL(2, C)

6.

SO(4)

7.

GL(2, C)

11.

Sp(1)

12.

Sp(2)

13.

O(2)

O(1, 1)

4.

O(3)

5.

SO(2)

9.

U(2)

10.

SU(3)

O(1, 3) 15.

O(2, 2)

14.

Chapter 5
Abelian Linear Groups
As a happy by-product, Lie theory gives us the structure of connected abelian
linear groups.

Definition 5.1 The Lie algebra L is said to be abelian if the Lie product is trivial,
ie
[X, Y ] = 0 for all X, Y L.
Proposition 5.1 If G is a linear group then
G abelian = LG abelian.
If in addition G is connected then
LG abelian G abelian.
Proof I Suppose G is abelian; and suppose
X, Y LG.
Then etX , etY commute for all t. If t is sufficiently small, tX and tY will lie in
the logarithmic zone U , so that
tX = log etX , tY = log etY
by Proposition ??. In particular tX, tY are expressible as power-series in etX , etY
respectively. Hence tX, tY commute; and so X, Y commute, ie LG is abelian.
Conversely, suppose G is connected and LG is abelian, ie
[X, Y ] = XY Y X = 0 X, Y LG.
Then eX , eY commute, by Proposition ??. and so therefore do any 2 products
eX1 . . . eXr , eY1 . . . eYs

(X1 , . . . , Xr , Y1 , . . . , Ys LG).

Hence G = G0 is abelian, by Proposition ??.

424III

51

424III

52

Proposition 5.2 If G is a connected abelian linear group then G is isomorphic to

a cylinder group, ie
G = Tr Rs ,
where T is the torus
T = R/Z.
In particular, the only compact connected abelian linear groups are the tori Tn .
Lemma 5.1 If G is a connected abelian linear group then the map
X 7 eX : LG G
is a surjective homomorphism (of abelian groups) with discrete kernel.
Proof of Lemma B The exponential map is a homomorphism in this case, by
Proposition ??. It is surjective, by Proposition ??, since
eX1 . . . eXr = eX1 +...+Xr

(X1 , . . . , Xr LG).

Its kernel is discrete, by Proposition ??; for the exponential map is one-one in the
logarithmic zone. C
Proof I By the Lemma,
G=

Rn
,
K

where K is a discrete subgroup of Rn .

We shall show by induction on n that such a subgroup has a Z-basis consisting
of m n linearly independent vectors
v1 , . . . , v m Rn ,
ie K consists of all linear combinations
a1 v 1 + . . . + am v m

(a1 , . . . , am Z).

Let v1 be the closest point of K to 0 (apart from 0 itself). We may suppose, on

choosing a new basis for Rn , that
v1 = (1, 0, . . . , 0).
Now let
p : Rn Rn1

424III

53

be the projection onto the last n 1 coordinates, ie

p(x1 , x2 , . . . , xn ) = (x2 , . . . , xn ).
Then the subgroup
pK Rn1
is discrete. For each point
(x2 , . . . , xn ) pK
arises from a point
v = (x1 , x2 , . . . , xn ) K;
and we may suppose, on adding a suitable integral multiple of v1 to v, that
1
1
< x1 .
2
2
But then v would clearly be closer to 0 than v1 (contrary to hypothesis) if x2 , . . . , xn
were all very small.
Applying the inductive hypothesis we can find a Z-basis for pK consisting of
linearly independent vectors
u2 , . . . , um Rn1 .
Choose
v 2 , . . . , v m Rn
such that
pvi = ui .
Then it is easy to see that v1 , v2 , . . . , vm are linearly independent, and form a Zbasis for K.
Again, on choosing a new basis for Rn we may suppose that
v1 = (1, 0, . . . , 0), v2 = (0, 1, . . . , 0), . . . , vm = (0, . . . , 1, . . . , 0),
ie
K = {(a1 , . . . , am , 0, . . . , 0) : ai Z}.
Then
G = Rn /K
= R/Z + . . . + R/Z + R + . . . + R
= Tm Rnm .
J

424III

54

Remark: We shall find this result extremely useful later, in studying the structure
of a general compact linear group G. For if we take any element g G then
the smallest closed subgroup of G containing g is abelian; and so the connected
component of I in this group must be a torus.
Summary: If G is abelian then so is LG; and the exponential map in this
case is a homomorphism, mapping onto the connected component G0 of G.
We deduce that every connected abelian linear group is of the form Tm Rn .

Chapter 6
The Lie Functor
To each linear group G we have associated a Lie algebra LG. But that is
only half the storyto complete it we must show that each homomorphism
G H of linear groups gives rise to a homomorphism LG LH of the
associated Lie algebras.

Definition 6.1 A homomorphism

f :LM
of Lie algebras over k is a linear map which preserves the Lie product, ie
1. f (aX) = a(f X)
2. f (X + Y ) = f X + f Y
3. f [X, Y ] = [f X, f Y ]
Proposition 6.1 Suppose
F :GH
is a continuous homomorphism of linear groups. Then there exists a unique homomorphism
f = LF : LG LH
of the corresponding Lie algebras such that
ef X = F (eX )
for all X LG.

424III

61

424III

62

Proof I Suppose X LG. Then the composition

t 7 etX 7 F (etX ) : R H
is a continuous homomorphism. Hence, by Proposition 2.5, there exists a unique
element f X LH such that
et(f X) = F (etX ) t R.
We must show that the 3 conditions of Definition 1 are satisfied.
1. This follows at once from the definition of f.
2. We saw in the proof of Proposition 3.1 that
(etX/m etY /m )m et(X+Y ) .
Applying the homomorphism F to each side,
(F (etX/m )F (etY /m ))m F (et(X+Y ) ),
ie
(etf X/m etf Y /m )m etf (X+Y ) .
But by the same Lemma,
(etf X/m etf Y /m )m et(f X+f Y ) .
Hence
etf (X+Y ) = et(f X+f Y )

t;

and so, by Proposition 2.5,

f (X + Y ) = f X + f Y.
3. We saw in the proof of Proposition 4.1 that
2

2 [X,Y

(etX etY etX etY )m et

The result follows from this as in (2).

J

Proposition 6.2 The assignment G LG is functorial, ie

424III

63

1. If
E : G H,

F :HK

are 2 continuous homomorphisms of linear groups then

L(F E) = (LF )(LE).
2. The identity on G induces the identity on LG,
L1G = 1LG .
Corollary 6.1 If the linear groups G and H are isomorphic then so are their Lie
algebras:
G
= H = LG
= LH.
Remark: This is a most important result, since it allows us to extend Lie theory
from linear to linearisable groups. Thus we can speak of the Lie algebra of a
topological group G, provided G is isomorphic to some linear groupwe need
not specify this linear group or the isomorphism, so long as we have established
that they exist. We shall return to this point later.
Proposition 6.3 Suppose
F :GH
is a continuous homomorphism of linear groups; and suppose G is connected.
Then F is completely determined by f = LF . More precisely, if
F1 , F2 : G H
are 2 such homomorphisms then
LF1 = LF2 = F1 = F2 .
Proof I Suppose T G. By Proposition 3.4,
T = eX1 . . . eXr .
Hence
F (T ) = F (eX1 . . . eXr )
= F (eX1 ) . . . F (eXr )
= ef X1 . . . ef Xr .

424III

Thus F (T ) is completely determined once f is known.

64

Remark: This result shows that if G is connected (and if it is not we can always
replace it by its connected component G0 ) then there is at most one group homomorphism
F :GH
corresponding to a given Lie algebra homomorphism
f : LG LH.
One might say that nothing is lost in passing from group homomorphism to Lie
algebra homomorphism.
Whether in fact f can be lifted to a group homomorphism F in this way is a
much more difficult question, which we shall consider in Chapter 7.
Proposition 6.4 Suppose
F :GH
is a continuous homomorphism of linear groups. Then
K = ker F
is a linear group; and
LK = ker(LF ).
Proof I Suppose G is a linear group, ie a closed subgroup of GL(n, R). Since
K is closed in G, it is also closed in GL(n, R). Thus K GL(n, R) is a linear
group.
Moreover
X LK =
=
=
=

etX K t R
F (etX ) = etf X = I
fX = 0
X ker f.

t R

Corollary 6.2 Suppose

F :GH
is a continuous homomorphism of linear groups. Then
LF injective ker F discrete.
In particular
F injective = LF injective.

424III

65

Proposition 6.5 Suppose

F :GH
is a continuous homomorphism of linear groups; and suppose H is connected.
Then
LF surjective F surjective.
Proof I Suppose first that f = LF is surjective; and suppose T H. Then by
Proposition 3.4,
T = eY1 . . . eYr ,
where Y1 , . . . , Yr LH. Since f is surjective we can find X1 , . . . , Xr LG such
that
Yi = f (Xi ) (i = 1, . . . , r)
Then
T = ef X1 . . . ef Xr
= F (eX1 ) . . . F (eXr )
= F (eX1 . . . eXr ).
Thus T im F ; and so F is surjective.
Now suppose conversely that F is surjective; and suppose Y LH. We must
show that there exists an X LG such that
f X = Y.
This is much more difficult.
Since f is linear, we can suppose Y so small that the line-segment
[0, 1]Y = {tY : 0 t 1}
lies inside the logarithmic zone in LH. For the same reason it suffices to find X
such that
f X = tY
for some non-zero t.
Our proof falls into 2 parts.
1: An enumerability argument. Let
C = F 1 (e[0,1]Y ) = {T G : F T = etY , 0 t 1}.

424III

66

We shall show that I is not isolated in C. If it were then every point of C

would be isolated. For suppose S, T C were close to one another; and
suppose
S = esY , T = etY ,
with s < t. Then
R = S 1 T
would be close to I; and it is in C since F (R) = e(ts)Y .
Thus if I were isolated in C we could find disjoint open subsets of M(n, R)
surrounding each element of C. But C is non-enumerable, since it contains
at least one point corresponding to each t I. Thusalways supposing
I isolated in Cwe would have a non-enumerable family of disjoint open
sets in M(n, R) = RN (where N = n2 ). But that leads to a contradiction.
For each subset will contain a rational point (ie a point with rational coordinates); and the number of rational points is enumerable. We conclude that
I cannot be isolated in C.
2: A logarithmic argument. Since I is not isolated, we can find T C arbitrarily
close to I. In particular we can choose T so that
1. T lies in the logarithmic zone of G, with say
X = log T ;
2. f X lies in the logarithmic zone of LH.
But then
F T = F (eX ) = eX ,
while on the other hand
F T = etY
for some t [0, 1]. Thus

ef X = etY .

Since f X and tY both lie in the logarithmic zone, it follows that

f X = tY,
as required.
J

The Corollary to Proposition 4 and Proposition 5 together give

424III

67

Proposition 6.6 Suppose

F :GH
is a continuous homomorphism of linear groups; and suppose H is connected.
Then
LF bijective ker(F )discrete and F surjective.
Suppose G, H, K are 3 linear groups; and suppose we are given homomorphisms
: G H, : H K.
Recall that the sequence
GHK
is said to be exact if
im = ker .
Proposition 6.7 An exact sequence
GHK
of linear groups yields a corresponding exact sequence
LG LH LK
of Lie algebras.
Proof I This follows at once from Propositions 6.4 and 6.5 above.

J Remark:

To summarise our last 3 results: if H is connected, then

1. LF injective ker F discrete;
2. LF surjective im F open;
3. LF bijective ker F discrete and F surjective.
Examples:
1. As we noted above, the functorial property of the Lie operator allows the
theory to be extended from linear to linearisable groups.
Consider for example the real projective group
PGL(n, R) = GL(n + 1, R)/R ,

424III

68

ie the group of projective transformations

P (T ) : PRn PRn
of n-dimensional real projective space PRn that is, the space of 1-dimensional
subspaces (or rays) in the (n+1)-dimensional vector space Rn+1 . Each nonsingular linear map
T : Rn+1 Rn+1
defines a projective transformation
P (T ) : PRn PRn :
two linear maps defining the same projective transformation if and only if
each is a scalar multiple of the other, ie
P (T ) = P (T 0 ) T 0 = aT.
As it stands, PGL(n, R) is not a linear group. However, we can linearise
it in the following way. Suppose T GL(n, R). Consider the linear map
X 7 T XT 1 : M(n + 1, R) M(n + 1, R).
It is evident that any scalar multiple aT will define the same linear map.
Conversely, suppose T, U GL(n + 1, R) define the same map, ie
T XT 1 = U XU 1

for all X M(n + 1, R).

Let
V = T 1 U.
Then
V XV 1 = X,
ie
V X = XV

X.

It follows that
V = aI,
ie
U = aT.
Thus we have defined an injective homomorphism
: PGL(n, R) GL(N, R),

424III

69

where N = (n + 1)2 , identifying the projective group with a linear group

G GL(N, R).
This identifies PGL(n, R) with a subgroup of GL(N, R). But that is not
quite sufficient for our purposes; we must show that the subgroup is closed.
(The image of an injective homomorphism certainly need not be closed.
Consider for example the homomorphism
R T2 : t 7 (a mod 1, b mod 1)
from the reals to the 2-dimensional torus. The image is a path going round
the torus like a billiard ball round a billiard table. There are 2 possibilities.
If the ratio a : b is rational, the path runs repeatedly over the same closed
subgroup of the torus. In particular, the homomorphism is not injective,
On the other hand, if the ratio a : b is irrational, then the path will never
return to the origin. The homomorphism is injective, and its image passes
arbitrarily close to every point of the torus. Thus the closure of the image is
the whole of the torus. However, it does not pass through every point of the
torus, since eg it will only cut the circle (0, y mod 1) enumerably often.
So the image group in this case is not closed.)
There are several ways of showing that im is closed. For example, we
shall see later that since SL(n + 1, R) is semisimple its image is necessarily
closed.
But perhaps the simplest approach is to identify the subgroup im . To this
end, observe that for each X GL(n + 1, R), the map
T : X 7 T 1 XT : M(n + 1, R) M(n + a, R)
is an automorphism of the algebra M(n+1, R), ie it preserves mutliplication
as well as addition and scalar multiplication. It also preserves the trace.
We shall show that every such automorphism is of this form, ie
im = Aut M(n + 1, R).
Since the property of being a trace-preserving automorphism of M(n +
1, R) can be defined by algebraic equations (albeit an infinity of them) the
automorphism group is closed in GL(N, R), so if we can prove that it is in
fact im we shall have achieved our objective.
Again, there are several ways of proving this. One approach, which might
appeal to the geometrically-minded but which we shall not pursue, starts

424III

610

by noting that we can represent subspaces of the projective space PRn by

projections (idempotents) in M(n+1, R). An automorphism of M(n+1, R)
then defines a transformation
PRn PRn
which sends subspaces to subspaces of the same dimension, in such a way
that all incidence relations are preserved. Now it is a well-known proposition of projective geometry that such a transformation is necessarily projective, ie it can be defined by a linear transformation T GL(n + 1, R). But
one could say that this approach is somewhat masochistic, since we have
thrown away a good deal of information in passing from an automorphism
of M(n+1, R) to the corresponding tranformation of projective space PRn .
The following proof may not be the simplest, but it has the advantage of
being based on ideas from the representation theory of finite groups, studied
in Part 1.
Let F be any finite or compact group having an absolutely simple n + 1dimensional representation; that is, one that remains simple under complexification.
For example, we saw in Part 1 that the symmetric group Sn+2 has such a
representationits natural representation in Rn+2 splitting into 2 absolutely simple parts
=1+
where is the representation in the (n + 1)-dimensional subspace
x1 + + xn+2 = 0.
We can turn this on its head and say that we have a finite subgroup F of
GL(n + 1, R) whose natural representation in Rn is simple.
Now suppose
: M(n + 1, R) M(n + 1, R)
is a trace-preserving automorphism of the algebra M(n + 1, R). (Actually
every automorphism of M(n + 1, R) is trace-preserving. But it is easier for
our purposes to add this condition than to prove it.) We want to show that
is an inner automorphism, ie of the form
= T
for some T GL(n + 1, R).

424III

611

To this end, note that the composition

: F GL(n + 1, R)
yields a second representation of F in Rn+1 .
The 2 representations , haves the same character (since preserves
trace). Hence they are equivalent, ie there exists a T GL(n + 1, R) such
that
(g) = T (g)
for all g F .
It follows that
(X) = T (X)
for all X M(n + 1, R) in the matrix algebra generated by the matrices of
F , ie all X expressible as linear combinations of the elements of F :
X = a1 F1 + + am Fm .
But we saw in Part 1 that every matrix X M(n + 1, R) is of this form, ie
the matrices in an absolutely simple representation span the matrix space.
(For a simple proof of this, consider the representation of the productgroup F G in the space of matrices M(n + 1, R). We know that this
representation is simple. Hence the matrices
(f1 )1 I(f2 ) = f11 f2
span the representation space M(n + 1, R).) Thus
(X) = T (X)
for all X M(n + 1, R), ie = T .
So we have identified PGL(n, R) with the group of trace-preserving automorphisms of M(n + 1, R). Since the property of being an automorphism,
and of preserving the trace, can be defined by polynomial equations (albeit an infinity of them), this automorphism group is a closed subgroup of
GL(N, R), as required.
Hence PGL(n, R) is linearisable, and we can speak of its Lie algebra
pgl(n, R).
By definition,
pgl(n, R) = LG,

424III

612

where G is the linear group above. However, we need not compute LG

explicitly; there is a much simpler way of determining pgl(n, R).
Consider the homomorphism
T 7 P (T ) : SL(n + 1, R) PGL(n, R).
If n is even then P is in fact an isomorphism. For on the one hand
T ker P = T = aI
= an = 1
= a = 1
since a R and n + 1 is odd; while on the other hand if P = P (T ) is a
projective transformation we can suppose that T SL(n + 1, R), since we
can always find a scalar a such that
det aT = an det T = 1.
Thus if n is even,
PGL(n, R) = SL(n + 1, R).
So it is evident in this case that PGL(n, R) is linearisablewe dont need
the long-winded argument aboveand that
pgl(n, R) = (n + 1, R).
In fact this last result still holds if n is odd. For in that case the homomorphism
: SL(n + 1, R) PGL(n, R)
is not bijective. However, it has kernel I; and its image is the subgroup of
PGL(n, R) of index 2, consisting of those projective transformations P (T )
defined by T GL(n + 1, R) with det T > 0. We can summarise this in
the exact sequence
1 C2 SL(n + 1, R) PGL(n, R) C2 1.
But now Proposition 6.7 above tells us that the corresponding Lie algebra
homomorphism
L : sl(n + 1, R) pgl(n, R)
is in fact an isomorphism.
So we find that in all cases
pgl(n, R) = sl(n + 1, R).

424III

613

2. Now consider the Euclidean group E(2), ie the isometry group of the Euclidean plane E 2 . As it stands, E(2) is not a linear group. However, on
choosing coordinates, we can identify E 2 with R2 ; and E(2) can then be
identified with the group of transformations
E : (x, y) 7 (ax + cy + e, bx + dy + f )
where
a c
b d

O(2).

We can extend this transformation E to a projective transformation

P (T ) : PR2 PR2 ,
namely that defined by the matrix

a c e

T = b d f
0 0 1
This defines an injective homomorphism
F : E(2) PGL(2, R),
allowing us to identify E(2) with a closed subgroup of the 2-dimensional
projective group. Since we have already established that the projective
group can be linearised, it follows that E(2) can be also.
Explicitly, we have identified E(2) with the group G of 3 3 matrices
described above. By definition,
e(2) = LG.
To determine LG, we adopt our usual technique. Suppose

p u x

X=
q v y LG.
r w z
Then

1 + tp
tu
tx

1 + tv
ty
I + tX = tq

tr
tw
1 + tz

424III

614

satisfies the conditions on G to the first order. Hence

p = r = v = w = z = 0, q + u = 0,
ie

0 q x

X = q 0 y
0 0 0
Conversely, it is readily verified that if X is of this form then
eX G.
The space LG has basis

0 1 0

L= 1 0 0
0 0 0

0 0 1

M = 0 0 0
0 0 0

By computation,
[L, M ] = N,

[L, N ] = M,

0 0 0

N = 0 0 1
0 0 0

[M, N ] = 0.

Thus
e(2) = hL, M, N : [L, M ] = N, [L, N ] = M, [M, N ] = 0i.

Chapter 7
Representations
Since a representation isfrom one point of viewjust a particular kind of
homomorphism, Lie theory certainly applies. But there is one small problem: the Lie algebra of a linear group is real, while we are interested almost
exclusively in complex representations. Overcoming this problem brings an
unexpected reward, by disclosing a surprising relation between apparently
unrelated groups. This allows us to extend the representation theory of compact groups to a much wider class of linear groups.

Definition 7.1 A representation of a Lie algebra L over k (where k = R or C) in

the vector space V over k is defined by a bilinear map
L V V,
which we denote by (X, v) Xv, satisfying
[X, Y ]v = X(Y v) Y (Xv).
Remark: Notice that we only consider real representations of real Lie algebras,
or complex representations of complex algebraswe do not mix our scalars. This
might seem puzzling, since:
1. we are primarily interested (as always) in complex representations of linear
groups; but
2. the Lie algebra of a linear group is always real.
The explanation is found in the following Definition and Proposition.
Definition 7.2 Suppose L is a real Lie algebra. We denote by CL the complex
Lie algebra derived from L by extension of the scalars (from R to C).
424III

71

424III

72

Remark: Suppose L has structure constants ckij with respect to the basis e1 , ..., en ,
ie
X
[ei , ej ] =
ckij ek .
k

Then we can take the same basis and structure constants for CL.
Proposition 7.1
1. Each real representation of the linear group G in U
gives rise to a representation L of the corresponding Lie algebra LG in
U , uniquely characterised by the fact that
(eX ) = eLX

X LG.

2. Each complex representation of G in V gives rise to a representation L

of the complexified Lie algebra CLG in V , uniquely characterised by the
fact that
(eX ) = eLX X LG.
In either case, if G is connected then is uniquely determined by L, ie
= = L = L.
Proof I The real and complex representations of G are defined by homomorphisms
G GL(n, R), G GL(n, C).
These in turn give rise to Lie algebra homomorphisms
LG gl(n, R), LG gl(n, C).
The first of these defines the required representation of LG by
(X, u) Xu.
The second needs a little more care.
Note first that gl(n, C) = M(n, C) has a natural structure as a complex Lie
algebra. If we use M(n, C) to denote this algebra then
gl(n, C) = RM(n, C),
where the realification RL of a complex Lie algebra L is defined in the usual
forgetful way.
Recall the following simple (if confusing) result from linear algebra. If U is a
real vector space and V a complex vector space then each real linear map
F : U RV

424III

73

extendsuniquelyto a complex linear map

F : CU V.
Applying this with U = LG and V = M(n, C), the real Lie algebra homomorphism
LG gl(n, C)
yields a complex Lie algebra homomorphism
CLG M(n, C).
This defines the required representation of CLG by
(X, v) Xv.
Finally, if G is connected then each g G is expressible as a product
g = eX1 . . . eXr .
But from the equation defining the relation between and L the action of eX on
U or V is defined by the action of X. Hence the action of LG or CLG completely
determines the action of each g G. J
Remark: By abuse of notation we shall call a representation of CL (where L
is a real Lie algebra) a complex representation of L. With this understanding
Proposition 1 can be summarised as follows:
Each representation of L defines a corresponding representation L of LG
in the same space. Moreover if G is connected then L uniquely determines , ie
= = L = L.

Corollary 7.1 Suppose L and L0 are 2 real Lie algebras. Then an isomorphism
between their complexifications
CL CL0
sets up a 1 1 correspondence between the complex representations of L and L0 .
Example: We shall show that, for each n, the Lie algebras
sl(n, R) = hX M(n, R) : tr X = 0i

424III

74

and
su(n) = hX M(n, C) : X = X, tr X = 0i
have the same complexification:
Csl(n, R) = Csu(n).
Certainly these 2 algebras have the same dimension:
dim sl(2, R) = dim su(n) = n2 1.
Moreover, we can regard both sl(n, R) and su(n) as real subalgebras of the
complex Lie algebra M(n, C); and the injections
sl(n, R) M(n, C), su(n) M(n, C)
define complex Lie algebra homomorphisms
: Csl(n, R) M(n, C), : Csu(n) M(n, C).
It is not obvious a priori that and are injections. However, that will follow if
we can establish that
im = im ;
dimC im = n2 1.
Indeed, this will also prove the desired result
Csl(n, R) = im = Csu(n).
But im is just the complex linear hull of sl(n, R) in M(n, C), ie the subspace
formed by the linear combinations, with complex coefficients, of the elements of
sl(n, R); and similarly im B is the complex hull of su(n).
But it is easy to see that these hulls are both equal to the complex subspace
V = {X M(n, C) : tr X = 0}.
For sl(n, R) this is a consequence of the elementary proposition that the complex solutions of a real linear equation are just the linear combinations, with complex coefficients, of real solutions.
For su(n), the result follows on noting that any element X V can be written
as
X = Y + iZ,

424III

with
Y =

75

X + X
X X
,Z =
2
2i

both in su(n).
In conclusion, since
dimC V = n2 1
we have established that
Csl(n, R) = Csu(n).
As a concrete illustration, consider the case n = 2. We have seen that
sl(n, R) = {X M(2, R) : tr X = 0} = hH, E, F i,
with
H=

1 0
0 1

E=

0 0
1 0

F =

0 1
0 1

while
su(n) = {X M(2, C) : X = X, tr X = 0} = hA, B, Ci,
with
A=

i 0
0 1

B=

0 1
1 0

C=

0 i
i 0

The isomorphism between the complexifications is defined by

A iH, B (E F ), C i(E + F ).
All the basic notions of group representation theory have Lie algebra analogues. These are summarised in
Definition 7.3
1. The representation of L in V is said to be simple if there
are no non-trivial stable subspaces, ie if U is a subspace of V such that
X L, u U = Xu U
then U = 0 or V .
2. The sum + of 2 representations and of L in U and V is the representation in the direct sum U V defined by
X(u + v) = Xu + Xv.

424III

76

3. The product of and is the representation in the tensor product U V

defined by
X(u v) = Xu v + u Xv.
4. The dual of is the representation in V defined by
(X, p) X 0 p.
5. A representation is said to be semisimple if it is expressible as a sum of
simple representations.
Proposition 7.2 Suppose , are representations of the connected linear group
G. Then
1. simple L simple
2. L( + ) = L + L
3. L() = (L)(L)
4. L( ) = (L)
5. semisimple L semisimple

Chapter 8
Simply Connected Linear Groups
To each homomorphism F : G H of linear groups the Lie functor associates a homomorphismf = LF : LG LH of the corresponding Lie
algebras. But which Lie algebra homomorphisms arise in this way? Which
can be lifted to group homomorphisms? This question is of a rather different
nature to those we have been considering. For while Lie theory is a local theory, this is a global question. Indeed, every Lie algebra homomorphism can
be lifted locally. The question is: do these local bits fit together? That depends, as we shall see, on the fundamental group (or first homotopy group)
1 (G) of the linear group G. If this homotopy group is trivialthat is, G
is simply-connected then every Lie algebra homomorphism f : LG LH
can be lifted.

Proposition 8.1 Suppose G and H are linear groups; and suppose G is simply
connected. Then every Lie algebra homomorphism
f : LG LH
can be lifted to a unique group homomorphism
F :GH
such that
LF = f.
Remark: Recall that a topological space X is said to be simply connected if it is
arcwise connected and if in addition every loop in X can be shrunk to a point, ie
every continuous map
u : S1 X

424III

81

424III

82

from the circumference S 1 of the 2-ball B 2 (the circle with its interior) can be
extended to a continuous map
u : B2 X
from the whole ball.
This is equivalent to saying that the first homotopy group (or fundamental
group) of X is trivial:
1 (X) = {e}.
Proof I Since G is connected, each T G is expressible in the form
T = eXr . . . eX1 ,
by Proposition 3.4. If F exists, F (T ) must be given by
F (T ) = F (eXr ) . . . F (eX1 )
= ef Xr . . . ef X1
In particular, the existence of F requires that
eXr . . . eX1 = I = ef Xr . . . ef X1 = I.
Conversely, if this condition is satisfied, then F (T ) is defined unambiguously by
(*); and the map F : G H defined in this way is clearly a homomorphism with
LF = f .
It is sufficient therefore to show that condition (**) is always satisfied. This
we do in 2 stages.
1. First we show that condition (**) is always satisfied locally, ie for sufficiently small X1 , . . . , Xr . This does not require that G be simply-connected,
or even connected. We may say that f always lifts to a local homomorphism, defined on a neighbourhood of I G.
2. Then we show that if G is simply-connected, every local homomorphism
can be extended to the whole of G.
These 2 stages are covered in the 2 lemmas below. But first we see how relations on linear groups, like those in (**) above, can be represented by closed
paths, or loops.
Let us call a path on G of the form
I = [0, 1] G : t 7 etX g

424III

83

an exponential path joining g to eX g. Then the relation

eXr . . . eX1 = I
defines an exponential loop starting at any point g G, with vertices
g0 = g, g1 = eX1 g, g2 = eX2 eX1 g, . . . , gr = eXr . . . eX1 g = g,
each successive pair gi1 , gi of vertices being joined by the exponential path
etXi gi1

(0 t 1).

(If we chose a different starting-point, we would get a congruent loop, ie a

transform P g of the first loop P by a group element g. In effect, we are only
concerned with paths or loops up to congruence.)
Each exponential path
etX g
in G defines an exponential path in H, once we have settled on a starting point h:
etX g 7 etf X h.
More generally, each path in G made up of exponential segmentslet us call it
a piecewise-exponential pathmaps onto a piecewise-exponential path in H,
starting from any given point.
In this context, condition (**) becomes: Every loop in G maps into a loop in
H. Or: if a path in G closes on itself, then so does its image in H.
Similarly, the local version reads: Every sufficiently small loop in G maps into
a loop in H. It is sufficient in this case to consider exponential triangles, made of
3 exponential paths. For the loop
Pr Pr1 . . . P0
can be split into the triangles
P0 Pr Pr1 , P0 Pr1 Pr2 , . . . , P0 P2 P1 .
In algebraic terms, given a relation
eXr . . . eX1 = I,
we set
eYi = eXi eXi1 . . . eX1

1 i r,

with
eYr = eY0 = I;

424III

84

the given relation then splits into the triangular relations

eYi = eXi eYi1 ,
ie
eYi eXi eYi1 = I.
If each of these relations is preserved by F , then the product of their images in H
gives the required relation:
ef Xr . . . ef X0 = (ef Yr eXr eY [r1 ) . . . (ef Y1 eX1 eY0 ) = I.
Lemma 8.1 Suppose G and H are linear groups; and suppose
f : LG LH
is a Lie algebra homomorphism. Then the map
eX 7 ef X
is a local homomorphism, ie there is a constant C > 0 such that if |X|, |Y |, |Z| <
C then
eX eY = eZ = ef X ef Y = ef Z .
Proof of Lemma B This falls into 2 parts:
1. 1. If the triangle is small of size (or side) d then the discrepancy in its
image, ie the extent to which it fails to close, is of order d3 .
2. 2. A reasonably small triangle, eg one lying within the logarithmic zone, can
be divided into n2 triangles, each of (1/n)th the size. Each of these maps
into a near-triangle with discrepancy of order 1/n3 . These sum to give a
total discrepancy of order n2 /n3 = 1/n. Since n can be taken arbitrarily
large the discrepancy must in fact be 0.
1. If
|X|, |Y |, |Z| < d
then
eX eY = eZ = Z = log(eX eY ).
But
eX eY

= (I + X + X 2 /2)(I + Y + Y 2 /2) + O(d3 )

= I + X + Y + XY + X 2 /2 + Y 2 /2 + O(d3 ).

424III

85

Hence
Z = log(eX eY )
= X + Y + XY + X 2 /2 + Y 2 /2 (X + Y )2 /2 + O(d3 )
= X + Y + [X, Y ]/2 + O(d3 ).
Since by hypothesis f preserves the Lie product it follows that
f Z = f X + f Y + [f X, f Y ]/2 + O(d3 ).
Hence (working backwards)
ef X ef Y = ef Z + O(d3 ).
2. We have not yet said exactly what we mean by the discrepancy D(R) of a
relation R on G. It is convenient to define D(R) to be the sup-norm of the
gap in H:
D(R) = |F (R)|0 ,
where
|T |0 = sup
x6=0

|T x|
.
|x|

This has the advantage, for our present purpose, that

|P 1 T P |0 = |T |0 .
This means that if we take a different vertex on a loop as starting-point, or
equivalently, take a conjugate form
eXi eXi1 . . . eX1 eXr . . . eXr+1 = I
of the given relation, the discrepancy remains the same.
It is clear that any triangle within the logarithmic zone can be shrunk to a
point within that zone. In fact, such a deformation can be brought about by a
sequence of elementary deformations, each consisting either of replacing
the path round 2 sides of a small exponential triangle by the 3rd side, or
conversely replacing 1 side by the other 2.
More precisely, if we start with a triangle of side < d then this can be
divided into n2 exponential triangles of side < d/n; and so the original
triangle can be shrunk to a point by at most n2 elementary deformations,
each involving an exponential triangle of side < d/n.

424III

86

Actually, we really need the converse constuction. Starting from the trivial
relation I = I, we can re-construct the original triangle by a sequence of
elementary deformations, each involving a small triangle of side < d/n.
The discrepancy caused by each of these will be of order 1/n3 by (1). This
perhaps calls for clarification. We can assume that each of these deformations involves the last edge of the loop, since the deformation is unaltered if
we change our base-vertex. Suppose then the relation
R = eXr . . . eX1 = I
is deformed into
eY eZ eXr1 . . . eX1 ,
where
eY eZ = eXr .
The new relation can be written
T R = (eY eZ eXr )(eXr . . . eX1 ) = I.
The corresponding gap in H is
F (T R) I = F (R)(F (T ) I) + (F (R) I);
and so
D(T R) |F (R)|0 D(T ) + D(R).
This shows (by induction) both that F (R) is bounded, and that the descrepancy changes at each deformation by an amount of order 1/n3 .
In sum therefore the n2 deformations will cause a change in the discrepancy
of order 1/n, ie an arbitrarily small change. Since the discrepancy was
initially 0 (with the triangle shrunk to a point), it must be 0 finally, ie
eX eY = eZ = ef X ef Y = ef Z .
C

Corollary 8.1 There is an open neighbourhood U of I in G such that every exponential loop in U maps into a loop in H.
Lemma 8.2 Suppose G and H are linear groups; and suppose G is simply connected. Then every local homomorphism
U H,
where U is a neighbourhood of I in G, has a unique extension to the whole of G.

424III

87

Proof of Lemma B Any exponential path

etX g
can be split into arbitrarily small sub-paths, with vertices
g, eX/m g, e2X/m g, . . . , eX g.
It follows that we can always suppose the edges of the various loops we consider
small enough to come within the ambit of Lemma 1.
Any loop can be shrunk to a point, by hypothesis. We can suppose this shrinkage occurs as a sequence of small shrinks,
Pr . . . P0 7 Pr0 . . . P00
say. These steps can in turn be split into a number of sub-steps, each involving a
small rectangular deformation, of the form
0
Pr . . . Pi+1 Pi Pi0 . . . P00 P0 Pr . . . Pi+1 Pi+1
Pi0 . . . P00 P0

Since by the Corollary to Lemma 1 each such deformation leaves the discrepancy
unchanged, and since the discrepancy finally vanishes, it must vanish initially, ie
eX1 . . . eXr = I = ef X1 . . . ef Xr = I.
C

Proposition 8.2 Suppose the linear group G is simply connected. Then every
representation of LG can be lifted uniquely to a representation 0 of G such
that
= L0 .
Proof I If is real then by Proposition 1 the Lie algebra homomorphism
: LG gl(n, R)
can be lifted (uniquely) to a group homomorphism
0 : G GL(n, R).
On the other hand, suppose is complex, ie a complex Lie algebra homomorphism
: CLG M(n, C).
Since LG < CLG, this restricts to a real Lie algebra homomorphism
: LG gl(n, C);
and the result again follows by Proposition 1.

424III

88

Corollary 8.2
1. Suppose L is a complex Lie algebra. If there exists a simply
connected compact Lie group G such that
L = CLG
then every representation of L is semisimple.
2. Suppose G is a linear group. If there exists a simply connected compact
linear group H with the same complexification,
CLG = CLH,
then every representation of G is semisimple.
Proof I
1. Every representation of L arises from a representation of G, by Proposition
2, which we know from Part B is semisimple.
2. Every representation of G arises from a representation of L = CLG, which
by (1) is semisimple.
J

Example: Every representation of the group SL(2, R) is semisimple, since its Lie
algebra sl(2, R) has the same complexification as the Lie algebra su(2) of the
simply connected compact linear group SU(2).

Chapter 9
The Representations of sl(2, R)
As we know, SU(2) and SO(3) share the same Lie algebra; and this algebra
has the same complexification, and so the same representation theory, as that
of SL(2, R). So in studying the Lie theory of any one of these groups we are
in effect studying all three; and we can choose whichever is most convenient
for our purpose. This turns out to be the algebra sl(2, R).

Proposition 9.1 The Lie algebra

sl(2, R) = hH, E, F : [H, E] = 2E, [H, F ] = 2F, [E, F ] = Hi
has just 1 simple representation (over C) of each dimension 1, 2, 3, . . .. If we
denote the representation of dimension 2j + 1 by Dj (for j = 0, 1/2, 1, 3/2, . . .)
then
Dj = hej , ej1 , . . . , ej i,
with
Hek = 2kek ,
Eek = (j k)ek+1 ,
F ek = (j + k)ek1 .
for k = j, j 1, . . . , j, setting ej+1 = ej1 = 0.
Remark: Note that we have already proved the existence and uniqueness of the
Dj in Part II. For
1. The representations of sl(2, R) and su(2) are in 11 correspondence, since
these algebras have the same complexification;
2. The representations of su(2) and SU(2) are in 11 correspondence, since
the group SU(2) is simply-connected;
424III

91

424III

92

3. We already saw in Part II that SU(2) possessed a unique representation Dj

of dimension 2j + 1 for each half integer j = 0, 1/2, 1, . . ..
However, we shall re-establish this result by purely algebraic means.
Proof I Suppose is a simple representation of sl(2, R) in the (finite-dimensional)
complex vector space V .
To fit in with our subsequent nomenclature, we shall term the eigen-values of
H (or rather, of H) the weights of the representation ; and we shall call the
corresponding eigen-vectors weight-vectors. We denote the weight-space formed
by the weight-vectors of weight by
W () = {v V : Hv = v}.
Suppose is a weight of ; say
He = e.
(Note that has at least 1 weight, since a linear transformation over C always
possesses at least 1 eigen-vector.) Since
[H, E] = 2E,
we have
(HE EH)e = 2Ee
= HEe Ee = 2Ee
= H(Ee) = ( + 2)Ee.
Thus either Ee = 0; or else Ee is also a weight-vector, but of weight + 2. In
any case,
e W () = Ee W ( + 2).
Similarly,
(HF F H)e = [H, F ]e = 2F e
= H(F e) = ( 2)F e.
Thus either F e = 0; or else F e is also a weight-vector, but of weight 2. In
any case,
e W () = F e W ( 2).
This is sometimes expressed by saying that E is a raising operator, which raises
the weight by 2; while F is a lowering operator, which lowers the weight by 2.

424III

93

A finite-dimensional representation can only possess a finite number of weights,

since weight-vectors corresponding to distinct weights are necessarily linearly independent. Let be the maximal weight, and let e be a corresponding weightvector. Then
Ee = 0,
since + 2 is not a weight.
Repeated action on e by F will take us 2 rungs at a time down the weightladder, giving us weight vectors
e, F e, F 2 e, . . .
of weights
, 2, 4, . . .
until finally the vectors must vanish (since there are only a finite number of
weights); say
F 2j e 6= 0, F 2j+1 e = 0,
for some half-integer j. Set
v = e, v2 = F e, . . . , v4j = F 2j e;
and let
v = 0 if > or < 4j.
Then
Hv = v ,
F v = v2
for = , 2, . . .
Since F takes us 2 rungs down the weight-ladder, while E takes us 2 rungs
up, EF and F E both leave us at the same level:
v W () = EF v, F Ev W ().
We shall show that these 2 new weight-vectors are in fact the same, up to scalar
multiples, as the one we started from, ie
EF v = a()v , F Ev = b()v .
Notice that each of these results implies the other, since
(EF F E)v = [E, F ]v
= Hv
= v .

424III

94

So eg the second result follows from the first, with

b() = a() .
At the top of the ladder,
F Ev = 0.
Thus the result holds there, with
b() = 0, ie a() = .
On the other hand, at the bottom of the ladder,
EF v4j = 0.
So the result holds there too, with
a( 4j) = 0.
We establish the general result by induction, working down the ladder. Suppose it proved for
v , v2 , . . . , v+2 .
Then
F Ev = F EF v+2
= F a( + 2)v+2
= a( + 2)v
Thus the result also holds for v , with
b() = a( + 2),
ie a() = a( + 2) + .
This establishes the result, and also gives a() by recursion:
a()
a( 2)
a( 4)

=
=
=
...
a() =

,
a() + 2
a( 2) + 4,
a( + 2) + .

Hence
a() = m + (m 2) + . . . +
= (m + 2)(m + )/4,

424III

95

while
b() = a()
= (m )(m + + 2)/4.
At the bottom of the ladder,
a( 4j) = 0 ie (2j + 1)( 2j) = 0.
Hence
= 2j.
Thus the weights run from +2j down to 2j; and
a() = (j /2 + 1)(j + /2),

b() = (j /2)(j + /2 + 1).

In particular,
Ev = EF e+2
= a( + 2)e+2
= (j /2)(j + /2 + 1)e+2 .
The space
U = hv2j , v2j2 , . . . , v2j i
spanned by the weight-vectors is stable under H, E and F . and is therefore the
whole of V , since the representation was supposed simple. On the other hand U is
simple; for any subspace stable under sl(2, R) must contain a weight-vector. This
must be a scalar multiple of one of the v ; and all the others can then be recovered
by the action of E and F .
This establishes the result; it only remains to prettify the description of Dj ,
by
Indexing the basis weight-vectors by = /2 in place of ;
Renormalising these vectors (now christened e ) so that
F e = (j + )e1 .
It then follows that
(j + + 1)Ee =
=
=
ie Ee =
as stated.

EF e+1
a( + 1)e+1
(j )(j + + 1)e+1
(j )e+1 ,

424III

96

Remark: Note in particular that all the weights of all the representations of sl(2, R)
are integral. The Lie algebra
su(2) = hH, U, V : [H, U ] = V, [H, V ] = U, [U, V ] = Hi
has just 1 simple representation (over C) of each dimension 1, 2, 3, . . . If we denote
the representation of dimension 2j + 1 by Dj (for j = 0, 1, 1, 3, . . .) then
Dj = hej , ej1 , . . . , ej i,
with
He = 2e ,
U e = (j )e+1 + (j + )e1 ,
V e = i(j )e+1 i(j + )e1 .

Proposition 9.2 Every representation of sl(2, R) is semisimple.

Proof I The representations of sl(2, R) are in 11 correspondence with the representations of su(2), since these 2 Lie algebras have the same complexification.
Moreover, the representations of su(2) are in 1-1 correspondence with the
representations of the group SU(2), since SU(2) is simply-connected.
But the representations of SU(2) are all semisimple, since this group is compact. So therefore are the representations of su(2), and hence of sl(2, R). J
Proposition 9.3 For all half-integers j, k,
Dj Dk = Dj+k + Dj+k1 + . . . + D|jk|
Remark: We have already proved this resultor rather the corresponding result
for SU(2)using character theory. But it is instructive to give an algebraic proof.
Proof I We know that Dj Dk is semisimple, and so a sum of Dj s.
Lemma 9.1 Suppose , 0 are representations of sl(2, R) in V, V 0 ; and suppose
e, e0 are weight-vectors in V, V 0 of weights , 0 respectively. Then the tensor
product e e0 is a weight-vector of 0 , with weight + 0 .

424III

97

Proof of Lemma B By the definition of the action of a Lie algebra on a tensor

product,
H(e e0 ) = (He) e0 + e (He0 )
= ( + 0 )(e e0 )
C

Corollary 9.1 If is the set of weights of the representation , and 0 that of 0 ,

then the set of weights of the product representation 0 is the sum-set
+ 0 = { + 0 : , 0 0 }.
Remark: When we speak of sets of weights, it is understood that each weight
appears with a certain multiplicity.
Proof of Proposition 3 (resumed). Let us denote the weight-ladder joining +2j to
2j by
L(j) = {2j, 2j 2, . . . , 2j}.
By the Corollary above, the weights of Dj Dk form the sum-set
L(j) + L(k).
We must express this set as a union of ladders. (Nb sum-sets must not be confused
with unions.)
A given set of weights (with multiplicities) can be expressed in at most 1 way
as a union of ladders, as may be seen by successively removing ladders of maximal
length.
To express the sum-set above as such a union, note that
L(k) = {2k, 2k} L(k 1).
We may assume that k j. Then
L(j) + {2k, 2k} = L(j + k) L(j k).
Hence
L(j) + L(k) = L(j) + ({2k, 2k} L(k 1))
= (L(j) + 2k, 2k) (L(j) + L(k 1))
= L(j + k) L(j k) (L(j) + L(k 1)).
It follows by induction on k that
L(j) + L(k) = L(j + k) L(j + k 1) . . . L(j k).
But this is the weight-set of
Dj+k + Dj+k1 + . . . + Djk .
The result follows.

Chapter 10
The Representations of su(3)
The representation theory of sl(2, R) (and su(2)) outlined above provides both
a model, and a starting-point, for the representation theory of the much larger
class of groups considered below. (We cannot at the moment define this class
preciselybut it includes the whole of the classical repertoire catalogued in
Chapter 1.)
As an illustration of the techniques involved, we take a quick look at the representations of sl(3, R). (This informal account will be properly proofed later.)
Recall that since su(3) has the same complexification as sl(3, R), we are at the
same time studying the representations of this algebra.
The only innovation in passing from sl(2, R) to sl(3, R) is that we must now
consider weights with respect, not just to a single element H, but to a whole
commuting family.
In general, suppose
H = {H1 , H2 , . . . , Hr }
is a family of commuting operators:
[Hi , Hj ] = Hi Hj Hj Hi = 0 i, j.
Then we say that a vector e is a weight-vector (always with respect to the given
family H) if
Hi e = i e for i = 1, . . . , r.
The weight of e is the r-tuple
= (1 , . . . , r )
We denote the space formed by the weight-vectors of weight (together with the
vector 0) by
W (1 , . . . , r ) = {v V : Hi v = i v for i = 1, . . . , r}.
424III

101

424III

102

Note that a commuting family always possesses at least 1 weight-vector. For

certainly H1 possesses an eigen-vector, with eigen-value 1 , say. But then the
eigen-space
U = {v V : H1 v = 1 v}
is non-trivial; and it is moreover stable under H2 , . . . , Hr , since
v U = H1 Hi v = Hi H1 v = 1 Hi v
= Hi v U.
We may therefore suppose, by induction on r, that H2 , . . . , Hr , acting on U , possess a common eigen-vector; and this will then be a common eigen-vector of
H1 , . . . , Hr , ie a weight-vector of the family H.
Example: Let us take an informal look at the representation theory of sl(3, R).
Although this algebra is 8-dimensional, it is convenient to work with the following
9-member spanning set rather than a basis, so that we can preserve the symmetry
between the 3 matrix coordinates:

0 0 0

H= 0 1 0

0 0 1

0 0 0

A= 0 0 1

0 0 0
0 0 0

D=
0 0 0
0 1 0

1 0 0

J = 0 0 0

0 0 1

0 0 0

B= 0 0 0

1 0 0
0 0 1

E=
0 0 0
0 0 0

1 0 0

K = 0 1 0

0 0 0

0 1 0

C= 0 0 0

0 0 0
0 0 0

F =
1 0 0
0 0 0

Notice that
H + J + K = 0;
this being the only linear relation between the 9 elements.

424III

103

The Lie products are readily computed:

[H, A] = 2A
[J, A] = A
[K, A] = A

[H, B] = B
[J, B] = 2B
[K, B] = B

[H, C] = C
[J, C] = C
[K, C] = 2C

[H, D] = 2D [H, E] = E
[H, F ] = F
[J, D] = D
[J, E] = 2E [J, F ] = F
[K, D] = D
[K, E] = E
[K, F ] = 2F
[A, D] = H
[A, E] = 0
[A, F ] = 0

[B, D] = 0
[B, E] = J
[B, F ] = 0

[C, D] = 0
[C, E] = 0
[C, F ] = K

[B, C] = D
[E, F ] = A

[C, A] = E
[F, D] = B

[A, B] = F
[G, E] = C.

The 3 elements (H, J, K) form a commuting family (since they are diagonal). All
weights and weight-vectors will be understood to refer to this family.
Notice that if (x, y, z) is such a weight then
H + J + K = 0 = x + y + z = 0.
Thus the weights all lie in the plane section
{(x, y, z) R3 : x + y + z = 0}
of 3-dimensional space.
We shall make considerable use of the natural isomorphism between the following 3 sub-algebras and sl(2, R):
L1 = hH, A, Di,

L2 = hJ, B, Ei,

L3 = hK, C, F i sl(2, R).

In particular, suppose (x, y, z) is a weight of (H, J, K). Then x is a weight of

L1 = hH, A, Di. But we saw above that all the weights of sl(2, R) are integral.
It follows that x is an integer. And so, similarly, are y and z. Thus all weights
(x, y, z) are integer triples.
Now suppose is a simple representation of sl(3, R) in V ; and suppose e is a
weight-vector of weight (x, y, z). Then
(HA AH)e = [H, A]e = 2Ae
ie
H(Ae) = (x + 2)Ae.

424III

104

Similarly
J(Ae) = (y 1)Ae,
K(Ae) = (z 1)Ae.
Thus
e W (x, y, z) = Ae W (x + 2, y 1, z 1),
e W (x, y, z) = Be W (x 1, y + 2, z 1)
= Ce W (x 1, y 1, z + 2)
= De W (x 2, y + 1, z + 1)
= Ee W (x + 1, y 2, z + 1)
= F e W (x + 1, y + 1, z 2).
Our argument is merely a more complicated version of that for sl(2, R). Let us
define a maximal weight to be one maximising x. Then if e is a corresponding
weight-vector we must have
Ae = 0, Ee = 0, F e = 0.
Our aim is to construct a stable subspace by acting on e with the operators A, B, C, D, E, F .
In fact the subspace spanned by the weight-vectors
W (j, k) = B j C k e
is stable under sl(3, R), ie for each operator X A, B, C, D, E, F
XW (j, k) = xW (j 0 , k 0 )
for some scalar x, and appropriate j 0 , k 0 .
This may readily be shown by induction on j + k. As an illustration, take
X = A, and suppose j > 0. Then
AW (j, k) = ABW (j 1, k)
= BAW (j 1, k) + [A, B]W (j 1, k),
and the inductive hypothesis may be applied to each term.
We conclude that there is at most 1 simple representation with a given maximal
weight.
A rather different point of view throws an interesting light on the weightdiagram, ie the set of weights, of a representation. Consider the restriction of

424III

105

the representation to L1 . From the representation theory of sl(2, R), the weights
divide into L1 -ladders
(x, y, z), (x 2, y + 1, z + 1), . . . , (x, y + x, z + x).
Since x + y + z = 0, this ladder in fact joins
(x, y, z) to (x, z, y),
and is sent into itself by the reflection
R : (x, y, z) 7 (x, z, y).
in the line
x = 0, y + z = 0.
It follows that the whole weight-diagram is sent into itself by R. Similarly (taking
L2 and L3 in place of L1 ), the diagram is symmetric under the reflections
S : (x, y, z) 7 (z, y, x),

T : (x, y, z) 7 (y, x, z);

and so under the group W formed by the identity 1, R, S, T and the compositions
ST : (x, y, z) 7 (z, x, y) and T S : (x, y, z) 7 (y, z, x).
We note that, starting from any non-zero weight , just 1 of the 6 transforms
g (with g W ) of a given non-zero weight lies in the chamber
{(x, y, z) : x + y + z = 0, x > 0, y < 0, z < 0}.
Our argument shows that, starting from any integer triple in this chamber, we
can construct a weight-diagram having as maximal weight, by taking the 6
transforms of , and filling in all the ladders that arise.
In this way it may be seen that there is indeed a simple representation of
sl(3, R) having a given integer triple (x, y, z), with x > 0, y < 0, z < 0, as
maximal weight.
In conclusion, we note that every representation of sl(3, R) is semisimple. For
the restrictions to L1 , L2 and L3 are semisimple, from the representation theory of
sl(2, R). Moreover, from that theory we see that H, J and K are each diagonalisable.
But a commuting family of matrices, each of which is diagonalisable, are simultaneously diagonalisable. (That follows by much the same argumentrestricting
to the eigenspaces of one of the operatorsused earlier to show that such a family possesses at least 1 weight-vector.) Thus every representation of sl(3, R) is
spanned by its weight-vectors. The semisimplicity of the representation follows
easily from this; for we can successively add simple parts, choosing at each stage
the simple representation corresponding to a maximal remaining weightuntil
finally the sum must embrace the whole representation.

Chapter 11
The Adjoint Representation
Definition 11.1 The adjoint representation ad of the Lie algebra L is the representation in L itself defined by
ad X(Y ) = [X, Y ].
Remark: We should verify that this does indeed define a representation of L. It is
clearly bilinear; so it reduces to verifying that
ad[X, Y ] = ad X ad Y ad Y ad X,
ie
[[X, Y ], Z] = [X, [Y, Z]] [Y, [X, Z]].
for all Z L. But this is just a re-arrangement of Jacobis identity.

424III

111

Chapter 12
Compactness and the Killing Form
As we know, every representation of a compact group carries an invariant
positive-definite quadratic form. When we find that the adjoint representation of a Lie algebra also carries an invariant form, it is natural to askat
least in the case of a compact linear groupwhether these are in fact the
same. If that is so, then we should be able to determine the compactness of
a linear group from its Lie algebra.

Definition 12.1 Suppose is a representation of the Lie algebra L in the vector

space V . Then the trace form of is the quadratic form on L defined by


T (X) = tr (X)2 .
In particular, the Killing form of L is the trace form of the adjoint representation, ie the quadratic form on L defined by


K(X) = tr (ad X)2 .

Theorem 12.1 Suppose G is a connected linear group.
1. If G is compact, then the trace form of every representation of LG is negative (ie negative-definite or negative-indefinite). In particular the Killing
form of LG is negative:
G compact = K 0.
2. If the Killing form on LG is negative-definite, then G is compact:
K < 0 = G compact.

424III

121

424III

122

Proof I 1. Suppose G is compact; and suppose is a representation of G in

V . Then we can find a positive-definite form P (v) on V invariant under Gand
therefore also under LG. By change of coordinates we may suppose that
P (x1 , x2 , . . . , xn ) = x21 + x22 + + x2n .
In other words, (G) O(n). But then
0 (LG) o(n) = {X M(n, R) : X 0 = X},
where 0 = L. Thus

X 2 = X 0 X,

where by abuse of notation we write X for 0 X; and so

K(X) = tr(X 2 ) = tr(X 0 X) 0.
2. It is much more difficult to establish the converse. Suppose K < 0.
Lemma 12.1 If the Killing form K of the Lie algebra L is non-singular, then the
Lie algebra homomorphism
ad : L der(L)
is in fact an isomorphism.
Proof I Firstly, ad is injective. For
X ker(ad) = ad(X) = 0
= K(X) = 0;
and by hypothesis K(X) = 0 only if X = 0.
Secondly, ad is surjective, ie every derivation of L is of the form ad(D), for
some D L.
For suppose d der(L), ie d : L L is a linear map satisfying
d([X, Y ]) = [d(X), Y ] + [X, d(Y )].
Consider the map
X tr ((ad X)d) .
Since K is non-singular, we can find D L such that
tr ((ad X)d) = K(X, D) = tr ((ad X)(ad D)) .

424III
Setting d0 = d ad(D),

123

tr ((ad X)d0 ) = 0

for all X L.
Now
ad(dY ) = d(ad Y ) (ad Y )d = [d, ad Y ]
for any derivation d of L, since
(ad(dY )) X =
=
=
=

[dY, X]
d ([Y, X]) [Y, dX]
d ((ad Y )X) (ad Y )(dX)
[d, ad Y ](X).

Hence, substituting d0 for d,

K(X, d0 Y ) =
=
=
=
=
=

tr ((ad X) ad(d0 Y ))
tr (ad X)[d0 , ad Y ])
tr ((ad X)d0 (ad Y )) tr ((ad X)(ad Y )d0 )
tr ((ad Y ad X ad X ad Y )d0 )
tr (ad[X, Y ]d0 )
0.

We conclude that d0 Y = 0 for all Y ; and so d0 = 0, ie d = ad D.

Corollary 12.1 If the linear group G is connected, and its Killing form K < 0,
then the homomorphism
Ad : G Aut(LG)0
is a covering.
Proof I The Lie algebra homomorphism associated to this group homomorphism
is just
ad : LG L (Aut(LG)) = der(LG).
J

We may assume, after suitable choice of basis, that

K(x1 , . . . , xn ) = x21 x2n .
Thus
Aut(LG) O(n)
is compact. The result will therefore follow if we can show that Z = ker(Ad) is
finite.
First we establish that Z is finitely-generated.

424III

124

Lemma 12.2 Suppose

: G G1
is a covering of a compact group G1 by a connected linear group G. Then we can
find a compact neighbourhood C of I in G such that
(C) = G1 .
Proof I Set
E = {g = eX : X LG, kXk 1}.
Since G is connected, it is generated by the exponentials eX ; thus
G = E E2 E3 . . .
Hence
G1 = (E) (E 2 ) . . .
Since G1 is compact, (E r ) = G1 for some r. We can therefore set C = E r .

Corollary 12.2 With the same assumptions,

G = CZ,
where Z = ker .
Lemma 12.3 Suppose
: G G1
is a covering of a compact group G1 by a connected linear group G. Then ker
is finitely-generated.
Proof I From above, G = CZ, where C is compact and Z = ker . Then C 2 is
compact, and so we can find a finite subset {z1 , . . . , zr } Z such that
C 2 = Cz1 Czr .
It follows that
C r Chz1 , . . . , zr i
for all r. Since G is connected, and the union
G=

C r is an open subgroup of G,

C r = Chz1 , . . . , zr i.

Consequently
Z = (C Z)hz1 , . . . , zr i.
Since C Z is finite (being compact and discrete), we conclude that Z is finitelygenerated. J

424III

125

Lemma 12.4 Suppose

: G G1
is a covering of a compact group G1 by a connected linear group G, with kernel
Z. Then any homomorphism
f :ZR
can be extended to a homomorphism
f : G R.
Proof I With the notation above, let G = CZ, with C compact. Let
u:GR
be a continuous non-negative function with compact support containing C:
u(c) > 0 for all c C.
If we define the function v : G R by
u(g)
zZ u(gz)

v(g) = P
then
X

v(gz) = 1

zZ

for each g G.
Now set
t(g) =

v(gz)f (z).

Note that
t(gz) =

v(gzz 0 )f (z 0 )

v(gz 0 )f (z 0 z 1 )

v(gz 0 )f (z 0 )

z0

z0

z0

v(gz 0 )f (z)

z0

= t(g) f (z).
Let us define the function T : G G R by
T (g, h) = t(gh) t(g).

424III

126

Then
T (gz, h) =
=
=
=

t(ghz) t(gz)
t(gh) f (z) t(g) + f (z)
t(gh) t(g)
T (g, h).

Thus T (g, h) depends only on (g) and h; so we have a function S : G1 G R

such that
T (g, h) = S(g, h).
We can now define the sought-for function by integrating S over G1 :
F (g) =

G1

S(g1 , g)dg1 .

To verify that F has the required properties, we note in the first place that
T (g, z) = t(gz) t(g)
= f (z).
Thus
S(g1 , z) = f (z)
for all g1 G1 , and so
F (z) = f (z).
Secondly,
T (g, hh0 ) = t(ghh0 ) t(g)
= t(ghh0 ) t(gh) + t(gh) t(g)
= T (gh, h0 ) + T (g, h).
Hence
S(g1 , hh0 ) = S (g1 (h), h0 ) + S (g1 , h) .
for all g1 G1 , and so on integration
F (hh0 ) = F (h) + F (h0 ).
J

We have almost reached the end of our marathon! We want to show that
Z = ker(Ad) is finite. Suppose not. We know that Z is finitely-generated. Thus
by the structure theory of finitely-generated abelian groups,
Z = Zr F,

424III

127

where F is finite. So if Z is not finite, it has a factor Z; and the natural injection
Z R extends to a homomorphism
f : Z R.
By our last lemma, this in turn extends to a homomorphism
F : G R.
Since G is connected, the image of this homomorphism is a connected subgroup
of R containing Z, which must be R itself, ie F is surjective.
The corresponding Lie algebra homomorphism
LF : LG LR
is therefore also surjective; so its kernel is an n 1-dimensional ideal of LG. We
can use the non-singular Killing form to construct a complementary 1-dimensional
ideal hJi.
M
LG = ker(LF ) hJi.
But if X ker(LF ),
[J, X] ker(LF ) hJi = {0},
since both are ideals. On the other hand, [J, J] = 0; so
[J, X] = 0
for all X LG; and so ad(J) = 0, and in particular
K(J, X) = 0
for all X, contradicting the non-singularity of K.

Remark: In view of the length and complexity of the proof above, a brief resume
may be in place.
We start with the homomorhism
Ad : G Aut(LG)0 .
We want to show that this is a covering. In Lie algebra terms, we have to
establish that the homomorphism
ad : LG L(Aut LG) = der(LG)
is an isomorphism. This is in fact true for any Lie algebra L with nonsingular Killing form.

424III

128

Injectivity follows at once from the fact that

ker(ad) = ZL,
since the radical of the Killing form of L contains ZL:
X ZL = K(X, Y ) = 0 for all Y.
Surjectivity is more difficult. We have to show that every derivation of L is
of the form ad(X) for some X L.
When L = LG, this is equivalent to showing that every automorphism of
LG in the connected component of the identity is of the form X 7 gXg 1
for some g G. This in turn implies that every automorphism of G in the
connected component of the identity is inner.
However, that is of little help in proving the result. Our proof was somewhat
formal and unmotivated. The result is perhaps best understood in the context
of the cohomology of Lie algebras and their modules (or representations), In
this context, the derivations of L consititute the 1-cocycles of the L-module
L, while the derivations of the form ad(X) form the 1-coboundaries:
Z 1 (L) = der(L),

B 1 (L) = ad(L).

Thus the result reflects the fact that H 1 (L) = 0 for a semisimple Lie algebra
L.
Having established that
Ad : G G1 = Aut(G)0
is a covering, it remains to be shown thatif K < 0this covering is finite.
The fact that K < 0 implies that G1 is compact. That is not sufficient
in itselfa compact group can have an infinite covering, as the covering
R T of the torus shows.
Again, our proof that ker(Ad) is finite was somewhat formal and unmotivated. And again, the result is probably best understood in the context of
cohomologyin this case the cohomology of groups.
For G is a central extension of G1 ; and such extensions correspond to the
second cohomology group H 2 (G1 ). Now if K is non-singular, H 2 (G1 , R) =
0; from which it follows that every essential extension of G1 is finite.
Proposition 12.1 Suppose G is a compact linear group. Then
LG = [LG, LG] ZLG.
Moreover, the Killing form vanishes on ZLG, and is positive-definite on [LG, LG].