Integration on Manifolds and Lie Groups

This note fills in some details for §5 in Chapter I of Bröcker–tom Dieck.

First, we review some basic material on integration on manifolds.

Integration on a smooth manifold Let M be an oriented smooth n-manifold,
and denote by Ccn (M ) the vector space of continuous R n-forms on M with compact
support. Our object here is to define the integral M : Ccn (M ) → R.
Take a smooth atlas on M consisting of orientation-preserving diffeomorphisms
hλ : Uλ ∼
= Uλ0 ⊂ Rn , where the open sets Uλ (for λ ∈ Λ) cover M . Take a partition of
unity,Pconsisting of continuous functions φλ : M → R with support supp(φλ ) ⊂ Uλ , so
that λ φλ = 1. (We do not need φλ to be smooth for current P purposes.)
n
Given ω ∈ Cc (M ), we decompose it as the sum ω = λ φλ ω; since every point
of M has a neighborhood on which only finitely many of the φλ are nonzero and
supp(ω) is covered by finitely many such neighborhoods, this sum is effectively finite:
all except finitely many terms are identically zero. We handle each term separately.
We pull back the n-form φλ ω by h−1 0 n
λ to the open set Uλ in R . The canonical
orientation of Rn orients the standard basis (e1 , e2 , . . . , en ) positively. We write the
coordinates on Rn as x = (x1 , x2 , . . . , xn ) (and recall that the coordinate xi is really
a function, the projection Rn → R to the i-th factor, so that dxi is a 1-form). The
result is
(h−1 ∗
λ ) (φλ ω) = aλ dx1 ∧ dx2 ∧ . . . ∧ dxn ,
where the function aλ is given explicitly as
aλ (x) = (aλ dx1 ∧ . . . ∧ dxn )x (e1 , . . . , en ) = φλ (p) ωp (T (h−1 −1
λ )e1 , . . . , T (hλ )en ) (1)
and p = h−1
λ (x).

Definition 2 We define the integral in terms of multiple integrals on Rn , as

Z XZ Z
ω= · · · aλ (x) dx1 dx2 . . . dxn .
M λ Uλ0

Remark If ωp (v1 , v2 , . . . , vn ) ≥ 0 for one (and hence any) positively oriented basis
of the tangent space Tp (M ), equationR (1) shows that aλ (hλ (p)) ≥ 0. If the condition
holds for all p ∈ M , we deduce that M ω > 0 unless ω is everywhere zero.
Remark If we reverse the orientation of M , the local chart hλ no longer preserves
orientation. We may replace it by r ◦ hλ , where r is the reflection of Rn that reverses
the first coordinate x1 only. Since r∗ dx1 = dr∗ x1 = −dx1 , we find
(h−1
λ
∗ ∗ ∗
◦ r) (φλ ω) = r (aλ dx1 ∧ dx2 ∧ . . . ∧ dxn ) = −(r aλ ) dx1 ∧ dx2 ∧ . . . ∧ dxn .

However,
Z Z Z Z
· · · aλ (r(x)) dx1 dx2 . . . dxn = · · · aλ (x) dx1 dx2 . . . dxn .
R
The net effect is to change the sign of M
ω.

110.620 Lie Grps and Lie Algs JMB File: intlie, Rev. A; 21 Mar 2007; Page 1
2 Integration on Manifolds and Lie Groups

Theorem 3 Given an oriented smooth n-manifold M , Definition 2 yields a canonical

R-linear function Z
: Ccn (M ) −−→ R.
M

Proof R-linearity is obvious. We have to show that the integral is independent of

the choices of atlas and partition of unity.
Take another atlas, consisting of orientation-preserving diffeomorphisms kµ : Vµ ∼
=
0 n
Vµ ⊂ P R ,Pwith associated partition of unity consisting of functions ψµ . We decompose
ω = λ µ φλ ψµ ω, so that φλ ψµ ω has compact support in Uλ ∩ Vµ . We pull back
each term,
(h−1 ∗
λ ) (φλ ψµ ω) = aλ,µ dx1 ∧ dx2 ∧ . . . ∧ dxn .
P
It is clear that aλ = µ aλ,µ . Similarly,
(kµ−1 )∗ (φλ ψµ ω) = bµ,λ dy1 ∧ dy2 ∧ . . . ∧ dyn ,
where we use coordinates (y1 , y2 , . . . , yn ) on this copy of Rn to avoid confusion. It
will be sufficient to prove that
Z Z Z Z
· · · bµ,λ (y) dy1 dy2 . . . dyn = · · · aλ,µ (x) dx1 dx2 . . . dxn .

We have the diffeomorphism θ = kµ ◦ h−1 ∼

λ : hλ (Uλ ∩ Vµ ) = kµ (Uλ ∩ Vµ ) between open
n
sets in R , hence
θ∗ (bµ,λ dy1 ∧ dy2 ∧ . . . ∧ dyn ) = aλ,µ dx1 ∧ dx2 ∧ . . . ∧ dxn , (4)
as both sides are pullbacks from φλ ψµ ω on Uλ ∩ Vµ ⊂ M . The left side is
(θ∗ bµ,λ )(θ∗ dy1 ) ∧ (θ∗ dy2 ) ∧ . . . ∧ (θ∗ dyn ).
Here, (θ∗ bµ,λ )(x) = bµ,λ (θ(x)). In coordinates, θ(x) = (θ1 (x), . . . , θn (x)) ∈ Rn , where
θj = yj ◦ θ. Then we can write
n
X ∂θj
θ∗ dyj = dθ∗ yj = d(yj ◦ θ) = dθj = dxi .
i=1
∂xi

Properties of ∧ show that (θ∗ dy1 ) ∧ (θ∗ dy2 ) ∧ . . . ∧ (θ∗ dyn ) expands to
J dx1 ∧ dx2 ∧ . . . ∧ dxn ,
where J(x) denotes the Jacobian determinant of θ at x. Thus equation (4) reduces
to aλ,µ (x) = bµ,λ (θ(x))J(x).
Meanwhile, the change of variables theorem in Rn shows that
Z Z Z Z
· · · bµ,λ (y) dy1 dy2 . . . dyn = · · · bµ,λ (θ(x))J(x) dx1 dx2 . . . dxn
Z Z
= · · · aλ,µ (x) dx1 dx2 . . . dxn

as required; since θ preserves the orientation of Rn , J > 0 and we do not need to

write |J(x)| here.

110.620 Lie Grps and Lie Algs JMB File: intlie, Rev. A; 21 Mar 2007; Page 2
 Integration on Manifolds and Lie Groups 3

Theorem 5 Suppose θ: P ∼ = M is an orientation-preserving diffeomorphism of

oriented
R ∗ smooth
R n-manifolds, and that ω is a continuous n-form on M . Then
P
θ ω = M ω.
Proof Take a smooth atlas {hλ : Uλ ∼
= Uλ0 ⊂ Rn : λ ∈ Λ} with partition of unity
{φλ : λ ∈ Λ} as above. There is an obvious atlas on P , consisting of open sets
Wλ = θ−1 (Uλ ) with local charts
θ h
Wλ −−→ Uλ −−λ→ Uλ0 ⊂ Rn ,
and the functions φλ ◦ θ = θ∗ φλ form a partition of unity on P . The pullback of
(θ∗ φλ )(θ∗ ω) on Wλ to Uλ0 is
((hλ ◦ θ)−1 )∗ ((θ∗ φλ )(θ∗ ω)) = (h−1 ∗ −1 ∗ ∗ −1 ∗
λ ) (θ ) θ (φλ ω) = (hλ ) (φλ ω).

Our choices make P θ∗ ω and M ω identical, term by term.

R R

Integration on a Lie group From now on, we assume that G is a compact Lie
group of dimension n. We wish R to integrate a continuous real-valued function f over
G. We write the integral as G f (g) dg, where g denotes the variable of integration
and dg should be considered a measure (instead of any kind of differential form) on
G, known as Haar measure. (However, we avoid doing any measure theory.)
Denote by C(G) the vector space of continuous real-valued functions on G. The
integral is required to have the following four properties:
(i) It is an R-linear function C(G) → R;
R
(ii) It is monotone: if f (g) ≥ 0 for all g, then G f (g) dg ≥ 0;
R R (6)
(iii) It is left-invariant: for any h ∈ G, G f (hg) dg = G f (g) dg;
R
(iv) It is normalized : G 1 dg = 1.

Theorem 7 Given a compact Lie group G, there exists a unique integral C(G) → R
that satisfies the axioms (6).
To prove existence, we treat G as an n-manifold and apply the preceding theory
of integration of n-forms. (Uniqueness will be included in Corollary 13.)
First, we need to orient G, which is easy. We choose an orientation oe of the
tangent space Te (G). We use the left translations lh to propagate this to an orientation
on the whole of G: given any basis {v1 , v2 , . . . , vn } of Th (G), for any h ∈ G, we define
oh (v1 , v2 , . . . , vn ) = oe (T (lh−1 )v1 , T (lh−1 )v2 , . . . , T (lh−1 )vn ).
This orientation is left-invariant: for any basis {v1 , v2 , . . . , vn } of Tg (G), we have
ohg (T (lh )v1 , T (lh )v2 , . . . , T (lh )vn ) = og (v1 , v2 , . . . , vn ),
and it is the only left-invariant orientation that extends oe .
Next, we need an n-form on G. We choose a positively oriented alternating n-
form ωe on the vector space Te (G), and similarly propagate it around G by taking
ωh = lh∗ −1 ωe for all h ∈ G, so that ω is left-invariant, lh∗ ω = ω for all h ∈ G.
R R
Definition 8 We define G f (g) dg = G f mω, where the number m is independent
of f and is chosen to make (6)(iv) hold.

110.620 Lie Grps and Lie Algs JMB File: intlie, Rev. A; 21 Mar 2007; Page 3
4 Integration on Manifolds and Lie Groups

The axioms (6) are readily verified. Axiom (i) is clear.

R By the Remark following
Definition 2, since ω is positively oriented everywhere, ω > 0 and m >R 0 can indeed
be chosen to make axiom (iv) hold; further, axiom (ii) holds because f mω ≥ 0 if
f ≥ 0 everywhere. For axiom (iii), we use lh∗ ω = ω and Theorem 5 to express the left
invariance as
Z Z Z Z Z
∗ ∗ ∗ ∗
f ◦ lh mω = lh f mω = lh f mlh ω = lh (f mω) = f mω.

Remark If we reverse the orientation of G, we change ωR to −ω, but m stays the same.
By the second Remark after Definition 2, the integral f (g) dg does not change.
It is easy to construct a right-invariant integral, by introducing an inverse. It
apparently requires a different measure, denoted by δg instead of dg.

Definition 9 We define the right-invariant integral

Z Z
f (g) δg = f (g −1 ) dg. (10)
G G

This clearly is right-invariant,

Z Z Z Z
−1 −1 −1
f (gh) δg = f (h g ) dg = f (g ) dg = f (g) δg.

The other three axioms are obvious.

We shall see later, in Corollary 13, that the two integrals in fact coincide, so that
δg = dg.
To prove this, we need a Fubini-type result, which has nothing to do with Lie
groups, or even manifolds. Let X and Y be compact metric spaces. Suppose we
are given “integrals” IX : C(X) → R and IY : C(Y ) → R that satisfy axioms (i) and
(ii) of (6). We define a partial integration over Y , IbY : C(X ×Y ) → C(X): given
F : X × Y → R, we define IbY F ∈ C(X) by (IbY F )(x) = IY Fx , where Fx ∈ C(Y ) is
given by Fx (y) = F (x, y). We similarly define IbX : C(X ×Y ) → C(Y ).

Lemma 11 For any F ∈ C(X ×Y ), IX IbY F = IY IbX F .

Proof We observe that if we equip the spaces C(X) etc. with the sup norm, the
hypotheses imply that IX and IbX become bounded linear operators with norm vol(X),
and IY and IbY have norm vol(Y ), where we introduce the volumes vol(X) = IX 1 and
vol(Y ) = IY 1.
By compactness, F is uniformly continuous. Given  > 0, choose δ > 0 such that
d(x1 , x2 ) < δ and d(y1 , y2 ) < δ imply that |F (x2 , y2 ) − F (x1 , y1 )| < . If d(x1 , x2 ) < δ,
kFx2 − Fx1 k <  and |(IbY F )(x2 ) − (IbY F )(x1 )| <  vol(X), which shows that IbY F does
indeed lie in C(X).
The result is obvious for a function of the form F (x, y) = φ(x)ψ(y); both sides
reduce by R-linearity to (IX φ)(IY ψ). For a general F , we approximate by sums of
such functions. We cover X by finitely many open sets Uλ of diameter less than δ,
take a partition of unity consisting of functions φλ with support in Uλ , and choose

110.620 Lie Grps and Lie Algs JMB File: intlie, Rev. A; 21 Mar 2007; Page 4
 Integration on Manifolds and Lie Groups 5

points xλ ∈ Uλ ; similarly, we cover Y by open sets Vµ with functions ψµ and points

yµ . We may then write
XX
F (x, y) = φλ (x)ψµ (y)F (x, y).
λ µ

We approximate F by the function

XX
E(x, y) = φλ (x)ψµ (y)F (xλ , yµ );
λ µ

then XX
|F (x, y) − E(x, y)| ≤ φλ (x)ψµ (y) |F (x, y) − F (xλ , yµ )|.
λ µ

For φλ (x) and ψµ (y) to be both nonzero, we must have x ∈ Uλ and y ∈ Vµ , hence
|F (xλ , yµ ) − F (x, y)| < . This yields, for all x and y,
XX
|F (x, y) − E(x, y)| ≤ φλ (x)ψµ (y) = .
λ µ

Thus kF − Ek ≤ , which implies that |IX IbY F − IX IbY E| ≤  vol(X)vol(Y ) and

|IY IbX F − IY IbX E| ≤  vol(X)vol(Y ). But E is an R-linear combination of the func-
tions φλ (x)ψµ (y), for which we have IX IbY E = IY IbX E. Then |IX IbY F − IY IbX F | ≤
2 vol(X)vol(Y ). Since  is arbitrary, we must have IX IbY F = IY IbX F .
We apply this to the case X = Y = G.

Theorem 12 Let IY be any integral on G that satisfies axioms (6), in particular,

left invariance, and IX be any right-invariant integral on G that satisfies the axioms
(with (iii) modified). Then IY = IX .

Corollary 13 There are exactly one left-invariant integral and one right-invariant
integral on G, and they are equal.

Remark Dealing directly with n-forms and orientations is more complicated and
is best avoided. A left-invariant n-form ω on G is not right-invariant in general,
unless G is connected; instead, the n-form ν ∗ ω, where ν: G → G denotes inversion, is
right-invariant, but satisfies ν ∗ ω = ±ω, where the sign can vary from component to
component of G.
Orientations behave the same way.
Proof of Theorem Given f ∈ C(G), we consider F (g, h) = f (gh). We note that
Fg (h) = f (gh) = (f ◦ lg )(h), or Fg = f ◦ lg . Integrating over h, we have
(IbY F )(g) = IY Fg = IY (f ◦ lg ) = IY f,
which is independent of g. Then IX IbY F = IX (IY f.1) = (IY f )(IX 1) = IY f . Similarly,
IY IbX F = IX f .

110.620 Lie Grps and Lie Algs JMB File: intlie, Rev. A; 21 Mar 2007; Page 5
6 Integration on Manifolds and Lie Groups

Theorem 14 Let G be a compact Lie group and f ∈ C(G). Then for any h ∈ G
and any automorphism φ of G, we have
Z Z Z Z Z
−1
f (g) dg = f (hg) dg = f (gh) dg = f (g ) dg = f (φ(g)) dg.

Proof The first two integrals are equal by left invariance. The fourth is our definition
of the right-invariant integral, which we have just seen is equal to the first. So the
left-invariant integral is also right-invariant, which gives the third integral.
R For the fifth, a different strategy is needed. We define the integral If =
f (φ(g)) dg. It clearly satisfies axioms (i), (ii) and (iv). For (iii), we have
Z Z Z
I(f ◦ lh ) = f (hφ(g)) dg = f (φ(φ (h)g)) dg = (f ◦ φ)(φ−1 (h)g) dg
−1

Z
= (f ◦ φ)(g) dg by left invariance, (6)(iii)
Z
= f (φ(g)) dg = If.
R
By uniqueness, If must coincide with the first integral, f (g) dg.

110.620 Lie Grps and Lie Algs JMB File: intlie, Rev. A; 21 Mar 2007; Page 6