Diff Geo

ELEMENTARY RIEMANNIAN GEOMETRY
VISHWAMBHAR PATI
Abstract. In these lectures, we cover some basic material on Riemannian Geometry.
1. Advanced Calculus
1.1. Derivatives. In single variable calculus, one says that a real-valued function f : (c, d) R is dierentiable
at a (c, d) if the limit
lim
h0
f(a +h) f(a)
h
exists. If it does, one calls this limit L the derivative of f at a, and denotes it by L =
df
dx
(a) or f
(a).
In order to generalise this to a real-valued function of several variables, i.e. a function (or map) f : U R,
where U R
n
is an open set, one cannot blindly carry over the above one-variable denition, because one
would have to divide the scalar f(a + h) f(a) by the vector h R
n
. But one may recast the one-variable
denition above in the following form:
lim
h0
_
_
_
_
f(a +h) f(a)
h
L
_
_
_
_
= 0
which is the same as saying:
lim
h0
|f(a +h) f(a) Lh|
|h|
= 0
This last fomulation easily generalises to functions of several variables. One merely has to note that the
scalar L must now be replaced by some map that can be applied to the vector h R
n
to yield a scalar. Also,
since multiplication by a scalar is a linear map from R to R, it is reasonable to require L : R
n
R to be a
linear map. This motivates the following:
Denition 1.1.1 (Dierentiability). Let U R
n
be an open set, and f : U R
m
be a map. One says that
f is dierentiable at a U if there exists a linear map L : R
n
R
m
satisfying:
lim
h0
|f(a +h) f(a) L.h|
|h|
= 0
The linear map L is called the derivative of f at a and denoted by L = Df(a). If f is dierentiable at all
points a U, we say f is dierentiable on U.
Remark 1.1.2. Another reformulation of the denition above is: there exists a linear map L such that for h
small enough,:
f(a +h) = f(a) +L.h +g(h)
where g(h) = o(|h|), viz.
lim
h0
g(h)
|h|
= 0
This formulation says that there is a linear approximation L to the map f, in the sense that the discrepancy
between f(a +h) f(a) and L.h is of order strictly higher than 1 in |h|, for small h.
1
2 VISHWAMBHAR PATI
Exercise 1.1.3.
(i): (Derivative is well-dened) Show that if a linear map L exists, in accordance with the denition 1.1.1
above, then it is unique.
(ii): Show that f is dierentiable at a implies that f is continuous at a.
(iii): (The Chain Rule) If f : U R
m
and g : V R
l
where U R
n
is open, V R
m
is open and contains
f(U), f is dierentiable at a U, and g is dierentiable at f(a) V , then the composite g f : U R
l
is dierentiable at a, and has the derivative Dg(f(a)) Df(a). (Hint: Use the remark 1.1.2).
(iv): Show that f : U R
m
is dierentiable at a U i each of the component functions f
i
(1 i m)
of f is dierentiable at a, and that
Df(a) =
_
_
_
_
_
_
_
_
Df
1
(a)
Df
2
(a)
.
.
.
Df
m
(a)
_
_
_
_
_
_
_
_
(v): If L : R
n
R
m
is a linear map, then its derivative is L at all points of R
n
(the best linear approximation
of a linear map is itself). More generally, for an ane map f : R
n
R
m
, (that is, f(x) = Lx +b, where
L : R
n
R
m
is a linear map, and b R
m
is some xed vector) the derivative Df(a) = L for all a R
n
.
(vi): Show that if f : U R
m
and g : U R
m
are dierentiable at a U, U R
n
an open set, then so is
f + g for xed , R. When m = 1, then show that the map fg (dened by fg(x) = f(x)g(x)) is
also dierentiable at a. If f(x) ,= 0 for all x U, show that x (f(x))
1
is also dierentiable at a. Thus
any polynomial function on R
n
is dierentiable at all points in R
n
, and a rational function (i.e. f = P/Q,
where P, Q are two polynomial functions) is dierentiable at all points of the open set R
n
\ Z(Q), where
Z(Q) is the closed zero-set Q
1
(0) of Q.
(vii): One can identify the vector space M(n, R) of n n real matrices with the euclidean space R
n
2
. Let
GL(n, R) M(n, R) be the open set of all real nonsingular matrices. Show that the bijective map of
GL(n, R) to itself dened by A A
1
is dierentiable at all points in GL(n, R). (Use Cramers formula
for the inverse of a matrix).
1.2. Directional and Partial Derivatives.
Denition 1.2.1 (Directional and partial derivatives). Let U R
n
, and f : U R
m
be dierentiable at a
point a U. Let v R
n
be some vector. The directional derivative of f along v is the quantity:
lim
t0
f(a +tv) f(a)
t
This quantity exists, and is easily shown to be Df(a)v (Exercise).
Let {e
i
}
n
i=1
be the standard basis of R
n
. The directional derivatives along these directions e
i
are called the
partial derivatives of f. If we write f = (f
1
, .., f
m
) in terms of its components, and denote:
f
i
x
j
(a) := Df
i
(a)(e
j
)
then with respect to the standard bases of R
n
and R
m
respectively, the linear map Df(a) can be represented
as the matrix:
Df(a) =
_
_
_
_
_
_
_
_
Df
1
(a)
Df
2
(a)
.
.
.
Df
m
(a)
_
_
_
_
_
_
_
_
=
_
_
_
_
f1
x1
(a)
f1
x2
(a) ...
f1
xn
(a)
. . ... .
. . ... .
fm
x1
(a)
fm
x2
(a) ...
fm
xn
(a)
_
_
_
_
ELEMENTARY RIEMANNIAN GEOMETRY 3
The mn matrix on the right is called the Jacobian matrix of f. Note that a matrix representation for Df(a)
can be done with respect to any bases of R
n
and R
m
respectively, but the Jacobian matrix is the representation
of Df(a) with respect to the standard bases.
Remark 1.2.2 (Caution). Note that for a function f as above, dierentiable or not, it is possible to dene
the partial derivative:
f
x
j
(a) := lim
t0
f(a +te
j
) f(a)
t
if it exists. Further, it is possible for all the partial derivatives
f
xj
(a) of a function f to exist at a point a
without the function f being dierentiable at a. For example, check that the function dened by:
f : R
2
R
(x, y)
xy
(x
2
+y
2
)
1
2
(x, y) ,= (0, 0)
(0, 0) 0
has both partial derivatives (= 0) at (0, 0), but is not dierentiable at (0, 0) (Why?). However, the reader is
urged to show that if all the partial derivatives of f exist and are continuous at a, then f is dierentiable at
a.
1.3. Higher Derivatives, smooth functions and maps. We note that the space of linear maps from R
n
to R
m
is itself a linear vector space, of dimension mn. If we denote this vector space by hom
R
(R
n
, R
m
), then
by choice of bases {e
i
} and {f
j
} of R
n
and R
m
respectively, we can identify this vector space with the vector
space of mn real matrices, which is isomorphic to R
mn
.
Denition 1.3.1 (C
r
-maps). Let U R
n
be open, and f : U R
m
be a map. If f is dierentiable on U
and the map:
Df : U hom
R
(R
n
, R
m
)
a Df(a)
is continuous, then we say f is C
1
. More generally, we inductively dene f to be C
r
if the map Df above is
C
r1
. If a function is C
r
for all r 1, then we say f is C
or smooth. By convention, a C
0
map means a
continuous map.
As an exercise, the reader may check that f is C
r
implies that all the mixed partial derivatives:
D
f
i
:=

i1
i2
..
in
x
i1
1
x
i2
2
..x
in
n
f
i
exist and are continuous for all 1 i m and all multi-indices := (i
1
, .., i
n
) with || :=
j
i
j
r.
1.4. Dieomorphisms.
Denition 1.4.1 (Dieomorphism). Let 1 r . A C
r
map f : U V , where U and V are open subsets
of some euclidean spaces, is called a C
r
-dieomorphism if there exists a C
r
map g : V U which satises
g f = Id
U
, f g = Id
V
. A C
-dieomorphism is called a smooth dieomorphism. A C

0
-dieomorphism
is called a homeomorphism.
Remark 1.4.2. The Jacobian of a C
r
-dieomorphism f : U V is pointwise invertible, as a linear map, for
r 1. For, by the chain rule 1.1.3 (iii), it follows that the linear maps Df(a) and Dg(f(a)) are inverses of
each other for each a U. Thus U and V have to be subsets of euclidean spaces of the same dimension if they
are C
r
-dieomorphic, for r 1. (For r = 0, the result is still true and is much harder to prove. It is called the
Brouwer Domain Invariance Theorem, and most books on algebraic topology will contain a proof. It requires
the use of homology theory.)
4 VISHWAMBHAR PATI
Exercise 1.4.3. Let f be a C
r
-map which is a C
1
- dieomorphism, for r 1. Then f is a C
r
-dieomorphism
(Use (vii) of the exercise 1.1.3). Show by an example that this conclusion is false if C
1
above is replaced by
C
0
. That is, give an example of a C
1
-map which is a homeomorphism but not a C
1
-dieomorphism.
Denition 1.4.4 (Local dieomorphism). Let f : U R
n
be a C
r
map, where U is an open set in R
n
. We
say that f is a C
r
-local dieomorphism at a U if there is a neighbourhood W of a such that f
|W
: W f(W)
is a C
r
- dieomorphism. Note that a map which is a C
r
-local dieomorphism at every point of U will also
have invertible Jacobian at every point of U, for r 1.
Example 1.4.5 (A local dieomorphism which is not a dieomorphism). Consider the map:
f : R
2
R
2
(x, y) (e
x
cos y, e
x
sin y)
This is nothing but the map z e
z
from C to C, written out in long-hand. It is clearly not a dieomorphism
because the points (0, 2n), n Z all map to (1, 0). However, it is not dicult to show that f restricted to
any open strip R (b, b + 2) is a dieomorphism. f maps this open strip dieomorphically to the half-slit
plane, i.e. R
2
minus the half-ray {(cos b, sin b) : 0}.
Example 1.4.6 (Polar Coordinates). There is also the (related) smooth map which is a local dieomorphism:
f : (0, ) R R
2
(r, ) (r cos , r sin )
Over any open strip (0, ) (, + 2), it is a dieomorphism on to a half slit plane. Note that the smooth
dieomorphism t e
t
taking R to (0, ) converts this example to the previous one.
1.5. The Inverse Function Theorem. As we noted above, if a map f : U R
n
is a local dieomorphism
at the point a U, then its derivative Df(a) at a is an invertible linear map. There is a fundamental result
which states that the converse is also true. That is:
Theorem 1.5.1 (Inverse Function Theorem). Let U be an open set in R
n
, and f : U R
n
be a C
r
map with
r 1. If the derivative Df(a) is invertible for a U, then f is a C
r
-local dieomorphism at a.
In order to prove this theorem, we need a lemma:
Lemma 1.5.2 (Banachs Contraction Mapping Theorem). Let B be a complete metric space with a metric d.
Let T : B B be a contraction mapping. That is, there is a constant 0 C < 1 such that:
d(Tx, Ty) Cd(x, y) for all x, y B
Then there is a unique xed point b B for T, viz. T(b) = b.
Proof of the Lemma: Just take an arbitrary point a B, and consider the sequence x
n
:= T
n
(a). Note
that by the hypothesis on T we have
d(x
n+2
, x
n+1
) Cd(x
n+1
, x
n
) ...C
n+1
d(x
1
, x
0
)
which easily implies that {x
n
} is a Cauchy sequence, since C < 1. Because B is complete, {x
n
} converges to
some limit b B. Also, T being a contraction map, is continuous, so:
Tb = T(lim
n
x
n
) = lim
n
Tx
n
= lim
n
x
n+1
= b
proving that b is a xed point. Its uniqueness is clear from the contraction property of T. 2
Proof of the Inverse Function Theorem:
Before we get into the proof, we make a simple remark. If f is an ane map, viz., a map of the form:
f(z) = Az +b
Figure 1. Inverse Function Theorem
where A : R
n
R
n
is linear, and b R
n
is some xed vector, then we would have the formula:
f(w) f(z) = Aw Az = Df(z)(w z)
so that if A = Df(z) were invertible, and we wanted the inverse image w of some y = f(w), we would have
the solution as:
w = z Df(z)
1
(f(z) y) ()
Now let us get back to a general map f as in the hypothesis. We can simplify the setting somewhat. By
translating in the domain and image, and composing with the linear map Df(a)
1
, we can assume, without
loss of generality that a = 0, f(a) = 0, and Df(a) = I, the identity map of R
n
. Let B = B(0, ) be the closed
ball of radius around the origin so that B U. We will suitably choose in the sequel.
For y R
n
, dene the map:
T
y
: B R
n
z z (f(z) y)
This map T
y
is motivated by the solution in (), in the sense that it would yield the inverse image of y in
case the map f were ane. Note that a xed point for this map T
y
would yield the inverse image of y. We
proceed to analyse T
y
so as to be able to apply the Banach contraction mapping theorem 1.5.2 above.
Now, since f is C
r
with r 1, the map z |I Df(z)|
op
(the operator norm of (I Df(z))) is a
continuous function of z. Thus by choosing suitably small, one can guarantee that:
|I Df(z)|
op
<
1
8
for z B (1)
We also choose small enough so that Df(z) is invertible all over B = B(0, ).
Now, we would like to make T
y
a contraction map. Note that for z, w U, z w U
1
=

wU
(U w)
where U w is the (w) translate of U. Clearly U
1
is open.
Dene the map:
g : U U
1
R
n
(w, h) g(w, h) := f(w +h) f(w) Df(w)h
Since f is C
1
, Df is continuous, and hence g is continuous (jointly in the variables w and h). Thus the
subset U
2
U U
1
dened by:
U
2
:= {(w, h) U U
1
: |g(w; h)| <
1
8
|h|}
6 VISHWAMBHAR PATI
is an open set. Since g(w, h) = o(|h|), by (1.1.2), it follows that (w, 0) U
2
for all w U. In particular
(0, 0) U
2
, and since U
2
is open, we can choose small enough so that B(0, /2) B(0, ) U
2
. If z, w
B
= B(0, /2), then z w B = B(0, ), so that we have, for as above:

|g(w, z w)| <
1
8
|z w| for all z, w B
= B(0, /2) (2)

Hence, for as above,
|T
y
(z) T
y
(w)| = |z w (f(z) f(w))|
|(I Df(w))(z w) g(w, z w)|
|I Df(w)|
op
|z w| +|g(w, z w)|
1
8
|z w| +
1
8
|z w|
1
4
|z w| for all z, w B
= B(0, /2) (3)

by using the equations (1) and (2).
Note that setting w = 0 in the above inequality and letting z B
, implies that
|T
y
(z) T
y
(0)| = |T
y
z y|
1
4
|z|

8
Thus if we let y B(0, /4), we have |T
y
(z)| /4 + /8 < /2. Thus for this choice of , and y B(0, /4),
the map T
y
maps B
= B(0, /2) to itself, and is a contraction map on B
.
So there exists a point b B
such that T
y
(b) = b (f(b) y) = b. That is f(b) = y. Since this xed point
is unique in B
, b is the unique inverse image of y in B
. We let V = B(0, /4) and W = f

1
(V ) B(0, /2).
By the foregoing, f : W V is a bijection. We need to show that the inverse map f
1
: V W is smooth.
Let y and y
= y +k V , and z = f
1
(y), w = f
1
(y
) = z +h W. The equation:
f(z) f(w) = z w (T
y
(z) T
y
(w))
implies that
|z w| |T
y
(z) T
y
(w)| |f(z) f(w)| |z w| +|T
y
(z) T
y
(w)|
which, in view of (3) above, implies that for z, w B
3
4
|z w| |f(z) f(w)|
5
4
|z w| (4)
for z, w B
. From (4) it follows that we can assume that h is very small if k is very small. (Note, as an aside,
that the injectivity of f over B
, as also the continuity of f

1
over V , also follow from (4)). The dierentiability
of f at z implies that:
f(z +h) f(z) = k = Df(z)h +g(h)
with g(h) = o(h). This equation can be rewritten, upon applying (Df(z))
1
all over, as:
Df(z)
1
(y +k y) = Df(z)
1
(k) = h g(k) = f
1
(y +k) f
1
(y) g(k)
where g(k) := Df(z)
1
g(h). This implies
f
1
(y +k) f
1
(y) = Df(z)
1
(k) + g(k)
It is clear that g(k) is o(|k|) in the light of the the inequality (4) above, and the fact that Df(z)
1
exists
and has operator norm bounded above for all z B
. This implies that f

1
is dierentiable at y V , with
derivative Df(z)
1
where z = f
1
(y). The continuity of D(f
1
) follows from the continuity of Df, the
continuity of f
1
and Cramers formula. Thus f is a C
1
-dieomorphism. From the exercise 1.4.3 it is a local
C
r
-dieomorphism, and the theorem is proved. 2
Remark 1.5.3.
(i): Note that the inverse function theorem concludes the local behaviour of a function (i.e. all over a
neighbourhood of a point a) from information at a point about its derivative Df(a).
(ii): As the examples 1.4.5 and 1.4.6 show, it is possible for a map f to be a local dieomorphism at each
point, without being a global dieomorphism. On the other hand, the reader can easily check that a map
f : R R which is a C
r
-local dieomorphism at each point (for r 1) is a global C
r
-dieomorphism.
Exercise 1.5.4. For each r 1, give an example of a C
r
-dieomorphism which is not a C
r+1
-dieomorphism.
1.6. The Implicit Function Theorems. There are two more important theorems that deduce the local
behaviour of a map from information about its derivative at a point. Both follow from the inverse function
theorem. First some denitions. As always, U is an open subset of R
n
.
Denition 1.6.1 (Submersion). Let r 1. A C
r
-map f : U R
m
is called a C
r
-submersion at a U if its
derivative Df(a) is surjective. This means that n m, and the Jacobian matrix of f at a is of rank m.
Denition 1.6.2 (Immersion). Let r 1. A C
r
-map f : U R
m
is called a C
r
-immersion at a U if its
derivative Df(a) is injective. This means that n m, and the Jacobian matrix of f at a is of rank n.
Example 1.6.3. An obvious example of a submersion is a linear map A : R
n
R
m
which is surjective. By
a change of basis, one can view this as the projection map onto the rst m coordinates. Similarly, for n m,
the inclusion of a linear m-dimensional subspace V R
n
into R
m
is an immersion. Again, by a change of basis
R
n
, this map can be viewed as the map keeping the rst n coordinates as they are, and inserting zeros in the
last mn coordinates.
The two implicit function theorems in the sequel say that, locally, a submersion (resp. immersion) is
equivalent to the two prototype models of submersions (resp. immersions) of the above example. More
precisely:
Theorem 1.6.4 (Implicit Function Theorem, submersive form). Let U R
n
be an open set, and f : U R
m
be a C
r
-map with r 1. Assume that f is a submersion at a, and f(a) = b R
m
, say. Then there exists a
C
r
-local dieomorphism : W (W), with W a neighbourhood of (b
1
, .., b
m
, 0, .., 0) = (b, 0) in R
n
, (W) a
neighbourhood of a contained in U, and (b, 0) = a, satisfying:
f (x
1
, ..., x
n
) = (x
1
, ..., x
m
) for (x
1
, .., x
n
) W
Proof: By rearranging the coordinates in the domain R
n
(noting that this can be absorbed into without
changing the statement), one can assume that the mm-submatrix:
A =
_
f
i
x
j
(a)
_
1i,jm
is of rank m. That is, it is invertible. Now, consider the map:
F : U R
n
(x
1
, .., x
n
) (f
1
(x
1
, .., x
n
), ..., f
m
(x
1
, .., x
n
), x
m+1
a
m+1
, .., x
n
a
n
)
Clearly, c := F(a) = (b, 0), and the Jacobian of F at a is:
DF(a) =
_
A
0 I
nm
_
8 VISHWAMBHAR PATI
Figure 2. The Implicit Function Theorem (Submersive form)
where I
nm
is the identity matrix of size n m. Since A is invertible, DF(a) is also invertible. Thus, by the
inverse function theorem 1.5.1, there is a neighbourhood (W) of a and a C
r
-dieomorphism : W (W)
where W is a neighbourhood of c, (c) = (b, 0) = a, and such that the composite:
F (x
1
, ...., x
n
) = (x
1
, ..., x
n
) for (x
1
, .., x
n
) W
By reading the rst m entries of the left hand side, we obtain:
f (x
1
, ..., x
n
) = (x
1
, ..., x
m
) for (x
1
, .., x
n
) W
This proves our assertion. 2
It isnt very transparent why the above theorem is called the Implicit Function Theorem. The following
corollary will clarify this.
Corollary 1.6.5. Let U and f be as in the statement of the theorem 1.6.4 above. Let a U, and b = f(a) as
above. Assume that the mm submatrix of Df(a) given by:
_
f
i
x
j
(a)
_
1i,jm
is nonsingular, as above. Then, there exists a neighbourhood V of (a
m+1
, .., a
n
) and a C
r
-map : V R
m
such that:
(i): (a
m+1
, .., a
n
) = (a
1
, .., a
m
).
(ii): f((x
m+1
, .., x
n
), x
m+1
, .., x
n
) = b for all (x
m+1
, .., x
n
) V .
(iii): In a small enough neighbourhood N of a, f
1
(b) N consists precisely of points of the kind
((x
m+1
, .., x
n
), x
m+1
, .., x
n
).
(That is, the implicit equation f(x
1
, .., x
n
) = b can be solved locally as x
i
=
i
(x
m+1
, .., x
m
) for the rst m
variables as C
r
functions of the last n m free variables.)
Proof: From the equation
F (x
1
, ...., x
n
) = (x
1
, ..., x
n
) for (x
1
, .., x
n
) W
in the proof of the theorem 1.6.4 above it follows, upon reading the last n m coordinates and the denition
of F, that:
j
(x
1
, .., x
n
) a
j
= x
j
for m+ 1 j n
In particular,
j
(b
1
, ...b
m
, x
m+1
a
m+1
, .., x
n
a
n
) = x
j
for m+ 1 j n
so that from the conclusion of the theorem above we have:
f(
1
(b, x
m+1
a
m+1
, .., x
n
a
n
), ..
m
(b, x
m+1
a
m+1
, .., x
n
a
n
), x
m+1
, .., x
n
) = b
for (b, x
m+1
a
m+1
, ..., x
n
a
n
) W. If we let V be the neighbourhood of (a
m+1
, .., a
n
) R
nm
which is the
inverse image of W under the map
: (x
m+1
, .., x
n
) (b
1
, .., b
m
, x
m+1
a
m+1
, .., x
n
a
n
)
and dene:
(x
m+1
, .., x
n
) := (
1
((x
m+1
, .., x
n
)), ..,
m
((x
m+1
, .., x
n
))
the conclusions (i) and (ii) follow. The last conclusion (iii) is left as an exercise. 2
Example 1.6.6. Let f : R
2
R be the map dened by f(x, y) = x
2
+ y
2
1. Then the Jacobian of f at
(1, 0) is the matrix:
Df(1, 0) = (2, 0)
which is clearly surjective. In fact
f
x
(1, 0) ,= 0 is invertible. By the conclusion of the previous corollary, one
should be able to nd a solution to the implicit equation f(x, y) = 0 with x getting expressed as a function of
y, and y in some neighbourhood of 0. Indeed, if we let V = (1, 1), and x = (y) = (1 y
2
)
1
2
, we have that
(0) = 1, and the corollary above is veried. Note that the neighbourhood V cannot be enlarged any further
for the conclusion to hold.
Theorem 1.6.7 (Implicit function theorem, immersive form). Let U R
n
be open, and f : U R
m
be a C
r
map which is an immersion at a U, with f(a) = b. Then, there exists a neighbourhood W of b in R
m
, and a
dieomorphism : W (W) such that (b) = (a, 0), and the composite:
f(x
1
, .., x
n
) = (x
1
, ..., x
n
, 0, .., 0) for all (x
1
, .., x
n
) f
1
(W)
Proof: As in the proof of the theorem 1.6.4 above, one constructs the auxiliary map:
F : U R
mn
R
m
(x
1
, .., x
m
) f(x
1
, .., x
n
) + (0, 0, .., x
n+1
, .., x
m
)
Apply the inverse function theorem to this map in a neighbourhood of (a, 0) U R
mn
. The details are left
as an exercise. 2
1.7. Real analytic functions and mappings.
Denition 1.7.1. Let U R
n
be an open set, and let f : U R be a smooth (C
) map. We say that f is

real analytic in a neighbourhood of a U if there exists a R > 0 such that the innite series:
i10,..,in0
1
i
1
!, .., i
n
!
_

i1
i2
..
in
f
x
i1
1
x
i2
2
..x
in
n
(a)
_
(x
1
a
1
)
i1
...(x
n
a
n
)
in
converges absolutely to f(x
1
, .., x
n
) for |x a| < R. The series above is then called the Taylor series of f
at a. A function which is real analytic at every point of U is called a real analytic or C
-function. A map
f : U R
m
is called a real analytic mapping or C
-mapping if each component function is a real analytic

function.
Example 1.7.2. Consider the function f : R R dened by f(x) = e
1
x
2
. It is easy to verify that this is a C
function, and that all the derivatives of f at x = 0 vanish. Thus the function has identically vanishing Taylor
series at 0, and cannot be analytic at 0. For, that would force f to be identically zero in a neighbourhood of
0, which it is not.
There are the obvious analogues of the inverse and implicit function theorems in the real analytic setting.
We omit these details, which can be found in [?], ?.
10 VISHWAMBHAR PATI
Figure 3. Coordinate changes on a dierentiable manifold
2. Differentiable Manifolds
From now on, we shall be dealing exclusively with C
or C
functions and maps unless otherwise stated.

2.1. Manifolds. We assume that the reader is familiar with several variable calculus, and the denition of a
dierentiable manifold. We briey recall a few concepts to x notation and terminology.
Denition 2.1.1. A smooth or C
dierentiable manifold X of dimension n is a paracompact, second count-

able, Hausdor topological topological space with an open covering {U
i
}
iN
such that there exist homeomor-
phisms
i
: U
i

i
(U
i
) satisfying:
(i):
i
(U
i
) are open subsets of R
n
for all i N.
(ii): For each i, j N, the coordinate change map:
j

1
i
:
i
(U
i
U
j
)
j
(U
i
U
j
)
is smooth (=C
, as a map between open subsets of R

n
.
The pair (
i
, U
i
) is called a coordinate chart, and the collection {(
i
, U
i
)}
iN
is called an atlas. Similarly, if all
the coordinate changes
j
1
i
are real analytic, we call it a real analytic manifold. Clearly every real analytic
manifold is a smooth manifold.
Remark 2.1.2. Because a manifold is locally homeomorphic to Euclidean space, it is locally metrizable.
Because of paracompactness, and the Smirnov Metrization Theorem (see [Mun], p. 260), it follows that a
manifold is metrizable. The second countability then becomes equivalent to having a countable dense subset.
We say that two smooth (resp. real analytic) atlases {(
i
, U
i
)} and {(
j
, V
j
)} are compatible if for each pair
i, j such that U
i
V
j
,= , we have:
j

1
i
:
i
(U
i
V
j
)
j
(U
i
V
j
)
is a smooth (resp. real analytic) map of open subsets of R
n
. One checks easily that compatibility of atlases on
a smooth (resp. real analytic) manifold is an equivalence relation, and an equivalence class of atlases on M is
called a smooth structure (resp. real analytic structure) on M. For all purposes, one can replace a given atlas
with a more convenient compatible atlas without changing anything. For example, given any open covering
{U
i
} of a smooth or real analytic manifold with some given atlas, one can nd another atlas {(
j
, V
j
)} which
is compatible with the original one, and such that the covering {V
j
} is ner than {U
i
}.
We recall some basic examples of dierentiable manifolds.
Example 2.1.3 (Open Subsets of Euclidean Space). Clearly, by taking just one chart U = R
n
, and = Id
R
n,
R
n
becomes a real analytic manifold of dimension n. Similarly, for any open subset U R
n
, letting i : U R
n
be the inclusion map, {(i, U)} is a real analytic atlas for U, making it a real analytic manifold of dimension n.
Similarly, any open subset of a smooth (resp. real analytic) manifold of dimension n is again a smooth (resp.
real analytic) manifold of dimension n in a natural way, with induced smooth (resp. real analytic) structure.
Example 2.1.4 (Manifolds dened by equations). Let U R
n
be an open subset, and let f : U R
m
be a
smooth (resp. real analytic) map, where m n. Let b R
m
be a regular value for f, i.e. that the Jacobian
matrix Df(p) := [
fi
xj
] is of rank m at each point p f
1
(a). By convention, b is a regular value if f
1
(b) = .
Then the inverse image X = f
1
(b), if nonempty, is a smooth (resp. real analytic) manifold of dimension
n m. (That is, the m equations f
i
= b
i
, i = 1, .., m determine a manifold of dimension n m, provided the
Jacobian criterion above is satised). This is a consequence of the implicit function theorem.
Let a f
1
(b) be a = (a
1
, .., a
n
). As noted in that theorem 1.6.5, after relabelling the coordinates, there is
a smooth (resp. real analytic) map : V U of a neighbourhood V of (a
m+1
, .., a
n
) such that
f((x
m+1
, ..., x
n
), x
m+1
, .., x
n
) = (a
1
, .., a
m
)
for (x
m+1
, .., x
n
) in this neighbourhood. Also, in a small neighbourhood of a, by the conclusion (iii) of 1.6.5,
the points of the kind:
((x
m+1
, ..., x
n
), x
m+1
, .., x
n
)
are precisely the points of f
1
(b) in this neighbourhood. Thus the smooth (resp. real analytic) chart for X in
a neighbourhood of a is provided by the map:
(x
m+1
, .., x
n
) ((x
m+1
, ..., x
n
), x
m+1
, .., x
n
)
Example 2.1.5 (Spheres). The sphere
S
n
R
n+1
= {(x
0
, .., x
n
) R
n+1
:
n
i=0
x
2
i
= 1}
is a real analytic manifold of dimension n, by virtue of the example 2.1.4 dened above.
Example 2.1.6 (Some Linear Groups). The set of n n matrices, denoted M(n, R) is just Euclidean space
of dimension n
2
, and thus a real analytic manifold by 2.1.3 above.
(i): (The General Linear Groups) The set GL(n, R) of non-singular n n matrices is an open subset of
M(n, R), and thus a real analytic manifold of dimension n
2
by 2.1.3 above. It is a group under matrix
multiplication, and is called the real general linear group. Similarly the complex general linear group,
consisting of non-singular n n matrices with complex entries, is a real analytic manifold of dimension
2n
2
.
(ii): (The Special Linear Groups) The real special linear group SL(n, R), and the complex special linear
group SL(n, C) are real analytic manifolds of dimension n
2
1 and
2n
2
2 respectively, by applying 2.1.4 above to the real analytic determinant map det : M(n, R) R
and det : M(n, C) C respectively. (Use Cramers rule to expand the determinant in terms of any row,
to compute the derivative of det with respect to any entry a
ij
and check that 1 is a regular value of det.)
(iii): (The Orthogonal and Special Orthogonal Groups) The real orthogonal group is dened as the subset:
O(n, R) := {A M(n, R) : AA
t
= I}
It is easily veried that this is precisely the group of linear transformations of R
n
which leave the standard
Euclidean inner product in R
n
invariant, i.e. A O(n, R) i Ax, Ay = x, y for all x, y R
n
. The
fact that this is a real analytic manifold of dimension
1
2
n(n 1) is seen as follows. Let S(n, R) denote
12 VISHWAMBHAR PATI
the linear subspace of M(n, R) consisting of symmetric n n real matrices, which is Euclidean space
R
1
2
n(n+1)
. Compute the derivative of the real analytic map:
: M(n, R) S(n, R)
A AA
t
at the point A M(n, R) to be D(A)X = AX
t
+XA
t
(by using the linear path A+sX starting at A).
Thus, if A O(n, R), we have D(A)(
1
2
Y A) = Y for any symmetric matrix Y S(n, R), and hence that
D(A) is surjective at each point of O(n, R) =
1
(I), and hence that I S(n, R) is a regular value of
. Thus by 2.1.4 above, O(n, R) is a real analytic manifold of dimension
1
2
n(n 1).
This group O(n, R) has two path-connected components, and which component an orthogonal matrix
A lies in is determined by whether the determinant det A = +1 or 1. (This is not entirely trivial, and
relies on the fact that every orthogonal matrix of determinant +1 can be brought, after an orthogonal
change of basis, into block diagonal form with
n
2
(resp.
n1
2
) 22 blocks if n is even (resp. odd, in which
case the last diagonal entry is 1). Each 2 2 block is of the form:
_
cos sin
sin cos
_
where R, viz. a planar rotation. Using this fact, it is an easy exercise to show that every orthogonal
matrix of determinant +1 can be joined by a smooth path to the identity matrix I.)
The connected component of O(n, R) which contains all orthogonal matrices of determinant +1 is
a normal subgroup of O(n, R) called the special orthogonal group and denoted SO(n, R). It is a real
analytic manifold of dimension
1
2
n(n 1). For notational convenience, we will always denote O(n, R)
(resp. SO(n, R)) as O(n) (resp. SO(n)). We have the obvious inclusions:
SO(n) O(n) GL(n, R)
SO(n) SL(n, R) GL(n, R)
(iv): (The Unitary and Special Unitary Groups) For a complex matrix A M(n, C), dene the adjoint of
A to be A
= A
t
. Dene the unitary group:
U(n) = {A M(n, C) : AA
= I}
It is easily seen that this is precisely the group of (complex) linear transformations of C
n
which leave
the hermitian inner product x, y :=
n
i=1
x
i
y
i
on C
n
invariant. By considering the map A AA
of
M(n, C) into the linear subspace H(n, C) M(n, C) of n n complex hermitian matrices, noting that
H(n, C) is the Euclidean space R
n
2
, and applying considerations similar to the ones in (ii) above, one
checks that U(n) is a real analytic manifold of dimension n
2
.
The group U(n) is path connected for all n (again because a unitary change of basis brings a given
unitary matrix into diagonal form, with each diagonal entry in U(1), i.e. the circle group of complex
numbers of modulus 1). The kernel of the group homomorphism det : U(n) U(1) is a normal subgroup
SU(n), called the special unitary group. It is a real analytic manifold of dimension n
2
1, as is seen
from 2.1.4 again. Clearly U(n), SU(n) are subgroups of SO(2n) respectively, after making the obvious
identication of C
n
with R
2n
, and noting that the determinant of an nn complex matrix considered as
a 2n 2n real matrix is the absolute square of its complex determinant. In fact we have the inclusions:
SU(n) U(n) GL(n, C)
SU(n) SL(n, C) GL(n, C)
SU(n) U(n) SO(2n)
Figure 4. Open Moebius strip
(v): (Products) Let M and N be smooth (resp. real analytic) manifolds of dimensions m and n respectively.
Then M N is a smooth (resp. real analytic) manifold of dimension m + n. For if {(
i
, U
i
)}
iN
and
{(
j
, V
j
)}
jN
are atlases for M and N respectively, then
{(
i
j
, U
i
V
j
)}
i,jN
is an atlas for M N.
(vi): (n-Torus) The n-fold product of spheres S
1
S
1
... S
1
is an analytic manifold of dimension n, called
the n-torus. The 2-torus looks like a bicycle tube and is pictured in Fig.3.
(vii): The Open Moebius strip Let be the map:
: R (1, 1) R
3
(s, t)
_
(2 +t cos
s
2
) cos s, (2 +t cos
s
2
) sin s, t sin
s
2
_
The image X of in R
3
is called the open Moebius strip. It is pictured in Fig. 4. It is a smooth manifold
of dimension 2, because R (1, 1) is a smooth manifold, and is a local homeomorphism. The reader
should verify that the two charts ((0, 2) (1, 1)) and ((, ) (1, 1)) constitute an analytic atlas
for X by computing the coordinate change.
A manifold of dimension 1 is called a curve and a manifold of dimension 2, a surface.
Exercise 2.1.7. Prove that the subset of R
3
given by revolving the circle of radius 1 centred at (2,0) in R
2
about the z-axis is the 2-torus S
1
S
1
. Show that the dening equation of this surface is:
(x
2
+y
2
+z
2
+ 3)
2
16(x
2
+y
2
) = 0
2.2. Projective Spaces and Grassmannians. We need to briey recall some facts about the quotient
topology. For more details, the reader is urged to consult any book on general topology, such as [Mun].
Let X be a topological space, and be an equivalence relation on X. Thus, one has a natural set of
equivalence classes, which is denoted by Y = X/ , and the obvious quotient map p : X Y which maps
each element x X to its equivalence class [x] = p(x). There is a unique natural topology on the set Y which
satises the following two properties:
(i): The map p : X Y is continuous.
14 VISHWAMBHAR PATI
(ii): If f : X Z is any continuous map which preserves the equivalence relation, (i.e. x y f(x) =
f(y)), there exists a unique continuous map g : Y Z which makes the diagram
X
f
Z
p
Y
g
commute.
It is easy enough to dene this topology, which is called the quotient topology. Just declare a set U Y
to be open i its inverse image p
1
(U) is open in X. The fact that this topology satises both the properties
listed above is an easy exercise. We remark in passing that any surjection p from a topological space X to a
set Y denes an equivalence relation on X (i.e. x y i p(x) = p(y)), and under this equivalence relation, the
set Y is just the set X/ . Thus we can again topologize the range set Y with the quotient topology. So when
we say p : X Y is quotient map, it just means that p is surjective and continuous, Y has the quotient
topology with respect to the equivalence relation dened by p.
Exercise 2.2.1. Show that a continuous surjection which is an open map is a quotient map. Similarly a
continuous surjection which is a closed map is also a quotient map. Verify that the map:
p : [0, ) S
1
t e
2it
is a quotient map which is neither a closed map nor an open map.
Exercise 2.2.2. Show that a continuous surjection which is a local homeomorphism at each point is a quotient
map. Hence the map t e
2it
or R to S
1
is a quotient map. Similarly, in the example (vii) of the open Moebius
strip in 2.1.6, the map is a quotient map.
Example 2.2.3 (Real projective space). Let X = S
n
, the n-dimensional sphere. Let be the equivalence
relation dened by x y i x = y. That is, each point is equivalent to itself, and to its antipodal point. The
resulting quotient space is called real projective space of dimension n, and denoted RP(n). Note that this is
just the space of all lines in R
n+1
passing through the origin, because each such line is determined by a unit
vector on it, and the unit vector in turn is determined only upto a sign.
The quotient map p : S
n
RP(n) is an open map, and if we restrict p to any open hemisphere, say of the
kind:
U
+
i
:= {(x
0
, .., x
n
) S
n
: x
i
> 0} or U
i
:= {(x
0
, .., x
n
) S
n
: x
i
< 0}
with i = 0, 1, ..n, then it is a homeomorphism onto its image V
i
= p(U
i
). These open sets then give an atlas
on RP(n). The point p(x
0
, .., x
n
) is denoted as [x
0
: .... : x
n
]. This notation means that for x
i
R with
n
i=0
x
2
i
= 1, the points: [x
0
: ... : x
n
] and [x
0
: ... : x
n
] are the same point in RP(n). The x
i
are called the
homogeneous coordinates of the point, and are indeterminate upto a common sign change of all of them.
To see that the atlas dened above is real analytic, note that open hemispheres U
i
are homeomorphic to
the open n-disc D
n
R
n
via the map:
(x
0
, ..., x
n
) (x
0
, ..., x
i
, ..x
n
)
where the hat denotes omission. Note that for a point [x
0
: .... : x
n
] p(U
+
i
) = p(U
i
) = V
i
, the homogeneous
coordinate x
i
is non-zero. Thus a homeomorphism
i
from V
i
to D
n
, would be well dened once we scale all
the homogeneous coordinates so as to make x
i
> 0. That is:
i
: V
i
D
n
[x
0
: .... : x
n
] (sgn (x
i
))(x
0
, .., x
i
, ., x
n
)
On V
i
V
j
, both homogeneous coordinates x
i
and x
j
are non-zero, and the coordinate change is given by:
1
j
:
j
(V
i
V
j
)
i
(V
i
V
j
)
(y
0
, .., y
j1
, y
j+1
, .., y
n
) (sgn y
i
)(y
0
, ..., y
i
, .., (1
y
2
l
)
1
2
, ..y
n
)
which is clearly an analytic map since sgn y
i
is analytic on
j
(V
i
V
j
) = D
n
\ {y
i
= 0}.
Exercise 2.2.4 (Another description of RP(n)). Let X = R
n+1
\ {0}, and dene an equivalence relation on
X by x y i x = y for some real number ,= 0. Show that the quotient space X/ is RP(n). Again, the
equivalence class of (x
0
, .., x
n
) X is the point in RP(n) denoted by [x
0
: .. : x
n
], with the understanding that
these homogeneous coordinates x
i
of that point are indeterminate upto a common non-zero scalar multiplier.
Exercise 2.2.5. Show that RP(1) is homeomorphic to S
1
. (Hint: consider the map z z
2
of the circle to
itself).
Example 2.2.6 (Grassmannians). The grassmannian is a generalisation of projective space. We take the set
of all k-dimensional R-vector subspaces of R
n
, i.e. the k-planes in R
n
that pass through the origin. This set
is called the grassmannian G
k
(R
n
). We make it a real analytic manifold as follows. Equip R
n
with the usual
euclidean inner product , , and for a xed k-subspace E G
k
(R
n
), let us denote:
U
E
= {F G
k
(R
n
) : F E
= {0}}
We claim that there is a bijection of U
E
onto the vector space hom
R
(E, E
). By the denition of U
E
, F U
E
i the kernel E
of the orthogonal projection

E
: R
n
E intersects F trivially, so F U
E
i
E|F
: F E
(the restriction of
E
to F) is an isomorphism. Dene:
E
: U
E
hom
R
(E, E
)
F
E
|F
(
E|F
)
1
: E E
Its inverse is given by
E
: hom
R
(E, E
) U
E
T {(x +Tx) E E
= R
n
: x E} = graph of T
It is easily checked that
E
and
E
are inverses of each other, and since E U
E
for each E, the sets U
E
constitute a covering of G
k
(R
n
). We can put a topology on each U
E
by pulling back the natural topology on
the k(n k)-dimensional vector space hom
R
(E, E
). That is, V U
E
is declared to be open in U
E
i
E
(V )
is open in hom
R
(E, E
). We need to check that the topologies thus induced on U

E
U
F
from U
E
and U
F
match. This will follow from the fact (proved below) that the coordinate changes are analytic dieomorphisms,
and in particular, homeomorphisms.
This will induce the topology on G
k
(R
n
) coherent with the topologies just obtained on the various U
E
s.
That is, V G
k
(R
n
) is declared to be open i V U
E
is open in U
E
for each E G
k
(R
n
). From the analysis
below, it will follow that for each E, F, the image
E
(U
E
U
F
) is open in hom
R
(E, E
). In particular, this
forces all the U
E
s to be open, and {(
E
, U
E
) : E G
k
(R
n
)} then denes a real analytic atlas.
So we proceed to examine the coordinate changes. Let E be the space spanned by the orthonormal basis
{e
i
}
k
i=1
. Complete this to {e
i
}
n
i=1
, an orthonormal basis of R
n
. Let F be another xed k-subspace in U
E
, and
let {f
i
}
k
i=1
be an orthonormal basis for F. For T hom
R
(E, E
), note that a basis for the k-subspace

E
(T)
is given by {e
i
+Te
i
}
k
i=1
. Such a subspace
E
(T) will belong to U
F
i the orthogonal projection
F
from it to
F is an isomorphism. That is, the projections onto F of the basis vectors (e
i
+Te
i
) of
E
(T), span F. That
is, the k k matrix of inner products:
[(T)
ij
] := [e
i
+Te
i
, f
j
]
is nonsingular. The determinant of this matrix is a polynomial in the entries of T hom
R
(E, E
), so setting
it ,= 0 gives an open subset of
E
(U
E
). This shows that
E
(U
E
U
F
) is open in
E
(U
E
) = hom
R
(E, E
), so
that U
E
U
F
is open in U
E
, with respect to the topology just dened on U
E
.
16 VISHWAMBHAR PATI
For convenience, complete the orthonormal basis {f
i
}
k
i=1
of F to an orthonormal basis {f
i
}
n
i=1
of R
n
. If
some k-dimensional subspace P is in U
E
U
F
, then it is
E
(T) for some T hom
R
(E, E
) and also
F
(S)
for some S hom
R
(F, F
). We need to show that S =

F

1
E
(T) is a real-analytic function of T. Since
{e
i
+Te
i
}
k
i=1
and {f
j
+Sf
j
}
k
j=1
both constitute bases for P, we must have a non-singular k k matrix [(T)
ij
]
satisfying:
(f
j
+Sf
j
) =
k
i=1
(T)
ji
(e
i
+Te
i
) for j = 1, ..., k ()
Taking inner products of both sides with f
m
, for 1 m k, and noting that Sf
j
, f
m
= 0 for m, j
{1, 2, .., k}, we get:
jm
=
k
i=1
(T)
ji
e
i
+Te
i
, f
m
As noted earlier, for T (U

E
U
F
), the matrix [(T)
im
] := [e
i
+Te
i
, f
m
] is nonsingular, so our matrix
[(T)
ij
] = [((T)
1
)
ij
]. We substitute this in () above, and take the inner product of both sides with
f
k+r
, r = 1, .., nk. This yields the matrix entries of S (with respect to the bases {f
j
}
k
j=1
of F and {f
k+r
}
nk
r=1
of F
) as:
S
jr
=
k
i=1
((T)
1
)
ji
e
i
+Te
i
, f
k+r
for 1 i k; 1 r n k
Since
ij
(T) is linear in the entries of T, its inverse is real analytic in T wherever (T) is invertible (i.e. all
over
E
(U
E
U
F
)). This shows that S =
F

1
E
(T) is real analytic in T.
Consequently, G
k
(R
n
) is a real analytic manifold of dimension k(n k).
Exercise 2.2.7. Show that the map O(n, R) G
k
(R
n
) which takes an orthogonal matrix A to the R-span of
its rst k columns is a continuous surjection, and hence that G
k
(R
n
) is compact.
Exercise 2.2.8. Show that the complex projective space:
CP(n) := {L C
n+1
: L a 1-dimensional C subspace of C
n+1
}
is a real analytic manifold of dimension 2n. Similarly, show that the complex grassmannian:
G
k
(C
n
) = {P C
n
: P a k-dimensional C subspace of C
n
}
is a real analytic manifold of dimension 2k(n k). It is a straightforward carrying over to C of the discussion
above. Indeed, one can dene a complex manifold of (complex dimension n) analogously to a smooth manifold,
by requiring the charts to be homeomorphic to open subsets of C
n
, and the coordinate changes to be holo-
morphic. Then the manifolds above turn out to be complex manifolds of (complex) dimension n and k(n k)
respectively.
We shall later see that grassmannians are examples of so called homogeneous spaces associated with the
orthogonal group O(n, R) and its closed subgroup O(k, R) O(n k, R), and so natural real analytic atlases
will result on them. There is yet a third way of doing it, by embedding G
k
(R
n
) as a real analytic submanifold
of some high-dimensional projective space.
3. Smooth mappings and tangent spaces
3.1. Smooth functions and mappings.
Denition 3.1.1 (Smooth and Analytic Mappings). We say that a continuous map
f : M N, where M, N are two smooth (resp. real analytic) manifolds is said to be smooth (resp. analytic)
if for each chart (, V ) of N and each chart (, U) of M, the composite map:
(U f
1
(V ))

1
U f
1
(V )
f
V

(V )
is a smooth (resp. real analytic) map of open sets of euclidean spaces. That this is well dened follows from
the fact that the coordinate changes of M and N are smooth (resp. real analytic). When N = R, we call
f a smooth (resp. real analytic) function. A smooth (resp. analytic) map between manifolds which has a
smooth (resp. analytic) inverse is called a smooth dieomorphism (resp. analytic dieomorphism). A (smooth
or analytic) dieomorphism of a neighbourhood of a point p in a manifold is called a local dieomorphism at
p.
It is easy to see that composites of smooth (resp. real analytic) maps are again smooth (resp. real analytic),
by the Chain Rule. The local charts of a smooth (resp. analytic) manifold are local dieomorphisms, by
denition.
Example 3.1.2. Let X = f
1
(a) be the manifold of dimension n m dened by taking the inverse image of
a regular value a R
m
under a smooth map f : U R
m
, where U is an open subset of R
n
, as described in
the example 2.1.4. Then the restriction to X of any smooth mapping g : U N, N any smooth manifold, is
again smooth. This is because, by the implicit function theorem, for a p X, there is a local dieomorphism
of a neighbourhood V around p which makes the following diagram commute:
X V V

(X V )
j
(V ) (5)
where the map j is the inclusion as the rst n m coordinates. This shows that the inclusion map i : X U
is a smooth map, and so g
|X
= g i is smooth. Similarly for the real analytic case.
An important example of a real analytic map is:
Example 3.1.3 (The exponential map). Let F denote R or C, and M(n, F) denote the vector space of n n
matrices with entries in F, and GL(n, F) the general linear group over F. We dene the exponential map:
exp : M(n, F) GL(n, F)
A exp A :=
i=0
A
i
i!
That this is a real analytic map on all of M(n, F) is easy to check, from the absolute convergence of the series
on the right over all of M(n, F). For an n n matrix A M(n, F), dene |A| = max
i,j
|A
ij
|. Then it is easy
to see that:
|AB| n|A| |B|
from which it follows that
_
_
A
k
_
_
n
k1
|A|
k
, so that:
_
_
_
_
_
k
i=0
A
i
i!
_
_
_
_
_
S
k
where S
k
is the k-th partial sum of the series e
nA
. Since M(n, F) is complete with respect to this norm | |
(which is equivalent to the usual Euclidean norm), it follows that the series converges absolutely over all of
M(n, F). In particular, exp is a real analytic map.
18 VISHWAMBHAR PATI
Exercise 3.1.4 (One parameter subgroups). Show that:
(i):
lim
t0
(exp (tA+tB) exp(tA) exp(tB) I)
t
2
=
1
2
[B, A]
where [B, A] = BAAB is the commutator of B and A.
(ii): det(exp A) = e
trA
(iii): The exponential map is a group homomorphism from the additive group RA to a subgroup of GL(n, F).
This subgroup, often denoted exp tA is called the one parameter subgroup generated by A.
(iv):
lim
t0
exp (tX)Aexp (tX) A
t
= [X, A]
For further details on the exponential map, see [BMST], 4.2.
On a smooth manifold (though not on a real analytic manifold) there is a method of patching up locally
dened smooth functions into globally dened ones. For this one needs smooth partitions of unity. We rst
need a little lemma, whose proof will be deferred till after the construction of partitions of unity below.
Lemma 3.1.5 (Bump Functions). Let U R
n
be an open set, and let W U be a non-empty open set such
that W U. Then there exists a smooth function : R
n
R such that:
(i): (x) 0 for all x R
n
.
(ii): (x) 0 for x , U.
(iii): (x) > 0 on W.
Example 3.1.6 (Partitions of Unity). Let M be any smooth manifold, and U = {U
i
}
i
be any open covering
of M. Then there exists a family of smooth functions {
i
}
i
on M satisfying:
(i):
i
(x) 0 for all x M.
(ii): For each x M, there exists a neighbourhood W of x such that all but nitely many
i
vanish on W.
(iii): For each i , the support of
i
, dened as:
supp
i
:= {y M :
i
(y) ,= 0}
is contained in U
i
.
(iv): For each x M,

i
i
(x) 1 (The sum is nite in view of (ii)).
Such a collection of functions {
i
}
i
is called a partition of unity subordinate to the open covering U =
{U
i
}
i
. These are constructed below.
First notice that if the open covering V = {V
j
}
j
is a renement of the given open covering of U = {U
i
}
i
,
(this means that each V
j
is contained in some U
i(j)
), then a partition of unity {
j
} subordinate to V will yield
a partition of unity {
i
} subordinate to U. For each j , x an i(j) so that V
j
U
i(j)
. Then dene
i
=
{j: i(j)=i}
j
(By denition, the empty sum is taken to be zero.) Then

i
i
(x) =

{j: i(j)=i}
j
(x) =

j
j
(x) 1
for each x M, verifying (iv) above. The other conditions (i), (ii) and (iii) are clear because the support of a
i
is contained in the union of the supports of all the
j
s satisfying i(j) = i.
So we may rene the given covering U into another one such that all the U
i
are coordinate charts, i.e.
dieomorphic to open subsets of R
n
via some dieomorphisms
i
(e.g. just take the common renement
{U
i
W
j
}, where W
j
are coordinate charts). Further renement will preserve this property.
Next, by the paracompactness of M, we may, without loss of generality, further rene this covering U to be
locally nite (i.e., each point x M has a neighbourhood W which meets only nitely many U
i
non-trivially.)
Going to a subcovering will preserve this property. Since subcoverings are renements, earlier properties will
be preserved.
Finally, we may pass to a subcovering which is countable. That is, we may assume that is a countable set,
because M is 2nd countable, and every open covering of M has a countable subcovering. Also by going to a
subcovering, we may assume that that U is an irredundant covering: U
i
,
j=i
U
j
for each i (Inductively
discard all U
i
s which dont satisfy this).
So, without loss of generality, we may assume that U satises:
(a): is countable, say = N
(b): U = {U
i
}
iN
is a locally nite covering of M.
(c): For each i N, U
i
is dieomorphic to an open set in R
n
via smooth dieomorphism
i
.
(d): U
i
,
j=i
U
j
for each i N.
By the condition (iv) above it follows that the non-empty closed set
C
1
= (
j=1
U
j
)
c
is contained in U
1
, and not contained in any other U
j
for j ,= 1. By the fact that our space M is a metric
space, and hence normal, there exists an open set W
1
U
1
such that:
C
1
W
1
W
1
U
1
Again by the corresponding fact for C
1
, W
1
is contained only in U
1
and not contained in any U
j
for j ,= 1.
Further, W
1
(
j=1
U
i
) = M. Suppose we have inductively dened open sets W
i
for 1 i m such that:
(W1): ,= W
i
W
i
U
i
for 1 i m.
(W2): (
m
i=1
W
i
)
_
j=m+1
U
j
_
= M
(W3): W
j
U
i
i j = i
We inductively construct W
m+1
as follows. Consider the closed set
C
m+1
=
_
(
m
i=1
W
i
)
_
j=m+2
U
j
_
c
This set is non-empty because it contains the nonempty set (
j=m+1
U
j
)
c
. Also it is contained in U
m+1
, and
no other U
j
. Thus there must be, by normality, an open subset W
m+1
such that:
C
m+1
W
m+1
W
m+1
U
m+1
It is left to the reader to verify that the inductive properties (W1), (W2) and (W3) are satised by W
m+1
.
Now (W2) above implies that W = {W
i
}
i=1
is again an open covering of M. (W1) and (W3) imply that
each W
i
is non-empty and contained in a unique U
j
, namely U
i
. This implies that W is also a locally nite
open covering of M. It is what is called a shrinking of U.
By the Lemma 3.1.5 above, there exists a C
function
i
on
i
(U
i
) whose support is contained in
i
(U
i
),
and which is everywhere 0 and strictly positive on
i
(W
i
). Dene the function, which we also call
i
, by:
i
: M R
x
i
(
i
(x)) if x U
i
0 if x , U
i
Clearly
i
is C
on all of M, and also

i
(x) > 0 for x W
i
and 0 for x , U
i
. By the fact that the set
V
i
:=
1
i
((0, )) contains W
i
, and that W
i
is contained in a unique U
i
, it follows that each V
i
is contained in
a unique U
i
. Note that {V
i
}
i=1
is also an open covering of M, since W is. From the local niteness of U, and
20 VISHWAMBHAR PATI
that each V
j
is contained in a unique U
j
, it follows that each point x M has a neighbourhood W such that
all but nitely many
i
are 0 on W. Finally, since each x M is in some W
i
, some
i
(x) > 0.
Thus we may dene, for each x M:
i
(x) =

i
(x)
j=1
j
(x)
These functions are easily veried to be the required partition of unity subordinate to U 2.
Proof of Lemma 3.1.5: We rst note that we may assume that the open set U is a bounded subset of R
n
.
This is because there is a smooth dieo : R
n
B(0, 1), and this dieomorphism will carry our open set U
into the bounded open subset (U) of B(0, 1). The open set (W) will have closure (W) contained in (U),
and the required bump function on U will be obtained by composing the corresponding bump function on (U)
with the dieo .
So we assume U is a bounded set, and hence has compact closure in R
n
. Thus W is a compact subset of U.
Let 0 < 2
1
:= d(W, U
c
) (d being euclidean distance), and V U be the subset:
V = {x U : d(x, U
c
) >
1
}
which clearly satises:
W W V V U
Now consider the continuous function:
f : R
n
R
+
x
d(x, V
c
)
d(x, W) +d(x, V
c
)
This function is identically = 1 on W, and vanishes identically outside V . To smoothen out this continuous
function we need to convolve it with approximate identities. We do this as follows.
Since f is continuous and of compact support (contained in U), it is uniformly continuous, so there exists a
2
> 0 such that
sup
xR
n
|f(x y) f(x)| <
1
2
for |y| <
2
()
Let = min(
1
,
2
). Consider the non-negative function:
: R
n
R
+
x A
exp
_

2
2
|x|
2
_
for |x| <
0 for |x|
where A
is a normalising constant such that

_
R
n
dx = 1
This function is compactly supported in B(0, ), and as 0,
becomes more concentrated around the

origin, and its peak value at the origin gets higher. Such a function is called an approximate identity, for
reasons to become clear soon. Now consider the function (convolution):
(x) = f
:=
_
R
n
f(x y)
(y)dy x R
n
Clearly (x) 0 for all x R
n
. Thus (i) of our lemma 3.1.5 is satised.
Note that if x U
c
, then for |y| the point xy will lie outside V , so f(xy) = 0, whereas for |y| > ,
(y) = 0. Thus we have (x) = 0 for all x U

c
. So (ii) of our lemma is satised.
Now we make a sup-norm estimate on . Using the fact that
_

(y)dy = 1, we have for x R

n
:
|(x) f(x)| =
_
R
n
(f(x y) f(x))
(y)dy
_
R
n
|f(x y) f(x)|
(y)dy
=
_
B(0,)
|f(x y) f(x)|
(y)dy
1
2
by using (). Thus
| f|
= sup
xR
n
|(x) f(x)|
1
2
Since f 1 on W, (x)
1
2
for x W. Thus (iii) of our lemma is satised. All we need to do is check
that is smooth. This is done by dierentiating under the integral sign in the denition of , and using the
smoothness and compact support of
. The details are left to the reader. 2

Remark 3.1.7.
(i): Unlike the smooth case, real analytic bump functions do not exist, because the zero-set of a real analytic
function on R
n
cannot have a non-empty interior, unless it is the zero function. For a real-analytic
function of one variable, this is clear, because the zeroes of a non-zero real analytic function are isolated
points (why?). In general, we note that if f is a real-analytic function on R
n
vanishing on an entire
neighbourhood of a point p, then by considering the restriction of the function to each hyperplane passing
through p, we see that each such restriction is 0 (by induction), so f 0.
(ii): Bump functions are also sometimes called cut-o functions.
Corollary 3.1.8 (Bump Functions on Manifolds). Let M be a smooth manifold, and let W, U be open sets
such that W U. Then there exists a smooth function : M [0, 1] such that (x) = 0 for x U
c
and
(x) = 1 for all x W.
Proof: Using 3.1.6, construct a smooth partition of unity {
1
,
2
} subordinate to the open cover {U
1
, U
2
},
where U
1
:= U and U
2
:= M \ W. Then =
1
is the required function. 2
3.2. Tangent space, tangent bundle. Having dened the notion of smooth maps between manifolds, one
needs to make sense of the derivative of such a smooth map. One could, of course, use local charts, and dene
the derivative of a map by composing with these local charts. But it isnt clear that this procedure will lead
to something global. We recall that in the case of R
n
, we looked at the behaviour of f along all directions (i.e.
all the directional derivatives). On a manifold M, we have to dene a linear space at each point x M which
is the totality of all the directions in M at that point x. This is called the tangent space at x, and we need
some technology to dene it.
Denition 3.2.1 (Function Germs). Let M be a smooth (resp. analytic) manifold, and x M a point. We
dene the germ of a smooth (resp. analytic) function at x to be the equivalence class of a pair (U, f) where U
is a neighbourhood of x and f is a smooth (resp. analytic) function f : U R under the following equivalence
relation: (U, f) (V, g) if there exists a neighbourhood W of x with W U V such that f(y) = g(y) for all
y W. The set of all germs at x, with pointwise addition and multiplication dened in the obvious manner,
is clearly an R-algebra, and is denoted O
M,x
, or simply O
x
when M is understood. The constant function
a denes a unique germ called the constant germ. Strictly speaking, we should write a germ as [(U, f)], but
in future, we will often denote a germ just by a letter like f, and suppress the domain of denition of its
representative (U, f) and the box brackets, for notational convenience.
22 VISHWAMBHAR PATI
Remark 3.2.2 (Smooth germs vs. analytic germs). There is a fundamental dierence between smooth and
analytic germs. An analytic germ f, say at 0 R
n
, is uniquely determined by all the Taylor coecients of f
at 0. On the other hand, all the Taylor coecients of the function f(x) = exp(x
2
) vanish at 0 R, but it
does not represent the zero germ in O
R,0
, for it is not identically 0 on any neighbourhood of 0. Indeed, the set
of all smooth germs is very much larger than the set of all real analytic germs.
Exercise 3.2.3. For a smooth manifold M, there is the R-algebra of smooth functions on M, denoted C
(M).
For a xed x M, there is also the natural homomorphism:
: C
(M) O
M,x
f [f]
taking the smooth function f to its germ [f] at x. Show that this map is surjective.
Exercise 3.2.4. For M a smooth (resp. real analytic) manifold, and x M, there is a natural evaluation
homomorphism of R-algebras:
e
x
: O
M,x
R
[f] f(x)
where O
M,x
is the R-algebra of smooth (resp. real analytic) germs at x. Show that e
x
is surjective, and its
kernel m
x
:= ker e
x
is a maximal ideal, indeed the unique maximal ideal, in O
M,x
. (A ring with a unique
maximal ideal, e.g. a ring like O
M,x
, is called a local ring).
Exercise 3.2.5. Let f : M N be a smooth (resp. analytic) map of smooth (resp. analytic) manifolds.
Show that the map:
f
: O
N,f(x)
O
M,x
[g] [g f]
is an R-algebra homomorphism satisfying f
(m
f(x)
) m
x
. (viz. it is a homomorphism of local rings).
It is clear that for any point x in any smooth (or real analytic) n-dimensional manifold M, the algebra O
M,x
of smooth (or real analytic) germs at x is isomorphic to the algebra O
R
n
,0
, by using a smooth (or analytic)
local chart around x. On this last algebra, one can dene the action of a vector X R
n
by means of the
directional derivative along X. More precisely, let (U, f) represent a germ in O
R
n
,0
, and dene:
X(f) := Df(0)X = lim
t0
f(tX) f(0)
t
Then it is readily veried that X : O
R
n
,0
R is R-linear, and that it acts on the product of germs via the
Leibniz formula:
X(fg) = g(0)X(f) +f(0)X(g)
Denition 3.2.6 (Derivations and tangent spaces). An R-valued R-linear map (=linear functional) on an R-
algebra which obeys the Leibniz formula is called an R-derivation of that algebra. It is also clear that all such
R-derivations form an R-vector space. The R-vector space of all derivations of the R-algebra O
M,x
is called
the tangent space of M at x. It is denoted T
x
(M), and elements in it are called tangent vectors.
Exercise 3.2.7. Prove that a derivation kills all constant germs.
Clearly, the R-derivations of O
R
n
,0
form a vector space, denoted T
0
(R
n
), the tangent space to R
n
at 0. From
the above, we have a natural map from R
n
to the vector space T
0
(R
n
), via the directional derivative.
We claim that this map is a vector space isomorphism. In fact, let us dene the derivations

xi
for i = 1, ., n,
called the coordinate partials by the formula:
_

x
i
_
|0
f := lim
t0
f(t e
i
) f(0)
t
=
f
x
i
(0)
(which is exactly e
i
(f) as dened above.) Note that
_

xi
_
|0
x
j
=
ij
, where x
j
are the germs of the coordinate
functions. This shows that these coordinate partials
_
_

xi
_
|0
_
n
i=1
are R-linearly independent. To show that
these coordinate partials span the vector space T
0
(R
n
), one needs the rst order Taylor formula for a smooth
function f.
Lemma 3.2.8 (1st-order Taylor formula). Let f be a smooth (resp. analytic) function in a neighbourhood U
of 0 in R
n
. Then we have:
f(x) = f(0) +
n
i=1
x
i
g
i
(x)
where x = (x
1
, .., x
n
) is any point in a small enough ball around 0 contained in U, and g
i
are smooth (resp.
analytic) functions on this ball, and satisfy g
i
(0) =
f
xi
(0).
Proof: Let f be smooth. Let > 0 be such that B(0, ) U. For a xed x B(0, ), consider the function
of one variable h(t) := f(tx). Then this function is smooth, and by the fundamental theorem of calculus, we
have:
f(x) f(0) = h(1) h(0) =
_
1
0
dh
dt
(t)dt
=
_
1
0
_
n
i=1
f
x
i
(tx)x
i
_
dt
=
n
i=1
x
i
g
i
(x)
where
g
i
(x) =
_
1
0
f
x
i
(tx)dt
Then g
i
(0) is easily seen to be
f
xi
(0), and are clearly smooth (resp. analytic) because f is smooth (resp.
analytic). 2
Corollary 3.2.9 (The tangent space to R
n
). The tangent space T
0
(R
n
) is spanned by the basis of corrdinate
partials, viz.
_
_

x
i
_
|0
_
n
i=1
and is therefore an R vector space of dimension n.
Proof: Let X T
0
(R
n
). Applying the derivation X to the Taylor formula in the lemma above, and using the
Leibniz formula, and that derivations kill all constant germs (exercise 3.2.7), we have
X(f) =
n
i=1
a
i
f
x
i
(0) =
n
i=1
a
i
_

x
i
_
|0
f
24 VISHWAMBHAR PATI
where a
i
= X(x
i
), x
i
being (germs of) the coordinate functions. This shows that the coordinate partials span
the tangent space T
0
(R
n
). They are linearly independent because on appplying them to the corrdinate function
germs x
i
, we have:
_

x
j
_
|0
x
i
=
ij
which proves the corollary. 2
If (, U) is a (smooth or real analytic) chart around a point x M, we can w.l.o.g. assume that (U) is
a neighbourhood of 0 = (x) in R
n
, and use this chart to make O
M,x
isomorphic as an R-algebra to O
R
n
,0
.
Then we make the following:
Denition 3.2.10 (Coordinate partials in a chart). Dene the coordinate partials with respect to the chart
(, U) as:
_

x
i
_
|x
f :=
_

x
i
_
|0
(f
1
) = lim
t0
(f
1
)(t e
i
) (f
1
)(0)
t
By the foregoing, this provides a basis for the R-vector of all derivations on the R-algebra O
M,x
of germs at x,
i.e. the tangent space T
x
(M). It is customary to denote the components of (x) as x
i
for all x U, and these
x
i
are called coordinate functions around x.
Exercise 3.2.11. Check that if (, V ) is another chart around the point x, with corresponding coordinate
functions y
j
, and if we denote the coordinate partials with respect to by
_

yj
_
|x
, then there is the Jacobian
formula :
_

x
i
_
|x
=
n
j=1
y
j
x
i
_

y
j
_
|x
where
yj
xi
is the (ji)-th entry in the Jacobian matrix D(
1
)((x)).
Exercise 3.2.12. In the notation of exercise 3.2.4, prove that the R-vector space T
x
(M) of R derivations of
O
M,x
is isomorphic to the vector space:
hom
R
_
m
x
/m
2
x
, R
_
the isomorphism taking a derivation X to the linear functional f X(f) for f m
x
/m
2
x
. Quite naturally, the
vector space m
x
/m
2
x
is called the cotangent space to M at x, and denoted by T
x
(M).
Denition 3.2.13. Let f : M N be a smooth or analytic map between manifolds. Then there is a natural
linear map called the derivative of f at x, denoted as
Df(x) : T
x
(M) T
f(x)
(N). It obeys the following:
(i): D(Id
M
)(x) = Id
Tx(M)
.
(ii): D(f g)(x) = Df(g(x))Dg(x)
It is dened as (Df(x)(X)) h = X(f
h) where h is any germ in O

N,f(x)
and f
h denotes the germ of h f

at x. Df is sometimes denoted f
. That Df(x)(X) is a derivation easily follows from the fact that X is a

derivation.
Exercise 3.2.14 (Derivative in local coordinates). Verify that the denition above makes sense, independent
of the representatives of the germs chosen. The best way to do this is to give a completely algebraic denition
for Df(x) by using the algebraic denition of the tangent space (in exercise 3.2.12) and the homomorphism f
dened in exercise 3.2.5. Also check that if we use local charts (, U) around x and (, V ) around f(x), then:
Df(x)
_

x
i
_
|x
=
n
j=1
f
j
x
i
((x))
_

y
j
_
|f(x)
with respect to the bases dened by the coordinate partials coming from and , and f
j
denotes the j-th
component of f
1
.
Example 3.2.15. Let M be a manifold dened by the inverse image of the regular value 0 under a smooth
or analytic map f : R
n
R
m
, as in example 2.1.4. Let x M be a point. We claim that the tangent space
T
x
(M) is isomorphic to the vector space:
ker Df(x) : T
x
(R
n
) T
f(x)
(R
m
)
(which is of dimension (nm) by the assumption that 0 is a regular value). Note that by the implicit function
theorem cited in 2.1.4 above, we know that T
x
(M) is of dimension n m. So if we can prove that the kernel
of Df(x) contains T
x
(M), we would have equality of the two, for dimensional reasons. But, if X T
x
(M), it
follows from the denition that for each germ (U, h) at 0 in R
m
, the composite map f
(h) = h f is a germ
in O
R
n
,x
which is constant (= h(0)) on the neighbourhood f
1
(U) M of x in M, and hence represents the
constant germ h(0) in O
M,x
. Thus any derivation in T
x
(M) will kill it, and hence (Df(x)(X))h = X(hf) = 0
for all X T
x
(M). Thus T
x
(M) ker Df(x), and we are done.
3.3. Tangent and other bundles.
Denition 3.3.1 (The Tangent Bundle, vector elds). Let M be a smooth or analytic manifold. We consider
the disjoint union of all the tangent spaces to M, viz.:
TM =
xM
T
x
(M)
We dene the projection map : TM M to be (T
x
(M)) = x. Let (, U) be any chart of M. We note
that the following local triviality condition holds. The diagram:
1
(U)

U R
n
pr
1
U
where
_
n
i=1
a
i
_

x
i
_
|x
_
= (x, a
1
, .., a
n
)
and
_
_

x
_
i|x
_
n
i=1
is the basis of T
x
(M) given by the coordinate partials with respect to as dened above,
commutes. Note that is a bijection, and since U R
n
is homeomorphic to the open set (U) R
n
R
2n
,
we topologise
1
(U) by requiring to be a homeomorphism. We need to check that if (, U) and (, V ) are
two charts, then the topology induced on the overlap
1
(U V ) from
1
(U) by this prescription (using the
chart (, U)) is the same as that induced from
1
(V ) by using the chart (, V ). This follows easily from the
Jacobian coordinate change formula for the coordinate partials of the exercise 3.2.11. Thus one can dene the
topology on TM to be the coherent topology from the family {
1
(U)}, i.e. W TM is open i W
1
(U)
is open in
1
(U) for all charts (, U). By denition then, TM becomes a manifold of dimension 2n, which is
analytic if M is analytic, and smooth if M is smooth. Verify that : TM M is an analytic (resp. smooth)
map. It is called the tangent bundle of M, and is called the bundle projection.
26 VISHWAMBHAR PATI
Denition 3.3.2 (Sections, vector elds). A smooth (resp. analytic) map s : M TM satisfying s = Id
M
is called a smooth (resp. analytic) section or vector eld on M. If U M is an open subset, and s : U
1
(U)
is a smooth (resp. analytic) map which satises s = Id
U
, then we say that s is a section over U or a vector
eld over U. If, for example (, U) is a chart, the coordinate partials

xi
give n linearly independent smooth
(resp. analytic) vector elds over U, called the coordinate elds on U, and every smooth (resp. analytic) vector
eld over U is of the form:
s(x) =
n
i=1
a
i
(x)
_

x
i
_
|x
where a
i
are smooth (resp. analytic) functions on U.
Exercise 3.3.3. Show that the tangent bundle of S
n
R
n+1
is the space:
TS
n
= {(x, v) S
n
R
n
: v, x = 0}
Use the two stereographic charts on S
n
to explicitly check local triviality of this bundle.
One can do various constructions with the tangent bundle. For example, one can dene:
Example 3.3.4 (Cotangent bundle). The cotangent bundle T
M is the union:
xM
T
x
(M)
where T
x
M is the R-dual of T
x
(M). As before, it can be given the structure of a manifold of dimension 2n.
In fact, if (, U) is a coordinate chart for M, then for each x U, a basis for T
x
(M) is given by the coordinate
dierentials {dx
i|x
}
n
i=1
dened by:
dx
i|x
_

x
j
_
|x
=
ij
which is nothing but the dual basis to the basis
_
_

xj
_
|x
_
n
j=1
of T
x
(M) dened in the last subsection. The
coordinate change from a chart (, U) (with coordinate functions x
i
) to another chart (, V ) (with coordinate
functions y
j
) will lead to the change of basis given by:
dx
i|x
=
n
j=1
_
x
i
y
j
(x)
_
dy
j|x
where
_
xi
yj
(x)
_
denotes the ji- th entry of the Jacobian matrix of D(
1
)((x)). The subsequent details
are left to the reader.
Notation : 3.3.5. From now on we will suppress the subscript x from the basis elements dx
i|x
of T
x
(M)
and
_

xj
_
|x
of T
x
(M), and simply denote them as dx
i
and
_

xj
_
respectively, for notational convenience.
Exercise 3.3.6. Write down the local formula for (Df(x))
dy
j
for a smooth map f : M N, analogous to
the formula 3.2.14.
Denition 3.3.7 (Derivative of a smooth map). As noted earlier in 3.2.13, a smooth map f : M N be-
tween smooth manifolds leads in natural way to a map Df(x) : T
x
(M) T
f(x)
(N). Putting these brewise
maps together leads to a map:
Df : TM TN
which is called the derivative of f. That it is a smooth map, with respect to the manifold structures dened
above on TM and TN follows from the formula 3.2.14. Further, it is linear on each bre, and is an example
of what we shall later call a bundle morphism. Likewise, the real analytic case.
3.4. Submersions, immersions, submanifolds.
Denition 3.4.1 (Immersions, submersions). A smooth (resp. analytic) map f : M N is called a smooth
(resp. analytic) immersion if Df(x) : T
x
(M) T
f(x
(N) is injective for each x M. (This implies dim M
dim N). It is called a smooth (resp. analytic) submersion if Df(x) is surjective for each x M (This implies
dim M dim N).
Proposition 3.4.2 (Local description of immersions and submersions). Let M and N be manifolds of dimen-
sions m and n respectively. Let f : M N be smooth map. If f is an immersion, then for each point x M
there exist charts (, U) around x and (, V ) around f(x) such that:
(i): (V ) is an open subset of R
m
, (V ) is an open subset of R
n
, and f(U) V .
(ii): The composite
(U)

1
U
f
V

(V )
is the inclusion map (x
1
, .., x
m
) (x
1
, .., x
m
, 0, ...0).
If f is a submersion, then for each point x M there exist charts (, U) around x and (, V ) around f(x)
such that:
(i): (V ) is an open subset of R
m
, (V ) is an open subset of R
n
, and f(U) V .
(ii): The composite
(U)

1
U
f
V

(V )
is the projection map (x
1
, .., x
m
) (x
1
, .., x
n
).
Likewise, in the real analytic setup.
Proof: Since both the assertions are purely local, we may assume that M and N are open sets in euclidean
space of the corresponding dimensions. In this case they are clear from the immersive and submersive forms
of the implicit function theorem 1.6.4 and 1.6.7. 2
Denition 3.4.3 (Regular Values). As in the case of euclidean space, we can dene a point y N to be a
regular value of a smooth map f : M N of manifolds to be a point such that Df(x) : T
x
(M) T
y
(N) is
surjective for all x f
1
(y). In particular, if f
1
(y) is empty, y is vacuously a regular value. For example a
submersion is a smooth map whose image consists entirely of regular values.
By applying local charts etc., one can again prove the manifold version of the example 2.1.4, viz. if y is
a regular value, and f
1
(y) is non-empty, then it is a smooth manifold of dimension dim M dim N. Not
surprisingly, in this event the tangent space to f
1
(y) at x is just the kernel of Df(x) : T
x
(M) T
y
(N).
Exercise 3.4.4. Let f : M N be a submersion. Show that f is an open map. If M is compact and N is
connected, show that f is surjective.
The contrast between open and closed sets is that A is an open subset of a space X i for each point x A,
there is a neighbourhood U of x in A such that U A is open in U. That is, the property of being open is
local at each point of A. However, the property of being closed is local at each point x X (Exercise: Prove
that C is closed in a topological space X i for each open covering {U
} of X, C U
is closed in U
for each
). If A is a set such that each point x A has a neighbourhood U in X such that U A is closed in U, we
say A is locally closed in X. This is weaker that it being closed. For example, the set (0, 1) 0 is locally closed
in R
2
, but not closed.
In view of the local nature of immersions described in the above proposition, all one can assert about an
immersion is that it is locally a closed map. In fact, immersions can behave quite badly, as is clear from the
following example:
28 VISHWAMBHAR PATI
Figure 5. Inverse images of regular and critical values
Example 3.4.5 (Skew line on the torus). Let be an irrational number. Then the map:
f : R S
1
S
1
t (e
it
, e
it
)
is an immersion, but not a closed map. In fact its image is dense in S
1
S
1
, and not even a locally closed
subset of S
1
S
1
! It is however, locally a closed map on M = R, as is forced by the proposition 3.4.2 above.
There is a nal result about regular values, namely
Theorem 3.4.6 (Sards Theorem). Let f : M N be a smooth map between smooth manifolds. Then the
set of critical values of f is a nowhere dense subset of N. The set of regular values of f is a dense subset of N.
Proof: In fact, the theorem also states that the set of critical values of f, called the critical set of f, is a subset
of measure zero in N. (Measure zero is easy to dene on a manifold, by using a countable locally nite atlas,
and using the notion of (Lebesgue) measure zero in R
n
.) If dim M < dimN, the critical set is precisely f(M)
and it is not dicult to show, using the Lipschitz property of a smooth map, that f(M) is a set of measure
zero in N. The proof of Sards theorem for dim M dim N is more complicated, and may be found in [Hir],
p. 69. 2
3.5. Embeddings and submanifolds.
Denition 3.5.1 (Embeddings). An immersion f : M N which is injective, and such that f : M f(M)
is a homeomorphism, is called an embedding. (Here f(M) has the induced (=subspace) topology from N).
(Caution: Some authors dene an embedding to be an injective immersion, which is more general than the
denition here. For example the immersion of the skew line on the torus of the example 3.4.5 above is a 1-1
immersion, but is not an embedding by our denition).
Denition 3.5.2 (Submanifolds). We say that N is a smooth (resp. real analytic) k-dimensional submanifold
of a smooth (resp. analytic) manifold M if there exists an atlas {(
i
, U
i
)}
iI
of M, compatible with the original
one, such that for each i with NU
i
,= , we have that
i
(NU
i
) is an open subset of R
k
R
n
}. In particular
a k-dimensional submanifold is a smooth (resp. real analytic) manifold of dimension k in its own right, and is
also locally closed in M. )
It is easy to see that if M N is a submanifold, then the inclusion map i : M N is an embedding, and
conversely, the image of an embedding is a submanifold.
For example, in all of the examples of 2.1.6, one can see that each of those manifolds gets dened as a
submanifold of some euclidean space. All the inclusions given at the end of (iii) and (iv) of 2.1.6 above are
submanifold inclusions. Similarly, the inverse image of a regular value (see denition 3.4.3 above) is also a
submanifold. There is a deep result, known as the Whitney Embedding Theorem which says that a smooth
manifold of dimension n can be realised as a smooth submanifold of R
2n
. For a proof of the lighter result that
a smooth compact manifold can be smoothly embedded in R
2n+1
, see [Hir], p. 24.
30 VISHWAMBHAR PATI
4. Vector Bundles
Before proceeding further, it is useful to abstract the structures of the tangent and cotangent bundles, and
introduce the notion of a smooth vector bundle on a manifold M.
4.1. Basic denitions.
Denition 4.1.1 (Vector bundles). (A smooth (resp. analytic) vector bundle of rank k on a smooth (resp.
analytic) manifold M is a smooth (resp. analytic) manifold E of dimension k + dim M and a smooth (resp.
analytic) surjection : E M such that each bre E
x
=
1
(x) is an R-vector space of dimension k. Further,
there exists an open covering {U
i
}
i
of M, and smooth (resp. analytic) maps (called local trivialisations or
bundle charts):
i
:
1
(U
i
) U
i
R
k
making the diagram:
1
(U
i
)
i
U
i
R
k
pr
1
U
i
commute. For each x U
i
, and each i, the restriction
i|Ex
is a vector space isomorphism to {x} R
k
. E is
called the total space of the bundle, and M the base space. is called the bundle projection. Sometimes we
will just call E a vector bundle over M, and denote it simply by E, when the projection and the base space
M are understood.
If there exists such an open covering with just a single element {M}, then the bundle is said to be a trivial
bundle. It is denoted by
k
, if it is of rank k. Note that for an analytic bundle to be trivial, there should exist
an analytic trivialisation all over M, and to be smoothly trivial, a smooth trivialisation all over M.
A vector bundle of rank 1 is called a line bundle
We leave it to the reader to guess what a smooth or holomorphic complex vector bundle is. Again a
holomorphic complex vector bundle can be either smoothly, or holomorphically trivial.
Figure 6. A non-trivial line bundle on S
1
Denition 4.1.2 (Sections of bundles, 1-forms). If : E M is a smooth (resp. analytic) vector bundle, a
smooth (resp. analytic) map s : M E, is called a smooth (resp. analytic) section of the bundle if s = Id
M
.
For U M an open set, a section over U is just a section of the restricted bundle: :
1
(U) U.
A smooth (resp. analytic) section of the cotangent bundle s : M T
M is called a smooth (resp. analytic)

1-form on M. Similarly, we can dene smooth (resp. analytic) 1-forms on an open set U M to be a section
of the cotangent bundle over U. For example, for a chart (, U), the basis element dx
i
dened above in 3.3.4
is a 1-form over U.
Exercise 4.1.3.
(i): Show that a smooth (resp. analytic) vector bundle E M of rank k is trivial i there exist k smooth
(resp. analytic) sections s
i
of this bundle which are everywhere linearly independent. That is, {s
i
(x)}
k
i=1
is a linearly independent set in E
x
for each x M.
(ii): Show that the tangent bundles of S
1
and S
3
are (real analytically) trivial.
(iii): Show that the tangent bundle of any Lie group is trivial. (A Lie group G is a smooth manifold which
is a topological group, and such that the map:
GG G
(x, y) xy
1
is smooth. It turns out that all Lie groups become analytic manifolds, and the map above map always
becomes analytic. All the classical groups listed at the outset are Lie groups. )
(iv): Show that the tangent bundle of S
2n+1
admits a nowhere-vanishing real analytic section, for all n.
Show that the tangent bundle of S
4n+3
has three everwhere linearly independent real analytic sections.
4.2. Constructions with vector bundles.
Example 4.2.1 (Morphism, quotients, direct sums, tensor products). There are obvious notions which carry
over from vector spaces to vector bundles. For example, a smooth (resp. analytic) morphism between two
vector bundles E and F over a manifold M is a smooth (resp. analytic) map : E F such that for each
x M, we have (i) (E
x
) F
x
and (ii) : E
x
F
x
is linear. A vector bundle morphism which has an inverse
that is a vector bundle morphism is called an isomorphism of vector bundles. It is an exercise to check that
this is equivalent to requiring the vector bundle morphism to be an isomorphism on each bre.
A sub-bundle E M of a bundle F M is a bundle with an inclusion E F which is a morphism of
bundles. Given such a pair, it is possible to form the quotient bundle F/E, whose bre at x is the quotient
space F
x
/E
x
. Given vector bundles E and F on M, there is the direct sum EF of (called the Whitney sum of
E and F) whose bre over x M is E
x
F
x
. Similarly, one denes the tensor product bundle E F. Another
construction is the bundle hom
R
(E, F) over M, whose bre at x M is the vector space hom
R
(E
x
, F
x
). The
bundle E
denotes the bundle hom

R
(E,
1
), and is called the dual bundle of E.
Finally, given a vector bundle E, one can form the r-th exterior power bundle
r
E, whose bre over x M
is the the r-th exterior power
r
E
x
of E
x
. The bundle charts for all of the above bundles all come out naturally
from those of E and F. If the original bundles were real analytic, all these various constructions will lead to
real analytic bundles. The details are left as an exercise.
Exercise 4.2.2.
(i): Giving a smooth (resp. analytic) morphism of smooth (resp. analytic) vector bundles E F is
equivalent to giving a smooth (resp. analytic) section of the bundle hom
R
(E, F).
(ii): As for vector spaces, the bundle hom
R
(E, F) is isomorphic to the vector bundle E
F.
(iii): Show that for any line bundle E, the bundle hom
R
(E, E) is a trivial bundle.
32 VISHWAMBHAR PATI
Example 4.2.3 (Tautological bundles over grassmannians). Let
n
be the trivial bundle of rank n over the
grassmannian G
k
(R
n
). That is, the bundle:
p :
n
(= G
k
(R
n
) R
n
) G
k
(R
n
)
where p is projection to the rst factor. Now we consider the subset:
k
n
:= {(P, v)
n
: v P}
and dene :
k
n
G
k
(R
n
) to be the restriction of p to E. Clearly the bre
1
(P) is precisely the k-
dimensional subspace P of R
n
. That is, the bre over the point P G
k
(R
n
) is the vector space P R
n
. One
needs to produce local analytic trivialisations for this bundle. Again, using the notation of 2.2.6, for a xed
E G
k
(R
n
), there is the map:
E
:
1
(U
E
) U
E
E
(P, v) (P,
E|P
(v))
which makes sense because
1
(U
E
) is precisely the set:
{(P, v) :
E|P
: P E is an isomorphism and v P}
If one uses the real analytic identication of U
E
with hom
R
(E, E
) via the chart

E
of example 2.2.6, then
the inverse of
E
is the map taking (T, w) to the element (T, w+Tw), which is certainly real analytic, and an
isomorphism on the bres (ie. from R
k
= E to P = graph T =
1
(P).)
The tautological bundle
1
n
on the space G
1
(R
n
) = RP(n 1) is often denoted
1
.
Exercise 4.2.4 (Tangent bundle of grassmannians and projective spaces).
(i): Let
k,
n
denote the bundle on G
k
(R
n
) whose bre at P G
k
(R
n
) is P
. Show that
k
n

k,
n
=
n
.
Also show that the tangent bundle of the grassmannian is isomorphic to the bundle hom
R
(
k
n
,
k,
n
).
(ii): Show that the following relation holds for the tangent bundle of RP(n):
T RP(n)
1

1
1
...
1
(n + 1) copies
where
1
is the dual bundle of
1
. (This dual of the tautological line bundle on projective space is called
the canonical line bundle). (Hint: Use (i) above and the part (iii) of 4.2.2 above.)
(iii): Show that
1
2
, the tautological line bundle on RP(1) S
1
is not a trivial bundle. (In fact this bundle
is homeomorphic to the Moebius strip, as dened in (vii) of 2.1.6. Why?)
4.3. Pullbacks, transition functions.
Denition 4.3.1 (Pullback). Let : E B be a smooth (resp. analytic) vector bundle of rank k, and let
f : X B be a smooth (resp. analytic) map. Consider the subset:
f
E := {(x, v) X E : f(x) = (v)}

Dene the map p : f
E X by p(x, v) = x (i.e. the restriction of the projection X E X). This bundle

is called the pullback of the bundle E X by f. Its bre over x X is precisely E
f(x)
=
1
(f(x)). If we let
f denote the map (x, v) v, then we have a commutative square:

f
E
f
E
p
X
f
B
To show that it is locally trivial is an easy exercise (in fact if (
i
, U
i
) is a bundle chart for E, then
i

f
will trivialise f
E over f
1
(U
i
)).
Exercise 4.3.2.
(i): Convince yourself that pullbacks commute with all the operations on vector bundles dened above (e.g.
tensor products, Whitney sums, homs, etc.).
(ii): If f : X B is the constant map, then f
E is a trivial bundle.
(iii): If there is a commutative square:
F
g
E
p
X
f
B
where p : F X is another vector bundle, and g is a smooth (resp. analytic) map which is linear on
each bre F
x
, then there is a unique smooth (resp, analytic) vector bundle morphism : F f
E (of
bundles on X) such that

f = g. (This is called the universal property of pullbacks. Any bundle having
this property with respect to f and E will become isomorphic to f
E, by virtue of the uniqueness of

the morphism in the denition).
(iv): If f : M N is a smooth (resp. analytic) map of smooth (resp. analytic) manifolds, then there is
a unique vector bundle morphism: TM f
TN (as bundles on M) which is denoted by f
. If f is an
immersion, this will realise TM as a sub-bundle of f
TN. If f is a submersion, it will realise f
TN as a
quotient of TM.
(v): Let f : S
n
RP(n) be the natural quotient map (f(x
0
, .., x
n
) = [x
0
: ... : x
n
]). Compute the bundles
f
1
and f
TRP(n) on S
n
.
There is another description of vector bundles, which is of great value in doing computations with them.
This is by means of the transition functions of that bundle.
Denition 4.3.3 (Transition functions). Let : E B be a smooth (resp. analytic) vector bundle of rank
k, with bundle charts {(
i
, U
i
)}. Then for x U
i
U
j
, we have the following diagram:
(U
i
U
j
) R
k
i
1
j
(U
i
U
j
) R
k
p
1
p
1
U
i
U
j
where both p
1
s are projections onto the rst factor. Since the top arrow
i

1
j
maps {x} R
k
linearly and
isomorphically to itself, this means that this map must be (x, v) (x, g
ij
(x)v), where g
ij
: U
i
U
j
GL(k, R)
is a smooth (resp. analytic) map. The functions {g
ij
: U
i
U
j
,= } are called the transition functions of the
bundle : E B.
Exercise 4.3.4.
(i): Show that the transition functions satisfy:
(a) g
ij
(x)g
ji
(x) = Id for all x U
i
U
j
; g
ii
= Id for all i
(b) g
ij
(x)g
jk
(x)g
ki
i
U
j
U
k
These are called the cocycle conditions.
(ii): Show that if there is an open covering {U
i
} of a manifold B, together with a collection of smooth (resp.
analytic) functions g
ij
: U
i
U
j
GL(k, R), which satisfy the cocycle conditions (a) and (b) of (i) above,
then there is a unique smooth (resp. analytic) bundle : E B whose transition functions are {g
ij
}.
34 VISHWAMBHAR PATI
(iii): Compute the transition functions of all the vector bundles encountered so far. Write down the tran-
sition functions of all the bundles that were constructed out of some given bundles (e.g. Whitney sums,
tensor products, exterior powers, homs, duals, pullbacks, etc.)
(iv): Let : E B be a smooth (resp. analytic) bundle of rank k, and let {(
i
, U
i
)} be its bundle charts.
Show that giving a smooth (resp. analytic) section s of this bundle is equivalent to giving a collection of
smooth (resp. analytic) functions s
i
: U
i
R
k
satisfying s
i
(x) = g
ij
(x)s
j
(x) for all x U
i
U
j
.
(v): Let E B and F B be two smooth (resp. analytic) vector bundles of rank k. Without loss of
generality, we can assume there is an open covering {U
i
} of B such that both bundles are trivialised over
these opens. Let g
ij
and h
ij
be their transition functions with respect to these trivialisations. Show that
E and F are isomorphic i there exist smooth (resp. analytic) functions s
i
: U
i
GL(k, R) (for all i)
such that:
g
ij
(x)s
j
(x) = s
i
(x)h
ij
(x) for all x U
i
U
j
Use this to formulate a criterion for a bundle to be trivial, in terms of its transition functions.
4.4. Tensors, dierential forms.
Example 4.4.1 (Tensors and and dierential forms). In view of the above, we can construct various associ-
ated bundles to the tangent bundle. For example the (k, l)-tensor power bundle : (
k
TM) (
l
T
M) M,
whose smooth sections are the k-contravariant s-covariant tensor elds on M, or the k-th exterior power bundle
:
_
k
T
M M, whose sections are called k-forms on M.

Example 4.4.2 (Local descriptions of tensors and forms). If (, U) is a coordinate chart on a smooth or an-
alytic manifold of dimension n, then the bundles TM and T
M get trivialised over U. Their local frames

(=basis sections) over U arising from this chart are {

xj
}
n
j=1
and {dx
i
}
n
i=1
. The associated bundle :
(
k
TM) (
l
T
M) M then acquires the trivialisation over U, given by the frame:

_

x
i1

x
i2
...

x
i
k
dx
j1
dx
j2
...dx
j
l
: 1 i
r
n 1 j
s
n
_
Thus if is a tensor eld of type (k, l), we have the following local expression for over U:
=
i1i2..i
k
j1j2...j
l
_

x
i1

x
i2
...

x
i
k
dx
j1
dx
j2
...dx
j
l
_
where
i1i2..i
k
j1j2...j
l
is a smooth real valued function on U. From the formulas
dx
j
=
m
x
j
y
m
dy
m
;

x
i
=
p
y
p
x
i
y
p
for another coordinate system (y
1
, .., y
n
) given by a chart (, V ), it follows that on U V :

p1....p
k
m1...m
l
=
i1,..,i
k
,j1,..,j
l
_
y
p1
x
i1
...
y
p
k
x
i
k
..
x
j1
y
m1
...
x
j
l
y
m
l
_
i1i2..i
k
j1j2...j
l
where
p1....p
k
m1...m
l
denotes the components of with respect to the coordinate bases arising from (, V ).
For k-forms, we trivialise the bundle
k
(T
M) by the frame:
{dx
I
:= dx
1
dx
i2
... dx
i
k
: i
1
< i
2
... < i
k
; 1 i
r
n}
We can thus express a k-form over the open set U as:
=
I
dx
I
where I = (i
1
, .., i
k
) is a multi-index with i
1
< i
2
< .. < i
k
, and
I
is a smooth real-valued function on U. If
one uses another chart (, V ), the corresponding formulas for the new components of on U V are:

J
=
J
A
JI
I
where A
JI
is the determinant of the k k-minor of the jacobian matrix
_
xi
yj
_
formed by the rows i
1
, .., i
k
and
columns j
1
, .., j
k
.
Denition 4.4.3. The vector space of all smooth k-forms on the manifold M will be denoted by
_
k
(M). Note
that
_
0
(M) = C
(M) is the space of all smooth functions on M. Clearly, by pointwise scalar multiplication,
_
k
(M) is a module over
_
0
(M).
4.5. Exterior derivative.
Denition 4.5.1 (Dierential of a smooth function). Note that for any open set U M, and a smooth func-
tion f : U R, there is a natural 1-form on U denoted by df, satisfying: df(X) = X(f) for a tangent vector
X T
x
(M) and x U. It is called the dierential of f.
Denition 4.5.2 (Exterior derivative). We dene the exterior derivative operator
d :
k
(M)
k+1
(M)
by rst dening it on
_
0
(M). An element of
_
0
(M) is a smooth function f on M, and we dene the 1-form
df by:
df(X) = Df(x)(X) = X(f)
for a tangent vector X T
x
(M) (as in the denition 4.5.1). Now one extends it to
_
k
(X) by requiring the
following condition:
d( ) = d + (1)
deg
d
where deg = p if is a p-form (the condition of a skew-derivation). It can be readily checked that this
denition makes sense. There is a local formula for d, in the exercise below.
Exercise 4.5.3.
(i): Let U be a coordinate patch corresponding to the local chart (, U), and let {x
i
} denote the corre-
sponding coordinate functions. As remarked earlier in 4.4.2, the coordinate 1-forms dx
i
form a basis of
_
1
(U), and thus a basis for
_
k
(U) is given by
{dx
I
:= dx
i1
dx
i2
.. dx
i
k
}
I
where I = (i
1
, .., i
k
) ranges over all k-multi-indices with 1 i
1
< i
2
< ...i
k
n. Show that if =
I

I
dx
I
is the expansion of over U, with
I
smooth functions, then d is given by:
d =
j,I
I
x
j
dx
j
dx
I
Verify that this formula is well-dened, independent of coordinate charts.
(ii): Prove that the composite:
k
(M)
d
k+1
(M)
d
k+2
(M)
is the zero map.
36 VISHWAMBHAR PATI
4.6. Pullbacks of dierential forms.
Denition 4.6.1. If f : M N is a smooth map, we have already introduced the bundle morphism f
:
TM f
TN. By taking duals, there will be a bundle morphism (of bundles over M):
f
: f
N T
M
which, upon taking exterior powers will yield a vector bundle morphism:
k
(f
) : f
(
k
T
N)
k
T
M
If
_
k
(M) is a k-form on N, it is a smooth section of
k
T
N, so tautologically results in a section

of f
(
k
T
N). We apply the above vector bundle morphism

k
(f
) to this pullback, and get a section of
k
T
M, i.e. an element of
_
k
(M). This form is called the pullback of , and denoted by f
, for notational
convenience. Similar considerations apply to analytic k-forms, analytic manifolds, analytic maps.
The only way to fathom the jargon above is to do the following:
Exercise 4.6.2.
(i): Show that for a 0-form (=function) g : N R, f
g = g f. Also show that f
dg = d(f
g).
(ii): Show that for smooth maps f : M N, g : N X, we have (g f)
= f
, by using the
corresponding property for pullbacks of bundles, duals etc. Also note that Id
= Id. Thus if f : M N
is a dieo, then f
is an isomorphism from
_
k
(N) to
_
k
(M) for all k.
(iii): Show that for dierential forms , on N, we have f
( ) = f
() f
()
(iv): Show that for a smooth map f : M N, and a dierential k-form on N, we have d(f
) = f
d.
(v): [Local formula for pullback] Let f be as above, and let (, U) (with coordinate functions x
1
, .., x
m
)
be a chart on M, and (, V ) a chart (with coordinate functions y
1
, .., y
n
) on N such that f(U) V .
Represent a k-form on the chart V as
|V
=

|I|=k

I
dy
I
, where
I
(y
1
, .., y
n
) are smooth functions.
Show that f
is represented on U by:
(f
)(x)
|I|=k
(
I
(f(x
1
, .., x
m
))df
I
for x U
where, for I = (i
1
, .., i
k
), df
I
denotes the k-form df
i1
df
i2
.. df
i
k
and f
i
denotes the i-th component
of f
1
4.7. Integration of dierential forms. The reason for introducing dierential forms is to globalise the
notion of integration from euclidean space to manifolds. For more details on this, the reader is urged to consult
the book [Spiv].
Denition 4.7.1 (Regular k-cubes and k-chains). Let M be a smooth manifold and let : U M be a
smooth map, where U R
k
is an open set containing the unit k-cube [0, 1]
k
. Then the map is called a
regular k-cube in M. The free abelian group on all the regular k-cubes is called the group of regular k-chains.
A regular k-chain is by denition a nite formal linear combination:
m
i=1
n
i
i
n
i
Z
where
i
are regular k-cubes in M. Note that the image of a regular k-cube is to be distinguished from a
regular k-cube. For example, if is the constant map, then the image of this regular k-cube is just a point.
For the (k 1)-cube [0, 1]
k1
in R
k1
, we dene the maps:
i
0
j
: [0, 1]
k1
[0, 1]
k
(t
1
, .., t
k1
) (t
1
, .., 0, ..., t
k1
) 0 at the j-th place
and
i
1
j
: [0, 1]
k1
[0, 1]
k
(t
1
, .., t
k1
) (t
1
, .., 1, ..., t
k1
) 1 at the j-th place
Denition 4.7.2 (Boundaries of regular cubes and chains). If is a regular k-cube in a manifold M, then
0
j
:= i
0
j
and
1
j
:= i
1
j
are regular k 1 cubes. We dene the boundary of to be the chain:
:=
k
j=1
(1)
j1
(
1
j

0
j
)
For a regular k-chain c =
i
n
i
i
, we dene c :=
i
n
i
i
.
The reader should do some pictures to convince herself that the signs have been put in to give it the correct
geometric meaning.
Exercise 4.7.3 (Boundary of a boundary). Verify that = 0
If is a smooth k-form on U, an open subset of R
k
containing [0, 1]
k
, then we may write = fdx
1
... dx
k
where f is a smooth function. Let
k
denote the identity regular k-cube in R
k
. We dene:
_
k
=
_
1
0
...
_
1
0
f(x
1
, .., x
k
)dx
1
dx
2
..dx
k
Denition 4.7.4 (Integration on chains). Let be a regular k- cube in a manifold M, and
_
k
(M) a
k-form. Since : U M is a smooth map, where U is a neighbourhood of [0, 1]
k
in R
k
, the pullback
is a
k-form on U (see denition 4.6.1). Thus, we dene
_
:=
_
where the right side is dened by the foregoing remarks. Finally, if c =
i
n
i
i
is a regular k-chain, we dene:
_
c
:=
i
n
i
_
i
Proposition 4.7.5 (Stokes Theorem). Let M be a smooth manifold, a k 1-form, and c a k-chain on M.
Then
_
c
d =
_
c
Proof: We rst prove the result for
k
and a (k 1)-form on a neighbourhood U of [0, 1]
k
in R
k
. In terms
of the basis (k 1)-forms, we may write
=
k
j=1
(1)
j1
j
dx
1
..
dx
j
.. dx
k
so that,
d =
_
_
k
j=1
j
x
j
_
_
dx
1
... dx
k
38 VISHWAMBHAR PATI
Then, successively using the Fundamental Theorem of Calculus on the j-th variable, we have
_
k
d =
_
1
0
.....
_
1
0
_
_
k
j=1
j
x
j
_
_
dx
1
dx
2
...dx
k
=
k
j=1
(1)
j1
_
1
0
...
_
1
0
(1)
j1
(
j
(x
1
, .., 1, ..x
k
)
j
(x
1
, .., 0, ..x
k
))dx
1
...
dx
j
..dx
k
=
k
j=1
(1)
j1
__
1
0
...
_
1
0
((i
1
j
)
(i
0
j
)
_
=
_
the last line following from the fact that (i

0
j
)
(dx
1
...

dx
l
... dx
k
) = 0 for k ,= l, so that (i
0
j
)
=
(
j
i
0
j
)dx
1
...
dx
j
..dx
k
, and likewise for (i
1
j
)
.
Now for a general regular k-cube in a manifold M, we have
_
d =
_
d =
_
k
d(
)
=
_
=
k
j=1
(1)
j1
_
_
i
j
1

_
i
j
0
_
=
k
j=1
(1)
j1
_
_
i
j
1

_
i
j
0
_
=
_
Finally, for regular k-chains, we obviously have the result by linearity. 2

Exercise 4.7.6. Consider the (n 1)-form
j
(1)
j1
x
j
dx
i
...
dx
j
.. dx
n
on R
n
, and let be the restriction (=pullback under inclusion) of this form to S
n1
. Compute
_
S
n1
. (Hint:
Express the sphere as the boundary of a suitable regular n-chain).
4.8. Orientability of bundles, manifolds.
Remark 4.8.1 (Triviality of real line bundles). A line bundle : E B is trivial i there exists a bundle
atlas {(
i
, U
i
)} for E such that the transition functions {g
ij
} corresponding to this trivialisation are all positive.
Only if is clear, since a trivial bundle can be trivialised with just one chart, for which g
11
= 1. Conversely, if
there exists such a system of charts with g
ij
(x) > 0 for all x U
i
U
j
, all i, j, we may write g
ij
(x) = exp(h
ij
(x))
for some smooth functions h
ij
: U
i
U
j
R. The cocycle conditions on g
ij
imply:
(a): h
ii
= 0 for all i; h
ij
(x) = h
ji
(x) for all x U
i
U
j
.
(b): h
ij
(x) +h
jk
(x) +h
ki
(x) = 0 for all x U
i
U
j
U
k
.
Extend each h
ij
by zero outside U
i
U
j
, and continue to denote them by h
ij
. Note that the condition (b),
i.e. h
ij
+h
jk
+h
ki
= 0 may no longer hold at all points, but only at points in U
i
U
j
U
k
.
Now take a partition of unity {
i
} subordinate to the covering {U
i
}, and dene:
t
i
(x) =
k
(x)h
ik
(x) for x U
i
Note that this is a nite sum. It is easy to check that these functions t
i
are smooth on U
i
. Since supp
k
U
k
,
for all x U
i
U
j
, we have
k
(x)(h
ij
(x)+h
jk
(x)+h
ki
(x)) = 0. This can be rewritten as
k
(x)(h
ik
(x)h
jk
(x)) =
k
(x)h
ij
(x) for all x U
i
U
j
. Then for x U
i
U
j
, we have:
t
i
(x) t
j
(x) =
k
(x)(h
ik
(x) h
jk
(x)) =
k
(x)h
ij
(x) = h
ij
(x)
Thus the collection of functions s
i
:= exp t
i
satisfy s
i
(x) = g
ij
(x)s
j
(x), and dene a nowhere vanishing section
of E. (See (iv) of 4.3.4).
Denition 4.8.2. We say that a vector bundle : E B of rank k is orientable if the top exterior power of
this bundle, namely :
_
k
E B is a trivial line bundle. This is clearly equivalent to asking (see exercise
4.3.4, part (iii) and the remark 4.8.1 above) that there exist a set of bundle charts so that the transition
functions for this line bundle, viz. det[g
ij
(x)] > 0 for all x U
i
U
j
, and all i, j. Clearly, a line bundle is
trivial i it is orientable.We say that a manifold M is an orientable manifold if its tangent bundle is orientable.
This is equivalent to requiring the existence of a compatible atlas so that the determinants of the jacobians
det
_
xi
yj
_
of all the coordinate changes are positive at all points of U
i
U
j
.
Example 4.8.3 (Tautological bundle on RP(n)). We claim that the tautological bundle
1
n+1
on RP(n) is not
an orientable (=trivial) bundle for all n. Note that since the restriction of a trivial bundle to a submanifold is
also trivial, it is enough to show that the bundle
1
2
on RP(1) is not a trivial bundle. The easiest way to do this
is to note that the removal of the zero-section RP(1) {0} from the trivial bundle RP(1) R will disconnect it
into the two connected components RP(1) (0, ) and RP(1) (, 0). If there were a bundle isomorphism
from
1
2
to the trivial bundle RP(1) R, it would carry the zero-section to the zero-section, and so if we denote
by Z the zero- section of
1
2
, the space
1
2
\ Z would be disconnected.
Recall that the bre
1
2,x
over x = [x
0
: x
1
] RP(1) is the line given by x, i.e. the set:
{(x, v) RP(1) R
2
: v
0
x
1
= v
1
x
0
}
We claim that
1
2
\ Z is connected. There are two bundle charts (
0
, U
0
) and (
1
, U
1
) where
U
i
= {[x
0
: x
1
] : x
i
,= 0} i = 0, 1
Both U
0
and U
1
are homeomorphic to R, as we have seen before, via the maps [x
0
: x
1
]
x1
x0
and
x0
x1
respectively.
The bundle charts are:
i
:
1
2|Ui
U
i
R
([x
0
: x
1
], (v
0
, v
1
)) v
i
for i = 0, 1. Since v
0
=
x0
x1
v
1
on U
0
U
1
, this shows that for [x
0
, x
1
] U
0
U
1
R
:= R\{0}, (after identifying

U
0
and U
1
with R as described above), the coordinate change map is:
0

1
1
(t, ) =
_
1
t
, t
_
Thus
1
2
\ Z is homeomorphic to two copies of R R
, with the point (t, ) in the rst copy identied with

(
1
t
, t) in the second copy, for t ,= 0. We leave it as an easy exercise for the reader to show that this space is
connected.
Exercise 4.8.4.
(i): Show that the open Moebius strip is not an orientable manifold. (If the tangent bundle of the Moebius
strip were orientable, its restriction E to the central circle of the Moebius strip would also be orientable.
Now, the tangent bundle T of this central circle is trivial, so orientable, and hence the quotient bundle
E/T would also be orientable. However, it is easy to check that this quotient bundle E/T on the central
circle ( RP(1)) is isomorphic to the line bundle
1
2
, which is non-trivial by the above example.)
(ii): Show that a projective space RP(n) is an orientable manifold i n is odd.
(iii): Show that for every manifold M, its tangent bundle is an orientable manifold.
(iv): Show that a complex manifold is an orientable manifold.
40 VISHWAMBHAR PATI
4.9. Integration again. Suppose M is a smooth orientable n-dimensional manifold, and is an n-form on
M. We would like to be able to dene
_
M
. One possibility is to express M as a regular n-chain, and apply
the denitions about integrations on chains. This would use the rather non-trivial fact that an orientable
n-manifold can be so expressed. An easier method is to use partitions of unity, coordinate charts, and apply
our knowledge of integration on R
n
.
So assume that is an n-form of compact support (i.e. vanishes outside a compact subset K M). For
example, if M is compact, then all n-forms on M are compactly supported. Let {(
i
, U
i
)} be a smooth oriented
atlas for M, viz. such that det(D(
i
1
j
) is everywhere positive on U
i
U
j
for all i, j. Let {
i
} be a partition
of unity subordinate to the covering {U
i
}. Then
i
is a compactly supported n-form, with support contained
inside U
i
. Then the pullback (
1
i
)
(
i
) is a compactly supported n- form on the open subset
i
(U
i
) of
R
n
. We can extend this to a compactly supported n-form (by zero) on all of R
n
. It is therefore of the form
fd
1
dx
2
.. dx
n
for some smooth compactly supported function on R
n
. Thus we dene the integral of this
form on R
n
to be
_
R
n
fdx
1
...dx
n
which makes sense by the compact support of f. Thus we nally dene:
_
M
=
i
_
R
n
(
1
i
)
(
i
)
Note that the support of being compact, will intersect only nitely many of the supports of
i
, because the
latter collection is locally nite from the denition of partitions of unity. Thus the sum on the right is a nite
sum.
Proposition 4.9.1. Let M be an orientable, smooth n-manifold, a compactly supported n-form on it. Then
the denition of
_
M
is independent of the choice of oriented atlases and partitions of unity.
Proof: We rst note that if = f(x
1
, .., x
n
)dx
1
... dx
n
is a compactly supported form on an open set U
of R
n
, and : U (U) is a smooth dieomorphism which preserves orientation (viz. det D(x) > 0 for all
x U), then
_
R
n
=
_
U
f(x
1
, .., x
n
)dx
1
...dx
n
=
_
(U)
f(
1
(
1
, ..,
n
))(det D)
1
d
1
, .., d
n
=
_
(U)
(
1
)
()
=
_
R
n
(
1
)
()
where we have used the change of variable formula from integral calculus, and (v) of 4.6.2.
Now let {(
i
, U
i
)} and {(
j
, U
j
)} be two countable atlases, both of whose coordinate change Jacobians are
everywhere positive. Thus, det D(
i

1
j
) will have the same sign for all i, j (on U
i
V
j
). Without loss of
generality, assume these are all positive (otherwise compose each
j
with a reection in R
n
). Let {
i
} and {
j
}
be partitions of unity subordinate to the coverings {U
i
} and {V
j
} respectively. Then, since
i

i
=
j

j
1,
we have:
_
M
=
i
_
R
n
(
1
i
)
(
i
)
=
j
_
R
n
(
1
i
)
(
i
j
)
=
j
_
R
n
(
1
j
)
(
1
i

j
)
(
i
j
)
=
j
_
R
n
(
1
j
)
(
i
j
) by remark above
=
j
_
R
n
(
1
j
)
(
j
)
which proves our assertion. 2
42 VISHWAMBHAR PATI
5. Metrics and Connections
5.1. Riemannian Metrics.
Denition 5.1.1 (Riemannian Metric). Let X be a smooth (resp. analytic) manifold, and let : E X be
a smooth (resp. analytic) vector bundle on X. A smooth (resp. analytic) Riemannian metric on this bundle
is a smooth (resp. analytic) section g of the tensor product bundle E
X such that g(x) E
x
E
x
is
a positive denite real valued bilinear form on E
x
, for each x X. It may be thought of as the family {g(x)}
of positive denite inner products on E
x
, varying smoothly (resp. analytically) with x X. For the sake of
convenience, we will denote g(x)(X, Y ) as X, Y
x
or even X, Y for X, Y E
x
.
A smooth (resp. analytic) Riemannian metric on the smooth (resp. analytic) tangent bundle : TX X
of X is said to be a smooth (resp. analytic) Riemannian metric on X. A pair (X, g), where X is a smooth (or
analytic) Riemannian manifold and g a Riemannian metric on X is called a Riemannian manifold. Usually g
is suppressed in the notation when it is understood. Note that a smooth manifold can have many Riemannian
metrics.
Since {dx
i
}
n
i=1
forms a local basis for the space of 1- forms on a chart U X, we may write:
g(x) =
i,j
g
ij
(x)dx
i
dx
j
(the right hand side is generally written as

i,j
g
ij
(x)dx
i
dx
j
for notational convenience) for all x U, where
[g
ij
(x)] is a positive denite symmetric matrix for each x U, and a smooth (resp. analytic) function of x U.
Clearly the matrix g
ij
is computed by g
ij
= g(

xi
,

xj
).
Proposition 5.1.2 (Existence of Smooth Riemannian Metrics). Let : E X be a smooth vector bundle
on a smooth manifold X. Then there exists a smooth Riemannian metric on it.
Proof: We rst note that if E X is a trivial bundle, i.e. if E = X V for some real vector space V ,
and = pr
1
, then we may take a positive denite inner product h on V and dene g to be the constant
section g(x) = h for all x X. For a general bundle, choose a system of bundle charts {(
i
, U
i
} such that the
restricted bundles : E
|Ui
U
i
are all trivial. Let h
i
be Riemannian metrics on these trivial bundles, and let
{
i
} be a partition of unity subordinate to the open covering {U
i
} (see example 3.1.6). Extend the sections
h
i
by 0 outside U
i
, and continue to denote these (discontinuous) sections of E
X by h
i
. Verify that
i

i
h
i
is a Riemannian metric on the bundle E, i.e. a smooth section of the bundle E
X. 2.
Note that in general, a Riemannian metric on any bundle will induce a Riemannian metric on all subbundles.
Note that due to the lack of analytic partitions of unity, the above procedure for constructing a smooth partition
of unity ops for the analytic case, and Riemannian metrics on analytic vector bundles and analytic manifolds
have to be constructed by other (global) means.
Exercise 5.1.3 (Bundle exact sequences).
(i): Let : E
1
E
2
be a smooth morphism of two smooth vector bundles over a manifold M. Show that if
the rank of the map
x
: E
1,x
E
2,x
is constant, independent of x M, then there are natural smooth
vector bundles ker (resp. Im ) whose bre at x M is ker
x
(resp. Im
x
). The quotient bundle
E
2
/Im is called Coker . If ker (resp. Coker ) is a bundle of rank 0, we call a monomorphism
(resp. epimorphism) of vector bundles.
We say a sequence of bundle morphisms:
E
1
E
2
E
3
is an exact sequence of bundles if Im = ker . We say a sequence:
0 E
1
E
2
E
3
0
is a short exact sequence of vector bundles if every pair of successive morphisms is an exact sequence.
This clearly means that is a monomorphism, is an epimorphism, and Im = ker .
(ii): Let E
2
be a Riemannian vector bundle, and let:
0 E
1
E
2
E
3
0
be a short exact sequence of vector bundles. Show that there is a vector bundle morphism : E
3
E
2
such that = Id
E3
. Conclude that E
2
E
1
E
3
. (Every short exact sequence of smooth bundles
splits, since every smooth bundle E
2
is a Riemannian bundle).
(iii): Let E, F be two smooth vector bundles with Riemannian metrics g and h respectively. Dene a bilinear
form , on the bundle hom(E, F) by:
T, S := trace
E
(T
S)
where T
is dened by the equation g(T
f, e) = h(f, Te). Verify that this is a Riemannian metric on the

bundle hom(E, F).
Example 5.1.4 (Riemmanian metric on the Sphere). We recall from the exercise 3.3.3 that the tangent bun-
dle of S
n
is the bundle:
TS
n
= {(x, v) S
n
R
n+1
: x, v = 0}
This makes it a sub-bundle of the trivial bundle S
n
R
n+1
, on which we have the obvious constant Rie-
mannian bundle metric from the Euclidean metric on R
n+1
. This induces the metric on TS
n
by dening
(x, v), (x, v
)
x
= v, v
where the right side is the Euclidean inner product.

Lemma 5.1.5 (Riemannian metric on submanifolds of R
n
). Let M R
n
be a submanifold of dimension m,
and let : U M be a smooth local chart (parametrisation) of M on an open set (U) M. Then,
with respect to the coordinate system (u
1
, .., u
m
) on (U) coming from the coordinate chart (
1
, (U)), the
Riemannian metric induced from R
n
is given by:
g((u
1
, ..., u
m
)) =
m
i,j=1
_
u
i
,

u
j
_
du
i
du
j
where , denotes the euclidean inner product on R
n
.
Proof: Clearly, the coordinate basis vector

ui
is, by denition D(

ui
), which is nothing but

ui
. Thus
g
ij
=
_

u
i
,

u
j
_
=
_
u
i
,

u
j
_
and this proves the assertion. 2
Exercise 5.1.6.
(i) Using the local parametrisation:
h : (0, ) (0, 2) S
2
(, ) (sin cos , sin sin , cos )
compute the metric on the open set h(U) (which is S
2
with a semi- great-circle removed) to be g(, ) =
d
2
+ sin
2
d
2
. Generalise to S
n
.
(ii) Write down a parametrisation of an open subset of the embedded torus in R
3
of the exercise 2.1.7 and
compute the metric coecients of the induced euclidean metric from R
3
. Do the same for the Moebius strip
of example 2.1.6 (vii).
44 VISHWAMBHAR PATI
Example 5.1.7 (Riemannian metric on the Grassmannian). Let M = G
k
(R
n
), and P M a point on it. By
2.2.6 and 4.2.4 (i), the tangent space T
P
(M) is naturally identied with the vector space hom
R
(P, P
). If
X : P P
is a linear map, we note that there is an adjoint of this map X
: P
P, which is dened by:

Xx, y = x, X
y for x P, y P
where , is the euclidean inner product on R

n
. We dene for tangent vectors X, Y T
P
(M), the inner
product
X, Y := g(P)(X, Y ) = trace
P
(X
Y ) =
k
i=1
Xe
i
, Y e
i
where e
i
is any orthonormal basis of the k-plane P. This is clearly symmetric and positive denite on T
P
(M).
We need to show that g(P) is a smooth function of P. We proceed as follows.
Claim: The tangent bundle TG
k
(R
n
) is isomorphic to the vector bundle hom(
k
n
,
k
n
) (see the example 4.2.3
for the denitions of the bundles
k
n
,
k
n
).
Proof of Claim: For brevity, let us denote
k
n
as , and
k
n
as
.
Since we cannot really use the denition of the tangent bundle of the Grassmannian in terms of derivations,
and since this bundle doesnt naturally sit as a sub-bundle of some simpler bundle, we have no option but
to proceed with local charts. We already know that this tangent bundle : TG
k
(R
n
) G
k
(R
n
) is locally
trivialised as follows:
1
(U
P
)

P
U
P
hom(P, P
)
X
Q
(Q, D(
P
)(Q)X
Q
)
where P is some xed point in G
k
(R
n
), Q is a variable point in U
P
, and X
Q
T
Q
(G
k
(R
n
)), and
P
is the
coordinate chart identifying U
P
with the euclidean space hom(P, P
) as dened in 2.2.6. This follows from

the way tangent bundles were trivialised in the denition 3.3.1, using the charts of the manifold.
Now we need to trivialise the bundle : hom(,
) G
k
(R
n
). Luckily, this bundle also trivialises over
U
P
. We need to note that for a variable point Q U
P
, the bre
1
(Q) = hom(Q, Q
). So our job is to
linearly and smoothly in Q, for varying Q U
P
, identify the varying vector space hom(Q, Q
) and the xed

vector space hom(P, P
). To do this we note that on U

P
U
Q
, we have the coordinate change map:
Q
(U
P
U
Q
)

1
Q
U
P
U
Q
P

P
(U
P
U
Q
)
which takes 0 Q
P
(Q). If we take the derivative of this map, we have:
hom(Q, Q
) = T
0
(
Q
(U
P
U
Q
)
D(
P
1
Q
)(0)
T
P
(Q)
(
Q
(U
P
U
Q
) = hom(P, P
)
which is the required linear isomorphism we seek. So trivialise the bundle hom(,
) by:
1
(U
P
)

P
U
P
hom(P, P
)
T
Q
(Q, D(
P

1
Q
)(0)T
Q
)
Thus one can nally dene the map:
1
(U
P
)

P

1
(U
P
)
X
Q
(
1
P

P
)X
Q
In an overlapping chart
1
(U
R
), such that Q U
R
U
P
we need to check that
R
(X
Q
) agrees with
P
(X
Q
).
But this is clear because:
R
(X
Q
) = (
1
R

R
)X
Q
=
1
R
(Q, D(
R
)(Q)X
Q
) =
_
D(
R

1
Q
)(0)
_
1
D(
R
)(Q)X
Q
=
_
D(
Q

1
R
)(
R
(Q)
_
D(
R
)(Q)X
Q
= D
Q
(Q)X
Q
and so also
P
(X
Q
) = D
Q
(Q)X
Q
by replacing R by P in each of the last bunch of equations. This proves
the claim. 2
Let us x a P, and consider the metric coecients of this metric in the trivialisation of TG
k
(R
n
) over U
P
.
As noted above, for Q in the open set U
P
, we have the isomorphism:
hom(Q, Q
)
D(
P
1
Q
)(0)
hom(P, P
)
whose inverse we may denote as
(Q) : hom(P, P
) hom(Q, Q
)
Also we have, for Q U
P
, the isomorphism:
P|Q
: Q P
whose inverse we denote by (Q). Thus if {e
i
}
k
i=1
is a xed basis of P, then {(Q)e
i
}
k
i=1
will be a basis for Q,
not necessarily orthonormal. Then the dual basis to {(Q)e
i
}
k
i=1
is the basis {((Q)
1
)
(e
i
)
}
k
i=1
.
Thus, for S, T hom(P, P
) we have:
g(Q)(S, T) = trace
Q
(((Q)S)
(Q)T)
=
k
i=1
((Q)e
i
)
(((Q)S)
(Q)T(Q)e
i
)
=
k
i=1
((Q)
1
)
(e
i
)
(((Q)S)
(Q)T(Q)e
i
)
which is clearly smooth in Q.
In fact, this metric has another nice property, i.e. that it is homogeneous.
We have an action of the group G = O(n, R) on M = G
k
(R
n
). If P G
k
(R
n
) is a k-subspace of R
n
, spanned
by some basis {e
i
}
k
i=1
, dene for g G the element gP to be the plane spanned by {ge
i
}
k
i=1
. That this is a
smooth action, i.e. that the map:
: GM M
(g, P) gP
is smooth is left to the reader, using the local charts dened on M. Thus, for any xed g G, the left
translation by g, denoted L
g
, is a dieomorphism, with inverse L
g
1. Since g is an orthogonal transformation,
(gP)
= g.(P
). The reader should verify that L

g
(U
P
) = U
gP
, and using the denition of
P
, that the
following diagram commutes:
U
P
P
hom(P, P
)
L
g
Ad(g)
U
g.P
gP
hom(gP, (gP)
)
where, for T hom(P, P
), Ad(g)T is dened as the composite:

gP
g
1
P
T
P
g
(gP)
From the commutativity of this diagram, it will follow that the diagram below commutes:
T
P
(M)

|T
P
(M)
hom(P, P
)
(DL
g
(P)) Ad(g)
T
gP
(M)

|T
gP
(M)
hom(gP, (gP)
)
46 VISHWAMBHAR PATI
where is the vector bundle isomorphism between the bundles TM and hom(,
) constructed above. Thus,

for tangent vectors X, Y T
P
(M) hom(P, P
), since g is an orthogonal transformation, we have:

trace
P
(X
Y ) =
k
i=1
Xe
i
, Y e
i
=
k
i=1
gXe
i
, gY e
i
=
k
i=1
gXg
1
ge
i
, gY g
1
ge
i
_
=
k
i=1
Ad(g)Xf
i
, Ad(g)Y f
i
= trace
gP
((Ad(g)X)
Ad(g)Y )
where {e
i
} is an orthonormal basis for P, so that {f
i
:= ge
i
} is an orthonormal basis for gP.
In other words, DL
g
is an isometry between T
P
(M) and T
gP
(M).
Denition 5.1.8 (Isometry). If
i
: E
i
M, with i = 1, 2 are two vector bundles with bundle metrics g
i
respectively, then a bundle isomorphism : E
1
E
2
is called a bundle isometry if:
g
2
(X, Y ) = g
1
(X, Y ) for all X, Y E
1,x
and all x M
A dieomorphism f : M
1
M
2
is called an isometry if the bundle map f
: TM
1
f
TM
2
(dened in
(iv) of 4.3.2) is a bundle isometry. (Note that the pullback of a bundle with a Riemannian metric acquires
a Riemannian metric in a tautological way). An immersion i : M N of Riemannian manifolds is called a
Riemannian immersion if the bundle morphism i
: TM i
TN satises g
M
(x)(X, Y ) = g
N
(x)(i
X, i
Y )
for all X, Y T
x
M, all x M. A Riemannian immersion which is an embedding is called a Riemannian
embedding.
Denition 5.1.9. Let G be a Lie group acting smoothly from the left (resp. analytically) on a smooth (resp.
analytic) manifold M. That is, there is a smooth (resp. analytic) map:
: GM M
(g, x) gx
satisfying g(g
x) = (gg
)x for all g, g
G, x M and e.x = x for all x M. Further assume that the action

is transitive, i.e. for each x M, the orbit of x, viz Gx = {gx : g G}, is all of M. By left translation
L
g
we mean the map x gx, which is a dieomorphism of M. We say that a Riemannian metric h is left
G-homogeneous if L
g
is an isometry with respect to h, for each g G. Similarly, we may dene right actions,
and right G- homogeneity.
Example 5.1.10. For example, we saw in 5.1.7 above that the Riemannian metric we constructed on G
k
(R
n
)
was a homogeneous metric. Verify that if E =
k
n
and F =
k,
n
(with the natural Riemannian metrics g and h
induced from the constant euclidean metric on the trivial bundle
n
on G
k
(R
n
), of which they are sub-bundles),
verify that the bundle metric thus resulting on hom(
k
n
,
k,
n
) TG
k
(R
n
) in accordance with the exercise 5.1.3
is exactly the homogeneous Riemannian metric constructed above in 5.1.7.
Exercise 5.1.11 (Homogeneity of the Riemannian metric on S
n
). Note that G = O(n) acts by left multipli-
cation on S
n1
. Verify that the Riemannian metric on TS
n1
constructed in 5.1.4 is homogeneous with respect
to this group action.
Example 5.1.12 (Invariant metrics on Lie groups). Since a Lie group acts smoothly on itself by both left
(resp. right translations), and the tangent bundle of a Lie group is trivial by part (iii) of the exercise 4.1.3
using left (resp. right) translations, we can put a left (resp. right) homogeneous (=left or right invariant)
metric on a Lie group by putting a positive denite bilinear form , on the tangent space T
e
(G) and left
(resp. right) translating it to all points of G. Note that for a Lie group, there is a unique left (resp. right)
translation L
g
(resp R
g
) taking e to a point g, so these trivilisations, and invariant metrics are well-dened.
This left-invariant (resp. right invariant) metric is the unique metric on G which makes each left translation
(resp. right translation) an isometry.
A metric on G which is both left and right invariant is called a bi-invariant metric. This is the unique metric
making each left and right translation an isometry.
We remark that a left (resp. right) G-invariant metric on G need not be right (resp. left) invariant in general.
We shall see below a condition for a metric , on the Lie algebra g := T
e
(G) to give rise to a bi-invariant
metric. It is purely a condition on the Lie algebra.
Clearly a left-invariant metric on G is bi-invariant i it is invariant under the adjoint action of G on itself
(namely g h := Ad(g)h = ghg
1
= (L
g
R
g
1)h for g, h G). Since the metric is left-invariant, and there is
the formula:
L
Ad(h)g
Ad(h) = (Ad(h) L
g
) g, h G
global Ad-invariance is equivalent to demanding that the linear map D(Ad(g))(e) : g g (which is denoted
(Ad(g))
for ease of notation) preserves the inner product on g = T

e
(G).
Here is an example of a left-invariant metric which is not bi-invariant. Consider G = GL(2, R), with
g = gl(2, R) being its Lie algebra. Verify that the bilinear form
X, Y := tr(XY
t
)
is symmetric, in fact positive denite, and denes an inner product on g, called the Hilbert-Schmidt inner
product (=the Euclidean inner product if we treat g as Euclidean 4-space). If we dene a left-invariant metric
by left-translating as above, we get a Riemannian metric on G. This Riemannian metric is not right invariant,
because ., . is not Ad-invariant. For let:
g =
_
1 1
0 1
_
, X =
_
0 0
0 1
_
from which it follows that Z := Ad(g)X = gXg
1
=
_
0 1
0 1
_
from which it follows that X, X = 1 ,=
Ad(g)X, Ad(g)X = 2.
Proposition 5.1.13. If G is a compact Lie group, then it has a bi-invariant metric.
Proof: Let d denote the right-invariant Haar-measure on G, and let ( , ) denote any positive denite inner
product on g. For X, Y g dene:
X, Y :=
_
G
((Ad(g))
X, Ad(g)
Y )d(g)
Since a smooth function of G is being integrated on a compact set, this expression makes sense. Since
Ad(g)Ad(h) = Ad(gh) and d(g) = d(gh), it is easily veried that ., . is a positive denite Ad-invariant inner
product on g. Extending it to a left-invariant metric by left-translations as in 5.1.12 above gives a bi-invariant
metric on G. 2
We shall be dealing with homogeneous manifolds and metrics in greater detail in a later subsection.
5.2. Arc length on a manifold. On a Riemannian manifold, there is a natural method of measuring dis-
tances. Let M be a smooth or analytic manifold, with a Riemannian metric , , whose restriction to T
x
(M)
is denoted ,
x
.
48 VISHWAMBHAR PATI
Denition 5.2.1. We dene a piecewise smooth curve c : [0, 1] M joining x and y in M to be a continuous
curve which is smooth when restricted to each sub-interval [a
i
, a
i+1
] of some nite partition of [0, 1] (with
0 = a
0
< a
1
< ... < a
N
= 1), and such that c(0) = x, c(1) = y. For such a piecewise smooth curve, the velocity
vector
dc
dt
:= c
_
d
dt
_
= Dc
_
d
dt
_
makes sense at all points c(t) except when t = a
i
. One then denes the arc
length of c by:
l(c) :=
N1
i=1
_
ai+1
ai
_
dc
dt
(s),
dc
dt
(s)
_1
2
c(s)
ds
Exercise 5.2.2. Show that the arc length of the composite of two piecewise smooth curves (which is also
piecewise smooth) is the sum of the arc lengths of the two curves.
Denition 5.2.3 (Metric on a Riemannian manifold). For x, y M, dene:
d(x, y) = inf{l(c) : c is a piecewise smooth curve joining x to y}
It is obvious that d(x, x) = 0 and d(x, y) = d(y, x). That the triangle inequality holds is an obvious
consequence of the fact that the composite of a piecewise smooth curve joining x to y with another one joining
y to z is a piecewise smooth curve joining x to z. We leave the verication that d(x, y) = 0 implies x = y to the
reader (using the fundamental theorem of calculus). This will be proved later, anyway. Thus a Riemannian
manifold becomes a metric space in a natural way. It is not clear at this point that the metric topology from
this metric coincides with the given topology on M. This is true, and will follow from later considerations
about geodesics.
Remark 5.2.4. Note that there may not exist any piecewise smooth curve joining x to y whose arc length is
equal to d(x, y). For example, if we take M = R
2
\ {0}, with the induced euclidean metric, then there is no
piecewise smooth curve joining x = (1, 0) to y = (1, 0) whose arc length is 2 = d(x, y). This matter will be
again resolved after the discussion on geodesics and completeness.
Remark 5.2.5. It is clear that if a Lie group G is acting smoothly on a smooth manifold, and , is a
Riemannian metric on M which is left G-homogeneous, then the metric d dened above is also G-homogeneous
(or G-invariant), that is each L
g
acts as an isometry with respect to the distance d.
5.3. Volumes on Riemannian manifolds. It is well-known that if we take n linearly independent vectors
v
1
, .., v
n
R
n
, then the (n-dimensional) volume of the parallelopiped spanned by {v
i
}
n
i=1
is just the number
det[v
ij
], where v
i
=

j
v
ij
e
j
. (If the reader isnt aware of this fact, then the best thing is to prove it by
induction, noting that the inductive denition of the volume of an n-dimensional parallelopiped is the length
of the perpendicular from the vertex v
n
to the base (n 1)-dimensional parallelopiped spanned by {v
i
}
n1
i=1
,
multiplied by the (n 1)-dimensional volume of this base).
More generally, if g is any positive denite inner product on a vector space V of dimension n, then the volume
of a parallelopiped spanned by {v
i
}
n
i=1
will be det[v
ij
], where v
i
=
j
v
ij
e
j
, with respect to the g-orthonormal
basis {e
i
}
n
i=1
. Thus we have:
Denition 5.3.1 (Volume form). The n-form (called the volume form on V , dened as := e
1
e
2
... e
n
,
when applied on an n-tuple of vectors (v
1
, .., v
n
), computes the volume of the parallelopiped spanned by {v
i
},
and has a sign 1 depending on whether the tuple is positively or negatively oriented with respected to the
orthonormal basis e
i
.
If f
i
is an arbitrary basis, and g
ij
:= g(f
i
, f
j
), then we write e
i
=

l
A
il
f
l
, and note that the relation
g(e
i
, e
j
) =
ij
leads to the relation AgA
t
= Id, from which it follows that det A = (det g
ij
)
1
2
. Now since
e
i
=
A
1
li
f
l
=
(A
1
)
t
il
f
l
, we have:
: = e
1
e
2
... e
n
= (det A)
1
(f
1
f
2
... f
n
)
= (det g
ij
)
1
2
(f
1
f
2
... f
n
)
If we assume that M is an orientable manifold, then one can get an atlas {(
i
, U
i
)} such that det D(
i
1
j
) > 0 for all i and j. In this case
n
T
M is trivial, and it would be convenient to have an everywhere positive

section of this bundle whose restriction to each bre
n
T
x
M gives the volume form on T
x
M in accordance
with the denition above. This is done as follows:
Denition 5.3.2 (Volume form on a Riemannian manifold). Let (, U) be a coordinate chart on a Riemann-
ian manifold M, with Riemannian metric , . Let [g
ij
] =
_

xi
,

xj
_
be the matrix of the metric in the
resulting coordinate system. Then the volume form dV is given on U by the formula:
dV = g
1
2
dx
1
... dx
n
where g = det [g
ij
].
It is easy to check that this form is globally well-dened, independent of charts, and everywhere positive.
The orientability of M comes in when we want to verify this, for if A = D(
i

1
j
) gives a change of the basis
{

xm
}, the quantity g
1
2
changes by | det A|.
Exercise 5.3.3. Let : E M be a smooth vector bundle, and let g := , be a Riemannian inner product
on it. Then show that if {e
i
}
k
i=1
is a local orthonormal frame (viz. orthonormal sections on some open set U),
we can dene a metric on the m-th exterior power
m
E by requiring the set
{e
I
:= e
i1
e
i2
... e
im
: I = (i
1
< i
2
... < i
m
), 1 i
r
n}
to be orthonormal. Show that this indeed denes a Riemannian metric, and satises:
v
1
v
2
... v
m
, w
1
w
2
... w
m
= det [v
i
, w
j
]
Note that with this denition, the volume form on an orientable Riemannian manifold is the section of
n
T
M
which is of unit length at all points, i.e. is an orthonormal frame of this trivial line bundle.
5.4. Vector elds, trajectories. Before we examine more examples of homogeneous metrics, we would like
to introduce some notions which are crucial to the sequel. The rst is the existence theorem on ordinary
dierential equations, and the denition of the exponential mapping on Lie groups.
Denition 5.4.1 (Integral curves or trajectories). Let M be a smooth manifold, U M an open subset, and
let X a smooth vector eld on U. A smooth curve, i.e. a smooth map c : (, ) U is said to be an integral
curve or trajectory of X if:
dc
dt
(t) := Dc(t)
_
d
dt
_
= X(c(t)) t (, )
If c(0) = x U, we say it is an integral curve through x.
50 VISHWAMBHAR PATI
We note here the fundamental existence theorem for ordinary dierential equations.
Theorem 5.4.2 (Picards Existence and Uniqueness Theorem for ODEs). Let X be a smooth vector eld on
an open subset U R
n
. Then:
(i): (Existence) For each x U, there exists an > 0 (depending on x) and a smooth map c
x
: (, ) U
such that:
(a)
dc
x
(t)
dt
= X(c
x
(t)) for all t (, )
(b) c
x
(0) = x
Such a c
x
is called a trajectory or integral curve of the vector eld X.
(ii): (Uniqueness) If c : (, ) U is another smooth map satisfying (a) and (b) of (i) above, then c c
x
on (, ) where = min {, }.
(iii): If K U is a compact set, then there exists a uniform
K
> 0 depending only on K such that
c
x
: (
K
,
K
) U is dened for all x K, and satises (a) and (b) of (i) above.
(iv): c
x
is continuous in the variable x. That is, if x
n
x for x
n
, x U, and c
xn
, c
x
are all dened on say
[, ], then c
xn
c
x
uniformly on [, ].
Proof: Choose an open set W U such that W U, W is compact, and x W. Since X is smooth, both
X and DX are continuous maps, so that there is a constant C such that:
|X(x)| C for all x W
|X(x) X(y)| C |x y| for all x, y W (6)
(The second equation follows from the mean value theorem, or 1st order Taylor Theorem, applied to the smooth
map X.)
Choose > 0 so that for the C as in (6), we have:
C
1
2
and B(x, C) W (7)
Now consider the space:
B
x
:= {c : [, ] W : c continuous, c(0) = x}
Dene a metric on B
x
by:
d(c
1
, c
2
) = sup
t[,]
|c
1
(t) c
2
(t)|
It is easily checked that B
x
is a complete metric space with respect to this metric.
If c satises (a) and (b) of statement (i) of our proposition, i.e. if:
dc
dt
(s) = X(c(s)) and c(0) = x
then it would follow, from the Fundamental Theorem of Calculus, that:
c(s) = x +
_
s
0
X(c(t))dt (8)
and conversely, if c satises the above integral equation, it would satisfy the ODE above, and the condition
c(0) = x.
So consider the integral operator T on B
x
dened by
(Tc)(s) = x +
_
s
0
X(c(t))dt
Note that since |Tc(s) x|
_
s
0
|X(c(s))| ds C, and B(x, C) W by the equation (7) above, hence
Tc(s) W for s [, ]. Also Tc is obviously continuous by its denition, and Tc(0) = x, hence Tc B
x
as
well. Note that c B
x
satises (a) and (b) of (i) i it satises the integral equation (8) above, i.e. i c
x
is a
xed point of T.
We now claim that T is a contraction mapping of the complete metric space B
x
. This is because for any
c, c B
x
and any s [, ] we have:
d(Tc, Tc) = sup
s[,]
|Tc(s) Tc(s)| sup
s[,]
_
s
0
|X(c(t)) X(c(t))| dt
C
_
s
0
|c(t) c(t)| dt C d(c, c)
1
2
d(c, c)
from the equations (6) and (7) above. Thus, by applying the Banach Contraction Mapping Theorem 1.5.2, we
have a xed point c, and this is the solution sought.
Since this xed point is unique, it follows that this c is the unique solution to the ODE above satisfying
c(0) = x. Let us denote it by c
x
, to signify that c
x
(0) = x.
Note that for all t [, ], c
x
(t) W since c
x
B
x
by its construction. A fortiori, we have c
x
([, ]) U,
since W U.
That c
x
is smooth in the variable s for a xed x follows from the integral relation:
c
x
(s) = x +
_
s
0
X(c
x
(t))dt
because the continuity of c
x
makes the integrand on the right continuous, so the relation above makes c
x
a C
1
function of s, which again by the relation above, makes c
x
C
2
, and so on by induction (this inductive process
of assuming it is C
k
, and concluding it is C
k+1
is known as bootstrapping.)
The uniqueness in part (ii) is clear, since both c
x
and c satisfy (a) and (b) of (i) on [, ], they both will
satisfy the same integral equation (8) for s [, ].
For (iii), we note that if K U is any compact set, and we let W U be an open set such that K W
W U, and let := d(K, W
c
) > 0, then B(x, C) W for all x K provided we choose C <

2
say. Thus
we may use such a W and such an > 0 in (7), and the proof above will now apply to all x K.
For (iv), i.e. continuity in the variable x also follows easily from the integral relation (8), For if is chosen
so that C <
1
2
, we have for s [, ], and d denoting the sup-distance introduced earlier on the interval
[, ]:
|c
x
(s) c
y
(s)| |x y| +Cd(c
x
, c
y
) for all s [, ]
d(c
x
, c
y
) |x y| +
1
2
d(c
x
, c
y
)
d(c
x
, c
y
) 2 |x y|
by using (6) in the second line, and hence continuity in x is clear. The proposition follows. 2
Smoothness in the variable x is, however, a delicate matter, and the reader is referred to Serge Langs
Dierential Manifolds p.66-82, for a complete proof. Also Langs Analysis II, p. 126-138. 2
Corollary 5.4.3. Let X and U be as in the Theorem 5.4.2 above. Assume that X is compactly supported,
i.e., there exists a compact set K U such that X(x) = 0 for all x , K. Then, for all x R
n
, there exists a
unique c
x
: R R
n
satisfying (a) and (b) of (i) of the Theorem 5.4.2. (A vector eld all of whose trajectories
can be dened for all t R is called a complete vector eld, and this corollary says that a compactly supported
vector eld is complete.
52 VISHWAMBHAR PATI
Proof: Extend the vector eld from U to a smooth vector eld on all of R
n
by setting X(x) = 0 for x , K.
Continue to call it x.
For x K
c
, the curve c
x
: R R
n
dened by c
x
(t) = x for all t R is clearly the unique trajectory with
initial point x K
c
.
For x K, by (iii) of the Theorem 5.4.2, there is a uniform > 0 such that the trajectory c
x
: (, ) R
n
is dened for all x K. Thus, combining with the rst para, we have c
x
: (, ) R
n
dened for all x R
n
.
For any such solution c
x
with x R
n
, let us extend it to a solution c : (, 3/2) R
n
as follows:
c(t) = c
x
(t) for t (, )
= c
cx(/2)
(t /2) for t (/2, 3/2)
It is easy to check by the uniqueness part (ii) of Theorem 5.4.2 that the two denitions above match on the
overlap (/2, ), so that c is smooth. Proceeding in this manner inductively, we have it dened on all of R.
(Exercise: Is c
x
(R) U for all x U? Same question with U replaced by K?) 2
Remark 5.4.4. Consider the vector eld X(t) 1(= 1.d/dt) on the open interval (0, 1). Then for x (0, 1),
we have c
x
(t) = x +t, and thus c
x
(t) U only for t (x, 1 x). Thus a uniform cannot be chosen for all
x U.
Remark 5.4.5. The uniqueness part breaks down completely if we do not assume that X is smooth. For,
dene the (continuous) vector eld X(x) = x
2/3
on the interval (, ) for any > 0. For an [0, ], consider
the curve:
c
(s) = 0 for t (, ]
= (1/3)
3
(t ) for t (, )
Then it is readily checked that for each [0, ], the curve c
satises c
(0) = 0 and
dc
(t)
dt
= X(c
(t)) for all

t (, ). Of course for ,= , c
,= c
, so there is at least a one parameter family of (in fact C

1
) solutions
with the same initial condition, and uniqueness breaks down quite badly. Also note that the solution c
0
which
is non-vanishing in any neighbourhood of 0 shows that even though X(0) = 0, there is a solution with initial
point x = 0 which does not stay put at 0 for any t > 0 (unlike the constant solution c
0 on (, )).
Denition 5.4.6. Note that by the uniqueness part (ii) of the Theorem 5.4.2, if c
x
is dened on any family
of open intervals I
with 0 I
for each , then it is also dened on I =
, because the solutions will

always match on any intersection I
by uniqueness. (Check!). Hence there is a maximal open interval

J(x) containing 0 such that c
x
: J(x) U is dened. For example, in the case of compactly supported vector
elds, we saw in Corollary 5.4.3 that J(x) = R for all x U.
Proposition 5.4.7. Let X and U be as in Theorem 5.4.2. Then the subset:
:= {(x, ) U R : J(x)}
is an open subset of U R (and hence open in R
n
R).
Proof: Let (x, ) , viz, J(x). Since J(x) is open, choose > such that [0, ] J(x). Then
c
x
: [0, ] U is certainly dened, since [0, ] J(x) by denition. Since K := c
x
([0, ]) is a compact subset
of U, there exists a relatively compact open set W such that
K W W U
Indeed, take =
1
2
d(K, U
c
), and let W = {y U : d(y, K) < }. One can, by the part (iii) of Theorem 5.4.2
choose a uniform > 0 such that c
z
: [, ] U is dened for all z W.
Find a partition 0 = t
0
< t
1
< t
1
< ....t
m
= so that t
i+1
t
i
< for all i, and set y
i
:= c
x
(t
i
). By the
sup-norm estimate in 5.4.2,
d(c
z
, c
x
) = sup
0s
|c
x
(s) c
z
(s)| 2 |x z|
for all x, z W, there is a neighbourhood N
0
around x = c
x
(0) = c
x
(t
0
) such that |c
z
(t
1
) c
x
(t
1
)| <
1
2
for
all z N
0
. Since c
x
(t
1
) K, it follows that d(c
z
(t
1
), K) <
1
2
< , so that c
z
(t
1
) W for z N
0
. Thus for
z
1
:= c
z
(t
1
), we have c
z1
(t) dened for all t [, ]. But, by uniqueness:
c
c(z1)(t)
(s) = c
z
(s +t)
so that we have c
z
dened on [0, t
2
] [0, t
1
+ ] for all z N
0
. Now, again, there is a neighbourhood N
1
of
y
1
= c
x
(t
1
) such that c
z
(t
2
) W for all z N
1
. Shrinking N
0
if necessary to ensure that c
z
(t
1
) N
1
for all
z N(0), we can repeat the above argument to show that for z N
0
, we have c
z
dened on [0, t
3
] [0, t
2
+].
Continuing nitely many times, we have for some neighbourhood N
0
of x that c
z
is dened on [0, t
m
] = [0, ]
and lies in W for all z N
0
. Thus the neighbourhood N
0
( , ) of (x, ) is contained in , for every
> 0, and is open. 2
Exercise 5.4.8. Show that for X and U as in Theorem 5.4.2, if c
x
(t
1
) = c
x
(t
2
) for some t
1
,= t
2
J(x), then
c
x
is a periodic trajectory, viz.
c
x
(t) = c
x
(t +t
2
t
1
)
for all t, and in particular, J(x) = R.
Denition 5.4.9. Let X and U be as in the statement of the Theorem 5.4.2, and let be the open set dened
in the Proposition 5.4.7. Then dene:
: U
(x, t) c
x
(t)
This map is called the ow of the vector eld X. (x, t) is often denoted
t
(x). Clearly
0
= (, 0) = Id
U
.
Exercise 5.4.10. Show that
t+s
(x) = (
t

s
)(x), whenever both sides make sense. In particular, if X is a
complete vector eld, then
t
: U U is dened for all t R, and is a dieomorphism with inverse
t
.
The Theorem 5.4.2 above globalises to manifolds as follows:
Theorem 5.4.11 (Existence and Uniqueness of Flows on Manifolds). Let M be a smooth manifold and U
M an open set. Let X be a smooth compactly supported vector eld on U, with support K U. (For example,
if M is itself compact, then X can be any smooth vector eld on M). Then there exists an > 0 and a smooth
map : U (, ) U satisfying (i), (ii) and (iii) of the previous theorem.
Proof: For each point x U, nd a neighborhood U(x) of x such that U(x) is contained in a chart, U(x) U,
and U(x) is compact. Let
x
be a smooth function on U such that supp
x
U(x), 0
x
1, and
x
is
identically = 1 on a neighbourhood W(x) U(x) with W(x) U(x) (see 3.1.8 for the construction of such a
function). Now consider the vector eld Y
x
=
x
X, which is smooth and compactly supported in U(x), and
coincides with X in W(x). Since U(x) is dieomorphic to an open subset of R
n
via the chart, by the previous
theorem 5.4.2 there exists a smooth mapping:
x
: U(x) (
x
,
x
) U(x)
satisfying (i), (ii) and (iii) of that theorem. Since K is compact, there exist {W(x
i
)}
m
i=1
such that K
m
i=1
W(x
i
). Let = min {
xi
}
m
i=1
. Dene (x, t) =
xi
(x, t) for x W(x
i
) and i = 1, .., m, and dene
(x, t) = x for x outside K and all t (, ). This is well dened because, by uniquenss,
xi
and
xj
are
both ows for X on the intersection W
(
x
i
) W(x
j
), and hence agree there by (iii) of the last theorem. Clearly
is the required ow. 2
54 VISHWAMBHAR PATI
5.5. Lie algebras, exponential mapping.
Denition 5.5.1 (Commutator of vector elds). Let M be a manifold, and let X, Y two smooth vector elds
on an open subset U M. We dene another vector eld denoted [X, Y ] on U by the prescription:
[X, Y ]f = X(Y (f)) Y (X(f))
for all smooth functions f on U. Note that X(f) and Y (f) are smooth functions on U, and hence the denition
makes sense. (Not quite, for one should be dening the action of the tangent vector [X, Y ](x) at each point
x U by its action on the germs f O
M,x
, but the denition is analogous to the one above, and we leave
those details to the reader. The local formula (iv) of the exercise 5.5.2 below will show that this can be carried
out, and doesnt depend on germ representatives chosen etc.)
Exercise 5.5.2. For vector elds X, Y, Z on an open subset U M, show that:
(i): [X, Y ] = [Y, X] (anticommutativity)
(ii): [X, fY ] = f[X, Y ] +X(f)Y for a smooth (resp. analytic) function f on M.
(iii): [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0 (Jacobi Identity)
(iv): Using smooth (resp. analytic) charts, show that if X =
n
i=1
a
i
xi
and Y =
n
j=1
b
j
xj
in the local
basis of coordinate elds, then the local expression for [X, Y ] is:
[X, Y ] =
i,j
_
a
i
b
j
x
i
b
i
a
j
x
i
_

x
j
and hence [X, Y ] is a smooth (resp. analytic) vector eld if X, Y are smooth (resp. analytic) vector
elds.
(v): If f : M N is a smooth (resp. analytic) dieomorphism, then [f
X, f
Y ] = f
[X, Y ]. (Note that

though f
(X) makes sense for a tangent vector for any smooth map f, it doesnt make sense for X a
vector eld, unless f is a dieomorphism.)
(vi): For the coordinate vector elds

xi
dened on U (for some chart (, U)), we have:
_

xi
,

xj
_
= 0 for
all 1 i, j n.
Denition 5.5.3 (Lie algebra, invariant vector elds). The tangent space at the identity of a Lie group, viz.
T
e
(G) is a very signicant object, often containing all the information about G, and is called the Lie algebra
of G, and denoted Lie(G) or g. Given a tangent vector X g, one can produce the left invariant vector eld
generated by X, denoted

X, by the formula:
X(x) = DL
x
(e)(X) = (L
x
)
(e)(X)
It clearly satises

X(L
y
(x)) = X(yx) = L
y
(x)(

X(x)), and also these are all the vector elds satisfying this
condition. Such vector elds are called left invariant vector elds on G. Similarly, one can dene right invariant
vector elds. Thus there is a 1-1 linear isomorphism between g and the R-vector space of all left (resp. right)
invariant vector elds on a Lie group. Similarly, we can dene left and right invariant 1-forms on G, and these
are in corespondence with the dual g
of the Lie algebra. As a consequence, one sees that the tangent and
cotangent bundles of a Lie group are trivial bundles.
To say that a vector eld on a Lie group is left-invariant is the requirement: L
g
(X) = X L
g
, and from
this and (v) of 5.5.2 above that the commutator of two left-invariant vector elds is a left-invariant vector eld,
and likewise for right- invariant vector elds. Thus the tangent space g = T
e
(G) is a real Lie algebra, i.e. is an
R-vector space with a bilinear pairing [ , ] : g g g satisfying:
(i): [X, Y ] = [Y, X]
(ii): [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0 (Jacobi identity).
If one takes the vector space (M) of all smooth or analytic vector elds on a smooth or analytic manifold,
and dene [ , ] to be the commutator (of vector elds), (M) will become an innite-dimensional Lie algebra.
Exercise 5.5.4. Let F = R or C. Show that the Lie algebra of GL(n, F) is M(n, F), which we shall henceforth
denote as gl(n, F). Show that the Lie algebra of SL(n, F) is the space of all trace zero elements in gl(n, F),
denoted sl(n, F). Compute the Lie algebras of O(n), SO(n), U(n), SU(n). (Use the result 3.2.15).
Theorem 5.5.5. Let G be a Lie group. Then there exists a smooth mapping:
exp : R g G
such that for a xed element X g, the curve t exp (tX) is a group homomorphism from the additive group
of reals to the group G. Furthermore, this smooth curve is a trajectory of the left invariant vector eld

X
generated by X.
Proof: Let X g. By applying cuto functions, local charts and the theorem 5.4.2 above, one can nd a
smooth curve c : (, ) G such that c(0) = e, and such that for t (, ), we have the equation:
dc
dt
(t) =

X(c(t))
where

X is the left-invariant vector eld generated by X. Let x = c(
2
). We rst remark that for s, t (, ),
such that s +t (, ), we have c(s +t) = c(s)c(t). This is because both the curves c(s +.) and c(s)c(.) take
the value c(s) at t = 0, and also both are trajectories for the vector eld

X. Thus by uniqueness of trajectories,
they agree wherever they are dened. For s (
2
,
3
2
) dene (s) := L
x
(c(s

2
)). We claim that c(t) = (t)
for

2
< t < . This is because for each such t, we have
dc
dt
(t) =

X(t), and also
d
dt
(t) = L
x
_
dc
dt
_
t

2
_
_
= L
x
_
X
_
c
_
t

2
___
=

X(L
x
c
_
t

2
_
)
=

X((t))
since

X is left invariant. Also, since x = c(
2
) we have (
3
4
) = L
x
c(
4
) = c(
2
)c(
4
) = c(
3
4
) by the remark in
the beginning. Thus we have extended the trajectory to (,
3
2
). We may proceed in this fashion to extend it
to all of R.
The fact that this smooth curve becomes a group homomorphism follows from the remark at the outset
(i.e. uniqueness). We denote c(t) by exp(tX), because again using uniqueness one easily shows the trajectory
corresponding to X takes the same value at time t as the trajectory corresponding to X at time t. The
smoothness of exp : Rg G follows from the fact that the trajectory of a smooth vector eld varies smoothly
if the vector eld is varied smoothly and the initial condition is varied smoothly (which is a part of 5.4.2, and
wasnt proved there). 2
Exercise 5.5.6.
(i): Verify that the exponential map dened earlier for the various linear Lie groups of Example 2.1.6 is the
exponential map dened above.
(ii): Show (by computing the derivative of exp at 0 and using the inverse function theorem) that the map
exp : g G maps a neighbourhood of 0 in g dieomorphically onto a neighbourhood of e in G. Thus we
have a natural chart around e in a Lie group (called the exponential chart, or exponential coordinates),
and by composing with left translations we get charts around each point in G.
56 VISHWAMBHAR PATI
(iii): For G = S
1
= SO(2), note that the exponential map exp : R S
1
may also be represented by t e
it
.
Show that for an irrational number , the subalgebra :
h = {(t, t) : t R}
exponentiates to the subgroup of T
2
= S
1
S
1
given by the skew line that was dened in 3.4.5.
Example 5.5.7 (Exponential map not necessarily surjective). The exponential map of a Lie group need not
be surjective. For example, let G = SL(2, R), with
sl(2, R) = {X gl(2, R) : tr X = 0}
its Lie algebra. Since any element A of SL(2, R) satises its characteristic polynomial, namely A
2
TA+I = 0,
where T = tr X, by taking the trace of this equation, we see that tr(A
2
) = 2 +T
2
2. Now, if an element
in SL(2, R) is of the form exp X for some X sl(2, R), it would follow that this matrix would be the square
of A = exp
X
2
, and hence would have trace 2. But, for example, the diagonal matrix with diagonal entries
2,
1
2
is an element of SL(2, R) of trace < 2. Hence, it isnt the exponential of anything.
However, it turns out that the exponential map of any compact connected Lie group is surjective. It is an
easy exercise, using the Jordan Canonical form, to prove that the exponential maps of GL(n, C) and SL(n, C)
are surjective. (Hint: The innite series for log(I +A) is a polynomial for A nilpotent.)
The reader may wonder whether something analogous to the exponential map constructed above exists on
all manifolds. The answer is contained in:
Proposition 5.5.8 (Flows on smooth manifolds). Let X be a compactly supported vector eld on a smooth
manifold M. Then there exists a smooth map:
: R M M
(t, x)
t
(x) = (t, x)
which satises:
(i):
0
(x) = x for all x M.
(ii):
dt(x)
dt
(0) = X(x) for all x M.
(iii): (t, x) = x for all t R and x , K = supp X.
(iv):
t+s
=
t

s
for all s, t R.
Such a family {
t
}
tR
is called a one- parameter family of dieomorphisms or ow corresponding to the vector
eld X.
Proof: We note that we have a map : [, ] M M as guaranteed by the theorem 5.4.11 above.
(For x , U, we just dene (t, x) = x for all t R.) We note that ( , ) satises (i), (ii) above, and also (iii)
for all t [, ], if we replace with . We also claim that (iv) is satised for all s, t [, ] such that
s +t [, ]. For brevity, denote (s, x) as
s
(x).
For, by denition of , (t, x) is c
x
(t), where c
x
(.) is the trajectory of the vector eld X satisfying c
x
(0) = x.
For a xed s, c
x
(s + t) is the trajectory of X satisfying c
x
(s + 0) = c
x
(s) =
s
(x). Thus by uniqueness, for
s, t, s +t [, ], we have:
c
s(x)
(t) = c
x
(s +t)
That is (t, (
s
, x)) =
s+t
(x). Now it is clear how to dene (t, x) for arbitrary t. It is just as in the theorem
5.5.5 above. Dene for t R, the positive integer N by the equation:
t = N(
2
) + where || <

2
and dene
t
:= (
2
)
N

. We leave it to the reader to verify that this satises all of the assertions above.
The reason that
t
is a dieo is that
t
is the inverse of
t
. 2
5.6. Lie Derivative. A smooth vector eld X on a manifold apriori only gives us a method of dierentiating
functions, to produce other functions. But it is possible to dierentiate other objects, such as tensors, and in
particular dierential forms with respect to X.
We rst note that even if X is not a compactly supported vector eld, we can multiply it with a cut-
o function
x
which is identically = 1 in some neighbourhood W(x) of x (see 3.1.8), and get a compactly
supported vector eld which agrees with X on W(x). Denote the ow (=1-parameter family of dieos) of this
vector by
X
t
. It is easy to see that this well-dened in a neighbourhood of x and small t, regardless of choice
of cut-o function.
Denition 5.6.1 (Lie derivative). For f
0
(M), a smooth function, dene:
L
X
f = X(f)
which clearly means, since
d
X
t
dt
(0) = X(x), that
(L
X
f)(x) = lim
t0
f(
X
t
(x)) f(x)
t
so that:
L
X
f = lim
t0
(
X
t
)
f f
t
If Y is another vector eld, then
X
t
being a dieomorphism, there exists the vector eld (
X
t
)
Y , as explained
in (v) of 5.5.2, whose value at x is precisely D
X
t
(
X
t
(x))(Y (
X
t
(x)). Thus we dene:
L
X
Y := lim
t0
(
X
t
)
Y Y
t
Similarly, for a 1-form , we dene:
L
X
:= lim
t0
(
X
t
)

t
Now we extend L
X
to tensors of type (k, l) by requiring that it acts as a derivation, that is, for two tensors
, , we have:
L
X
( ) = L
X
+ L
X
Since every tensor is locally the sum of decomposable tensors, this denition makes sense.
Remark 5.6.2. The reason for the opposite signs of t in the denitions of L
X
for 1-forms and vector elds
is that (Y ) is a function. By the denition of pullback of a form under a map f, we have [(f
)](x)Z =
(f(x))(Df(x)Z) for a tangent vector Z T
x
(M). Thus if we let the vector eld Z := (
X
t
)
Y , then Z(x) =
D
X
t
(
X
t
(x))Y (
X
t
(x)). Thus the number (
X
t
)((
X
t
)
X)(x) is the same as the number (

X
t
(x))(D
X
t
D
X
t
)(
X
t
(x))Y (
which is exactly (
X
t
(x))(Y (
X
t
x). This is, of course, the function
X
t
((Y )). In other words, we have the
relation of functions:
((
X
t
)
)(
X
t
Y ) = (
X
t
)
((Y ))
and the formula for the Lie derivatives of vector elds, 1-forms, and functions match up to give:
L
X
((Y )) = (L
X
)(Y ) + (L
X
Y )
Proposition 5.6.3 (Lie derivative of a vector eld). For smooth vector elds X, Y on a smooth manifold M,
we have L
X
Y = [X, Y ].
Proof: For brevity, let us denote
X
t
by
t
. For a smooth function f, and a point x M, we have:
[(
t
Y )(x)]f = [D
t
(Y (
t
(x))]f
= Y (
t
(x))(f
t
)
58 VISHWAMBHAR PATI
so that :
[(
t
Y )(x)](f
t
) = Y (
t
(x))f
so that,
(Y
t
)f Y (f
t
) = (
t
Y Y )(f
t
)
= t(L
X
Y )(f
t
) +o(t
2
)
= t(L
X
Y )f +o(t
2
)
where the second equality results from the denition of Lie derivative, and the third from the fact that f
t
=
f + tXf + o(t
2
). Adding and subtracting Y f to the left hand side of the rst equation above, the equation
above may be rewritten as:
_
(Y f)
t
Y f
t
_
_
Y (f
t
) Y f
t
_
= (L
X
Y )f +o(t)
Taking the limit as t 0, we get L
X
Y = X(Y f) Y (Xf) = [X, Y ]f. 2
Corollary 5.6.4 (The adjoint action of a Lie group). Let G be a Lie group, and let X g = T
e
(G). Then
we have the map Ad(exp(tX)) : G G which takes g to exp(tX)g(exp(tX). Its derivative Ad(exp tX)
=
D(Ad(exp tX) is an automorphism of g. We have the equation:
lim
t0
Ad(exp tX)
Id
t
= ad X := [X, ]
Further, if we consider the map:
: GG G
(x, y) xyx
1
y
1
Then
1
2
(D
2
)(e, e) : g g g
is the map (X, Y ) [X, Y ].
Proof: Let Y g, and let

Y (resp.

X) denote the left-invariant vector elds corresponding to Y (resp. X).
Note that exp(tX) is the trajectory of the vector eld

X passing through the identity element e at t = 0.
This follows easily from the fact that the exponential map is a homomorphism. We need to see what the
trajectories passing through other points are. We claim that R
exp tX
(x) = x. exp(tX) is the trajectory of

X
passing through x at t = 0. This is because:
d(xexp tX)
dt
(0) = L
x
_
d exp tX
dt
(0)
_
= L
x
(X) = L
x
(
X(0)) = X(x)
by the left invariance of

X. Thus, the ow of

X is given by
X
t
= R
exp tX
, which we abbreviate as
t
.
Consequently:
(
t,
Y )(e) = DR
exp(tX)
Y (e. exp tX) = DR

exp(tX)
Y (exp tX)
= DR
exp(tX)
DL
exp tX
(Y ) = D(R
exp(tX)
L
exp tX
)(Y )
= D(Ad(exp tX))Y = Ad(exp tX)
Y
since the left invariance of

Y implies

Y (exp tX) = DL
exp tX
(Y ).
From the previous proposition it follows that:
[X, Y ] = [
X,
Y ] = L
X
Y = lim
t0
Ad(exp tX)
Y Y
t
which proves the rst assertion.
The second assertion follows easily from the rst by noting that (x, y) = xyx
1
y
1
= (L
x
(Ad y)x
1
),
and is left as an exercise (using (ii) of 5.6.6 below). 2
Lemma 5.6.5 (Characterisation of one-parameter subgroups). A smooth homomorphism:
: R G
into a Lie group G (or sometimes the image of ) is called a 1-parameter subgroup of G. We claim that is a
one-parameter subgroup of G i (t) = exp tX for some X g = T
e
(G).
Proof: The if part is clear. So suppose : R G is a one parameter subgroup of G. Then let us denote:
X =
d
dt
(0)
which is an element of G. We claim that (t) is a trajectory of the left invariant vector eld

X. This is because:
d
dt
(s) = D
_
d
dt
|s
_
=
d(s +t)
dt
(0) =
d((s)(t))
dt
(0)
=
d(L
(s)
((t))
dt
(0) = (L
(s)
)
_
d(t)
dt
_
(0)
=

X((s))
Further, (0) = e. The curve (t) := exp tX is also a trajectory of

X which satises (0) = e. By the
uniqueness of trajectories, it follows that . 2
Corollary 5.6.6. For a Lie group G:
(i):
(Ad exp tX)
= exp t(ad X)
where ad X gl(g) is the linear map Y [X, Y ], so that its exponential is the power series
i=0
(adX)
i
i!
,
an element of GL(g).
(ii): The following formula holds:
(Ad exp tX)(exp sY ) = (exp tX)(exp sY )(exp tX)
1
= exp(s exp (t ad X)Y )
Note that exp (t ad X) is an element of GL(g), so exp (t ad X)Y g, and its exponential is an element of
G.
(iii): G is an abelian Lie group i g is an abelian Lie algebra. That is [X, Y ] 0 for all X, Y g.
Proof: To see (i), we consider the 1-parameter subgroup:
: R GL(g)
t (Ad exp tX)
Note that
d
dt
(0) = lim
t0
(Ad exp tX)
I
t
= ad X
by the corollary 5.6.4 above. Thus, by the lemma 5.6.5 above, we have the result.
To see (ii), x a t and X. For a general Y g consider the one-parameter subgroup in G:
: R G
s Ad(exp tX)(exp sY ) = (exp tX)(exp sY )(exp tX)
1
Clearly, from (i) it follows that
d
ds
(0) = (Ad exp tX)
(Y ) = (exp tad X)Y , by the foregoing, so that by the

lemma 5.6.5 above, it follows that:
(s) = exp(s exp(t adX)Y )
thus proving (ii).
60 VISHWAMBHAR PATI
For (iii), note that the only if part is trivial, because the adjoint action of G on itself is the identity map,
and so its derivative (Ad exp tX)
is the identity map for all t, and so by (i) of 5.6.4, adX = 0 for all X, and
thus g is abelian.
Conversely, if g is abelian, the part (ii) proved above shows that
(Ad exp tX)(exp sY ) = (exp tX)(exp sY )(exp tX)
1
= exp(sY )
for all s, t R and X, Y g.
This shows that all elements in the image of exp : g G commute. By the inverse function theorem, (see
(ii) of 5.5.6) every element in a neighbourhood of e is in this image, so all elements in a neighbourhood of the
identity commute. Since every connected topological group is generated by any neighbourhood of the identity,
G is abelian. 2
Exercise 5.6.7 (Derivative of a Lie homomorphism). Let : G H be a smooth homomorphism of Lie
groups. Show that there is a unique linear map of Lie algebras, which is denoted
:= D(e) from g to
h, which is a homomorphsim of Lie algebras, i.e. [
X,
Y ] =
[X, Y ]. Show that if , are two Lie

homomorphisms from G to H, and G is connected, then
implies = . (Hint: Just look at one

parameter subgroups of G.)
5.7. Homogeneous manifolds. Let G be a Lie group of dimension n, and H G be a closed subgroup. It
is a theorem due to Cartan-Lie (see [BMST], p.58 Thm 4.5 for the case when G is a linear Lie group) that H is
also a Lie group, say of dimension m. Look at the space of left cosets, denoted G/H. This is just the space of
orbits of the smooth right action of H on G by multiplication. It is given the quotient topology. Since G can
be given a right G-invariant Riemannian metric (see example 5.1.12), this metric will also be right H-invariant.
Now let d be the distance on G, in accordance with 5.2.3, which will also be right H-invariant by 5.2.5. That
is, it will satisfy d(ah, bh) = d(a, b) for all a, b G, and h H.
Because H is closed, one can dene a distance on G/H as:
(aH, bH) := inf
hH
(d(a, bh)) = d(a, bH) = inf
h,h
H
d(ah, bh
) = (bH, aH)
Using the right H-invariance of the distance d, it is an easy exercise to check that denes a distance on G/H,
which therefore makes it Hausdor and paracompact. It is second countable (=separable) because the image
of a countable dense subset G under the natural quotient map : G G/H is a countable dense subset
of G/H.
Proposition 5.7.1 (Homogeneous manifolds). For a Lie group G and a closed subgroup H, the space G/H,
as dened above, is a smooth manifold of dimension dim Gdim H = n m.
Proof:
Note that the space M := G/H is homogeneous in the following sense: given any point x = gH in M, there
is a homeomorphism, also denoted L
g
, taking the identity coset o = H to x = gH. Thus if we can prove that
a neighbourhood of the identity coset o is homeomorphic to an open subset of R
nm
, and also show that the
coordinate changes are smooth, we will be done.
Let h g be the Lie algebra of H. It is a fact (see [BMST], p.58) that the exponential map exp : g G
takes h to H. Let m be a vector space complement to h inside g, so that g = h m. Dene the map:
: mh G
(X, Y ) exp X exp Y
It is easily checked that the derivative D(0) is the identity map g g so that, by the inverse function
theorem, there exists a neighbourhood

U of 0 in g on which is a dieomorphism. Let (U
1
) (
U) be a
smaller neighbourhood such that it is symmetric (i.e. such that y (U
1
) y
1
(U
1
), and such that
(U
1
)(U
1
) (
U). Thus it follows that if y

1
, y
2
(U
1
), we have y
1
1
y
2
(
U). Now take a neighbourhood

of the form WV U
1
, where W and V are neighbourhoods of 0 in m and h respectively. Let U := (WV ) =
exp(W). exp(V ). We claim that (U) is an open neighbourhood of the identity coset o = eH in G/H.
Recall that with the quotient topology on G/H, a set is open in G/H i its inverse image under is open
in G. But
1
((U)) is just
hH
R
h
(U), which is clearly open. Consider the map : W (U) dened by
(X) (exp X)H. That is, =
|m
. Clearly is continuous.
We claim that its inverse is the continuous map:
: (U) W
(y) pr
m
1
(y)
To see that this map is well-dened, note that if (y
1
) = (y
2
) with y
1
, y
2
U, then by writing y
1
=
(X
1
, Y
1
) and y
2
= (X
2
, Y
2
) with X
i
W and Y
i
V , we have (exp X
1
)
1
exp X
2
(
U) H, so that
exp X
2
= exp X
1
exp Z for some Z

U h. That is (X
2
, 0) = (X
1
, Z). Since was injective on

U, it
follows that X
1
= X
2
= X and Z = 0. Thus y
i
= exp X exp Y
i
for i = 1, 2. Thus
1
(y
i
) = (X, Y
i
), and
pr
m
1
(y
i
) = X is well dened.
62 VISHWAMBHAR PATI
Hence (
1
, V ), where V := (U) is a local chart around o. Local charts around other points are given
by composing with left translations. That is, for a g G, we dene gV = L
g
(V ) = (gU), and
1
g
by the
diagram:
V
L
1
g
gV
1

1
g
W W
Thus on the overlap W
g
:=
1
(V gV ), the coordinate change
1
g
is the composite:
W
g
V gV
L
1
g
V g
1
V

1
W
g
1
which we denote by . Again, since = , we conclude from the diagram:
V gV
L
1
g
V g
1
V

U gU
L
1
g
U g
1
U

W
g
W
g
1
that the composite is also the composite
1
L
1
g
of the above diagram, (where

L
g
denotes left translation
on the group G. This is clearly smooth.
Thus we have dened a smooth atlas, and X = G/H is a smooth manifold of dimension n m. 2
Example 5.7.2 (G-manifolds as homogeneous spaces). Let M be a smooth manifold of dimension m, and let
G be a Lie group of dimension n acting smoothly on it. (See denition 5.1.9). We dene the isotropy group
(or stabiliser) of x to be the subgroup of G given by:
G
x
= {g G : gx = x}
Recall that the action is said to be transitive if M consists of a single orbit, i.e. M = Gx for each x M.
We call the action a good action if it is a transitive action, and if for some x M, the derivative at the identity
e G of the map
x
: G M dened by
x
(g) = gx, namely:
D
x
(e) : T
e
(G) T
x
(M)
is surjective. Since we have a commutative square:
T
e
(G)
Dx(e)
T
x
(M)
L
g
L
g
T
g
(G)
Dx(g)
T
gx
(M)
and L
g
: G G and L
g
: M M are dieomorphisms, it follows from the square above that the lower
horizontal arrow is surjective for all g G, i.e. each point of M is a regular value of the map
x
.
Thus
1
x
(x) = G
x
is a closed submanifold of G, i.e. a closed Lie subgroup of G of dimension n m, by the
(submersive) implicit function theorem 1.6.4.
Claim: M is dieomorphic to the homogeneous manifold G/G
x
of 5.7.1.
For, x a point x M, and again consider the map
x
. Since
x
is continuous, and takes gh and g to the
same point gx for each h G
x
, it descends to a continuous map : G/G
x
M by the property of the quotient
topology. This map is easily checked to be a bijection. On the other hand, since D
x
is surjective at each
point g G, the (submersive form of) the implicit function theorem implies that
x
is locally a projection,
and hence an open map. Thus is also open, and we have that is a homeomorphism.
By the example 3.2.15, the tangent space to G
x
at e is the kernel:
ker D
x
(e) : T
e
(G) T
x
(M)
Thus, if m is the vector space complement to T
e
(G
x
) that was chosen in the proposition 5.7.1 above, it follows
that D
x
(e)
|m
is an isomorphism. That is, the derivative of the map:
m
exp
G
x
M
at 0 m, is an isomorphism. Hence, by the inverse function theorem, this composite is a local dieomorphism.
But on a neighbourhood W of 0 in m, this map is also the composite:
W

V

M
where is the local chart around o G/G
x
constructed in 5.7.1. Thus is smooth around o G/G
x
, and
in fact a local dieomorphism of a neighbourhood V of o with a neighbourhood of x in M. By applying left
translations on G/G
x
and M, we see that is a local dieomorphism at all points of G/G
x
. Since is a
homeomorphism which is a local dieomorphism at all points, it is a dieomorphism. This proves the claim.
2
Exercise 5.7.3. Show that the natural left action of G on G/H as dened in 5.7.1 is a smooth and good
action.
One can obviously dene a right G-action in analogous fashion, and carry all of the above over to right actions.
Given any left action, one can get a right action by dening xg := g
1
x, and vice versa, so that everything
above can be reformulated for right actions. On a Lie group, there is the left action of left multiplication, and
the right action of right multiplication, and these actions are distinct if the group is non-abelian. Indeed, the
adjoint action of G on itself (which is a left action) measures the discrepancy between the left and right
actions, as was discussed in the preceding subsection.
5.8. Homogeneous metrics. In the last subsection, we saw two alternative descriptions of a homogeneous
manifold, i.e. as a coset space of a Lie group with respect to a closed subgroup, or as a manifold with a smooth
good left-action by a Lie group G.
Note that the Riemannian metric we got on G/H at the beginning of the last subsection need not be left
G-invariant. It is quite natural to ask how one can put a left G-invariant (= G-homogeneous) Riemannian
metric on such a manifold.
Proposition 5.8.1 (Homogeneous metrics). Let G and H be as in 5.7.1. Assume there exists a Riemann-
ian metric , which is left G-invariant and right H- invariant. Then there exists a (left) G-invariant (or
homogeneous) Riemannian metric on G/H.
Proof: Consider the quotient map, which is a submersion:
: G G/H
g gH
which induces the exact sequence of tangent bundles:
: TG
T(G/H) 0
The kernel of this bundle morphism is the bundle G, whose bre
g
at g G is given as the kernel of
D(g) : T
g
(G) T
(g)
(G/H). Let
g
be the orthogonal complement of
g
in T
g
(G), with respect to the given
Riemannian metric , . It follows from the discussion of 5.7.1 that
,g
:
g
T
(g)
(G/H) is an isomorphism
for each g G, and so
|
:
T(G/H)
is an isomorphism of vector bundles. Thus its inverse gives an isomorphism of
T(G/H) with the subbundle

of TG.
64 VISHWAMBHAR PATI
Notice that both bundles and
T(G/H) are mapped to themselves by the derivative of the right translation

R
h
, for h H. Further, by right H-invariance of , , the isomorphism above commutes with R
h
. (All this
is a rerun of part (ii) of exercise 5.1.3.) Thus, by restricting the left G-invariant Riemannian metric on TG, we
have a left G-invariant Riemannian metric (call it ( , )) on the bundle
T(G/H). Because the original metric

was invariant under R
h
for all h H, this induced metric ( , ) is also right H-invariant. Identifying with
T(G/H), we have an orthogonal decomposition

TG
T(G/H)
which is preserved under R
h
for all h H.
Now, the restricted bundle
T(G/H)
|gH
is a trivial bundle, and for h H, (R
h
) carries the bre of this
bundle at g isomorphically to the bre at gh. Since the metric ( , ) is invariant under R
h
, this metric naturally
descends to a Riemannian metric on T(G/H), which is left G-invariant. 2
Corollary 5.8.2. If G is a Lie group, and H is compact, then there always exists a left G-invariant Riemannian
metric on G/H. In particular, if G is itself compact, then we have such a metric.
Proof: H being compact, a positive denite bilinear form on g (which automatically yields a left G-invariant
Riemannian metric on G) can be averaged out with respect to the adjoint H-action Ad(h)
for h H, as in
the proposition 5.1.13, to yield a left G right H invariant metric on G. 2
Exercise 5.8.3. Show that if G is a Lie group, and H a connected closed subgroup, and , is a positive
denite bilinear form on g which satises:
(ad Z)X, Y +X, (ad Z)Y = 0
for all Z h, and all X, Y g, then the homogeneous manifold G/H acquires a left G-invariant Riemannian
metric. (Hint: use the formula for Ad(exp(tZ))
and the condition for an H-bi-invariant metric discussed in

example 5.1.12).
Example 5.8.4 (The Sphere). It is easy to realise the unit sphere S
n
R
n+1
as a homogeneous space as
follows. Consider natural action of the SO(n + 1) on R
n+1
given by left matrix multiplication with a column
vector. Since SO(n + 1) preserves lengths, this action restricts to an action of SO(n + 1) on S
n
. If we let
x = (0, 0, .., 1) be the unit vector of the x
n+1
-axis (the north pole of S
n
), we can always nd a matrix
A SO(n + 1) such that Ax = y for any given y S
n
(why?), so that this action is transitive. It is easy to
check that it is smooth. To see that it is a good transitive action, one just has to verify that the map:
x
: SO(n + 1) S
n
which maps a matrix A SO(n + 1) to its last column Ax has derivative of full rank at the identity. In a
neighbourhood U of the identity, we write A(t) = exp tX (by (ii) of 5.5.6) where X is a skew-symmetric matrix
of trace 0, and thus:
D
x
(e)(X) = lim
t0
_
(exp tX)x x
t
_
= Xx
Note that X being skew-symmetric, Xx, x = X
t
x, x = x, Xx which implies Xx is orthogonal to x,
and hence lies in T
x
(S
n
) (see exercise 3.3.3 above). Since any tangent vector v T
x
(S
n
) is just a vector with
last entry zero, one can view it as the last column of the skew-symmetric matrix X which has v as the last
column, v as the last row, and zeros elsewhere. Thus v = Xx, and our map is surjective, and the transitive
action is good. Clearly, the isotropy subgroup of x is the group SO(n) on the rst n coordinates. By the
foregoing example 5.7.2, the sphere S
n
is dieomorphic to the homogeneous space SO(n + 1)/SO(n).
There is the natural the bi-invariant metric arising from X, Y = tr(X
t
Y ) on SO(n + 1). With respect to
this metric, the Lie algebra o(n + 1) of SO(n + 1) decomposes as an orthogonal direct sum:
o(n + 1) = o(n)
m
where
m :=
_
_
_
_
_
_
_
_
0 0 0 ... v
1
0 0 0 ... v
2
.. .. .. .. ....
.. .. .. .. v
n
v
1
v
2
... v
n
0
_
_
_
_
_
_
: v
i
R
_
_
For two elements X, Y m, their inner product X, Y = trX
t
Y is just twice the euclidean inner product of
their non-vanishing (last) columns, which is exactly twice the inner product on R
n
= T
o
(S
n
), where o is the
identity coset x = (0, 0, .., 1). Thus the natural metric on S
n
is (twice) the one coming from the proposition
5.8.1. (Since both metrics are left SO(n + 1)-invariant, they are completely determined by their values at
x = o).
Example 5.8.5 (Grassmannians). The orthogonal group O(n) acts on the grassmannian G
k
(R
n
), as explained
in 5.1.7. That this action is transitive and, in fact, good is easily veried. The isotropy group of the point
o = span(e
1
, .., e
k
) is the subgroup H = O(k) O(n k). Again, the bi-invariant Riemannian metric coming
from the positive denite bilinear form X, Y = tr(X
t
Y ) yields an orthogonal decomposition:
o(n) = h
m
where
m =
__
0 X
t
X 0
_
: X hom(o, o
)
_
The natural identication of T
o
(G
k
(R
n
) with m, and the fact that the restriction of the metric tr(X
t
Y ) to m
is exactly twice the Riemannian metric introduced on T
o
(G
k
(R
n
)) in 5.1.7, proves that metric to be equal to
twice the one coming from 5.8.1, by left SO(n) invariance of both these metrics.
Example 5.8.6 (Hyperbolic 2-space). We let H
2
denote the upper-half plane in R
2
= C, dened as:
H
2
= {z C : Imz > 0}
On this 2-dimensional manifold we put the Riemannian metric given by:
g(z) = g(x +iy) =
dx
2
+dy
2
y
2
There is an action of the group G = SL(2, R) on H
2
dened as follows:
: SL(2, R) H
2
H
2
(
_
a b
c d
_
, z)
az +b
cz +d
Clearly this makes sense, since it is easily checked that:
Im
_
az +b
cz +d
_
= |cz +d|
2
Imz
which is > 0 if Imz > 0. It is also easily checked to be smooth. It is transitive because the element:
_
y
1
2
xy
1
2
0 y
1
2
_
takes i =
1 to x +iy H
2
. We need to check that the action is a good action. That is, the map:
:=
i
: G H
2
dened by (g) = gi has surjective derivative at the identity (see example 5.7.2 above). Consider the matrix
in g = sl(2, R) given by:
X =
_
0
0
_
66 VISHWAMBHAR PATI
Clearly
exp tX =
_
e
t
0
0 e
t
_
so that (exp tX)i = e
2t
i, so that
(D(e))(X)i =
d
dt t=0
(e
2ti
) = 2i
Similarly, consider the matrix Y sl(2, R) dened by:
Y =
_
0
0
_
so that
exp tY =
_
cosh t sinh t
sinh t cosh t
_
so that
(D(e))(Y ) =
d
dt t=0
(exp tY )i = 2
Thus D(e) has image C = T
i
(H
2
) and is surjective. To identify H
2
as a G-space with G = SL(2, R) we just
need to determine the isotropy subgroup of the point i, which is easily checked to be the subgroup:
K := SO(2) =
__
cos sin
sin cos
_
: R
_
which is just the circle group. Thus, as a manifold H
2
= SL(2, R)/SO(2).
We now claim that the group G = SL(2, R) acting on H
2
by left translations is acting by isometries. For
this, note that the metric is given by:
g(z) =
dzdz
(Imz)
2
If we let g =
_
a b
c d
_
SL(2, R) and denote gz =
az+b
cz+d
by w, we compute:
dw =
dz
(cz +d)
2
Since
Imw = Im
_
az +b
cz +d
_
= |cz +d|
2
Imz
it follows that
dzdz
(Imz)
2
=
dwdw
(Imw)
2
so that G is acting by isometries, and the hyperbolic 2-space H
2
has a G invariant metric.
The next question is obviously whether there is a left-G and right-K invariant metric on G = SL(2, R)
which descends to the hyperbolic metric above, in accordance with proposition 5.8.1.
As usual, dene the bilinear form X, Y = tr(XY
t
). This will lead to a left G-invariant Riemannian metric
on G. All we need to do is check that it is right K-invariant, and in fact K-bi-invariant. But this is obvious,
because for g K, g
1
= g
t
, and Adg(X) = gXg
t
for g K. Thus
X, Y = tr(XY
t
) = tr(gXY
t
g
1
) = tr(gXg
t
gY
t
g
t
) = Adg(X), Adg(Y )
proving left-G and right-K invariance of the resulting Riemannian metric on G, and by proposition 5.8.1, we
get a left G-invariant (homogeneous) Riemannian metric on G/K = H
2
.
Now, we just need to verify that this metric coincides with the hyperbolic metric above on H
2
. For this we
note that g has the polar decomposition
sl(2, R) = k p
where
k = so(2) =
_
X sl(2, R) : X = X
t
_
, p =
_
X sl(2, R) : X = X
t
_
and by our proposition 5.7.1, the vector space p gets identied with the tangent space T
i
(H
2
). The identication
is via D
i
(e) which, from the rst para above is the mapping
D
i
(e) : m T
i
(H
2
)
X :=
_

_
2i + 2
which shows that the tr(XX
t
) = 2(
2
+
2
) = 2 (, ), (, ), which is exactly half the square of the hyperbolic
length of the vector D(e)(X) T
i
(H
2
). Thus, as before, the G-invariant Riemannian metric from 5.8.1 is the
hyperbolic metric upto a scalar multiple.
Remark 5.8.7. We remark that the above time-tested positive denite bilinear form X, Y = tr(XY
t
) is not
AdG invariant, as is seen from the example 5.1.12. In fact:
Claim: There does not exist any positive denite AdG invariant bilinear form on g = sl(2, R).
Proof:
Note that there is the decomposition:
g = sl(2, R) = RY RH RX
where
H =
_
1 0
0 1
_
X =
_
0 1
0 0
_
Y =
_
0 0
1 0
_
and it is easily calculated that [H, X] = 2X, [H, Y ] = 2Y, [X, Y ] = H. We recall that, as in the exercise
5.8.3, an AdG invariant bilinear form , would have to satisfy:
0 =
_
d
dt
_
t=0
Ad exp tX(Y ), Ad exp tX(Z) = [X, Y ], Z +Y, [X, Z] for all X, Y, Z g
So that we would have:
2 X, X = [H, X], X = X, [H, X] = 2 X, X
which would imply X, X = 0 and thus , is not positive denite. (In fact, it is an exercise to show that
any non-degenerate symmetric bilinear form on g = sl(2, R) is a non-zero multiple of the (non-degenerate
but indenite) symmetric bilinear form A, B = tr(AB). This last form is indenite because I, I = 2, and
X Y, X Y = 2, where X, Y sl(2, R) are the matrices dened above.) This proves the claim. 2
Example 5.8.8 (Disc Model of Hyperbolic 2-space). There is another useful Poincare Disc Model of hyper-
bolic space. We equip the open unit disc:
D
2
= {z C : |z| < 1}
with the Poincare metric:
g(z) = g(x +iy) = 4
dx
2
+dy
2
(1 x
2
y
2
)
2
=
4dzdz
(1 |z|
2
)
2
This Riemannian manifold is isometric to H
2
above via the Cayley transform mapping:
: H
2
D
2
z
z i
z +i
The map
1
maps w to i
_
1+w
1w
_
. The imaginary axis in H
2
maps to the real axis in D
2
, and the distinguished
point (identity coset) i =
1 H
2
maps to the origin in D
2
. It is a straightforward calculation to verify that
is an isometry. The good transitive action of SL(2, R) gets replaced by a good transitive action of SU(1, 1)
on D
2
. The group SU(1, 1) is dened by:
SU(1, 1) =
__
a b
b a
_
GL(2, C) : |a|
2
|b|
2
= 1
_
68 VISHWAMBHAR PATI
and this acts on D
2
by z
az+b
bz+a
. Again, one checks that this action leaves the Poincare metric above invariant.
The isotropy of the origin is the circle subgroup S
1
SU(1, 1) of elements with b = 0 and |a| = 1.
Exercise 5.8.9 (Equivalence of both models). Construct the explicit isomorphism between SL(2, R) and SU(1, 1).)
Example 5.8.10 (Hyperbolic 3-space). Let G = SL(2, C) and K = SU(2). View R
4
as the non-commutative
(so be careful in the calculations below!) algebra of quaternions. Let H
3
be the subspace of H = R
4
dened
by:
H
3
= {x +iy +jw +kt H : t = 0, w > 0}
This is the upper-space of R
3
R
4
, and is a Riemannian manifold with the metric given by:
g(x +iy +jw) =
dx
2
+dy
2
+dw
2
w
2
We claim that H
3
is the homogeneous manifold SL(2, C)/SU(2), and the metric above is G = SL(2, C)
invariant. First let us dene a transitive action of SL(2, C) on H
3
as follows:
: GH
3
H
3
(
_
a b
c d
_
, r) (ar +b)(cr +d)
1
(Exercise: show that cr + d ,= 0 for r H
3
). Here we are regarding C as the subalgebra of H consisting of
quaternions of the type x + iy. Let us see that this is well-dened. Regard a quaternion as z + jw where
z, w C. Note that the quaternion multiplication rules imply that jz = zj for z C, and hence that:
(z +jw)(z
1
+jw
1
) = (zz
1
ww
1
) +j(wz
1
+zw
1
)
There is also the conjugation involution of H which takes r = x +iy +jw +tk to r = x iy jw tk. If we
express r = z + jw with z, w C, then r = z jw = z wj. It is easily checked that this conjugation is an
anti- homomorphism, i.e.:
rs = s r
for all r, s H. For any quaternion r ,= 0, it is easy to check that its inverse r
1
is just
r
r
2
.
To see that s = (ar +b)(cr +d)
1
H
3
for r H
3
, note that for r = z +jw H
3
, we have 2w = jr rj.
So we have to compute the quantity js sj. Since
s =
(ar +b)(r c +d)
|cr +d|
2
so that
s =
(cr +d)(r a +b)
|cr +d|
2
from which we compute (using brute force and ad bc = 1) that
js sj =
jcrb +jdr a br cj ardj
|cr +d|
2
=
2w
|cr +d|
2
which is clearly positive for w > 0. If we denote s = q +ju, then this formula implies that:
u =
w
|cr +d|
2
Since the matrix inverse:
_
a b
c d
_
1
=
_
d b
c a
_
it follows that:
w =
u
|a cs|
2
and hence that:
u
2
w
2
=
|a sc|
2
|cr +d|
2
()
an equation which we use later.
To see that the action is good, is a simple verication as in the case of H
2
above. As before, let denote
the map G H
3
taking g to g.j. Then for
X =
_
0
0
_
R > 0
we nd that
D(e)(X) =
_
d(exp tX.j)
dt
_
(0) =
_
d(e
2tj
)
dt
_
(0) = 2j
whereas for
Y =
_
0
0
_
, and Z =
_
0 i
i 0
_
, R
we have:
D(e)(Y ) =
_
d(exp tY.j)
dt
_
(0)
=
_
d(cosh tj + sinh t)(j sinh t + cosh t
dt
_
(0)
=
_
d(sinh 2t +j)
dt
_
(0) = 2
and similarly D(e)(Z) = 2i, which clearly shows that D(e) maps the subspace p sl(2, R) surjectively onto
T
j
(H
3
). Here p denotes the subspace of g = sl(2, R) consisting of traceless hermitian matrices.
For r = z +jw H
3
, where w R is strictly positive, the matrix:
_

w
z
w
0
1
w
_
takes j H
3
to r = z +jw. Thus the action is transitive, and by the foregoing, a good action.
To compute the isotropy of j H
3
, we have the equation:
(aj +b)(cj +d)
1
= j
which implies that aj +b = jcj +jd = c +dj so that b = c and a = d. But this group:
__
a b
b a
_
: a, b C
_
is precisely the subgroup SU(2) SL(2, C). Thus H
3
is dieomorphic to the manifold SL(2, C)/SU(2) by
5.7.2.
Finally, to check that the action leaves the metric above invariant, note that on dierentiating the equation
s(cr +d) = ar +b) we obtain ds(cr +d) +sc(dr) = adr, which implies:
(a sc)dr = ds(cr +d)
The conjugate of this equation is
dr(a sc) = (cr +d)ds
and multiplying the two equations, and noting that real quantities commute with everything, we get:
|a sc|
2
drdr = |cr +d|
2
dsds
and using () above we have that:
drdr
w
2
=
dsds
u
2
which implies the invariance of the metric since drdr = dx
2
+dy
2
+dw
2
.
70 VISHWAMBHAR PATI
The nal issue is whether this metric comes from a left-G and right-K invariant Riemannian metric on
G = SL(2, C) in accordance with 5.8.1. As in the example of SL(2, R) in 5.8.6 above, we have a direct sum
decomposition of Lie algebras:
g = sl(2, C) = su(2) p
where su(2) is the Lie algebra k of K, consisting of all traceless skew-hermitian complex matrices (i.e. X
=
X), and p those of tracelss hermitian ones. Under D(e), p gets identied with T
j
(H
3
). We leave it as an
exercise for the reader to prove that the positive denite bilinear form X, Y = tr(XY
) (where Y
:= Y
t
) on
g = sl(2, C) is AdK bi-invariant, and the corresponding homogeneous metric resulting from 5.8.1 on M = H
3
is a scalar multiple of the above hyperbolic one.
Both the hyperbolic spaces above are examples of Riemannian symmetric spaces of non-compact type.
5.9. Classication of compact surfaces. The hyperbolic 2-space above is the source of a great deal of
beautiful mathematics, unifying Lie theory, Riemannian geometry, complex analysis and number theory. For
example, there is the following fundamental:
Theorem 5.9.1 (Riemann mapping theorem). A simply connected complex manifold of (complex) dimension
one is biholomorphically (or conformally) equivalent to C, H
2
or S
2
. (Here S
2
is the Riemann sphere C,
and can be made into a complex manifold of dimension one in a natural way. It can also be viewed as the
complex projective space CP(1).)
The reader may consult Ahlfors book Complex Analysis, or Rudins Real and Complex Analysis for a proof
of this deep fact. The original proof given by Riemann was a heuristic one, and it took the better part of a
century for a rigorous proof (due to Ahlfors).
As a consequence, one can classify (upto dieomorphism) all compact connected smooth manifolds of (real)
dimension 2. First one classies the orientable ones. If M is a compact connected orientable manifold, one can
give it a smooth Riemannian metric, and then prove the existence of local isothermal parameters. That is, in
a neighbourhood of every point, there is a coordinate system (u, v), with respect to which the metric g takes
on the form (u, v)(du
2
+ dv
2
), where is a strictly positive function. This remarkable fact is due to Gauss,
and a proof maybe found in Springers Riemann Surfaces.
If on an overlapping region, we have isothermal coordinates (x, y), then the reader can easily verify that the
relation (u, v)(du
2
+dv
2
) = (x, y)(dx
2
+dy
2
), coupled with the extra fact that the jacobian of the coordinate
change (u, v) (x, y) is positive (which can be assumed without loss of generality by the orientability of M)
shows that this coordinate change obeys the Cauchy-Riemann equations:
u
x
=
v
y
u
y
=
v
x
on the overlap, which makes M a comlex manifold of dimension 1.
By the Riemann mapping theorem above, its simply-connected covering

M is therefore either CP(1), C or H
2
,
and the group of covering transformations must be a subgroup of the group of holomorphic automorphisms
Aut
h
(
M).
For the reader who is familiar with a bit of one-variable complex analysis, it is not dicult to show that:
Aut
h
(CP(1)) = PGL(2, C) = GL(2, C)/{I}
Aut
h
(C) = {z z + : C
, and C}
Aut
h
(H
2
) = PSL(2, R) = SL(2, R)/{I}
where the groups PGL(2, R) and PSL(2, R) act via fractional linear transformations z
_
az+b
cz+d
_
.
So now the question boils down to nding out which subgroups of the above automorphism groups
Aut
h
(
M) can act properly discontinuously on the complex manifolds

M above. Per force, such subgroups
have to be discrete subgroups. In fact, by covering space theory, it will turn out that is isomoprhic to
the fundamental group
1
(M). Suitably classifying these discrete subgroups of the various automorphism
groups listed above will lead to a classication of compact connected complex manifolds of dimension 1, upto
holomorphic or conformal equivalence.
This is a deeper question than what we are interested in. The fact that a compact connected orientable
surface becomes a complex manifold of dimension 1 leads to the fact that it is triangulable, that is, it is a
two dimensional simplicial complex. This fact is proved in the book Riemann Surfaces by Ahlfors and Sario.
Now, using triangulability, one can give combinatorial arguments to show that a compact connected ori-
entable surface M is homeomorphic to one of the following:
(i): S
2
, if M is simply connected.
(ii): T
2
#T
2
#...#T
2
(g-times), if M is not simply connected.
Here # denotes connected sum. That is, for two surfaces X, Y , their connected sum X#Y is dened as the
surface obtained by removing an open disc from each, thus yielding X
1
= X \ D
2
and Y
1
= Y \ D
2
, and then
gluing X
1
and Y
1
along the boundary circles of the holes. One has to check that the homeomorphism type of
this object is well-dened. The number g is called the genus of M, and is a topological invariant. The surface
of genus g above is often called a g-handle.
The proof of this topological classication maybe found in W. Masseys book Introduction to Algebraic
Topology. As it turns out, the smooth (dieomorphic) classication is the same as the topological classication,
because each of the surfaces listed above admits a unique dierentiable structure. (See Hirschs Dierential
Topology, Chapter 9, for a direct classication of smooth orientable surfaces using Morse Theory).
The sphere S
2
is already simply connected. The simply connected cover of the genus-1 surface, i.e. the
torus T
2
, is R
2
= C, and in this case is any subgroup of C of rank 2 (e.g. Z+iZ), called a lattice. In fact all
lattices in C are of the form = Z+jZ where j is a complex number such that Imj > 0. It turns out that for
two lattices = Z+jZ and
= Z+j
Z the tori C/ and C/
are biholomorphically equivalent i j
=
aj+b
cj+d
for some
_
a b
c d
_
SL(2, Z). The simply connected cover of all compact connected orientable surfaces of
genus 2 is the upper half plane H
2
.
Each of the surfaces listed above also has a nice Riemannian metric. We saw earlier that S
2
has a left-
SO(2) invariant Riemannian metric, viz. the one induced from R
3
. Similarly, since the lattices C act by
translations, they are isometries with respect to the euclidean metric on C, and thus T
2
= C/(Z+iZ) acquires
a at metric, denoted by d
2
+d
2
.
Finally for the surfaces above of genus 2, the universal cover H
2
has the hyperbolic metric, and since the
group is a subgroup of SL(2, R), it is a group of hyperbolic isometries, and the quotient naturally acquires
the quotient hyperbolic metric.
The homogeneity of the metrics on S
2
, C = R
2
and H
2
will result in their curvature (to be dened later)
being constant functions (in fact, 1, 0, and -1 respectively). Thus the quotient surfaces above with these
quotient metrics will also have constant curvature.
The compact non-orientable surfaces, of course cannot be complex manifolds, since all complex manifolds
are orientable. However their orientable double covers will belong to the above list. They are all of the following
type:
RP(2)#RP(2)#...#RP(2) k copies
(which is called the k-crosscap). The involutions of the orientable surfaces M above, which give rise to
the non-orientable surfaces M/{1, }, are all isometries of M with respect to the metrics dened above. (For
example, if M = S
2
, the is the antipodal map, and M/{1, } = RP(2).) Thus the quotients, viz. the
crosscaps all acquire nice quotient metrics.
72 VISHWAMBHAR PATI
6. Connections
6.1. Principal G-bundles. In the sequel G will always a smooth Lie group, and g its Lie algebra. L
g
and
R
g
denote left and right translations by g G, and Adg := L
g
R
g
1.
Denition 6.1.1 (Principal G-bundles). A smooth principal G-bundle on a smooth manifold M is a smooth
manifold P of dimension dim G + dim M such that G acts freely and smoothly from the left on it. Further,
there is a smooth surjection : P M such that each bre P
x
=
1
(x) is a free and transitive G-space, and
a G-orbit of the action of G on P. Further, there exists an open covering {U
i
}
i
of M, and smooth maps
(called local trivialisations or bundle charts):
i
:
1
(U
i
) U
i
G
making the diagram:
1
(U
i
)
i
U
i
G
pr
1
U
i
commute. For each x U
i
, and each i, the restriction
i|Px
is a G-equivariant (=commuting with the left
action of G) map to {x} G. P is called the total space of the bundle, and M the base space. is called the
bundle projection. G is called the structure group of the principal bundle. (Physicists call it the restricted
gauge group)
If there exists such an open covering with just a single element {M}, then the bundle is said to be a trivial
bundle. In this case, P = M G, and the left action of G on P is via left-multipication on the second factor.
It is easy to check that local triviality forces to be open, and thus is a quotient map, and M P/G,
where P/G is the orbit space of the left G-action on P.
For a principal G-bundle, we can dene transition functions, analogous to what we did for vector bundles
in 4.3.3. Namely, on the overlap U
i
U
j
, the left G-equivariant bundle coordinate change:
(U
i
U
j
) G

1
i
1
(U
i
U
j
)
j
(U
i
U
j
) G
must carry (x, g) to (x, g.g
ij
(x)), where g
ij
(x) G. Thus, for each i, j, we have a smooth map:
g
ij
: U
i
U
j
G
such that the following cocycle conditions hold:
(a): g
ii
= Id, g
ij
(x) = g
ji
(x)
1
for all i, j.
(b): g
ij
(x)g
jk
(x)g
ki
i
U
j
U
k
.
A smooth map s : M P is called a section of the bundle if s = Id
M
. As in the part (iv) of the exercise
4.3.4, this is equivalent to giving smooth functions s
i
: U
i
G for each i, satisfying s
j
(x) = s
i
(x)g
ij
(x) for all
x U
i
U
j
.
Lemma 6.1.2. A principal G-bundle : P M is (smoothly) trivial i it has a smooth section s : M P.
Proof: If the bundle is trivial, then the section x (x, e), (where e G is the identity element) is a section
of the trivial bundle M G M, and a G-bundle isomorphism : M G P of this bundle to a principal
G-bundle : P M would result in the section x (x, e) of that bundle.
Conversely, if : P M admits a section, the isomorphism:
: M G P
(x, g) gs(x)
is a G-bundle isomomorphism of P with the trivial G-bundle. 2.
Example 6.1.3. All the examples of homogeneous manifolds:
: G G/H
constructed so far (see the section 5.7) are examples of principal H-bundles. The only dierence with the
denition above is the fact that in those examples, the closed subgroup H was acting from the right by right
multiplication. One can easily convert it into a left action by dening h g := gh
1
. To see local triviality, one
needs to show local triviality of the bundle only around a neighbourhood of the identity coset o = e.H G/H.
In the notation of proposition 5.7.1, a neighbourhood of the identity coset is ((WV )), which is dieomorphic
to the neighbourhood W of the origin in m. The open set
1
(W) in G is just exp(W)H. We may therefore
nd a trivialisation of
1
(W) W by setting:
: exp(W) H
1
(W)
(exp X, h) (exp X)h
which commutes with the left action (=right multiplication) of H. Indeed, on the open set exp(W) exp(V ),
we have seen in the proposition above that this is essentially the dieomorphism of that proposition. The
H-equivariance proves it is a dieomorphism all over
1
(W), and is the trivialisation sought.
Example 6.1.4 (Frame bundle of a vector bundle). Let : E M be a smooth vector bundle of rank k. We
dene the set of all frames in the bre E
x
to be set of all ordered tuples (v
1
, .., v
k
) of k linearly independent
vectors in E
x
(=ordered bases or frames of E
x
), and denote it by F(E)
x
. Now, dene the frame bundle F(E)
to be the disjoint union:
F(E) =
xM
F(E)
x
and the projection p : F(E) M by mapping an element in F(E)
x
to x. There is a left action of G =
GL(k, R) on each F(E
x
), where g = [g
ij
] G takes the frame (v
1
, .., v
k
) of E
x
to the frame (w
1
, .., w
k
), where
w
j
=

g
1
ij
v
i
. (The funny denition stems from the fact that tuples of basis elements are to be treated as
row vectors, and the right action of G on row vectors has to be converted to a left action). We note that
if
i
:
1
(U
i
) U
i
R
k
is a smooth bundle atlas for E, then we may dene:
i
: p
1
(U
i
) U
i
GL(k, R)
(v
1
, .., v
k
) F(E)
x
(x, pr
R
k(
i
(v
1
), ...,
i
(v
k
))
where we are naturally identifying all frames in R
k
with G (by writing the elements of a frame as columns of
a matrix). These
i
then become G-equivariant bijections, and we transport the topology from U
i
G, and
check that they are compatible on overlaps. With this topology, it is easy to check that F(E) is a smooth
manifold, and p : F(E) M is a prinicpal GL(k, R) bundle. The bundle F(E) is also sometimes denoted
GL(E).
Example 6.1.5. If : E M is a smooth vector bundle with a Riemannian metric , , we may dene
the bundle of orthonormal frames O(E) in a fashion analogous to F(E) above. O(E)
x
is then the set of all
orthornormal frames in E
x
with respect to the inner product ,
x
on E
x
. The discussion above goes through,
mutatis mutandis, and exhibits p : O(E) M as a principal O(k)-bundle.
In analogous fashion, if the vector bundle : E M is an oriented bundle (i.e. there is a nowhere vanishing
section s of
k
(E)), we can construct a principal SL(k, R) bundle p : SL(E) M by choosing only those
frames (v
1
, .., v
k
) of E
x
which satisfy v
1
.. v
k
= s(x). If it is an oriented Riemannian vector bundle, then we
can form the oriented orthonormal frame bundle SO(E) by choosing only those orthonormal frames (v
1
, .., v
k
)
of E
x
which satisfy v
1
.. v
k
= s(x).
Exercise 6.1.6. Dene the set:
V
k
(R
n
) = {(v
1
, .., v
k
) (R
n
)
k
: v
i
, v
j
=
ij
}
74 VISHWAMBHAR PATI
of orthonormal k-frames in R
n
. Show that it is a homogeneous manifold O(n)/O(nk). It is called the Stiefel
manifold. For example V
1
(R
n
) = S
n1
. Show that the natural map:
: V
k
(R
n
) G
k
(R
n
)
(v
1
, .., v
k
) span(v
1
, ..., v
k
)
is a principal O(k) bundle. Prove that it is isomorphic to the orthonormal frame bundle O(
k
n
) of the tautological
bundle
k
n
G
k
(R
n
). In particular, the principal O(1) = Z
2
orthonormal frame bundle of the tautological line
bundle
1
RP(n 1) is the natural quotient mapping:
: S
n1
RP(n 1)
Exercise 6.1.7 (Orthonormal frame bundle of TS
n
). Show that the orthonormal (resp. oriented orthonor-
mal) frame bundles of the oreiented Riemannian vector bundle TS
n
S
n
are : O(n + 1) S
n
(resp.
: SO(n + 1) S
n
), where maps the orthogonal (resp. special orthogonal) matrix A to its last column
Ae
n+1
.
It is perhaps not yet clear why principal G-bundles were introduced. It is true that we could have continued
our discussion by just sticking with vector bundles. However, it is important to note that various dierent
vector bundles may have one underlying principal G-bundle. Also, as we saw in the last subsection, dierent
structures on a vector bundle (e.g. an orientation, Riemannian bundle metric) lead to reduction of the structure
group G (for example, from GL(k, R) to SL(k, R), or O(k, R)) of the principal G-bundle. All this is claried
in the sequel.
6.2. Associated bundles to a principal bundle. Let G be a Lie group. For the sake of brevity, we denote
a set X with a left G action as a G-space. If X is a manifold, we require the action to be smooth, if X is
a vector space, we require each left translation L
g
to be linear and the homomorphism G GL(X) taking
g L
g
to be smooth. If X is another Lie group H, we again require the action to be smooth. In all of these
situations, we have a homomorphism or representation:
: G Aut(X)
mapping g to L
g
, and Aut(X) is to be intepreted according to the context. Just note that when X = H, a
group, L
g
is only usually a dieomorphism, and not a homomorphism. For example, if G is a subgroup of H,
then left multiplication by g G is not a homomorphism H H.
Denition 6.2.1 (Associated bundles). Let : P M be a principal G-bundle, and let X be a G space
in the sense above. Introduce an equivalence relation on P X by coordinatewise left action of G (called the
diagonal action). That is (a, x) (ga, gx) for all g G. Denote the equivalence class of (a, x) as [p, x].
Then the associated bundle with bre X is dened as follows:
P
X = P X/G = {[a, x] : a P, x X}
It has a natural quotient topology. Dene the projection map p : P
X M by p([a, x]) = (a). This is

clearly well dened since (a) = (ga) for all g G.
A typical bre, say y M, will be p
1
(y) a X, where a is some xed element of
1
(y). (Since the
action of G on P is free, (a, x) is a well-dened representative of the equivalence class [a, x]). We leave it to
the reader to show that if P has bundle charts over open sets U
i
M, then over U
i
this bre bundle will be
isomorphic to the trivial bundle U
i
X.
Example 6.2.2. If X is a vector-space of dimension k, and G acts on V by linear automorphisms (i.e. there
is a smooth representation : G GL(V )), then the associate bundle p : P
rho
V M is a vector bundle
of rank k. If maps into SL(V ), then this associated vector bundle becomes an oriented vector bundle. If V
has an inner product , , and maps into O(V ), then this vector bundle acquires a Riemannian metric. If
maps into SO(V ), it becomes an oriented vector bundle with Riemannian metric. If
g
ij
: U
i
U
j
G
are the transition functions of the given principal bundle , show that the transition functions of these associated
vector bundles are g
ij
.
Proposition 6.2.3. Let : E M be a smooth vector bundle. Let V denote the bre E
x
of some xed point
x M. We x a linear isomorphism (= identication) : R
k
V . Thus get an action of GL(kR) on V , by the
(natural) left multiplication on column vectors. That is, we x an isomorphism : GL(k, R) GL(V ). Let
F(E) denote the frame bundle of E, constructed in 6.1.4. Then the associated vector bundle F(E)
V M
is isomorphic to the original vector bundle E M.
Proof: A point (a, v) F(E) V gives a frame a = (v
1
, .., v
k
) in y = (a) M, and a vector (
i

i
e
i
)
where {e
i
} is the standard basis of R
k
. Dene the map:
: F(E) V E
(a, v) F(E)
y
V
i
v
i
E
y
We note that if g = [g
ij
] GL(k, R), then ga is the frame of E
y
given by w
j
=

i
g
1
ij
v
i
. Also g.v =
(
j

j
e
j
), where
j
=
j
g
jl
l
. Then
(ga, gv) =
j
w
j
=
i,l,j
g
jl
l
g
1
ij
v
i
=
i,l
il
l
v
i
=
i
v
i
This proves that descends to a map : F(E)
V E. The verication that this is an isomorphism of

vector bundles is left to the reader. 2
Exercise 6.2.4. Prove the analogous facts for an oriented (resp. Riemannian vector bundle) E M, and
the principal bundles SL(E) (resp. O(E)) constructed in 6.1.5.
Exercise 6.2.5. In the notation of the preceding exercise, show how the various associated bundles
k
(E),
E
, (
k
(E)) (
l
(E
)) may be constructed as associated bundles to the prinicpal bundle F(E) M and

suitable representations of GL(k, R) on vector spaces (such as
k
(V ) etc.) associated with V .
Denition 6.2.6 (Reduction of structure group). Let : P M be a principal G-bundle, and let H be a
subgroup of G. Thus there is the natural left action of H on G by left multiplication, and this makes G into a
left H-space, and as usual we denote the structure map by . Note that the associated bundle Q
G acquires
a left action of G by g.[x, b] := [x, b.g
1
]. In fact, it becomes a principal G-bundle with this action of G (Prove).
We say that the principal G-bundle P M admits reduction of structure group to H if there exists a principal
H-bundle Q H such that the associated bundle Q
G is isomorphic to P as a principal G-bundle.

For example, in the situation of an oriented (resp. Riemannian) vector bundle, the frame bundle F(E)
admitted a reduction of structure group from GL(k, R) to SL(k, R) (resp. O(k)). The reader should verify
that F(E) SL(E)
GL(k, R) (resp O(E)
GL(k, R).)
For further details about prinicpal bundles and associated bre bundles, the reader may consult D. Husemollers
book Fibre Bundles or Kobayashi and Nomizus Foundations of Dierential Geometry, Vol 1.
76 VISHWAMBHAR PATI
6.3. The Maurer-Cartan Form. Let G be a Lie group (of dimension say m), and g its Lie algebra. We have
already seen that a trivialisation of T(G) can be obtained by choosing a basis {X
i
}
m
i=1
of g, and generating the
m everywhere linearly independent sections

X
i
, the left-invariant vector elds corresponding to X
i
. Similarly,
if we take an element of g
, we can generate a left-invariant 1-form by the prescription:

(L
g
)
(g) = (e) =
That is, L
g
= L
g
1
. Again, starting with a basis {
i
} of g
will lead to left invariant 1-forms {

i
}
trivialising T
G, the cotangent bundle of G. If we denote by

1
(G) the space of 1-forms on G, then we have
the natural identication:
:
0
(G) g

1
(G)
f f
which preserves the left
0
module structure (multiplication with smooth functions) of both sides.
Now, tensor product
1
(G) g (=the space of g valued 1-forms), which we denote
1
(G, g), therefore
becomes isomorphic to
0
(G) g
g. Now, for any vector space, there V is a unique element

V
V
V ,
which can be denoted by

i
e
i
e
i
, where {e
i
} is any basis of V , and {e
i
} the corresponding dual basis.
This element is well- dened independent of choice of basis, and corresponds to the identity map Id
V
under
the canonical isomorphism V
V hom
R
(V, V ). Similarly, there is the constant function 1 on G that is a
distinguished element of
0
(G).
Denition 6.3.1. Under the above identication:
1
(G, g)
0
(G) g
g
the g-valued 1-form in
1
(G, g) corresponding to the distinguished element 1
g
in the right hand side is
called the Maurer-Cartan 1-form of G. If we choose a basis {X
i
} of g, and denote by
i
its dual basis, then
g
=
i

i
X
i
. This corresponds under above to the left-invariant g valued 1-form:
=
i
X
i

i
We call it the Maurer-Cartan form of G, because it is the unique left invariant 1-form on G whose value at
e G is the (also unique) form
g
.
We can similarly dene the right invariant Maurer-Cartan form of a Lie group G to be the unique g valued
one form on G which is invariant under right translations, and agrees with
g
at the identity. This right
invariant Maurer-Cartan 1-form will not agree with the left invariant one unless the group G is abelian.
We can compute the Maurer-Cartan form of GL(n, R), for example.
Example 6.3.2 (Maurer-Cartan form of GL(n, R)). Note that since G = GL(n, R) is an open subset of g =
gl(n, R) = R
n
, there are global coordinates g
ij
available on G. We let g denote the g-valued function given by
restricting the identity map of g to the open subset G. Thus there are also the globally dened 1-forms dg
ij
,
and indeed the g-valued 1-form dg := [dg
ij
] on G.
If h = [h
ij
] is a xed matrix, then for a B GL(k, R) we have:
g
ij
(L
h
B) = (h.B)
ij
=
l
h
il
g
lj
(B)
which shows that L
h
g = g L
h
= h.g and similarly R
h
g = g R
h
= g.h
Hence
L
h
dg = d(L
h
g) = d(h.g) = h.dg
and similarly R
h
dg = dg.h So we see that dg is neither left nor right invariant. But since the function g on G
also undergoes the same eect under left and right translations,
the 1-form
R
:= dg.g
1
on G (the dot . again means matrix multiplication) will be right-invariant, and
similarly the 1-form
R
= g
1
dg will be left-invariant.
Now we just have to examine their values at I G. But at the identity we have the basis {E
ij
} of
T
e
(G) = gl(n, R). Our one form dg.g
1
at the identity is just the matrix [dg
ij
], which maybe written as
i,j
E
ij
dg
ij
. Note that, in Euclidean space g, E
ij
=

gij
are the natural basis of coordinate partials, and
dg
ij
are the dual basis. Thus
L
and
R
are, by uniqueness, the left and right invariant Maurer-Cartan forms
of G respectively. .
Note that L
R
= hdg.g
1
h
1
= Ad(h)
R
, and similarly R
L
= Ad(h
1
)
(
L
), where the adjoint
automorphism Ad(h)
: g g is being applied pointwise. This motivates the:

Exercise 6.3.3. Prove the analogous fact for the left and right Maurer- Cartan forms of any Lie group G.
(Just examine the eect of L
h
(resp. R
h
) on right-invariant (resp. left invariant) 1- forms.)
Remark 6.3.4 (Maurer-Cartan forms of linear lie groups). It is clear that the restriction (=pullback under
inclusion) of the Maurer-Cartan form of G to a Lie subgroup H will be the Maurer-Cartan form of H. Thus we
can restrict the left (resp. right) Maurer-Cartan form of GL(n, R) to any of the classical subgroups SL(n, R),
O(n), SO(n) and get their left (resp. right) Maurer-Cartan forms. For example, for H = SO(2) GL(2, R),
we have:
dg.g
1
=
_
sin cos
cos sin
_
d.
_
cos sin
sin cos
_
=
_
0 1
1 0
_
d
When one identies the Lie-algebra R of SO(2) as the subspace spanned by
_
0 1
1 0
_
, this becomes the
well-known volume form d of the circle group, which is right-invariant. The left Maurer-Cartan form for
SO(2), viz. g
1
dg also comes out to be the same, since SO(2) is abelian!
6.4. Connections on principal bundles.
Denition 6.4.1. Let : P M be a smooth principal G-bundle. The vertical tangent bundle, denoted
T
v
P P is dened by the short exact sequence of vector bundles:
0 T
v
P TP

TM 0
That is, it is the kernel bundle of the surjective bundle morphism
= D : TP
TM (that
is surjective
follows from local triviality).
Lemma 6.4.2. Let : P M be a principal G-bundle, and L
g
denote the left-translation map x gx of P.
Then:
(i): The vertical tangent bundle is mapped isomorphically to itself by L
g
= DL
g
for each g G. That is
L
g
(T
v
x
P) = T
v
gx
P = (L
g
T
v
P)
x
.
(ii): For each x P, T
v
x
(P) = T
x
(Gx). That is, the vertical tangent space at each point x P is the
tangent space to the orbit Gx.
(iii): T
v
P is a trivial vector bundle with bre g.
Proof:
(i) is clear since L
g
= , which implies that L
TM =
TM and
L
g
.
For (ii), note that for x P,
1
((x)) = Gx, and since (x) is a regular value of , the kernel of D(x) is
precisely the tangent space
1
((x)) by 3.2.15.
For (iii), we recall that the left-action of G on P is given by a smooth map:
: GP P
x (g, x) := g.x
78 VISHWAMBHAR PATI
Thus it has a derivative D(e, x) : g T
x
(P) T
x
(P). Note that for a xed x P,
x
:= (, x) : G Gx
is a left G-equivariant dieomorphism, and therefore maps g isomorphically onto T
x
(Gx) = T
v
x
P T
x
P.
In particular, given an X g, we get a vector eld on P, which we also denote as

X, dened by

X(x) :=
x
(X). Now, for an h G, we have
hx
(Ad(h)g) = hgh
1
.hx = hg.x = L
h

x
(g)
Thus
L
h

X = L
h
(D
x
(e)X) = D
hx
(e)(Ad(h)
X) =

(Ad(h)
X)
Thus

X is not a left-invariant vector eld, but obeys the Ad-invariance property:
L
h

X =

Ad(h)
X)
Denition 6.4.3 (Maurer-Cartan section of the vertical bundle). By the last proposition, there is a smooth
and canonical isomorphism of vector bundles:
hom(T
v
P, T
v
P) hom(T
v
P, g) (T
v
P)
g
Denote by
v
the section of (T
v
P)
g which corresponds to the identity section of hom(T

v
P, T
v
P) under
this identication. If we choose a basis {X
i
} of g, and denote by e
i
the Ad-invariant sections of T
v
P which
satisfy e
i
(
X
j
) =
ij
, then
v
is given by:
v
=
i
e
i
X
i
Again, as in the case of the Maurer-Cartan form of a Lie group, the reader may verify that this expression for
v
holds for all choices of basis for g. Clearly,
v
is Ad-invariant under left translations L
g
. That is, for a xed
g G,
L
v
= Ad(g)
v
where the right side denotes pointwise multiplication. It will be called the Maurer-Cartan section of (T
v
P)
g.
Remark 6.4.4. It is natural to ask what the Maurer-Cartan section of (T
v
P)
g has to do with the left

and right Maurer-Cartan forms of G introduced in 6.3.1. The exercise 6.3.3 suggests that the transformation
property of the Maurer-Cartan section
v
above, under left translations, is the same as that of the right
Maurer-Cartan form
R
of the group G. Indeed, the reader can convince herself that if we identify the orbit
Gx with G, via the G-equivariant map
x
, then the pullback (under
x
) of the Maurer-Cartan section
v
is
the same as the right Maurer-Cartan section
R
of G.
Denition 6.4.5. A connection on a principal G-bundle : P M is a vector bundle morphism:
: TP T
v
P
such that:
(i):
|T
v
P
= Id
T
v
P
(ii): L
g
= L
g

(We recall from (i) of the lemma above that L
g
preserves the vertical tangent bundle, so this condition
makes sense).
Notation : 6.4.6. In the sequel, for g G, we will denote the automorphism Ad(g)
: g g as Ad(g) for
notational convenience.
Proposition 6.4.7. Let : P M be a principal G-bundle. The following are equivalent:
(i): A connection on the principal G-bundle P M.
(ii): A smooth sub-bundle T
h
P TP (called the horizontal bundle) such that (a) L
g
(T
h
x
) = T
h
gx
P for all
x P, g G, and (b) T
h
P T
v
P = TP.
(iii): A smooth section of T
P g (=smooth Lie algebra-valued 1-form) satisfying: (a) L
g
= Ad(g)
and (b) (X) =
v
(X) for every vertical tangent vector X T
v
P. (Note that in (a) Ad(g) denotes the
1-form whose value on a tangent vector v is Ad(g)((v))).
Proof:
(i) (ii) Clearly, if is a smooth connection, then for each x P, the kernel bundle:
T
h
P := ker : TP T
v
P
will satisfy T
h
P T
v
P = TP, because, by denition, is a bundle-splitting map for the inclusion T
v
P TP,
and hence (b) of (ii) follows. Also, for v T
h
x
P = ker
x
: T
x
P T
v
x
P, (L
g
v) = L
g
((v)) = 0, by the
property (ii) in 6.4.5. Thus L
g
(v) ker
gx
, which implies (a) of (ii)
(ii) (iii) Since TP = T
v
P T
h
P, we have an isomorphism
T
P g = (T
v
P)
g (T
h
P)
g
We already have the smooth Maurer-Cartan section
v
of (T
v
P)
g, by 6.4.3. Dene the section of the

left hand bundle by = (
v
, 0). Then (b) of (iii) is trivial. The fact that L
v
= Ad(g)
v
, and the fact that
L
g
preserves the bundles T
h
P and T
v
P, and hence the direct sum decomposition TP = T
v
P T
h
P, implies
that L
g
= Ad(g).
(iii) (i) Let be a smooth form satisfying (a) and (b), and for a Y T
x
P, dene (Y ) =
(x)(Y )(x)
T
v
x
P. Note that (x)(Y ) g, and it makes sense to dene the Ad-invariant vector eld from it, as explained
in (iii) of 6.4.2.
Then
(L
g
Y ) =

(gx)(L
g
Y )(gx) =

L
g
(x)(Y )(gx)
=

Ad(g)((x)(Y ))](gx)
= L
g
[
(x)Y (x)]
= L
g
((Y ))
which proves (ii) of the denition of a connection. For some basis X
i
of g, the vector elds

X
i
span T
v
P at
each point, and furthermore
v
=
i
e
i
X
i
. Thus:
(x)(
X
j
) =
v
(
X
j
) =
i
e
i
(
X
j
) X
i
=
ij
X
i
= X
j
. Hence (
X
j
) =
X
j
) =

X
j
which proves (i) in the denition of a connection. 2
Corollary 6.4.8. The connection : TP T
v
P and the connection form are related by the formula:
(Z) ==
v
((Z)) for Z TP
Proof: This is clear from the step (ii) (iii) in the proof of the above proposition, for if Z TP, then
Z = ((Z), Z (Z)) in the decomposition TP = T
v
P T
h
P, and so (Z) =
v
((Z)). 2.
Denition 6.4.9. Let : P M be a principal G-bundle. A G- equivariant automorphism : P P such
that = (i.e. a bundle isomorphism of the principal bundle) is called a gauge transformation. The set of
gauge transformations is a group, called the full gauge group, and denoted by G.
80 VISHWAMBHAR PATI
Remark 6.4.10. If we let {(
i
, U
i
)} be a G-bundle atlas for : P M, we can dene smooth local sections
i
: U
i

1
(U
i
) of the bundle P
|Ui
by setting
i
(x) =
1
i
(x, e). We note that these
i
will satisfy the
following patching condition on U
i
U
j
:
i
(x) = (
1
j

j
)
1
i
(x, e) =
1
j
(x, g
ij
(x)) = g
ij
(x)
1
j
(x, e) = g
ij
(x)
j
(x)
where g
ij
: U
i
U
j
G are the bundle transition functions introduced in the rst subsection 6.1.
Now, given a gauge transformation , we note that for x U
i
, (g.
i
(x)) = g(
i
(x)) by G- equivariance of
, and hence is completely determined by (
i
). Also, since (
i
(x)) is in the same bre as
i
(x), there is a
unique
i
(x) G such that (
i
(x)) =
i
i
(x), and again is uniquely determined by these maps
i
: U
i
G.
Finally, for x U
i
U
j
:
i
(x)
i
(x) = (
i
(x)) = (g
ij
(x)
j
(x))
= g
ij
(x)(
j
(x)) = g
ij
(x)
j
(x)
j
(x)
= (Adg
ij
(x)(
j
(x)))
i
(x)
which shows that the maps
i
patch up to give a section of the associated adjoint bundle P
Ad
G. Hence a
gauge transformation is equivalent to a section of the adjoint G-bundle of G.
Denition 6.4.11. Let : TP T
v
P be a smooth connection, and let : P P be a gauge transformation.
Since preserves the G-orbit of a point, its derivative
maps the vertical tangent space T

v
x
P to the vertical
tangent space T
v
(x)
P. Also the We dene the pullback connection
on : P M to be the composite:
TP

1
TP

T
v
P

T
v
P
Note that
T
v
P
= Id
T
v
P
implies the same fact for
, thus making
also a connection.
This gives an action of the full gauge group G on the space of all connections on P. Before we proceed
further, here is another fruitful way of describing a gauge transformation.
Remark 6.4.12. There is a 1 1-correspondence between the full gauge group G and the set:
{g : P G smooth : g(hx) = Adh.g(x) = hg(x)h
1
for all x P}
which is a group via pointwise multiplication. For, if G, then since preserves bres, and G acts freely on
the bres, for each x P:
(x) = g
1
(x).x
for a unique g
(x) G. From the G-equivariance of , it follows that g

1
(hx).hx = (hx) = h(x) =

hg
1
(x)h
1
.hx, which implies that g
(hx) = Ad(h)g
(x) for h G and x P. Similarly, one easily checks

that the map g
of G to the set above is a group homomorphism for:

(x) = (g
1
(x)x) = g
1
(x)(x) = g
1
(x)g
1
(x).x = (g
(x)g
(x))
1
x
Since it is clearly a bijection, the map g
is an isomorphism of groups. The reader may easily check from

the denitions that g(
i
(y))
1
=
i
(y) for y U
i
, where
i
: U
i
G are dened in 6.4.10. From this relation,
and the fact that g(hx) = Adh.g(x), the reader may also check that g : P G is smooth.
Lemma 6.4.13. We recall that we have trivialised the vertical tangent bundle in 6.4.2, via the map X

X
for X g. If : P P is a gauge transformation, then
X(x)) =

X((x))
Proof: By denition
x
(g) = gx, and thus by the G-equivariance
x
(g) = g(x) =
(x)
. Thus:
X(x)) = D(x)(D
x
(e)(X)) = D
(x)
(e)(X) =

X((x))
Corollary 6.4.14. Let
v
be the Maurer-Cartan section of the vertical cotangent bundle (T
v
P)
g dened in
6.4.3. If G, we have the isomorphism of bundles
: T
v
P
T
v
P, and hence the dual isomorphism
(T
v
P)
(T
v
P)
, and the induced isomorphism, which we also denote as
(T
v
P)
g (T
v
P)
g.
Then, we claim:
(
v
) =
v
for all x P.
Proof: We let X
i
be a basis of g, and e
i
be the dual basis to the elds

X
i
, retaining the notation of 6.4.3.
The lemma above implies that
X
i
) =

X
i
.
By denition:
(
e
i
)(x)(
X
j
(x)) = e
i
((x))(
(X
i
(x))) = e
i
((x))(
X
j
((x)) =
ij
which shows that (
e
i
)(x) = e
i
(x). Then, using the explicit expression
v
=
i
e
i
X
i
for the Maurer-Cartan
section (see 6.4.3), the result follows. 2.
Proposition 6.4.15 (Eect of G on the connection form). Let be the connection form of a connection ,
and
that of
, where G. Also let g := g
: P G be the smooth map associated with as described

in 6.4.12 above. Then:
= (Ad g) +g
R
where
L
is the left-invariant Maurer-Cartan 1-form on G.
Proof: We recall from the corollary 6.4.8 above that (Z) =
v
((Z)), for Z T
x
P. If G is a gauge
transformation, then:
(Z) =
v
(
(Z)) =
v
(
(Z))
Since annihilates horizontal vectors, we need the vertical component of
1
(Z) in the connection . Note

that
1
is the composite:
P
(g
,1)
GP

P (9)
so we need to compute the (vertical component of) the derivative D(g(x), x) where, for notational convenience,
we drop the subscript on g
. For any g G, there is the commutative square:

G P

P
L
g
Id L
g
G P

P
which yields the relation:
L
g
D(1, x) = D(g, x) (L
g
, Id)
Hence if (Y, Z) T
(g,x)
(GP) = T
g
GT
x
P is a tangent vector at (g, x), we have:
D(g, x)(Y, Z) = L
g
D(1, x)(L
1
g
Y, Z)
Now we recall the right-invariant Maurer-Cartan form
R
of G, as dened in 6.3.1. Note that Adg
1
R
(Y ) :=
R
(L
1
g
Y ) = L
1
g
(Y ), because L
1
g
Y T
1
G = g and
R
acts on it as identity. Thus we have the formula:
D(g, x)(Y, Z) = L
g
D(1, x)(Adg
1
R
(Y ), Z) (10)
If we just want the vertical component, we apply to the formula above, and using the fact that commutes
with L
g
we have:
(D(g, x)(Y, Z)) = (L
g
D(1, x)(Adg
1
R
(Y ), Z)
= L
g
_
(
x
(Adg
1
R
(Y )) + (D(1, x)(0, Z)
_
= L
g
((

Adg
1
R
(Y )) +L
g
(Z)
= (L
g
(

Adg
1
R
(Y )(x)) +L
g
(Z)
= (gx)(
R
(Y )) +L
g
(Z)
=

R
(Y )(gx) +L
g
(Z) (11)
82 VISHWAMBHAR PATI
since D(1, x)(0, Z) = Z, L
g
W =
Adg.W for any W g (by proof of (iii) in Lemma 6.4.2), and
R
(Y ) is a
vertical vector eld, and so maps it to itself. Thus, from (9) above, we have:
(
1
(Z)) = (D(g, x)(Dg(x)Z, Z))

=

(
R
(Dg(x)Z))(g(x)x) +L
g(x)
(Z) (12)
Hence,
(x)(Z) =
v
(
(Z))
= (
v
)(x)(
1
(Z))
=
v
((x))(
1
(Z)) (13)
=
v
(g(x)x)

R
(Dg(x)Z)(g(x)x) +
v
L
g(x)
(Z)
=
R
(Dg(x)Z)) +
v
(L
g(x)
(Z))
= g
R
(Z) + (L
g(x)
v
)((Z))
= g
R
(Z) + (Adg(x)
v
)((Z))
= g
R
(Z) + Adg(x)((Z))
where we used the corollary 6.4.14 in the third line, equation (12) in the fourth line, the fact that
v
(y)(
W(y)) =
W for all W g in the fth line, and and the fact that L
v
= Adg.
v
from 6.4.3. in the sixth line. This
proves the proposition. 2
Remark 6.4.16. From the equation (13) in the proof above, it follows that:
(Z) =
v
(
1
(Z)) = (
1
(Z)) = (
1
)
(Z)
That is,
= (
1
)
Denition 6.4.17 (Local description of a connection). Let {

i
, U
i
} be a bundle atlas for the principal G-
bundle : P M, and let
i
: U
i

1
(U
i
) be the smooth local sections dened in the remark 6.4.10
above. Let be the Lie algebra-valued 1-form of a connection on : P M as in (iii) of 6.4.7. We denote
the pullback 1-forms
i
, which are sections of T
U
i
g, as
i
. Since the inverse of the trivialisation map
1
i
: U
i
G P
|Ui
is dened by (x, g) L
g

i
(x), and that L
g
= Adg(), it is clear that all the
i
completely determine .
It is clear that the local forms
i
dened above will have some compatibility formula on the overlaps U
i
U
j
.
This follows from the proposition 6.4.15 above. More precisely:
Proposition 6.4.18 (Compatibility condition for the local forms
i
). Let
i
denote the local connection 1-
forms on U
i
for the connection as in 6.4.17 above. Then:
i
= (Adg
ij
)
j
+ (g
ij
R
) for all x U
i
U
j
In particular, if G = GL(n, R), we recall that
R
= dg.g
1
, and hence g
ij
R
= (dg
ij
.g
1
ij
), which yields:
i
= (Adg
ij
)
j
+dg
ij
.g
1
ij
for all x U
i
U
j
Proof: Consider the principal G-bundle P
ij
:= P
|UiUj
=
1
(U
i
U
j
) on U
ij
:= U
i
U
j
. We can dene a
gauge transformation of this bundle by:
(h
j
(x)) = h
i
(x) = hg
ij
(x)
j
(x)
This would mean that if g
: P
ij
G is the corresponding map, then we have g
(
j
(x)) = g
ij
(x). Now since
i
=
1
j
we have:
i
=
j
((
1
)
) =
j
(
)
using the remark 6.4.16 above. Now from the proposition 6.4.15 above, we substitute for
to get
i
(x) =
_
j
(Adg
+g
R
)
_
(x)
= Adg
(
j
(x))(
j
)(x) +
_
(g

j
)
R
_
(x)
= Adg
ij
(x)
j
(x) + (g
ij
R
)(x)
which proves the proposition. 2
Exercise 6.4.19. Show that a collection of smooth g-valued 1-forms
i
on U
i
which satisfy:
i
= Ad(g
ij
)
j
+g
ij
R
on the overlaps U
i
U
j
denes a connection on the bundle : P M.
Proposition 6.4.20 (Existence of a connection). Let : P M be a principal G-bundle. Then there exists
a connection on it.
Proof: Let us rst note that a connection certainly exists on a trivial bundle P = M G. Just the pullback
:= p
R
of the right-invariant Maurer-Cartan form
R
on G under the second projection:
p : M G G
will do the job. For any x = (m, g) P, we have that
x
: G P is the map h (m, h.g) = (m, R
g
h), so
that for X g, the vector eld

X is dened by
p
X(m, g)) = p
(0, R
g
X) =

X(g)
where

X is the right-invariant vector eld generated on G by X. Thus:
X =
R
(
X) =
R
(p

X) = (
X)
which shows that and
v
agree on vertical vectors. Further, the transformation property of under L
g
follows from the corresponding fact for
R
and the fact that p preserves the left G-action.
Now let : P M be any principal G-bundle, with
i
: P
|Ui
U
i
G the G-equivariant trivialisations.
Clearly, we get connection 1-forms
i
on P
|Ui
=
1
(U
i
) by setting
i
:=
i
()
where is a connection 1-form on the trivial bundle U
i
G as constructed above.
Now let
i
be a partition of unity subordinate to U
i
. Dene the 1-form by:
(x) =
i
((x))
i
(x) for x P
where
i
(x) is taken to be 0 outside
1
(U
i
). The fact that L
g
= Adg. follows from the corresponding
property of the
i
. Also, for a vertical vector X, we have (X) =

i

i
i
(X) =

i

i
v
(X) =
v
(X), since
i
(X) =
v
(X) for all i and X a vertical vector. 2
84 VISHWAMBHAR PATI
Example 6.4.21. A simple situation where a natural connection arises on a principal G-bundle : P M
is when P has a Riemannian metric , that is invariant under left translations. One can simple dene the
horizontal subspace T
h
x
P as the orthogonal complement (T
v
x
P)
for each x P. The fact that , is left-

invariant shows that L
g
(T
h
x
P) = T
h
gx
P for each x P. We can easily compute the connection form for this
connection. Let {X
i
}
n
i=1
be a basis for g, and let

X
i
be the corresponding basis of vertical vector elds, and e
j
the basis of 1-forms dual to this basis. Then let Y
i
be the basis of vertical vector elds obtained by applying
the (smooth) Gram-Schmidt process to

X
i
. Clearly (Z) =
i
Z, Y
i
Y
i
. Thus
(Z) =
v
((Z)) =
j
( e
j
(Z) X
j
) =
i,j
A
ij
(x) Z, Y
i
X
j
for Z T
x
P, where A
ij
(x) := e
j
(Y
i
(x)) is the matrix of the Gram-Scmidt operation.
Example 6.4.22 (Connections on homogeneous manifolds). We have seen already in the example 6.1.3 that
the projection : P = G G/H = M onto a homogeneous manifold is a principal H-bundle. Any right
H-invariant Riemannian metric on G will result in a natural connection on this principal bundle in accordance
with the example 6.4.21 above. If the Riemannian metric on G is left-G and right H-invariant, and arising
from an inner product , on g, we can explicitly compute the connection form as follows. Let {X
i
}
n
i=1
be an
orthonormal basis for g with respect to , such that {X
i
}
m
i=1
form an orthonormal basis for h. Then, since
the left action of H on G is given by R
1
h
, the map
g
: H G is the map h gh
1
, i.e. the map L
g
I,
where I is the inversion map of H. Thus for X h, we get that

X
P
(g) = L
g
(X) =
X(g), where

X on the
right hand side is the left invariant vector eld on G generated by X. Clearly, then
ij
= e
i
(
X
P
j
) =
_
X
i
,

X
j
_
=
_
X
i
,

X
P
j
_
For a Z T
g
G, we have (Z) =
m
i=1
_
Z,

X
P
i
_

X
P
i
, which implies that
(Z) =
v
((Z)) =
m
i=1
_
Z,

X
i
_
X
i
If
L
denotes the left G-invariant Maurer-Cartan form on G, the reader may easily verify that this connection
form is nothing but (p
L
), where p : g h is the orthogonal projection with respect to , .
Exercise 6.4.23 (Stiefel bundles). Consider the principal O(k) bundle:
: V
k
(R
n
) G
k
(R
n
)
introduced in the exercise 6.1.6. The bi-invariant metric on O(n) generated by the inner product X, Y =
tr(X
Y ) on the Lie algebra o(n) descends to a left-O(n) right-O(k)-invariant metric on V

k
(R
n
) = O(n)/O(n
k). (Note that for an orthonormal k-frame (v
1
, ..., v
k
), where each v
i
is an n-column vector, there is a
left action of O(n) by (v
1
, .., v
n
) (Av
1
, .., Av
k
), and also the right action by O(k) given by (v
1
, .., v
k
)
(
j
B
j1
v
j
, ...,
j
B
jk
v
j
) for B O(k). It is the right action of O(k) which yields the principal bundle above).
Compute the connection form of this bundle for the connection given in the example 6.4.21 above, and also
the sections
i
and local connection forms
i
, after noting that a system of trivialising charts can be given on
the open sets U
P
introduced in 2.2.6.
Exercise 6.4.24. Let : P M be a principal G-bundle with a connection , and let f : N M be a
smooth map of manifolds. Note that the pullback G-bundle f
P is a bundle which completes the commutative

square:
f
P

P

N
f
M
where is G-equivariant. Show that there is a natural pullback connection denoted f
on f
P which satises
the following property:
(T
h
x
(f
P)) = T
h
(x)
P
Show that the local connection forms of f
P are given by f
i
: f
1
(U
i
) g, where
i
: U
i
g are the local
connection forms of (as in 6.4.17).
Example 6.4.25 (Canonical Cartan Connection on a Lie Group). Let G be a Lie group. We consider the
trivial principal G-bundle:
1
: GG G
(g, h) g
with the G-action being g.(h, k) = (g, (h, k)) = (h, gk). This means that for x = (h, k) P, the map
x
: G P is given by g (h, gk), i.e.
x
= (h, R
k
), and hence for X g,
x
(X) = (0, R
k
X) = (0,

X(k)) T
h
(G) T
k
(G)
where

X denotes the right invariant vector eld generated by X. If we denote by
2
: P G the second
projection (h, k) k, the vertical tangent bundle of P is given by T
v
P =
2
TG. Thus the Maurer-Cartan
section
v
of the vertical bundle is the pullback section
R
of the right-invariant Maurer-Cartan 1-form
R
of G. Since P = GG, TP =
1
(TG)
2
(TG) in a canonical manner, and by setting the natural connection
: TP T
v
P to be (X, Y ) Y , we get the connection form
(X, Y ) =
v
(Y ) =
R
(X, Y )
is therefore constant along horizontal slices Gk, and satises the correct transformation law under L
g
by
the corresponding fact for
R
as noted in the exercise 6.3.3.
Incidentally, one gets the tangent bundle TG of G as the associated bundle to this trivial prnicipal bundle
by taking the adjoint representation:
: G GL(g)
g Ad(g)(:= Ad(g)
)
For, we just map the element [(h, k), X] (P
g)
h
to the element L
h
(Ad(k
1
)X) =

(Ad(k
1
)X)(h)
T
h
G. Then [g.(h, k), (g)X] = [(h, gk), Ad(g)X] will get mapped to
L
h
(Ad(gk)
1
Adg X) = L
h
(Ad(k
1
)X)
and this map is well dened independent of representatives. We leave it as an exercise for the reader to dene
the inverse isomorphism.
6.5. Horizontal lifting of vector elds and paths.
Denition 6.5.1 (Horizontal elds and lifts). Let U be an open subset of P, where : P M is a principal
G-bundle with a connection . We say a vector eld Y on U is horizontal if Y (x) T
h
x
P for each x U.
Equivalently, if (Y ) 0 at all points of U, which is the same as requiring (Y ) 0 on U (since
v
= ).
Note that for a xed g G, L
g
Y is a horizontal eld i Y is a horizontal eld.
If X is a smooth vector eld on an open set U M, we say a smooth horizontal vector eld

X on
1
(U)
is a horizontal lift of X if
X) = X. Note that in view of the remark above, we will have

X(hx) = L
h
X(x),
since both sides are horizontal vectors at hx mapping to X((x)) under
, and the only things in the kernel

of
are vertical vectors.

86 VISHWAMBHAR PATI
Remark 6.5.2. Every vector eld X on an open subset U of M has a horizontal lift. For, this vector eld
leads to the pullback section
X of
TM over
1
(U), dened by
(X)(x) = X((x)). Since ker
= T
v
P,
: T
h
P
TM is an isomorphism of vector bundles. Thus the smooth vector eld

X :=
1
X) is the
required horizontal lift.
Proposition 6.5.3 (Horizontal lifts of smooth paths). Let : P M be as above, with a connection . Let
c : [0, 1] M be a smooth path in M. Let x P be a point in the bre
1
(c(0)). Then there exists a unique
path c : [0, 1] P such that:
(i): c(0) = x.
(ii):
dc
dt
(t) is horizontal for all t.
This path c will be called the horizontal lift of c beginning at x. For each g G, the path L
g
c will be the
horizontal lift of c beginning at gx, i.e. all the horizontal lifts of c are left translates of each other.
Proof: Clearly, we can assume that c extends to a smooth map c : (, 1 +) M. Let us denote (, 1 +)
as I. We consider the pullback G-bundle Q := c
P on I. We therefore have a commutative pullback square:

Q

P

I
c
M
where is a smooth G-equivariant isomorphism of the bre Q
t
=
1
(t) with the bre P
c(t)
=
1
(c(t)). Let
denote the pullback connection on Q, as in exercise 6.4.24.

On I we have the natural coordinate eld Y =
d
dt
such that c
(
d
dt
) =
dc
dt
. By the remark 6.5.2 above, there
is a horizontal lift

Y of this eld Y to Q. Also, as remarked at the end of 6.5.1, the horizontal vector eld

Y
is invariant under L
g
for all g G.
Claim: For each x Q
0
=
1
(0), there is a smooth path a
x
: [0, 1] Q such that :
(i): a
x
(0) = x
(ii): a
x
(t) = t for all t [0, 1], and
(iii): a
x
(t) is a trajectory of

Y .
(iv): a
gx
(t) = L
g
a
x
(t) for all t [0, 1].
Proof of claim: Fix some x
0

1
(0). Then, by the existence theorem for ODEs (5.4.2), there is the germ
of a smooth path a
x0
: (, ) Q which satises (i), (ii) and (iii) for all t [0, ). Also note that for all
g G, we can dene a
gx0
: (, ) Q as L
g
a
x0
. Since

Y is invariant under L
g
, and every x Q
0
is of the
form gx
0
by the uniqueness part of 5.4.2, it is clear that a
gx0
will also satisfy the properties (i) through (iv),
for x = gx
0
.
Say a subinterval [0, ] or [0, ) with 1 is good, if there exists a
x
: [0, ] Q satisfying (i) through
(iv) above for all t [0, ]. The remarks above show that the collection of good (right closed or right open)
subintervals is non-empty. Order it by inclusion, and observe that the union of a chain provides an upper
bound for that chain. Hence, by Zorns lemma, there is a maximal element [0, b] or [0, b). In either case [0, b) is
good. If b < 1, one can take a point y
1
(b), and construct a path : (2, 2) Q by 5.4.2 which will be
a trajectory of

Y , and satisfy (0) = y. Since [0, b) is good, there is a path a
x
: [0, b) Q satisfying (i) through
(iv) above all over [0, b]. Since a
x
(b)
1
(b), there is a group element g such that g(()) = a
x
(b).
Then the path dened on [0, b +2) by t aking it to be a
x
on [0, b), and L
g
(t b) on [b, b +2) is a smooth
path (by uniqueness in 5.4.2, because a
x
and L
g
(t b) must agree for the overlap [b 2, b]). This path
satises (i) through (iv) on [0, b +2), and shows that [0, b +2) is also good. This contradicts the maximality
of [0, b] (or [0, b)), and hence b = 1. Thus the claim follows. 2
Now to prove the proposition, let x P, and set y =
1
(x) Q. The smooth path a
y
, where a
y
is the
path constructed in the above claim, is easily checked to be the required horizontal lift of c. 2.
It is useful to write down the dierential equation that the horizontal lift of a smooth path satises. We do
this in the next lemma.
Lemma 6.5.4 (ODE for horizontal lifts). Let {(
i
, U
i
)} be a bundle atlas for the principal G-bundle : P
M. Let
i
(x) =
1
i
(x, 1) be the smooth local sections over U
i
, and
i
=
i
the local connection forms. Let
i
(t) denote
i
(c(t)). Then, for a path c : [a, b] U
i
, the horizontal lift c(t) beginning at z is given by:
c(t) = g(t)
i
(t)
where z = g(0)
i
(0) and
R
_
dg
dt
_
+ Adg(t)
i
(c(t))
_
dc
dt
_
= 0
In the particular case when G is a linear group, we get the equation:
dg
dt
+L
g(t)
i
(c(t))
_
dc
dt
_
= 0
Proof: Since c(t) is a path in
1
(U
i
), we can certainly write
c(t) = g(t)
i
(t)
for a uniquely dened smooth function g : [a, b] G. Now, we need to express the condition (
dc
dt
) = 0.
Letting dots denote t-derivative
d
dt
, we have:
(
c(t)) =
v
(
c(t))
=
v
(D(g(t),
i
(t))( g(t),
i
( c(t)))
=
v
_
R
( g(t)) +L
g(t)
i
( c(t))
_
=
R
( g(t)) + Adg(t)(
i
( c(t)))
=
R
( g(t)) + Adg(t)
i
( c(t))
where we have used the equation (11) from the proposition 6.4.15 in the third line, and put dots to denote
d
dt
.
This proves the lemma. The second part is clear since by 6.3.2,
R
= dg.g
1
, and by denition dg( g(t)) = g(t).
2
6.6. Connections on vector bundles, parallel transport. Connections provide a method of dierentiating
sections of vector bundles. To do this, one needs the notion of what sections are constant, i.e. have zero
derivative. For this one introduces parallel transport or parallel translation. So let : P M be a principal G-
bundle, and let : G GL(V ) be a representation of G on a vector space V . For notational simplicity we will
write (g)v as g.v or even gv, where g G and v V . Let be a connection on P, and let , (resp.
i
) denote
the g-valued connection 1-forms on P (resp. U
i
) as in the previous subsections. Let p : E(= P
V ) M be
the associated vector bundle, as in example 6.2.2.
Let y M and v E
y
= p
1
(y). That is v = [x, w] where x P
y
=
1
(y). Let c : [0, 1] M be a smooth
path with c(a) = y, and c : [a, b] P be a horizontal lift of c starting at x, as in 6.5.3. Note that [c(a), w] = v.
88 VISHWAMBHAR PATI
Denition 6.6.1. The vector dened by P
c
t
v := [c(t), w] is called the parallel translate of v by t along the
path c. Note that p(P
c
t
(v)) = (c(t)) = c(t), so that P
c
t
v E
c(t)
. It is easily checked that this mapping:
P
c
t
: E
c(a)
E
c(t)
is linear. In fact, as an exercise the reader can check that:
(i): P
c
0
= Id, and P
c
t
P
c
s
= P
c
t+s
, whenever both sides are dened.
(ii): P
c
t
is an isomorphism of E
c(0)
onto E
c(t)
. (Its inverse is P
c
t
which just denotes P
d
t
, where d : [0, t]
M is the path d(s) = c(a +t s)).
For notational convenience we shall write P
t
instead of P
c
t
. We remark that for two distinct paths joining the
same two points y, y
, the two corresponding parallel transport isomorphisms from E

y
to E
y
will be distinct.
Let c be a smooth path as above. Clearly, any frame (v
1
, .., v
k
) of a bre p
1
(y) = E
y
can be expressed as
v
i
= [x, e
i
], where (e
1
, .., e
k
) is a frame for V (why?). Denote v
i
(t) := P
t
v
i
= [c(t), e
i
]. This will be a frame
for E
c(t)
. If v =

k
i=1
a
i
v
i
, then we have P
t
v =

k
i=1
a
i
v
i
(t). The frame ( v
1
(t), ..., v
k
(t)) is called the parallel
frame along c with initial value (v
1
, .., v
k
).
Denition 6.6.2 (Covariant derivative). Let p : E M be a vector bundle, with E = P
V , and : P M
a principal G-bundle with a connection . Let s : U E be a smooth section of E over the open set U M,
and X T
y
M be a tangent vector at y U. We dene the covariant derivative
X
s to be the vector in E
y
given by the formula:
X
s = lim
t0
P
c
t
(s(c(t)) s(c(0))
t
where c : (, ) U is any smooth curve with c(0) = y and c(0) = X.
We need to verify that the denition above makes sense, i.e., is independent of the curve c that was chosen.
This will follow as soon as we write down equations. Meanwhile, let us just note that if ( v
j
(t) = P
c
t
v
j
) is the
parallel transport of v
j
where (v
1
, ..., v
k
) is a frame for E
y
, then we may express s(c(t)) =
j
s
j
(t) v
j
(t), from
which it follows that P
c
t
s(c(t)) =
j
s
j
(t)v
j
, so that:
X
s =
j
ds
j
dt
(0)v
j
From this it easily follows that if f : U R is a smooth function, then for the smooth section fs we have the
Leibniz Formula:
X
(fs) = f(c(0))
X
s +X(f)s (14)
Now, to show that
X
does not depend on the choice of path germ c satisfying c(0) = y and c(0) = X, we
need only check this for a suitable basis of smooth sections over U, in the light of the Leibniz formula above.
For this purpose, without loss of generality, we assume that U is a coordinate chart, and : U g
is the local connection form on U. Let {v
i
= [x, e
i
]}
k
i=1
be a frame of E
y
, where e
i
is a frame of V , and
x P
y
. Let c be a path germs at y with c(0) = y, and c(0) = X. Let c denote its horizontal lift beginning
at x. Let : U
1
(U) be the natural section coming from the bundle trivialisation, and w.l.o.g. take
x = (0) := (c(0)). Then, we have a natural frame (basis of smooth sections) of E all over U coming from ,
namely
w
i
(z) = [(z), e
i
] for z U
Proposition 6.6.3 (Covariant derivative in local coordinates). With the setting above, the covariant deriva-
tive
X
w
j
is given by:
X
w
j
=
i
((X))
ji
w
i
(c(0))
where := D(1) : g gl(V ) is the derived representation. (see the exercise 5.6.7).
Proof: Note that s
j
(c(0)) = v
j
. The parallel translated frame is:
v
i
(t) = [c(t), e
i
]
We write c(t) = g(t)(c(t)), as before. Then:
w
i
(t) = [(c(t)), e
i
] = [g(t)(c(t)), (g(t))e
i
]
= [c(t),
j
(g(t))
ji
e
j
]
=
j
(g(t))
ji
[c(t), e
j
]
=
j
(g(t))
ji
v
j
(t)
where [(g)
ij
] is the matrix of (g) with respect to the basis {e
i
} of V .
Since P
t
v
j
(t) = v
j
by denition, we have P
t
w
i
(c(t)) =
j
(g(t))
ij
v
j
. Thus can compute:
X
w
j
=
i
_
(g(t))
ji
(t) (g(0))
ji
t
_
v
i
=
i
D
ji
(1)(
dg
dt
(0))v
i
since g(0) = 1. Now, by the ODE of 6.5.4,
dg
dt
+L
g(t)
(c(t))
_
dc
dt
_
= 0
where g(0) = 1. Thus (
dg
dt
(0)) = (c(0))( c(0)) = (y)(X). If we denote the derivative of : G GL(V ) at
the identity by , we then have the formula:
X
w
j
=
i
((X))
ji
v
i
Since v
i
= w
i
(c(0)), we may as well write this as:
(
X
w
j
) =
i
((X))
ji
w
j
(c(0))
This formula clearly shows that
X
w
j
depends only on c(0) and X = c(0). 2
Remark 6.6.4. From the local formula of the above proposition, it is clear that
cX
s = c
X
s
for all c R. In particular, if X is a tangent vector eld on U, we may dene, for a smooth section s : U E,
another section
X
s by the formula (
X
s)(y) :=
X(y)
s. The formula above shows that this is also a smooth
section of E on U. In other words, we have a linear operator:
: (U, TM)
R
(U, E) (U, E)
X s
X
s
where (U, E) (resp. (U, TM)) denotes the C
(U) module of smooth sections of E over U (resp. vector

elds over U). This operator is linear with respect to multiplication by a smooth function f C
(U) in the
rst (=X) variable, but obeys the Leibniz rule with respect to the such a multiplication in the second (=s)
variable.
90 VISHWAMBHAR PATI
Corollary 6.6.5. A smooth connection on : P M gives rise, for every open subset U of M, the covariant
derivative operator:
: (U, E)
1
(U) (U, E)
s s
where (U, E) denotes the C
(U)-module of smooth sections of E at y M. It satises the following Leibniz

Formula):
(fs) = fs +df s
Proof: Clear from the remark above, by dening (s)(X) =
X
s for a tangent vector eld X, and noting
that df(X) = X(f). 2
Conversely, given a covariant derivative operator on a vector bundle E, one can create a connection on the
associated frame bundle. This is the content of the:
Exercise 6.6.6. Let E be a smooth vector bundle of rank k. Fix a vector bundle atlas {(
i
, U
i
)} for E, and let
(for example) {s
i
(x)} denote the frame {
1
i
(x, e
1
), ..,
1
i
(x, e
k
)}. Note that on U
i
U
j
, we have the formula
s
i
(x) = g
ij
(x)s
j
(x) where g
ij
(x) GL(k, R). Suppose there is given a covariant dierentiation operator
satisfying the Leibniz formula in 6.6.5 above. Set
s
i
=
i
.s
i
where
i
is a matrix of 1-forms on U, and if we denote its component 1-forms by
kl
i
, and s
i
= (s
1
i
, .., s
k
i
) then
(
i
.s
i
)
k
:=
l

kl
i
s
l
i
. Using the Leibniz formula, show that the collection of gl(k) valued 1-forms
i
satisfy
the transformation property of 6.4.17, and hence dene a connection on the principal GL(k) frame bundle of
E.
Exercise 6.6.7 (Covariant derivative w.r.t. Cartan connection). Let G be a Lie group, and let be the Car-
tan connection introduced in the example 6.4.25. For two left invariant vector elds

X,
Y , compute the covariant

derivative
X
Y (on the tangent bundle TG).

6.7. Metric connections. Let p : E M be a smooth vector bundle of rank k, and a Riemannian bundle
metric , . We will denote the inner product on the bre E
y
by , . Thus we may form the bundle of
orthonormal frames : P M, which will be a principal O(k) bundle. Its bre over y will be the set of all
orthonormal frames in E
y
. (That is E will admit a reduction of structure group from GL(k, R) to O(k) in the
sense of 6.2.6.
Denition 6.7.1. A connection on the principal O(k)-bundle of a Riemannian vector bundle E will be called
a metric connection or Riemannian connection.
There is a formula that connects covariant dierentiation of sections on E with respect to a metric connection,
and the metric itself. That is:
Proposition 6.7.2. Let E be as above, and let X T
y
(M), and and two smooth sections of E over U,
an open neighbourhood of y in M. Then
X
, +,
X
= X, (15)
for the covariant derivative
X
with respect to any metric connection.
Proof:
If c is a path germ at y M, the horizontal lift

c(t) beginning at any point x P
y
will be a path of
orthonormal frames. If v
i
= [x, e
i
], for some orthonormal frame e
i
V , then the parallel translated frame
along c is v
i
(t) = [c(t), e
i
].
The isomorphism P
V E is the map [(w

1
, .., w
k
),
j
a
j
e
j
]

j
a
j
w
j
, as noted in 6.2.3. Thus
v
i
(t) = c
i
(t). Consequently
v
i
(t), v
j
(t)
c(t)
= c
i
(t), c
j
(t)
c(t)
=
ij
since c(t) is an orthonormal frame for all t. That is,
P
c
t
v
i
, P
c
t
v
j
c(t)
= v
i
, v
j
and parallel translation along each curve is an isometry with respect to , . In particular, for any sections ,
of E, we have:
P
c
t
(c(t)), P
c
t
(c(t))
_
c(0)
= (c(t)), (c(t))
c(t)
Taking
d
dt |t=0
, and setting c(0) = X shows that:
X
, +,
X
= X,
Remark 6.7.3. If e = (e
1
, .., e
k
) is a local orthonormal frame of E over some open set U, then we recall that
X
e
i
=
il
(X)e
l
where : U o(k) is the local connection form over U. Since e
i
, e
j
is the constant matrix
ij
, the relation
(15) above implies that:
ij
(X) +
ji
(X) = 0
which is precisely the statement that is o(k) valued, and hence a skew-symmetric matrix.
Clearly, since every principal bundle admits many connections, there is nothing unique about a metric
connection. However, on the tangent bundle of a Riemannian manifold, there is a unique connection if one
imposes an extra (torsion free) condition, which is the content of the following subsections.
6.8. Covariant dierentiation on Vector Bundles. If one wishes to avoid considering principal G-bundles
altogether, it is possible to directly dene connections on vector bundles.
We recall some notation rst. If : E M is a vector bundle over a smooth manifold M, and U M
is an open set, the R vector space of smooth sections of this bundle over U will be denoted (U, E), as
usual. Multiplication by smooth functions on U makes this vector space a module over C
(U). The space of

dierential 1-forms on U is denoted by
_
1
(U), which is precisely (U, T
(M)), and therefore also a module

over C
(U).
Denition 6.8.1 (Connection on a vector bundle). Let : E M be a vector bundle. A connection or
covariant derivative on this vector bundle is an R-linear operator:
: (U, E)
1
(U) (U, E)
s s
for each open set U M. This operator should further satisfy the following Leibniz formula:
(fs) = fs +df s (16)
for all f C
(U), all s (U, E), and all U M open. A connection on the tangent bundle TM of a
manifold M is called a connection on M.
Note that this denition is just the axiomatisation of what we dened in 6.6.5 of the previous subsection.
92 VISHWAMBHAR PATI
Let X T
x
(M) be a tangent vector. We note that there is a natural operation of contraction dened by:
T
x
(M) T
x
(M) R
(X ) i
X
:= (X)
which leads to a natural (pointwise) operator:
T
x
(M)
1
(U) (U, E) E
x
(X s) (i
X
(x))s(x) = (x)(X)s(x)
which we also denote i
X
( s), for notational convenience.
If U M is an open set, then the above operators extend as natural (local) operators over U:
(U, TM)
1
(U) C
(U)
(X ) i
X
:= (X)
which satises i
fX
= fi
X
= i
X
f for all f C
(U), by denition.
Hence we have the corresponding (local) operator:
(U, TM)
1
(U) (U, E) (U, E)

(X s) (i
X
)s
where the smooth function i
X
on U has the obvious pointwise denition (i
X
)(x) = i
X(x)
((x)). Also note
that we have
i
fX
( s) = fi
X
( s) = i
X
( fs)
for all smooth functions f C
(U), all
_
1
(U) and all s (U, E).
Denition 6.8.2 (Covariant derivative with respect to a tangent vector). Let be a connection on a smooth
vector bundle : E M. Let x U, where U M is an open set. For a tangent vector X T
x
(M), dene:
X
: (U, E) E
x
s
X
s = i
X
(s)
and similarly dene for a vector eld X (U, TM), the operator:
X
: (U, E) (U, E)
s
X
s = i
X
(s)
where i
X
are the two operators dened above. It is trivial to check the following, using the facts about i
X
mentioned above, and the Leibniz rule, that:
(i):
fX
s = f
X
s
(ii): (Leibnitz formula)
X
(fs) = f
X
s +X(f)s
for all for all smooth functions f C
(U), and all s (U, E). This is again exactly what we dened in 6.6.2
of the last subsection.
Remark 6.8.3. A connection on the bundle : E M can also be dened just as a global R-linear operator
operator
: (M, E)
1
(M) (M, E)
which satises the Leibnitz rule (fs) = fs+df s for all f C
(M) and all s (M, E). This is because

one can use cut-o functions etc. to dene the operator on all open subsets U M. The Leibnitz formula
above tells us how to deal with multiplication by a function, and the details are left to the reader.
Remark 6.8.4 (Local description of a connection on a vector bundle). If {s
j
}
k
j=1
is a frame (= basis of sec-
tions) of the rank-k vector bundle : E M on some open set U M, then we can dene a k k-matrix of
1-forms [
ij
] on U by:
s
j
=
ij
s
i
which is sometimes written in more compact notation simply as
s = .s
where s denotes the frame {s
j
}
k
j=1
, and denotes the matrix [
ij
], and the dot denotes matrix tensor product.
These 1-forms will change by a formula, if we change the frame to another frame. More precisely, if g is
an invertible matrix of smooth functions, and we have another frame t := gs (matrix multiplication!), with
corresponding matrix , then:
.t = .gs = t = (gs) = gs +dg.s = g(.s) +dg.s
by Leibnitz rule, so that:
g = g +dg
and hence
= gg
1
+ (dg)g
1
where dg is the k k-matrix whose (ij)-th entry is dg
ij
, and all multiplications are matrix multiplications.
This is the same formula we had encountered in 6.4.15 (see also the exercise 6.6.6).
From the transformation formula above, it also follows that the matrix of connection 1-forms
dened
from some local frames s
on trivialising open sets {U
} coming from some bundle atlas will not patch up and

give a global matrix of 1-forms on M. If we write s
= g
, the term dg
.(g
)
1
in the formula shows
that
(x) does not depend just on g
(x) and
(x), but also on dg
.
6.9. The Levi-Civita Connection on a Riemannian Manifold.
Proposition 6.9.1 (Levi-Civita connection). Let M be a Riemannian manifold with Riemannian metric g.
For tangent vectors X, Y T
x
(M), denote g(x)(X, Y ) = X, Y . Then there is a unique connection on M (i.e.
on the tangent bundle : TM M which satises:
(i):
X
Y
Y
X = [X, Y ] for all smooth vector elds X, Y on M (torsion free condition).
(ii): XY, Z =
X
Y, Z +Y,
X
Z (compatible with metric).
It is called the Levi-Civita connection on M.
Proof: Let (, U) be a local chart on M, and let

xj
denote the corresponding coordinate vector elds on
U. For notational convenience, we will denote

xj
by
j
. Recall that these coordinate elds commute, i.e.
[
i
,
j
] = 0 for 1 i, j n.
Unfortunately, these coordinate elds
i
will not constitute an orthonormal frame. However any connection
on the principal O(n) frame bundle of TM will result in a covariant derivative operator on TM, which in turn
will result in a connection on the principal GL(n)-bundle of all frames in TM, call it : P M. Using the
section : U
1
(U) dened by (x) = (
1,x
, ..,
n,x
), we can pullback the connection form on P to the
local connection form
on U. For notational convenience, we call it . Clearly the connection is completely

determined on U once we specify (
i
) gl(n, R).
If is a covariant derivative operator on M, it is completely determined on U by
i
for i = 1, .., n. These
in turn are completely determined on U by the vector elds
i
j
. We may expand this last eld as:
j
=
k
ij
k
94 VISHWAMBHAR PATI
where we use the Einstein summation convention, i.e. repeated indices are summed over 1 through n. These
k
ij
are smooth functions on U, and are called the Christoel symbols or connection coecients of the
connection. They are related to the matrix of connection 1-forms as follows:
k
ij
= (
i
)
jk
where is the local connection form on U dened above in remark 6.8.4. This is clear from the relation
j
= (.)
j
=
k

jk
k
, as in the remark 6.8.4 ( and also proposition 6.6.3).
Let g = , denote the Riemannian metric. On U, this metric is completely determined by the symmetric
matrix of smooth functions [g
ij
(x)] where g
ij
(x) =
i
,
j
x
. Since (ii) above is to hold, we get
k
g
ij
=
i
,
j
+
i
,
=
l
ki
g
lj
+
l
kj
g
il
The above equation is called the First Christoel Identity. By cyclically permuting i, j, k we get the equations
and:
k
g
ij
=
l
ki
g
lj
+
l
kj
g
il
i
g
jk
=
l
ij
g
lk
+
l
ik
g
jl
j
g
ki
=
l
jk
g
li
+
l
ji
g
kl
Now, since the connection is to be torsionless, we have:
j

j
i
= [
i
,
j
] = 0
so that
k
ij
=
k
ji
. Using this, and the symmetry of g
ij
, we subtract the second of the above equations from
the sum of the rst and third to obtain:
1
2
(
k
g
ij
+
j
g
ik
i
g
jk
) =
l
jk
g
li
Multiplying the equation above by g
im
, where [g
im
] is the inverse matrix of g
ij
, and summing over i, we obtain:
m
jk
=
1
2
g
im
(
k
g
ij
+
j
g
ik
i
g
jk
) (17)
This determines the Christoel symbols, and hence the connection form in terms of the metric g
ij
on any
coordinate patch U. and is called the second Christoel Identity. Hence the connection is uniquely determined
by the conditions (i) and (ii), and is the Levi-Civita connection. 2
Remark 6.9.2 (Torsion of a connection). Let be any connection on a manifold, viz., a connection on its
tangent bundle. Then, for vector elds X, Y on M, dene the torsion of the connection by:
(X, Y ) =
X
Y
Y
X [X, Y ]
A priori, the denition above just tells us that is an R- bilinear map (M, TM) (M, TM) (M, TM).
However we rst make the:
Claim: The tangent vector [(X, Y )](x) T
x
(M) depends only on X(x) and Y (x). In particular, gives a
well dened bundle morphism TM TM TM, that is, it is a tensor of type (1, 2), and because of skew
symmetry, a section of TM
_
2
(T
(M)) (=a TM-valued 2-form on M).

To see the claim, we rst note that for smooth functions f, g on M, we have:
(fX, gY ) = fg(X, Y )
Because (X, Y ) = (Y, X) by denition, it is enough to see that (X, fY ) = f(X, Y ). This follows because
(X, fY ) =
X
(fY )
fY
(X) [X, fY ]
= (f
X
Y +X(f)Y ) f
Y
X (X(f)Y +f[X, Y ])
= f (
X
Y
Y
X f[X, Y ]) = f(X, Y )
by using (i) and (ii) of 6.8.2, and (ii) of 5.5.2. In fact, this same formula holds for vector elds X, Y on an
open set U, and smooth functions f, g C
(U).
Now let X be a vector eld that vanishes at a point x, where x U
for some coordinate patch U
. Letting
i
be the coordinate elds on U
, we may write X on U
as:
X =
n
i=1
f
i
i
where f
i
are smooth functions on U
, and f
i
(x) = 0. Then
(X, Y ) =
i
(f
i
i
, Y ) =
i
f
i
(
i
, Y )
by the observations above. Thus
[(X, Y )](x) =
i
f
i
(x)(
i,x
, Y (x)) = 0
By skew-symmetry, it similarly follows that [(X, Y )](x) = 0 if Y (x) = 0. Hence, if X(x) =

X(x), and Y (x) =
Y (x) for some smooth vector elds X,

X, Y,

Y , then we have [(X, Y )](x) = [(
X,
Y )](x) =: (X(x), Y (x)),

and hence the claim.
Example 6.9.3 (Connections on a Lie group). We recall the (left-invariant) Cartan connection on a Lie group
G (see 6.4.25 and 6.6.7). It is the unique connection on TG satisfying:
Y = [
X,
Y ] =
[X, Y ]
for X, Y g, and

X,
Y the corresponding left- invariant vector elds. The properties ((i) and (ii) of 6.8.2) of
a connection then determines this connection on all smooth vector elds on G.
This natural Cartan connection on a Lie group is not torsionless. In fact, from the denition above it follows
that for left invariant vector elds

X and

Y we have:
Y
Y
X [
X,
Y ] = [
X,
Y ]
There is an easy way to remedy this, of course. If we dene
X
Y =
1
2
[
X,
Y ], then we nd that this new

connection is torsionless.
The Cartan connection does however satisfy a formula analogous to that of a metric connection. There is a
symmetric bilinear form dened on the Lie algebra g called the Killing form, as B(X, Y ) = tr(adXadY). It is
an easy exercise to check (using the Jacobi Identity) that it has the following ad-invariance property:
B((adX)Y, Z) +B(Y, (adX)Z) = B([X, Y ], Z) +B(Y, [X, Z]) = 0
This will lead to a bi-invariant symmetric form on the tangent bundle by setting
_
X,
Y
_
:= B(X, Y ). Then,
for left-invariant vector elds, since
_
Y ,

Z
_
is a constant function on G, we will have:
_
Y ,

Z
_
+
_
Y ,
X
Z
_
= 0 =

X
_
Y ,

Z
_
(18)
Unfortunately, in general, this symmetric form could be degenerate, or even identically 0 (e.g. if the group is
abelian). For a compact semisimple Lie group (e.g. SO(n), U(n)) however, this Killing form is negative denite,
and so changing its sign yields a bi-invariant Riemannian metric on G. In that case, setting
X
Y =
1
2
[
X,
Y ]
would yield a torsionless connection, and in view of the formula (18) above, yield the unique Levi-Civita
connection on G.
However, in the general case, it is useful to write down the Levi-Civita connection on any Lie group, with
respect to a left-invariant metric , on G, and work the above (compact) case out as a special case. Let
{X
i
}
n
i=1
be an orthonormal basis of g, with respect to , , so so that we have a global smooth left-invariant
orthonormal frame eld of TG given by the left-invariant vector elds

X
i
. The structure constants c
k
ij
of the
Lie algebra are dened by the formula:
[X
i
, X
j
] =
k
c
k
ij
X
k
96 VISHWAMBHAR PATI
which implies c
k
ij
= [X
i
, X
j
], X
k
. (They satisfy certain obvious relations, i.e. c
k
ij
= c
k
ji
, and another
quadratic identity coming from the Jacobi identity).
Let us express
Xi
X
j
=
k
ij

X
k
where
k
ij
are smooth functions on G (It will turn out presently that they are constants). The torsion free
condition immediately implies that:
i
jk
i
kj
= c
i
jk
(19)
The metric connection condition implies that:
k
ji
+
i
jk
=
_
Xj
X
i
,

X
k
_
+
_
X
i
,
Xj
X
k
_
=

X
j
_
X
i
,

X
k
_
=

X
j
X
i
, X
k
= 0 (20)
Permuting i, j, k cyclically, we get the equations:
k
ji
+
i
jk
= 0
i
kj
+
j
ki
= 0
j
ik
+
k
ij
= 0
By subtracting the third form the sum of the rst two, and using (19), we have:
i
jk
+
i
jk
= c
j
ki
+c
k
ij
which, together with the equation (19) implies that
i
jk
=
1
2
_
c
i
jk
c
j
ki
+c
k
ij
_
In particular, the
i
jk
s are all constants, and we get the formula:
Xj
X
k
=
i
_
c
i
jk
c
j
ki
+c
k
ij
_

X
i
(21)
If the inner product is bi-invariant, then c
k
ij
= [X
i
, X
j
], X
k
= X
j
, [X
k
, X
i
] = c
j
ki
, and in this case our
formula (21) reads as:
Xj
X
k
=
1
2
i
c
i
jk

X
i
=
1
2
[
X
j
,

X
k
] =
1
2
[X
j
, X
k
]
which is the formula we had stumbled upon earlier. In fact, for a compact semisimple group, every bi-invariant
metric is a scalar multiple of B above, and the formula above is therefore just the statement of ad-invariance
of the Killing metric B.
In particular, if the Lie group is abelian, then with respect to the Levi-Civita connection with respect to a
left invariant metric,the left-invariant vector elds will have zero covariant derivatives (ie. will be covariantly
constant) with respect to any tangent vector at any point.
Exercise 6.9.4. Compute the Christoel symbols of the Levi-Civita connection on the sphere S
2
, and the
hyperbolic spaces H
2
, H
3
.
6.10. Geodesics. Let M be a Riemannian manifold, and let be a connection on the principal GL(n, R)-frame
bundle P of TM, leading to a covariant dierentiation operator on the vector bundle TM. If c : (a, b) M is
a smooth curve in M, then it is clear that the pullback of the tangent bundle c
TM will have the corresponding

pulled back principal bundle c
P on (a, b) as its principal frame bundle. It will also have the pullback connection
c
introduced in the exercise 6.4.24. Note that

d
dt
is a smooth vector eld on (a, b).
Denition 6.10.1 (Covariant derivative along a curve). A vector eld X along the curve c is dened to be a
smooth section of c
TM. (In particular X(t) T

c(t)
M for all t. However note that X is not a function of the
point c(t) but of t. If c crosses itself, X will be multivalued at the point c(t)). If X is a vector eld along
the curve c, we dene the covariant derivative along the curve c to be:
DX
dt
:= d
dt
X
with respect to the above pulled back connection. It will again be a smooth section of c
TM on (a, b). We will

say that the vector eld X along c is parallel along c if
DX
dt
= 0. It is an exercise to check that if c is an arc,
i.e. doesnt cross itself, then being parallel along c is equivalent to saying that X(t) = P
c
t
X(0).
Lemma 6.10.2 (Local expression for
D
dt
). If we let c : (a, b) U, be a smooth curve, where U is a coordinate
chart, and let X be a vector eld along c, then:
DX
dt
=
j
_
_
dX
j
dt
+
k,i
j
ki
dc
k
dt
X
i
_
_
j
where
j
are the pullback sections c
_

xj
_
of c
TM on (a, b), and X(t) =
n
i=1
X
i
(t)
i
.
Proof: Writing X(t) =
i
X
i
(t)
i
, where
i
are the pullbacks of the coordinate elds

xi
by c, we have:
DX
dt
= d
dt
(
i
X
i
i
)
=
i
_
dX
i
dt

i
+X
i
D
dt
(
i
)
_
from the Leibniz formula.
Now, the if the local connection form on U is , then the connection form on (a, b) for the pullback connection
is c
(by the exercise 6.4.24). Thus, for the frame

i
of c
TM, we have, by 6.6.3,

D
dt
(
i
) = d
dt
(
i
)
=
j
(c
)
_
d
dt
_
ij
j
=
_
dc
dt
_
ij
j
=
j,k
dc
k
dt

_

x
k
_
ij
j
=
j,k
j
ki
(c(t))
dc
k
dt

j
where the last line follows from the denition of
j
ki
as (

x
k
)
ij
(e.g. in the proposition 6.9.1). Thus we have
DX
dt
=
j
_
_
dX
j
dt
+
k,i
j
ki
dc
k
dt
X
i
_
_
j
98 VISHWAMBHAR PATI
proving the lemma. 2
Denition 6.10.3. If c : (a, b) M is a smooth curve, then the velocity vector
dc
dt
is naturally a section of
c
TM, and the vector eld

D
dt
_
dc
dt
_
along c is called the covariant acceleration or geodesic curvature of c. If we
denote
dc
dt
as the velocity eld T along c, then this quantity is often denoted
T
T, by an abuse of language.
We say c is a geodesic for the connection if
D
dt
_
dc
dt
_
0.
Example 6.10.4. If we look at the canonical Cartan connection on a Lie group G, (see 6.4.25), then by the
exercise 6.6.7, we have
X
X = [
X,

X] = 0. Since the integral curves to the left invariant vector eld

X have
velocity elds along them to be the restriction of

X (to them), it is clear that these integral curves are geodesics
for the Cartan connection. Thus, for the Cartan connection, the one parameter groups (exp(tX)) , and all
xed left translates g. exp(tX), for X g, are geodesics.
From 6.9.3 it also follows that these are all the geodesics for the Levi-Civita connection of a bi-invariant
metric. In general, however, they will not be geodesics for the Levi-Civita connection with respect to a left
invariant metric. In fact, there is the following :
Exercise 6.10.5. Using the standard basis X
1
= H, X
2
= X, X
3
= Y of the group SL(2, R) introduced in
5.8.7, compute the connection coecients
k
ij
(by using the computation in 6.9.3) for the Levi-Civita connection
of the left-invariant metric coming from the metric X
i
, X
j
=
ij
on sl(2, R). Check that
X2
X
2
is not zero,
and hence the one parameter group exp(tX
2
) is not a geodesic for this connection.
Proposition 6.10.6. Let M be a Riemannian manifold, and X and Y be vector elds along a smooth curve
c : (a, b) M. Then for all t (a, b), X(t), Y (t) T
c(t)
M, and we can dene the pullback Riemannian metric
X, Y
t
:= X(t), Y (t). Then, for
D
dt
with respect to the Levi-Civita connection on M:
d
dt
X, Y
t
=
_
DX
dt
, Y
_
t
+
_
X,
DY
dt
_
t
Proof: For the pullback bundle c
TM, the pullback c
of the Levi-Civita connection on M, which is a

metric connection on the principal orthogonal frame bundle P of TM, will be a connection on the pullback
orthogonal frame bundle c
P. But c
P is precisely the orthogonal frame bundle of c
TM with respect to the

pullback Riemannian bundle metric on TM. Thus c
is a metric connection. The result follows from the

proposition 6.7.2 applied to the metric connection c
and the resulting covariant derivatives in the Riemannian

vector bundle c
TM on the manifold (a, b). 2

Remark 6.10.7 (Parallel transport in the Levi-Civita connection). Let X be a Riemannian manifold equipped
with the Levi-Civita connection. Then P(t) : T
c(0)
(X) T
c(t)
is an isometry with respect to the corresponding
metrics ,
c(0)
and ,
c(t)
on the domain and range respectively.
Proof: This follows because if v, w T
c(0)
(X), and we denote their parallel translated vector elds along c by
V (t) and W(t) respectively, then:
d
dt
P
t
(v), P
t
(w)
c(t)
=
DV (t)
dt
, W(t)
c(t)
+V (t),
DW(t)
dt

c(t)
= 0
because the connection is compatible with the metric and V (t) and W(t) are parallel. This shows that the
function P
t
(v), P
t
(w)
c(t)
is constant in t, and thus equal to P
0
(v), P
0
(w)
c(0)
= v, w
c(0)
, proving our
assertion 2
We have the following proposition:
Proposition 6.10.8 (Existence of geodesics). Let M be a smooth manifold, and be any connection on it
(i.e. on the tangent bundle TM). Let x M and let v T
x
(M) be a tangent vector. Then there exists an
> 0 and a smooth curve:
c : [, ] M
such that:
(i): c is a geodesic.
(ii): c(0) = x
(iii): c(0) = v
Proof: Let x lie in the coordinate patch U, and let x
j
denote the coordinate functions with respect to this
chart. The conditions (i) and (ii) above lead to the system of 2n rst order ordinary dierential equations (for
Z
j
(t) and c(t):
dc
j
dt
(t) = Z
j
(t) (22)
dZ
k
dt
(t) +
k
ij
dc
i
dt
Z
j
= 0 (23)
from the local formula in the lemma 6.6.3. The initial conditions are c
j
(0) = x
j
, Z
j
(0) = v
j
. This can also be
recast as the system of n second order ODEs for c(t):
d
2
c
k
dt
2
(t) +
k
ij
dc
i
dt
dc
j
dt
= 0
with the initial conditions c(0) = x, c(0) = v. By the existence and uniqueness theorem for ODEs, there is a
unique solution for this system for t [, ] and some > 0, thus proving the proposition. 2
Example 6.10.9 (Euclidean Space). Since the metric is
ij
in the standard coordinates on R
n
, all Christoel
symbols
k
ij
= 0. Thus the geodesic equations say that the second derivative of c(t) is zero, so that c(t) = ta+b
for some a, b R
n
, i.e. a straight line.
Remark 6.10.10. The example of X = (0, 1) and the tangent vector
d
dt
at say a point x X shows that as
x is closer to the end points, the guaranteed by the previous proposition will shrink.
Denition 6.10.11 (Exponential map on a manifold). Let M be a smooth manifold with any smooth con-
nection . We denote the curve c(t) of the proposition 6.10.8 above by Exp
x
(tv). The notation suggests the
following exercise: The geodesic curve c(t) dened on [, ] with c(0) = x, c(0) = v is the same as the geodesic
curve c
1
(t) = c(Rt) dened on [
R
,

R
] satisfying c
1
(0) = x, c
1
(0) = Rv. So we see that Exp
x
(tv) = c(t) is the
same as c
1
(
t
R
) = Exp
x
((
t
R
)(Rv)) and the denition of Exp
x
makes sense.
In fact we have the following stronger statement, which follows from the fact the solution to the above
system of geodesic ODEs varies smoothly in the initial conditions, as stated in the proposition 5.4.2:
Proposition 6.10.12. Let M be a smooth manifold with x X a point on it. Let be a connection, with
the corresponding covariant derivative on TM. Then there exists an open neighbourhood U of the origin in
T
x
(X), and a neighbourhood V of x in M such that the map:
Exp
x
: U V
v Exp
x
(v)
is a smooth dieomorphism of U onto V . Thus one gets a natural chart (V, Exp
1
x
) around each point x in X.
These are known as exponential or geodesic coordinates.
100 VISHWAMBHAR PATI
Corollary 6.10.13. If V is the neighbourhood x in the proposition above, then each point of V can be joined
to x by a unique geodesic. There is a stronger theorem of J.H.C. Whitehead which asserts that for the Levi-
Civita connection on a Riemannian manifold, each point x X has a geodesically convex neighbourhood, i.e.
a neighbourhood V such that every pair of points in V can be joined by a unique minimal length (among
all curves joining these points in X) geodesic which lies entirely in V . For a proof, see the book Notes on
Dierential Geometry by Hicks, p. 134.
Remark 6.10.14 (Geodesics in the Levi-Civita connection). Suppose X is a Riemannian manifold, and is
the unique torsionless Levi-Civita connection on it as guaranteed by 6.9.1. Then there is a nice geometric
interpretation of the geodesics for this connection. Consider the Energy Functional of a smooth path c :
[0, 1] X which is dened as:
E(c) =
_
1
0
c(t), c(t) dt
Suppose c(0) = x X and c(1) = y X, and we perform a one parameter smooth variation of the path c with
xed end points x and y. That is, we have a smooth map:
: (, ) [0, 1] X
(s, t) (s, t)
(s, 0) = x (s, 1) = y s, (0, t) = c(t) t
Let us abbreviate the vectors D(s, t)(

s
) and D(s, t)(

t
) as
s
and
t
respectively. Then let us nd the
condition for c to be energy extremising. This means that:
s
(E((s, ))
|s=0
=
_
1
0
t
,
t
dt
|s=0
= 2
_
1
0
t
,
t
dt
|s=0
= 2
_
1
0
s
,
t
dt
|s=0
= 2
_
1
0
s
,
t
dt 2
_
1
0
s
,
t
t
dt
|s=0
where we use the torsionless condition to get
s
t
=
t
s
, and the compatibility of the metric to get the
last line above. By the fundamental theorem of calculus, the rst integral above is just
s
,
t
|s=0,t=1
s
,
t
|s=0,t=0
which is equal to zero since
s
vanishes at t = 0 and t = 1 because the end-points are xed. Setting the second
integral equal to zero is the condition:
_
1
0
s
,
c(t)
c(t)
_
dt
is zero. If E(c) is to be extremised at c, this must hold good for all variations , and in particular a variation
which is localised in an arbitrarily small neighbourhood of any point t [0, 1]. This implies that
c(t)
c(t) =
D
dt
(
dc
dt
) = 0, which is precisely the geodesic equation.
With a little more eort, one can show that geodesics extremise the energy functional not only over smooth
variations, but also piecewise smooth variations (see Milnors Morse Theory, 12). One could similarly want
to characterise the curves for which the length functional:
L(c) =
_
1
0
c(t), c(t)
1
2
dt
is extremal. We nd on solving this variational problem that the extremal curves for this functional are just
reparametrisations of geodesics, which is logical since the reparametrisation of a curve does not change its
length. On the other hand, for a geodesic c(t), we have:
d
dt
c(t), c(t) = 2
c(t)
, c(t)
_
= 0
by the geodesic equation. Thus a geodesic has constant speed at all times (where speed at t is the magnitude
of the velocity vector c(t), i.e. c(t), c(t)
1
2
. It is customary to scale the parameter by a constant so that this
constant speed is = 1, in which case it is parametrised by the arc length parameter:
s =
_
s
0
dt =
_
s
0
c(t), c(t)
1
2
dt
which is the length of the segment c([0, s]).
We note here that if G is a Lie group with a left invariant metric , , the corresponding map Exp dened
above may not tally with the exponential map exp dened earlier. In fact:
Proposition 6.10.15 (Exp and exp for a Lie group). Let g be a left invariant Riemannian metric on a Lie
group G, corresponding to a positive denite inner product , on the Lie algebra g. Then the exponential
map Exp
e
at the identity e G coincides with exp precisely when , satises:
Z, [Y, X] +[Y, Z], X = 0 X, Y, Z g
(For example, this will happen if g is bi-invariant (=Ad-invariant)).
Proof: We let

X,

Y ,

Z denote the left invariant vector elds corresponding to X, Y , Z respectively. Then we
have
d
dt
(exp tX) =

X etc. For the curve exp tX to be a geodesic, it is necessary and sucient that:
_
X,
Y
_
= 0
for all Y . That is:
X
_
X,
Y
_
X,
X
Y
_
=
_
X,
X
Y
_
= 0
since the rst term is 0 by the left invariance of , . Since the connection is torsionless the last expression
above is just
X,
Y
X
_
+
_
X, [
Y ,

X]
_
=
_
X, [
Y ,

X]
_
since the rst term is
1
2
Y
_
X,

X
_
= 0 by the left invariance of , . Thus, for all Z g, we have:
_
X +

Z, [[
X +

Z],
Y ]
_
= 0
which implies
_
Z, [
X,
Y ]
_
+
_
[
Z,
Y ],

X
_
= 0
which implies the required condition by removing the tildes since everything is left invariant. 2
Corollary 6.10.16. On SL(2, R), the maps Exp
e
and exp are distinct. This is because every non-degenerate
form satisfying the condition of the proposition above is a multiple of the Cartan-Killing form on sl(2, R),
which is not positive (or negative) denite.
Now we will show that geodesics are locally minimising for the distance d dened in 5.2.3. We will also
nally prove the fact (mentioned after 5.2.3) that the topology of M coincides with the metric space topology
w.r.t. the distance d. We need some denitions and lemmas.
Denition 6.10.17. Let f be a smooth function on a Riemannian manifold M, with Riemannian metric , .
We dene a vector eld grad f called the gradient of f by the formula:
grad f(x), Z = Z(f) = df(Z) for all Z T
x
M and all x M
As an exercise, the reader can verify that in a coordinate chart U with local coordinates x
i
, and local metric
coecients g
ij
, with inverse matrix g
kl
, we have the formula:
grad f =
k,l
g
kl
_
f
x
l
_

x
k
which shows that grad f is a smooth vector eld.
Lemma 6.10.18. Let f be a smooth function on an open subset U M, and x M be a point.
(i): For each smooth curve c : [, r] U we have:
f(c(t)) f(c()) =
_
t
grad f, c(s) ds
(ii): If c : (0, r] U is a trajectory of grad f, then
f(c(t)) = f(c()) +E
t
(c)
where E
t
(c) denotes the energy of c between and t ( =

_
t
| c(s)|
2
ds).
(iii): If |grad f| 1 at all points of U, then for any trajectory c : (0, r] U of grad f, we have:
f(c(t)) f(c()) = t
(iv): Let f be as in (iii) above. Let : [0, 1] M be a PS path, and ((0, 1]) U and (0) = x U.
Assume that f : U R is continuous on U. Then we have
L
t
0
() :=
_
1
0
| (s)| ds f((t)) f(x)
In particular, if c : (0, r] U is a trajectory of grad f with lim
t0
c(t) = x, then for all PS paths
satisfying (0) = x = c(0), (r) = c(r), the arc lengths of and c satisfy the inequality:
L
r
0
() r = L
r
0
(c)
Proof: For a smooth curve c, the denition of gradf yields the relation grad f, c(s) = c(s)(f) =
d(fc)
ds
, so
that by the fundamental theorem of Calculus, we have:
f(c(t)) f(c()) = =
_
t
d(f c)
ds
=
_
t
grad f, c(s) ds
This implies (i).
(ii) is clear from (i), because for a trajectory c of grad f, we have c(s) = grad f(c(s)).
For (iii), note that if |grad f| 1 at all points of U, then for any trajectory c : (0, r] U of grad f, we
have E
t
(c) = t , and we are done by (ii).

For (iv), we take the limits as 0 in all of the above. Then, if is any other PS path starting at x, (i)
will imply that
f((r)) f(x) =
_
r
0
c(s), (s) ds
_
r
0
| (s)| ds = L
r
0
()
by the Cauchy-Schwarz inequality. Thus if (r) = c(r) where c is a trajectory of grad f, we get (using (iii)
above) f((r)) f(x) = f(c(r)) f(x) = r = L
r
0
(c), and (iv) follows. 2
Proposition 6.10.19. Let M be a Riemannian manifold, and let x M. Assume that the exponential map
:= Exp
x
: V U is a dieomorphism, where V is an open neighbourhood of 0 T
x
M, and U is an open
neighbourhood x in M. Let B(0, r) V be an closed ball or radius r. Then:
(i): for each v T
x
(M) with |v| = 1, and each r, the path Exp
x
(sv) for 0 s , is the path of
minimal arc length joining x to Exp
x
(v).
(ii): For < r, the image Exp
x
(B(0, )) is precisely the open ball {y M : d(x, y) < }, where d is
the distance of 5.2.3. In particular, a fundamental system of neighbourhoods of x M is given by
{B
d
(0, )}
0<<r
, and the metric topology on M dened by d coincides with the topology of M (as
claimed in 5.2.3).
Proof: We have already remarked that the derivative DExp
x
(x) is the identity map, and so by the inverse
function theorem, := Exp
x
is a local dieomorphism on some neighbourhood V of 0 in T
x
M, and for some
r > 0, we have B(0, r) V .
Consider the smooth function f : U \ {x} R dened by f(y) =
_
_
1
(y)
_
_
, where |v|
2
, as usual, denotes
v, v
x
. For v with |v| = 1, we will denote the geodesic (tv) = Exp
x
(tv) by c
v
(t). We claim that:
grad f(c
v
()) = grad f((v)) = c
v
() (24)
for all v with |v| = 1, and all r.
For a vector Z T
v
(T
x
(M)) = T
x
(M), a path in T
x
(M) with initial velocity Z is the path (v +sZ). We
consider the 1-parameter variation of the geodesic (tv) given by (s, t) = (tv +sZ)). We have the natural
vector elds along the geodesic c
v
(t) = (tv) dened by:
s
(t) :=
(t, 0)(

s
) =
_
d
ds
_
|s=0
((tv +sZ))
t
(t) :=
(t, 0)(

t
) =
d
dt
((tv)) = c
v
(t)
Now we have:
grad f,
(Z)
cv()
:=
(Z)(f) = D(v)(Z)(f)
=
_
d
ds
_
s=0
(v +sZ, v +sZ
x
)
1
2
=
1
2
(|v|)
1
(2 v, Z
x
)
= v, Z
cv(0)
since |v| = 1.
We have to show that the last quantity above is the same as:
c
v
(),
(Z)
(v)
= c
v
(),
(Z)
cv()
Note that
s
(1) =
(Z), and
s
(0) = Z. Also
t
(0) = v, and
t
(1) = c(). Thus
c
v
(),
(Z)
cv()
v, Z
cv(0)
=
1
(
t
(1),
s
(1)
1
t
(0),
s
(0)
0
)
=
1
_
1
0
d
dt
(
t
(t),
s
(t)
t
) dt
=
1
__
1
0
(
t
t
,
s
t
+
t
,
t
t
) dt
_
= 0 +
1
_
1
0
t
,
s
t
dt
=
1
2
_
1
0
_
d
ds
_
s=0
t
,
t
t
dt
= 0
where we used the fact that
t
is the velocity eld of a geodesic, so has zero acceleration, and that
s
and
t
commute, in the third line, and that it has constant norm, in the last line. This proves the claim.
We now observe that the vector eld grad f on U \ {x} has the geodesics c
v
(t) for |v| = 1 and 0 < t r as
its trajectories. This is because of the equation (24). Thus by (ii) and (iii) of the lemma 6.10.18, we have:
f(c
v
()) f(c
v
()) = E
(c
v
) =
In particular, E
0
(c
v
) = = L
0
(c
v
), because f is continuous on U \ {x}, and we may take limits as 0.
This says that the energy of c
v
over any time interval [0, ] is the same as its arc-length.
For any PS curve : [0, r] U with (0) = x and () = c
v
(), (iv) of 6.10.18 will imply that L
0
()
r = L
0
(c
v
). This implies (i) of the proposition.
From (i) it follows that for < r that B(0, ) V , the set Exp
x
(B(0, )) {y M : d(y, x) < }. On
the other hand, if y is in the right hand set, say d(x, y) < , we have a PS path : [0, 1] M such that
(0) = x, (1) = y, and with L
1
0
() = < . If y , Exp
x
(B(0, )), then since (0) = x, the path would
have to intersect Exp
x
(S
), the image of the sphere of radius at some point (a) = b for a < 1. But then
d(b, x) = , so by (i) above, the arc length L
a
0
() > = L
1
0
(), a contradiction. Thus y Exp
x
(B(0, )).
This proves our assertion (ii), and the proof of the proposition is complete. 2.
We also have the following related:
Lemma 6.10.20. If U is as above, then for a smooth path : [a, b] U \ {x}, we have the inequality
d((b), x) d((a), x) L
b
a
()
with equality holding i is a reparametrisation of a radial geodesic joining (a) to (b).
Proof: With the function f constructed in the last proposition, we have for any y = Exp
x
(rv) and |v| = 1
that d(y, x) = r = f(y). Thus the inequality follows from (iii), (iv) of 6.10.18.
To see when equality holds, write (t) = Exp
x
(r(t)v(t)), where |v(t)| = 1 for all t. Call the function
Exp
x
(rv(t)) as G(r, t). Then (t) = G(r(t), t), and:
(t) =
G
r
r
(t) +
G
t
so that, because
G
t
is orthogonal to
G
r
, and
_
_
G
r
_
_
= 1, we have:
| (t)|
2
= |r
(t)|
2
+
_
_
_
_
G
t
_
_
_
_
2
|r
(t)|
2
Thus
| (t)| |r
(t)| r
(t)
and integrating, we have the inequality above, after noting that f(Exp
x
(r(t)v(t))) = r(t). If equality holds in
this, then we must have
G
t
0 for all t, that is v(t) is a constant vector v, and also r
(t) is positive, i.e. r(t)

is monotone. This yields (t) = Exp
x
(r(t)v), i.e. it is a reparametrised geodesic. 2
Exercise 6.10.21. Let x, y M, and let be a PS path joining x and y which is of minimal length among
all PS paths joining x and y. (That is, the arc length L() = d(x, y). Then is a geodesic. Such geodesics are
called minimal geodesics.
Remark 6.10.22. Suppose x, y are two points on a Riemannian manifold M. Suppose that x and y are joined
by a unique geodesic c. (For example if x, y lie in a geodesically convex neighbourhood). Let be an isometry
which leaves x and y xed. Since isometries preserve the metric, they preserve the Levi-Civita connection, and
hence preserve geodesics. So (c) is another geodesic which joins x to y, and by the uniqueness of c, must be
c. Thus c is left pointwise xed by . This often enables us to determine geodesics on a Riemannian manifold
without messing our hands with the non-linear system of geodesic ODEs, as we shall see below.
Example 6.10.23 (Geodesics on the Sphere). Let S
n
R
n+1
be given the Riemannian metric induced from
the Euclidean metric on R
n+1
. Then the orthogonal reection R
H
about any hyperplane H through the origin
(i.e. n-dimensional vector subspace) on R
n+1
is an isometry of S
n
(because it is an isometry of R
n+1
mapping
S
n
to itself!) So if x, y are two suciently near points (in fact, lying in the same open hemisphere), then there
will be a unique geodesic c joining these two points, and consequently by remark 6.10.22 above, will be left
pointwise xed by the reection R
H
where H is the hyperplane H containing x and y. Thus the geodesic c
will be contained in S
n
H = S
n1
H
. Now one can proceed with the hyperplane in S
n1
H
inductively, until one
gets c to lie on circle, which is the successive intersection of S
n
with a sequnce of such hyperplanes, i.e. the
intersection of S
n
with a 2-dimensional subspace of R
n+1
. Thus (the image of) c is the arc of a great circle
through x and y. In fact, since any segment of a geodesic is a geodesic, the geodesics on S
n
are precisely
segments of great circles on S
n
, parametrised by arc length. Note that for two points x, y S
n
which are not
antipodal, i.e. y ,= x, there are exactly two geodesics joining x to y, one of minimal length and the other of
maximal length. For two antipodal points, there is a whole family of geodesics parametrised by S
n1
joining
them. A geodesically convex neighbourhood of x S
n
is the open hemisphere centered at x (i.e. determined
by the hyperplane x
).
Example 6.10.24 (Geodesics in H
2
). We note that x x and y y is an isometry of H
2
(see example
5.8.6 for notation). This leaves the y-axis (call it L) pointwise xed, and so by the remark 6.10.22, all small
geodesic segments between any two points p, q L will consist of the segment of L between p and q. We can
actually determine the parametrisation by taking p =

1 = (0, 1) as a reference point. The length of the
segment c(t) = (0, 1 +t(y 1)) between p = (0, 1) and q = (0, y) is precisely:
L(c) = s =
_
1
0
(y 1)
y
, (y 1)
y
1
2
dt
=
_
1
0
y 1
1 +t(y 1)
dt
= log y
Thus, reparametrising by arc length s, we have c(s) = (0, e
s
). This is therefore a geodesic.
To determine other geodesics, we hit this curve by various isometries from G = SL(2, R). Under the
isometry:
g =
_
a b
c d
_
the point iy = (0, y) goes to the point:
w =
1
c
_
d
d
2
+c
2
y
2
+
icy
d
2
+c
2
y
2
+a
_
by an easy calculation using adbc = 1. Thus w
a
c
=
1
c
e
i
where = tan
1 cy
d
. This shows that the image of
L is a semicircular arc centred about (
a
c
, 0) of radius
1
c
. Clearly by choosing a and c suitably, we can capture
any semicircular arc centred about any desired point on the x-axis and of any desired radius, so all of these
(suitably parametrised) are geodesics. Finally, since translation by a real number b is an isometry, (in fact
corresponding to the element:
_
1 b
0 1
_
of SL(2, R)), we get all vertical lines as geodesics by translating the geodesic L around. These may be treated
as semicircular arcs meeting the x-axis orthogonally (that is, having their centre on the x-axis), and of
innite radius. Thus all semicircular arcs of nite or innite radius, meeting the x-axis orthogonally (suitably
parametrised) are geodesics. It turns out that these are all the geodesics.
Example 6.10.25 (Geodesic coordinates in the Disc Model D
2
). We recall the example 5.8.8 above. Our
metric on D
2
is (after dropping the factor of 4 for convenience):
g(z) =
dzdz
(1 |z|
2
)
2
Since the imaginary axis (suitably parametrised) constituted a geodesic in H
2
, its image under the Cayley
transform map , i.e. the x-axis will constitute a geodesic in D
2
. Since rotations in SU(1, 1), i.e. the matrices:
_
e
i
0
0 e
i
_
are isometries of D
2
which are rotations, they carry the ray
{(t, 0) : t R
+
}
to any given ray through the origin in D
2
. Thus all rays starting at the origin in D
2
, suitably parametrised, are
geodesics. To see what the parametrisation is, we compute the length of the radial segment (t, 0). Its velocity
is (1, 0) =
x
, and thus its speed at t is (1 t
2
)
1
. Thus its length is:
s(t) =
_
t
0
1
1 t
2
dt =
1
2
log
_
1 +t
1 t
_
Thus t = tanh s. Thus Exp(sv) = (tanh s)v for any unit vector v T
0
(D
2
). This gives us a geodesic polar
coordinate chart on D
2
. Since the metric at a point te
i
is:
dx
2
+dy
2
(1 t
2
)
2
=
dt
2
+t
2
d
2
(1 t
2
)
2
and ds = (1 t
2
)
1
dt, and
t
(1t
2
)
=
1
2
sinh s cosh s =
1
2
sinh (2s), our metric reads
ds
2
+
1
4
(sinh 2s)
2
d
2
in these geodesic polar coordinates.
Example 6.10.26 (Geodesics in H
3
). By using the isometries which reect through the 2-dimensional hyper-
planes H that pass through the w-axis (see example 5.8.10), and using the remark 6.10.22, it follows that all
geodesics in H
3
will lie on such hyperplanes. The induced Riemannian metric on any such hyperplane will be
the 2-dimensional metric of the hyperbolic plane H
2
. Thus it will be a geodesic of H
2
. Thus by the previous
example, the geodesics of H
3
are precisely semicircular arcs orthogonal to (centred at points of) the w = 0
hyperplane.
6.11. Completeness and the Hopf-Rinow Theorem. In this subsection, we shall prove that M is a
complete metric space with respect to d if all geodesics are innitely extendable (i.e. M is geodesically
complete).
Denition 6.11.1. We say that a Riemannian manifold M is geodesically complete if for each x M, the
geodesic Exp
x
(tX) is dened for all t R, and all X T
x
(M).
Proposition 6.11.2 (Hopf-Rinow). If M is connected and geodesically complete, then:
(i): Each pair of points in M can be joined by a minimal length geodesic.
(ii): The closure of every bounded set (with respect to the metric d) is compact. In particular, M is a
complete metric space with respect to d.
Proof: Let x, y M, which is assumed to be geodesically complete. We rst prove (i).
Around x, there is a closed ball B(x, ) such that the map Exp
x
: B(0, ) B(x, ) is a dieomorphism,
by the proposition 6.10.19 above. If y B(x, ), we are done by 6.10.19. If not, we proceed as follows.
Let d(x, y) = r with r > . We denote, as before, the sphere:
S
= Exp
x
({v T
x
M : |v| = })
Because S
is compact, and y , B(x, ), there is a point z S
of minimum distance from y. Write z =

Exp
x
(v) where |v| = 1. Since M is geodesically complete, Exp
x
(rv) is dened, and is a point in M. We
claim that Exp
x
(rv) = y. Since the arc length of the curve c
v
(t) = Exp
x
(tv) from 0 to r (= L
r
0
(c
v
)) is
r = d(x, y), it will follow that this is the required minimal geodesic.
We rst make the:
Claim:
d(c
v
(t), y) = r t for all 0 t r
Then setting t = r would prove that y = c
v
(r). First we note that this claim is true for t = . For, every PS
curve from x to y must meet S
, so:
r = d(x, y) = min{L
1
0
(c) : c a PS curve with c(0) = x, c(1) = y}
= min{L
a
0
(c) +L
1
a
(c) : c(a) S
}
+d(z, y)
where z = c
v
() = Exp
x
(v) is the point we chose above. This implies that d(c
v
(), y) r . On the other
hand, r = d(x, y) d(x, c
v
()) + d(c
v
(), y) = + d(c
v
(), y) by the triangle inequality. This shows that our
claim is valid for t = , and indeed for all t , by varying to smaller .
Let s be the supremum of all the t [0, r] for which the claim above is valid. By continuity, it will follow
that the claim is also valid for t = s. If we show that s = r, then we are done. Suppose that s < r. Then
c
v
(s) ,= y. Again choose a small enough ball B(c
v
(s), ) around c
v
(s) so that y lies outside it. Let S
(c
v
(s))
denote the image of {w : |w| = } under Exp
cv(s)
. Let p S
(c
v
(s)) be the point nearest to y in that sphere.
We claim that p = Exp
x
((s + )v). For we have by writing p = Exp
cv(s)
(w) etc. and the identical argument
to the one in the last para above that:
d(c
v
(s), y) d(p, y) +
The triangle inequality implies d(c
v
(s), y) d(c
v
(s), p) +d(p, y) = +d(p, y). Thus d(p, y) = d(c
v
(s), y) .
From this, and the equality d(c
v
(s), y) = r s, it follows that d(p, y) = r s . Thus d(x, p) d(x, y)
d(y, p) = r (r s ) = s + .
Now choosing the PS curve dened by c
v
(t) for t s, and the curve Exp
cv(s)
(tw) for t , we have a
curve of arc length s + joining x to p. Thus two facts result:
(i): d(x, p) = s + .
(ii): is a length-minimising curve from x to p.
From the exercise 6.10.21, it follows that is a smooth geodesic. Since it coincides with Exp
x
(tv) for
0 t s, it coincides all the way upto s +, that is p = Exp
x
((s +)v). That is p = c
v
(s +). We had already
seen above that d(p, y) = r s , so this means d(c
v
(s + ), y) = r s , contradicting the maximality of
s. Hence s = r, and (i) is proved.
To see (ii), it is enough to show that the closed ball B(x, r) is compact (for some xed x M) for all r > 0.
From (i), it follows that the ball B(0, r) is precisely the image Exp
x
(B(0, r)) for all r > 0. But the last set
is compact from the continuity of the map Exp
x
, and by Heine-Borel for T
x
(M) R
n
. Since every Cauchy
sequence is a bounded sequence, and since all compact subsets of a metric space are sequentially compact, it
follows that every Cauchy sequence is convergent, and M is complete as a metric space. This proves (ii), and
the proposition. 2
Proposition 6.11.3. Every compact connected Riemannian manifold M is geodesically complete.
Proof: For each point x M, we use the remark 6.10.13 to get a neighbourhood U(x) of x which is geodesically
convex. Since M is compact, there is a Lebesgue number for this covering {U(x)}
xM
. That is, every closed
ball of radius in the manifold is contained in some U(x). Now if y M is any point, it follows that every
point z B(y, ) can be joined to y by a geodesic of minimal length. This means that Exp
y
: B(0, ) B(y, )
is a dieomorphism for every y M.
Let x M, and v T
x
(M) with |v| = 1. We will show that Exp
x
(tv) M for all 0 t m, for every
positive integer m. This will clearly imply geodesic completeness.
Let x
1
:= Exp
x
(v). Since d(x
1
, x) = , we have x B(x
1
, ). Thus x = Exp
x1
(v
1
) for some v
1
T
x1
(M)
with |v
1
| = 1. We claim that Exp
x
(tv) = Exp
x1
((t )v
1
) for all 0 t 2. The segments Exp
x
(tv) for
0 t and Exp
x1
((t )v
1
) for 0 t are both minimal length geodesics joining x and x
1
, and hence
coincide for t [0, ]. Thus, by uniqueness, it follows that Exp
x
(tv) = Exp
x1
((t )v
1
) for all 0 t 2.
Now repeat the argument by replacing x above with x
1
, and x
1
by Exp
x1
(v
1
) = Exp
x
(2v), and we are
through by induction. 2
Corollary 6.11.4. If M is a compact connected Riemannian manifold, then for each point x M, the map
Exp
x
: T
x
(M) M is surjective.
Proof: By 6.11.3 above, M is geodesically complete. By 6.11.2, every point y can be joined to x by a minimal
geodesic, which implies that y = Exp
x
(rv) for some unit vector v T
x
(M) and r = d(x, y). 2
Corollary 6.11.5. Let G be a Lie group with a bi-invariant metric (for example if G is compact). Then the
exponential map exp : g G (dening 1-parameter subgroups) is surjective.
Proof: For G, the maps Exp and exp coincide, as pointed out in the proposition 6.10.15, the geodesics Exp
e
(tv)
through 1 G coincide with the one parameter subgroups exp(tv) for v g. Since 1-parameter subroups are
dened for all t, it follows that G is geodesically complete. We are again through by the Hopf-Rinow theorem
6.11.2. 2
7. Curvature
7.1. Curvature form on a principal bundle. Let : P M be a principal G-bundle, with a connection
, and associated g-valued connection 1-form . We also have the trivialisations
i
:
1
(U
i
) U
i
G, and
the naturally arising smooth sections
i
: U
i

1
(U
i
) which are dened by
i
(
i
(x)) = (x, 1) for x U
i
.
They are related by the transition formula g
ij
(x)
j
(x) =
i
(x) for x U
i
U
j
. (see the subsections 6.3 and
6.4). The local connection 1-forms
i

1
(U
i
) g were dened as
i
.
Notation : 7.1.1. Let , be smooth g-valued 1-forms on an open set U P for some smooth manifold P.
Then we may write =
i

i
X
i
and =
j

j
X
j
, where {X
i
} is a basis for g, and
i
,
j
are 1-forms on U.
We dene a g-valued 2- form on U by:
[, ] =
i,j
(
i

j
)[X
i
, X
j
]
Note that by this denition [, ] is not zero, but equal to

i<j

i

j
[X
i
, X
j
]. In fact, the reader can
easily check that:
[, ] = [, ] (25)
We also note that [, ] is nothing but the image of under the composite map:
(
1
(U) g) (
1
(U) g)
1
(U)
1
(U) g g
[ . ]

2
(U) g
Hence it does not depend on the basis {X
i
} chosen for g.
For example, if G = GL(n, R) (or for that matter any group of matrices), and we use the basis of elemen-
tary matrices {E
ij
}
1i,jn
for g = gl(n), then writing =

i,j

ij
E
ij
, =

k,l

kl
E
kl
, and noting that
[E
ij
, E
kl
] =
jk
E
il
li
E
kj
, we see (using the repeated index Einstein summation convention) that
[, ] = (
ij

kl
(
jk
E
il
li
E
kj
))
= (
ij

jl
) E
il
(
ki

ij
) E
kj
=
il
E
il
where
il
:=
j
(
ij

jl
+
ij

jl
)
If = ,
il
is the il-th matrix entry of the matrix product 2 , which is matrix multiplication but with
usual product of scalar entries replaced by the wedge products. Some authors denote our [, ] as .
More generally, one denes [, ] as a g-valued (p +q)-form, for , g-valued p and q-forms respectively on
U, as follows:
[, ] =
i,j
(
i

j
)[X
i
, X
j
]
= (1)
pq+1
j,i
(
j

i
)[X
j
, X
i
]
= (1)
pq+1
[, ] (26)
Also, if =
i

i
X
i

p
(U) g is a g valued p-form, then d will be the g-valued (p +1)-form

i
d
i
X
i
.
That is, it is the image of under the map
d Id :
p
(U) g
p+1
(U) g
applied to .
Exercise 7.1.2. Show that for a g-valued 1-form ,
[, [, ]] = 0
(Hint: use the Jacobi identity of g.)
Lemma 7.1.3. With the above denitions of d and [ , ], we have:
d[, ] = [d, ] + (1)
p
[, d]
for
p
(U) g,
q
(U) g.
Proof: By the R bilinearity of both sides, we may take = X and = Y where , are R-valued
1-forms and X, Y g. Then [, ] = [X, Y ], so that d[, ] = d [X, Y ] + (1)
p
d[X, Y ] =
[d, ] + (1)
p
[, d], and we are done. 2
Lemma 7.1.4. If is a 1-form on an open subset U of a smooth manifold M, and X, Y smooth vector elds
on U, then
d(X, Y ) = X((Y )) Y ((X)) ([X, Y ])
Proof: We rst leave it as an exercise for the reader to verify that the right hand side satises linearity with
respect to multiplication of functions, viz:
fX((gY )) gY ((fX)) ([fX, gY ]) = fg (X((Y )) Y ((X)) ([X, Y ]))
for smooth functions f, g. Now we need verify the formula of the lemma on any any coordinate chart U, and
by the observation above, take X =
i
=

xi
and Y =
j
=

xj
. Then [X, Y ] 0. Say i < j, without loss of
generality, because both sides are skew-symmetric in X and Y .
Write =
j

j
dx
j
, so that X((Y )) =
i
j
and Y ((X)) =
j
i
, so that the right side of our formula is
j

j
i
. On the other hand, by the denition of exterior derivative:
d =
j
d
j
dx
j
=
i<j
(
i
j

j
i
)dx
i
dx
j
so that
d(X, Y ) = d(
i
,
j
) =
i
j

j
i
thus proving the lemma. The same formula is checked to be true for g-valued 1-forms. 2
Corollary 7.1.5. Let
R
be the right-invariant Maurer- Cartan form on a Lie group G. That is,
R
(

X) = X
for X g and

X the right- invariant vector eld on G generated by X. (see 6.3.1). Then
d
R
=
1
2
[
R
,
R
]
Proof: Let {X
i
} be a basis for g, so that
R
=
j
e
j
X
j
for e
j
the right-invariant R valued 1-forms satisfying
e
i
(

X
j
) =
ij
. Let i < j. Then, by the lemma 7.1.4 above:
d
R
(

X
i
,

X
j
) =

X
i
(X
j
)

X
j
(X
i
)
R
([

X
i
,

X
j
])
= 0 0
R
([

X
i
,

X
j
])
Let : G G be the inversion map, i.e. (g) = g
1
. Then L
g
(h) = h
1
g
1
= R
g
1 (h).
Thus we have
L
g
= R
g
1
. Applying both sides to X g, and noting that
(e) = Id, we
get

X(g) =
X(g
1
) =
X((g)). That is,
X) =
X . From this it follows that [

X
i
,

X
j
] =
[
X
i
),
X
j
)] =
X
i
,

X
j
] =
[X
i
, X
j
] =
[X
i
, X
j
].
Thus we have:
d
R
(

X
i
,

X
j
) =
R
(
[X
i
, X
j
]) = [X
i
, X
j
]
On the other hand
[
R
,
R
] =
k,l
( e
k
e
l
)[X
k
, X
l
]
from which it follows that
1
2
[
R
,
R
](

X
i
,

X
j
) = [X
i
, X
j
]. Since the vector elds

X
i
constitute a C
(G) basis
for all the vector elds on G, we are done.
We note that an analogous argument with

X
i
will imply a similar formula for
L
with a changed sign , viz.
d
L
+
1
2
[
L
,
L
] = 0. 2
Lemma 7.1.6. Let
0
(G)g, where G is a Lie group and g its Lie algebra. Let Ad g denote the smooth
g-valued function whose value at g G is Ad g((g)). Then:
d(Ad g ) = Ad g(d) + [
R
, Ad g()]
where
R
is the right-invariant Maurer-Cartan form of G.
Proof: By the R linearity of both sides, we may assume that = fX, where f
0
(G) is asmooth function,
and X g. Then Ad g() = fAd g(X). Let Z g, and

Z the corresponding right-invariant vector eld on Z.
Then (exp(tZ).g) will be a smooth curve passing through g, and with velocity R
g
(Z) =

Z(g) at t = 0. Then
d(Ad g)(

Z(g)) =
d
dt |t=0
(f(exp(tZ)g)(Ad exp(tZ)g)X)
=

Z(g)(f)(Ad g)X +f(g)
d
dt |t=0
Ad exp(tZ)(Ad g)X
= df(

Z(g))(Ad g)X + [Z, f(g)(Ad g)X]
=
_
Ad g(d) + [
R
, Ad g ]
_
(

Z(g))
which proves our formula. 2
Corollary 7.1.7. Let be a g-valued 1-form on a manifold P, and let g : P G be a smooth function. If
X
i
are as above, and =
i

i
X
i
, where
i

1
(P), we denote by Ad g the g-valued 1- form whose value
at x P is

i

i
(x)Ad g(x) X
i
. Then:
d(Ad g ) = Ad g d + [g
R
, Ad g ]
Proof: Again, as before, we may assume =
i
X
i
, where
i
is a R-valued 1-form on P. Then, if we
let denote the g-valued constant function X
i
on G, we have g
(Ad g ) = Ad(g(x))X
i
, so that Ad g =
i
g
(Ad g ). Then d = 0, and d(Ad g ) = [

R
, Ad g ]. Thus:
d(Ad g ) = d
i
g
(Ad g )
i
d(g
(Ad g )
= Ad g d
i
g
(d(Ad g ))
= Ad g d
i
g
[
R
, Ad g ]
= Ad g d
i
j
g
e
j
[X
j
, Ad gX
i
]
= Ad g d +
j
(
i
g
e
j
)[Ad gX
i
, X
j
]
= Ad g d + [Ad g , g
R
]
This proves the lemma. 2
Let be a connection on the principal G- bundle with connection form . Let : P P a gauge
transformation, and let
denote the connection form of the gauge-transformed connection
. We recall
from 6.4.12 and 6.4.15 that:
= (Ad g) +g
R
Proposition 7.1.8 (Curvature form). In the setting above, we have the identity:
d
+
1
2
[
] = Ad g
_
d +
1
2
[, ]
_
(27)
The g-valued 2-form := d +
1
2
[, ] is called the curvature of the connection .
Proof: We have the formula above:
= (Ad g) +g
R
For brevity denote
1
= (Ad g). Then by the lemmas 7.1.7 and 7.1.5 we have:
d
= d
1
+d(g
R
)
= Ad g d + [g
R
,
1
] +g
d
R
= Ad g d + [g
R
,
1
] +
1
2
[g
R
, g
R
]
On the other hand we have:
1
2
[
] =
1
2
[
1
,
1
] +
1
2
[
1
, g
R
] +
1
2
[g
R
,
1
] +
1
2
[g
R
, g
R
]
=
1
2
[Ad g , Ad g ] + [g
R
,
1
] +
1
2
[g
R
, g
R
]
=
1
2
Ad g[, ] + [g
R
,
1
] +
1
2
[g
R
, g
R
]
Thus we get:
d
+
1
2
[
] = Ad g
_
d +
1
2
[, ]
_
This proves the proposition. 2
Corollary 7.1.9. Let : P M be as above, and let
i
: U
i

1
(U
i
) be the local sections dened above
on the trivialising coordinate opens U
i
. Denote the pullbacks
i
()
2
(U
i
) g as
i
. Then:
i
= Ad g
ij
j
That is,
i
patch up to give a global smooth section of
2
(T
M) P
Ad
g.
Proof: As in the proof of proposition 6.4.18, view the transition functions g
ij
as dening a gauge transformation
of the principal bundle P
|UiUj
, and use the proposition 7.1.8 above. 2
Remark 7.1.10 (Tensoriality of ). We recall here that if P M is a principal G-bundle, with transition
functions g
ij
: U
i
U
j
G coming from the local trivilisations, then for any associated vector bundle, the
transition functions would be (g
ij
. (See the example 6.2.2). In particular the vector bundle E = P
Ad
g
would have transition functions Ad (g
ij
) GL(g). To give a global section of this bundle, we would have to
give a collection of smooth functions s
i
: U
i
g which satisfy s
i
= Ad (g
ij
)s
j
on U
i
U
j
.
If we want to give a global section s of (T
(M)) E, then for every smooth vector eld X of M, we would

have s(X) as a smooth section of E (by pointwise evaluation), so that we would need a collection of maps
s
i
(X) : U
i
g which obeyed (i) s
i
(X) is linear in X, and its value at x U
i
depends only on X(x), and (ii)
s
i
(X) = Ad (g
ij
)s
j
(X). Similarly for sections of
_
p
(T
M) E.
The transformation formula for
i
in 6.4.18 precludes them from patching up to give a global section of
T
M E, or for that matter, any bundle on M. The term g
ij
R
show that derivatives of transition functions
occur, and the values of
i
,
j
at x U
i
U
j
are not related by an automorphism depending only on the point
x.
By contrast, the forms
i
dened above, by virtue of 7.1.9, patch up to give a global smooth section of
_
2
(T
M) E.
Proposition 7.1.11 (Bianchi Identity). The curvature form obeys the formula:
d [, ] = 0
Proof: Since = d +
1
2
[, ], we get on taking d, and using the lemma 7.1.4 above:
d =
1
2
[d, ]
1
2
[, d]
= [,
1
2
[, ]]
1
2
[
1
2
[, ], ]
= [, ]
where we have used the exercise 7.1.2 to get [, [, ]] = 0, and the skew commutativity [, ] = (1)
2.1+1
[, ] =
[, ] from 7.1.1. 2
7.2. Curvature and covariant dierentiation. Let p : E = P
V M be an associated smooth vector

bundle to the principal bundle : P M with a connection . In 6.6.2, 6.6.5 we introduced the covariant
dierentiation operator:
: (U, E)
1
(U) (U, E)
coming from the connection. If U is a coordinate chart, and : U
1
(U) is a smooth local section, we had
the natural associated sections w
i
(z) = [(z), e
i
] of E
|U
:= p
1
(U) (see the discussion immediately preceding
proposition 6.6.3).
If we denote the frame (w
1
, .., w
k
) of E
|U
by w, then the proposition 6.6.3 related the local connection form
by the formula:
w
i
=
j
( (
))
ji
w
j
Notation : 7.2.1. Since the g-valued curvature form on P is dened as:
= d +
1
2
[, ]
we will have
= d
+
1
2
[
]. Thus, applying , we have:

(
) = d(
) +
1
2
[
]
For the sake of convenience, call this gl(k, R) valued 1-form
on U as , and the gl(k, R) 2-form
as . We also noted at the end of notation 26, that for a matrix valued 1-form , [, ] = 2 , where the
right side is the matrix wedge product.
Then we have the two neater looking equations:
X
w
i
=
j
(X)
ji
w
j
(28)
= d + (29)
The second equation above means that (X, Y ) = d(X, Y ) +(X)(Y ) (Y )(X), where the products
on the right hand side denote matrix products.
Proposition 7.2.2. With the notations of 7.2.1 above, we have:
j
(X, Y )
ji
w
j
=
_
Y

Y
X

[X,Y ]
_
w
i
for smooth vector elds X, Y on U.
Proof: By the formula (29) above, and using the repeated summation convention,
Y
w
i
= ((Y ))
ji
w
j
so that, using Leibniz formula for covariant dierentiation,
Y
w
i
= X((Y )
ji
)w
j
+ (Y )
ji
X
w
j
= X((Y ))
ji
w
j
+ (Y )
ji
(X)
lj
w
l
= [(X((Y )) + (X)(Y )]
ji
w
j
Thus we will have:
(
X
Y
w
Y
X
)w
i
= [(X((Y )) Y ((X)) + [((X))((Y )) (Y )(X)]]
ji
w
j
= [(d(X, Y ) + ([X, Y ]) + [ ](X, Y )]
ji
w
j
= (X, Y ))
ji
w
j
+
[X,Y ]
w
i
We recall the remark made at the end of denition 6.6.1 that parallel transport P
c
t
depends on the curve c.
We will presently see how curvature measures this dependency.
Proposition 7.2.3. Let p : E M be a smooth vector bundle of rank k, with its parallel transport P coming
from a connection (on the underlying prinicpal bundle). Let U be a coordinate patch, and let X =
x1
and
Y =
X2
be two coordinate vector elds, with (x
1
, x
2
, .., x
n
)) being some coordinates on U. At any point
x = (a
1
, .., a
n
), we denote the trajectory (a
1
+ t, a
2
, .., a
n
) of X by a(t), and the corresponding trajectory
(a
1
, a
2
+s, .., a
n
) of Y , by b(s). Then for any smooth section of E on U:
(X, Y ) = lim
s,t0
P
a
t
P
b
s
P
b
s
P
a
t
st
Proof: The statement means that we are joining the point x = (a
1
, .., a
n
) to x+te
1
+se
2
= (a
1
+t, a
2
+s, .., a
n
)
by two paths: one being the path a from x to x+te
1
followed by b from x+te
1
to x+te
1
+se
2
, and the other
path is the one with the roles of a and b reversed.
We recall the denition of covariant dierentiation, which implies that:
P
b
s
(x +te
1
+se
2
) = (x +te
1
) +s
Y
(x +te
1
) +o
2
(x +te
1
; s)
= (x +te
1
) +s
Y
(x +te
1
) +o
2
(x; s) +o(st)
where the symbol o() as usual, denotes a quantity which goes to zero faster than (i.e. lim
0

1
o() 0).
Now one iterates this and writes a similar Taylor formula for P
a
t
P
b
s
. Then take the dierence with
P
b
s
P
a
t
, and take limits on dividing by st to obtain (
X
Y

Y
X
) = (X, Y ), since [X, Y ] = 0. We
leave the details as an exercise. 2
Denition 7.2.4. A connection for which 0 is called a at connection.
7.3. Some curvature computations.
Example 7.3.1 (Curvature of Cartan connection on a Lie group). We recall from the example 6.4.25 that the
connection form for the canonical Cartan connection is just =
R
, where
2
: P = G G G is the
second projection, and
R
is the right-invariant Maurer-Cartan 1-form of G. Thus = d +
1
2
[, ] =
2
_
d
R
1
2
[
R
,
R
]
_
= 0 by the lemma 7.1.5 above. Hence the canonical Cartan connection has no curva-
ture, i.e. is a at connection.
The above showed that the 2-form on P is 0. Alternatively, one could use the formula
X
Y = [
Y ,

X],
and the proposition 7.2.2 above, together with the Jacobi identity to show that the curvature form
is 0
on M = G.
Denition 7.3.2 (Riemannian curvature). Let be Levi-Civita connection on a Riemannian manifold. The
curvature form on M, taking values in gl(n, R) satises:
(X, Y )
ji
w
j
= (
X
Y

Y
X

[X,Y ]
)w
i
as note in the proposition 7.2.2, where w
i
is the natural frame in a coordinate chart coming from , which in
this case is the coordinate frame
i
. Thus, if Z =
i
Z
i
i
is a smooth vector eld, we get the equation:
(
X
Y

Y
X

[X,Y ]
)Z =
j
(
ji
(X, Y )Z
i
)
j
We denote, by standard convention, the vector eld on the right hand side as R(X, Y )Z. That is R(X, Y )
i
=
R(X, Y )
i
=
ji
j
. By the remark 7.1.10, it follows that R denes a global smooth section of
2
(T
M) F(TM))
Ad
gl(n, R) =
2
(T
M) hom(TM, TM)
(Convince yourself of the bundle identity F(TM)
Ad
gl(n, R) = hom(TM, TM)) R is called the Riemann
curvature tensor.
In other words, the Riemann curvature tensor is a smooth section R( , ) of
_
2
T
M hom(TM, TM), such

that for tangent vectors X, Y T
x
(M) and x U
i
, the matrix of the linear transformation R(X, Y ) : T
x
M
T
x
M with respect to the frame w
j
of TM
|Ui
is
i
(X, Y ).
Remark 7.3.3. Many authors (e.g. Milnor, Hicks) directly dene R(X, Y )Z = (
X
Y

Y
[X,Y ]
)Z,
for vector elds X, Y, Z. Then, to show that it is actually a tensor, one needs to verify that R(fX, gY )(hZ) =
fgh(R(X, Y )Z) for smooth functions X, Y, Z. This is is a routine verication using the Leibniz formula for ,
denition of commutator, etc. We dont need this verication by 7.1.10 and the proposition 7.2.2 above.
Proposition 7.3.4 (Properties of Riemannian curvature). Let X, Y, Z, W be tangent vectors in T
x
M, where
M is a Riemannian manifold with Riemannian metric , . Then:
(i): R(X, Y )Z = R(Y, X)Z
(ii): R(X, Y )Z +R(Y, Z)X +R(Z, X)Y = 0
(iii): R(X, Y )Z, W = R(X, Y )W, Z
(iv): R(X, Y )Z, W = R(Z, W)X, Y
Proof: (i) is clear from the formula R(X, Y )Z =
j
(
ij
(X, Y )Z
j
)w
i
for Z =
j
Z
j
w
j
, because is a 2-form.
(ii) is a routine verication using the fact that
X
Y
Y
X = [X, Y ] and the Jacobi identity.
For (iii), we have, for vector elds X, Y, Z, W:
Y
Z, W = X
Y
Z, W
Y
Z,
X
W
= X(Y Z, W Z,
Y
W)
Y
Z,
X
W
= XY Z, W
X
Z,
Y
W Z,
X
Y
W
Y
Z,
X
W
Interchanging X and Y and subtracting, we have:
Y
Z, W
Y
X
Z, W = [X, Y ] Z, W
X
Y
W, Z +
Y
X
W, Z
Now using that [X, Y ] Z, W =
[X,Y ]
Z, W
_
+
Z,
[X,Y ]
W
_
, we have (iii).
(iv) is also straightforward, and left as an exercise. 2
Example 7.3.5 (Riemannian Curvature of Lie Groups). Suppose G is a Lie-group with a bi-invariant metric
, . From 6.9.3 we have the formula:
Y =
1
2
[
X,
Y ]
Thus
R(
X,
Y )
Z =
1
4
_
[
X, [
Y ,

Z]] [
Y , [
X,

Z]]
_
1
2
[[
X,
Y ],

Z]
=
1
4
[
Z, [
X,
Y ] +
1
2
[
Z, [
X,
Y ]]
=
1
4
[
Z, [
X,
Y ]] (30)
Thus the Riemannian curvature of such a Lie group is a left-invariant tensor, as it should be, since the
metric is left invariant. It also shows how curvature measures the degree of non-abelian-ness of G. That is,
if all triple commutators in G vanish (as happens in e.g. abelian groups, or more generally 2-step nilpotent
groups), then the Levi-Civita connection for a bi-invariant metric on G is a at connection. Another useful
formula is:
R(X, Y )Z, W =
1
4
_
[
Z, [
X,
Y ]], W
_
=
1
4
_
[
X,
Y ], [
Z,
W]
_
(31)
Exercise 7.3.6. Using 6.9.3, compute the Riemann curvature tensor of a Lie-Group G with respect to a left-
invariant Riemannian metric on G.
There is an explicit formula relating the curvature tensor R with the Christoel symbols of (17).
Proposition 7.3.7 (Curvature in local coordinates). Let U be a coordinate patch, and
i
for i = 1, .., n be
the resulting coordinate vector elds on U. Note that they all commute on U. Writing (with repeated index
summation convention):
R(
i
,
j
)
k
= R
h
ijk
h
Then:
R
h
ijk
=
i
h
jk
h
ik
+
h
il
l
jk
h
jl
l
ik
where
i
jk
are the Christoel symbols of the connection.
Proof: We have the relation (17), by which:
k
=
l
jk
l
so that
j
=
i
(
l
jk
)
l
+
l
jk
l
= (
i
h
jk
+
l
jk
h
il
)
h
Interchanging the roles of i and j, subtracting, and noting that [
i
,
j
] = 0, we have the result. 2
Denition 7.3.8 (Sectional curvature). Let X, Y T
x
(X) be two tangent vectors. We dene the sectional
curvature in the plane of X and Y as the quantity:
K(X, Y )(x) =
R(X, Y )X, Y
X, X Y, Y X, Y
2
=
R(X, Y )Y, X
X, X Y, Y X, Y
2
(the denominator represents the area of the parallelogram spanned by X and Y in T
x
(X)). Note that:
R(X + Y, X + Y )X + Y, X + Y = ( )
2
R(X, Y )X, Y
by the skew-symmetry properties (i) and (iii) of proposition 7.3.4. The denominator
X, X Y, Y X, Y
2
also changes by the same factor ( )
2
on changing X to X + Y , and Y to
X + Y . Hence the quantity K(X, Y ) depends only on the plane spanned by X and Y .
Remark 7.3.9. Note that in M is 1-dimensional, its Riemann tensor is zero because R(X, X)X = R(X, X)X
by 7.3.4, so it vansihes identically. For a surface, note that by 7.3.4, R(X, X) = R(Y, Y ) = 0. So only R(X, Y )
can be non-zero. Again, by (iii) of 7.3.4, R(X, Y )X, X = 0, and similarly R(X, Y )Y, Y = 0. Thus the only
non-zero quantity is possibly R(X, Y )X, Y . That is the sectional curvature captures the Riemann curvature
tensor, and is a scalar.
Example 7.3.10. If G is a Lie group with a bi-invariant metric, then the sectional curvature (with respect to
the corresponding Levi-Civita connection is given by:
K(
X,
Y ) =
1
4
[X, Y ], [X, Y ]
for X, Y G orthonormal vectors, from 31 above. As a corollary, the Lie algebra g is abelian i all the sectional
curvatures of G are zero. In particular, if the group is connected, then since G is generated by a neighbourhood
of e, and since Ad exp(tX) = exp(tad X), it will follow that:
Corollary 7.3.11. A connected Lie group with bi-invariant metric is abelian i all its sectional curvatures are
zero.
Example 7.3.12 (Sectional curvatures of S
n
). We rst recall the parametrisation:
h : (0, ) (0, 2) S
2
(, ) (sin cos , sin sin , cos )
of S
2
, and note that the Riemannian metric on this coordinate patch U := h((0, ) (0, 2)) is given by
d
2
+ sin
2
d
2
, from (i) of 5.1.6. Thus e
1
=
and e
2
= (sin )
1
constitute an orthonormal frame of TS

2
on U. This means a section : U
1
(U) of the (oriented) orthonormal frame bundle of TS
2
, which we
know is : SO(3) S
2
by the exercise 6.1.7. Since g is the space of all 2 2 skew-symmetric matrices, we
see that local connection form on U will be a smooth map:
=
_
0
12
12
0
_
where
12
is a scalar valued 1-form on U. Also, by (29), we have the equations:
e
1
=
12
e
2
e
2
=
12
e
1
Further, we have the torsion free condition:
e1
e
2
e2
e
1
=
12
(e
2
) +
12
(e
1
) = [e
1
, e
2
]
Now [e
1
, e
2
] = [
, (sin )
1
] = cot e
2
. Thus
12
= cot e
2
= (cot sin )d = cos d. Thus, by
(29)
= d +
= d
_
0 cos d
cos d 0
_
+
_
0 cos d
cos d 0
_
_
0 cos d
cos d 0
_
=
_
0 sin d d
sin d d 0
_
+ 0
=
_
0 e
1
e
2
e
1
e
2
0
_
Thus, since e
i
is an orthonormal basis,
K(e
1
, e
2
) = R(e
1
, e
2
)e
2
, e
1
= (e
1
, e
2
)e
2
, e
1
= (e
1
e
2
) e
1
, e
1
= 1
(One could have also done the above calculation with Christoel symbols and the local formula for R.) This
shows that the sectional curvature of S
2
is the constant 1. (This follows from the symmetry of S
2
, for U could
be rotated to any other location. Or alternatively, since the (scalar) sectional curvature is continuous on S
2
,
and 1 on U, and U is dense in S
2
, it is 1 on S
2
.
For S
n
, one can argue that for two unit vectors X, Y in T
x
(S
n
), the quantity R(X, Y )Y lies entirely in the
span of X, Y . One successively uses the isometric reections about hyperplanes containing x, XandY , (similar
to the argument for geodesics on S
2
), and nally shows that this vector lies in the span of x, X and Y , but
since it is a tangent vector to S
n
, it must lie entirely in the span of X and Y . Then the sectional curvature
K(X, Y ) is precisely the sectional curvature of the intersection of the plane containing x, X, Y with S
n
, which
is an S
2
of radius 1, with its usual metric. Thus it is 1 at all points of S
n
.
Example 7.3.13 (Sectional curvature of H
2
). We recall the metric:
dx
2
+dy
2
y
2
so that
i
,
j
= y
2
ij
for 1 i, j 2. We could again mimic the above argument for S
2
to compute and
(which is a neater way of going about it and left to the reader), and instead, just like the ancients, do the
calculation with Christoel symbols.
Recalling that
j
=
k
ij
k
and dierentiating:
0 =
1
1
,
1
= 2
1
1
,
1

1
11
= 0
Similarly the relation
2
1
,
1
= 2y
3
yields
1
21
= y
1
. Dierentiating all the inner products
i
,
j
partially with respect to x and y yields all the Christoel symbols:
1
11
=
2
12
=
2
12
=
1
22
= 0
2
11
=
1
12
=
1
21
2
22
= y
1
From which we get
1
= y
1
1
,
1
1
= y
1
2
so that
1
= y
2
2
Thus R(
1
,
2
)
1
,
2
= y
4
. Since
1
,
1

2
,
2
= y
4
and
1
,
2
2
= 0, we get the sectional curvature of
H
2
to be 1.
7.4. Surfaces in R
3
. Let M R
3
be a smooth submanifold of dimension 2. We further assume that the
manifold is orientable. We give M the induced Riemannian metric from the euclidean metric on R
3
(as we did
for S
2
). We would like to compute curvatures, covariant derivatives etc. for M, in terms of the parametric
representation of M over trivialising charts h(U), where h = (h
1
, h
2
, h
3
) is a dieo (local parametrisation) of
M for some open subset U R
2
(e.g. the spherical polar coordinates of S
2
above.)
Because the manifold M is orientable, we can nd a system of charts {(h
i
, U
i
)} such that U
i
are opens in
R
2
, and det(h
1
j
h
i
) > 0 (on h
1
i
(h
i
(U
i
) h
j
(U
j
)) for all i and j. Thus for each such parametrisation
h = (h
1
, h
2
, h
3
) : U h(U) M R
3
the vectors:
u
:=
h
1
u

1
+
h
2
u

2
+
h
3
u

3
v
:=
h
1
v

1
+
h
2
v

2
+
h
3
v

3
where
1
=

x
etc. are the standard coordinate elds on R
3
, constitute an oriented basis for T
x
(M) for all
x h(U). In particular, the cross-product vector
n(u, v) =

u
v
|
u
v
|
is a unit vector in R
3
, normal to M (i.e. T
x
(M)) at all points x = h(u, v) U. In fact, this vector is globally
well-dened independent of charts, as the reader can readily verify, from the orientability of M and the choice
of the atlas {(h
i
, U
i
)} as an oriented atlas. It is called the unit normal eld to M.
Denition 7.4.1 (Gauss Map). The smooth map
n : M S
2
is called the Gauss map. This map captures all the curvature properties of M,as we shall see.
Denition 7.4.2 (1st and 2nd fundamental forms). The induced Riemannian metric on M, in terms of the
basis
u
,
v
introduced above, is
g(h(u, v)) = g
11
du du +g
12
du dv +g
21
dv du +g
22
dv dv
It is easily seen that
g
11
=
u
,
u
= |(
u
h
1
,
u
h
2
,
u
h
3
)|
2
g
22
=
v
,
v
= |(
u
h
1
,
u
h
2
,
u
h
3
)|
2
g
12
= g
21
=
u
,
v
= (
u
h
1
,
u
h
2
,
u
h
3
), (
v
h
1
,
v
h
2
,
v
h
3
)
where , always denotes the euclidean inner product in this subsection.
This 2 2 symmetric matrix (of the induced Riemannian metric) is called the 1st fundamental form.
Let us consider the Gauss map n : M S
2
. We will drop the hat on n for brevity.
Note that for x M, Dn(x) : T
x
(M) T
n(x)
S
2
. But T
n(x)
S
2
= (Rn(x))
= T
x
(M). Thus Dn(x) :
T
x
(M) T
x
(M). It is called the Weingarten map at x, and denoted S
x
. For Y, Z T
x
(M), the bilinear form:
S
x
(Y, Z) := S
x
Y, Z
is called the second fundamental form. We shall see presently that it is symmetric, and encodes covariant
dierentiation on M.
Lemma 7.4.3. Let M be as above, with the normal vector eld n as dened above. For a tangent vector
X to R
3
, we denote by
X
the covariant dierentiation on R
3
with respect to the (Levi Civita connection)
of the euclidean metric, and for X T
x
(M), D
X
denotes the (Levi- covariant derivative with respect to the
(Levi-Civita connection) of the Riemannian metric g induced on M. Note that we can write , without
ambiguity. Let S : TM TM denote the Weingarten map. Then:
(i): S(X) =
X
n for X TM.
(ii): S(X, Y ) := S(X), Y = S(Y ), X = S(Y, X) for all X, Y TM.
(iii):
X
Y = D
X
Y S(X, Y )n for X, Y TM.
Proof: We rst note that (i) has to be interpreted, because n is only a vector eld (section of the normal
bundle TM
) dened on M, and not on an open subset of R

3
. We just write n =
i
n
i
i
, where n
i
are smooth
functions on M, and
i
are the coordinate elds on R
3
. Then we mean:
u
n :=
3
i=1
u
(n
i
)
i
+
3
i=1
i
and similarly for
v
n, and this makes sense since
u
is a vector eld on M, acting on smooth functions n
i
on
M, and the eld above is well dened at points of M (though we dont know yet whether it is tangent to M.)
Now,
u
i
is a linear combination of
i
j
. But
i
j
=
k
ij
k
= 0 since all the Christoel symbols for
R
3
are 0, by (17)
Hence
u
n =
3
i=1
(
ni
u
)
i
= Dn(
u
) = S(
u
), and similarly for
v
n. This proves (i).
To see (ii), note that Y, n 0 for a vector eld Y on M. Thus 0 = XY, n =
X
Y, n + Y,
X
n =
X
Y, n + Y, S(X). That is, S(X, Y ) = S(X), Y =
X
Y, n. Thus from the torsionlessness of the
connection S(X, Y ) S(Y, X) = [X, Y ], n. But since X, Y are vector elds on M, [X, Y ] is a vector eld
on M, and thus [X, Y ], n 0. This proves (ii).
To see three, dene a covariant derivative operator on TM by D
X
Y =
X
Y S(X, Y )n. The Leibniz
formula easily follows from the Leibniz formula for . Further
D
X
Y D
Y
X =
X
Y +S(X, Y )n
Y
X S(Y, X)n = [X, Y ]
by the fact that is torsionless, and S is symmetric. Thus D is torsionless. To see compatibility with the
metric, for vector elds X, Y, Z on M,
XY, Z =
X
Y, Z +Y,
X
Z
= D
X
Y S(X, Y )n, Z +Y, D
X
Z S(X, Z)n
= D
X
Y, Z +Y, D
X
Z
since n is orthogonal to Y and Z. Thus D
X
is the Levi-Civita connection on TM, by the fact that the
Levi-Civita connection is unique (see 6.9.1). This proves the lemma. 2
Corollary 7.4.4. Let h = (h
1
, h
2
, h
3
) : U M be the parametric representation (coordinate chart) above.
Then the second fundamental form is given by:
S(
u
,
u
) =
_
2
h
u
2
, n
_
S(
u
,
v
) = S(
v
,
u
) =
_

2
h
uv
, n
_
S(
v
,
v
) =
_
2
h
v
2
, n
_
Proof: In the course of proving (ii) in the above proposition, we saw that S(X, Y ) =
X
Y, n. Since
X =
u
:= h
(

u
) =
i
(
u
h
i
)
i
then by the rst para in the proof of (i) above
u
=
3
i=1
u
(
u
h
i
)
i
which proves that S(
u
,
u
) =
u
u
, n =
_
2
h
u
2
, n
_
. The proofs of the other identities are identical. 2
Example 7.4.5. Consider the example of the sphere, with the parametrisation of (i) in 5.1.6. Thus:
h : (0, ) (0, 2) S
2
(u, v) (sin ucos v, sin usin v, cos u)
Clearly n(x) = n(h(u, v)) = h(u, v).
Then
u
=
u
h = (cos ucos v, cos usin v, sin u)
v
=
v
h = (sin usin v, sin ucos v, 0)
2
u
h = (sin ucos v, sin usin v, cos u)
u
h = (cos usin v, cos ucos v, 0)
2
v
h = (sin ucos v, sin usin v, 0)
So that the matrix of S with respect to {
u
,
v
} is
_
1 0
0 sin
2
u
_
That is, the second fundamental form is precisely the metric (=1st fundamental form) which it should be since
the Weingarten map S is the identity map.
Denition 7.4.6 (Principal curvatures). The 2 eigenvalues
1
(x),
2
(x) of the symmetric linear (Weingarten)
map S
x
: T
x
(M) T
x
(M) (= the eigenvalues of the matrix of the second fundamental form [S
x
(e
i
, e
j
)] where
e
i
are an orthonormal frame of T
x
(M)) are real numbers at each x, in fact smooth functions of x, and are
called the principal curvatures of M at x.
1
2
(
1
(x) +
2
(x)) is called the mean curvature at x. The product
1
(x)
2
(x) is called the scalar or Gauss curvature at x.
For example, in the case of the sphere S
2
above, the principal curvatures were both =1.
Exercise 7.4.7. Show that the Gauss curvature is
det[S(fi,fj)]
det[fi,fj]
where {f
1
, f
2
} is any frame of T
x
(M).
Proposition 7.4.8 (Theorema Egregium of Gauss). Let M be an orientable surface in R
3
as above. Then
the Gauss curvature is equal to the sectional curvature of the induced metric on M. In particular, the Gauss
curvature, which is dened in terms of the embeddings h, is an intrinsic isometric invariant of the Riemannian
manifold M.
Proof: By the remark 7.3.9, we just need to calculate R(X, Y )Y, X for a frame X, Y of T
x
(M). It is most
convenient to take X and Y to be coordinate elds (e.g.
u
,
v
), so that [X, Y ] = 0.
By (iii) of Lemma 7.4.3, and S(X) =
X
n, we have:
Y
Y =
X
(D
Y
Y S(Y, Y )n)
=
X
D
Y
Y S(Y, Y )
X
n X(S(Y, Y ))n
= D
X
D
Y
Y S(D
Y
Y, X)n S(Y, Y )S(X) X(S(Y, Y ))n
= D
X
D
Y
Y S(Y, Y )S(X) (X(S(Y, Y ) +S(D
Y
Y, X)) n
Thus, since n, X = 0, we have:
Y
Y, X = D
X
D
Y
Y, X S(Y, Y )S(X, X) (32)
Similarly
X
Y =
Y
(D
X
Y S(X, Y )n)
= D
Y
D
X
Y S(D
X
Y, Y )n Y (S(X, Y ))n S(X, Y )
Y
n
= D
Y
D
X
Y S(X, Y )S(Y ) (S(D
X
Y, Y ) +Y (S(X, Y ))) n
Taking inner product with X, and noting n, X = 0, we have
X
Y, X = D
Y
D
X
Y, X S(X, Y )
2
(33)
Subtracting (33) from (32), we get:
Y
Y
Y
X
Y, X = D
X
D
Y
Y D
Y
D
X
Y, X S(Y, Y )S(X, X) +S(X, Y )
2
But since [X, Y ] = 0, the left hand side is the sectional curvature of R
3
in the plane spanned by X, Y , and
hence 0. The right side is R(X, Y )Y, X , where is the determinant of the matrix of S with respect to
the basis {X, Y }. Dividing by X, X Y, Y X, Y
2
, on both sides and using the exercise 7.4.7, we are done.
2
Proposition 7.4.9 (Gauss-Mainardi-Codazzi relations). In the setting of the last proposition, we have the
following relations:
(i): D
X
S(Y ) D
Y
S(X) = S([X, Y ]).
(ii): R(X, Y )Z = S(Y, Z)S(X) S(X, Z)S(Y )
Proof: We have, from (i) of the lemma 7.4.3 that S(X) =
X
n. Thus:
D
X
S(Y ) = D
X
(
Y
n) =
X
(
Y
n) S(X,
Y
n)n
=
X
Y
n
X
n,
Y
n n
Interchanging X, Y and subtracting, and noting that on R
3
, atness implies
X
Y

Y
X
=
[X,Y ]
, we
have that
D
X
S(Y ) D
Y
S(X) =
[X,Y ]
n = S([X, Y ])
which proves (i).
For (ii), let X, Y, Z be vector elds on M. We recall the proof of (32) in 7.4.8, and show in identical fashion
that
Y
Z = D
X
D
Y
Z S(Y, Z)S(X) (X(S(Y, Z)) +S(D
Y
Z, X))n
Interchanging X, Y , and subtracting, we get:
[X,Y ]
Z = D
[X,Y ]
Z S([X, Y ], Z)n = D
X
D
Y
Z D
Y
D
X
Z S(Y, Z)S(X) +S(X, Z)S(Y ) n
where is some scalar. Taking the tangential component of this equation kills all the terms containing n, and
we get:
D
[X,Y ]
Z = D
X
D
Y
Z D
Y
D
X
Z S(Y, Z)S(X) +S(X, Z)S(Y )
from which we have (ii). 2
7.5. Gauss-Bonnet for surfaces.
Denition 7.5.1. A geodesic triangle in a Riemannian manifold is a triangle whose edges are geodesics. Simi-
larly geodesic polygons can be dened. If x
1
M is a vertex of a geodesic triangle x
1
x
2
x
3
, at which the geo-
desic edges and meet, with (1) = (0) = x, the external angle
1
at x
1
is dened as cos
1
_

(0), (1)
|
(0)| (1)
_
.
The internal angle at x
1
is dened as
1
=
1
.
Theorem 7.5.2 (Local Gauss-Bonnet). Let x
1
x
2
x
3
be a geodesic triangle in a coordinate chart U of an
oriented surface M, with external angles
i
=
i
(where
i
are the internal angles) at x
i
for i = 1, 2, 3.
Let k be the scalar (=sectional) curvature function of M, and let dA denote the volume (=area) form on M.
Then:
_
kdA = 2
3
i=1
i
=
3
i=1
Proof: We let
1
,
2
,
3
denote the geodesic segments x
1
x
2
, x
2
x
3
, x
3
x
1
respectively. We let c : S
1
M denote
the PS curve which is
i
on the segments [0, s
1
], [s
1
, s
2
], [s
2
, 2] of [0, 2] respectively. For simplicity we assume
i
are parametrised by arc length, so that the arc length of c is 2, and
i
is of unit length everywhere. (This
can be done by scaling the metric for example, so that k and dA get scaled by equal and opposite amounts,
thus leaving both left and right sides of the equation unchanged). We will denote the velocity vector, which is
dened at all s except 0, s
1
, s
2
, by T.
Let e
1
, e
2
be an orthonormal frame eld on U. Then, the connection form on U is a skew-symmetric 2 2
matrix with entry
21
=
12
. Thus
T
e
1
=
12
(T)e
2
and hence
12
(T) =
T
e
1
, e
2
. If we integrate the left side on the 1-chain c =
1
+
2
+
3
, we get by Stokes
formula that
_
c
12
=
_
(d
12
)
=
_
12
=
_
kdA
=
i
_
i
T
e
1
, e
2
Now we need to understand the last expression. Since

i
are geodesics, we have:
0 =
T
T =
T
(T, e
1
e
1
+T, e
2
e
2
)
=
2
i=1
(T(T, e
i
)e
i
+T, e
i

T
e
i
)
so that, on taking inner-product with e
1
, e
2
, and noting that
T
e
2
is a multiple of e
1
, and
T
e
1
is a multiple
of e
2
, we have the two equations:
T
e
2
, e
1
T, e
2
+T(T, e
1
) = 0 (34)
T
e
1
, e
2
T, e
1
+T(T, e
2
) = 0 (35)
Now since
T
e
2
, e
1
=
12
(T) =
T
e
1
, e
2
, we multiply the rst of the equations above with T, e
2
, the
second with T, e
1
and subtract the rst from the second, and use T, e
1
2
+T, e
2
2
= |T|
2
= 1, to get:
T
e
1
, e
2
= T, e
2
T(T, e
1
) T, e
1
T(T, e
2
)
But the right hand side is precisely the pullback T
(xdy ydx) = T
, where = xdy ydx is the volume

form of S
1
, and T : c S
1
is the piecewise smooth map, taking c(s) to (T(c(s)), e
1
(c(s)) , T(c(s), e
2
(c(s)).
It is a result of Hopf (called the Umlaufsatz) that this integral, viz.
_
c
T
= 2
3
i=1

i
. This proves the
result. 2
Proposition 7.5.3 (Global Gauss-Bonnet). Let M be a compact oriented surface. Then:
_
M
kdA = 2(M)
where (M) is the Euler characteristic of M.
Proof: It is a theorem (deep) that a surface as above can be triangulated by means of geodesic triangles. That
is, there is a nite subset V M (called vertices), and geodesic triangles (homeomorphic to 2-simplices) called
faces with vertices coming from V , such that M is the union of these triangles, and such that a pair of triangles
can meet at either at a single vertex, or along a unique geodesic common edge, or not meet at all. Also every
edge is a common edge of exactly two faces. And nally, we can choose such a triangulation to be ner than
a covering by open charts. Then, if we denote the number of vertices, edges, faces by v, e, f respectively, one
denition of the Euler characteristic is (M) = v e +f. (That it does not depend on the triangulation, and
is a topological invariant follows from homology theory). If we count three edges to each face, then we would
get 3f, but in this counting, each edge would have been counted twice, since it is a common edge to two faces.
Thus, 2e = 3f, or e =
3f
2
.
Denote by F the set of faces. Then note that at each vertex, the sum of the internal angles
i
at x of all
the faces having x as a vertex, is 2. Thus:
_
M
kdA =
x1x2x3F
(
1
+
2
+
3
)
= 2v f = 2(v
3f
2
+f)
= 2(v e +f) = (M)
thus proving the assertion. 2
Remark 7.5.4. To compute (M), it is not necessary to have a geodesic triangulation, any triangulation will
do. The result above says that there is a global topological obstruction to putting a metric of prescribed scalar
curvature on a compact oriented manifold. For example, since (S
2
) = 2, one cannot put a Riemannian metric
on S
2
whose scalar curvature is everywhere 0. Similarly, on a torus T
2
, whose Euler characteristic is zero,
one cannot put any Riemannian metric whose scalar curvature is everywhere positive, or everywhere negative.
It turns out that the result above generalises to oriented Riemannian manifolds of all even dimension. Odd
dimensional oriented Riemannian manifolds have Euler characteristic zero.
8. Jacobi Fields
8.1. Variations of geodesics. Let M be a Riemannian manifold, and x M be a point. Let : [0, a] M
be a geodesic (parametrised by arc-length), with (0) = x. A variation of geodesics is a smooth map:
: (, ) [0, a] M
(s, t) (s, t)
such that (0, t) = (t) and for each s,
s
:= (s,
)
is a geodesic. For example, if (t) = Exp
x
(tX) for some
tangent vector X T
x
(M), and if W T
x
(M) is another tangent vector, we can consider the family
s
of
geodesics:
s
(t) = (s, t) := Exp
x
(tX +sW)
parametrised by s (, ), with
0
= , and
s
(0) = Exp
x
(sW).
Denition 8.1.1 (Jacobi Field of a variation). By the Jacobi eld of the variation , we mean the vector eld
along dened as:
J(t) :=
(s, t)
s |s=0
Proposition 8.1.2. The Jacobi eld J(t) satises the following 2nd- order linear dierential equation:
D
2
J
dt
2
+R(J, T)T = 0
where T(t) := (t). The Jacobi eld J(t) is uniquely determined by the values J(0) and J(a).
Proof: To get the dierential equation, note that the vector elds
s
:=
(s, t)(
s
) and
t
:=
(s, t)(
t
),
which are vector elds along (, ) (0, a) (i.e. smooth sections of
TM) commute (why?). Note also

that the restrictions of
s
and
t
to (0) (0, a) are precisely J(t) and T(t) respectively. In particular, J and
T commute. Thus:
R(
s
,
t
)
t
=
D
ds
D
t
dt

D
dt
D
t
ds
=
D
dt
D
t
ds
But
Dt
ds

Ds
dt
= [
s
,
t
] = 0, so that we have, on setting s = 0, the relation:
D
2
J
dt
2
+R(J, T)T = 0
which is the required ODE. The uniqueness given the boundary conditions follows from the existence and
uniqueness theorem of linear 2nd-order ODEs. 2
Denition 8.1.3 (Jacobi Field). Let : [0, a] M be a geodesic, and let J be a smooth vector eld along
. Then we say J is a Jacobi eld if J satises:
D
2
J
dt
2
+R(J, T)T = 0
where T = (t). Clearly, J is uniquely determined by either (i) J(0) and J(a) or (ii) J(0) and
DJ
dt
(0).
More generally, we shall say J is a broken Jacobi eld if there exists a subdivision 0 = t
0
< t
1
< t
2
< ... <
t
k
= a of [0, a] such that J
|[ti,ti+1]
is a smooth Jacobi eld for each i. We do not require J to be even continuous
on [0, a]. If J is a broken Jacobi eld along , it is uniquely determined by J(0), J(a), J(t
i
+) and J(t
i
) for
i = 1, .., k 1.
Remark 8.1.4 (Broken vs smooth Jacobi elds). Let J be a broken Jacobi eld along a geodesic , and in
the notation of the denition above, assume that for each i, we have (i)J(t
i
+) = J(t
i
), and (ii)
DJ
dt
(t
i
+) =
DJ
dt
(t
i
). Then the broken Jacobi eld J is smooth. This is clear, by the hypothesis above, the 0-th and 1-st
order derivatives from the left and right agree at t
i
, and by the ODE of 8.1.2, all derivatives of J of order 2
from the left and right also agree at the points t
i
.
Lemma 8.1.5. Let : [0, a] M be a geodesic, with (0) = x. Then every Jacobi eld along is the Jacobi
eld of some geodesic variation.
Proof: Let J be a Jacobi eld along . We rst assume (t) = Exp
x
(tX). This is possible if a is small enough,
for example such that lies entirely in a neighbourhood U of x such that Exp
x
: U
1
U is a dieomorphism
for some neighbourhood U
1
of 0 in T
x
M, and aX U
1
.
Let : (, ) M be a path such that (0) = x, and (0) = J(0). Let X(s) be a smooth vector eld
along such that:
X(0) = X
DJ
dt
(0) =
DX
ds
(0) (36)
(prove that such a vector eld exists). Consider the variation:
(s, t) = Exp
(s)
(tX(s))
Note that for a xed s, (s, ) is a geodesic, so that is a geodesic variation. Clearly, by (36), (0, t) =
Exp
x
(tX) = (t). Also,
(s, 0)
s
=
Exp
(s)
(0)
s
= (s) (37)
Now, we also have
t
(s, 0) = X(s) =
t
(s, 0)
so that
Dt(s,0)
ds
=
DX(s)
ds
. But since [
s
,
t
] = 0, we have:
D
s
(s, 0)
dt
=
D
t
(s, 0)
ds
=
DX(s)
ds
Denote the Jacobi eld of the variation by J
1
(t) =
s
(0, t). Since
Ds(0,t)
dt
=
DJ1
dt
(t), the above calculation
implies
DX(s)
ds
(0) =
DJ1
dt
(0). But by (36),
DX(s)
ds
(0) =
DJ
dt
(0), so we have:
DJ
1
dt
(0) =
DJ
dt
(0) (38)
Both J and J
1
satisfy the 2nd order Jacobi ODE of 8.1.2. The equation (37) implies that J
1
(0) = (0) = J(0).
Combined with (38), the uniqueness part of the ODE theorem implies that J(t) J
1
(t) for all t [0, a].
In the general case, we break [0, a] into a partition 0 = t
0
< t
1
< .. < t
k
= a such that each segment [t
i
, t
i+1
]
is a small geodesic segment of the kind considered above. We construct geodesic variations over each segment
[t
i
, t
i+1
], and patch them up. The details are left to the reader. 2
Lemma 8.1.6 (Second Variation Formula). Let : [0, a] M be a continuous curve, and let 0 = t
0
< t
1
<
... < t
k
= a be a partition of [0, a] such that
|[ti,ti+1]
is a smooth geodesic, for each 0 i k 1. (That is,
is a broken geodesic), with velocity eld T(t) on the intervals of smoothness [t
i
, t
i+1
]. Let:
: (, ) (, ) [0, a] M
(r, s, t) (r, s, t)
be a two parameter variation which is continuous, and smooth on (, )(, )[t
i
, t
i+1
] for each 0 i k1,
and such that (0, 0, t) = (t). Let Y (t) and Z(t) denote the (piecewise smooth) vector elds along dened
by

r
(0, 0, t) and

s
(0, 0, t) respectively. Denote by:
E(r, s) :=
k1
i=0
__
ti+1
ti
t
(r, s, t),
t
(r, s, t) dt
_
the energy functional. Then,
1
2
E
(Y, Z) :=
1
2
2
E
rs
(0, 0) =
Proof:
We note that the vector elds along (, )(, )[t
i
, t
i+1
] dened by
r
:=

r
,
s
:=

s
, and
t
:=

t
are smooth sections of
TM, and commute, for each 0 i k 1. Dierentiating once:

E(r, s)
s
= 2
i
_
ti+1
ti
_
D
t
s
,
t
_
dt
So that
1
2
2
E(r, s)
rs
=
i
__
ti+1
ti
_
D
t
s
,
D
t
r
_
dt +
_
ti+1
ti
_
D
r
D
t
s
,
t
_
dt
_
=
i
__
ti+1
ti
_
D
s
t
,
D
r
t
_
dt +
_
ti+1
ti
_
D
r
D
s
t
,
t
_
dt
_
=
i
__
ti+1
ti
_
d
dt
_
r
,
D
s
t
_
r
,
D
2
s
t
2
__
dt +
_
ti+1
ti
_
D
r
D
s
t
,
t
_
dt
_
Now denote the jumps:
i
_
Y,
DZ
t
_
:=
_
Y (t
i
+),
DZ
t
(t
i
+)
_
_
Y (t
i
),
DZ
t
(t
i
)
_
and use the relation
D
r
D
s
t
=
D
t
D
s
r
+R(
r
,
t
)
s
and observe that:
_
D
t
D
s
r
,
t
_
=
d
dt
_
D
s
r
,
t
_
_
D
s
r
,
D
t
t
_
so that when s = r = 0,
t
= T, and since
T
T = 0, we have:
1
2
2
E
rs
(0, 0) =
_
Y (a),
DZ
t
(a)
_
_
Y (0),
DZ
t
(0)
_
k1
i=1
i
_
Y,
DZ
t
_
k1
i=0
_
ti+1
ti
__
Y,
D
2
Z
t
2
_
T, R(Y, T)Z
_
Now, from the relation T, R(Y, T)Z = T, R(Z, T)Y = Y, R(Z, T)T (identities (iii) and (iv) of the
proposition 7.3.4), we nally have:
1
2
2
E
rs
(0, 0) =
_
Y (a),
DZ
t
(a)
_
_
Y (0),
DZ
t
(0)
_
k1
i=1
i
_
Y,
DZ
t
_
k1
i=0
_
ti+1
ti
__
Y,
D
2
Z
t
2
+R(Z, T)T
__
and the proposition follows. 2
I.S.I., Bangalore 560 059

Diff Geo

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Diff Geo

Загружено:

Авторское право:

Доступные форматы

ELEMENTARY RIEMANNIAN GEOMETRY

-dieomorphism is called a smooth dieomorphism. A C

= B(0, /2), then z w B = B(0, ), so that we have, for as above:

= B(0, /2) (2)

= B(0, /2) (3)

= B(0, /2) to itself, and is a contraction map on B

, b is the unique inverse image of y in B

. We let V = B(0, /4) and W = f

, as also the continuity of f

. This implies that f

) map. We say that f is

-mapping if each component function is a real analytic

functions and maps unless otherwise stated.

dierentiable manifold X of dimension n is a paracompact, second count-

, as a map between open subsets of R

of the orthogonal projection

Its inverse is given by

). We need to check that the topologies thus induced on U

), note that a basis for the k-subspace

). We need to show that S =

As noted earlier, for T (U

on all of M, and also

is a normalising constant such that

becomes more concentrated around the

(y) = 0. Thus we have (x) = 0 for all x U

(y)dy = 1, we have for x R

. The details are left to the reader. 2

h) where h is any germ in O

h denotes the germ of h f

. That Df(x)(X) is a derivation easily follows from the fact that X is a

M is called a smooth (resp. analytic)

denotes the bundle hom

) via the chart

E := {(x, v) X E : f(x) = (v)}

E X by p(x, v) = x (i.e. the restriction of the projection X E X). This bundle

f denote the map (x, v) v, then we have a commutative square:

E, by virtue of the uniqueness of

TN (as bundles on M) which is denoted by f

TN. If f is a submersion, it will realise f

M M, whose sections are called k-forms on M.

M get trivialised over U. Their local frames

M) M then acquires the trivialisation over U, given by the frame:

N, so tautologically results in a section

N). We apply the above vector bundle morphism

) to this pullback, and get a section of

g = g f. Also show that f

where the right side is dened by the foregoing remarks. Finally, if c =

the last line following from the fact that (i

Finally, for regular k-chains, we obviously have the result by linearity. 2

:= R\{0}, (after identifying

, with the point (t, ) in the rst copy identied with

X such that g(x) E

is dened by the equation g(T

f, e) = h(f, Te). Verify that this is a Riemannian metric on the

where the right side is the Euclidean inner product.

is a linear map, we note that there is an adjoint of this map X

P, which is dened by:

where , is the euclidean inner product on R

) as dened in 2.2.6. This follows from

) and the xed

). To do this we note that on U

). The reader should verify that L

), Ad(g)T is dened as the composite:

) constructed above. Thus,