X

V φ−1
x
DIFFERENTIAL GEOMETRY φ−1 (x)
U

ADAM KELLY – MATHEMATICAL TRIPOS PART II

Definition 1.4 (Chart)

A diffeomorphism φ : U → V , where U is an open
set of Rk is a parameterisation of the neighbor-
1. Differential Topology hood V . The inverse diffeomorphism φ−1 : V → U
is called a chart on V .
1.1. Manifolds. The idea of a manifold is to general-
ize our typical boring and flat ‘Euclidean space’ with
Of course, if we have manifolds X and Z with Z ⊆ X
an object where everything isn’t necessarily flat. For
we say that Z is a submanifold of X. In this case, the
example, we can take some surfaces as drawn below.
codimension of Z in X is dim X − dim Z.
1.2. Tangent Spaces and Derivatives. This is a course
in differential geometry, and we’ve pinned down the ‘ge-
ometry’ part, now it’s time to pin down the ‘differential
part’. We want to generalise the notion of a derivative
for smooth functions on manifolds.
The idea you should have in your head of derivatives are
that they are the best linear approximation to a given
function. From this perspective, defining the derivative
We really care about surfaces where, in a sufficiently on manifolds requires us to answer two questions: what
small region around each point, things look like we are space should be our ‘linear approximation’ to a mani-
in normal flat Euclidean space. This, along with a de- fold that our derivative should act on, and then what is
sire for things to be nice and smooth (as we shall define) our linear approximation? Answering the first of these
will manifest into our basic notion of a manifold. questions will lead us to the notion of tangent spaces.

Definition 1.1 (Smooth) x + D0 φ x + Tx X

For an open subset U ⊆ Rn , we say a function x
X
f : U → Rm is smootha if it has continuous partial
U
derivatives of all orders. Affine tangent plane
aFor a function f : X → Rm with X not open, we say Tx X
that f is smooth if the above definition holds for some open D0 φ
neighborhood Vx containing x for each x ∈ X. Tangent space

Definition 1.5 (Tangent Space of a Manifold)

Definition 1.2 (Diffeomorphism)
Let X ⊆ Rn be an arbitrary smooth manifold, and
A smooth function f : X → Y is a diffeomor- let x ∈ X. Choose a parameterisation φ : U → X
phism if it’s a bijection with smooth inverse. If with φ(0) = x where U is an open set of Rk . Then
a diffeomorphism between X and Y exists we say thinking of φ as a map from U to RN , we define
they are diffeomorphic. the tangent space Tx X to be the image of the
derivative of φ at 0, D0 φ(Rk ).

Definition 1.3 (Manifold)

In this definition we had to pick some parameterisa-
We say that X ⊆ Rn is a k-dimensional man- tion φ, but one would hope that our resulting space is
ifold if each x ∈ X has a neighborhood V ⊆ X independent of this choice. Indeed, this is the case.
diffeomorphic to an open set of Rk .
Lemma 1.6 (Tangent Spaces are Paramaterisation
With neighborhoods around each point looking like Rk , Independent)
we can sensibly imagine that one could locally put a Tx X does not depend on the parameterisation φ
coordinate system on such neighborhoods, one coming chosen.
from the diffeomorphism to Rk . This motivates the
definition of a chart on a manifold.

Date: May 29, 2023. Email [email protected].

1
2 ADAM KELLY – MATHEMATICAL TRIPOS PART II

Proof. Suppose ψ : V → X is another choice of Theorem 1.8 (Inverse Function Theorem)

parameteristaion, and suppose without loss of gen-
Suppose that f : X → Y is a smooth map whose
erality that φ(0) = ψ(0) = x. By shrinking both U
derivative dfx at the point x is an isomorphism.
and V , we may assume that φ(U ) = ψ(V ). Then
Then f is a local diffeomorphism at x.
the map h = ψ −1 ◦φ : U → V is a diffeomorphism.
Now we write φ = ψ ◦h and we differentiate. Using
the chain rule we have: D0 φ = D0 ψ ◦ D0 h. Since Proof Sketch. This can be proven using the inverse
D0 h is an invertible linear map, it follows at once function theorem for smooth functions between open
that D0 φ and D0 ψ have the same image.  sets of Euclidean space, as was shown in Analysis
II. 
We are now ready to define the derivative of a smooth
map f : X → Y between arbitrary manifolds, and we 1.3. Regular values and Sard’s theorem. Let f :
will see that it comes rather naturally. We would hope X → Y be a smooth map between manifolds. Let C
that for X and Y open subsets of euclidean space, our be the set of all points x ∈ X such that dfx : Tx X →
new definition should coincide with our previous defini- Tf (x) Y is not surjective.
tion of the derivative. Also, we would expect that the
chain rule should still hold. This is going to force our Definition 1.9
definition.
A point in C will be called a critical point. A
Let φ : U → X and ψ : V → Y be parameterisations, point in f (C) will be called a critical value. A
with φ(0) = x, ψ(0) = y and U, V being open subsets of point in the complement of f (C) will be called a
Rk and Rl respectively. Then for small enough U , the regular value.
following diagram commutes:
Remark. Note that if dim X < dim Y , then C = X
X
f
Y and the preimage of a regular value is the empty set.
φ ψ
h=ψ −1 ◦f ◦φ Theorem 1.10 (Preimage Theorem)
U V
Let y be a regular value of f : X → Y with
Then D0 ψ, D0 φ, D0 h are all well defined, and as we dim X ≥ dim Y . Then the set f −1 (y) is a sub-
said previously we would hope the chain rule still holds, manifold of X with dim f −1 (y) = dim X − dim Y .
which implies that the following square of linear maps
commutes:
Proof. Let x ∈ f −1 (y). Since y is a regular value,
dfx
the derivative dfx maps Tx X onto Ty Y . The kernel
Tx X Ty Y of dfx is a subspace K of Tx X of dimension p :=
dφ0 dψ0
dim X − dim Y . Suppose X ⊂ RN and let T :
RN → Rp be any linear map such that Ker(T ) ∩
dh0
Rk Rl K = {0}. Consider the map F : X → Y × Rp
given by
Then since D0 φ is an isomorphism, only one definition F (z) = (f (z), T (z)).
of Dx f will meet our requirements, namely
The derivative of F is given by
Dx f = D0 ψ ◦ D0 h ◦ D0 φ−1 .
dFx (v) = (dfx (v), T (v))
Again, one would hope that this definition of Dx f is which is clearly nonsingular by our choice of T . By
independent of φ and ψ, and this is the case (and the the inverse function theorem, F is a local diffeo-
same proof as before works). Also, we have forced the morphism at x, i.e. F maps some neighbourhood
chain rule still to hold. U of x diffeomorphically onto a neighbourhood V
of (y, T (x)). Hence F maps f −1 (y) ∩ U diffeo-
Theorem 1.7 (Chain Rule) morphically onto ({y} × Rp )∩V which proves that
f −1 (y) is a manifold with dim f −1 (y) = p. 
If f : X → Y and g : Y → Z are smooth maps
between manifolds, then
The case dim X = dim Y is particularly important. The
Dx (g ◦ f ) = Df (x) g ◦ Dx f. theorem says that if y is a regular value of f , then
f −1 (y) is a 0 -dimensional manifold, i.e. collection of
Proof. Omitted.  points. If X is compact, this collection must be finite
(why?) so we obtain:
1.2.1. The Inverse Function Theorem. Let f : X → Y
be a smooth map between manifolds. We say that f is a Corollary 1.11
local diffeomorphism at x if f maps a neighbourhood
of x diffeomorphically onto a neighbourhood of f (x).
 DIFFERENTIAL GEOMETRY 3

Let f : X → Y be a smooth map between mani- such that

folds of the same dimension. If X is compact and dfA (H) = HAt + AH t = C.
y is a regular value, f −1 (y) consists of a finite set
of points. Since C is symmetric, we can write C = 21 C +
1 t t 1
2 C so if we can solve HA = 2 C we are done.
We can actually say a bit more: Multiplying by A on the right and using AAt = I
we obtain H = 12 CA which solves dfA (H) = C.
Thus I is a regular value of f .
Theorem 1.12 (Stack of Records Theorem)
Let f : X → Y be a smooth map between mani- Note that O(n) is both a group and a manifold.
folds of the same dimension with X compact. Let y In fact, the group operations are smooth, that is,
be a regular value of f and write f −1 (y) = {x1 , . . . , xk }. the maps (A, B) 7→ AB and A 7→ A−1 = At are
Then there exists a neighbourhood U of y in Y such smooth. (Why?) A group that is manifold, and
that f −1 (U ) is a disjoint union V1 ∪ · · · ∪ Vk , where whose group operations are smooth, is called a Lie
Vi is an open neighbourhood of xi and f maps each group.
Vi diffeomorphically onto U .
1.3.1. Sard’s Theorem. The Preimage theorem raises
the question: is it easy to find regular values? How
Proof. By the inverse function theorem we can pick are abundant are they?
disjoint neighbourhoods Wi of xi such that f maps
Wi diffeomorphically onto a neighbourhood of y. Recall that an arbitrary set A in Rn has measure zero
Observe that f (X − ∪i Wi ) is a compact set which if it can be covered by a countable number of rectan-
does not contain y. Now take gular solids with arbitrary small total volume. In other
words given ε > 0, there exists a countable collection
U := ∩i f (Wi ) − f (X − ∪i Wi ) .
{R1 , R2 , . . .} of rectangular solids in Rn , such that A is
 contained in ∪i Ri and
X
Vol (Ri ) < ε.
If we let #f −1 (y) be the cardinality of f −1 (y), the the-
i
orem implies that the function y 7→ #f −1 (y) is locally
Let X be manifold. An arbitrary subset A ⊂ X has
constant as y ranges over regular values of f .
measure zero if, for every local parametrization φ of
The Preimage theorem gives a particularly nice way of X, φ−1 (A) has measure zero in Euclidean space.
producing manifolds. For example, if we consider the
The following deep theorem, tells us that there are
map f : Rn+1 → R given by f (x) = |x|2 we can check
plenty of regular values.
easily that 1 is a regular value of f (do it!). Hence
S n = f −1 (1) is a smooth manifold of dimension n. By
switching our mental wavelength a little we can produce Theorem 1.14 (Sard’s Theorem, 1942)
quite interesting examples as follows: The set of critical values of a smooth map f : X →
Y has measure zero.
Example 1.13 (Orthogonal Group)
Let O(n) be the group of orthogonal matrices of Since a set of measure zero cannot contain a non-empty
size n × n, i.e., a matrix A ∈ O(n) if and only if open set we obtain:
AAt = I, where I is the identity matrix. The space
2
of all n × n matrices M (n) is just Rn , so we can Corollary 1.15
2
think of O(n) as living inside Rn . We will show The regular values of any smooth map f : X → Y
that O(n) is a manifold of dimension n(n−1)
2 . are dense in Y .

Let S(n) ⊂ M (n) be the space of all symmetric

1.3.2. Morse Functions. See the first example sheet.
matrices. Since it is a vector space, is clearly a
submanifold of M (n) and it has dimension n(n+1)
2 . 1.4. Transversality. We know that the solutions of
the equation f (x) = y form a smooth manifold, pro-
Let f : M (n) → S(n) be the smooth map f (A) = vided that y is a regular value of f : X → Y . Suppose
AAt . Since O(n) = f −1 (I) it suffices to show that now that we replace y by a submanifold Z ⊂ Y and we
I is a regular value of f . We compute: ask, when is f −1 (Z) a submanifold of X ?
f (A + sH) − f (A)
dfA (H) = lim
s→0 s Definition 1.16 (Transversal)
(A + sH)(A + sH)t − AAt
= lim A smooth map f : X → Y is said to be transver-
s→0 s sal to a submanifold Z ⊂ Y if for every x ∈ f −1 (Z)
t t
= HA + AH . we have
Fix A with AAt = I. We must prove that given any Image (dfx ) + Tf (x) Z = Tf (x) Y.
C ∈ S(n) = TI S(n) there is H ∈ M (n) = TA M (n)
4 ADAM KELLY – MATHEMATICAL TRIPOS PART II

We write f t Z. The boundary ∂Hn is defined to be the hyperplane xk =

0 in Rk .
Note that if Z = {y}, the notion of transversality re-
duces to the notion of regular value. Definition 1.20
A subset X ⊂ RN is called a smooth k-manifold
With this definition, we can now state a more general
with boundary if each x ∈ X has a neighbourhood
version of the Preimage theorem.
diffeomorphic to an open set in Hk . As before,
such a diffeomorphism is called a chart on X. The
Theorem 1.17 boundary of X, denoted ∂X, is given by the set of
If the smooth map f : X → Y is transversal to points that belong to the image of ∂Hk under some
a submanifold Z ⊂ Y , then f −1 (Z) is submani- local parametrization. Its complement is called the
fold of X. Moreover, f −1 (Z) and Z have the same interior of X, Int(X) = X − ∂X.
codimension.
Remark (Warning). Do not confuse the boundary or
Proof (Non-examinable, sketch only). It is not hard interior of X as defined above with the topological no-
to see that Z can be written in a neighbourhood tions of interior and boundary as a subset of RN .
of a point y = f (x) as the zero set of a collection
of functions h1 , . . . , hr , where r is the codimen- The tangent space is defined as before, so that Tx X is
sion of Z in Y . Let H := (h1 , . . . , hr ). Then near a k-dimensional vector space even for points x ∈ ∂X.
x, f −1 (Z) is the zero set of the function H ◦ f . The interior of X is a k-manifold without boundary and
Thus if 0 ∈ Rr is a regular value of H ◦ f we are ∂X is a manifold without boundary of dimension k − 1
done. But dHy ◦ dfx is surjective if and only if (this requires a proof!).
Image (dfx ) + Tf (x) Z = Tf (x) Y Here is an easy way of generating examples.
since dHy : Ty Y → Rr is onto with kernel Ty Z. 
Lemma 1.21
Let X be a manifold without boundary and let
Example 1.18 f : X → R be a smooth function with 0 as a regular
Consider f : R → R2 given by f (t) = (0, t) and let value. Then the subset {x ∈ X : f (x) ≥ 0} is a
Z be the x axis in R2 . Then f is transversal

to Z, smooth manifold with boundary equal to f −1 (0).
but for example the map h(t) = t, t2 is not.

Proof. The set where f > 0 is open in X and is

An important special case occurs when f is the inclusion therefore a submanifold of the same dimension as
of a submanifold X of Y and Z is another submanifold X. For a point x ∈ X with f (x) = 0, the same
of Y . In this case the condition of transversality reduces proof of the Preimage Theorem 1.13 shows that
to x has a neighbourhood diffeomorphic to a neigh-
Tx X + Tx Z = Tx Y bourhood of a point in Hk . 
for every x ∈ X ∩ Z. This condition is quite easy to
visualize and when it holds we say that X and Z are As an easy application of the lemma, consider the unit
transversal (we write X t Z). We now have: ball B k given by all x ∈ Rk such that |x| ≤ 1. By
considering the function f (x) = 1 − |x|2 it follows that
Theorem 1.19 B k is a smooth manifold with boundary S k−1 .
The intersection of two transversal submanifolds of
Y is a submanifold of codimension given by Theorem 1.22
codim(X ∩ Z) = codim X + codim Z. Let f : X → Y be a smooth map from an m-
manifold with boundary to an n-manifold, where
m > n. If y is a regular value, both for f and
One of the main virtues of the notion of transversal- for the restriction of f to ∂X, then f −1 (y) is a
ity is its stability i.e. it survives after small pertur- smooth (m − n)-manifold with boundary equal to
bations of the map f . You can convince yourself of f −1 (y) ∩ ∂X
this property by looking at pictures of transversal sub-
manifolds of Euclidean space. Another very important
virtue is genericity: smooth maps may be deformed by Proof. Recall that being a submanifold is a local
arbitrary small amounts into a map that is transversal property so without loss of generality we can sup-
to Z. (This is non-obvious and we refer to Chapter 2 pose that f : Hm → Rn with y ∈ Rn a regular
in the book by Guillemin and Pollack for details.) value. Consider z ∈ f −1 (y). If z belongs to the
interior of Hm , then as in the Preimage theorem
1.5. Manifolds with Boundary. Consider the closed 1.13f −1 (y) is a smooth manifold near z.
half-space
Hk := (x1 , . . . , xk ) ∈ Rk : xk ≥ 0 .

 DIFFERENTIAL GEOMETRY 5

Let z be now in ∂Hm . Since f is smooth, there is the same dimension as Y and ∂X = ∅. If y is a
a neighbourhood U of z in Rm and a smooth map regular value for both f and g, then
F : U → Rn such that F restricted to U ∩ Hm is f . #f −1 (y) = #g −1 (y) (mod 2).
By shrinking U if necessary we can assume that F
has no critical points in U (why?). Hence F −1 (y)
is a smooth manifold of dimension m − n. Proof. The proof relies on the following impor-
tant fact, the ‘Classification of 1-manifolds’: Every
Let π : F −1 (y) → R be the projection (x1 , . . . , xm ) 7→ compact connected 1-manifold is diffeomorphic to
xm . Observe that the tangent space of F −1 (y) at [0, 1] or S 1 .
a point x ∈ π −1 (0) is equal to the kernel of
As every compact manifold is the disjoint union
dFx = dfx : Rm → Rn .
of finitely many compact connected manifolds we
Hence 0 must be a regular value of π since we are have that the boundary of any compact 1-manifold
assuming that y is a regular value of f restricted consists of an even number of points.
to ∂Hm .
Let F : X × [0, 1] → Y be a smooth homotopy
But F −1 (y) ∩ Hm = f −1 (y) ∩ U is the set of all between f and g. Assume for the time being that
x ∈ F −1 (y) with π(x) ≥ 0 and by Lemma 1.25, is y is also a regular value for F . Then F −1 (y) is a
a smooth manifold with boundary equal to π −1 (0). compact 1-manifold with boundary equal to

F −1 (y)∩(X×{0}∪X×{1}) = f −1 (y)×{0}∪g −1 (y)×{1}.
It is not hard to guess what is the appropriate version Thus the cardinality of the boundary of F −1 (y) is
of this theorem for the more general case of a map f : just #f −1 (y) + #g −1 (y). By the corollary above
X → Y and a submanifold Z of Y . Suppose X has #f −1 (y) = #g −1 (y)(mod2).
boundary, but Y and Z are boundaryless. The next
theorem is stated without proof. If y is not a regular value of F we proceed as
follows. From the Stack of records theorem 1.15
we know that w 7→ #f −1 (w), w 7→ #g −1 (w) are
Theorem 1.23
locally constant as w ranges over regular values.
Suppose that both f : X → Y and f | ∂X : ∂X → Thus there are neighbourhoods V and W of y, con-
Y are transversal to Z. Then f −1 (Z) is a manifold sisting of regular values of f and g respectively for
with boundary given by f −1 (Z) ∩ ∂X and codi- which
mension equal to the codimension of Z. #f −1 (w) = #f −1 (y)
for all w ∈ V , and
1.6. Degree Modulo 2. Let X be a smooth bound-
#g −1 (w) = #g −1 (y)
aryless manifold. Then X × [0, 1] is a manifold with
boundary ∂X = X × {0} ∪ X × {1}. for all w ∈ W . By Sard’s theorem we can choose
a regular value z of F in V ∩ W . Then
Definition 1.24 #f −1 (y) = #f −1 (z) = #g −1 (z) = #g −1 (y)( mod 2)
Two maps f, g : X → Y are called smoothly ho- as desired. 
motopic if there exists a smooth map F : X ×
[0, 1] → Y with F (x, 0) = f (x) and F (x, 1) = g(x)
for all x ∈ X. The map F is called a smooth ho- Lemma 1.27 (Homogeneity Lemma)
motopy between f and g. The relation of smooth Let X be a smooth connected manifold, possibly
homotopy is an equivalence relation. (Why?) The with boundary. Let y and z be points in Int(X).
equivalence class to which a map belongs is its ho- Then there exists a diffeomorphism h : X → X
motopy class. Let ft : X → Y be the 1-parameter smoothly isotopic to the identity such that h(y) =
family of maps given by ft (x) = F (x, t). z.

Proof. Let us call two points y and z “isotopic”

Definition 1.25 if there exists a diffeomorphism h isotopic to the
The diffeomorphism f is smoothly isotopic to g if identity that maps y to z. It is evident that this
there exists a smooth homotopy F : X × [0, 1] → Y is an equivalence relation. Since X is connected,
from f to g such that for each t ∈ [0, 1], ft maps X it suffices to show that each equivalence class is a
diffeomorphically onto Y . open set.
To prove that equivalence classes are open we will
construct an isotopy ht of Rk such that h0 is the
Lemma 1.26 (Homotopy Lemma) identity, each ht is the identity outside the open
Let f, g : X → Y be smooth maps which are unit ball around the origin, and h1 (0) is any spec-
smoothly homotopic. Suppose X is compact, has ified point in the open unit ball. This will suffice,
6 ADAM KELLY – MATHEMATICAL TRIPOS PART II

since we can now parametrize a small neighbour- The identity map of a compact boundaryless man-
hood of y in Int(X) and use ht to construct an ifold X has deg2 = 1 and the constant map has
isotopy that will move y to any point nearby. deg2 = 0. Therefore they are never homotopic.
When X = S n , this implies that there is no smooth
Let ϕ : Rk → R be a smooth function such that
map f : B k+1 → S k which restricts to the iden-
ϕ(x) > 0 for |x| < 1 and ϕ(x) = 0 for |x| ≥ 1
tity on S k (i.e. there is no retraction). Indeed,
(why do they exist?). Given a unit vector u ∈ Rk
if such a map exists it would give rise to a homo-
consider the ordinary differential equation in Rk
topy F : S k × [0, 1] → S k , F (x, t) = f (tx) which
given by
dx is a homotopy between the constant map and the
= uϕ(x). identity.
dt
Let Ft : Rk → Rk be the flow of this differential
equation, i.e., for each x ∈ Rk , the curve t 7→ Ft (x) The previous example yields the smooth version of a
is the unique solution passing through x. Standard famous fixed point theorem.
theorems on differential equations tell us that:
Theorem 1.30 (Smooth Brouwer Fixed Point Theo-
(1) Ft is defined for all x ∈ Rk and t ∈ R and
rem)
smooth;
Any smooth map f : B k → B k has a fixed point.
(2) F0 is the identity;
(3) Ft+s = Ft ◦ Fs . Proof. Suppose f has no fixed point. We will con-
Clearly for each t, Ft leaves all points outside the struct a map g : B k → S k−1 which restricts to the
unit ball fixed. For appropriate choices of u and identity on S k−1 which contradicts Example 1.35.
t, Ft will map the origin to any point in the open For x ∈ B k , let g(x) ∈ S k−1 be the point where the
unit ball. (I may replace this proof by the one line segment starting at f (x) and passing through
given by Guillemin and Pollack on page 142 of their x hits the boundary. One can write a formula for
book which does not use flows.)  g to show smoothness. 

In what follows suppose that X is compact and with- In fact, any continuous map f : B k → B k has a fixed
out boundary and Y is connected and with the same point (Brouwer fixed point theorem). One can prove
dimension as X. Let f : X → Y be a smooth map. this using the smooth version approximating f by poly-
nomials. Indeed, by the Weierstrass approximation the-
Theorem 1.28 (Degree mod 2) orem, given ε > 0, there is a polynomial P such that
|f (x) − P (x)| < ε for all x ∈ B k . Set Q(x) = P (x)/(1 +
If y and z are regular values of f then
ε). Now Q maps B k into B k and |f (x) − Q(x)| < 2ε
#f −1 (y) = #f −1 (z)(mod2). for all x ∈ B k . Suppose f (x) 6= x for all x. Then
This common residue class is called the degree the continuous function |f (x) − x| must take a positive
mod 2 of f (denoted deg2 (f )) and only depends minimum τ on B k . If now let ε = τ /2 the Q from
on the homotopy class of f . above does not have a fixed point, which contradicts
the smooth Brouwer fixed point theorem.

Proof. Given y and z, let h be the diffeomorphism 1.6.1. Intersection Numbers Modulo 2. What we said
smoothly isotopic to the identity such that h(y) = above about degree modulo 2 can be generalized to the
z given by the Homogeneity Lemma. Observe that case of a smooth map f : X → Y and Z a submanifold
z is also a regular value of h ◦ f . Since h ◦ f is ho- of Y . Suppose that:
motopic to f , the Homotopy Lemma tells us that (1) X is compact without boundary;
#(h ◦ f )−1 (z) = #f −1 (z) (mod 2).
(2) Z is closed and without boundary;
But (h ◦ f )−1 (z) = f −1 h−1 (z) = f −1 (y) and

therefore (3) f t Z

#f −1 (y) = #f −1 (z) (mod 2) (4) dim X + dim Z = dim Y .

as desired. Let g be smoothly homotopic to f . By Under these conditions f −1 (Z) is a closed 0-dimensional
Sard’s theorem, there is a point y ∈ Y which is submanifold of X and hence it consists of finitely many
a regular value for both f and g (why?). By the points. Define the mod 2 intersection number of the
Homotopy Lemma map f with Z, I2 (f, Z), to be the cardinality of f −1 (Z)
deg2 (f ) = #f −1 (y)( mod 2) = #g −1 (y)( mod 2) = degmodulo
2 (g)
2. An analogue of the Homotopy Lemma (for
intersection numbers) also holds here: if f0 and f1 are
which completes the proof.  maps transversal to Z and homotopic, then I2 (f0 , Z) =
I2 (f1 , Z). In fact, we can define I2 (f, Z) for a map f
which is not necessarily transversal to Z. Using that
Example 1.29 transversality is generic, we can find a map g homo-
topic to f such that g t Z and we now set I2 (f, Z) :=
 DIFFERENTIAL GEOMETRY 7

I2 (g, Z). It does not matter which g we choose as long The next theorem tells us that we did not lose much
as g is homotopic to f , thanks to the Homotopy Lemma by restricting out attention to manifolds as subsets of
for intersection numbers. Euclidean space.

Let us have a closer look at the important special case Theorem 1.33 (Whitney’s embedding theorem)
in which X itself is a compact submanifold of Y and A smooth n-manifold X can be embedded into
Z a closed submanifold of complementary dimension. R2n+1 .
In this case f is the inclusion map X ,→ Y , and if
X t Z, I2 (X, Z) := I2 (f, Z) is just #(X ∩ Z) mod 2.
The proof of this theorem can be found in most books
If I2 (X, Z) 6= 0, it means that no matter how X is
about manifolds. In fact, Whitney proved the much
deformed, we cannot move it away completely from Z.
harder result that X can be embedded in R2n .

Example 1.31 2. OLD

1 1 1
Let Y be the 2-torus S × S . Let X = S × 2.1. Orientability and Tangent Planes. Let’s now
{1} and Z = {1} × S 1 . Then I2 (X, Z) = 1 and use the framework we have developed to discuss a geo-
we cannot smoothly pull the circles apart. When metric property of surfaces: orientability.
dim X = 21 dim Y , we can consider I2 (X, X) the
self-intersection number modulo 2. If X is the cen-
Definition 2.1 (Orientation-Preserving Maps)
tral curve in a Möbius band, then I2 (X, X) = 1.
Let V, V 0 be open sets in R2 , and let f : V → V 0 be
a diffeomorphism. We say that f is orientation-
1.7. Abstract Manifolds and Whitney’s Embed- preserving if Df |x has positive determinant for
ding Theorem. One can actually define manifolds with- all x ∈ V .
out making any reference to the ambient space RN . You
will study manifolds in this more abstract setting in
Part III and in the Riemann Surfaces course. In any Definition 2.2 (Orientability)
case, here is the definition: An abstract smooth surface Σ is orientable if it
admits an atlas {(Ui , φi )} where the transition maps
Definition 1.32 are all orientation-preserving. Such an atlas is a
orientation of Σ.
An n-dimensional smooth manifold is a second count-
able Hausdorff space X together with a collection
of maps called charts such that: We will now show that orientability is a ‘topological’
property of surfaces, in that it is preserved by diffeo-
(1) a chart is a homeomorphism φ : U → φ(U ), morphisms.
where U is open in X and φ(U ) is open in Rn
(2) each point x ∈ X belongs to the domain of Theorem 2.3
some chart; If Σ1 and Σ2 are diffeomorphic abstract smooth
n
(3) for charts φ : U → φ(U ) ⊂ R and ψ : V → surfaces, then Σ1 is orientable if and only if Σ2 is
φ(V ) ⊂ Rn , the map φ ◦ ψ −1 : ψ(U ∩ V ) → orientable.
φ(U ∩ V ) is smooth;
Proof. Let f : Σ1 → Σ2 be diffeomorphism. Sup-
(4) the collection of charts is maximal with re-
pose Σ2 is orientable and equipped with an ori-
spect to the properties above.
ented smooth atlas. Consider the atlas on Σ1 of
A set of charts satisfying the first three properties charts of the form (f −1 (U ), φ ◦ f |f −1 (U ) ), where
is called an atlas. An atlas can always be enlarged (U, ψ) is a chart at f (p) in the oriented atlas for
uniquely to give a maximal atlas as in the defini- Σ2 . Then the transition map between two such
tion. If in the definition we require the maps φ◦ψ −1 charts is exactly a transition map between charts
just to be of class C k (k ≥ 1) then we say that we in the Σ2 atlas. 
have a manifold of class C k . (Recall that a map is
of class C k if it has continuous partial derivatives So what does orientability feel like, in a geometric sense?
up to order k.) To discuss this, we are going to need the idea of tangent
planes.

The definition is set up so that it is plain how to define Definition 2.4 (Tangent Plane)
a smooth map between manifolds. An immersion is
a smooth map f : X → Y such that dfx is injective Let Σ be a smooth surface in R3 , and p ∈ Σ. Let
for all x ∈ X. A submersion is a smooth map f : σ : V → U ⊆ Σ be an allowable parameterisation
X → Y such that dfx is surjective for all x ∈ X. An of Σ near o, so V is an open subset of R2 and U
embedding is a smooth map f : X → Y which is an is open in Σ, such that σ(0) = p. The tangent
immersion and a homeomorphism onto its image.
8 ADAM KELLY – MATHEMATICAL TRIPOS PART II

plane Tp Σ to p at Σ is the image of Dσ|0 ⊆ R3 , A smooth surface in R3 is orientable if and only if

which is a two-dimensional vector subspace of R3 . it is two sided.
The affine tangent plane is p + Tp Σ, which is an
affinea subspace of R3 . Proof. Omitted. 
aAn affine subspace of a vector space is a translate of a
linear subspace.
Some illustrations of an orientable and non-orientable
surface are shown below.
p + Tp Σ
No continuous choice of norma
p
Σ
Affine tangent plane

Tp Σ
Torus (Orientable) Möbius Strip (Non-Orientable)

3. Geometry of Surfaces in R3
Tangent plane
3.1. First Fundamental Form. Having developed some
R3
ways to discuss surfaces particularly in R3 , we are go-
We should check that this notion makes geometric sense, ing to start discussing more geometric notions such as
in that it’s independent of the choice of parametrisation length and angles.
used. Recall that for a curve γ : (a, b) → R3 , the length of γ
is given by
Lemma 2.5 Z b
Tp Σ is independent of the choice of allowable param- L(γ) = kγ 0 (t)k dt.
a
terisation.
Since this is independent of paramterisation1, we can
naturally parameterise a curve with respect to arc length
Proof. Suppose σ : V → U and σ̃ : Ṽ → Ũ are (as long as γ 0 (t) 6= 0). Now, given a curve on some
allowable parameterisations with σ(0) = σ̃(0) = smooth surface in R3 , we want to find a nice way to
p. Then there exists a transition map σ −1 ◦ σ̃ compute lengths.
which is a diffeomorphism of open sets in R2 . So
let Σ be a smoth surface in R3 , and let σ : V → U ⊆ Σ
D (σ −1 ◦ σ̃) 0 is an isomorphism, and since σ̃ = σ◦
be an allowable parameterisation. Let γ : (a, b) → R3
(σ −1 ◦ σ̃) the images of Dσ̃|0 and Dσ|0 agree. 
be smooth and have its image be contained in U . Then
writing γ(t) = σ(u(t), v(t)), and so γ 0 (t) = σu u0 (t) +
We can then sensibly define the normal to a smooth σ v 0 (t), giving
v
surface.
kγ 0 (t)k = Eu0 (t)2 + 2F u0 (t)v(t) + Gv 0 (t)2 ,
Definition 2.6 for functions
If Σ is a smooth surface in R3 and p ∈ Σ, the nor-
mal direction to Σ at p is (Tp Σ)⊥ , the Euclidean E = hσu , σu i; F = hσu , σv i = hσv , σu i; G = hσv , σv i,
orthogonal complement to the tangent plane at p. where h·, ·i is our Euclidean inner product. These values
encapsulate the notion of length for our surface.
We can see that for every p ∈ Σ, there exist exactly two
normalized normal vectors, and from this, we finally get Definition 3.1 (First Fundamental Form)
our interpretation of orientability.
The first fundamental form of Σ in the param-
eterisation σ is the expression
Definition 2.7
E du2 + 2F du dv + G dv 2 .
A smooth surface in R3 is two-sided if it admits
a continuous global choice of unit normal vector.
Note that if γ has image in the image of σ, then
Z bp
L(γ) = E(u0 )2 + 2F u0 v 0 + G(v 0 )2 dt.
Lemma 2.8 (Interpretation of Orientability) a

1Feel free to check this

 DIFFERENTIAL GEOMETRY 9

With length encapsulated by the first fundamental form, q

we get a nice way to capture when two surfaces (locally)
have the same ‘geometry’ in a length sense. hn(p), q − pi
Σ
Definition 3.2 (Isometric)
p n(p)
Let Σ, Σ0 be smooth surfaces in R3 . We say that
they are isometric if there exists a diffeomorphism
f : Σ → Σ0 that preserves the length of al curves.
Tp Σ

Let Σ be a smooth surface in R3 , and let σ : V → U ⊆

Definition 3.3 (Locally Isometric)
Σ be an allowable parameterisation. Consider a point
We say that Σ, Σ0 are locally isometric near points p = σ(u, v), and some nearby point q = σ(u + h, v + `).
p ∈ Σ and q ∈ Σ0 if there exists open neighbour- Then by Taylor’s theorem we have
hoods U of p and V of q such that U and V are
isometric. σ(u + h, v + `) = σ(u, v) + hσu (u, v) + `σv (u, v)
1 2
h σu u(u, v) + 2h`σuv (u, v) + `2 σvv (u, v)


+
2
Lemma 3.4 + O(h3 , `3 ).
Smooth surfaces Σ, Σ0 in R3 are locally isometric Recalling that the tangent plane is given by Tp Σ =
near p ∈ Σ and q ∈ Σ0 if and only if there exist span{σu , σv }, we get the perpendicular distance from
allowable parameterisations σ : V → U ⊆ Σ and q = σ(u + h, v + `) to the affine tangent plane Tp Σ + p
σ 0 : V → U 0 ⊆ Σ0 such that the first fundamental is given by
forms are equivalent. 1
hn, σ(u + h, v + `) − σ(u, v)i = hn, σuu ih2 + 2hh, σuv ih` + hn, σvv
2
Proof. Omitted.  + O(h3 , l3 ).
and so we can see that these inner products give us this
difference. This inspires our definition of the second
3.2. Area. With length pinned down by the first fun-
fundamental form.
damental form, it’s not too hard to pin down area.
Definition 3.6 (Second Fundamental Form)
Recall that a parallelogram spanned by two vectors v, w
has area |v × w| = hv, vihw, wi − hv, wi2 . So if we have The second fundamental form of Σ in the al-
an allowable parameterisation σ : V → U ⊆ Σ with lowable parameterisation σ is
σ(0) = p, and consider σu , σv ∈ Tp Σ, the square of L du2 + 2M du dv + N dv 2
the area of the infinitesimal parallelogram spanned by where L = hn, σuu i, M = hn, σuv i, and N = hn, σvv i,
σu , σv is given by ×σv
and n is the unit normal given by n = kσσuu ×σ .
vk

hσu , σu ihσv , σv i − hσu , σv i2 = EG − F 2 . We will say more about how this relates to curvature in
the coming sections. It’s good to know (but not proven
in this course) that the two fundamental forms com-
This gives us a sensible way to define area. pletely determine a smooth oriented connected surface
in R3 , up to rigid motion. This is the fundamental the-
Definition 3.5 (Area) orem of surfaces in R3 .
Let Σ be a smooth surface in R3 , and let σ : V → 3.4. The Gauss Map and Curvature. We are now
U ⊆ Σ be an allowable parameterisation. Then going to define what our notion of curvature actually
Z p
is. This is done using the Gauss map.
area(U ) = EG − F 2 du dv.
V
Definition 3.7 (Gauss Map)
Let Σ be a smooth oriented surface in R3 . The
3.3. Second Fundamental Form. Intuitively, the first Gauss map n : Σ → S 2 is the map p 7→ n(p),
fundamental form pins down how length works along where the normal vector is normalised and hence
our surface. We now, through the second fundamental lies in the unit sphere.
form, will try to pin down how curved our surface is (as
an embedding of R3 ).
Definition 3.8 (Gauss Curvature)
We are (informally) going to look at curvature at a point
as measuring how much we deviate from the tangent
plane by moving a small amount away from our point.
10 ADAM KELLY – MATHEMATICAL TRIPOS PART II

Let Σ be a smooth surface in R3 . The Gauss cur- Informally, Gauss curvature is an infinitesimal measure
vature κ is κ : Σ → R given by of how much the Gauss map n distorts area.
 
κ(p) = det Dn|p
Theorem 3.14 (Theorema Egregium)
The gauss curvature is isometry invariant.
We note that this is smooth since the Gauss map is
smooth2 and is also always well defined even for non-
orientable Σ, as Σ is always locally orientable.
Theorem 3.15 (Gauss-Bonnet)
We can compute κ by using the first and second funda- If Σ is a compact smooth surface in R3 , then
mental forms, but it will require looking at these from Z
a different perspective. κ dAΣ = 2πχ(Σ).
Σ

A Different Perspective on the Fundamental Forms. The

first fundamental form specifies 4. Geodesics
4.1. Definitions. We are now going to discuss a par-
We can compute what κ is exactly. TODO: FILL IN
ticular geometric phenomena, geodesics, which are curves
THE DETAILS. And we arrive at
with (in some sense) are the shortest path between
LN − M 2 points.
κ= .
EG − F 2
Definition 4.1 (Energy)
Definition 3.9 The energy of a curve γ is given by
3
Z b
A smooth surface in R with vanishing gauss cur- 2
E(γ) = kγ 0 (t)k dt.
vature everywhere is called flat. a

Definition 3.10 Definition 4.2 (One-Parameter Variation)

Let Σ be a smooth surface in R3 , and let p ∈ Σ. Let γ : [a, b] → Σ, where Σ is a smooth surface
We say that in R3 . A one-parameter variation (with fixed-
(i) elliptic if κ(p) > 0; endpoints) of γ is a smooth map Γ : (−ε, ε) ×
[a, b] → Σ, such that if γs = γ(s, ·), then γ0 (t) =
(ii) hyperbolic if κ(p) < 0; γ(t) and γs (a) and γs (b) are independent of s.
(iii) parabolic if κ(p) = 0.

Definition 4.3 (Geodesic)

Lemma 3.11 A smooth curve γ : [a, b] → Σ is a geodesic ifa,
In sufficiently small nbhd of elliptic point, surface for every variation (γs ) of γ with fixed endpoints,
d
lies entirely to one side. we have ds s=0
E(γs ) = 0.
aAlternatively, γ is a critical point of the energy functional
on curves from γ(a) to γ(b).
Theorem 3.12
Compact smooth surface has elliptic point. 4.2. The Geodesic Equations. So how do we actu-
ally find geodesics? It turns out to be relatively easy,
Now an interpretation of Gauss curvature. using the geodesic equations.
Suppose we have some curve γ with image contained in
Theorem 3.13 an allowable parameterisation σ. Then for sufficiently
Let Σ be a smooth surface in R3 , and let p ∈ Σ small s, we can write γs (t) = σ(u(s, t), v(s, t)). Now let
such that κ(p) 6= 0. Let U be an open neighbor- the first fundamental form with respect to σ be
hood of p, and a decreasing sequence Ai ⊆ U of E du2 + 2F du dv + G dv 2 .
neighborhoods that ‘shrink to p’ in the sense that Then by definition,
for all ε > 0, Ai ⊆ B(p, ε) for sufficiently large i. Z b
Then, E(γs ) = E u̇2 + 2F u̇v̇ + Gv̇ 2 dt.
areaS 2 (n(Ai )) a
|κ(p)| = lim .
i→∞ areaΣ (Ai ) Differentiating under the integral sign and integrating
by parts, we get
Z b
2Follows by writing down an expression for n(p) and noting d ∂u ∂v
E(γs ) = A +B dt
that it’s the composition of smooth functions. ds s=0 a ∂s ∂s
s=0
 DIFFERENTIAL GEOMETRY 11

where A plane in R3 containing the z-axis is a plane of sym-

metry, hence meridians are geodesics3. Not all parallels
d are geodesics however.
A = Eu u̇2 + 2Fu u̇v̇ + Gu v̇ 2 − 2 (E u̇ + F v̇)
dt
d Lemma 4.8
B = Ev u̇2 + 2Fv u̇v̇ + Gv v̇ 2 − 2 (F u̇ + Gv̇)
dt A parallel given by u = u− 0 is a geodesic when pa-
rameterised at constant speed if and only if f 0 (u0 ) =
These give us the geodesic equations. 0.

Theorem 4.4 (Geodesic Equations) Proof. Omitted. 

A smooth curve γ(t) = σ(u(t), v(t)) is a geodesic if
and only if it satisfies the geodesic equations Consider a curve γ(t) on Σ, making an angle θ with a
d Eu u̇2 + 2Fu u̇v̇ + Gu v̇ 2 parallel of radius ρ = f .
(E u̇ + F v̇) =
dt 2
d Ev u̇2 + 2Fv u̇v̇ + Gv v̇ 2 Theorem 4.9 (Clairaut’s Relation)
(F u̇ + Gv̇) =
dt 2 If γ is a geodesic, then ρ cos θ is constant along γ.

Unlike length, energy is sensitive to reparameterisation, 4.4. Surfaces of Constant Curvature. Our result
but Cauchy-Schwarz gives us that L(γ)2 ≤ E(γ)(b − a), looks like this:
with equality if and only if γ is parameterized propor-
tional to arc length. Proposition 4.10
Let Σ be a smooth surface in R3 . Then,
Corollary 4.5 (i) If κ ≡ 0, then Σ is locally isometric to (R2 , du2 +
If γ has constant speed and locally minimizes length, dv 2 )
then it is a geodesic. Further, if γ minimizes en-
ergy globally, then it minimizes length globally and (ii) If κ ≡ 1, then Σ is locally isometric to (S 2 , du2 +
is parameterized with constant speed. cos2 u dv 2 ).

4.5. Riemannian Metrics.

For some examples, the geodesics in the plane consist of
4.6. The Length Metric. So far, the geometry of our
straight lines, and the geodesics in the sphere consist of
the curves of intersection between the sphere and planessurfaces has been induced by the space that surface has
through the origin. been embed in (which has so far been R3 ). Of course,
it’s possible to do geometry on surfaces that do not arise
In R , the geodesics are not just locally shortest but in this way. One way of specifying the geometry of a
2

they are also locally straightest. We would intuitively surface is with a Riemannian metric.
expect this to hold on other surfaces too. That is, we
would expect the change in the tangent vector to be as Definition 4.11 (Riemannian Metric)
small as possible subject to the constraint that we lie Let V ⊆ R2 be an open set. An (abstract) Rie-
on the surface. mannian metric is a smooth map from V to the
set of positive definite symmetric billinear forms,
Theorem 4.6 given by  
Let Σ be a smooth surface in R3 . A smooth curve E(v) F (v)
v 7→ ,
γ : [a, b] → Σ is a geodesic if and only if γ̈(t) is F (v) G(v)
everywhere normal to the surface Σ. such that E, G, EG − F 2 > 0.

Proof. Omitted.  If v is a vector at p ∈ C, we can compute its infinitesmal

length by
 
2 E(v) F (v)
4.3. Surfaces of Revolution. One common surface kvk = v > v.
F (v) G(v)
we come across is a surface of revolution, where we ro-
tate say η(u) = (f (u), 0, g(u)) where ν is smooth and So, if γ : [a, b] → V is smooth,
Z bp
injective and f (u) > 0 about the z axis.
L(γ) = E u̇2 + 2F u̇v̇ + Gv̇ 2 dt,
a
Definition 4.7 where γ(t) = (u(t), v(t)).
A circle obtained by rotating a point of η is called
a parallel. A curve obtained by rotating η itself
by a fixed angle about the z-axis is a meridian. 3DETAIL HERE!