Additional notes for Geometry 2021

Gergely Bérczi, Cristiano Spotti

Version 8 Feb 2021

2

These extra notes cover some additional material for the 2021 Geometry class that
is not (fully) contained in Do Carmo book. Parts of them are going to be based on
the Supplerende noter til geometri by A. du Plessis and A. Swann used in a previous
version of the class. The notes will be continuously updated.
Chapter 1

Introduction

Objects of study: Curves and surfaces (mostly) in R3 via differential calculus.

Pre-History: Decartes (coordinates), Newton (calculus), and even before (ancient

Greeks).

History: foundational results of Gauss in Disquisitiones Generales Circa Superfi-

cies Curvas, (1828).

Main themes:

• Intrinsic properties (e.g., length, areas, etc...) vs extrinsic properties (e.g., how
curves and surfaces bend in space). Relations between such concepts, e.g., Gauss
Egregium Theorem: some a-priori extrinsic property, the Gauss curvature K, is
actually intrinsic.

• Local properties (e.g., curvatures) vs global properties (e.g., planar simple curve
enclosing the
R biggest area). Relations between such concepts, e.g., Gauss-Bonnet
Theorem: S K = topology(S), S compact surface.

This lead to (after Riemann): abstract manifolds, Riemannian metrics. (Applica-

tions: e.g., Einstein’s theory of General Relativity).

3
4 CHAPTER 1. INTRODUCTION
Chapter 2

Fundamental theorem of spatial

curves

Let γ : I → R3 be a curve in arc-length parametrization (|γ 0 | = 1). To measure its

extrinsic geometry recall that we have defined the following two scalar quantities:

• The curvature k, measuring the variation of tangent lines: denoting the tangent
vector as t := γ 0 , we have t0 =: k n with k ≥ 0 and, for k 6= 0, n(⊥ t) the unit
normal .

• The torsion τ , measuring the variation of the osculating tn-plane (k 6= 0): de-
noting with b := t ∧ n the binormal vector, we have b0 =: τ n.

The vectors (t, n, b) are orthonormal along γ (Frenet trihedron) and satisfy the
following ODE:

0
t = k n

n0 = −k t − τ b

 0
b =τn

Question 2.0.1. Do k and τ completely determin a spatial curve?

The answer is yes, up to the natural ambient isometries (rigid motions):

Theorem 2.0.2. Fundamental Theorem of Spatial Curves. Let I ⊆ R be an

interval and let k : I → R+ and τ : I → R be any two smooth functions. Then there
exists a curve γ : I → R such that its curvature k γ (resp. its torsion τ γ ) is equal to k
(resp. τ ). Moreover, any other curve satisfying the same conditions is of the form

γ̃(s) = Aγ(s) + v,

with A orthogonal matrix and v ∈ R3 , i.e., γ̃ is obtained from γ via a rigid motion.

Remark 2.0.3. Note that we cannot expect more than the above uniqueness up to
rigid motions, since curvature and torsion are invariant under such transformations
(exercise).

5
6 CHAPTER 2. FUNDAMENTAL THEOREM OF SPATIAL CURVES

PROOF IDEAS: the above equation for the Frenet trihedron is a (non-autonomous)
linear ODE for x(s) = (t(s), n(s), b(s)) ∈ R9 , i.e., of the form x0 (s) = B(s)x(s) with
matrix B(s) whose entrances are given by + or − the curvature and torsion (or zeros).
By Picard’s theorem and global existence criterion (see below), we have a solution on
all I of this ODE, that is, we have found the Frenet trihedron of the curve we are
looking for! So we can simply define the curve by ”integrating its speed”:
Z s
γ(s) := t(s̃)ds̃.
s0

It is now a little computation to check that such curve satisfies the desired property.
Moreover, one can argue its uniquenes by observing that any such curves can be moved
to have the same initial point and same initial Frenet trihedron via a rigid motion.
Then uniqueness follows also easily.

Let’s now give a bit more details, starting with a very quick sketchy overview of
local (and global) solution to ODE, that is equations of the form x0 (s) = f (s, x(s))
with f : I × Rn → Rn and initial datum x(0) = x0 , 0 ∈ I ⊆ R (Cauchy problem). The
main local existence theorem is the following:
Theorem 2.0.4. Picard Theorem. Suppose f as above is continuous in [−a, a] ×
B̄x0 (b), where B̄x0 (b) is the closed ball of center x0 and radius b in Rn . Moreover,
assume f is uniformly Lipschitz in the variables x, i.e., ∃L > 0

|f (s, x1 ) − f (s, x2 )| ≤ L|x1 − x2 |,

for all s ∈ [−a, a] and x1 , x2 ∈ B̄x0 (b). Then there is  > 0 and a solution x(s) to the
ODE for s ∈ [−, ].
Proof. By the Fundamental Theorem of Calculus (FTC) our problem is equivalent to
find a fixed point of the following operator Φ:
Z s
x(s) 7→ x0 + f (s̃, x(s̃))ds̃ =: Φ(x(s)).
0

Take  < min{a, b/ max |f |, L−1 }, where max |f | is finite on [−a, a] × B̄x0 (b) (by Weier-
strass theorem since f is continuous), and define the closed set

B := {x(s) ∈ C 0 ([−, ], Rn ) | sup |x(s) − x0 | ≤ b},

in the Banach space of continuous function with the sup norm || − ||∞ . Then Φ sends
B to itself, since
Z s
|Φ(x(s)) − x0 | ≤ |f (s̃, x(s̃))|ds̃ ≤ max |f | ≤ b,
0

and is a contraction, i.e.,

Z s
|Φ(x1 (s))−Φ(x2 (s))| ≤ |f (s̃, x1 (s̃))−f (s̃, x2 (s̃))|ds̃ ≤ L sup |x1 (s)−x2 (s)| < ||x1 −x2 ||∞ .
0
 7

Then by Banach fix point theorem there is a (unique) solution in B (which indeed is
automatically differentiable by FTC).

For global solutions we can state, without proof, this useful criterion:

Proposition 2.0.5. In the above convention, if f is a smooth function such that for
all compact subsets K ⊂ I we have sup{| ∂∂f xi
(s, x)| | s ∈ K, x ∈ Rn } < ∞, then the
solution x(s) to the ODE exists for all s ∈ I.

Going back to our fundamental theorem we immediately see that the hypothesis
of both Picard and the above criterion are satisfied: indeed, being a linear ODE we
can estimate the norm of the derivaties (hence also the Lipschitz constant) in term of
|B(s)| (see also Proposition 3.2.4 and its Corollary), which is bounded on any compact
subset since k and τ are bounded by Weierstrass theorem. Hence, taking as initial
conditions the standard o.n. basis (e1 , e2 , e3 ) we get a solution to the ODE, which
we denote as x(s) = (t(s), n(s), b(s)). This is clearly smooth by taking higher and
higher derivatives of the ODE and using the smoothness hypothesis of k and τ : e.g.,
x00 = (Bx)0 exists since B is smooth and x differentiable.
It is easy to check that (t(s), n(s),
R s b(s)) remains an o.n. basis (exercise). Thus, if we
define the smooth curve γ(s) := s0 t(s̃)ds̃, we have that γ 0 (= tγ ) = t and |γ 0 | = |t| = 1,
i.e., γ is in arc-length parametrization, which in turns implies that, since γ 00 = t0 = kn
with n ⊥ t, |n| = 1, n = nγ and k = k γ . Similarly, since bγ = tγ ∧ nγ = t ∧ n = b,
(bγ )0 = b0 = τ n = τ γ nγ , i.e., τ = τ γ .
Finally, for the uniqueness, we can argue as follow: by a rigid motion we can move
any two curves γ and γ̃ satisfying the desired property to start from the same point,
having there the same Frenet trihedron and satisfying the same Frenet ODE with
solutions x(s) and x̃(s) (hence with x(0) = x̃(0)). Hence

1
0
|x − x̃|2 = (t − t̃).(t0 − t̃0 ) + (n − ñ).(n0 − ñ0 ) + (b − b̃).(b0 − b̃0 ) =
2

= k(t − t̃).(n − ñ) − τ (b − b̃).(n − ñ) − k(t − t̃).(n − ñ) + τ (b − b̃).(n − ñ) = 0.
Since x(0) = x̃(0), this implies x(s) = x̃(s) for all s (this could have also been inferred
using the more general uniqueness of ODE). In particular, γ 0 = t = t̃ = γ̃ 0 , i.e.,
γ = γ̃ + c; but γ(0) = γ̃(0) implies c = 0, so γ = γ̃. This concludes the proof of the
fundamental theorem.
8 CHAPTER 2. FUNDAMENTAL THEOREM OF SPATIAL CURVES
Chapter 3

Review of multivariable
differentiability

3.1 Differentiability
Definition 3.1.1. Let U ⊆ Rn be an open set and x0 ∈ U . A function f : U → Rm
is said to be differentiable at x0 if there exists a linear map L ∈ Hom(Rn , Rm ) such
that:
|f (x) − f (x0 ) − L(x − x0 )|
lim = 0.
x→x0 |x − x0 |
If so, we denote such linear map L as Dfx0 , and we call it the differential of f at x0 .

Heuristically, a function is differentiable at x0 if it can be nicely approximated near

that point by a linear function.

Remark 3.1.2. • Observe we could replace Rn and Rm above with any normed
vector spaces (e.g., with an infinite dimensional Banch spaces) to give a notion
of differentiability in this more general context (Fréchet differentiability).

• It is customary to denote the matrix representation of Dfx0 as (Jfx0 )ij =

ei .Dfx0 ej , with (ei ) std basis, and refer to it as the Jacobian of f at x0 .

• We say that f is differentiable if it is differentiable at any point x ∈ U . If,

moreover, the function x 7→ Jfx0 ∈ Rm×n is continuous, we say that f is of class
C 1 on U .

• It follows directly from the definition that a differentiable function at x0 must

be continuous at x0 .

A couple of easy examples:

• f : Rn → Rm is linear, then Dfx0 = f for all x0 ∈ U , (and it is clearly C 1 ).

• γ : I → Rn a curve. Then, by arguing component-wise, Dγt0 (v) = γ 0 (t0 )v for all

t0 ∈ I and v ∈ R. Indeed, putting L(t − t0 ) := γ 0 (t0 )(t − t0 ) in the definition we
just get the usual definition of derivative!

9
10 CHAPTER 3. REVIEW OF MULTIVARIABLE DIFFERENTIABILITY

Observe that the definition of differentiability can be rewritten by saying that f is

differentiable at x0 if and only if there is a function R : U → Rm which is continuous
at x0 with R(x0 ) = 0, and which satisfies

f (x) − f (x0 ) = L(x − x0 ) + |x − x0 |R(x).

This is useful to prove the next few elementary propositions.

Proposition 3.1.3. The differential at x0 is unique

Proof. Suppose there exists Li such that f (x) − f (x0 ) = Li (x − x0 ) + |x − x0 |Ri (x),
for i = 1, 2. For any v ∈ Rn with |v| = 1, xt = x0 + tv ∈ U if t is small enough. Hence,
letting t → 0:
 
xt − x0
L1 (v) − L2 (v) = (L1 − L2 ) = R2 (xt ) − R1 (xt ) → 0,
|xt − x0 |

i.e., L1 = L2 , that is the differential is unique.

Proposition 3.1.4. A linear combination of differentiable functions is differentiable,

with differential given by the linear combination of the differentials.

Proof. Let h := af + bg with a, b ∈ R and f, g differentiable at x0 . Then:

h(x)−h(x0 )−(aDfx0 +bDgx0 )(x−x0 ) = a|x−x0 |Rf (x)+b|x−x0 |Rg (x) = |x−x0 |(aRf (x)+bRg (x)).

Clearly the function (aRf (x)+bRg (x)) is continuous at x0 and its value there vanishes,
i.e., h is differentiable at x0 with Dhx0 = aDfx0 + bDgx0 .

Proposition 3.1.5. Chain rule. A composition of differentiable functions is differ-

entiable, with differential given by the composition of differentials.

Proof. Let h := g ◦ f with f differentiable at x0 and g differentiable at f (x0 ). Then:

h(x) − h(x0 ) = Dgf (x0 ) (f (x) − f (x0 )) + |f (x) − f (x0 )|Rg (f (x))
= Dgf (x0 ) (Dfx0 (x − x0 ) + |x − x0 |Rf (x))+
+ |(Dfx0 (x − x0 ) + |x − x0 |Rf (x))|Rg (f (x))
= (Dgf (x0 ) ◦ Dfx0 )(x − x0 ) + |x − x0 |R(x),
x−x0
with R(x) := Dgf (x0 ) (Rf (x)) + |Dfx0 ( |x−x 0|
) + Rf (x)|Rg (f (x)) clearly continuous at
x0 and vanishing there, i.e., h is differentiable at x0 with Dhx0 = Dgf (x0 ) ◦ Dfx0 .

Corollary 3.1.6. If f and its inverse are differentiable, then (Df −1 )f (x0 ) = (Dfx0 )−1 .

Proof. I = D(f −1 ◦ f ) = Df −1 ◦ Df , since the differential of the identity is the

identity.

Corollary 3.1.7. Assume f is differentiable at x0 . Then Dfx0 (v) = (f ◦ γ)0 (0), with
γ any curve such that γ(0) = x0 and γ 0 (0) = v, e.g., γ(t) = x0 + tv.
3.2. PARTIAL DERIVATIVES AND DIFFERENTIABILITY 11

Proof. Since Dγ0 (1) = γ 0 (0) = v, we have Dfx0 (v) = Dfx0 (Dγ0 (1)) = D(f ◦ γ)0 (1) =
(f ◦ γ)0 (0).

Remark 3.1.8. This last lemma is going to be ”behind” the definition of tangent
vectors for abstract manifolds.

3.2 Partial derivatives and differentiability

Definition 3.2.1. Consider a function f : U ⊆ Rn → Rm and define the partial
derivative of f w.r.t. xj at p ∈ Rn to be the vector:

∂f f (p1 , . . . , pj + t, . . . , pn ) − f (p)
(p) := lim ∈ Rm ,
∂xj t→0 t

if such limit exist.

∂f
Proposition 3.2.2. If f is differentiable at p, then ∂xj (p) exist for all j = 1, . . . , n,
∂f i
and the Jacobian matrix representing its differential is given by Jfp = ( ∂x j
(p)).

∂f ∂f i
Proof. By the last lemma of previous section Dfp (ej ) = ∂xj (p). Hence ∂xj (p) =
ei .Dfp (ej ) = (Jfp )ij .

Remark 3.2.3. The existence of partial derivatives at a given point are not sufficient
for differentiability! Indeed, take f (x, y) = |xy|. Then ∂f ∂f
p
∂x (0) = ∂y (0) = 0, since
the function is identically zero on the axes. By the above proposition if f would be
differentiable in zero would have zero Jacobian, i.e., zero differential. By definition of
differentiability, this implies that f|p|
(p)
→ 0 as p → 0. However, if we take pt = (ta, tb)
√
|ab|
for a, b 6= 0, and let t → 0, we have that f|p(ptt|) = √a2 +b2 6= 0, giving a contradiction.
We will see soon a criterion on partial derivatives which ensures differentiability.
1
Recall that the Frobenius norm of a matrix A is given by |A| := ij (A2ij ) 2 (this is
P

just the usual norm of A seen as an element of Rm×n ). Note that, by Cauchy-Schwarz
|Av|2 = i (Ai .v)2 ≤ i |Ai |2 |v|2 = |A|2 |v|2 , i.e., |Av| ≤ |A||v|. This is used here:
P P

Proposition 3.2.4. Let f : U → R differentiable in U and, given a, b ∈ U , assume

that the segment [a, b] := {(1 − t)a + tb | t ∈ [0, 1]} is all contained in U (e.g., if U is
convex). Then
|f (b) − f (a)| ≤ sup |Jfp ||b − a|.
p∈[a,b]

Proof. Let γ(t) := (1 − t)a + tb the curve parameterizing the segment, and define
h := f ◦ γ. Then h0 (t) = Jfγ(t) γ 0 (t) = Jfγ(t) (b − a). Hence, by FTC:

|f (b) − f (a)| = |h(1) − h(0)| ≤ sup |h0 (t)| ≤ sup |Jfp ||b − a|.
t∈[0,1] p∈[a,b]
12 CHAPTER 3. REVIEW OF MULTIVARIABLE DIFFERENTIABILITY

Corollary 3.2.5. Let f : U → R be a differentiable function on a convex set U , and

assume supp∈U |Jfp | ≤ M . Then f is Lipschitz with Lipschitz constant bounded by M .
Remark 3.2.6. The hypothesis of the above corollary holds true for any open set V
with compact closure in U if f is required to be C 1 , by Weirstrass’ theorem.
Proposition 3.2.7. Criterion of differentiability. Let f : U → Rm , with U open,
such that all partial derivatives exist for all x ∈ U and are continuous at p ∈ U . Then
f is differentiable in p.
∂fi
Proof. We claim that the linear map L := ( ∂x j
(p)) is indeed the differential of f at
p. Take r small enough such that Bp (r) ⊆ U and, given x ∈ B(p, r), define points
zj := (p1 , . . . , pj , xj+1 , . . . , xn ) ∈ Bp (r), for j = 1, . . . , n − 1, and z0 := x, zn := p.
Hence rewrite:

| nj=1 f (zj−1 ) − f (zj ) − ∂x∂f

P
|f (x) − f (p) − L(x − p)| j
(p)(xj − pj )|
= .
|x − p| |x − p|

Clearly to prove the result would be sufficient to prove that, for each j,
∂f
|f (zj−1 ) − f (zj ) − (p)(xj − pj )| ≤ |xj − pj |Rj (x),
∂xj

for Rj a non-negative continuous at p functions going

P
to zero as x → p: the RHS of
j |xj −pj |Rj (x)
P
the above expression would then be bounded by |x−p| ≤ j Rj (x) → 0.
To see that, consider the segments λj (t) = (p1 , . . . , pj−1 , (1 − t)pj + txj , . . . , xn )
connecting zj−1 to zj and define the function of one variable

∂f
γj (t) := f (λj (t)) − t (p)(xj − pj ).
∂xj
∂f
Since ∂xj exist in U , γj is differentiable in t and
 
∂f ∂f
γj0 (t) = (λj (t)) − (p) (xj − pj ).
∂xj ∂xj

Hence
∂f
|f (zj−1 ) − f (zj ) − (p)(xj − pj )| = |γj (1) − γj (0)| ≤ sup |γj0 (t)|.
∂xj t∈[0,1]
 
∂f ∂f
So to conclude we have to show that Rj (x) := supt∈[0,1] | ∂xj (λj (t)) − ∂xj (p) |→0
∂f
as x → p. But this is clear since, by hypothesis, ∂xj is continuous at p.

3.3 Higher order derivatives.

Given a differentiable function f : U → Rm we can define a map Df : U → Hom(Rn , Rm ) ∼
=
Rm×n simply as Df (p) := Dfp .
3.3. HIGHER ORDER DERIVATIVES. 13

Definition 3.3.1. A function f : U → Rm is twice differentiable at x0 if it exists

an open set V ⊆ U containing x0 such that f is differentiable in V and such that the
function Df is differentiable at x0 .
Clearly D(Df )x0 ∈ Hom(Rn , Hom(Rn , Rm )). We can equivalently think to it as a
bilinear map, called Hessian, D2 fx0 : Rn × Rn → Rm ,

D2 fx0 (v, w) := ((D(Df )x0 )v)w,

v, w ∈ Rn .
Proposition 3.3.2. If f is twice differentiable in x0 then the Hessian is symmetric,
i.e., D2 fx0 (v, w) = D2 fx0 (w, v) for all v, w ∈ Rn .
Remark 3.3.3. In matrix representation, expressing (as we did in the previous sec-
∂f ∂f
tion) the differential map as p 7→ Jfp = ( ∂x j
(p)), where ∂x j
(p) ∈ Rm is a column of
the Jacobian matrix, we get
   
2 ∂ ∂f
D fp (ei , ej ) = (p) ∈ Rm ,
∂xi ∂xj
 
∂ ∂f ∂2f
with ei standard basis. It is costumary to denote ∂x i ∂xj (p) as ∂xi ∂xj (p), and they
are called second (order) partial derivatives. Hence the above proposition equivalently
2f 2f
says that for a twice differentiable function at p we have ∂x∂i ∂x j
(p) = ∂x∂j ∂x i
(p), i.e.,
the second derivatives commutes.
In general the existence of the second partial derivatives is not sufficient for a
function to be twice differentiable: for example, it is easy to check that the function

xy(x2 − y 2 )
f (x, y) = ,
x2 + y 2
∂ f 2 ∂ f 2
extended continuous in 0 by f (0) := 0, satisfies ∂y∂x (0) = −1 and ∂x∂y (0) = 1, hence
by the above proposition the function cannot be twice differentiable in zero. However
we have the following criterion:
Theorem 3.3.4. Schwarz (or Clairaut) Theorem. Let f : U → R be a function
2f
for which the first and second partial derivatives exist and assume ∂x∂i ∂x j
are continuous
∂2f ∂2f
at p for all i, j. Then ∂xi ∂xj (p) = ∂xj ∂xi (p) for all i, j and, moreover, f is twice
differentiable at p.
The proofs of the above theorem and of the symmetry of the Hessian are elementary
but quite involved: they will be both discussed in the exercise session as a group reading
activity.
Finally we can take even more than two derivatives, via an inductive definition
generalizing the notion of twice differentiability:
Definition 3.3.5. For any k > 1 a function f : U → Rm is k-times differentiable at
x0 if it exists an open set V ⊆ U containing x0 such that f is (k − 1)-differentiable in
14 CHAPTER 3. REVIEW OF MULTIVARIABLE DIFFERENTIABILITY

V and such that the function

Dk−1 f : V → Homk−1 (Rn , Rm ),

is differentiable at x0 .

Here Homk−1 (Rn , Rm ) denotes k − 1 multilinear maps (i.e., maps of the form
L(v1 , . . . , vk−1 ) which is linear in each variable) and Dl f is defined inductively as
(Dl f )p (v1 , . . . vl ) = (D(Dl−1 f )p v1 )(v2 , . . . , vl ).

Definition 3.3.6. A function is of class C k if it is k-differentiable (k ≥ 1) and p 7→

Dk f is continuous on U . It is of class C ∞ if it is C k for all k (also said to be smooth).

As before, in matrix representation w.r.t. the standard basis we have

   
∂ ∂f
Dk fx0 (ei1 , . . . , eik ) = ... . . . (x0 ),
∂xi1 ∂xik

∂k f
which we will simply denote as ∂xi1 ...∂xik (x0 ) (k-th order partial derivatives), and there
is no ambiguity in permuting the order of differentiation if f is of class C k (in particular
if it is smooth) by Schwarz theorem.
Chapter 4

Inverse Function Theorem

Theorem 4.0.1. Inverse Function Theorem. Let f : U ⊆ Rn → Rn be a smooth

(resp. C k for k ≥ 1) on an open set U . Suppose that for a point u0 ∈ U , the differential
Dfu0 is invertible. Then there exists a smooth (resp. C k ) inverse function g : V → Ũ
for f|Ũ between open neighborhoods V of f (u0 ) and Ũ of u0 , i.e., f is a local smooth
(resp. C k ) diffeomorphism near u0 ,

Remark 4.0.2. • A trivial example: f (u) = u2 is clearly not invertible globally

(is not-injective). However, Dfu0 = 2u0 6= 0 for u0 6= 0, so away from zero its
differential is invertible, hence it is locally invertible for u 6= 0. Indeed we can take
√ √
g1 (v) = v to be the local smooth inverse of f for u > 0 and g2 (v) = − v for
u < 0. Note that it would be impossible to find a local inverse in a neighborhood
of zero (no matter how small is the neighborhood, the function f is never injective
there).

• In general the above theorem is only an existence theorem for the local inverse
function (i.e., it may be not possible to write down the local inverse function
explicitly!)

• Note that even if the function f (u) = u3 is clearly (globally) invertible (and a
homeomorphism of R), its inverse is not differentiable in zero, that is f is not
a local diffeomorphism near zero (and indeed Df0 = 0 so the implicit function
theorem does not apply there).

• The proof we are presenting is actually based on Banach fixed point theorem and
it is going to be valid if we replace Rn with any Banach space X, even infinite
dimensional.

IDEA OF THE PROOF: Assume u = 0, f (0) = 0 and Df0 is the identity. Thus,
given small v we would like to find a unique u (which we would call g(v)) such that
v = f (u). By rearranging, this is equivalent to look for a (unique) fix point of the
function:
kv (u) := v + u − f (u).
Since Df0 = 1, kv (u) ≈ v + cu2 (this ”quadratic behavior” is imprecise, but just
imagine Taylor’s expansion). For very small u1 , u2 , we expect something like that to

15
16 CHAPTER 4. INVERSE FUNCTION THEOREM

hold: |kv (u1 ) − kv (u2 )| ≈ |c||u21 − u22 | ≤ c|u1 + u2 ||u1 − u2 | < |u1 − u2 |, i.e., kv is a
contraction! So by Banach fixed point theorem we get our unique small u solution. It
is then not too hard to make the above precise (and check that the inverse we obtain
this way is also smooth). Indeed:

Proof. Assume first that u0 = 0, f (0) = 0 and Df0 = 1 and define h(u) := f (u) − u,
noting that h(0) = 0 and Dh0 = 0. Then the set

1
W := {u ∈ U | det(Jfu ) 6= 0, |Jhu | < },
2
is an open neighborhood of zero, since f is smooth (note that the two defining functions
of W are then continuous). Pick an  so small such that B0 (3) ⊂ W . Since here
|Jhu | < 21 , |h(u0 ) − h(u)| ≤ 12 |u − u0 | on its compact subset B0 (2). Hence, since
h(0) = 0, h(B0 (2)) ⊆ B0 ().
Now take any v ∈ B0 () and define a map by kv (u) := v − h(u) (as we did in the
above proof sketch). Such map is a contraction form the complete set B0 (2) to itself.
Indeed |kv (u)| ≤ |v| + |h(u)| ≤ 2, and |kv (u0 ) − kv (u)| = |h(u0 ) − h(u)| ≤ 21 |u − u0 |.
Thus, by Banach fixed point theorem, there exists a unique point uv ∈ B0 (2) such
that kv (uv ) = uv , i.e., such that f (uv ) = v. Note that if v is in the open ball B0 ()
then uv is in the open ball B0 (2), since the above reasoning can be applied for all
0 < . Hence define g(v) := uv and V := B(0, ), and Ũ := g(V ) = f −1 (V ) ∩ B0 (2).
We now claim that f : Ũ → V is a homeomorphism. For this, the only thing that
remains to be proven is that g = f −1 is indeed continuous. Since h is Lipschitz, for any
u, u0 ∈ B0 (2), we have 21 |u0 − u| ≥ |f (u0 ) − f (u) − (u0 − u)| ≥ |u0 − u| − |f (u0 ) − f (u)|,
that is |f (u0 ) − f (u)| ≥ 21 |u0 − u|. But then for v, v 0 ∈ V , |g(v 0 ) − g(v)| = |uv0 − uv | ≤
2|f (uv0 ) − f (uv )| = 2|v 0 − v|, i.e., g = f −1 is Lipschitz, hence, in particular, continuous.
Next we claim g is differentiable with differential equal to Dgv = (Dfuv )−1 . Take
u = g(v) and u0 = g(v 0 ) for v, v 0 ∈ V . By the differentiability of f in u, v 0 − v =
Dfu (u0 − u) + Ru (u0 )|u0 − u|. By applying to both members (Dfu )−1 (which is indeed
invertible for small u), we obtain g(v 0 ) − g(v) = (Dfu )−1 (v 0 − v) + Sv (v 0 )|v 0 − v| for
0
Sv (v 0 ) = − |g(v|v)−g(v)|
0 −v| (Dfu )−1 (Rg(v) (g(v 0 )), which is continuous and tends to zero as
v 0 → v (since g is Lipschitz). Hence g is differentiable for all v ∈ V . Also note that
higher order differentiability is then clear because Dg = (Df )−1 and taking the inverse
is a smooth map: if Df is differentiable then Dg is also differentiable and so on.
Finally, to prove the theorem for a general f , just note that by defining a new
function f˜ := S ◦ f ◦ T , where we T (u) = u + u0 and S(v) = (Dfu0 )−1 (v − v0 ), and
observing that f˜(0) = S(f (u0 )) = (Dfu0 )−1 (v0 −v0 ) = 0, and Df˜0 = DS ◦Dfu0 ◦DT =
(Dfu0 )−1 ◦ Dfu0 ◦ 1 = 1, we obtain a local smooth inverse g̃ for f˜ near 0. Then
g := T ◦ g̃ ◦ S is a local smooth inverse for f near u0 .

4.0.1 Implicitly defined planar curves

Let f : U ⊆ R2 → R be a smooth function, and define the following subset of the plane

Cf := {(x, y) ∈ U | f (x, y) = 0} ⊆ R2 .
 17

For example: f (x, y) = x2 + y 2 − 1.

Proposition 4.0.3. In the above notation, assume Dfp is not zero at a point p =
(x0 , y0 ) ∈ Cf , hence assume w.l.o.g. that ∂f
∂y (p) 6= 0. Then, for a small , there exists
a smooth (parametrized) curve γ : (x0 − , x0 + ) → R2 of the (graph) form γ : x 7→
(x, γ2 (x)) such that γ(x0 ) = p and f (γ(x)) = f (x, γ2 (x)) = 0 for all x ∈ (x0 − , x0 + )
(in particular γ(x) gives a parametrization of Cf near p).

Remark 4.0.4. • This is a special two dimensional case of the so-called implicit
function theorem, whose prove is essentially identical (see exercises). The y
component γ2 of the curve above is called implicit function.

• In the above circle example,

√ points for
√ which y > 0 (resp. < 0), γ2 can explicitly
written as γ2 (x) = 1 − x (resp. − 1 − x2 ) for x ∈ (−1, 1). Note that here we
2

are not covering the points

p (±1, 0) (to do so, we can instead parametrize with a
function of y given by ± 1 − y 2 ).

Proof. Define a function F : (x, y) 7→ (x, f (x, y)) ∈ R2 . Clearly DF(x0 ,y0 ) is invertible,
since detJF(x0 ,y0 ) = ∂f
∂y (p) 6= 0. Hence, by the Implicit Function Theorem, it exists a
local inverse G of F near (x0 , y0 ), with Gi , i = 1, 2 its components. Hence:

(x, y) = F ◦ G(x, y) = F (G1 (x, y), G2 (x, y)) = (G1 (x, y), f (G1 (x, y), G2 (x, y))),

which implies G1 (x, y) = x and f (x, G2 (x, y)) = y for (x, y) near F (x0 , y0 ) = (x0 , 0).
The smooth curve γ : x 7→ (x, G2 (x, 0) =: γ2 (x)) is then the desired one, since
f (x, γ2 (x)) = 0 for x ∈ (x0 − , x0 + ) with  small enough.
18 CHAPTER 4. INVERSE FUNCTION THEOREM
Chapter 5

Global results for curves

In this section we discuss few global theorem for curves. For planar curves we will fully
discuss the proof of the very classical Isoperimetric Inequality, while for spatial curves
we will very briefly discuss (without proofs!) the Fenchel and Fáry-Milnor Theorems.

5.0.1 Green’s Theorem and the isoperimetric inequailty

The following is a very famous beautiful result, which in particular shows that the
circle is the curve shape that encloses the biggest area:

Theorem 5.0.1. Isoperimetric Inequality. Let γ be a simple closed plane curve of

length l. If A denotes the area of the region bounded by γ, then

4πA ≤ l2 ,

and equality holds if and only if γ is a circle.

Remark 5.0.2. • It is clear that the circle (x − x0 )2 + (y − y0 )2 = r2 satisfies the

equality case: indeed its area is πr2 , while its length 2πr.

• The fact that for a simple closed curve there is a well-defined notion of region
bounded by γ is a consequence of Jordan’s theorem (see Topology class). One
can also weakening the hypothesis by requiring the curve to be just piecewise C 1
(that is continuous and made by C 1 arcs).

• This is also known as the (Dido’s problem): in his epic Aeneid the Latin poet
Vergilius (70-19 B.C.) tells the story of Queen Dido. In short, after leaving her
home country, she was granted by a local king of the Libyan cost as much land
as could be enclosed by the hide of a bull. The clever Dido cut the hide into
narrow strips, join them together and then enclosed with it a circle (to become
the town of Carthage), hence ”solving” the problem above of enclosing the most
area given a curve with fixed length.

• This problem is at the origin of the branch of Analysis called Calculus of Vari-
ations, and it was solved rigorously only in the XIX century (J. Steiner 1838).
The proof we discuss in the class is by E. Schmidt in 1938.

19
20 CHAPTER 5. GLOBAL RESULTS FOR CURVES

In the proof of the isoperimetric inequality is used the following other important
result. Here we are assuming a very basic familiarity (Riemann’s integration is fine
for us) with integrals for functions of several variables (e.g., Poulsen, Lecture notes,
Chapter 10).

Theorem 5.0.3. Green’s Theorem Let R ⊆ R2 be a compact region with boundary

a simple closed curve γ. Let f, g be two real valued smooth functions defined in an
open neighborhood of R. Then
Z   Z
∂g ∂f
− dxdy = f dx + gdy,
R ∂x ∂y γ

where the last line integral is defined as

Z Z
f dx + gdy := f (γ(t))γ10 (t) + g(γ(t))γ20 (t)dt,
γ I

where γ : I → R2 has a anticlockwise (i.e., (t, n) gives a positive basis) parametrization

with γi its components.

Remark 5.0.4. • If you would consider a clockwise parametrization, the sign of

the value of the line integral would be simply reversed.

R R
• Green’s theorem is the 2D case of the important Stokes’ Theorem M dα = ∂M α
for abstract manifolds (whose discussion is beyond the content of this class), a
higher dimensional generalization of the Fundamental Theorem of Calculus.

As a corollary of Green’s Theorem, by taking f (x, y) = −y and g(x, y) = x we

immediately obtain that areas can be computed by integrating along the boundary:

Corollary 5.0.5. The area of a region R as above is given by the line integral
Z
1
A(R) = xdy − ydx
2 γ

We start by presenting the proof of Green’s Theorem for a graph type region

R = {(x, y) ∈ R2 | x ∈ [x1 , x2 ], , y ∈ [y1 , h(x)]},

for a smooth function y = h(x), and for functions f, g which are zero for all points of
the boundary of R beside the points on the graph of h:

Ry
Proof. Special case: Define G(x, y) := y1 g(x, u)du. Clearly G(x, y1 ) = 0 and
G(xi , y) = 0 for i = 1, 2. Then:
 21

!
Z   Z x2 Z h(x)  
∂g ∂f ∂g ∂f
− dxdy = − dy dx
R ∂x ∂y x1 y1 ∂x ∂y
Z x2
∂G
= (x, y) − f (x, y)]y=h(x)
[ y=y1 dx
x1 ∂x
Z x2
∂G
= (x, h(x)) − f (x, h(x))dx
x1 ∂x
Z x2
d ∂G
= (G(x, h(x)) − (x, h(x))h0 (x) − f (x, h(x))dx
x1 dx ∂y
Z x2
∂G
x2
= [G(x, h(x)]x1 − (x, h(x))h0 (x) + f (x, h(x))dx
x1 ∂y
Z x2
∂G
=− (x, h(x))h0 (x) + f (x, h(x))dx.
x1 ∂y

On the other end, noting that we are traveling backward along the graph of h since
we have taken the anticlockwise orientation:
Z Z x1
f (x, h(x)) + g(x, h(x)h0 (x) dx


f dx + gdy =
γ x2
Z x2
∂G
=− f (x, h(x)) + (x, h(x))h0 (x)dx,
x1 ∂y

hence proving the theorem in this special case.

We are ready to discuss the general statement, that can be reduced to the above
special case using a so-called partition of unity trick:

Proof. Note that at each point of the boundary γ we can apply the inverse function
theorem so that γ is locally the graph of a function y = h(x) (or x = h̃(y)), e.g., if
γ10 (t) 6= 0, then (x, γ2 (γ1−1 (x)), for x near γ1 (t).
By compactness, we can cover R with finitely many squares Qi in which we can
describe the curve locally as a graph if Qi ∩γ 6= ∅. W.l.o.g we can assume that the balls
of 1/3 of the radius of the inscribed circle to Qi still a cover of R. Next we claim we
can find smooth functions φi : R2 → [0, 1] with compact support in Qi , i.e., such that
the set {x ∈ R2 s.t. φi (x) 6= 0} has compact closure in Qi , and such that i φi (p) = 1
P

for p ∈ R. For x near the center pi of the inscribed ball!Bi of radius ri (pi ) define the
1
(bump) smooth functions ϕi (x) := exp − r2 (p ) , extended smoothly to zero
i i
4
−|x−pi |2
for x such that |x − pi | ≥ 12 ri (pi ). Then just take φi := Pϕiϕi .
i
Now take the cover of R given by Ri := Qi ∩ R, and consider on Ri the function
fi := φi f and gi := φi g. Note that if Qi ∩ γ 6= ∅, then thedata on the  region Ri are
R ∂gi ∂fi
of the type considered before, while if Qi ∩ γ = ∅ then Ri ∂x − ∂y dxdy = 0 since
22 CHAPTER 5. GLOBAL RESULTS FOR CURVES

fi and gi are of compact support on Ri and hence they vanish on all the boundary of
P ∂ P ∂ P
Ri . Thus, noting that i φ = 1 and ∂x i φ = ∂y i φ = 0 on R = ∪Ri , we get:
Z   Z  
∂g ∂f ∂g ∂f X
− dxdy = − φi dxdy
R ∂x ∂y Ri ∂x ∂y
i
X Z  ∂gi ∂fi 
= − dxdy
Ri ∂x ∂y
i
XZ Z
= fi dx + gi dy = f dx + gdy.
i γ∩Ri γ

For the proof of the isoperimetric inequality, which indeed uses the corollary of the
Green’s theorem given above, we simply follow do Carmo 1.7.

5.0.2 Fenchel and Fáry-Milnor theorems

The proofs of the next two global results require some results from the theory of
surfaces in R3 which we will see later on in the class. So here we simply state them to
have at least some interesting examples of how global theorems for spacial curves may
look like.
First few definitions. Let γ be a simple closed spatial curve. We define the total
curvature as: Z l
K(γ) := k(s)ds,
0
with k ≥ 0 the spatial curvature of the curve γ and l its length.
We say that a planar curve is convex if its trace is all contained in one of the two
(closed) half spaces determined by any tangent line.

Theorem 5.0.6. Fenchel Theorem. Let γ be a simple closed spatial curve. Then
K(γ) ≥ 2π with equality if and only if it is a planar convex curve.

Remark 5.0.7. • This theorem was proven in 1929 by the German mathematician
Werner Fenchel, who worked for many years at the University of Copenhagen.

• Note that for a planar curve we could consider R l also the signed curvature k,
1
and then the (possibly negative) quantity 2π 0 k(s)ds. It is not hard to show
1
Rl
(exercise) that 2π 0 k(s)ds is always an integer, and it is actually a theorem of
Hopf that for a simple closed plane curve is value is indeed equal to ±1 (depending
on the orientation).

The proof uses an auxiliary tube surface around the curve. If you are interested,
you can find the proof in Do Carmo 5.7.
It is possible to somehow strength this result as follow: we say that a simple closed
spatial curve is unknotted if we can continuously deform it to the standard closed circle
(cos(t), sin(t), 0) through a one parameter family of simple closed spatial curves. If
this is not the case, we say that γ is knotted.
 23

Theorem 5.0.8. Fáry-Milnor Theorem. For a knotted curve K(γ) ≥ 4π.

Remark 5.0.9. • Equivalently, this theorem says that if the total curvature is
small (< 4π) then the curve is unknotted. However, note that there can be
unknotted curves whose total curvature is bigger than 4π, i.e., the viceversa of
the theorem doesn’t hold.

• This theorem (conjectured by Karl Borsuk) was proved in 1949 independently

by the hungarian mathematician István Fáry and by (at that time 18 years old
undergraduate!) John Milnor.

See Do Carmo 5.7 for its proof, or Milnor’s original paper On the Total Curvature
of Knots, Annals of Mathematics, Vol. 52, No. 2 (1950).