Geometry of Surfaces

2023–24

Dr David Sheard
King’s College London

E-mail: david.1.sheard@kcl.ac.uk

This is Dini’s surface, we will meet it again in remark 5.4.8.

Contents

I Curves 4
1 What is a curve? 4
1.1 Introducing curves 4
1.2 Arc length 6
1.3 Unit speed reparametrisation 7

2 How much does a curve curve? 11

2.1 Curvature 11
2.2 Plane curves 14
2.3 Space curves 21

II Surfaces 26
3 What is a surface? 26
3.1 Concerning maps 26
3.2 Surfaces as sets and maps 28
3.3 Regular surfaces and tangent spaces 32
3.4 Examples of surfaces 34

4 First fundamental form 42

4.1 The definition 42
4.2 Lengths, angles, and areas on a surface 45
4.3 Maps of surfaces 48
4.4 Conformal, isometric, and area-preserving maps 50

5 Curvature of surfaces 56
5.1 The Gauss map 56
5.2 The Weingarten map and the curvatures of a surface 59
5.3 The Weingarten map in local coordinates 63
5.4 Curvature in local coordinates 66
5.5 Geometric interpretation of the Gaussian curvature 71

6 Curvature of curves on surfaces 74

6.1 Normal and geodesic curvature 74
6.2 The principal and normal curvatures 77
6.3 Geodesics 79
6.4 Geodesics as shortest paths 83

–i–
7 Gauss’ Theorema Egregium 88
7.1 The theorem 88
7.2 Applications of the Theorema Egregium 93

8 The Gauss-Bonnet Theorem 95

8.1 Local versions 95
8.2 Global surfaces 102
8.3 The Gauss-Bonnet theorem for global surfaces 104

– ii –
Acknowledgements
This course was introduced by Andrew Pressley many years ago, and taught from his
original notes ever since. Those notes have been turned into a textbook too. The present
set of notes owe a great deal to the originals both in their overall structure and much of
their contents. I also owe much to several other textbooks in preparing these notes, in
particular Do Carmo’s “Differential geometry of curves and surfaces”. My thanks go to
Antonio d’Alfonso del Sordo for his comments and help finding errors.
Every effort has been made to limit mathematical and typographical errors, however
I would be very happy to hear of and correct any mistakes which you may find.
David Sheard, 2024

Notation

|·| absolute value of a scalar

∥·∥ norm of a vector
f′ derivative of a one variable function f with respect to its argument
γ̇ (usually) the derivative of a curve γ with respect to arc length
◦ composition of functions
# cardinality of a set
Aσ (R) area of the surface σ(R)
αi interior angle of a curvilinear polygon
b binormal of a space curve
C (trace of) a curve
D2 topological disk
dAσ area element of a surface
dp f differential of a smooth map f at p
−dp G Weingarten map of an orientable surface
E first coefficient of Ip
e evolute of a curve
e number of edges in a triangulation
ei edge of a curvilinear polygon
F second coefficient of Ip
f number of triangles in a triangulation
Φ reparametrisation map of a surface
ϕ reparametrisation map of a curve
φ angle between the unit tangent of a plane curve and a fixed vector
G third coefficient of Ip
G Gauss map of a(n orientable) surface
Γijk Christoffel symbol
γ (parametrisation of) a curve
γτ smooth variation of a curve γ
H mean curvature of a surface
Ip first fundamental form
IIp second fundamental form
i involute of a curve
int(γ) interior of a curvilinear polygon
K Gaussian curvature of a surface
κ curvature of a curve
κ1,2 principal curvatures of a surface
κg geodesic curvature of a curve in a surface
 κn normal curvature of a curve in a surface
κs signed curvature of a plane curve
L first coefficient of IIp
ℓ total length of a curve
M second coefficient of IIp
N third coefficient of IIp
Np normal to a surface in R3 at p
n normal to a space curve
ns signed normal to a plane curve
P curvilinear polygon
∂P boundary of a curvilinear polygon
R set of real numbers
ρ radius of curvature
Sn n-dimensional sphere
S(τ ) action of a smooth variation γ τ
s(t) arc length of a curve parametrised by t
S2 a surface
Σ2g compact orientable surface of genus g
σ a local parametrisation of a surface
σu partial derivative of σ with respect to u
T triangulation of part of a surface
Ti curvilinear triangle in a triangulation
Tp S 2 tangent space to S 2 at p
t the parameter of a curve
⊺
transpose of a matrix or vector
t tangent vector to a curve
t1,2 principal directions of a surface
τ torsion of a space curve
θi exterior angle of a curvilinear polygon
U an open subset of R2
v number of vertices in a triangulation
vi vertex of a curvilinear polygon
Xn a space which is n dimensional (n = 1 is a curve, n = 2 is a surface)
χ Euler characteristic
Z set of integers

3
 PART
Curves
The first part of this course, despite its name, will focus on curves rather than surfaces.
I
There are two reasons for this. The first reason is that, since curves are one dimensional
and surfaces are two dimensional, the theory of curves is simpler while still allowing us
to explore some of the ideas and techniques which we will use to study surfaces. The
second reason is that the theory of curves can be applied directly to study the theory of
surfaces. In particular, it will turn out to be a very useful technique to study a surface
by studying the curves which lie in that surface.

1 What is a curve?
Our first task will be to decide exactly what we mean by a curve. There are two ap-
proaches. Perhaps the most natural approach would be to consider a set of points in some
bigger space, R2 or R3 say, which together form something suitably one-dimensional.
Asked to give an example of a curve, one might suggest a line, circle, or parabola

y = mx + c y 2 + x2 = r 2 y = x2

and in each of these case we would really be talking about a curve as a set: the set of
points which satisfy a certain equation. In the case of the circle, say, the curve would be
the set
C = {(x, y) ∈ R2 | y 2 + x2 = r2 }.

Such curves were discussed in Calculus 1 when doing implicit differentiation. Another
way of representing curves discussed in that course is using a parameter, for example
the circle C is described by
(
x(t) = cos(t)
; for t ∈ [0, 2π).
y(t) = sin(t)

This is the basis for the less obvious, but it turns out much more useful approach:
to define curves as continuous maps or functions from (part of) R into some bigger space
(so that the image of this map is the set considered by the first approach). One way to
think about this is as of a particle moving through Rn and tracing out the curve as it
goes. The parameter, t in the example above, would then be the time elapsed during
the particles journey.
The principle advantage of this is that we can apply the tools of calculus to these
maps to study the curves by differentiating the function. Time and again, we will find
properties of maps, rather than of sets, will be extremely useful for studying curves.

1.1 Introducing curves

The words ‘function’ and ‘map’ are often used somewhat interchangeably to refer to
some form of correspondence between sets X → Y in which each element of X is sent
to exactly one element of Y . In the fields of geometry and topology, the word ‘map’ has
a clear distinction from ‘function’.
Remark 1.1.1. A map is a continuous function.

4
Introducing curves What is a curve?

The meaning of ‘continuous’ in a general topological context is beyond the scope of

this course. For us, all maps will be between subsets of Rn , and so continuity has its
standard meaning from real analysis1 . 1
Recall a function f : Rm → Rn
We can now give the formal definition of a curve for the purpose of this course. is continuous if for all x0 ∈ Rm ,
limx→x0 f (x) = f (x0 ); or equiva-
lently, if for all ε > 0 there exists
Definition 1.1.2. A curve in Rn is a map γ : (a, b) → Rn where −∞ ≤ a < b ≤ ∞; δ > 0 such that |x − y| < δ implies
ie a, b is an open interval in R which may be finite, infinite, or the whole of R. The |f (x) − f (y)| < ε.
trace of a curve γ is the image {γ(t) | t ∈ (a, b)}.

The set of points idea of a curve is exactly the trace of a curve defined as a function.
Suppose that C ⊂ Rn is a set of points, and there is a curve γ whose trace is C, then we
say that γ parametrises the set C.

Example 1.1.3. We can parametrise the parabola given by y = x2 by

γ : (−∞, ∞) → R2 : t 7→ (t, t2 ).

Indeed, this satisfies γ2 (t) = γ1 (t)2 , and t runs over all possible x values. This is
by no means the only valid parametrisation of the parabola—we could also take
γ(t) = (t3 , t6 ) for t ∈ R, or even γ(t) = (tan(t), sec2 (t) − 1) for t ∈ (−π/2, π/2)
which satisfy the same conditions.
We could not parametrise the parabola by γ(t) = (t2 , t4 ), since while γ2 (t) = γ1 (t)2
does hold true, t2 only takes on non-negative values, so this choice for γ only
parametrises the right half of the parabola, see figure 1.

Example 1.1.4. Now we can try to parametrise the circle given by y 2 + x2 = 1.

√ √
After a bit of rearranging, we get y = 1 − x2 (or y = − 1 − x2 ). Thus we could
take p
γ : (−1, 1) → R2 : t 7→ (t, 1 − t2 ).
Figure 1. The curves
This, however, would NOT be a parametrisation of the circle, since γ only traces parametrised by (t, t2 ) top,
√ and (t2 , t4 ) bottom.
out the upper semicircle. If we took, γ(t) = (t, − 1 − t2 ) we would have much
√
the same problem, and γ(t) = (t, ± 1 − t2 ) is not a function (since it is not single
valued), so can also be ruled out.
Instead, we need to find a function γ(t) = (γ1 (t), γ2 (t)) which satisfies γ1 (t)2 +
γ2 (t)2 = 1 and where γ1 and γ2 both range over [−1, 1] as t varies. There is an
easy solution: take γ1 (t) = cos(t) and γ2 (t) = sin(t) since cos2 (t) + sin2 (t) = 1.
We could take t ∈ (−∞, ∞), however it in fact suffices to take t ∈ (a, b) for any
interval whose length is greater than 2π.
Question 1.1. Why would it not be sufficient to take t ∈ (0, 2π) to parametrise the
circle?

Curves are vector valued functions and we shall be studying them by using the tools
of calculus, so it is well to recall some of the elementary notions and notations from
vector calculus. If we have a curve given by

γ(t) = (γ1 (t), γ2 (t), . . . , γn (t)) ∈ Rn ,

5
Arc length What is a curve?

then we can differentiate it component-wise, assuming each γi (t) is sufficiently smooth:

dγ dγ1 dγ2 dγn

 
γ′ = = , ,...,
dt dt dt dt
d γ
2
d γ1 d γ2 d2 γn
 2 2

γ ′′ = 2 = , , . . . ,
dt dt2 dt2 dt2
etc . . .

We may also sometimes use the notation γ̇, γ̈, etc. . . . In this course we will usually only
require that our curves are at most three times differentiable, ie the third derivative of
each of the components exists. However, so as not to worry too much about precisely
how smooth our curves need to be for this or that result, we will usually just assume
that γ is smooth2 . 2
Recall, this means we can differen-
tiate γ as many times as we like.

Definition 1.1.5. Let γ(t) be a smooth curve, then the first derivative γ ′ (t) is called
the tangent vector to γ at t.

Geometrically, the tangent vector γ(t0 ) points in the direction of the trace of curve
at t = t0 , but its magnitude also encodes information about the parametrisation, in
particular how fast γ(t) moves along the curve at t = t0 . This can be seen by considering
a small change ∆t, as in figure 2. By definition,

γ(t0 + ∆t) − γ(t0 )

γ ′ (t0 ) = lim ,
∆t→0 ∆t

but, γ(t0 + ∆t) − γ(t0 ) is a vector which gets closer and closer to pointing along the Figure 2. The vector γ(t0 +∆t)−
tangent line to γ(t) at t0 as ∆t → 0. γ(t0 ) approaches the tangent line
as ∆t → 0.
The following result is intuitive, and straightforward to prove with a little thought.

Proposition 1.1.6. If the tangent vector to a smooth curve is constant then the
trace of the curve is (part of) a straight line.

Proof. Fix a smooth curve g such that γ ′ (t) = a for all t. Then
Z Z
γ(t) = γ (t)dt =
′
adt = at + b,
Figure 3. Parametrisation of a
straight line.
where b is a constant (vector) of integration. If a = 0 then γ(t) = b is a single
point, which is certainly part of a straight line. Alternatively, if a ̸= 0 then
γ(t) = at + b which is the parametric equation of the straight line through b in
the direction a, see figure 3. ■
Question 1.2. Is the converse true? Does every curve whose trace is (part of) a straight
line have constant tangent vector?

1.2 Arc length

Suppose we want to compute the length of the trace of some curve γ(t); by taking small
increments ∆t in t we can approximate the curve by a sequence of line segments, as in
figure 4.
The length of such a line segment, starting at γ(t0 ) is ∥γ(t0 + ∆t) − γ(t0 )∥, and if
∆t > 0 is small

∥γ(t0 + ∆t) − γ(t0 )∥ Figure 4. Approximating a curve

∥γ(t0 + ∆t) − γ(t0 )∥ = ∆t ≈ ∥γ ′ (t0 )∥∆t.
∆t by line segments.

6
Unit speed reparametrisation What is a curve?

We could sum up all of these lengths and take the limit as ∆t → 0, which turns the sum
into an integral3 . This motivates the following definition. 3
The fact that γ is smooth
(or at least continuously differen-
tiable) is essential, otherwise this
Definition 1.2.1. The arc length of a smooth curve γ(t) starting at γ(t0 ) is the
limit/integral may not exist. Frac-
function Z t tals are examples where this can go
s(t) = ∥γ ′ (τ )∥dτ. wrong, for example, the famous Koch
t0 snowflake is a curve where the length
of any interval along the curve is in-
Since γ is smooth, γ ′ exists and is continuous, an hence ∥γ ′ (τ )∥ is Riemann inte- finite.

grable, and this expression makes sense. Note that s(t0 ) = 0, and in general s(t) is
positive or negative depending on whether t is greater than or less than t0 .
We often think of length as a number, but here it is defined as a function. This
will be essential in the next section, but we can recover the total length of any section
of a curve by evaluating s(t), for example, the total length of a curve γ : (a, b) → Rn is
ℓ(γ) = s(b) − s(a).

Example 1.2.2. Consider the curve γ : R → R2 given by γ(t) = (et cos(t), et sin(t)).
This is called the logarithmic spiral, figure 5. Its tangent vector is

γ ′ = (et (cos(t) − sin(t)), et (sin(t) + cos(t))),

so

2 cos(t)
∥γ ′ ∥2 = e2t (cos2 (t) − ( (((
(( + sin2 (t) + sin2 (t) + 2 sin(t)
sin(t) (((( cos(t)
(( + cos2 (t))

= 2e2t (cos2 (t) + sin2 (t)) = 2e2t .

Fix t0 = 0, then the arc length of γ is given by

Z t Z t√ √
s(t) = ∥γ ′ (τ )∥dτ = 2eτ dτ = 2(et − 1).
0 0

Question 1.3. What is the total length of the logarithmic spiral for t restricted to
(−∞, 0); what about for (0, ∞)?
Remark 1.2.3. As a direct application of the Fundamental Theorem of Calculus,
notice that the derivative of the arc length s(t) of a curve γ(t) starting at t = t0 is
Figure 5. The logarithmic spiral.
ds d t
Z
= ∥γ (τ )∥dτ = ∥γ (t)∥.
′ ′
dt dt t0

This justifies the discussion following definition 1.1.5 where we said that the magni-
tude of the tangent vector measures how fast the point γ(t) moves along the curve
as t varies.
1.3 Unit speed reparametrisation
In example 1.1.3 we saw that a curve—in the sense of a set of points—may admit many
different parametrisations. We can ask how these parametrisations relate to each other,
and whether there is a ‘best’ choice of parametrisation.

Definition 1.3.1. A curve γ e : (ã, b̃) → Rn is a reparametrisation of a curve

γ : (a, b) → R if there is a smooth function ϕ : (a, b) → R (called the reparametri-
n

sation map) such that for all t ∈ (a, b)

dϕ
i) ̸= 0, and
dt
ii) γ
e (ϕ(t)) = γ(t).

7
Unit speed reparametrisation What is a curve?

The second condition ensures that γ e and γ have the same trace, while first condi-
tion implies ϕ is strictly monotonic, and hence invertible. Recall the Inverse Function
Theorem.

Theorem 1.3.2 (Inverse Function Theorem). Let ϕ : (a, b) → R be a smooth function

dϕ
such that ̸= 0, then
dt
a) the image of ϕ is an open interval, say (ã, b̃),

b) ϕ : (a, b) → (ã, b̃) is a bijection, and

c) the inverse ϕ−1 : (ã, b̃) → (a, b) is smooth with derivative

−1
dϕ−1 dϕ

(t̃ ) = (t) ,
dt dt

where t̃ = ϕ(t).

If γ
e is a reparametrisation of γ, then it follows that for all t̃ ∈ (ã, b̃),

γ(ϕ−1 (t̃ )) = γ
e (ϕ(ϕ−1 (t̃ ))) = γ
e (t̃ ),

so γ is a reparametrisation of γ
e with reparametrisation map ϕ−1 ; in other words, being
a reparametrisation is an equivalence relation on the set of curves γ : (a, b) → Rn .

Example 1.3.3. In example 1.1.4, we gave the parametrisation of a circle γ(t) =

(cos(t), sin(t)). Consider also the parametrisation

e (t) = (sin(t), cos(t)).

γ

π dϕ
We will show that γ
e is a reparametrisation of γ. Let ϕ(t) = − t. Then =
2 dt
−1 ̸= 0 and

e (ϕ(t)) = (sin(π/2 − t), cos(π/2 − t)) = (cos(t), sin(t)) = γ(t),

γ

as required.
Now that we can reparametrise curves, we can ask what is the ‘best’ parametrisation?
This is a subjective question which will depend on context and purpose. Nevertheless, a
particularly simple case could be when a parametrisation has constant speed, and if we
choose the speed to be constant, why not make that constant 1.

Definition 1.3.4. A smooth curve γ is a unit speed curve if ∥γ ′ (t)∥ = 1 for all t.

Mathematically, unit speed curves have two particularly useful properties.

Lemma 1.3.5. Let γ be a unit speed curve, with parameter t, then (up to an
additive constant) this parameter is the arc length.

Proof. The arc length of γ is

Z t Z t
s(t) = ∥γ ′ (τ )∥dτ = dτ = t − t0 .
t0 t0

■
For this reason, we will often use the parameter s for unit speed curves (as opposed
to t or any other parameter), and denote the derivative of a curve with respect to s, this

8
Unit speed reparametrisation What is a curve?

special choice of parameter, by a dot, γ̇, rather than a prime γ ′ .

Lemma 1.3.6. Let γ(s) be a unit speed curve, then the vectors γ̇ and γ̈ are per-
pendicular.

Proof. Since ∥γ̇∥ = 1, γ̇ · γ̇ = 1. Differentiating both sides with respect to s gives

γ̈ · γ̇ + γ̇ · γ̈ = 2γ̇ · γ̈ = 0 =⇒ γ̇ · γ̈ = 0,

ie γ̇ and γ̈ are perpendicular. ■

Does every smooth curve have a reparametrisation which has unit speed? The answer
turns out to be no, but the following proposition tells us exactly when it is possible to
find a unit speed reparametrisation.

Proposition 1.3.7. A smooth curve γ(t) has a unit speed reparametrisation if and
only if its tangent vector γ ′ (t) is never 0, ie γ ′ (t) ̸= 0 for all t.

Proof. Suppose first that γ(t) has a unit speed reparametrisation γ e , with
reparametrisation map ϕ. We have γ e (ϕ(t)) = γ(t), and so differentiating with
respect to t and applying the chain rule,

γ dϕ
de dγ de
γ dϕ dγ
= ⇒ = .
dϕ dt dt dϕ dt dt

de
γ
Since γ
e (ϕ(t)) has unit speed, = 1, and so
dϕ

dϕ dγ
= .
dt dt

dϕ dγ
Since ϕ is a reparametrisation map, ̸= 0 and hence and in turn the
dt dt
dγ
tangent vector vanishes nowhere.
dt
dγ
Conversely, assume that the tangent vector ̸= 0 for all t. By remark 1.2.3, the
dt
arc length of γ satisfies
ds dγ
= > 0.
dt dt
Therefore s is a reparametrisation map, and we can define γ
e by setting γ
e (s(t)) =
γ(t). By the chain rule

de
γ de
γ ds de
γ de
γ ds
= ⇒ =
dt ds dt dt ds dt

but also by the definition of γ

e and remark 1.2.3

de
γ dγ ds
= = .
dt dt dt

ds
Equating these and dividing through by ̸= 0,
dt

de
γ
= 1,
ds

as required. ■

9
Unit speed reparametrisation What is a curve?

In view of this proposition, we make the following definition.

Definition 1.3.8. Let γ : (a, b) → Rn be a smooth curve. For t ∈ (a, b), γ(t) is
called a regular point if γ ′ (t) ̸= 0. Then γ is called regular if γ(t) is a regular
point for all t ∈ (a, b).

Question 1.4. Which of the parametrisations given in example 1.1.3 are regular?
We have the following immediate corollary of lemma 1.3.5.

Corollary 1.3.9. Let γ(t) be a regular curve and γ

e a unit speed reparametrisation
of γ with reparametrisation map ϕ. Then if s is the arc length of γ (starting at
any point),
ϕ(t) = ±s(t) + c (1.1)

for some constant c. Conversely, if ϕ(t) is a reparametrisation map given by

eq. (1.1), the corresponding reparametrisation γ
e of γ has unit speed.

Example 1.3.10. Recall from example 1.2.2, that for the logarithmic spiral
parametrised by
γ(t) = (et cos(t), et sin(t))

we computed that ∥γ ′ (t)∥2 = 2e2t . This is non-zero for all t, so γ is regular, and
√
moreover, the arc length starting at t = 0 is s(t) = 2(et − 1). Inverting this, we
find that  
s
t = ln √ + 1 ,
2
and hence a unit speed reparametrisation of γ(t) is

e (s) = γ(t(s))
γ
         
s s s s
= √ + 1 cos ln √ + 1 , √ + 1 sin ln √ + 1 .
2 2 2 2

Example 1.3.11. The twisted cubic is a curve in R3 parametrised by

γ(t) = (t, t2 , t3 ); t ∈ R.

We can compute
p Figure 6. The twisted cubic.
γ ′ (t) = (1, 2t, 3t2 ), ∥γ ′ (t)∥ = 1 + 4t2 + 9t4 ̸= 0,

so γ is regular. The arc length starting at t = 0 is given by

Z tp
s(t) = 1 + 4τ 2 + 9τ 4 dτ,
0

which cannot be expressed in terms of elementary functions (functions such as

exponentials, logarithms, polynomials, etc.), it is an example of an elliptic integral.
By proposition 1.3.7 we know that this curve has a unit speed reparametrisation,
but it is not possible to write it down explicitly in a nice form.
A set of points may be parametrised both by regular and non-regular curves.

Example 1.3.12. The twisted cubic can also be parametrised by γ e (t) = (t3 , t6 , t9 ).
Indeed, both parametrisations satisfy the conditions y = x and z = x3 , so have
2

e ′ (t) = (3t2 , 6t5 , 9t8 ) which is

trace {(x, y, z) ∈ R3 | y = x2 , z = x3 }. However, γ
zero at t = 0, and hence γe (t) is not regular.

10
 How much does a curve curve?

2 How much does a curve curve?

2.1 Curvature
We want to understand how much a curve curves or bends, in other words to quantify the
fact that a line doesn’t curve, a large circle (eg the equator of the Earth) curves but very
slowly, and a small circle (eg a horse on a merry-go-round) curves quickly. Curvature in
this intuitive sense is really a property of the trace of a curve, so we should define it so
that the curvature does not depend on the choice of parametrisation.
How should we quantify this curvature? In proposition 1.1.6 we saw that a curve
with constant tangent vector is a straight line, and so should have zero curvature. More
Figure 7. Image: Tournesol (CC
generally, we might imagine that if a curve curves slowly, then the tangent vector will BY-SA 4.0 DEED)
change only slowly, whereas a quickly changing tangent vector signifies a large curvature.
Thus, the curvature for a curve γ might be related to the rate of change of the tangent
vector γ ′ . The problem with this is that the tangent vector depends on the choice of
parametrisation, and even in the case of a straight line, may change over time unless the
line is parametrised with constant speed (cf question 1.2).
One solution to this is to use the special choice of parametrisation we explored in
the previous section.

Definition 2.1.1. Let γ be a unit speed curve with parameter s, then its curvature Figure 8. Image: Mark (CC BY
at the point γ(s) is given by κ(s) = ∥γ̈(s)∥. 2.0 DEED)

The symbol κ is the Greek letter kappa. Note that κ is exactly the size of the rate
of change of the tangent vector γ̇(s). Since we have fixed the parametrisation to have
unit speed, the magnitude of γ̇ is constant, so this rate of change is just the amount that
the tangent vector is changing direction.
We should check that this definition indeed captures our intuitive idea of curvature
in the cases of lines and circles, as well as that it doesn’t depend on the choice of unit
speed parametrisation.
Example 2.1.2. Suppose we have a line ℓ which passes through the point b and goes
in the direction a. If we assume that ∥a∥ = 1, then γ(s) = as + b is a unit speed
parametrisation (∥γ̇(s)∥ = ∥a∥ = 1), so the curvature is κ(s) = ∥γ̈(s)∥ = ∥0∥ = 0,
as expected.
Example 2.1.3. Consider the circle or radius R centred at (x0 , y0 ), with parametri-
sation  s  s 
γ(s) = x0 + R cos , y0 + R sin .
R R
Then
 s  s  r s s
γ̇(s) = − sin , cos ⇒ ∥γ̇(s)∥ = sin2 + cos2 =1
R R R R

so this is indeed a unit speed parametrisation. The curvature is

1 1
 s  s 
κ(s) = ∥γ̈(s)∥ = − cos , − sin
R R R R
1 1 1
r  s   s 
= cos2 + 2 sin2 = .
R2 R R R R

This matches our intuition since if the circle is large, R is big, and 1/R is small.

Proposition 2.1.4. The definition of curvature does not depend on the choice of
unit speed parametrisation.

11
Curvature How much does a curve curve?

Proof. Recall from corollary 1.3.9 that, for a unit speed curve γ(s), if γ
e (u) is a
unit speed reparametrisation of γ, then u = ±s + c for some constant c. By the
chain rule
de
γ deγ du deγ
= =± = ±γ e˙ (u), (2.1)
ds du ds du
So the curvature of γ
e is

¨ d ˙ (2.1) d de
γ
e(u) = ∥γ
κ e (u)∥ = γ = ± ; using γ
e (u(s)) = γ(s)
du du ds
e

d ds d
= ± γ̇(s) = ± γ̇(s) = ∥γ̈(s)∥ = κ(s)
du du ds

the curvature of γ. ■
This is all well and good if we have a unit speed parametrisation of a curve, but we
have already seen, for example with example 1.3.11, that even relatively simple curves
may not possess unit speed parametrisations which can be easily written down. What
we want is a formula for κ which does not assume that γ is parametrised with unit speed.

Proposition 2.1.5. Let γ(t) be a regular point of a smooth curve in Rn , then the
curvature of γ at γ(t) is given by

∥γ ′′ (γ ′ · γ ′ ) − γ ′ (γ ′ · γ ′′ )∥
κ(t) = .
∥γ ′ ∥4

In the special case that the curve is in R3 , this expression simplifies to

∥γ ′′ × γ ′ ∥
κ(t) = .
∥γ ′ ∥3

Here · and × represent the scalar and vector products respectively.

Proof. Since γ(t) is regular, it possesses a unit speed reparametrisation γ
e (s)
by proposition 1.3.7. We will denote differentiation with respect to t by , and
′

differentiation with respect to s by a dot. By the chain rule

ds γ′
e˙
γ = γ′ e˙ = ′
or γ (2.2)
dt s
and

d ˙ (2.2) d γ ′
 
¨
κ = ∥γ
e∥ = e =
γ
ds ds s′
dt d γ dt γ ′′ s′ − γ ′ s′′ dt 1
 ′  
= = ; using = ′
ds dt s′ ds (s′ )2 ds s
γ ′′ s′ − γ ′ s′′ s′ γ ′′ (s′ )2 − γ ′ s′′ s′
= = . (2.3)
(s )
′ 3 s′ (s′ )4

Next notice that

1
2 2 2
ds 2 ds ds

e˙
γ ′ · γ ′ = ∥γ ′ ∥2 = γ e˙
= γ = = (s′ )2 .

dt dt dt

Differentiating both sides of this with respect to t gives

ds d2 s ds d2 s
γ ′′ · γ ′ + γ ′ · γ ′′ = 2 ⇒ γ ′ · γ ′′ = = s′ s′′ .
dt dt2 dt dt2

12
Curvature How much does a curve curve?

Substituting these into (2.3) yields

∥γ ′′ (γ ′ · γ ′ ) − γ ′ (γ ′ · γ ′′ )∥
κ= ,
∥γ ′ ∥4

as claimed.
In the special case that the curve is in R3 , recall the BAC-CAB property of the
vector triple product:

a × (b × c) = b(a · c) − c(a · b).

Comparing this to the numerator, we see that

γ ′′ (γ ′ · γ ′ ) − γ ′ (γ ′ · γ ′′ ) = γ ′ × (γ ′′ × γ ′ ).

Moreover, γ ′ and γ ′′ × γ ′ are perpendicular, so

∥γ ′ × (γ ′′ × γ ′ )∥ = ∥γ ′ ∥∥γ ′′ × γ ′ ∥,

and hence
∥γ ′ ∥∥γ ′′ × γ ′ ∥ ∥γ ′′ × γ ′ ∥
κ= ′
= .
∥γ ∥ 4 ∥γ ′ ∥3
■
All of this depends on the curve being regular. Clearly, the above expression for the
curvature is not well-defined when γ ′ (t) = 0; however, in the case of a curve which is
not regular, we can restrict its domain to some interval where γ ′ (t) ̸= 0 and then this
formula works fine.

2πh
Example 2.1.6. Consider a circular helix with axis the z-axis, and parametrised by

γ(θ) = (R cos(θ), R sin(θ), hθ)

R
for constants R > 0 and h ∈ R, where θ ∈ R. This curve lies in the cylinder of
radius R centred on the z-axis. Note that in the case h = 0 we recover a circle in
the (x, y)-plane.
We can compute the curvature as follows. Let us begin by computing
p
γ ′ (θ) = (−R sin(θ), R cos(θ), h) ⇒ ∥γ ′ ∥ = R2 + h2 ̸= 0 for all θ.
Figure 9. A circular helix.
This curve does not have unit speed so we cannot use definition 2.1.1, but it is
regular so we can use the formula in proposition 2.1.5.

γ ′′ (θ) = (−R cos(θ), −R sin(θ), 0),

so

γ ′′ × γ ′ = (−Rh sin(θ), Rh cos(θ), −R2 cos2 (θ) − R2 sin2 (θ))

= (−Rh sin(θ), Rh cos(θ), −R2 )

and p
∥γ ′′ × γ ′ ∥ = R h2 + R2 .

Therefore √
∥γ ′′ × γ ′ ∥ R h2 + R2 R
κ= = √ 3 = h2 + R2 .
∥γ ′ ∥3 h +R
2 2

Notice that this curvature is constant. If h = 0, γ(θ) is a circle and we get the

13
Plane curves How much does a curve curve?

correct curvature 1/R. If we take the limit as either R → 0 or h → ±∞, the

curve approaches a straight line (the z-axis or the vertical line through (R, 0, 0)
respectively), and in both the cases, the curvature approaches 0.

2.2 Plane curves

What we have said so far holds for curves in any dimension, but now we will take a
short diversion to focus on curves in the plane, where we can give a slightly more refined
notion of curvature called the signed curvature. Given a smooth curve γ in Rn , at any
point γ(t), there is an (n − 1)-dimensional space of vectors which are perpendicular to
the curve (or more precisely, to γ ′ (t)) at γ(t). For example, in R3 , there is a 2D plane
perpendicular to the curve, see figure 10. When n = 2, this means that there is just a Figure 10. The set of vectors per-
one dimensional space perpendicular to the curve, so up to a choice of sign, there is only pendicular to a curve in R3 form a
plane.
one choice for a unit normal vector to the curve, and we will exploit this fact.

Definition 2.2.1. Let γ : (a, b) → R2 be a unit speed curve in R2 , and denote its
tangent vector by
t = γ̇,

in particular t is the unit tangent vector. Let ns be the unit vector obtained by
rotating t by π/2. Then ns is the signed normal vector to γ, shown in figure 11.
More generally, if γ e ′ /∥e
e is a regular curve, t = γ γ ′ ∥ and ns is again defined by Figure 11. The signed normal
rotating t by π/2. vector to a plane curve.

Recall from lemma 1.3.6 that, for a unit speed curve, γ̇ and γ̈ are always perpendi-
cular—in particular, this means that γ̈ is parallel to ns .

Definition 2.2.2. The signed curvature of a unit speed plane curve γ at γ(s) is
the scalar κs (s) such that γ̈(s) = ṫ = κs ns .
More generally, the signed curvature of a regular curve γ
e at γ
e (t) is defined to be
the signed curvature of any unit speed reparametrisation γ at γ(s(t)).

Unlike the curvature κ, which is always positive by definition, κs may be positive,

zero, or negative. In fact, using ∥ns ∥ = 1 it follows that

κ = ∥γ̈∥ = ∥κs ns ∥ = |κs |.

So the two notions of curvature agree, except that the sign of κs also encodes information
about the direction that γ is curving in, see figure 12.
Figure 12. The sign of κs indi-
We can say precisely the sense in which the signed curvature measures the rate of
cates whether the curve bends to
change of the direction of the tangent vector. the right (κs < 0) or left (κs > 0).

Proposition 2.2.3. Fix a unit vector a ∈ R2 . Let γ(s) be a unit speed plane curve,
and let φ(s) be the angle between a and t = γ̇(s) measured anticlockwise. Then

dφ
κs = .
ds

The definition of φ(s) is illustrated by figure 13. Note, if we change vector a to some
new vector a′ , then this will change the function φ by a constant—the angle between a
and a′ , which goes to 0 when we differentiate by s. Therefore, this formula for κs indeed
does not depend on the choice of a.
Proof. Recall that a and t are unit vectors, so a · t = cos(φ). Differentiating the Figure 13. The definition of φ.
dφ
right hand side with respect to s gives − sin(φ), and doing the same to the left
ds

14
Plane curves How much does a curve curve?

hand side
0
dφ
−
· t + a · ṫ = a · γ̈ = a · (κs ns ) = κs (a · ns ).
sin(φ) = ȧ



ds
On the other hand, ns is obtained from t by rotating it by π/2, so
 π
a · ns = cos φ + = − sin(φ).
2
dφ dφ
Altogether then, − sin(φ) = −κs sin(φ), that is, κs = . ■
ds ds

Corollary 2.2.4. Let γ : R → R2 be a unit speed closed curve, ie there is some

T > 0 such that γ(s + T ) = γ(s) for all s ∈ R. Take T to be the smallest such
value. Then there is some integer w such that
Z T
κs (s)ds = 2πw.
0

The integer w is called the winding number and counts the number of times the curve
circles round before getting back to the start. Winding numbers have several applications
beyond this course, including yielding a topological proof of the Fundamental Theorem
of Algebra4 , and a proof that it is impossible to smoothly deform as circle in the plane 4
For the interested reader, check out
which is oriented anticlockwise into a circle which is oriented clockwise5 . this explanation.
5
Here is a classic video explaining
Proof. By the fundamental theorem of calculus and proposition 2.2.3,
circle and sphere eversion: turning
them inside out.
T T
dφ
Z Z
κs (s)ds = ds = φ(T ) − φ(0).
0 0 ds

Since γ(T ) = γ(0) and γ is smooth, we must have t(T ) = t(0) and hence φ(T ) −
φ(0) must be some integer multiple of a full rotation. ■
It might be tempting to think that this integral is always 0, but the next example
will show that this is not always the case.
Example 2.2.5. Consider the circle of radius R parametrised with unit speed by
 s  s 
γ(s) = R cos , R sin .
R R
This has curvature 1/R by example 2.1.3, and is parametrised going anticlockwise,
so by figure 12, κs > 0. Finally, we know that |κs | = κ, hence κs = 1/R in this
case. Moreover, the period of the curve, T , is 2πR, hence
T 2πR
1 2πR
Z Z
κs (s)ds = ds = = 2π.
0 0 R R

If, instead, we had parametrised the circle in the clockwise direction, κs = −1/R
and the integral would have been −2π. In each case, the winding number is ±1.
Question 2.1. Can you sketch the trace of a smooth closed curve with winding number
0?

The last point we will discuss regarding curves in the plane is the extent to which a
curve is determined by its signed curvature. Let’s begin with an analogy. Suppose you
wanted to navigate your way across London from King’s to Marble arch. If I told you
to start standing outside the Strand Building, facing West along the Strand, and told to
walk forward while bending your steps left or right by such-and-such an amount at each
instant, do you think you would make it to your destination (assuming you were able to

15
Plane curves How much does a curve curve?

follow my somewhat eccentric instructions)? You probably would. Not only would your
final destination be assured, but these instructions would also exactly determine your
path to get there.
In the case of a plane curve, this is analogous to fixing some starting point γ(s0 ),
some starting direction γ̇(s0 ), and the signed curvature κs (s) as a function of s. If the
example of walking across London is to be believed, we might expect this information
to uniquely determine the plane curve—and this turns out to be true.

Theorem 2.2.6 (Fundamental Theorem of Plane Curves). Let k : (a, b) → R be a

smooth function, x, v ∈ R2 be vectors, with ∥v∥ = 1, and s0 ∈ (a, b). Then there
is a unique unit speed curve γ : (a, b) → R2 whose signed curvature is k such that
γ(s0 ) = x and γ̇(s0 ) = v.

Remark 2.2.7. An equivalent formulation of this Theorem is that if k is any smooth

function then there exists a curve γ with signed curvature k, and if γ e is another
unit speed curve with signed curvature k, then there is a rigid motion of R2 (ie a
translation combined with a rotation) which sends γ(s) to γe (s).
Proof. To prove existence, inspired by proposition 2.2.3, define
Z s
φ(s) = k(u)du + φ0
s0

where φ0 is the angle between (1, 0) and v, measured anticlockwise, so that v =

(cos(φ0 ), sin(φ0 )). Then set
Z s Z s 
γ(s) = cos(φ(t))dt, sin(φ(t))dt + x.
s0 s0

By definition, γ(s0 ) = x and

γ̇(s) = (cos(φ(s)), sin(φ(s))).

This is the unit vector making angle φ(s) with (1, 0) so γ has unit speed, and
moreover
γ̇(s0 ) = (cos(φ0 ), sin(φ0 )) = v.

Finally, by proposition 2.2.3,

dφ(s) d s
Z 
κs (s) = = k(u)du + φ0 = k(s),
ds ds s0

as required.
To prove uniqueness, assume that γ e (s) is another unit speed smooth curve with
signed curvature κ es (s) = k(s), and which satisfies γ e˙ (s0 ) = v. Since
e (s0 ) = x and γ
∥γ˙
e (s)∥ = 1 there is some function φ(s)
e such that

e˙ (s) = (cos(φ(s)),
γ e sin(φ(s))),
e

dφ(s) dφ(s)
and again proposition 2.2.3 implies that = k(s) = . By the funda-
e
ds ds
mental theorem of calculus, there is some constant such that φ(s)
e = φ(s) + c, but
since

(cos(φ(s0 ) + c), sin(φ(s0 ) + c)) = (cos(φ(s

e 0 )), sin(φ(s
e 0 ))) = v
= (cos(φ(s0 )), sin(φ(s0 )))

16
Plane curves How much does a curve curve?

it must be that c = 2nπ for some n ∈ Z. In particular, γ e˙ (s) = γ̇(s).

Finally
d
(e e˙ (s) − γ̇(s) = 0
γ (s) − γ(s)) = γ
ds
implies that γ
e (s) = γ(s)+c for some constant c, but x = γ
e (s0 ) = γ(ss )+c = x+c,
so c = 0 and we are done. ■
Any smooth function can be the signed curvature of some plane curve, but even
simple functions can lead to complicated curves, as this next example shows.
Example 2.2.8. Let us suppose that κs (s) = s for all s ∈ R and try to find a curve
with this curvature. The existence part of the proof of theorem 2.2.6 is constructive
in the sense that it tells us how to find the curve γ given κs , so let’s follow that
recipe. Fix s0 = 0, v = (1, 0) so φ0 = 0, and x = 0; then set
s s
s2
Z Z
φ(s) = κs (u)du = udu = ,
0 0 2

and define s s
t2 t2
Z   Z   
γ(s) = cos dt, sin dt .
s0 2 s0 2
Similar to the arc length of the twisted cubic we computed in example 1.3.11, these
integrals cannot be evaluated in terms of elementary functions6 . We can however 6
They are called the Fresnel inte-
numerically approximate them, and hence plot the curve, which is called Cornu’s grals.

spiral, or sometimes Euler’s spiral, see figure 14.

Figure 14. Cornu’s spiral.

In a similar vein to proposition 1.1.6, we can now use the fundamental theorem of
plane curves to give a characterisation of of circles.

Corollary 2.2.9. Let γ : (a, b) → R2 be a regular curve with constant non-zero

curvature. Then γ parametrises (part of) a circle.

The following lemma will help us prove this.

Lemma 2.2.10. Let γ : (a, b) → R2 be a unit speed curve, then ṅs = −κs t.

17
Plane curves How much does a curve curve?

Proof. The pair {t, ns } form an orthonormal basis of R2 , so there are scalars λ
and µ such that ṅs = λt + µns . Since ns is a unit vector, ṅs · ns = 0, so µ = 0.
We have that ṅs · t = λ; since ns and t are orthogonal, ns · t = 0. Differentiating
with respect to s, we get

0 = ṅs · t + ns · ṫ = λ + ns · κs ns = λ + κs ⇒ λ = −κs ,

as required. ■
Proof. (of corollary 2.2.9) Possibly after a reparametrisation, we can assume with-
out loss of generality that γ = γ(s) is a unit speed parametrisation. Since the
signed curvature is a smooth function, and hence continuous, it follows from the
fact that κ = |κs | that either κs = κ or κs = −κ for all s; in either case it is
constant. Consider the quantity γ + κ1s ns , taking the derivative with respect to s
we get
d 1 1 1
 
γ + ns = t + ṅs = t + (−κs t) = t − t = 0,
ds κs κs κs
here we have used lemma 2.2.10. Hence a = γ + 1
κs n s is a constant vector, and

1 1
∥γ − a∥ = − ns = .
κs κ

This shows that the trace of γ is contained the circle of radius 1/κ centred at a. ■
In this proof, we have seen that a curve with constant curvature κ is a circle with
radius 1/κ. More generally, if the curvature of a regular curve at a point is κ, then near
that point, the curve will look like (part of) a circle of radius 1/κ, see figure 15. This
motivates the following definition.

Definition 2.2.11. Let γ : (a, b) → R2 be a regular curve. Then the radius of

1 Figure 15. A curve looks locally
curvature of γ at γ(t) is ρ(t) = . The centre of curvature of γ at γ(t) is like part of a circle near γ(s) of
κ(t)
1 radius 1/κ.
the point γ(t) + ns (t).
κs (t)

1
Note that the sign of κs always ensures that the vector ns points to the correct
κs (t)
side of the curve. We can use this to define a new curve given any regular curve.

Definition 2.2.12. Let γ : (a, b) → R2 be a regular curve, the evolute of γ is a

curve e : (a, b) → R2 traced out by the centre of curvature of γ, ie

1
e(t) = γ(t) + ns (t).
κs (t)

Example 2.2.13. Consider the ellipse with parametrisation

γ(t) = (a cos(t), b sin(t))

Figure 16. The ellipse for a = 3
and b = 2 together with its evo-
for a, b > 0 constants, and let’s compute its evolute. Since γ does not have unit lute.
speed, we need to compute

γ′ (−a sin(t), b cos(t))

t= =p .
∥γ ′ ∥ a2 sin2 (t) + b2 cos2 (t)

18
Plane curves How much does a curve curve?

Then rotating this by π/2,

(−b cos(t), −a sin(t))

ns = p .
a2 sin2 (t) + b2 cos2 (t)

Finally, using the formula derived on the exercise sheet for week 2, the signed
curvature turns out to be
ab
κs (t) = .
(a2 sin (t) + b2 cos2 (t))3/2
2

Putting this all together, the evolute is

1
e(t) = γ(t) + ns (t)
κs (t)
*1
(a2 sin2 (t) + b2 cos2 (t))
3/2
(−b cos(t), −a sin(t))

= (a cos(t), b sin(t)) + ((( (
ab p
a2 sin2 (t)
(( +(
b2 cos2 (t)
((((
1
= (a2 b cos(t) − a2 b sin2 (t) cos(t) − b3 cos3 (t),
ab
ab2 sin(t) − a3 sin3 (t) − ab2 sin(t) cos2 (t))
1
= (a2 b cos(t)(1 − sin2 (t)) − b3 cos3 (t),
ab
ab2 sin(t)(1 − cos2 (t)) − a3 sin3 (t))
1
= ((a2 b − b3 ) cos3 (t), (ab2 − a3 ) sin3 (t))
ab
a2 − b2 b2 − a2
 
= cos3 (t), sin3 (t) ,
a b

which is an example of a Lamé curve, and is shown in figure 16. Notice that while
γ is a regular curve, its evolute is not even smooth. It has four cusps 7 , ie points 7
This is not special to the ellipse,
where the curve is not smooth. it happens for all closed curves, and
is a consequence of the so-called Four
Another way to construct a new curve from an old one, which turns out to be closely Vertex Theorem.
related to the evolute, is called the involute.

Definition 2.2.14. Let γ : (a, b) → R2 be a regular curve, and fix t0 ∈ (a, b). Then
the involute of γ at γ(t0 ) is the curve i : (a, b) → R2 parametrised by

i(t) = γ(t) − s(t)t(t),

where s(t) is the arc length of γ starting at t0 .

Geometrically, one can think of the involute as follows. Take a taught piece of string
laid along the curve so that the end of the string is at γ(t0 ). The involute is the curve
traced out by this endpoint as the string is pulled away from the curve, see figure 17.
The length of the string pulled away from γ when you get as far as γ(t) is the length
of the curve between γ(t0 ) and γ(t), ie s(t). This string meets the curve at a tangent, so
it is parallel to t, but this vector points in the opposite direction, so we pick up a minus
sign. Therefore, the end of the string, which traces out the involute, is at the position
γ(t) − s(t)t(t).
Example 2.2.15. Consider the circle γ(s) = (cos(s), sin(s)), let’s find its involute.
The unit tangent vector is

t = γ̇ = (− sin(s), cos(s))

19
Plane curves How much does a curve curve?

Figure 17. Unwrapping a piece of string from a curve to trace out the involute.

Then the involute of γ at γ(0) is

i(s) = γ(s) − st(s) = (cos(s) + s sin(s), sin(s) − s cos(s)).

This curve is shown in figure 18.

Figure 18. The involute (orange) of the unit circle (blue). The involute is traced out by the
end of a piece of string (multicoloured) as it is unwound from the circle.

The connection between the evolute and involute of a curve is given by the following
theorem, which we will not prove.

Theorem 2.2.16. Let γ : (a, b) → R2 be a plane curve with involute i(t) at γ(t0 )
for some t0 ∈ (a, b). Then the evolute of i(t) is γ(t).

Thus the evolute of an involute is the original curve.

Question 2.2. What do you think happens if instead you take the involute of the
evolute? Do different choices of involute change the curve that you end up with?
Example 2.2.17. Continuing the example of the circle γ(s) = (cos(s), sin(s)) with
involute
i(s) = γ(s) − st(s) = (cos(s) + s sin(s), sin(s) − s cos(s)),

20
Space curves How much does a curve curve?

we can compute the evolute of this involute. Note that

d
i(s) = (− sin(s) + sin(s) + s cos(s), cos(s) − cos(s) + s sin(s))
ds
= (s cos(s), s sin(s)),

has norm equal to s, so even though the circle we started with had unit speed, its
involute does not. Nevertheless, the unit tangent is

d d

ti := i(s) i(s) = (cos(s), sin(s))
ds ds

and so the signed normal of the involute is nsi = (− sin(s), cos(s)). Finally, to com-
pute the signed curvature of i, κsi , we can notice that since i curves anticlockwise,
as shown in figure 18, it must have positive signed curvature (compare with fig-
ure 12), ie the signed curvature equals the curvature. We can then use the formula
in proposition 2.1.5, which eventually gives

1
κsi = .
s
Therefore, the evolute of i(s) is

1
ei (s) = i(s) + nsi
κsi
= (cos(s) + s sin(s),
  sin(s) − s cos(s))

 + s(−sin(s),
 
 cos(s))

= (cos(s), sin(s)),

as expected.

2.3 Space curves

Having looked at curves in the plane, we now turn our attention to curves in 3-space.
We saw that in two dimensions, the (signed) curvature is enough to entirely determine
the shape of a curve. In three dimensions this is no longer true. In fact, we have already
seen this since both the circle (example 2.1.3) and the circular helix (example 2.1.6) can
be viewed as curves in R3 , and both have constant curvature. Therefore, curvature alone
is not enough to completely describe a curve. In this section, we will introduce a second
type of curvature called the torsion which, together with the curvature κ, is enough to
fix the shape of a curve in three dimensions.
We will extend the notation of definition 2.2.1 to R3 , and write t = γ̇ for the unit
tangent vector of a regular curve γ : (a, b) → R3 .

Definition 2.3.1. Let γ : (a, b) → Rn be a unit speed curve with non-zero curvature
κ(s) at γ(s). The principal normal of γ at γ(s) is the vector

1
n(s) = ṫ(s).
κ(s)

Recall that, by definition the curvature is κ = ∥ṫ∥, so n is a unit vector. By

lemma 1.3.6, t · ṫ = 0, so t and n are perpendicular unit vectors. We can therefore
complete {t, n} to an orthonormal basis for R3 .

Definition 2.3.2. Again, let γ have unit speed with non-zero curvature at γ(s).
Then the binormal vector to γ at γ(s) is

b(s) := t(s) × n(s).

21
Space curves How much does a curve curve?

Now {t, n, b} is a right-handed orthonormal basis for R3 wherever the curvature of γ

is non-vanishing, see figure 19. In particular, the following are immediate consequences
of the definition
b = t × n, n = b × t, t = n × b.

Proposition 2.3.3. Let γ : (a, b) → R3 be a unit speed curve with non-zero curva-
ture κ for all s ∈ (a, b). Then there exists a function τ : (a, b) → R such that

ḃ(s) = −τ (s)n(s). Figure 19. The tangent, normal,

and binormal form an orthonor-
Proof. Since b is a unit vector, the same argument as in lemma 1.3.6 shows that mal basis of R3 at every point
b and ḃ are perpendicular. Also, by the definition of b, along the curve.

*0
ḃ =  + t × ṅ = t × ṅ

ṫ ×
n

since
ṫ × n = κn × n = 0.

But if ḃ = t × ṅ, then ḃ must be perpendicular to t. Any vector in R3 which is

perpendicular to b and t must be parallel to n, which proves the existence of τ . ■

Definition 2.3.4. Let γ : (a, b) → R3 be a unit speed curve with non-zero curvature
κ for all s ∈ (a, b), then the function τ in proposition 2.3.3 is called the torsion of
γ.

Lemma 2.3.5. The torsion does not depend on the choice of unit speed parametri-
sation for γ.

Proof. This is left as an exercise. ■

Just as for curvature, we can extend the definition of torsion to include all regular
curves by defining the torsion of a regular curve γ at the point γ(t) to be the torsion of
a unit speed reparametrisation γe of γ at γ
e (s(t)), so long a the curvature at this point is
non-zero. As with proposition 2.1.5, it would be useful to have a formula for the torsion
of a curve which does not require us first to find a unit speed reparametrisation.

Proposition 2.3.6. Let γ : (a, b) → R3 be a regular curve with non-vanishing cur-

vature at γ(t) for all t ∈ (a, b). Then the torsion is given by

(γ ′ × γ ′′ ) · γ ′′′
τ= .
∥γ ′ × γ ′′ ∥2

We omit the proof of this formula which is a somewhat tedious computation.

Example 2.3.7. Let’s return to the example of the circular helix (see example 2.1.6)
parametrised by γ(θ) = (R cos(θ), R sin(θ), hθ), for R > 0 and h constants. We saw
that the curvature is given by R/(h2 +R2 ) ̸= 0 so the torsion is defined everywhere.
We also saw that
p
γ ′ (θ) = (−R sin(θ), R cos(θ), h) ⇒ ∥γ ′ ∥ = h2 + R2 .
√
Since this is constant, the arc length is s(θ) = h2 + R2 θ (up to an additive
√
constant). Setting a = 1/ h2 + R2 , a unit speed reparametrisation is

e (s) = (R cos (as) , R sin (as) , ahs) .

γ

22
Space curves How much does a curve curve?

Thus t̃ = (−aR sin(as), aR cos(as), ah) and ˜ṫ = (−a2 R cos(as), −a2 R sin(as), 0),
so κ e ∥ = ∥˜ṫ∥ = a2 R. By definition,
¨
e = ∥γ

1˜
e=
n ṫ = (− cos(as), − sin(as), 0),
κ
e
and
e = t̃ × n
b e = (ah sin(as), −ah cos(as), aR).

Finally,
ė = (a2 h cos(as), a2 h sin(as), 0) = −a2 he
b n,
h
so τ (s) = a2 h = .
h2 + R2
Recall that the case that h = 0 corresponds to a circle, which lies in a plane, and in
this case, the torsion is 0. This is not a coincidence.

Proposition 2.3.8. Let γ : (a, b) → R3 be a regular curve with nowhere vanishing

curvature. Then the trace of γ lies in a plane if and only if the torsion is identically
zero.

Proof. Without loss of generality, we can assume that γ has unit speed. Suppose
that τ (s) = 0 everywhere. By the definition of torsion, ḃ = 0, so b is a constant
vector. Then
d
(γ · b) = γ̇ · b +  ḃ = t · b = 0,
γ ·

ds
so γ · b is a constant, say d. This means that the trace of γ lies in the plane
r · b = d.
Conversely, assume that the trace of γ lies in the plane r · a = d for some constants
a and d. We many assume ∥a∥ = 1, then differentiating the equation γ · a = d
with respect to s gives

d
0= (γ · a) = γ̇ · a +  ȧ = t · a
γ ·

ds
and again
d
0= (t · a) = ṫ · a +  ȧ = κn · a
t · ⇒ 0 = n · a.
ds
Since both t and n are both perpendicular to a and both perpendicular to b (by
definition) it follows that a and b are parallel, see figure 20. Since both of these
are unit vectors, and b is a continuous function of s (γ is smooth), either b(s) = a
or b(s) = −a for all s. In either case, b is constant, hence ḃ = 0 and τ = 0 by
definition. ■
Since {t, n, b} form a basis, any vector in R3 can be written in terms of these vectors.
Between the definitions of curvature and torsion, we have expressions for ṫ and ḃ in terms
of this basis. We may ask how to write ṅ in terms of this basis. This turns out to be Figure 20. A curve lying in the
straightforward to compute, but taken together, these expressions are very useful and plane orthogonal to a.
go by the name the Frenet-Serret equations.

Theorem 2.3.9 (Frenet-Serret Equations). Let γ : (a, b) → R3 be a unit speed curve

with nowhere vanishing curvature. Then

ṫ = κn


ṅ = −κt + τb

ḃ =

− τn

23
Space curves How much does a curve curve?

Proof. The first equation follows from the definition of the principal normal,
definition 2.3.1; while the third is the definition of torsion, see proposition 2.3.3.
We are left to prove the second equation.
Since {t, n, b} form a basis, there are scalars λ, µ, and ν such that

ṅ = λt + µn + νb.

Taking the dot product with t and using orthonormality, we have

1 0 0
ṅ · t = λ  t + µ
t ·
*
 n t + ν
·
: b t = λ;
·
:

similarly, ṅ · n = µ and ṅ · b = ν. Since n is a unit vector, ṅ · n = 0, so µ = 0.

Perpendicularity implies t · n = 0, and differentiating both sides gives

0 = ṫ · n + t · ṅ = κ
n :+1 λ = κ + λ
·n
 ⇒ λ = −κ.

Similarly, differentiating b · n = 0 gives

0 = ḃ · n + b · ṅ = −τ 
n :+1 ν = −τ + ν
·n
 ⇒ ν = τ.

Therefore, ṅ = −κt + τ b. ■
Compare this to lemma 2.2.10. As a simple application of the Frenet-Serret equa-
tions, we can generalise corollary 2.2.9 to characterise all space curves which parametrise
(parts of) circles.
Figure 21. A curve lying in
the intersection of a sphere and a
Corollary 2.3.10. Let γ : (a, b) → R3 be a unit speed curve with constant non-zero plane.
curvature and zero torsion. Then γ parametrises (part of) a circle.

Question 2.3. Is the converse of this true?

Proof. By proposition 2.3.8 and its proof, since the torsion vanishes, γ lies in the
plane orthogonal to b, which is a constant vector. Consider the quantity8 γ + κ1 n. 8
Compare this to the definition of
Taking the derivative with respect to s, we get the evolute for plane curves, defini-
tion 2.2.12.
d 1 1 1 0
 
γ + n = t + ṅ = t + (−κt +  τ b) = t − t = 0,
ds
>
κ κ κ

where we have used the fact that κ is a constant and the second Frenet-Serret
equation. Hence a = γ + κ1 n is a constant vector, and

1 1
∥γ − a∥ = − n = .
κ κ

This shows that the trace of γ is contained in the sphere of radius 1/κ centred at
a. The intersection of a plane with a sphere is a circle. ■
Question 2.4. The last line of this proof is slightly incomplete, since, in general, the
intersection of a plane with a sphere could be either empty, a single point, or a circle.
In fact, the plane and circle intersect in an equator of the circle, why is that?

As with plane curves, it is possible to prove, using the Frenet-Serret equations, that
the shape of a space curve is determined exactly by its curvature and torsion. This proof
is rather involved, and uses results about the existence and uniqueness of solutions to
ordinary differential equations. For that reason we will omit it here.

24
Space curves How much does a curve curve?

Theorem 2.3.11 (Fundamental Theorem of Space Curves). Let κ : (a, b) → (0, ∞)

and τ : (a, b) → R be smooth functions, x, v ∈ R3 be vectors, with ∥v∥ = 1, and
s0 ∈ (a, b). Then there is a unique unit speed curve γ : (a, b) → R3 whose curvature
is κ, torsion is τ , and which has γ(s0 ) = x and γ̇(s0 ) = v.

25
 PART
Surfaces
In part I, we have developed some of the theory of curves in Rn using the tools of calculus.
II
In particular, we have been studying the smooth image of maps R ⊃ (a, b) → Rn ,
especially for n = 2 or 3.
In this part of the course, we will study mathematical objects which are one di-
mension higher than curves—surfaces. These will be the images of certain maps R2 ⊃
U → Rn . Again, we wish to use calculus so these maps will need to be sufficiently
differentiable. We will also typically restrict ourselves to R3 .

3 What is a surface?
You are already familiar with some surfaces, for example, planes, spheres, and cones.

ax + by + cz = d x2 + y 2 + z 2 = r 2 x2 + y 2 = z 2

When studying surfaces, there are several different ways to study them (topological,
differentiable, algebraic), and sometimes these lead to slightly different definitions which
may sometimes be contradictory. In this course, we will meet topological and smooth
surfaces. We will also meet local parametrisations, which may be smooth or regular.
The collection of objects described by smooth surfaces turns out to be subtly different
from the collection described by local parametrisations. It will be important to have a
clear understanding of the differences between all of these, and how they relate to each
other.
As in the case of curves, we can think of surfaces either as sets of points which are
‘2–dimensional’ in some sense, or as functions from some subset of R2 . We will use the
second approach; therefore, it is important (before we start) to have the appropriate
language to describe such maps.

3.1 Concerning maps

Recall the convention that a map is a continuous function. We will now introduce some
more terminology. The notions of surjective, injective, and bijective functions/maps are
as important in this subject as any other part of mathematics. We will also need a
related, slightly weaker, notion than injective. The following definition can be made in
a more general topological setting, but we will restrict ourselves to subsets of Rn .

Definition 3.1.1. Let U ⊂ Rn , a map f : U → Rm is locally injective, if for all

p ∈ U there is an open neighbourhood Vp ⊂ U such that p ∈ Vp and f |Vp (ie the
restriction of f to Vp ) is injective. Then we say that f is an immersion.

26
Concerning maps What is a surface?

Any injective map is locally injective (just take Vp = U for all p ∈ U ). Recall that
injective maps are sometimes called embeddings.
Example 3.1.2.
1. The function f : R → R : x 7→ x2 is not locally injective since any open
neighbourhood V0 of the point 0 ∈ R contains an interval [−ε, ε] for some
ε > 0. Then f (−ε) = ε2 = f (ε), so f is not injective on V0 .

2. Let S1 be the unit circle in R2 . The function

Figure 22. The image of g.

g : S1 → R2 : (cos(t), sin(t)) 7→ sin(3t)(cos(t), sin(t)),

fails to be injective since at t = 0, π/3, . . . , 5π/3, and 2π, g(t) = (0, 0).
Nevertheless, it is easy to check that around any of these points u, g is
injective on the interval (u − π/3, u + π/3). Hence, g is locally injective, it is
an immersion of the circle into the plane, but not an embedding.

3. Again, let S1 be the unit circle in R2 . The function

h : S1 → R2 : (cos(t), sin(t)) 7→ (sin2 (3t) + 1)(cos(t), sin(t)), Figure 23. The image of h.

is injective, and hence, h is an embedding of the circle into the plane.

4. As a final example involving the circle, consider the map

j : R → R2 : t 7→ (cos(t), sin(t)).

The image of this map is the unit circle, around which the real line winds
infinitely many times. It fails to be injective at every point since for any
r ∈ R, j(r + 2kπ) = j(r) for all k ∈ Z. Nevertheless it is locally injective
because if j is restricted to any interval of the form (a, a + 2π) then it is
Figure 24. Part of R (orange) be-
injective.
ing mapped by j onto the circle.

Definition 3.1.3. Let U ⊂ Rn and V ⊂ Rm , a homeomorphism9 is a function 9

Be careful, a ‘homeomorphism’ is
U → V which is continuous, bijective, and has a continuous inverse. If such a not the same as a ‘homomorphism’,
which is a function between algebraic
homeomorphism exists, than U and V are homeomorphic.
objects which preserves the algebraic
structure.
This essentially means that U and V are the same up to continuous changes.
Remark 3.1.4. For subsets of Rn , the fact that the inverse is continuous follows
from the fact that the map itself is continuous and bijective. This third condition
is necessary in more general topological settings so it is included here.
Recall that for U ⊂ Rn and V ⊂ Rm , a function f : U → V such that

f (x1 , x2 , . . . , xn ) = (f1 (x1 , . . . , xn ), f2 (x1 , . . . , xn ), . . . , fm (x1 , . . . , xn )),

is smooth if all partial derivatives of all orders of every fi exist.

Definition 3.1.5. Let U ⊂ Rn and V ⊂ Rm , a diffeomorphism is a function U →

V which is smooth, bijective, and has a smooth inverse. If such a diffeomorphism
exists, than U and V are diffeomorphic.

In this case, a diffeomorphism means that U and V are the same up to smooth
changes.
In the following sequence of subsets of the plane, two are diffeomorphic, and three
are homeomorphic.

27
Surfaces as sets and maps What is a surface?

Since every differentiable function is continuous, every diffeomorphism is a homeo-

morphism, but the converse is not true.

3.2 Surfaces as sets and maps

As when we began to think about curves, there are two ways to think about surfaces in
Rn . The first is as a subset which ‘looks’ 2-dimensional, whereas the second is as a map
from part of R2 into Rn . Before giving the definition, recall that a subset V of Rn is
open is for all v ∈ V there is ε > 0 such that u ∈ V |u − v| < ε ⊂ V .


Definition 3.2.1. A subset10 S 2 of Rn is a topological surface if for all p ∈ S 2 10

Be careful: the superscript 2
there us an open set V ⊂ Rn such that p ∈ V and there is a homeomorphism in S 2 indicates this subset is 2-
dimensional, it is not a power.
fp : U → V ∩ S 2 , where U ⊂ R2 is open.
A topological surface is a smooth surface if every homeomorphism fp can be
chosen to be a diffeomorphism.

Example 3.2.2.

a) S2 a sphere11 b) The outside of a cube 11

Throughout this course, Sn will
denote an n-dimensional sphere.
Whereas S 2 represents some surface,
not necessarily a sphere.

This is a smooth surface This is a topological surface

c) T2 a torus d) K2 a Klein bottle

This is a smooth surface This is not a topological surface

28
Surfaces as sets and maps What is a surface?

e) B2 an open disc f) D2 a closed disk

This is a smooth surface This is not a topological surface

In this course, we want to study surfaces using calculus, and so having corners like
the cube is a problem because we cannot differentiate there. Most of the time, having
our surfaces be twice continuously differentiable will be sufficient, but to avoid being
overly pedantic, we will usually just assume our surfaces are smooth.
This approach to surfaces, thinking of them as subsets of Rn , is not very amenable
to the tools of calculus, because we cannot differentiate a set, we can only differentiate
functions. For this reason, we want to think of surfaces as smooth maps instead.

Definition 3.2.3. A local parametrisation of a surface is a smooth, locally injec-

tive map σ : U → Rn from an open subset U ⊂ R2 .

σ S2
σ(U )

Figure 25. A cartoon of what a local parametrisation of a surface might look like.

Local parametrisations, as illustrated in figure 25, allow us to study local properties

of a surface, ie those properties which depend only on the point you are at in the surface,
or perhaps a small neighbourhood of that point.
At the end of this course, we will briefly discuss some global properties of surfaces,
but these often fall under the study of topology.

Notation 3.2.4. Let the coordinates on U ⊂ R2 be (u, v). Then the partial deriva-
tives of σ : U → Rn can be written
∂ ∂
σ = σu σ = σv
∂u ∂v

∂2 ∂2 ∂2 ∂2
σ = σ uu σ = σ uv σ = σ vu σ = σ vv
∂u∂u ∂u∂v ∂v∂u ∂v∂v
. . . and so on. Recall that in our case, when σ is smooth, Schwartz’ Theorem says
that σ uv = σ vu .

29
Surfaces as sets and maps What is a surface?

Example 3.2.5 (A plane). A plane Π in Rn is spanned by two linearly independent

vectors, say p, q ∈ Rn . If Π contains a point a, then a local parametrisation of the
plane is given by
σ : R2 → Rn : (u, v) 7→ a + up + vq.

Compare this to figure 26. Smoothness is also easy to verify:

σ u = p, σ v = q, σ uu = 0, σ uv = 0, ...

(Local) injectivity is also straightforward. Suppose σ(u, v) = σ(u′ v ′ ), then

0 = σ(u, v) − σ(u′ v ′ ) = (a + up + vq) − (a + u′ p + v ′ q)

= (u − u′ )p + (v − v ′ )q. Π
p
q
Linear independence implies that u − u′ = 0 and v − v ′ = 0, ie u = u′ and v = v ′ ,
as required.
a
In this case, one local parametrisation allows us to describe the whole surface. This
is usually not possible, as the next example illustrates. 0
Example 3.2.6 (The sphere). Any point on the sphere can be described using a
Figure 26. Parametrising a
longitude (the angle φ the point makes from some reference meridian) and latitude
plane.
(the angle θ the point makes from the equator), see figure 27. You can convince
z
yourself that a point with longitude and latitude (φ, θ) on the unit sphere S2 in R3
has Cartesian coordinates

(cos(θ) cos(φ), cos(θ) sin(φ), sin(θ)). θ y

Is the map σ : R2 → R3 : (φ, θ) 7→ (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ)) a local φ

parametrisation of the sphere? Each of the coordinates is a smooth function, so
that is not the problem. What about local injectivity? x
The preimage of the north pole (0, 0, 1) is {(φ, π/2+2kπ) | φ ∈ R, k ∈ Z}, a family
of lines in R2 , see figure 28. Any neighbourhood of any point on one of these lines
necessarily contains other points on that line, so σ is not locally injective. A similar Figure 27. Parametrising a
problem occurs with the south pole as well. To fix this problem, we need to restrict sphere by longitude and latitude.

the domain of σ. If we take R × (−π/2, π/2) = {(φ, θ) | −π/2 < θ < π/2} as the
domain of σ instead, we now have a smooth, locally injective map.
π
This now does not cover the sphere, it misses the two points (0, 0, ±1). If we choose
another parametrisation σ ′ which misses two different points, then we could cover π
every point on the sphere.
Question 3.1. Can you rigorously verify that
 π π
σ: R × − , → R3 : (φ, θ) 7→ (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ))
2 2
is indeed locally injective everywhere? Figure 28. The preimage under
σ of the north (pink, solid) and
Remark 3.2.7. It turns out that there does not exist an injective local parametrisa- south (purple, dashed) poles
tion which covers the whole sphere, you need to use at least two local parametrisa-
tions. The proof requires tools from topology which we will not see in this module.
On the other hand, some subsets of Rn which are the image of a local parametrisa-
tion are not smooth surfaces, if the local parametrisation fails to be injective. For
example, there exists a surjective local parametrisation σ : R2 → R3 whose image
is the Klein bottle K2 from example 3.2.2. We can still study the Klein bottle using
the methods we develop in this course.

30
Surfaces as sets and maps What is a surface?

Definition 3.2.8. A local parametrisation σ

e: U
e → Rn is a reparametrisation of
a local parametrisation σ : U → Rn if there exists a diffeomorphism Φ : U → U
e
(called a reparametrisation map) such that for all (u, v) ∈ U

e (Φ(u, v)) = σ(u, v).

σ

A schematic representation of a reparametrisation is shown in figure 29, where the

equation σ
e (Φ(u, v)) = σ(u, v) means that the result is the same whichever way around
the diagram you go.

Φ S2

σ
e
U
e

y
Figure 29. A cartoon of what a reparametrisation of a surface looks like.

Example 3.2.9 (A reparametrisation of S2 ). Let S2 be the unit sphere in R3 and x

consider the parametrisation
p
e: U
σ e → R3 : (u, v) 7→ (u, 1 − u2 − v 2 , v),
Figure 30. The image of the
for U
e = {(u, v) ∈ R2 | u2 + v 2 < 1}. The image of σ
e lies in S2 since parametrisation σ
e.

σ (u, v)∥2 = u2 + (1 − u2 − v 2 ) + v 2 = 1.
∥e

Notice also that U

e is open, σ
e is injective, and it is also smooth, since
   
−u −v
e u = 1, √
σ ,0 e v = 0, √
σ ,1
1 − u2 − v 2 1 − u2 − v 2
√
which are continuous on U e since 1 − u2 − v 2 ̸= 0, and the same holds for all
orders. Therefore, σ
e is a local parametrisation of S2 , see figure 30.
Now consider the local parametrisation from example 3.2.6 with a restricted do-
main, ie
σ : U → R3 : (φ, θ) 7→ (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ))

for U = {(φ, θ) ∈ R2 | φ ∈ (0, π), θ ∈ (−π/2, π/2)}.

Claim: e is a reparametrisation of σ.
σ

31
Regular surfaces and tangent spaces What is a surface?

To prove this, we need to find a diffeomorphism Φ : U → U

e such that

e (Φ(φ, θ)) = σ(φ, θ).

σ
| {z }
=:(u,v)

This implies that

 
u
 cos(θ) cos(φ)

 √ 
1 − u2 − v 2 = cos(θ) sin(φ) .
 

v sin(θ)


From the first and third equalities, we can define Φ(φ, θ) = (cos(θ) cos(φ), sin(θ)),
which is smooth. Moreover,
p q
1 − u2 − v 2 = 1 − cos2 (θ) cos2 (φ) − sin2 (θ)
q
= 1 − cos2 (θ)(1 − sin2 (φ)) − sin2 (θ)
q
= 1 − (cos2 (θ) + sin2 (θ)) + cos2 (θ) sin2 (φ)
= cos(θ) sin(φ),

hence σ
e (Φ(φ, θ)) = σ(φ, θ).
Now consider Ψ : Ue → U defined by
   
u
Ψ(u, v) = arccos √ , arcsin(v) .
1 − v2

This is smooth on U
e (using the fact that this is open), and, moreover,
! !
cos(θ) cos(φ)
Ψ(Φ(φ, θ)) = arccos p , arcsin(sin(θ))
1 − sin(θ)2
cos(θ) cos(φ)
   
= arccos , θ = (φ, θ),
cos(θ)

and
   
u
Φ(Ψ(u, v)) = cos(arcsin(v)) cos(arccos √ ), sin(arcsin(v))
1 − v2
q 
u
= 1 − sin2 (arcsin(v)) √ , v = (u, v).
1 − v2

Thus, Ψ is a smooth inverse of Φ, so Φ is a diffeomorphism, and hence a reparametri-

sation map.

3.3 Regular surfaces and tangent spaces

One useful tool in our study of surfaces will be the understanding we have gained about
curves, which we will apply by studying curves that live in a given surface.

Definition 3.3.1. Let σ : U → Rn be a local parametrisation of a surface S 2 . A

smooth curve in S 2 is a smooth curve

γ : (a, b) → Rn : t 7→ σ(u(t), v(t)),

in particular, the trace of γ lies in S 2 .

Notice that γ is the image under σ of a plane curve γ

e : (a, b) → U ⊂ R2 , see figure 31.

32
Regular surfaces and tangent spaces What is a surface?

(a, b)
U
γ σ S2
e

Figure 31. A smooth curve lying in a surface.

Definition 3.3.2. Let σ : U → Rn be a local parametrisation with image S 2 , and

let p ∈ S 2 be a point. The tangent space to S 2 (or σ) at p, Tp S 2 is the set of
tangent vectors to all smooth curves on S 2 through p, at p. Or in other words,

Tp S 2 := {γ ′ (t0 ) | γ : (a, b) → S 2 , γ(t0 ) = p}.

The tangent space of a surface is illustrated in figure 32.

γ ′1 Tp S 2
γ ′2
p p
γ ′3 γ3 γ2
γ1

Figure 32. The tangent space to a surface S 2 at the point p.

Proposition 3.3.3. The tangent space of a local parametrisation σ at p ∈ σ(U ) is

a vector space and it is spanned by {σ u , σ v } evaluated at p.

If (u0 , v0 ) ∈ U is such that σ(u0 , v0 ) = p, then

Tp S 2 = spanR {σ u (u0 , v0 ), σ v (u0 , v0 )}.

Proof. Let γ be a smooth curve in S 2 and write

γ(t) = σ(u(t), v(t))

for some smooth functions u(t) and v(t). Then, applying the chain rule,

γ ′ = σ u u′ + σ v v ′ ∈ span{σ u , σ v }.

Conversely, consider aσ u + bσ v ∈ span{σ u , σ v } and define

γ(t) = σ(u0 + at, v0 + bt)

for t ∈ (−ε, ε), ε > 0 sufficiently small.12 Then γ is a smooth curve in S 2 , γ(0) = 12
Since U is open, such an ε exists
σ(u0 , v0 ) = p and γ ′ (0) = aσ u + bσ v . ■ so that γ is a well defined curve in
S2.
Since a surface is supposed to locally look like a piece of R2 , we would expect its
tangent space to be a plane, ie two dimensional, at every point. It turns out that out

33
Examples of surfaces What is a surface?

definition of smooth surface is not quite strong enough to guarantee this.

Example 3.3.4 (Degenerate tangent spaces). Consider the local parametrisation
σ : R2 → R3 : (u, v) 7→ (u2 , u3 , v), which is smooth and injective, shown in fig-
ure 33. Then

σ u = (2u, 3u2 , 0) σ v = (0, 0, 1)

so at all points in the surface where u = 0, the tangent space is the span of
{(0, 0, 0), (0, 0, 1)}, which is one dimensional.
To rule out situations like this, we make the following stronger definition.

Definition 3.3.5. A local parametrisation σ : U → Rn is regular if {σ u , σ v } are

linearly independent at all points in U . If the image lies in R3 , σ is regular if Figure 33. The surface para-
σ u × σ v ̸= 0 for all (u, v) ∈ U . metrised by σ(u, v) = (u2 , u3 , v).

If σ is regular, then Tp S 2 is a plane at every point, and we call the tangent space
the tangent plane.

Proposition 3.3.6. The tangent plane does not depend on the choice of parametri-
sation.

Proof. Let σ e: U
e → Rn be a reparametrisation of σ : U → Rn with reparametri-
sation map Φ : U → U
e : (u, v) 7→ (e
u, ve). Then σ
e (Φ(u, v)) = σ(u, v) and applying
the chain rule

σu = u e ũ + veu σ
eu σ e ṽ
σv = u e ũ + vev σ
ev σ e ṽ ,

so the basis of Tσ(u,v) σ(U ) can be written in terms of the basis Tσ̃(ũ,ṽ) σ
e (U
e ). Reg-
ularity means that these are both 2-dimensional vector spaces, and so they must
coincide. ■
Notice that if we write the Jacobian of the reparametrisation map as
!
u
eu uev
JΦ = ,
veu vev

then this proof shows that if we have a vector a in the tangent space Tp S 2 written in
terms of the basis B = {σ u , σ v } as
!
a
[a]B = aσ u + bσ v = ,
b

then we can re-write it in the basis Be = {e e ṽ } by multiplying by JΦ , ie

σ ũ , σ

[a]Be = JΦ⊺ [a]B .

3.4 Examples of surfaces

The study of surfaces is rich in examples and pictures. In this section, we will build up a
zoo of such examples—a menagerie to be our playground in which to illustrate and test
out the general theory of surfaces.

34
Examples of surfaces What is a surface?

Example 3.4.1 (Ruled surfaces). A ruled surface is a surface which is swept out by
a line moving through Rn . Alternatively it can be thought of as a union of lines
parametrised by a space curve, see figure 34.
If the space curve is γ(u), and for each u, the line through γ(u) has direction vector
δ(u) then a parametrisation of the ruled surface is

σ(u, v) = γ(u) + vδ(u).

To check regularity, we compute

σ u = γ ′ + vδ ′ σ v = δ,

so σ is regular if and only if γ ′ + vδ ′ and δ are linearly independent. In a small

neighbourhood of γ, ie for v ∈ (−ε, ε) for ε > 0 sufficiently small, it suffices to take
δ and γ ′ to be linearly independent.
The simplest example of a ruled surface is a plane, where γ is taken to be a line,
and δ is constant. Setting γ(u) = a+up and δ = q, we recover the parametrisation
in example 3.2.5.
The next couple of examples are special cases of ruled surfaces.
Example 3.4.2 (Generalised cylinder). Let γ(u) = (f (u), g(u), 0) be a curve in the
plane, the generalised cylinder corresponding to γ is the surface swept out by
Figure 34. A ruled surface
moving γ through R3 in the direction perpendicular to the plane, see figure 35. parametrised by the curve γ
Equivalently, it is the ruled surface obtained by taking δ = (0, 0, 1). In particular, (blue), and the lines shown in
we can define black.
σ(u, v) = (f (u), g(u), v).

It is straightforward to check that

σ is smooth ⇔ γ is smooth
σ is locally injective ⇔ γ is locally injective

For regularity, notice that that σ u = γ ′ lies in the (x, y)-plane, while σ v = δ does
not, so these vectors are linearly independent, and σ if and only if γ ′ ̸= 0, ie if and
only if γ is regular.
Note:

a) In example 3.3.4 we considered a non-regular example of a generalised cylin-

der, where γ(u) = (u2 , u3 ).

b) If γ(u) = (cos(u), sin(u)) is parametrises the unit circle, then we recover the
standard circular cylinder

σ(u, v) = (cos(u), sin(u), v).

Example 3.4.3 (Generalised cone). Let γ : (a, b) → R3 : u 7→ (f (u), g(u), 1) again be

a plane curve, now lying in the plane {(x, y, 1) | x, y ∈ R}. This time, as we move
it in the direction perpendicular to the plane, we can rescale the curve by a factor
Figure 35. An example of a gen-
proportional to the distance moved. The surface which is swept out is a generalised eralised cylinder.
cone, shown in figure 36. As a ruled surface, δ(u) is the vector from the origin to
γ(u), giving the parametrisation

σ(u, v) = γ(u) + vγ(u) = (v + 1)γ(u).

We may as well replace v + 1 7→ v and write σ(u, v) = vγ(u) = (vf (u), vg(u), v).

35
Examples of surfaces What is a surface?

If γ is smooth and locally injective, then so is σ except for v = 0, where local

injectivity fails. To check regularity, we compute:

σ u = vγ ′ σ v = γ.

Here, σ u lies in the (x, y)-plane while γ does not, so σ is regular as long as γ
is regular and v ̸= 0. In other words, we can choose as the domain for σ, U =
(a, b) × (R − {0}).

Example 3.4.4 (Quadratic surfaces). A quadratic surface (sometimes called a

quadric surface) is the vanishing set of a quadratic polynomial
Figure 36. An example of a gen-
q(x, y, z) = a11 x2 + 2a12 xy + 2a13 xz + a22 y 2 + 2a23 yz + a33 z 2 eralised cone.

+ b1 x + b2 y + b3 z + c (3.1)

where aij , bk , c ∈ R; so S 2 = {(x, y, z) ∈ R3 | q(x, y, z) = 0}. Any quadratic

polynomial like this can be written in the form
 
  x  
q(x, y, z) = x y z A y  + b · x y z + c,
 
z

where A is a 3 × 3 symmetric matrix, b ∈ R3 , and c ∈ R are all constant.

Recall that any symmetric matrix is diagonalisable by an element of O(3). After
such a change of basis, and possibly translating in R3 , it is possible to show that
every quadratic surface is one of a handful in standard form (listed below) up to
a rigid motion of R3 . Some polynomials lead to degenerate surfaces, subsets of R3
which are really points or lines for example. If we ignore these, it is possible to
show that every quadratic surface is one of the following.

a) Ellipsoid b) Hyperboloid of one sheet

x2 y2 z2 x2 y2 z2
+ + =1 + − =1
a2 b2 c2 a2 b2 c2

36
Examples of surfaces What is a surface?

c) Hyperboloid of two sheets d) Elliptic paraboloid

x2 y2 z2 x2 y2
− 2
− 2 + 2 =1 2
+ 2 =z
a b c a b

e) Hyperbolic paraboloid f) Quadratic cone

x2 y2 x2 y2 z2
2
− 2 =z 2
+ 2 − 2 =0
a b a b c

See example 3.4.3.

g) Elliptic cylinder h) Hyperbolic cylinder

x2 y2 x2 y2
+ =1 − =1
a2 b2 a2 b2

See example 3.4.2. See example 3.4.2.

37
Examples of surfaces What is a surface?

i) Parabolic cylinder j) Plane

z=0
x2
=y
a2

See example 3.4.2. See example 3.2.5.

Finding parametrisations for each of these is straightforward but tedious. For (f)–
(j) see the previous examples. Here are parametrisations for a couple of the others.

a) For the ellipsoid, we can take

σ(θ, φ) = (a cos(θ) cos(φ), b cos(θ) sin(φ), c sin(θ)).

b) For the hyperboloid of one sheet, we can take

σ(θ, φ) = (a cosh(φ) cos(θ), b cosh(φ) sin(θ), c sinh(φ)).

Example 3.4.5 (Surfaces of revolution). A surface of revolution S 2 is the surface

swept out by a curve in the (x, z)-plane which is rotated around the z-axis, see
figure 37. Suppose γ is given by

γ(u) = (f (u), 0, g(u)),

then
Figure 37. An example of a
σ(u, v) = (f (u) cos(v), f (u) sin(v), g(u))
surface of revolution. Meridians
is a parametrisation of S 2 . (blue) and parallels (orange) are
shown.
The meridians are the curves in S 2 for which v is constant, and the parallels are
the curves for which u is constant.
As usual, we want to check when a surface of revolution is regular.

σ u = (f ′ cos(v), f ′ sin(v), g ′ ) σ v = (−f sin(v), f cos(v), 0),

so

σ u × σ v = (−f g ′ cos(v), −f g ′ sin(v), f f ′ cos2 (v) + f f ′ sin2 (v))

= (−f g ′ cos(v), −f g ′ sin(v), f f ′ ).

Then
∥σ u × σ v ∥2 = f 2 g ′2 + f 2 f ′2 = f 2 (f ′2 + g ′2 ).

Hence, σ is regular if and only if f (u) ̸= 0 (ie γ does not intersect the z-axis), and
γ is itself regular (so f ′2 + g ′2 ̸= 0).

38
Examples of surfaces What is a surface?

Some special cases:

• Spheres, or more generally ellipsoids with two of a, b, and c equal.

• Hyperboloids of one or two sheets with a = b.

• Elliptic paraboloids with a = b.

• Circular cones.

• Circular cylinders.
It is sometimes that case that the same surface can be described in several different
ways: for example the hyperboloid of one sheet is also a ruled surface, and if it is circular,
it is also a surface of revolution. This is shown in the next example.
Example 3.4.6. Consider the ellipse in the (x, y)-plane which satisfies

x2 y2
2
+ 2 = 1,
a b Figure 38. The construction of
ℓθ .
for a, b ∈ R constant. Take two copies of this ellipse in R which lie in the planes
3

z = ±1, parametrised by

γ ± (θ) = (a cos(θ), b sin(θ), ±1).

Fix an angle θ0 ∈ [0, 2π) and consider the family of lines ℓθ which pass through
γ + (θ) and γ − (θ + θ0 ) as θ varies, see figure 38. Such a line has directions

δ(θ) = γ + (θ) − γ − (θ + θ0 )
= (a cos(θ) − a cos(θ + θ0 ), b sin(θ) − b sin(θ + θ0 ), 2).

Hence, the resulting ruled surface has parametrsation

Figure 39. The surface built out
of the lines ℓθ .
σ(θ, t) = γ + (θ) + tδ(θ)
= (a(cos(θ) + t(cos(θ) − cos(θ + θ0 ))),
b(sin(θ) + t(sin(θ) − sin(θ + θ0 ))), 1 + 2t)
= : (x(θ, t), y(θ, t), z(θ, t)).

By visualising the surface, see figure 39, we might guess that it is a hyperboloid of
one sheet (see example 3.4.4(b)), but let’s prove it. Consider the quantity

x(θ, t)2 y(θ, t)2

2
+ = cos2 (θ) + 2t cos(θ)(cos(θ) − cos(θ + θ0 ))
a b2
+ t2 (cos(θ) − cos(θ + θ0 ))2
+ sin2 (θ) + 2t sin(θ)(sin(θ) − sin(θ + θ0 ))
+ t2 (sin(θ) − sin(θ + θ0 ))2
= 1 + 2t − 2t(cos(θ) cos(θ + θ0 ) + sin(θ) sin(θ + θ0 ))
+ t2 (cos2 (θ) − 2 cos(θ) cos(θ + θ0 ) + cos2 (θ + θ0 )
+ sin2 (θ) − 2 sin(θ) sin(θ + θ0 ) + sin2 (θ + θ0 ))
= 1 + 2t + 2t2
− (2t + 2t2 )(cos(θ) cos(θ + θ0 ) + sin(θ) sin(θ + θ0 ))
| {z }
=cos(θ−(θ+θ0 ))=cos(−θ0 )=cos(θ0 )

= 1 + (2t + 2t )(1 − cos(θ0 )).

2

39
Examples of surfaces What is a surface?

Notice that this does not depend on θ, and can be related to z(θ, t)2 = (1 + 2t)2 =
1 + 4t + 4t2 as

x(θ, t)2 y(θ, t)2 1 1

+ = 1 + (1 + 4t + 4t2 )(1 − cos(θ0 )) − (1 − cos(θ0 ))
a2 b 2 2 2
z2 1 2
= 1 + 2 − 2; c : =
2
.
c c 1 − cos(θ0 )

Therefore
x(θ, t)2 y(θ, t)2 z2 1 c2 − 1
+ − = 1 − = ;
a2 b2 c2 c2 c2
if we define

c2 − 1 2 c2 − 1 c2 − 1
a2 = a2 ; b = b2 ; c2 = c2 = c2 − 1,
c2 c2 c2
then we see that σ parametrises the surface

x(θ, t)2 y(θ, t)2 z2

2 + 2 − 2 = 1,
a b c

which is indeed a hyperboloid of one sheet.

Remark 3.4.7. Notice that if we replace θ0 by −θ0 in the argument above, essen-
tially the same computation shows that we parametrise the same surface, but using
a different family of lines, see figure 40. This shows that the hyperboloid of one Figure 40. The hyperboloid of
sheet is a doubly ruled surface (a ruled surface in two essentially different ways). one sheet can be viewed as a ruled
In fact, it can be shown that every doubly ruled surface is a quadratic surface. surface in two different ways.

Question 3.2. Can you find another surface among those discussed in example 3.4.4
(other than the plane) which is a doubly ruled surface?
Application 3.4.8 (Ruled surfaces and architecture). Ruled surfaces are often
used by architects in the design of buildings because they can create complicated
shapes which are relatively easy to construct using straight structural components.

Application 3.4.9 (Quadratic surfaces and special relativity). A bilinear form

B : V ⊕ V → R on a (real finite dimensional) vector space V can be represented by
a matrix B (after a basis for V has been fixed) such that
Figure 41. A roof designed as a
ruled surface. Image: Hasanisawi
B(u, v) = u⊺ Bv. (CC BY-SA 4.0 DEED)

Associated to any bilinear form is a quadratic form

QB : V → R : v 7→ B(v, v) = v⊺ Bv.

The level sets of QB , ie {v ∈ V | QB (v) = k} for some k ∈ R constant, are

quadratic surfaces in V . In 2D special relativity, space-time is modelled as (1 + 2)-
dimensional vector space {(t, x, y) | t, x, y ∈ R} ∼
= R3 equipped with the Lorenzian
form  
c2 0 0
L =  0 −1 0  ,
 
0 0 −1 Figure 42. The light cone in spe-
cial relativity.
where c is the speed of light in a vacuum. In this space-time, light travels along
lines where QL = 0, ie the null or light cone. Then paths whose tangent vector
has QL (t) > 0 are called time-like, and if QL (t) < 0 the path is called space-like,

40
Examples of surfaces What is a surface?

see figure 42. The surfaces where QL = k are


k < 0 The hyperboloid of one sheet


k = 0 The circular cone

k > 0 The hyperboloid of two sheets


as shown in figure 43.

Figure 43. Level surfaces of QL .

41
 First fundamental form

4 First fundamental form

We want to be able to compute and understand geometric properties on surfaces such
as distances, angles, and areas. We know how to define these in the Euclidean plane, for
example by identifying it with R2 equipped with the standard inner product. The inner
product can then be used to define distance and angles. In particular, if −·− : R2 ×R2 →
R is the standard inner product, then

• the distance between x and y is ∥x − y∥ = (x − y) · (x − y),

p

• the angle between x and y measured at the origin is

 
x·y
θ = arccos .
∥x∥∥y∥

In order to transfer what we already know about the geometry of the flat plane to a
curved surface S 2 , we can use the tangent space Tp S 2 to approximate S 2 at any point
p by a flat space. If S 2 is regular, Tp S 2 is a copy of R2 , a vector space with a natural
choice of basis, {σ u , σ v }, by proposition 3.3.3.
We just need to fix a choice of inner product on Tp S 2 for each p to define distances
and angles, and in order for these definitions to be compatible with the smooth structure
of the surface, we want this inner product to change smoothly as p moves about in the
surface. In principle, any such choice of inner product can be taken, and indeed, this
level of generality is taken in Riemannian geometry. In this course, however, we will
make a very special choice of inner product using the fact that S 2 lives inside Rn .

4.1 The definition

Any two vectors in Tp S 2 ⊂ Rn are also (free) vectors in Rn , and hence we can compute
their inner product with respect to the standard one on Rn . This induces an inner
product, called the first fundamental form on Tp S 2 by restricting the inner product on
Rn , see figure 44.

Definition 4.1.1. The first fundamental form of a regular local parametrisation

σ : U → Rn is the inner product Ip (−, −) induced on the tangent planes Tp S 2 by
the Euclidean inner product on Rn .

Figure 44. The tangent plane to a regular surface in Rn can be viewed as a subspace, with an
induced inner product.

Proposition 4.1.2. The first fundamental form indeed gives an inner product on
each tangent plane Tp S 2 of a regular local parametrisation, and in terms of the

42
The definition First fundamental form

basis {σ u , σ v }, it is given by the matrix

!
σu · σu σu · σv
Ip = .
σv · σu σv · σv

Proof. Denote by − · − the standard inner product on Rn , and let Ip (−, −) be

the function induced on Tp S 2 × Tp S 2 by − · −. In other words, this means that
for any x, y ∈ Tp S 2 , Ip (x, y) = x · y. It follows immediately that Ip (−, −) is a
positive definite symmetric bilinear form, ie an inner product.
We can compute this inner product in terms of the basis {σ u , σ v }. Suppose x =
xu σ u + xv σ v and y = yu σ u + yv σ v , where xu , xv , yu , yv ∈ R. Then

Ip (x, y) = Ip (xu σ u + xv σ v , yu σ u + yv σ v )
= xu yu Ip (σ u , σ u ) + xu yv Ip (σ u , σ v )
+ xv yu Ip (σ v , σ u ) + xv yv Ip (σ v , σ v )
= xu yu σ u · σ u + xu yv σ u · σ v + xv yu σ v · σ u + xv yv σ v · σ v
! !
  σ ·σ σ ·σ yu
u u u v
= xu xv
σv · σu σv · σv yv
= x⊺ Ip y,

as required. ■

Definition 4.1.3. For a regular local parametrisation σ : U → Rn , the coefficients

of the first fundamental form are

E := σ u · σ u , F := σ u · σ v = σ v · σ u , G := σ v · σ v ,

so that !
E F
Ip = .
F G

Example 4.1.4. Consider a plane parametrised by σ : R2 → R3 : (u, v) 7→ a+ub+vc

for some a, b, and c in R3 with {b, c} linearly independent (see example 3.2.5).
Then σ u = b and σ v = c so

E = ∥b∥2 , F = b · c, G = ∥c∥2 ,

and if p = σ(u, v) !
∥b∥2 b · c
Ip = .
b · c ∥c∥2

If we choose {b, c} to be an orthonormal basis for the plane, then

!
10
Ip = .
01

Remark 4.1.5. As this example shows, choosing different parametrisations of the

same surface will (in general) lead to different coefficients of the first fundamental
form—different values of E, F , and G. Nevertheless, the form itself is the inner
product on the tangent plane induced by the standard inner product on R3 . The
tangent plane doesn’t depend on the parametrisation, and so the first fundamental
form won’t change. What changes is the basis in which it is expressed.

43
The definition First fundamental form

Proposition 4.1.6. Let σe: Ue → Rn be a reparametrisation of σ : U → Rn with

reparametrisation map Φ : U → U e : (u, v) 7→ (e
u, ve). Then the coefficients of the
first fundamental forms of σ and σ
e are related by
! !
E F ⊺ E F
e e
= JΦ e e JΦ ,
F G F G
!
u
eu uev
where JΦ = is the Jacobian of the function Φ.
veu vev

Proof. This is a straightforward, if tedious, exercise in linear algebra. ■

Example 4.1.7. Consider the sphere with parametrisation σ : (−π/2, π/2) × R →
R3 given by
σ(θ, φ) = (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ)),

as in example 3.2.6. Then

σ θ = (− sin(θ) cos(φ), − sin(θ) sin(φ), cos(θ))

σ φ = (− cos(θ) sin(φ), cos(θ) cos(φ), 0)

and we can compute

E = sin2 (θ) cos2 (φ) + sin2 (θ) sin2 (φ) + cos2 (θ) = 1,
F = sin(θ) cos(φ) cos(θ) sin(φ) − sin(θ) cos(φ) cos(θ) sin(φ) + 0 = 0,
G = cos2 (θ) sin2 (φ) + cos2 (θ) cos2 (φ) + 02 = cos2 (θ).

Then !
1 0
Ip = .
0 cos2 (θ)

Example 4.1.8. Consider a generalised cylinder, as in example 3.4.2, with

parametrisation σ(u, v) = (f (u), g(u), v). Then

σ u = (f ′ , g ′ , 0) σ v = (0, 0, 1),

and

E = f ′2 + g ′2 , F = 0, G = 1,

ie !
f ′2 + g ′2 0
Ip = .
0 1

If σ is regular, γ(u) = (f (u), g(u)) is regular. If we choose a unit speed parametri-

sation for γ, then f ′2 + g ′2 = 1 and we get the same first fundamental form as the
plane.
Example 4.1.9. The Klein bottle shown in example 3.2.2 cannot be embedded in
R3 , only immersed, however it can be embedded in four dimensions, and one such
parametrisation is given by

σ : R2 → R 4
: (θ, φ) 7→ ((1 + cos(θ)) cos(φ), (1 + cos(θ)) sin(φ),
sin(θ) cos(φ/2), sin(θ) sin(φ/2)).

44
Lengths, angles, and areas on a surface First fundamental form

Even though this is not in R3 , we can still compute the first fundamental form in
the same way. In particular,

σ θ = (− sin(θ) cos(φ), − sin(θ) sin(φ),

cos(θ) cos(φ/2), cos(θ) sin(φ/2))
σ φ = (−(1 + cos(θ)) sin(φ), (1 + cos(θ)) cos(φ),
1 1
− sin(θ) sin(φ/2), sin(θ) cos(φ/2)).
2 2
Therefore

E = sin2 (θ) cos2 (φ) + sin2 (θ) sin2 (φ) + cos2 (θ) cos2 (φ/2) + cos2 (θ) sin2 (φ/2)
= sin2 (θ) + cos2 (θ)
=1
F = (1 + cos(θ)) sin(θ) cos(φ) sin(φ) − (1 + cos(θ)) sin(θ) sin(φ) cos(φ)
1 1
− sin(θ) cos(θ) cos(φ/2) sin(φ/2) + sin(θ) cos(θ) sin(φ/2) cos(φ/2)
2 2
=0
G = (1 + cos(θ))2 sin2 (φ) + (1 + cos(θ))2 cos2 (φ)
1 1
+ sin2 (θ) sin2 (φ/2) + sin2 (θ) cos2 (φ/2)
4 4
1
= (1 + cos(θ)) + sin (θ)
2 2
4

Of course, we cannot visualise this surface in R4 , but we can project it to R3 , and

one such projection (where it is no longer embedded) is shown in figure 45.

4.2 Lengths, angles, and areas on a surface

Figure 45. The projection of the
The first fundamental form can be used to generalise the notions of length, angle, and
Klein bottle in R4 to R3 obtained
area from the plane to any regular surface parametrisation. We begin with lengths. by restricting to the second, third,
and fourth coordinates.
Proposition 4.2.1. Let σ : U → Rn be a regular local parametrisation, and
γ : (a, b) → Rn a curve on S 2 = σ(U ) given by γ(t) = σ(u(t), v(t)). Then the
length of γ is given by
Z b q Z b p
ℓ(γ) = Iγ(t) (γ ′ , γ ′ )dt = u′2 E + 2u′ v ′ F + v ′2 Gdt.
a a

Proof. By the chain rule,

γ ′ = u′ σ u + v ′ σ v ,

which lies in the tangent plane of S 2 by proposition 3.3.3. Then

! !
E F u′
∥γ ∥ = γ · γ = γ Iγ(t) γ = (u , v )
′ 2 ′ ′ ′⊺ ′ ′ ′
= u′2 E + 2u′ v ′ F + v ′2 G;
F G v′

here we have used the definition of the first fundamental form: that it is the
standard inner product restricted to Tp S 2 , and the fact that the coefficients in
definition 4.1.3 express the first fundamental in the same basis as we have expressed
γ ′ . The proposition follows by substituting this into the definition of length, see
definition 1.2.1. ■
Recall the definition of the angle between two curves which intersect.

45
Lengths, angles, and areas on a surface First fundamental form

Definition 4.2.2. Let γ i : (ai , bi ) → Rn for i = 1, 2 be two curves, and suppose they
meet, ie there is t1 ∈ (a1 , b1 ) and t2 ∈ (a2 , b2 ) such that γ 1 (t1 ) = γ 2 (t2 ) = p. Then
the angle between γ 1 and γ 2 at p is the angle θ ∈ [0, π) which satisfies

γ ′1 (t1 ) · γ ′2 (t2 ) = ∥γ ′1 ∥∥γ ′2 ∥ cos(θ).

For curves lying in a surface, the dot product can be replaced by the first fundamental
form, so

Ip (γ ′1 , γ ′2 ) Ip (γ ′1 , γ ′2 )
   
θ = arccos = arccos .
∥γ ′1 ∥∥γ ′2 ∥ Ip (γ ′1 , γ ′1 )1/2 Ip (γ ′2 , γ ′2 )1/2

Finally, to compute the area of (some part of) a surface we can start by thinking
about a small region. For a local parametrisation σ : U → Rn with coordinated u and v
in U , consider a small rectangle with corner (u0 , v0 ) and with sides of length ∆u and ∆v,
respectively. The image of this small rectangle will be approximately a parallelogram
in σ(U ) with corner σ(u0 , v0 ) and with sides ∆uσ u and ∆vσ v (evaluated at (u0 , v0 )),
see figure 46. The area of this parallelogram (and hence approximately the area of the
image of the rectangle) is ∥∆uσ u ∥∥∆vσ v ∥ sin(θ), where θ is the angle between ∆uσ u
and ∆vσ v .

Figure 46. Approximating the area of a surface with small parallelograms.

If we set γ 1 (u) = σ(u∆u + u0 , v0 ) and γ 2 (v) = σ(u0 , v∆v + v0 ), then ∆uσ u and
∆vσ v are γ ′1 and γ ′2 . We can use the expression for the angle between curves to re-
express

∥∆uσ u ∥∥∆vσ v ∥ sin(θ) = ∆u∥σ u ∥∆v∥σ v ∥ 1 − cos2 (θ)

p
s 
Ip (
∆uσ
 u ,  v ) 2
∆vσ
= ∆u∥σ u ∥∆v∥σ v ∥ 1 −
∥∆uσ
 u |∥ ∆vσ
 v∥
q
= ∆u∆v ∥σ u ∥2 ∥σ v ∥2 − Ip (σ u , σ v )2
p
= ∆u∆v EG − F 2 .

Let R ⊂ U be an open13 subset of U , and suppose we want to compute the area of 13

We need the integral below to be
σ(R). We could cover R by small rectangles as above, sum over these rectangles and well defined. In general, that means
that R must be measurable in the
take a limit—this is precisely the definition of a (Riemann) integral:
sense of measure theory, but for our
Xp purposes, assuming R is open (which
Area(σ(R)) ≈ EG − F 2 ∆u∆v. implies it is measurable) is sufficient.
R

This motivates the following definition.

Definition 4.2.3. Let σ : U → Rn be a regular local parametrisation, and let R ⊂ U

46
Lengths, angles, and areas on a surface First fundamental form

be open. Then the area of σ(R) is

ZZ p
Aσ (R) := EG − F 2 dudv.
R

Remark 4.2.4. Notice that representing the first fundamental form as a matrix in
the basis {σ u , σ v }, we can re-write the area as
ZZ q
Aσ (R) := det(Ip )dudv,
R

where p = σ(u, v). In the special case that the surface is in R3 , then we can write
down the area of the parallelogram, and hence the area, using the cross product:
ZZ
Aσ (R) := ∥σ u × σ v ∥dudv.
R

Proposition 4.2.5. The length of a curve on a surface, the angle between two curves,
and the area of a region on a surface do not depend on the choice of parametrisation.

Proof. The length of a curve was defined in section 1.2 without reference to
a surface, so certainly does not depend on the parametrisation of the surface the
curve happens to lie in. Similarly, the angle between two curves is define with
respect to the standard inner product on the ambient space Rn , and so does not
depend on the parametrisation of the surface.
In proposition 4.1.6, we saw that Ip = JΦ⊺ Ĩp JΦ , so
q q q
det(Ip ) = det(JΦ⊺ Ĩp JΦ ) =
det(JΦ⊺ ) det(Ĩp ) det(JΦ⊺ )
q q
= det(Ĩp ) det(JΦ⊺ )2 = det(Ĩp )| det(JΦ )|.

Recall that in a double integral, the change of variables associated to a map Φ is

given by | det(JΦ )|dudv = de
ude
v , and therefore
ZZ q ZZ q
Aσ (R) = det(Ip )dudv = det(Ĩp )de
ude
v = Aσ̃ (Φ(R)).
R Φ(R)

■
Example 4.2.6. Let’s use this to compute the surface area of the unit sphere. From
example 4.1.7 we know that the local parametrisation

σ(θ, φ) = (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ))

for (θ, φ) ∈ (−π/2, π/2) × R has first fundamental form

!
1 0
Ip = .
0 cos2 (θ)

If we restrict the domain of φ to (0, 2π), then σ covers the whole sphere except
for one line of longitude (σ(θ, 0) together with the north and south poles), so set
R = (−π/2, π/2) × (0, 2π). This curve of points does not contribute to the area,

47
Maps of surfaces First fundamental form

so we can say that the area of the sphere is

ZZ q
Aσ (R) = det(Ip )dθdφ
R
π
Z 2π Z 2
= 1 × cos2 (θ) − 02 dθdφ
p
0 −π
2
π
Z 2π Z 2
= cos(θ)dθdφ = 4π,
0 −π
2

as expected.

4.3 Maps of surfaces

We have considered maps from a very simple surface, namely R2 , to a surface (the
definition of a local parametrisation), and from curves to surfaces. Now we will consider
maps between surfaces more generally. Let σ 1 : U1 → Rn1 and σ 2 : U2 → Rn2 be two
local parametrisations, and set Si2 = σ i (Ui ).

Definition 4.3.1. A map f : S12 → Rm is smooth if (f ◦ σ 1 ) : U1 → Rm is smooth.

A map f : S12 → S22 is a smooth map between surfaces if there is a smooth map
F : U1 → U2 such that f ◦ σ 1 = σ 2 ◦ F :

F
U1 U2

σ1 σ2

S12 S22
f

Additionally, f is a diffeomorphism if F can be chosen to be a diffeomorphism,

see figure 47.

Figure 47. A smooth map f between surfaces.

Notice that we have already seen an example of this: a reparametrisation. In this

case f : S 2 → S 2 is the identity and F = Φ.

48
Maps of surfaces First fundamental form

Φ
U1 U2

σ1 σ2

S2 S2
id

Example 4.3.2. Consider the unit sphere S2 = {(x, y, z) | x2 + y 2 + z 2 = 1} and the

circular cylinder of height 2 and radius 1, C 2 = {(x, y, z) | x2 + y 2 = 1, |z| < 1},
and define a smooth map by projecting the sphere (minus the north and south
poles) radially to the cylinder, as in figure 48

f : S2 \{(0, 0, ±1)} → C 2
!
x y
: (x, y, z) 7→ ,p ,z .
x2 + y 2 x2 + y 2
p

If we parametrise S2 \{(0, 0, ±1)} by

σ : U = (−π/2, π/2) × R → R3 : (θ, φ) 7→ (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ))

and the cylinder by

e : U → R3 : (θ, φ) 7→ (cos(φ), sin(φ), sin(θ))

σ

then we can set F : U → U to be the identity, and you can easily check that
e (θ, φ) = f (σ(θ, φ)), in other words f ◦ σ = σ
σ e ◦ F.
Given a smooth map between surfaces, there is an induced map between the tangent
spaces called the differential of the map.

Definition 4.3.3. Let σ : U → Rn be a local parametrisation, and set S 2 = σ(U ).

Also, let f : S 2 → Rm be a smooth map. Then for all p ∈ S 2 , the differential of
f at p is the function Figure 48. The unit sphere con-
dp f : Tp S 2 → Rm tained in a cylinder, and projected
radially outwards.
defined as follows. For v ∈ Tp S 2 , let γ(t) = σ(u(t), v(t)) be a curve in S 2 with
γ(0) = p and γ ′ (0) = v. Define γe (t) = f (γ(t)) to be the corresponding curve in
Rm , and set
dp f (v) := γ e ′ (0).

This definition is illustrated by figure 49. As a smooth map between surfaces is, in
particular, a smooth map, the differential is defined in this case too.

Figure 49. The definition of the differential of a map.

49
Conformal, isometric, and area-preserving maps First fundamental form

Lemma 4.3.4. The differential is well-defined in the sense that such a curve γ exists
for any v (see the proof of proposition 3.3.3), and the function does not depend on
the choice of γ.

Proposition 4.3.5. The differential of a smooth map f : S 2 → Rm is a linear map

between vector spaces, and is defined on the basis {σ u , σ v } of Tp S 2 by

dp f (σ u ) = (f ◦ σ)u and dp f (σ v ) = (f ◦ σ)v .

If f is a smooth map between surfaces S12 → S22 , then the image of Tp S12 lies in
Tf (p) S22 .

This is illustrated by figure 50.

Figure 50. The differential of a smooth map between surfaces defines a linear map between
tangent spaces.

Example 4.3.6. We can continue example 4.3.2, and compute the differential of the
map f : S2 → C 2 . For a point p = σ(u0 , v0 ) ∈ S2 , the bases of Tp S2 and Tf (p) C 2
are given by {σ θ , σ φ } and {e e φ }. Recall that f ◦σ = σ
σθ , σ e , so by proposition 4.3.5,

dp f (σ θ ) = (f ◦ σ)θ = σ
eθ

and

dp f (σ φ ) = (f ◦ σ)φ = σ
eφ

so in this case, the linear map dp f : Tp S2 → Tf (p) C 2 can be expressed as a matrix

in terms of those bases as !
10
dp f = .
01

This is quite special behaviour, and arises because the map F from U → U is the
identity.

4.4 Conformal, isometric, and area-preserving maps

From the point of view of geometry, the most interesting smooth maps between surfaces
are those which preserve the geometric properties we discussed in section 4.2—this is like
group theory where we study maps between groups which preserve the multiplication
(homomorphisms).

Definition 4.4.1. Let σ : U → Rn be a regular local parametrisation with S 2 =

σ(U ) and σe: U
e → Rm be a regular local parametrisation with σ e (U
e ) = Se2 . A
smooth map f : S → Se is conformal at a point p if it preserves the angle
2 2

between any two curves in S 2 which meet at p. It is conformal if it is confor-

mal at every point in S 2 . It is locally conformal if every point p in S 2 has a

50
Conformal, isometric, and area-preserving maps First fundamental form

neighbourhood Vp on which f is conformal.

Note that, by our definition of angle, it is always the angle measured anticlockwise
in the tangent plane. Conformal maps therefore preserve signed angles, and a map with
preserves the size of angles, but reverses their direction is not conformal—such maps are
sometimes called anti-conformal.

Proposition 4.4.2. Let σ : U → Rn be a regular local parametrisation with S 2 =

σ(U ) and σ e: Ue → Rm be a local parametrisation with σ e (U
e ) = Se2 . A smooth
map f : S 2 → Se2 is conformal if and only if there is some smooth positive valued
function λ2 : S 2 → (0, ∞) such that for all p ∈ S 2

Ĩf (p) (dp f (v1 ), dp f (v2 )) = λ2 (p)Ip (v1 , v2 ) .

for all v1 , v2 ∈ Tp S 2 , where Ip is the first fundamental form of σ at p and Ĩf (p) is
the first fundamental form of σ e at f (p).

The idea is that, the angle between vectors is preserved if and only if the map
dp f only rotates or scales the vectors uniformly, in which case the scale factor is some
λ(p) ̸= 0.

Theorem 4.4.3. Let σ : U → S 2 and σ e : U → Se2 be regular local parametrisations

with the same domain U ⊂ R2 , with σ injective, and suppose that the coefficients
of the first fundamental forms of S 2 and Se2 satisfy

E = λ2 E,
e F = λ2 Fe, and G = λ2 G,
e

for some non vanishing function λ : U → R. Then the map f = σ

e ◦σ −1 : σ(U ) → Se2
is a locally conformal map.

Any local parametrisation σ : U → Rn can be thought of as a smooth map R2 ⊃ U →

S = σ(U ), and we have already seen in example 4.1.4 that the first fundamental form
2

of the plane (with respect to the standard parametrisation by orthonormal coordinates)

is !
10
I(u,v) = .
01

This motivates the following definition.

Definition 4.4.4. A regular local parametrisation σ : U → Rn is conformal or

isothermal if the coefficients of the first fundamental form satisfy E = G and
F = 0.

Theorem 4.4.5. Any regular surface can be covered by a collection of conformal

local parametrisations.

The proof of this of this is very difficult, and will not be covered here.

Corollary 4.4.6. Any two regular embedded surfaces are locally conformal.

Proof. Using theorem 4.4.5, we can find conformal local parametrisations of S 2

and Se2 , σ : U → S 2 and σ
e : U → Se2 with coefficients of the first fundamental
form satisfying E = G, F = 0, E e = G,
e and Fe = 0. Since the first fundamental
form is an inner product, E and E are positive. Set λ2 = E/E
e e > 0 and apply

51
Conformal, isometric, and area-preserving maps First fundamental form

theorem 4.4.3. ■
So far we have looked at maps of surfaces which preserve angles, what about maps
which preserve lengths?

Definition 4.4.7. Let σ : U → Rn be a regular local parametrisation with S 2 =

σ(U ) and σ e: U
e → Rm be a regular local parametrisation with σ e (U
e ) = Se2 . A
smooth map f : S → Se is a local isometry if it takes any smooth curve in S 2 to
2 2

a curve in Se2 of the same length. In this case, S 2 and Se2 are locally isometric.
A local isometry is an isometry if it is a diffeomorphism (in particular, if it is
bijective), in which case S 2 and Se2 are isometric.

Example 4.4.8. The plane is locally isometric to the generalised cylinder of any
regular curve. Let γ : (a, b) → R3 : s 7→ (f (s), g(s), 0) be any regular (locally
injective) curve, which we can assume, without loss of generality, is parametrised
by arc length. Let σ, σ
e : U → R3 for U = (a, b) × R be local parametrisations given
by
σ(u, v) = (0, u, v) (part of a plane)

e (u, v) = (f (u), g(u), v)

σ (generalised cylinder)

and define h : σ(U ) → σe (U ) : (0, u, v) 7→ (f (u), g(u), v) to be a smooth map from

the plane to the cylinder. This is smooth, but it is not a diffeomorphism unless γ
is injective.
Consider an arbitrary smooth curve in the plane µ : (c, d) → σ(U ) given by
µ(t) = σ(u(t), v(t)), and define µ e (t) = h(µ(t)) = (f (u(t)), g(u(t)), v(t)) be the
corresponding curve in σe (U ). The tangent to this curve is

e ′ (t) = (u′ f ′ (u), u′ g ′ (u), v ′ ) ,

µ

so

µ′ (t)∥2 = u′2 (f ′2 + g ′2 ) + v ′2 = u′2 + v ′2 = ∥µ′ (t)∥2 ;

∥e

where we have used the fact that γ has unit speed to say that f ′2 + g ′2 = 1. Then,
we can compute the length of µe to be
Z d Z d
′
µ) =
ℓ(e µ (t)∥dt =
∥e ∥µ′ (t)∥dt = ℓ(µ).
c c

So h maps any curve in the plane to a curve in the cylinder of the same length, so
h is a local isometry.
It is possible to characterise local isometries, similar to the way that theorem 4.4.3
characterises locally conformal maps.

Theorem 4.4.9. Two surfaces S12 and S22 are locally isometric if and only if they
have injective regular local parametrisations σ 1 : U → S12 and σ 2 : U → S22 with
the same domain U ⊂ R2 such that E1 = E2 , F1 = F2 , and G1 = G2 .

Proof. First, assume that σ 1 and σ 2 have the same coefficients of the first
fundamental form. Define f : σ 1 (U ) → σ 2 (U ) : σ 1 (u, v) 7→ σ 2 (u, v). For a curve
(u(t), v(t)) defined in U for t ∈ (a, b), let γ 1 (t) = σ 1 (u(t), v(t)) and γ 2 (t) =

52
Conformal, isometric, and area-preserving maps First fundamental form

σ 2 (u(t), v(t)). Then f (γ 1 (t)) = γ 2 (t) and by proposition 4.2.1,

Z b p
ℓ(γ 1 ) = u′2 E1 + 2u′ v ′ F1 + v ′2 G1 dt
a
Z b p
= u′2 E2 + 2u′ v ′ F2 + v ′2 G2 dt
a
= ℓ(γ 2 ),

so f is a local isometry.
Conversely, let σ ei : U
ei → S 2 be injective regular local parametrisations for i = 1
i
and 2; note we do not assume that U e1 = Ue2 . Let f : S 2 → S 2 be a local isometry,
1 2
−1
and define F : U e1 → U e2 by F = σe2 ◦ f ◦ σ e 1 , which is smooth. By restricting Ue1
and U2 if necessary, we can assume that F is bijective, and hence a diffeomorphism:
e

F
U=U
e1 U
e2
σ2
σ1 = σ
e1 σ
e2

S12 S22
f

Define σ 1 = σ e 1 and σ 2 = σ e 2 ◦ F to be local parametrisations of S12 and S22 respec-

tively, both with U = U e1 as their domain. In particular, σ 2 is a reparametrisation
of σ
e 2 with reparametrisation map F .
By assumption, for any curve γ 1 in S12 , γ 2 = f ◦ γ 1 is a curve with the same length
in S22 . If we set γ 1 (t) = σ 1 (u(t), v(t)), then
Z b p Z b p
u′2 E1 + 2u′ v ′ F 1 + v ′2 G1 dt = u′2 E2 + 2u′ v ′ F2 + v ′2 G2 dt.
a a

As this holds for all curves14 , this implies that the integrands must coincide, and 14
How could you make this abso-
hence lutely rigorous?

u′2 E1 + 2u′ v ′ F1 + v ′2 G1 = u′2 E2 + 2u′ v ′ F2 + v ′2 G2 .

Letting (u0 , v0 ) ∈ U , then taking

1. γ 1 (t) = σ 1 (u0 + t, v0 ) implies E1 = E2 ,

2. γ 1 (t) = σ 1 (u0 , v0 + t) implies G1 = G2 ,

3. γ 1 (t) = σ 1 (u0 + t, v0 + t) implies E1 + 2F1 + G1 = E2 + 2F2 + G2 , and hence

F1 = F2 ;

as required. ■
Example 4.4.10. Notice that with this theorem, examples 4.1.4 and 4.1.8 immedi-
ately imply example 4.4.8.

Corollary 4.4.11. A local isometry is a locally conformal map, and locally preserves
area.

Proof. Angles and area can be expressed using the first fundamental form. A
local isometry implies there are local parametrisations with the same coefficients
of the first fundamental form, and hence the same angles and areas. ■
We have seen that maps which preserve length necessarily preserve angles and area,
and maps which preserve angles do not necessarily preserve lengths (take the map R2 →
R2 which scales everything by a factor of 2). We are left to consider maps which preserve

53
Conformal, isometric, and area-preserving maps First fundamental form

area. We will concern ourselves with just looking at a special example of such a map.

Theorem 4.4.12 (Archimedes’ Tombstone Theorem). Let S2 be the unit sphere and
C 2 the unit circular cone, and define the smooth map
!
x y
f : S \{(0, 0, ±1)} → C : (x, y, z) 7→ p
2 2
,p ,z .
x2 + y 2 x2 + y 2

Then f is area preserving.

We considered this map in example 4.3.2, it is the map which projects the sphere
minus the north and south poles radially to the cylinder from the z-axis, see figure 48.
Archimedes proved two important results about the sphere inscribed by a cylinder. The
first is that the ratio of the surface area of the sphere to the surface area of the curved
part of the cylinder is 1 : 1, and the second is that the ratio of the volume of the sphere
to the volume of the cylinder is 2 : 3. He did this in a treatise called On the sphere and
the cylinder, and he was so pleased with it, that he asked for a sphere and cylinder to
be inscribed on his tombstone15 , hence the name of the theorem above. 15
The Roman politician and ora-
Of course, Archimedes used very clever and detailed elementary methods to prove tor Cicero claimed to have found the
tomb with its geometrical inscription
his results—we have the power of calculus and so will be able to give a much shorter
in the first century CE near Syra-
proof. Note that theorem 4.4.12 is more general than saying that the areas of the sphere cuse, Sicily (almost 3 centuries after
and cylinder are the same, it says that if R ⊂ U is any open subset in the domain of a Archimedes lived). The location is no
local parametrisation of S2 , then the area of σ(R) and of f (σ(R)) is the same. longer known.

Proof. Parametrising the sphere minus the poles by

σ : U → S2 : (θ, φ) 7→ (cos(θ) cos(φ), cos(θ) sin(φ), sin(θ)),

where U = (−π/2, π/2) × R, then σe = f ◦ σ : U → C 2 is a parametrisation of the

cylinder given by
e (θ, φ) = (cos(φ), sin(φ), sin(θ)).
σ

By example 4.1.7, the first fundamental form of the sphere is

E = 1, F = 0, and G = cos2 (θ).

For the cylinder, we have

e θ = (0, 0, cos(θ)) σ
σ e φ = (− sin(φ), cos(φ), 0),

and so
e = cos2 (θ),
E Fe = 0, and G
e = 1.

These are not the same, but, EG − F 2 = cos2 (θ) = E

eGe − Fe2 , so for any open
R ⊂ (U )
ZZ p ZZ q
Aσ (R) = EG − F 2 dθdφ = E e − Fe2 dθdφ = Aσ̃ (R),
eG
R R

as claimed. ■

Application 4.4.13 (Cartography). Making accurate maps of the Earth is very

difficult—in fact, later in the course, we will prove that it is mathematically impos-
sible to accurately represent lengths, since for any region on a sphere, there does
not exist an isometry to the plane. Figure 51. The Mercator map
It is possible to map the Earth conformally, the one of the most common map projection. Image: Lars H. Ro-
hwedder (CC BY-SA 3.0 DEED).

54
Conformal, isometric, and area-preserving maps First fundamental form

projections—the Mercator projection—does exactly this, see figure 51. This pro-
jection does not represent areas very well, and famously, on a Mercator map, Green-
land appears to be as big as the whole of continental Africa, while in reality, it has
less that 1/14th the land area. Some anachronistic criticisms of this projection say
that this was a deliberate attempt by 16th century colonialists to inflate the relative
importance of Europe and Africa. This is not true—the colonialists needed good
maps for navigation, which in those days meant travelling by sea using a compass,
and to take bearings, they needed maps which worked well with angles.
On the other hand, if we wanted to create a map which represented areas well,
we can use Archimedes’ Tombstone Theorem: project the Earth to a cylinder, and
then using example 4.4.8, the cylinder is locally isometric to the plane, so we can
unroll the cylinder to get a map, see figure 52. The trade-off for a map which
represents areas well is that it distorts lengths and angles, so it wouldn’t be of
much use for navigation.

Figure 52. The map of the Earth obtained from Archimedes’ Tombstone Theorem.

55
 Curvature of surfaces

5 Curvature of surfaces
In this section, we will start to think about how to measure the degree to which a surface
is curved, and the way that it curves. For example the ellipsoid and the hyperbolic
paraboloid are clearly both curved, but the way in which they curve is quite different.

Figure 53. The ellipsoid and hyperbolic paraboloid are both curved, but in qualitatively
different ways.

For curves, the way we measured curvature was by looking at the rate of change of
the unit tangent vector, or equivalently, of the unit normal vector. For surfaces, we will
mirror this approach by looking at the rate of change of a unit tangent vector to the
surface. For the same reasons as discussed at the start of section 2.2 (see figure 10), to
have a single normal direction to a 2D surface in Rn , we have to restrict to n = 3, ie
surfaces in three dimensions.

5.1 The Gauss map

Assumption 5.1.1. In this section, all surfaces live in R3 .

In R3 , a vector is normal, that is to say perpendicular to, a regular surface if it is

perpendicular to its tangent plane. By proposition 3.3.3, the tangent plane is spanned
by {σ u , σ v }, so it suffices to choose the normal vector to be perpendicular to both of
these vectors. Such a vector is given exactly by the cross product. Regularity ensures
this cross product is non-zero, see definition 3.3.5. Therefore we can normalise to get a
unit normal vector everywhere, which motivates the following definition.

Definition 5.1.2. Let σ : U → R3 be a regular local parametrisation of a surface

S 2 , then the unit normal to σ to p = σ(u0 , v0 ) is the vector

σu × σv
Np = (u0 , v0 ).
∥σ u × σ v ∥

Of course, there are two choices for a unit normal, this definition ensures that
{σ u , σ v , Np } is a right handed basis for R3 , shown in figure 54.

Example 5.1.3. Consider the sphere S2 parametrised by

Figure 54. The cross product of
σ(θ, φ) = (cos(φ) cos(θ), cos(φ) sin(θ), sin(φ)). σ u and σ v is perpendicular to the
surface.
Then the tangent plane is spanned by

σ θ = (− cos(φ) sin(θ), cos(φ) cos(θ), 0)

σ φ = (− sin(φ) cos(θ), − sin(φ) sin(θ), cos(φ)),

56
The Gauss map Curvature of surfaces

and we can compute

σ θ × σ φ = (cos2 (φ) sin(θ), cos2 (φ) cos(θ),

cos(φ) sin(φ)(cos2 (θ) + sin2 (θ)))
= cos(φ)(cos(φ) sin(θ), cos(φ) cos(θ), sin(φ)).

This vector has norm

∥σ θ × σ φ ∥ = cos(φ),

and so
Np = (cos(φ) sin(θ), cos(φ) cos(θ), sin(φ)).

Notice that in this very special case, we have Np (θ, φ) = σ(θ, φ), which makes
sense considering the geometry of the sphere, see figure 55.
The unit normal associates to every point in U a unit vector to σ(U ). This is not
quite the same as saying that it associates a unit vector to every point of the surface
S 2 = σ(U ) since σ is not necessarily injective. It is useful to have a unit normal
associated to each point of the surface, and so we make the following definition.
Figure 55. The unit normal to
the unit sphere.
Definition 5.1.4. A smooth surface S 2 is orientable if it can be covered by regular
local parametrisations {σ i : Ui → S 2 }i∈I such that

1. for all points p ∈ S 2 , if p = σ i (ui , vi ) = σ j (uj , vj ); where i, j ∈ I, (ui , vi ) ∈

Ui , and (uj , vj ) ∈ Uj ; then
p = Np ,
N(i) (j)

where N(i) is the unit normal of σ i and N(j) is the unit normal of σ j .

2. The vector Np varies continuously as p varies over S 2 .

Remark 5.1.5.
• If σ is injective and S 2 = σ(u) then S 2 is orientable.

• If the image of σ intersects transversally, which is to say, like the Klein bottle
in example 3.2.2(d) or in figure 45, then σ is not orientable.
Example 5.1.6. The sphere is orientable, since we can cover it with local parametri-
sations such that they all have the unit normal pointing outwards (or all pointing
inwards).
Example 5.1.7. The Möbius strip, which is discussed in exercise 5 on sheet 4 can
be covered by the parametrisation σ : R × − 12 , 12 given by

σ(θ, v) = (cos(θ) − v cos(θ) cos(θ/2), sin(θ) − v sin(θ) cos(θ/2), v sin(θ/2)).

Starting at a point p, and moving the unit normal along the surface around the
loop once until you return to the other side, the result is the unit tangent vector
pointing in the opposite direction—so there is no continuous choice of unit normal,
and the Möbius strip is not orientable, see figure 56.

If a surface S 2 is orientable then we can make a smooth choice of normal vector

over the whole surface which does not depend on the parametrisation, ie we can choose
Nσ(u0 ,v0 ) so that it depends only on p = σ(u0 , v0 ) ∈ S 2 , and not on (u0 , v0 ) ∈ U . This
allows us to make the following definition.
Figure 56. A Möbius strip is a
non-orientable surface.

57
The Gauss map Curvature of surfaces

Definition 5.1.8. Let S 2 be an orientable regular surface and Np a smooth choice

of unit normal vector for all p ∈ S 2 . Then the Gauss map of S 2 is the smooth
map
G : S 2 → S2 : p 7→ Np .

In the case that S 2 is orientable and covered by a single specified local parametrisa-
tion σ, by convention, we choose Np to be the unit normal defined in definition 5.1.2.
Assumption 5.1.9. For the rest of this section we assume that all surfaces are
orientable.
Example 5.1.10.

The Gauss map of a

1.
plane is constant.

The image of the Gauss

map of a generalised
2.
cylinder is (part of) a
great circle.

3. By example 5.1.3, the Gauss map of the sphere with our usual choice of
parametrisation is the identity map G : S2 → S2 .

Example 5.1.11 (Adapted from question A6 on the 2019 Summer exam). Compute
the image of the Gauss map of the surface parametrised by σ(u, v) = (u, v, u2 +v 2 ). Figure 57. The unit normal
This surface is a paraboloid. Let’s first answer the question by reasoning entirely to the paraboloid with parametri-
geometrically, and then verify our answer by explicit computation. sation σ(u, v, u2 + v 2 ) is inward-
pointing.
The unit normal Np is either inward or outward facing. By definition of the cross-
product, {σ u , σ v , Np } forms a right-handed basis for R3 , and so we can deduce
from figure 57. Viewing more of the paraboloid, we can see that in the middle, the
unit normal points straight up, and as you move further away, it points more and
more horizontal, see figure 58. Since the gradient of a parabola is unbounded but
finite at all points, the normal will get arbitrarily close to horizontal, but never
reach it. Together with the fact that the paraboloid is rotationally symmetric, it
follows that the image of the Gauss map is the open upper hemisphere.
We can verify this by direct computation. The partial derivatives of σ are

σ u = (1, 0, 2u)
Figure 58. The unit normal to a
σ v = (0, 1, 2v), cross-section of the paraboloid.

so to compute the unit normal, we have

σ u × σ v = (−2u, −2v, 1) ∥σ u × σ v ∥ = 1 + 4(u2 + v 2 );

p
⇒

therefore,
(−2u, −2v, 1)
Np = p .
1 + 4(u2 + v 2 )

58
The Weingarten map and the curvatures of a surface Curvature of surfaces

Since the z-coordinate is always positive, the normal always points up, so the image
lies in the upper hemisphere. Rewriting u = r cos(θ) and v = r sin(θ), we see that

(−2r cos(θ), −2r sin(θ), 1)

Np = √ ,
1 + 4r2

and, as θ varies, this sweeps around a full circle. Moreover, at r = 0 and taking
the limit as r → ∞, the z-coordinate is, respectively,

1 1
√ = 1 and lim √ = 0,
1 + 4r2 r=0
r→∞ 1 + 4r2

so the Gauss map is surjective onto the open upper hemisphere. Using either
method, the answer is G(σ(R2 )) = {(x, y, z) ∈ S2 | z > 0}.

5.2 The Weingarten map and the curvatures of a surface

The Gauss map of an orientable surface encodes the way the normal varies across the
surface. Following the general strategy to define curvature of curves, it makes sense to
try to define the curvature of a surface by looking at the rate of change of the normal
across the surface. As surfaces are two dimensional, this is not going to be a single
number, because it has to encode the rate of change of the normal in two independent
directions.
One way to understand the rate of change of the normal would be to look at the
differential of the Gauss map. This defines a linear map

dp G : Tp S 2 → TG(p) S2 ,

which, in local coordinates, is given by a 2 × 2 matrix.

Before analysing the behaviour of this map rigorously, it may help first to get a
qualitative understanding for what it is doing. On the left in the figure below, we have
plotted part of the negative paraboloid parametrised by σ(u, v) = (u, v, −(u2 + v 2 )),
together with its tangent plane at p = σ(0, 0), and the vectors σ u (blue) and σ v (purple).
On the right is the unit sphere with the image of the Gauss map shown in orange, the
tangent plane to the sphere at G(σ(0, 0)), and the images of σ u and σ v under the
differential of the Gauss map.

Recall from proposition 4.3.5 that these vectors are given by

dp G(σ u ) = (G ◦ σ)u = (Np )u and dp G(σ v ) = (G ◦ σ)v = (Np )v ,

so these vectors are exactly the derivatives of the normal vector in the u and v directions.

59
The Weingarten map and the curvatures of a surface Curvature of surfaces

In this case, the paraboloid curves away from σ(0, 0) quite quickly in all directions, so
dp G stretches the vectors.
If we do the same with a different surface which is less curved—but still curved in
the same way, then dp G shrinks vectors rather than stretching them. This can be seen
in the image of the Gauss map itself, which covers a much smaller portion of the sphere.

The effect of the differential of the Gauss map when the paraboloid curves up rather
than down in all directions is to stretch all the vectors by the same amount as in the
first case, but also reverse their direction, as the normal now varies by the same amount,
but in the opposite direction.

For a plane, we have already seen that the Gauss map is constant, so its differential
is zero—this corresponds, in some sense, to the curvature of the plane being zero.

60
The Weingarten map and the curvatures of a surface Curvature of surfaces

Finally, if we take a surface like the hyperbolic paraboloid which curves up in one
direction, and down in another, the differential of the Gauss map encodes this by leaving
one vector pointing in the same direction, but reversing the other. One way to think of
this is that it sends a right handed basis for Tp S 2 to a left handed basis for TG(p) S2 . We
will see later that this corresponds to negative curvature.

Having developed some intuition for the geometry of the Gauss map and its dif-
ferential, let’s take a more mathematical approach. The first thing to note is that by
example 5.1.3, the normal to the unit sphere at the point p is the point p itself (see
figure 55). This means that TG(p) S2 is the plane perpendicular to
2
NSG(p) = G(p) = Np

which is perpendicular to Tp S 2 . In other words

TG(p) S2 = Tp S 2 .

This allows us to make the following definition.

Definition 5.2.1. Let S 2 be an orientable surface, then the Weingarten map

(sometimes also called the shape operator) is given by16 16
The negative sign here is a
convention which will make some
−dp G : Tp S 2 → Tp S 2 . of the expressions we obtain later
more geometrically intuitive, see re-
mark 5.4.5.
By proposition 4.3.5, the Weingarten map is determined by its image on a basis for
Tp S 2 and in terms of {σ u , σ v } it is given by

− dp G(σ u ) = −(G ◦ σ)u = −(Np )u , and

− dp G(σ v ) = −(G ◦ σ)v = −(Np )v ,

The following lemma will help us perform calculations with the Weingarten map.

Lemma 5.2.2. With notation as above, we have

(Np )u · σ u = −Np · σ uu , (Np )u · σ v = −Np · σ uv ,

(Np )v · σ u = −Np · σ vu , (Np )v · σ v = −Np · σ vv .

61
The Weingarten map and the curvatures of a surface Curvature of surfaces

Proof. By definition, Np is perpendicular to σ u and σ v , so Np ·σ u = 0 = Np ·σ v .

Differentiating Np · σ u = 0 with respect to u gives

0 = (Np )u · σ u + Np · σ uu ,

and with respect to v gives

0 = (Np )v · σ u + Np · σ uv ,

which imply the first two equalities. The second two follow similarly by differenti-
ating 0 = Np · σ v . ■

Proposition 5.2.3. The Weingarten map is symmetric with respect to the first
fundamental form (in other words it is self-adjoint), which means that for any
w1 , w2 ∈ Tp S 2
Ip (−dp G(w1 ), w2 ) = Ip (w1 , −dp G(w2 )).

Proof. By the linearity of dp G and Ip , it suffices to check the claim for w1 , w2 ∈

{σ u , σ v }. If w1 = w2 the claim is clear, so we just need to check w1 = σ u and
w2 = σ v (the other case follows by the symmetry of Ip ). We can compute:

Ip (−dp G(σ u ), σ v ) = −dp G(σ u ) · σ v

= −(Np )u · σ v lemma 5.2.2
= Np · σ vu
= −(Np )v · σ u
= −dp G(σ v ) · σ u
= Ip (σ u , −dp G(σ v )).

Corollary 5.2.4. The Weingarten map −dp G : Tp S 2 → Tp S 2 is given by a symmet-

ric matrix with respect to the basis {σ u , σ v }, and therefore has real eigenvalues17 . 17
Geometrically, the presence of
non-real complex eigenvalues for a
linear map from the plane to itself in-
This symmetric matrix is of central importance, and we will compute it explicitly
dicates that the map rotates the plane
shortly. First we can use this corollary to extract geometric information about the (and perhaps scales it). This result
surface. shows that dp G does not rotate the
plane.
Definition 5.2.5. Let S 2 be an orientable surface, and let p ∈ S 2 . Then the eigen-
values of the Weingarten map −dp G, κ1 and κ2 , are called the principal curva-
tures of S 2 at p. If κ1 ̸= κ2 then the principal directions are the corresponding
eigenvectors t1 and t2 .

Later, when we consider the curvature of curves lying in a surface, we will be able
to give a concrete geometric interpretation for these eigenvalues and vectors. For now,
we can content ourselves that −dp G measures the rate of change of the unit normal to
a surface, and the eigenvalues encode key information about this linear map. Given the
eigenvalues there are two other natural choices of real numbers we can compute, which
turn out to carry significant geometric information.

Definition 5.2.6. With notation as above, the Gaussian curvature of S 2 at p is

the determinant of the Weingarten map

K = det(−dp G) = κ1 κ2 ,

62
The Weingarten map in local coordinates Curvature of surfaces

and the mean curvature is half the trace of the Weingarten map

1 κ1 + κ2
H= trace(−dp G) = ,
2 2
which can also be thought of as the average of the principal curvatures.

Without an explicit matrix representation of −dp G, it can be tricky to compute

these curvatures, so we will now turn our attention to that task.

5.3 The Weingarten map in local coordinates

To understand the Weingarten map, and inspired by proposition 5.2.3, we make the
following definition.

Definition 5.3.1. Let S 2 be an orientable surface with regular local parametrisation

σ : U → S 2 . The second fundamental form of σ at p = σ(u0 , v0 ) is

IIp : Tp S 2 × Tp S 2 → R : (w1 , w2 ) 7→ Ip (−dp G(w1 ), w2 ).

By proposition 5.2.3, the second fundamental form is a symmetric bilinear form 18 18

Recall, a symmetric bilinear form
on the tangent plane of S 2 . Note, in general it is not an inner product because it may on a real vector space V is a map
⟨−, −⟩ : V × V → R which satisfies
fail to be positive definite.
⟨x + ay, z⟩ = ⟨x, z⟩ + a⟨y, z⟩ and
Example 5.3.2. If S 2 is a plane, we have already seen that the Gauss map is ⟨x, y⟩ = ⟨y, x⟩ for all x, y, z ∈ V and
a ∈ R.
constant, so the differential of G (and hence the Weingarten map) is the 0 map.
Therefore, for any w1 , w2 ∈ Tp S 2

IIp (w1 , w2 ) = Ip (−dp G(w1 ), w2 ) = Ip (0, w2 ) = 0 · w2 = 0.

Remark 5.3.3. In order to understand the geometric meaning of the second funda-
mental form, it is useful to consider the associated quadratic form

QIIp : Tp S 2 → R
: w 7→ IIp (w, w) = Ip (−dp G(w), w),

which essentially measures how much the Weingarten map moves w.

As with the first fundamental form, it is often useful to express the the second
fundamental form as a matrix with respect to the basis {σ u , σ v } of Tp S 2 .

Definition 5.3.4. Let S 2 be an orientable surface with regular local parametrisation

σ : U → S 2 . The coefficients of the second fundamental form of σ are

L := Ip (−dp G(σ u ), σ u ),
M := Ip (−dp G(σ u ), σ v ) = Ip (−dp G(σ v ), σ u ),
N := Ip (−dp G(σ v ), σ v )

so that !
L M
IIp = .
M N

Beware, we are using capital N here for two very different quantities which should not
be confused: Np is the unit normal to a surface and is a vector, while N is a coefficient
of the second fundamental form, and hence a scalar.

63
The Weingarten map in local coordinates Curvature of surfaces

Proposition 5.3.5. The coefficients of the second fundamental form are given by

L = Np · σ uu ,
M = Np · σ uv = Np · σ vu ,
N = Np · σ vv

Proof. These are straightforward to compute using lemma 5.2.2, for example

L = Ip (−dp G(σ u ), σ u )
= −(Np )u · σ u
= Np · σ uu .

Note that we already saw the expressions for M in the proof of proposition 5.2.3.
The proof of N is similar. ■
Example 5.3.6. We can compute the second fundamental form of a surface of rev-
olution. Let
σ(u, v) = (f (u) cos(v), f (u) sin(v), g(u)),

where f (u) > 0 and f ′2 + g ′2 = 1 (ie this is the surface of revolution of a regular
unit speed curve). Then

σ u = (f ′ cos(v), f ′ sin(v), g ′ ) and σ v = (−f sin(v), f cos(v), 0),

so

σ u ×σ v = (−f g ′ cos(v), −f g ′ sin(v), f f ′ ) =⇒ ∥σ u ×σ v ∥ = f 2 (g ′2 + f ′2 ) = f,

p

and the unit normal is

σu × σv
Np = = (−g ′ cos(v), −g ′ sin(v), f ′ ).
∥σ u × σ v ∥

We need the second derivatives of σ in order to compute the coefficients of the

second fundamental form, which are given by

σ uu = (f ′′ cos(v), f ′′ sin(v), g ′′ )
σ uv = (−f ′ sin(v), f ′ cos(v), 0)
σ vv = (−f cos(v), −f sin(v), 0).

Finally,

L = Np · σ uu = −f ′′ g ′ (cos2 (v) + sin2 (v)) + f ′ g ′′ = f ′ g ′′ − f ′′ g ′

M = Np · σ uv = f ′ g ′ (cos(v) sin(v) − cos(v) sin(v)) = 0
N = Np · σ vv = f g ′ (cos2 (v) + sin2 (v)) = f g ′ ,

so the second fundamental form is

!
f ′ g ′′ − f ′′ g ′ 0
IIp = .
0 f g′

We can consider a couple of special cases.

1. The unit sphere S2 is a surface of revolution with f (u) = cos(u) and g(u) =

64
The Weingarten map in local coordinates Curvature of surfaces

sin(u). Therefore its second fundamental form is given by

!
1 0
IIp = .
0 cos2 (u)

2. The circular cylinder with radius 1 is a surface of revolution with f (u) = 1

and g(u) = u, so !
00
IIp = .
01

Remark 5.3.7 (Geometric interpretation of the coefficients of the second fundamental

form). Consider a regular surface patch σ : U → R3 , and for (u0 , v0 ) ∈ U . Let’s
think about how the surface behaves near σ(u0 , v0 ). If we consider a small pertur-
bation ∆u, ∆v near (u0 , v0 ), then we can use Taylor’s Theorem to write the surface
near σ(u0 , v0 ) as

σ(u0 + ∆u, v0 + ∆v) = σ(u0 , v0 )

+ σ u ∆u + σ v ∆v
1
+ (σ uu ∆u2 + 2σ uv ∆u∆v + σ vv ∆v 2 )
2
+ terms of order 3

This means that, to first order, the surface is approximated by its tangent plane,
spanned by σ u , σ v , which should not be surprising. To second order, we can
measure how much the surface differs from its tangent plane by taking the dot
product of σ(u0 + ∆u, v0 + ∆v) − σ(u0 , v0 ) with Np . Recall from the geometry of
the dot product, this is the orthogonal projection of the surface to Np , see figure 59.
This gives

(σ(u0 + ∆u, v0 + ∆v) − σ(u0 , v0 )) · Np

1
= (σ uu · Np ∆u2 + 2σ uv · Np ∆u∆v + σ vv · Np ∆v 2 ) + · · ·
2
1
= (L∆u2 + 2M ∆u∆v + N ∆v 2 ) + · · · ,
2
so the coefficients of the second fundamental form measure the extent to which the
surface curves away form its tangent plane.
As promised, we can now return to the Weingarten map, and in particular, how to
express it as a symmetric matrix in terms of the coordinates {σ u , σ v }.
Figure 59. The amount a surface
Theorem 5.3.8. In terms of the basis {σ u , σ v } of Tp S 2 , the Weingarten map is deviates from its tangent plane
given by the matrix is measured by the second funda-
−dp G = I−1
p IIp .
mental form.

Writing out these matrices using the coordinates of the fundamental forms gives
! !
1 G −F L M
−dp G =
EG − F 2 −F E M N
!
1 GL − F M GM − F N
= .
EG − F 2 EM − F L EN − F M

65
Curvature in local coordinates Curvature of surfaces

Proof. Write !
a11 a12
−dp G =
a21 a22

so that

− dp G(σ u ) = a11 σ u + a12 σ v , (5.1)

− dp G(σ v ) = a21 σ u + a22 σ v . (5.2)

Taking the dot product of eq. (5.1) with σ u gives

−dp G(σ u ) · σ u = a11 σ u · σ u + a12 σ v · σ u = a11 E + a12 F ;

but by lemma 5.2.2

−dp G(σ u ) · σ u = −(Np )u · σ u = Np · σ uu = L,

so a11 E + a12 F = L. Taking the dot product of eq. (5.1) with σ v yields

a11 F + a12 G = M.

Playing the same trick with eq. (5.2) gives

a21 E + a22 F = M and a21 E + a22 F = N.

We can combine these four equations in the matrix equation

! ! !
E F a11 a21 L M
=− ,
F G a12 a22 M N

from which the theorem follows. ■

5.4 Curvature in local coordinates
Corollary 5.4.1. Let S 2 be an orientable surface, and σ : U → S 2 a regular local
parametrisation. Let E, F , G, L, M , and N be the coefficients of the first and
second fundamental forms of σ. Then

a) The Gaussian curvature is given by

LN − M 2
K= .
EG − F 2

b) The mean curvature is given by

EN − 2F M + GL
H= .
2(EG − F 2 )

c) The principal curvatures satisfy det(IIp − κIp ) = 0 and are roots of

κ2 − 2Hκ + K = 0,
√
hence κ1,2 = H ± H 2 − K.

d) The principal directions are λσ u + µσ v where

λ2 (EM − F L) + λµ(EN − GL) + µ2 (F N − GM ) = 0,

66
Curvature in local coordinates Curvature of surfaces

ie λ and µ are the roots of

 
µ2 −λµ λ2
det  E F G  = 0.
 
L M N

Proof.
a) We have

det(IIp ) LN − M 2
K = det(−dp (G)) = det(I−1
p IIp ) = = .
det(Ip ) EG − F 2
!
1 GL − F M GM − F N
b) Since −dp G = , we have
EG − F 2 EM − F L EN − F M

1 1 1
H= trace(−dp G) = ((GL − F M ) + (EN − F M ))
2 2 EG − F 2
EN − 2F M + GL
= .
2(EG − F 2 )

c) By definition, the principal curvatures, as eigenvalues, satisfy

0 = det(κI − (−dp G)) = det(κI − I−1

p IIp ),

multiplying by det(Ip ) and using det(AB) = det(A) det(B) gives the result.
For the second claim, the principal curvatures satisfy the quadratic

0 = (κ − κ1 )(κ − κ2 ) = κ2 − (κ1 + κ2 )κ + κ1 κ2 = κ2 = κ2 − 2Hκ + K.

Finally, apply the quadratic formula to compute κ1 and κ2 .

d) An eigenvector satisfies
! ! !
λ λ λ
−dp G = I−1
p IIp = κi .
µ µ µ

Multiplying on the left by Ip gives

! !
Lλ + M µ Eλ + F µ
= κi .
Mλ + Nµ F λ + Gµ

Taking the ratio of the two components of these vectors, it follows that

Lλ + M µ Eλ + F µ
= .
Mλ + Nµ F λ + Gµ

Cross multiplying and expanding out the brackets gives

(Lλ + M µ)(F λ + Gµ) = (M λ + N µ)(Eλ + F µ),

λ LF + λµ(LG + M F ) + µ2 M G = λ2 M E + λµ(M F + N E) + µ2 N F,
2

so
0 = λ2 (EM − F L) + λµ(EN − GL) + µ2 (F N − GM ).

The final claim follows by expanding out the determinant explicitly.

■

67
Curvature in local coordinates Curvature of surfaces

Example 5.4.2. We have seen that any plane with an orthonormal parametrisation
has ! !
10 0 0
Ip = and IIp = ,
01 0 0

see examples 4.1.4 and 5.3.2. Hence the Weingarten map is

!
00
−dp G = ,
00

which has eigenvalues κ1 = κ2 = 0. These are the same, so the principal directions
are not defined, and the Gaussian and mean curvatures are

κ1 + κ2
K = κ1 κ2 = 0 and H = = 0.
2
Example 5.4.3. In example 5.3.6, we computed that for the surface of revolution of
a regular curve (f (u), 0, g(u)) for f (u) > 0 and f ′2 + g ′2 = 1 (so it has unit speed)
given by
σ(u, v) = (f (u) cos(v), f (u) sin(v), g(u)),

we have

σ u = (f ′ cos(v), f ′ sin(v), g ′ ) and σ v = (−f sin(v), f cos(v), 0).

Therefore, the coefficients of the first fundamental form are

E = σ u · σ u = f ′2 + g ′2 = 1,
F = σ u · σ v = 0,
G = σv · σv = f 2 ,

so !
1 0
Ip = .
0 f2

In that example, we also computed that

!
f ′ g ′′ − f ′′ g ′ 0
IIp = .
0 f g′

Therefore, the Weingarten map is given by

−dp G = I−1
p IIp
! ! !
1 f2 0 f ′ g ′′ − f ′′ g ′ 0 1 f 2 (f ′ g ′′ − f ′′ g ′ ) 0
= 2 = 2
f 0 1 0 fg ′ f 0 f g′
!
f ′ g ′′ − f ′′ g ′ 0
=
0 g ′ /f

We can either use the definitions directly, or the formulae in corollary 5.4.1. As the
Weingarten map is diagonal for this local parametrisation σ, it’s probably easier
just to use the definitions.

• The principal curvatures are the eigenvalues of −dp G, so

g′
κ1 = f ′ g ′′ − f ′′ g ′ and κ2 = ,
f

68
Curvature in local coordinates Curvature of surfaces

the order of κ1 and κ2 is not important.

• Assuming the principal curvatures do not coincide, the principal directions

are the eigenvectors of −dp G, which up to a non-zero scalar are

t1 = σ u and t2 = σ v .

• The Gaussian curvature is given by

(f ′ g ′′ − f ′′ g ′ )g ′
K = κ1 κ2 = .
f

Recall that by assumption, f ′2 + g ′2 = 1; differentiating this with respect to

u gives
0 = 2f ′ f ′′ + 2g ′ g ′′ =⇒ g ′ g ′′ = −f ′ f ′′ .

Substituting this above gives

−f ′2 f ′′ − f ′′ g ′2 f ′′ (f ′2 + g ′2 ) f ′′
K= =− =− .
f f f

• The mean curvature is given by

κ1 + κ2 f ′ g ′′ − f ′′ g ′ + g ′ /f f (f ′ g ′′ − f ′′ g ′ ) + g ′
H= = =
2 2 2f
We can use this example to compute the curvatures of the sphere.
Example 5.4.4. The sphere of radius R with local parametrisation

σ(θ, φ) = (R cos(φ) cos(θ), R cos(φ) sin(θ), R sin(φ))

can be viewed as the surface of revolution of the curve (R cos(φ),

 φ  0, R sin(φ)). So
 φ 
that this has unit speed, we can reparametrise it by R cos , 0, R sin ,
R R
so φ φ
f (φ) = R cos and g(φ) = R sin .
R R
The derivatives of these are
φ φ
f ′ = − sin g ′ = cos
R R
1 φ 1 φ
f = − cos
′′
g = − sin
′′
R R R R
so by example 5.4.3

1 φ 1 1 cos R
φ
1
φ

• κ1 = cos2 + sin2 = , and κ2 = φ = .
R cos R


R R R R R R
• These are the same, so the principal directions are not defined.

• The Gaussian curvature is

1
K = κ1 κ2 = .
R2

• The mean curvature is

κ1 + κ2 1 2 1
 
H= = = .
2 2 R R

69
Curvature in local coordinates Curvature of surfaces

Remark 5.4.5. In the field of geometric analysis, and particularly geometric flow,
geometers study how surfaces can be deformed using their mean curvature. They
define the mean curvature vector H = HNp and move every point in the surface
with a velocity H, see figure 60. The choice of the sign of H is essential.
In 2002–3, Grigori Perelman published a proof of the Poincaré Conjecture, the
first of the Clay Institute’s $1,000,000 millennium problems to be solved. This
conjecture says that any three-manifold which is closed, connected, and simply
connected is homeomorphic to S3 . His proof relied on a more complicated version
of mean curvature flow, called Ricci curvature flow.

Figure 60. In two dimensions, mean curvature flow is called curve shortening flow, and is
slightly easier to visualise.

We have so far seen the sphere with constant positive Gaussian curvature, and the
plane with constant zero Gaussian curvature. We could wonder what a surface with
constant negative Gaussian curvature looks like. We can use the example of the surface
of revolution to construct such a surface.
Example 5.4.6. By example 5.4.3, to find a surface with constant Gaussian curva-
ture K = −1, say, it suffices to find a regular curve γ(u) = (f (u, 0, g(u))) which
f ′′
satisfies f ′2 + g ′2 = 1 and K = − = −1. This second equation is equivalent to
f
f ′′ = f , which has general solution19 f (u) = Aeu + Be−u , for some constants A 19
This is a homogeneous second or-
and B. We just want to find one surface, so for simplicity, let’s choose f (u) = eu . der linear ODE.

Next, we need to find g(u) which satisfies 1 = g ′2 + f ′2 = g ′2 + e2u , in other words

√
g ′ = 1 − e2u (up to a sign which we do not care about). Notice that this only
has a solution for u ≤ 0. This can be solved by directly integrating

du 1
Z p
g(u) = 1 − e2u du; set x = eu , so = .
dx x
1
Z p
= 1 − x2 dx (5.3)
x
1 − x2 1
Z
= √ dx
x 1 − x2
1 1 −x
Z Z
= √ dx + √ dx
x 1 − x2 1 − x2
1 1 d p 1 dx 1
Z Z
= √ dx + 1 − x2 dx set y = , so = − 2.
x 1−x 2 dx x dy y
−1
1 1
Z r p
= −y
1− 2 dy + 1 − x2
y y 2
1
Z p
=− p dy + 1 − x2
y −1
2
p
= −arcosh(y) + 1 − x2 + c we are looking for a solution,
p
= 1 − x2 − arcosh(x−1 ) so let’s set c = 0
p
= 1 − e2u − arcosh(e−u ).

70
Geometric interpretation of the Gaussian curvature Curvature of surfaces

√
Therefore, the surface of revolution of γ(u) = (eu , 0, 1 − e2u − arcosh(e−u )), for
u < 0, has constant negative curvature K = −1, see figure 61.
This surface is called the pseudosphere, which means “false sphere”, because it
pretends to be a sphere by having constant non-zero curvature, but manifestly is
not a sphere.

Remark 5.4.7. The unit speed curve we found in this example is called a tractrix,
and it admits another nice geometric description. If we set z = g(u) and x = f (u),
then by (5.3) √
dz 1 − x2
=
dx x
so the equation of the tangent line to the curve at P = (x0 , z0 ) is

1 − x20
p
z − z0 = (x − x0 ).
x0
Figure 61. The pseudosphere is
a surface with constant negative
Let Q = (0, z1 ) be the point where the tangent line meets the z-axis, see figure 62.
curvature.
At Q we have

1 − x20
p q
z1 − z0 = (0 − x0 ) = − 1 − x20 .
x0

Now, the distance between P and Q is

|P Q| = (0 − x0 )2 + (z1 − z0 )2
p
q
= 20 + (1 − x
x
 20 ) = 1


is constant. This can be interpreted as follows. A tractor tows a bail of hay using
a rigid rod of length 1. The tractor drives in a straight line up the z-axis, and the
rod starts perpendicular to the direction of motion, lying along the x-axis. The Figure 62. The tangent to the
path traced out by the bail of hay is exactly the tractrix, see figure 63. tractrix at P meets the z-axis at
Q.

Remark 5.4.8. The surface on the cover of these lecture notes is called Dini’s
surface, named after Ulisse Dini, an Italian mathematician. It is another example
of a surface with constant negative curvature, and can be thought of as a twisted
version of the pseudosphere (more precisely, a generalised helicoid generated by the
tractrix). It can be parametrised by
    v  
σ(u, v) = a cos(u) sin(v), a sin(u) sin(v), a cos(v) + ln tan + bu ,
2
for constants a and b. It has Gaussian curvature
1
K=− ,
a2 + b2

so has curvature −1 when a2 + b2 = 1.

5.5 Geometric interpretation of the Gaussian curvature
Figure 63. The tractrix is the
In remark 5.3.7, by using Taylor’s Theorem, we saw that a surface is approximated by path followed by an object towed
its tangent plane up to first order, and to second order by terms related to the second by a rigid rod.
fundamental form. We have also seen that to compute the various types of curvature
of a local parametrisation σ, we have to compute the second derivatives. This suggests
that surfaces defined by second order polynomials are, in a sense, the simplest surfaces
which can express the various possible different types of curvature. Such surfaces are
the quadratic surfaces we looked at in example 3.4.4. Let us look again at these surfaces

71
Geometric interpretation of the Gaussian curvature Curvature of surfaces

from the point of view of curvature.

Consider the quadratic surface which is the set of solutions to the polynomial z =
ax + by 2 for a, b ∈ R constants. This has local parametrisation
2

σ(u, v) = (u, v, au2 + bv 2 ),

and partial derivatives

σ u = (1, 0, 2au), σ v = (0, 1, 2bv),

σ uu = (0, 0, 2a), σ uv = (0, 0, 0), σ vv = (0, 0, 2b).

at the point p = (0, 0, 0)

E = 1, F = 0, G = 1,

Np = (0, 0, 1), Figure 64. An elliptic point.

and
L = 2a, M = 0, N = 2b.

Therefore, the Weingarten map is given by

!−1 ! !
10 2a 0 2a 0
−dp G = = ,
01 0 2b 0 2b

which has eigenvalues κ1 = 2a and κ2 = 2b.

It follows that near a point p on a surface S 2 with principal curvatures κ1 and κ2 , Figure 65. A hyperbolic point.
the surface looks like the quadratic surface, to second order

1
z= (κ1 x2 + κ2 y 2 ).
2
This motivates the following definitions.

Definition 5.5.1. Let S 2 be an orientable surface and p ∈ S 2 a point in the surface.

• If κ1 κ2 > 0 then p is an elliptic point, and the surface looks locally like an
elliptic paraboloid. Figure 66. A parabolic point.

• If κ1 κ2 < 0 then p is a hyperbolic point, and the surface looks locally like
a hyperbolic paraboloid paraboloid.

• If κ1 = 0 and κ2 ̸= 0 then p is a parabolic point, and the surface looks

locally like a parabolic cylinder.
Figure 67. A planar point.
• If κ1 = κ2 = 0 then p is a planar point, and the surface looks locally like a
plane.

These definitions are illustrated in figures 64 to 67.

We can see what this kind of classification looks like in practice by looking at the
example of a torus.

Example 5.5.2. Let’s look at the torus, T2 , which can be viewed as the surface
of revolution of a circle which is off-set along the x-axis. In particular, we can
Figure 68. The torus is the sur-
consider the circle centred at (2, 0, 0) with radius 1, as shown in figure 68. This
face of revolution of an offset cir-
has unit speed parametrisation cle.

γ(u) = (2 + cos(u), 0, sin(u)),

72
Geometric interpretation of the Gaussian curvature Curvature of surfaces

and so the torus has parametrisation σ : R2 → R3 given by

σ(u, v) = ((2 + cos(u)) cos(v), (2 + cos(u)) sin(v), sin(u)),

and shown in figure 69. Setting f (u) = 2 + cos(u) and g(u) = sin(u), then using
example 5.4.3, we can compute

κ1 = f ′′ g ′ − f ′ g ′′ = − cos2 (u) − sin2 (u) = −1

and
g′ cos(u)
κ2 = − =− .
f 2 + cos(u)
We can classify the points on the torus according to definition 5.5.1. Since κ1 ̸= 0,
the torus has no planar points. We have

cos(u)
κ1 κ2 = ,
2 + cos(u)

and the denominator is strictly positive for all u, so the type of point just depends
on the sign of cos(u), in particular, a point p is
π π
1. elliptic if and only if − < u − 2kπ < for some k ∈ Z,
2 2
π
2. parabolic if and only if u − kπ = for some k ∈ Z,
2
π 3π
3. hyperbolic if and only if < u − 2kπ < for some k ∈ Z.
2 2
Notice that this only depends on the u, which makes sense because we could imagine
rotating the torus in the v direction, where it is symmetric, so we would not expect
the curvature properties to change. This classification is shown in figure 70.

Figure 69. A torus, the v coordi-

nate parametrises the torus along
the blue circles, and the u coordi-
nate along the red circles.

Figure 70. The elliptic (green),

parabolic (blue), and hyperbolic
(orange) points of a torus.

73
 Curvature of curves on surfaces

6 Curvature of curves on surfaces

In the last section, we developed several notions of curvature of a surface using the Wein-
garten map. In this chapter, we will take a completely different approach to curvature
on a surface, namely studying the curves on a surface, and their curvature, and see that
we recover exactly the same notions of curvature, particularly in terms of the princi-
pal curvatures. This confirms that we really have hit on a fundamental an important
property of surfaces.
In fact, we will decompose the curvature of a curve on a surface into two parts: the
amount that the curve curves within the surface itself (or if you prefer, in the directions
of the tangent plane), and the amount the curve curves in the direction perpendicular
to the surface.
At the end of the section, we will consider, in detail, those curves which do not curve
in the direction of the surface itself. These are called geodesics and generalise the idea
of straight lines in the plane to general curved surfaces.
Throughout, we maintain the assumption that all surfaces are orientable and live in
R .
3

6.1 Normal and geodesic curvature

Let γ(s) = σ(u(s), v(s)) be a curve in a surface S 2 with local parametrisation σ : U →
S 2 , and assume that it is parametrised with unit speed. As we have seen before, γ̇ =
u̇σ u + v̇σ v , which lies in the tangent plane, and therefore {Nγ(s) , γ̇(s)} is orthonormal.
We can extend this to a right-handed orthonormal basis of R3 by considering

{Nγ(s) , γ̇(s), Nγ(s) × γ̇(s)}.

Recall (lemma 1.3.6) that since ∥γ̇∥ = 1, we have that γ̇ and γ̈ are perpendicular. It
follows that there are real functions κn (s) and κg (s) such that

γ̈(s) = κn Nγ(s) + κg (Nγ(s) × γ̇(s)). (6.1)

Definition 6.1.1. For a unit speed curve γ in a surface parametrised by σ, the

normal curvature is κn = γ̈ · Nγ(s) , and the geodesic curvature is κg =
γ̈ · (Nγ(s) × γ̇(s)). For a regular curve, the normal and geodesic curvatures are
defined to be the normal and geodesics of any unit speed reparametrisation.

Recall that the curvature of an unit speed curve, γ, thought of as just a space curve,
is κ = ∥γ̈∥, ie the rate of change of the tangent to γ. The normal curvature measures the
amount of this curvature in the direction of the normal to the surface, and the geodesic
curvature measures the curvature in the direction of Nγ(s) × γ̇(s) ∈ Tγ(s) S 2 .

Lemma 6.1.2. For a unit speed curve, with notation as above, we have

κ2 = κ2n + κ2g ,

and κn = κ cos(ψ) and κg = ±κ sin(ψ), where ψ is the angle between γ̈ and Nγ(s) .

Proof. By the orthonormality of {Nγ(s) , Nγ(s) × γ̇(s)}, eq. (6.1) immediately

implies that κ2 = ∥γ̈∥2 = κ2n + κ2g . Using the fact that γ̈ = κn (recall n is the
principal normal of γ), we have

κn = γ̈ · Nγ(s) = κn · Nγ(s) = κ∥n∥∥Nγ(s) ∥ cos(ψ) = κ cos(ψ),

and hence κg = ±κ sin(ψ) by the first claim. ■

74
Normal and geodesic curvature Curvature of curves on surfaces

Figure 71. The normal and geodesic curvatures measure the orthogonal component of γ̈
parallel and perpendicular to the surface.

Drawing γ̈ in the plane spanned by {Nγ(s) , Nγ(s) × γ̇(s)}, this lemma becomes
clearer, see figure 71. We can, of course, find a way to compute these curvatures for
regular curves without first computing a unit speed reparametrisation. We take the
geodesic curvature first.

Proposition 6.1.3. Let γ be a regular curve on a surface (not necessarily

parametrised by arc length), then the geodesic curvature at p = γ(t) is given
by
γ ′′ · (Np × γ ′ )
κg = .
∥γ ′ ∥3

Proof. Let γ e=γ e (s) be the unit speed reparametrisation of γ, where s = s(t) be
d d
the arc length of γ; and as usual, denote with ′ and by˙. Using γ
e (s(t)) = γ(t)
dt ds
we can show that
d ds ˙
γ ′ (t) = γ e (s(t)) = e (s),
γ
dt dt
and  2
d ds ˙ d2 s ˙ ds
 
γ ′′ (t) = e (s) = 2 γ
γ e (s) + γ¨
e (s).
dt dt dt dt
Using these we can compute
 2 ! 
d2 s ˙ ds ds ˙
 
γ · (Np × γ ) =
′′ ′
γ (s) + ¨
γ (s) · Np × γ (s)
dt2 dt dt
e e e

 0  2
 
ds  d s ˙
2
ds
 :
˙ 
¨
= e (s) · (N
γ p×γe (s)) + e˙ (s))
e (s) · (Np × γ
γ
dt dt2  dt
3
ds

= κg .
dt

ds
Recalling from remark 1.2.3 that = ∥γ ′ ∥, the claim follows. ■
dt
Somewhat surprisingly, the normal curvature at p of a curve in a surface depends
only on the tangent vector to the curve at p, and not on the curve itself, see figure 72.

Proposition 6.1.4 (Meusnier’s Theorem). Let S 2 be a surface with regular local

parametrisation σ, and let γ(s) = σ(u(s), v(s)) be a unit speed curve in a surface Figure 72. All curves in a sur-
face through p with the same tan-
passing through p ∈ S 2 . Then the normal curvature of γ at p is given by gent vector v have the same nor-
mal curvature.
κn = Lu̇2 + 2M u̇v̇ + N v̇ 2 = γ̇ ⊺ IIp γ̇ = Ip (−dp G(γ̇), γ̇).

75
Normal and geodesic curvature Curvature of curves on surfaces

In particular, this only depends on the second fundamental form of the surface, and
the tangent vector to γ at p.

Proof. This follows from a straightforward calculation:

dγ̇ d
κn = Np · γ̈ = Np · = Np · (σ u u̇ + σ v v̇)
ds ds
= Np · ((σ uu u̇ + σ uv v̇)u̇ + σ u ü + (σ vu u̇ + σ vv v̇)v̇ + σ v v̈)
= Lu̇2 + 2M u̇v̇ + N v̇ 2 ,

using proposition 5.3.5 and the fact Np ·! σ u = 0 = Np · σ v . In terms of the basis

u̇
{σ u , σ v }, γ̇ = σ u u̇ + σ v v̇ is given by , and multiplying out, we get
v̇
! !
L M u̇
(u̇ v̇) = Lu̇2 + 2M u̇v̇ + N v̇ 2 ,
M N v̇

which gives the second equality. The third equality is simply the definition of the
second fundamental form. ■
Therefore, it makes sense to talk about the normal curvature in a direction at p,
rather than of a curve through p.

Corollary 6.1.5. Let γ be a regular (not necessarily unit speed) curve in a surface
with local parametrisation σ. Then the normal curvature of γ at p = γ(t) is

κn = Ip (−dp G(t), t) = t⊺ IIp t,

γ′
where t = is the unit tangent to γ at p.
∥γ ′ ∥

Example 6.1.6. Consider the unit sphere S2 and let γ be a line of latitude, in other
words γ is a unit speed reparametrisation of the coordinate curve γ(t) = σ(θ0 , t)
for some fixed θ0 , where
Figure 73. The curve γ(t) =
σ(θ, φ) = (cos(φ) cos(θ), sin(φ) cos(θ), sin(θ)). σ(θ0 , φ) on the sphere.

This is shown in figure 73. Recall from example 5.3.6, that this parametrisation of
the sphere has second fundamental form
!
1 0
IIp = .
0 cos2 (θ)

We can write explicitly that

γ(t) = (cos(t) cos(θ0 ), sin(t) cos(θ0 ), sin(θ0 )),

Figure 74. A cross-section of the
so γ ′ (t) = (− sin(t) cos(θ0 ), cos(t) cos(θ0 ), 0), and sphere, showing γ̈ in relation to
Nγ(s) and Nγ(s) × γ̇(s). No-
∥γ ′ (t)∥ = cos(θ0 ). tice that for the parametrisation
in this example, the unit normal
is inward-pointing.
Alternatively, we can write γ ′ (t) in terms of {σ θ , σ φ } as

γ ′ (t) = 0σ θ (θ0 , t) + σ φ (θ0 , t) = σ φ (θ0 , t),

76
The principal and normal curvatures Curvature of curves on surfaces

and hence
σ φ (θ0 , t)
t(t) = .
cos(θ0 )
Therefore,

1 N (t, θ0 ) cos2 (θ0 )

κn = t⊺ IIp t = σ ⊺φ IIp σ φ = = = 1.
cos2 (θ 0) cos (θ0 )
2 cos2 (θ0 )

This is illustrated in figure 74.

We have seen that the normal curvature of a curve in a surface depends only on its
tangent vector, but there are many curves through a given point on a surface with a
given tangent vector (recall figure 73). Among these curves there is a special choice of
curve.

Definition 6.1.7. Let S 2 be a surface and p ∈ S 2 and point in S 2 . Let Π be a plane

which contains p and is perpendicular to Tp S. Any curve which parametrises the
intersection Π ∩ S 2 is called a normal section to S 2 at p.

This is illustrated in figure 75.

Proposition 6.1.8. Let γ be a normal section to S 2 at p, then at p, its curvature

is κ = ±κn , and hence κg = 0.
Figure 75. A normal section of a
We will see in section 6.3 that normal sections are not the only curves in S 2 which surface.
have this property.
Proof. We can assume that γ is parametrised with unit speed. Since the trace of
γ lies in Π, so does γ̈. By construction, Π is perpendicular to Tp S 2 , and so contains
Np . Therefore, γ̈ and Np are both perpendicular to γ̇, and hence parallel. The
angle between them is ψ = 0 or π, and so by lemma 6.1.2, κn = ±κ and κg = 0. ■

6.2 The principal and normal curvatures

For this reason, it would be useful to have an orthonormal basis for Tp S 2 , say {e1 , e2 },
because then we would have a simple way to describe the directions at p: any direction
corresponds to a unique unit vector, which in turn can be written as cos(θ)e1 + sin(θ)e2
for some unique θ ∈ [0, 2π). We cannot use the basis we have been using up until now for
this purpose, because in general {σ u , σ v } is not orthonormal. To find a natural choice
of orthonormal basis, we will use the following fact from linear algebra.

Lemma 6.2.1. Let A : V → V be a self-adjoint linear map on a two-dimensional

inner product space (V, ⟨−, −⟩, then V has an orthonormal basis of eigenvectors of
A.

Proof. Since A is self-adjoint, it has real eigenvalues λ1 and λ2 . If λ1 = λ2 then

every vector in V is an eigenvector, and we can choose any orthonormal basis to
be an eigenbasis.
Assume therefore that λ1 ̸= λ2 , and let e1 and e2 be unit eigenvectors correspond-
ing to λ1 and λ2 respectively, we just need to show they are orthogonal. Indeed

λ1 ⟨e1 , e2 ⟩ = ⟨λ1 e1 , e2 ⟩ = ⟨Ae1 , e2 ⟩

= ⟨e1 , Ae2 ⟩ = ⟨e1 , λ2 e2 ⟩ = λ2 ⟨e1 , e2 ⟩.

Since λ1 ̸= λ2 , the only way this is possible is if ⟨e1 , e2 ⟩ = 0, so e1 and e2 are

orthogonal. ■

77
The principal and normal curvatures Curvature of curves on surfaces

Proposition 6.2.2. Let κ1 and κ2 be the principal curvatures of a surface S 2 at a

point p ∈ S 2 .

1. If κ1 ̸= κ2 then principal vectors t1 and t2 corresponding to κ1 and κ2 are

orthogonal.

2. If κ1 = κ2 then every non-zero tangent vector to S 2 at p is an eigenvector of

the Weingarten map.

In either case, we can choose and orthonormal basis {t1 , t2 } for Tp S 2 of eigenvectors
of the Weingarten map.

Proof. We already saw in proposition 5.2.3 that the Weingarten map is self-
adjoint on (Tp S 2 , Ip (−, −)), so we can apply lemma 6.2.1, and in the case κ1 =
κ2 , choose any orthonormal basis to be an eigenbasis. Otherwise, choose unit
principal directions {t1 , t2 }, which are eigenvectors of the Weingarten map, to be
the orthonormal basis for Tp S 2 . ■
Using this basis, we can compute the normal curvature in terms of the principal
curvatures.

Theorem 6.2.3. (Euler’s Theorem) Let v be a direction in Tp S 2 , ie a unit vector,

and fix an orthonormal basis of Tp S 2 consisting of eigenvectors of the Weingarten
map, {t1 , t2 }. Then the normal curvature of S 2 in the direction v is given by

κn = κ1 cos2 (θ) + κ2 sin2 (θ),

where θ is the angle between v and t1 measured anticlockwise.

Proof. We can write v as

v = cos(θ)t1 + sin(θ)t2 ,

where θ is the angle between v and t1 . The normal curvature in this direction can
be computed using proposition 6.1.4:

κn = Ip (−dp G(v), v)
= Ip (−dp G(cos(θ)t1 + sin(θ)t2 ), cos(θ)t1 + sin(θ)t2 )
= Ip (cos(θ)κ1 t1 + sin(θ)κ2 t2 ), cos(θ)t1 + sin(θ)t2 )
= κ1 cos2 (θ) + κ2 sin2 (θ),

where we have used the face that t1 and t2 are eigenvectors of −dp G with eigen-
values κ1 and κ2 . ■
For the first time, this now allows us to give a properly geometric interpretation of
the principal curvatures and principal vectors of a surface.

Corollary 6.2.4. The principal curvatures at a point p are the maximum and mini-
mum values of the normal curvature of all unit speed curves in the surface passing
through p. The unit principal vectors are the tangent vectors to the curves which
realise these maximum and minimum values.

This also motivates the choice of notation t1 and t2 for principal directions.
Proof. First consider the case that κ1 ̸= κ2 , without loss of generality, we may

78
Geodesics Curvature of curves on surfaces

assume that κ1 > κ2 . Then

κn = κ1 cos2 (θ) + κ2 sin2 (θ) = κ1 (1 − sin2 (θ)) + κ2 sin2 (θ)

= κ1 − (κ1 − κ2 ) sin2 (θ),
| {z }
>0

which implies κn ≤ κ1 with equality if and only if θ = 0 or π, ie if and only if v

is parallel to t1 . A similar argument shows that κn ≥ κ2 with equality if and only
if v is parallel to t2 .
In the case that κ1 = κ2 , the κn = κ1 = κ2 and every unit vector is a principal
vector. ■
Remark 6.2.5. Often, this is taken to be the definition of the principal curvatures,
and then one must prove that they are in fact the eigenvalues of the Weingarten
map.
Euler’s theorem also provides a different interpretation of the name mean curvature.
By definition, it is the mean or average of the principal curvatures, but we can also prove
that it is the mean of the normal curvature over all directions in the surface.

Corollary 6.2.6. Let κn (θ) be the normal curvature of a surface S 2 in the direction
making an angle θ to the principal direction t1 . Then the mean curvature is the
mean of κn over all directions, ie

1 2π
Z
H= κn (θ)dθ.
2π 0

Proof. This is a simple computation using theorem 6.2.3.

1 2π
1 2π
Z Z
κn (θ)dθ = κ1 cos2 (θ) + κ2 sin2 (θ)dθ
2π 0 2π 0
:0 :0
1 2π
1 +
cos(2θ) 1 −
cos(2θ)
Z
 
= κ1 + κ2 dθ,
2π 0 2 2

here we have used that the integral of cos over a full period vanishes. Continuing,

1 κ1 + κ2
Z 2π
= dθ
2π 0 2
1 κ + κ2
=  1
2π
2π 2



κ1 + κ2
=
2
= H,

as required. ■

6.3 Geodesics
We have spent quite a bit of effort in the last section looking at the normal curvature.
In particular, we saw that it is somewhat independent of the curve itself (see proposi-
tion 6.1.4), and so can be used to re-derive curvature properties of the surface itself, such
as the principal and mean curvatures.
We turn now to think about the geodesic curvature. Unlike the normal curvature,
the geodesic curvature is highly dependent on the curve and the way it sits inside the
surface. Recall that the geodesic curvature measures the amount the curve bends in the
direction of the tangent plane to the surface. In particular, we will see that curves with
zero geodesic curvature—curves which in some sense follow the contours of the surface

79
Geodesics Curvature of curves on surfaces

most closely—have very special properties. They play the role that straight lines do in
the plane, and (locally) minimise the distance between points in the surface.
Inspired by the definition of the geodesic curvature for unit speed curves, κg =
γ̈(s) · (Nγ(s) × γ̇(s)), we make the following definition.

Definition 6.3.1. A smooth curve γ on a surface is a geodesic if γ ′′ is perpendicular

to the surface everywhere.

The term geodesic comes from the archaic word geodesy, which was the scientific
name for the study of the size and shape of the Earth. Being perpendicular to the
surface is equivalent to being parallel to the unit normal Np , see figure 76. The first Figure 76. A curve γ is a
observation is that, up to multiplying the parameter of the curve by a constant, geodesics geodesic if γ ′′ is always perpendic-
have unit speed. ular to the surface.

Proposition 6.3.2. Geodesics have constant speed.

d ′ 2
Proof. Since the square of the speed of γ is ∥γ ′ ∥2 = γ ′ · γ ′ , we have ∥γ ∥ =
dt
2γ · γ . If γ is a geodesic then γ is parallel to Np , which is, by definition,
′ ′′ ′′

perpendicular to the tangent plane Tp S 2 , which contains γ ′ . Therefore, γ ′ · γ ′′ = 0

and the derivative of the square of the speed is zero, and hence the square of the
speed, and the speed as well, is constant. ■

Proposition 6.3.3. A curve γ in a surface is a geodesic if and only if it has constant

speed and κg = 0 everywhere.

Proof. If γ is a geodesic then it has constant speed by proposition 6.3.2, and

using proposition 6.1.3
γ ′′ · (Np × γ ′ )
κg = = 0, Figure 77. Lines in the plane are
∥γ ′ ∥3 geodesics.
since γ ′′ is parallel to Np which is perpendicular to Np × γ ′ , so the dot product
in the numerator vanishes.
Conversely, if κg = 0 then γ ′′ ·(Np ×γ ′ ) = 0 implies γ ′′ is perpendicular to Np ×γ ′ .
Since ∥γ ′ ∥2 is constant, γ ′ · γ ′′ = 0 implies that γ ′′ is perpendicular to γ ′ . Hence
γ ′′ is parallel to Np , so γ is a geodesic. ■
Example 6.3.4. In the plane, any straight line γ parametrised with constant speed
is a geodesic, since γ ′′ = 0 is perpendicular20 to Tp S 2 , see figure 77. 20
This may sound odd, because the
More generally, any line which is contained in a surface, if parametrised with con- zero vector lies in Tp S 2 , but recall
the definition of v being perpendicu-
stant speed, will be a geodesic. This applies, in particular to ruled surfaces, where
lar to w is v · w = 0, so the zero
we can find lots of lines in the surface, like generalised cones and cylinders, see vector is perpendicular to every vec-
figure 78. In the case of doubly ruled surfaces such as the hyperboloid of one sheet tor (including itself).
(see example 3.4.6), we can immediately find two geodesics through every point in
the surface, see figure 40.

We have already seen in proposition 6.1.8 that if γ is a normal section to S 2 at p,

then the geodesic curvature of γ at p is zero. The next proposition follows immediately.

Proposition 6.3.5. Let γ : (a, b) → S 2 be a constant speed curve in a surface S 2

which is a normal section to S 2 at γ(t) for all t ∈ (a, b), then γ is a geodesic.

Example 6.3.6. Recall that a great circle on a sphere is the intersection of the
sphere with a plane through its centre. Since such a plane is perpendicular to the
sphere at all points of intersection, any constant speed parametrisation of a great

80
Geodesics Curvature of curves on surfaces

Figure 78. The lines in a ruled surface are geodesics.

circle is a geodesic in S2 , see figure 79.

Example 6.3.7. The intersection of a generalised cylinder with a plane Π perpen-

dicular to its axis can be parametrised as a geodesic, see figure 80. Indeed, if
σ(u, v) = (f (u), g(u), v) with f ′2 + g ′2 = 1, then it is straightforward to check that
Np = (g ′ , −f ′ , 0) which is perpendicular to the z-axis, so Np is parallel to Π, and
the intersection is a normal section at every point.
These examples we have seen so far: lines and normal sections in surfaces, fall a long
way short of allowing us to determine all geodesics in a surface. The following theorem
gives a characterisation of geodesics as solutions to a system of PDEs.

Theorem 6.3.8 (Geodesic equations). A curve γ(t) = σ(u(t), v(t)) on a surface

parametrised by σ is a geodesic if and only if Figure 79. Great circles are
geodesics on a sphere.
d 1
(Eu′ + F v ′ ) = (Eu u′2 + 2Fu u′ v ′ + Gu v ′2 ), and
dt 2
d 1
(F u′ + Gv ′ ) = (Ev u′2 + 2Fv u′ v ′ + Gv v ′2 ).
dt 2

Proof. By definition, γ is a geodesic if and only if γ ′′ is perpendicular to σ u and

σ v , ie γ ′′ · σ u = 0 = γ ′′ · σ v , at all points along γ. By writing γ ′ = u′ σ u + v ′ σ v ,
we can see that γ ′′ · σ u = 0 if and only if

d ′
 
0= u σu + v σv · σu
′
dt
Figure 80. Normal sections of a
d d
 
cylinder perpendicular to the axis
= (u σ u + v σ v ) · σ u − (u′ σ u + v ′ σ v ) · σ u
′ ′
dt dt are geodesics.

by the Leibniz rule. Note that

(u′ σ u + v ′ σ v ) · σ u = u′ σ u · σ u + v ′ σ v · σ u = Eu′ + F v ′ ,

while
d
σ u = u′ σ uu + v ′ σ uv .
dt
Putting this together, we get

d
0= (Eu′ + F v ′ ) − (u′2 σ u · σ uu + u′ v ′ (σ u · σ uv + σ v σ uu ) + v ′2 σ v · σ uv )
dt | {z } | {z } | {z }
= 12 Eu =Fu = 12 Gu

d 1
= (Eu′ + F v ′ ) − (Eu u′2 + 2Fu u′ v ′ + Gu v ′2 ),
dt 2

81
Geodesics Curvature of curves on surfaces

which implies the first equation. A similar computation, starting with γ ′′ · σ v = 0,

yields the second equation. ■

Corollary 6.3.9. Local isometries map geodesics to geodesics.

Proof. Let f : S12 → S22 be a local isometry. By theorem 4.4.9, there are local
parametrisations σ i : U → Si2 , for i = 1, 2, which have the same first fundamental
forms. The choice of these parametrisations do not affect the normals to S12 and
S22 up to sign, nor γ ′′ where γ is a curve in Si2 , so the geodesics of Si2 are not
effected by the choice of these local parametrisations. Now the corollary follows
from the geodesic equations, since these show that the geodesics depend only on
the coefficients of the first fundamental form, and their derivatives, which are the
same for σ 1 and σ 2 . ■
After expanding out the left hand sides of the geodesic equations and rearranging,
you can check that they are equivalent to
! !
u′′ 1 −1 (Gu − 2Fv )v ′2 − Eu u′2 − 2Ev u′ v ′
= Ip . (6.2)
v ′′ 2 (Ev − 2Fu )u′2 − Gv v ′2 − 2Eu u′ v ′

If we set the right hand side equal to

!
f (u, v, u′ , v ′ )
,
g(u, v, u′ , v ′ )

where f and g are smooth functions, then we can apply the existence and uniqueness of
solutions to systems of ODEs. In the present setting, this goes as follows.

Theorem 6.3.10 (Picard–Lindelöf Theorem). Let f (u, v, u′ , v ′ ) and g(u, v, u′ , v ′ ) be

smooth functions. For every x ∈ R2 , and every y ∈ R2 there exists a unique curve
µ(t) = (u(t), v(t)) defined for t ∈ (−ε, ε), for some ε > 0, such that µ(0) = x,
µ′ = y, and the coordinates u(t) and v(t) satisfy
! !
u′′ f (u, v, u′ , v ′ )
= .
v ′′ g(u, v, u′ , v ′ )

This theorem is stated for first order ODEs with a single function in Introduction
to Dynamical Systems. The generalisation to this setting is relatively straightforward.
The proof is beyond the scope of this course.

Corollary 6.3.11 (Existence and uniqueness of geodesics). Let p ∈ S 2 and v ∈ Tp S 2 ,

then for some ε > 0, there exists a unique geodesic γ(t) = σ(u(t), v(t)) on σ,
defined for t ∈ (−ε, ε), with γ(0) = p and γ ′ (0) = v.

In other words, there exists a unique geodesic through every point, in every direction,
with every speed in S 2 , and in particular a unit speed geodesic through every point in
every direction.
We can use this to characterise all geodesics in some of the previous surfaces which
we have considered.
Example 6.3.12. There is a line parametrised with any speed, in every direction,
through any point in the plane. Hence, all geodesics in the plane are lines.
Example 6.3.13. Every geodesic in a generalised cylinder is the image of a straight
line in the plane under the local isometry in example 4.4.8. This follows from
corollaries 6.3.9 and 6.3.11. In particular, every geodesic in the circular cylinder is

82
Geodesics as shortest paths Curvature of curves on surfaces

• a vertical line (recall example 6.3.4),

• a meridian circle (recall example 6.3.7),

• a circular helix.

These are illustrated in figure 81.

Example 6.3.14. There is a great circle in every direction and through every point
of a sphere, so all geodesics on a sphere are great circles.

Example 6.3.15. Consider a surface of revolution parametrised by

σ(u, v) = (f (u) cos(v), f (u) sin(v), g(u)),

Figure 81. The geodesics on a
with f (u) > 0 and, as usual, f ′2 + g ′2 = 1. Recall from example 5.4.3 that this has circular cylinder.
first fundamental form !
1 0
Ip = .
0 f2

so the geodesic equations become

d 1 0 1  0 0 2f f ′
′+F
(Eu v ′ ) = (
E>
u u′2
+ 2 F
u>
u ′ ′
v + u ) ⇐⇒ u = f f v , and
G>
 v ′2 ′′ ′ ′2
(6.3)
dt 2
0 f2 0 0′ ′ 0
d 1  d 2 ′
(
F u + Gv
′
 ′) = (v u + 2
E> ′2
Fvu v + 
> v v ) ⇐⇒
G
> ′2
(f v ) = 0. (6.4)
dt 2 dt
If we assume that the geodesic has unit speed, then

1 0 f2
1 = ∥γ ∥ = Eu
 + 2
′ 2 ′2
F u v + Gv
′ ′
 ′2 = u′2 + f 2 v ′2 . (6.5)

From these considerations, we can conclude the following.

(6.3)
a) Every unit speed meridian v = v0 is a geodesic since v ′ = 0 =⇒ u′′ = 0, so
u′ is a constant, and by (6.5), equal to 1. Hence γ(t) = σ(t + u0 , v0 ) satisfies
the geodesic equations (6.4) is automatically satisfied. Figure 82. Meridians (blue) and
parallels at the stationary points
b) A parallel u = u0 is a geodesic if and only if f ′ (u0 ) = 0. Indeed, for ( =⇒ ) of f (orange), are geodesics on a
surface of revolution.
(6.5) (6.3)
u = u0 =⇒ 1 = f (u20 )v ′2 =⇒ v ′ = ±(f (u0 ))−1 ̸= 0 =⇒ f ′ (u0 ) = 0.

0 0
Conversely, for (⇐=), f (u0 ) = 0 =⇒ 
′ ′′
u>
 = f7′ v ′2 , so (6.3) holds. Also,
f
(6.5) implies v = ±(f (u0 ))
′ −1
as before, which is constant. Thus, f 2 v ′ is
constant and (6.4) holds. Therefore the geodesic equations hold.
Therefore, a parallel is a geodesic if and only if f is stationary at that point.

c) It is possible to find all geodesics in a regular surface of revolution, although

not necessarily in closed form. See, for example, Example 5 in section 5-5 of
“Differential geometry of curves and surfaces” by Manfredo do Carmo.

The geodesics of type (a) and (b) are shown in figure 82.

6.4 Geodesics as shortest paths

At the start of the previous section we said that geodesics generalise the idea of straight
lines in the plane to curved surfaces. In the plane, lines provide the shortest paths to get

83
Geodesics as shortest paths Curvature of curves on surfaces

from point A to point B. This is not quite the case for geodesics, but they can instead
be though of as locally minimising distances between points.
In some cases like the plane, and more generally, in surfaces with negative Gaussian
curvature, geodesics always minimise distance. However, in some cases, like the sphere,
geodesics in some sense, maximise the distance, see figure 83.
We can make all of this formal using calculus of variation.

Definition 6.4.1. Let γ : (a, b) → S 2 be a constant speed smooth curve. Fix some
closed interval [c, d] ⊂ (a, b). A smooth variation of γ on [c, d] is a one-parameter
family, γ τ , of curves in S 2 given by Figure 83. Part of a great circle
joining two points on the sphere
which goes the long way round is a
(−δ, δ) × (a, b) → S 2 : (τ, t) 7→ γ τ (t),
geodesic which does not minimise
distance.
which satisfies

• γ τ (t) is smooth in τ and t,

• γ 0 (t) = γ(t) for all t ∈ (a, b),

• γ τ (c) = γ(c) and γ τ (d) = γ(d) for all τ ∈ (−δ, δ).

Figure 84. A smooth variation γ τ (light green) of a curve γ (dark green) on [c, d].

An illustration of a smooth variation is shown in figure 84. Given a smooth variation

of a curve, we define the notion of an action (sometimes called the energy), which is a
function which measures how the length of γ τ depends on τ . In fact, this is an example
of a functional, a function which takes a function as its input, ie γ τ .

Definition 6.4.2. Let γ τ be a smooth variation of γ on [c, d], then the action is
the function S : (−δ, δ) → R given by
Z d
  ∂
S(τ ) = ℓ γ τ = γ (t) dt.
[c,d]
c ∂t τ

The fact that geodesics locally minimise distances is encapsulated by the following
theorem.

Theorem 6.4.3. A smooth curve γ in a surface is a geodesic if and only if

d
S(τ ) τ =0 = 0 for all [c, d] ⊂ (a, b) and all smooth variations γ τ of γ on [c, d].
dτ

At first, it may seem to be insufficient to check just the first derivative of the action
vanishes—this guarantees that γ is a stationary point, but what stops it from being a
local maximum, say? The reason is that S has no local maxima. Indeed suppose that γ
locally maximised the length of paths between two points—then we could always make

84
Geodesics as shortest paths Curvature of curves on surfaces

it wiggle and meander more to get a longer path, which would be a contradiction.
Proof. Let σ be a local parametrisation of the surface and set

γ τ (t) = σ(u(τ, t), v(τ, t)).

∂
To simplify notation, but somewhat unusually, we will use ′ to denote when
∂t
convenient. Before attempting the proof, we will rewrite the derivative of the
action in a different form. We can compute

=:g(τ,t)
d d d
Z d Z d z }| {

1/2
S(τ ) = ∥γ ′τ (t)∥dt = Eu′2 + 2F u′ v ′ + Gv ′2 dt
dτ dτ c dτ c
1 d
Z d Z
∂   ∂g
= g(τ, t)1/2 dt = g(τ, t)−1/2 dt,
c ∂τ 2 c ∂τ

where we have used the fact that g is smooth to exchange the derivative and the
integral. We can further compute
′
 ′ ′
∂v ′

∂g ∂E ′2 ∂F ′ ′ ∂G ′2 ′ ∂u ∂u ′ ′ ∂v
= u +2 uv + v + 2Eu + 2F v +u + 2Gv ′
∂τ ∂τ ∂τ ∂τ ∂τ ∂τ ∂τ ∂τ
     
∂u ∂v ∂u ∂v ∂u ∂v
= Eu + Ev u′2 + 2 Fu + Fv u′ v ′ + Gu + Gv v ′2
∂τ ∂τ ∂τ ∂τ ∂τ ∂τ
2
 2 2
∂2v

′ ∂ u ∂ u ′ ′ ∂ v
+ 2Eu + 2F v +u + 2Gv ′
∂τ ∂t ∂τ ∂t ∂τ ∂t ∂τ ∂t
∂u ∂v
= (Eu u′2 + 2Fu u′ v ′ + Gu v ′2 ) + (Ev u′2 + 2Fv u′ v ′ + Gv v ′2 )
∂τ ∂τ
∂2u ∂2v
+ 2(Eu′ + F v ′ ) + 2(F u′ + Gv ′ ) .
∂τ ∂t ∂τ ∂t
Just considering the contribution to the integral coming from these last two terms,
and using integration by parts, we get

1 d 2 2
Z  
′ ∂ u ′ ∂ v
g(τ, t)−1/2 2 (Eu ′
+ F v ) + 2 (F u ′
+ Gv ) dt
2 ∂τ ∂t ∂τ ∂t




c
0 0
  d
∂u


∂v



= g(τ, t) −1/2
(Eu + F v )
′ ′

+ (F u + Gv )
′ ′


∂τ
∂τ t=c
Z d h
∂ i ∂u
− g(τ, t)−1/2 (Eu′ + F v ′ )
c ∂t ∂τ
∂ h i ∂v 
+ g(τ, t)−1/2 (F u′ + Gv ′ ) dt,
∂t ∂τ

∂u ∂v
here, the two partial derivatives and vanish at t = c and t = d by the third
∂τ ∂τ
point in the definition of a smooth variation: γ τ (c) = γ(c) and γ τ (d) = γ(d) are
constant as functions of τ .
∂g
Bringing back in the other terms of , we have shown that
∂τ

d
Z d 
∂u ∂v
S(τ ) = U +V dt,
dτ c ∂τ ∂τ

85
Geodesics as shortest paths Curvature of curves on surfaces

where
1
U (τ, t) = g(τ, t)−1/2 (Eu u′2 + 2Fu u′ v ′ + Gu v ′2 )
2
∂ h i
− g(τ, t)−1/2 (Eu′ + F v ′ ) ,
∂t
and
1
V (τ, t) = g(τ, t)−1/2 (Ev u′2 + 2Fv u′ v ′ + Gv v ′2 )
2
∂ h i
− g(τ, t)−1/2 (F u′ + Gv ′ ) .
∂t
We are now, finally, ready to prove the theorem. First, assume that γ is a geodesic.
The instances U (0, t) = 0 and V (0, t) = 0 are exactly the geodesic equations, using
the fact that g(0, t) = ∥γ ′0 ∥2 is a constant, recall theorem 6.3.8.
d
Hence S(τ ) τ =0 = 0.
dτ Z d 
∂u ∂v
Conversely, assume that U +V dt = 0 when τ = 0 for all smooth
c ∂τ ∂τ
variations of γ. We need to prove that U (0, t) = 0 = V (0, t) for all t ∈ [c, d].
Assume that U (0, t0 ) ̸= 0 for some t0 ∈ [c, d], and suppose that it is positive
at this point. By the continuity of U , there exists some ε > 0 such that for
all t ∈ (t0 − ε, t + ε), U (0, t) > 0. Choose a smooth function ϕ which has the
properties21 that 21
The fact that such a function ex-
ists is not immediately obvious. Such
• ϕ(t) > 0 if t ∈ (t0 − ε, t + ε), and functions do exist, and are called
“bump functions”. One such function
• ϕ(t) = 0 if t ̸∈ (t0 − ε, t + ε). can be constructed as follows. Let
 2
Write γ(t) = σ(u0 (t), v0 (t)) and consider the variation e−1/t t>0
θ(t) = ,
0 t≤0
γ τ (t) = σ(u0 (t) + τ ϕ(t), v0 (t)) =: σ(u(τ, t), v(τ, t)), which has the property that all
derivatives exist and are equal to 0 at
∂u ∂v t = 0. Now set ψ(t) = θ(1+t)θ(1−t),
shown in figure 85. Then = ϕ(t) and = 0. Hence,

∂τ ∂τ and finally, set ϕ(t) = ψ t−tε
0
.
Z d  Z t0 +ε
∂u ∂v
0= U +V dt = U (0, t) ϕ(t) dt > 0,
c ∂τ ∂τ τ =0 t0 −ε | {z }|{z}
>0 >0

where the first equality is the assumption—but this gives a contradiction. The
same contradiction arises for U (0, t0 ) < 0 with ϕ(t) ≤ 0. Thus, U (0, t) = 0. A
similar argument shows that V (0, t) = 0. This means that the geodesic equations
hold, and hence γ is a geodesic. ■

Figure 85. The smooth variation γ τ considered at the end of the proof.

86
Geodesics as shortest paths Curvature of curves on surfaces

Corollary 6.4.4. Let γ be a shortest curve between p1 and p2 in a surface, then it

is a geodesic.

d
Proof. S(τ ) has a global minimum at τ = 0 so S(τ ) = 0. ■
dτ τ =0

The converse is not true, as the example of a great circle on the sphere which ‘goes
the long way round’ shows.
Remark 6.4.5. Shortest paths/geodesics may not exist between two points in a
surface. Take for example S 2 = R2 \{(0, 0)}. Then there is no shortest path from
p1 = (−1, 0) to p2 = (1, 0). The line between these points minimises the distance,
2, but does not stay in the surface since it contains the point (0, 0). On the other
hand, we can find piecewise smooth curves in S 2 with length 2 + ε for all ε > 0.
Let γ ε be the curve which is the line between p1 and p2 except that it traces a
semicircle or radius ε around (0, 0), see figure 86. This curve has length

ℓ(γ ε ) = 2(1 − ε) + πε = 2 + (π − 2)ε,

so limε→0 ℓ(γ ε ) = 2.

ε
γε
(−1, 0) (0, 0) (1, 0)

Figure 86. The path γ ε in R2 \{(0, 0)}.

Remark 6.4.6. If a surface S 2 is a closed subset of R3 , then there always exists a

shortest path between two points, and hence a geodesic between any two points.
Examples of such surfaces include the plane, the sphere, and the torus. In general,
however, geodesics can be very complicated and difficult to compute, see for ex-
ample figure 87. The geodesics on a torus were studied in detail in “The Geodesic
Lines on the Anchor Ring” by Gilbert Ames Bliss, published in the Annals of
Mathematics, vol. 4, no. 1, 1902, pp. 1–21.

Application 6.4.7. Geodesics have dozens of applications, including in engineer-

ing, navigation, graphics rendering, and GPS. They play a central role in Einstein’s
theory of general relativity, where the universe is modelled as a curved 4D surface Figure 87. An example of a com-
(a manifold) called space-time, and objects moving freely under gravity (ie without plicated geodesic on the torus.
other forces acting on them) follow geodesic paths through space-time. The amount
that space-time curves is determined by the amount of matter (really mass-energy)
in the universe, and the way that matter moves is determined by the way that
space-time curves.
From this point of view, there is no such thing as the force of gravity—what we
perceive and feel as a force is just the effect of object following geodesics. This
model gives extremely accurate predictions about the way the universe behaves
on large scales. In 2022, the satellite MICROSCOPE verified one of the founding
principles of general relativity (the weak equivalence principle) to an accuracy of
10−15 .

87
 Gauss’ Theorema Egregium

7 Gauss’ Theorema Egregium

In 1827, Karl Friedrich Gauss proved what he himself described as a “remarkable the-
orem”. At the time, scientific writing was carried out principally in Latin, which is the
language he used to make this remark. That Latin phrase has stuck as the name given
to this result: the Theorema Egregium 22 . 22
Gauss’ papers on the subject have
been translated into English, and are
freely available online. You should be
7.1 The theorem
in a position to try to read and un-
To understand the significance of this theorem, and in particular, what makes it remark- derstand these, if you re interested.
able, we need to think about the difference between what is measured by the first and
second fundamental forms.
We used the first fundamental form to define quantities such as length, angles, and
area. These are geometric properties which depend only on the surface itself, and not
on the way the surface sits in R3 . Put another way, suppose we drew a triangle on a
flat piece of paper—this triangle would have certain edge lengths, certain interior angles
which add to π, and some area. If we now bent and twisted the paper without stretching
it (paper does not usually stretch), then the triangle would become curved on a curved
surface, but its edge lengths, interior angles, and area would stay the same.
On the other hand, the second fundamental form measures how the surface sits and
curves in R3 . You can think of this either in the way that the second fundamental form
is defined in terms of the Weingarten map; or, as we did in remark 5.3.7, view it as
measuring how much the surface curves away from its tangent plane at each point.
This distinction between the types of properties which depend just on the geometry
in the surface itself (ie just on the first fundamental form), and properties which also
depend on the way the surface sits in R3 (ie on the second fundamental form) has a
name.

Definition 7.1.1. A (local) property of a surface is intrinsic if it only depends on

distances and angles measured in the surface, in other words, only on the first fun-
damental form. Other properties, which depend on more than the first fundamental
form, are called extrinsic.

Proposition 7.1.2. A property of a surface is intrinsic if and only if it is preserved

by local isometries.

Proof. By theorem 4.4.9, two surfaces are locally isometric if and only if they
have local parametrisations with the same domain and the same coefficients of the
first fundamental form. Since intrinsic properties can be computed just using the
first fundamental form, the proposition follows. ■
We have already seen some examples of intrinsic properties. Tautologically, lengths
and angles are intrinsic, and so is surface area by definition 4.2.3. In addition, we showed
that geodesics are intrinsic in corollary 6.3.9, since the geodesic equations only involved
the coefficients of the first fundamental form and their derivatives.
Properties which we’ve met and which, at least on the surface,23 appear to be ex- 23
pun intended
trinsic are all of the curvatures which we defined using the Weingarten map: principal
curvatures, Gaussian curvature, and mean curvature. Not only are these defined in terms
of the second fundamental form, they measure the amount the surface curves through
space, and so appear to be the archetype of extrinsic properties. In general, this is
true—with one very notable exception—hence the remarkability of Gauss’ theorem.

Theorem 7.1.3 (Theorema Egregium). The Gaussian curvature depends only on the
first fundamental form.

88
The theorem Gauss’ Theorema Egregium

Corollary 7.1.4. The Gaussian curvature is intrinsic, and hence preserved by local
isometries.

Recall corollary 5.4.1, where we showed that

LN − M 2
K= .
EG − F 2

This theorem means that, while the second fundamental form cannot be computed just
using the first fundamental form, its determinant can be.
The proof of theorem 7.1.3 will involve proving several lemmata.

Lemma 7.1.5. For a local parametrisation σ, we have the following.

1 1
1. σ uu · σ u = Eu 4. σ uu · σ v = Fu − Ev
2 2
1 1
2. σ uv · σ u = Ev 5. σ uv · σ v = Gu
2 2
1 1
3. σ vv · σ u = Fv − Gu 6. σ vv · σ v = Gv
2 2
Proof. These are straightforward to check from the definitions of E, F , and G.
As an example

1 1
Fv − Gu = (σ u · σ v )v − (σ v · σ v )u
2 2
1
=( (·(
σ uv σ(v + σ u · σ vv −
2( (·(
σ uv σ(
v,

2
which is equation (3). ■

For a regular surface in R3 , {σ u , σ v , Np } forms a basis of R3 , and so it is possible

to express vectors such as σ uu , σ uv , and σ vv in terms of this basis, eg σ uu = aσ u +
bσ v + cNp . Using the fact that Np is a unit vector and orthogonal to the other two
basis vectors, we know that the Np components of this vector, c, is given by the first
coefficient of the second fundamental form (see proposition 5.3.5). The other coefficients
are defined to be th Christoffel symbols.

Definition 7.1.6. The Christoffel symbols of a regular local parametrisation,

Γijk , are the remaining coefficients of σ uu , σ uv , and σ vv in terms of the basis
{σ u , σ v , Np }. In particular,

σ uu = Γ111 σ u + Γ211 σ v + LNp ,

σ uv = Γ112 σ u + Γ212 σ v + M Np ,
σ vu = Γ121 σ u + Γ221 σ v + M Np ,
σ vv = Γ122 σ u + Γ222 σ v + N Np .

Note that all three of i, j, and k in the Christoffel symbol Γijk are indices, none of
them represents a power. The way to remember how the indices work is to match up
u ↔ 1 and v ↔ 2, and the lower two indices correspond to the partial derivatives on the
left hand side, which the upper index corresponds to the partial derivative of σ u or σ v
on the right hand side.
Note also that, since σ uv = σ vu , the Christoffel symbols are symmetric in their
lower indices, ie Γijk = Γikj .

89
The theorem Gauss’ Theorema Egregium

Lemma 7.1.7. The Christoffel symbols can be expressed in terms of E, F , and G,

and their partial derivatives.

Proof. We can combine the identities in lemma 7.1.5 with the definitions of the
Christoffel symbols. Taking identity (1) gives

1
Eu = σ uu · σ u = (Γ111 σ u + Γ211 σ v + LNp ) · σ u
2
= Γ111 E + Γ211 F + 0,

and identity (4) gives

1
Fu − Ev = σ uu · σ v = (Γ111 σ u + Γ211 σ v + LNp ) · σ v
2
= Γ111 F + Γ211 G + 0.

In matrix form, this becomes

! ! !
1
2 Eu E F Γ111
= ,
Fu − 21 Ev F G Γ211

or equivalently,
! ! !
Γ111 1 G −F 1
2 Eu
= .
Γ211 EG − F 2 −F E Fu − 21 Ev

In the same way, we can deduce from identities (2) and (5) that
! ! ! !
Γ112 Γ121 1 G −F 1
2 Ev
= = ,
Γ212 Γ212 EG − F 2 −F E 1
2 Gu

and from (3) and (6) that

! ! !
Γ122 1 G −F Fv − 12 Gu
= .
Γ222 EG − F 2 −F E 1
2 Gv

Lemma 7.1.8. The determinant of the second fundamental form can be written as

LN − M 2 = σ uu · σ vv − σ uv · σ uv + T,

where T is a sum of terms involving just E, F , and G, and their derivatives.

Proof. In order to bring in the Christoffel symbols, we can rewrite LN − M 2 in

terms of the unit normal, and then use the definition of the Γijk ’s.

LN − M 2 = LNp · N Np − M Np · M Np
= (σ uu − Γ111 σ u − Γ211 σ v ) · (σ vv − Γ122 σ u − Γ222 σ v )
− (σ uv − Γ112 σ u − Γ212 σ v ) · (σ uv − Γ112 σ u − Γ212 σ v )
= σ uu · σ vv − Γ122 σ uu · σ u − Γ222 σ uu · σ v − Γ111 σ vv · σ u − Γ211 σ vv · σ v
+ Γ111 Γ122 E + Γ111 Γ222 F + Γ211 Γ122 F + Γ211 Γ222 G
− σ uv · σ uv + 2Γ112 σ uv · σ u + 2Γ212 σ uv · σ v

2
2
− Γ112 E − 2Γ112 Γ212 F − Γ212 G.

90
The theorem Gauss’ Theorema Egregium

We have already seen that all these extra terms can be written in terms of the
coefficients of the first fundamental form in lemmas 7.1.5 and 7.1.7. ■
We can now prove Gauss’ Theorema Egregium, theorem 7.1.3.
Proof. By corollary 5.4.1 and lemma 7.1.8, we know that

σ uu · σ vv − σ uv · σ uv + T
K= ,
EG − F 2

where T is a sum of terms involving just E, F , and G, and their derivatives. Finally,
notice that, using lemma 7.1.5,

1 1
 
Fv − Gu − Evv = (σ vv · σ u )u − (σ uv · σ u )v
2 u 2
=( σ uvv
(( ·σ
((u + σ vv · σ uu − (σ uvv
(( ·σ
((u − σ uv · σ uv

= σ uu · σ vv ,

so
Fv − 12 Gu u − 12 Evv + T


K= ,
EG − F 2
■
By keeping more careful track of all the terms which appear in this proof, it is
possible to deduce the following formula for the Gaussian curvature in terms of just the
coefficients of the first fundamental form, and their derivatives.

Corollary 7.1.9. The Gaussian curvature of a surface with coefficients of the first
fundamental form E, F , an G is given by
 
− 1 Evv + Fuv − 12 Guu 12 Eu Fu − 12 Ev 0 12 Ev 1
2 Gu
1  2
K= 1
Fv − 2 Gu E F 1
− 2 Ev E F .

(EG − F 2 )2

1 1
2 Gv F G 2 Gu F G

This is often too cumbersome to be useful for computing K by hand, but there are
some special cases.

• If F = 0, then

1
    
∂ G ∂ E
K=− √ √ u + √ v .
2 EG ∂u EG ∂v EG

• If E = 1 and F = 0, then

1 ∂ 2 √ 
K = −√ G
G ∂u2

Remark 7.1.10. While lemma 7.1.5 to corollary 7.1.9 is examinable, you will not
be expected to remember any of the formulae or identities in the exam. If you are
expected to use any of these formulae, they will be given to you.
Remark 7.1.11. You may recognise one of these expressions from walking around
the maths department. In the fourth and fifth floors, there are several equations
written on the walls, some of which relate to this course. Which of the following
do you recognise? (One of these is the Global Gauss-Bonnet Theorem, which we
have not yet seen)

91
The theorem Gauss’ Theorema Egregium

In section 5.1, we made the assumptions that all surfaces would be orientable and
live in R3 , and we have maintained these assumptions up to this point. The reason was
that to define first the Gauss and then the Weingarten map, we needed a well-defined
notion of unit normal vector. With Gauss’ Theorema Egregium, we now know that these
assumptions are not necessary to define the Gaussian curvature.

Definition 7.1.12. Let σ : U → Rn be a regular local parametrisation for n ̸= 3,

then the Gaussian curvature of σ is defined by corollary 7.1.9.

Alternatively, if σ(U ) ⊂ Rn is locally isometric to a surface S 2 ⊂ R3 , then its

Gaussian curvature is the same as the Gaussian curvature of S 2 .
Example 7.1.13. Consider the family of surfaces Sθ2 in R3 with local parametrisa-
tion

σ θ (u, v) = ( cos(θ) sinh(v) sin(u) + sin(θ) cosh(v) cos(u),

− cos(θ) sinh(v) cos(u) + sin(θ) cosh(v) sin(u),
u cos(θ) + v sin(θ)),

where (u, v) ∈ R2 and the parameter θ ∈ R.

A standard, but somewhat tedious, computation shows that the first fundamental
form is given by !
cosh2 (v) 0
Iθp = .
0 cosh2 (v)

In particular, this does not depend on θ. It follows that all of the surfaces Sθ2 are
locally isometric to each other by theorem 4.4.9. In particular, S02 is the helicoid
(ruled surface where γ is the z-axis, and δ is a circular helix), while S 2π is the
2
catenoid (ie the surface of revolution of the catenary curve x = cosh(z)).

The helicoid S02 . The catenoid S 2π .

2

These surfaces, which look very different, are nevertheless locally isometric, and
so have the same geodesics and Gaussian curvature. In fact, we can compute the
Gaussian curvature using corollary 7.1.9, with F = 0 and E = G = cosh2 (v). Then

92
Applications of the Theorema Egregium Gauss’ Theorema Egregium

Eu = Gu = 0 and Ev = Gv = 2 cosh(v) sinh(v). Hence

1
    
∂ G ∂ Ev
K=− √ √ u + √
2 EG ∂u EG ∂v EG
!
1 ∂
2
cosh(v)
 sinh(v)

=− 0+
2

cosh 2
(v) ∂v cosh2 (v)
1 ∂
=− tanh(v)
cosh (v) ∂v
2

1
=− .
cosh4 (v)

The family Sθ2 gives a continuous transformation of the helicoid into the catenoid
by surfaces which are all locally isometric, this is shown in figure 88.

Figure 88. The family of surfaces Sθ2 interpolating between a helicoid and a catenoid.

7.2 Applications of the Theorema Egregium

In this section, we will discuss two application of the Theorema Egregium to the real
world.
Recall that, in definition 5.5.1, we classified points on a surface according to their
Gaussian and principal curvatures as one of elliptic (K > 0), hyperbolic (K < 0), or
parabolic/planar (K = 0). The Theorema Egregium means that under a local isome-
try, elliptic points must remain elliptic; hyperbolic points must remain hyperbolic; but,
in principle, a planar point may become parabolic, or vice versa. This has a rather
surprising application.
Application 7.2.1 (Eating pizza). A pizza is flat, and therefore has Gaussian
curvature K = 0. Cooked pizza dough is not stretchy, so while a slice of pizza may
bend as you pick it up, this bending will not change distances, and so is a local
isometry. Gauss’ Theorema Egregium says that however the slice of pizza bends,
its Gaussian curvature will still be K = 0.
If held naïvely by the crust, the end of the pizza slice is liable to droop down under
gravity, and the toppings will fall off. Along its length, the slice of pizza with have
a non-zero principal curvature κ1 , and so for the the Gaussian curvature κ1 κ2 to
be zero, the principal curvature along the crust must be κ2 = 0.
In order to avoid the toppings falling off the pizza, we want the end of the pizza to
stay up, or in other words we want the principal curvature along the length of the
pizza to be κ1 = 0. By the Theorema Egregium, we can force this to be the case

93
Applications of the Theorema Egregium Gauss’ Theorema Egregium

by introducing a curve into the crust of the pizza, ie by making κ2 non zero.
This is shown in figure 89. Most people know this trick, almost intuitively, but
now you know the mathematics behind why it works.

Figure 89. It is possible to manipulate th principal curvatures of a slice of pizza to stop the
toppings from falling off.

For the second application, we return to the subject of cartography—that is to

say, making maps of the Earth— which we discussed in application 4.4.13. There we
saw that the Mercator projection allows us to create a conformal map of the Earth,
and Archimedes Tombstone Theorem allows us to create area-preserving maps of the
Earth. We can ask ourselves whether it is possible to create maps of the Earth which
accurately represent distances? Gauss’ Theorema Egregium gives a negative answer to
that question.
Application 7.2.2. Any map of any region of the Earth’s surface distorts dis-
tances.
Proof. Model the Earth as a sphere S 2 of radius r = 6.3 × 106 m, and let R ⊂ S2
be a region. Suppose that f : R → R2 is a smooth function which scales the lengths
of all curves in R by some constant factor24 ε > 0. The image f (R) would be a 24
Maps are not usually 1 : 1, the
map of R which did not distort distances. Compose f with the function shrink the region they represent so
that a large area can be represented
1 on a piece of paper.
h : R2 → R2 : p 7→ p.
ε

This gives a function h ◦ f : R → R2 which preserves distances, which is, by def-

inition, a local isometry. Gauss’ Theorema Egregium then implies that R and
h(f (R)) must have the same Gaussian curvature. But of the sphere of radius r,
the Gaussian curvature is K = 1/r2 ̸= 0 (by example 5.4.4), whereas the plane
has Gaussian curvature K = 0. This is a contradiction, so no such function f can
exist. ■

94
 The Gauss-Bonnet Theorem

8 The Gauss-Bonnet Theorem

We will finish this course by discussing (one of) the deepest, most beautiful, and im-
portant results in the theory of smooth surfaces. The Gauss-Bonnet Theorem relates
the Gaussian curvature averaged over a region R of the surface with the total geodesic
curvature along the boundary of the region ∂R, and the topology of the surface within
the region, see figure 90.

topological
ZZ Z
KdAσ + κg ds = 2π ×
R ∂R invariant Figure 90. A region R on a sur-
face.
One of the most remarkable things about this is that the right hand side is invariant
under homeomorphisms—continuous invertible transformations of the surface—and can
be computed even if the surface is not smooth.25 In particular, you do not need to use 25
If you have taken the course Ge-
any of the theory we have developed in this course to compute it. On the other hand, ometric Topology, you can already
compute the right hand side, because
the left hand side uses everything we have developed, and is highly dependent on the
it is essentially the Euler character-
smooth structure of the surface. In this way, the Gauss-Bonnet Theorem provides a istic of the region.
bridge linking the worlds of geometry and topology.

Figure 91. A bridge between geometry and topology.

We will in fact meet several versions of the Gauss-Bonnet Theorem which apply in
more and more general settings. We will first consider the case that the region R is a
polygon in the surface, which was proved by Pierre Ossian Bonnet in 1848. The special
case of this—when R is a triangle with geodesic edges—was proved earlier by Gauss in
the same 1827 paper where he proved his Theorema Egregium. In both of these cases,
the topological invariant is 1. We will finally prove the full version for global surfaces
where the topological invariant can be more complicated.
The Gauss-Bonnet Theorem can be thought of, in a sense, as a generalisation of the
Fundamental Theorem of Calculus, or Stoke’s Theorem, where an integral over a region
is related to the integral (or in the case of FTC, the sum) over the boundary. In a certain
sense, you can also throw Cauchy’s Residue Theorem from complex analysis into this
class of results.
In turn, the Gauss-Bonnet Theorem has important generalisations to higher dimen-
sions. These include the Chern–Gauss–Bonnet Theorem, which is a direct generalisation
to even dimensional manifolds, the Riemann–Roch Theorem for complex manifolds, and
the Atiyah–Singer Index Theorem.

8.1 Local versions

Before tackling the general version of the Gauss-Bonnet Theorem, we will consider vari-
ous local versions. In each case, the important simplification we make is that the region

95
Local versions The Gauss-Bonnet Theorem

R under consideration will be topologically simple, so we can focus on the geometric

aspects of the theorem. The topological aspects will arise in the last section.

Definition 8.1.1. A (topological)26 surface D2 is a (topological) disk if it is home- 26

Recall definition 3.2.1.
omorphic to
{(x, y) ∈ R2 | x2 + y 2 < 1}.

In what follows, the word disk should be understood to mean a topological disk.
Some examples are shown in figure 92.

Figure 92. The first three are all examples of (topological) disks, the fourth is not.

We will need one more topological notion for now.

Definition 8.1.2. A subset X ⊂ Rn is path connected if, for any two points
p, p′ ∈ X, there exists a continuous map p : [0, 1] → X such that p(0) = p and
p(1) = p′ . A path connected component of a set X is a path connected subset
Y ⊂ X such that if Y ′ is a path connected subset satisfying Y ⊂ Y ′ ⊂ X, then
Y = Y ′.

The path connected components of a set are the maximal path connected subsets of
X, see figure 93. There are several notions we must introduce before we can state the
first version of the Gauss-Bonnet Theorem rigorously. We start by defining polygons on
a surface. Figure 93. Each of the path con-
nected components of this subset
Definition 8.1.3. Let S 2 be a path connected surface, and γ : [a, b] → S 2 a piecewise of the plane has been given a dif-
smooth, continuous, and injective function which satisfies ferent colour.

• γ(a) = γ(b),

• the complement of γ, S 2 \ γ([a, b]), has two path connected components, at

least one of which is a (topological) disk,

• if a = a0 < a1 < a2 < · · · < an−1 < an = b is a partition of [a, b] such that
γ is smooth on each open interval (ai , ai+1 ), then γ is a regular curve when
restricted to each interval and

lim γ ′ (t) ̸= 0 and lim γ ′ (t) ̸= 0.

t→ai + t→ai+1 −

Fix a component of the complement of γ, D2 , which is a topological disk27 , then 27

It turns out that, except when S 2
P = γ([a, b]) ∪ D2 is s curvilinear polygon, γ is called its boundary, sometimes is homeomorphic to a sphere S2 , ex-
actly one of the path connected com-
written ∂P , and D2 is its interior, sometimes written int(P ) or int(γ).
ponents will be a disk.

Example 8.1.4. Figure 94 shows three surfaces, each containing a piecewise smooth
closed curve. In the sphere, the curve together with the maroon region is an
example of a curvilinear polygon, since γ is injective, and P is a disk. It happens
that γ together with the blue region is also a curvilinear polygon.
In the second surface, the pseudosphere, γ satisfies all of the necessary conditions,

96
Local versions The Gauss-Bonnet Theorem

but neither the maroon nor the blue region is a disk, and so γ is not the boundary
of a curvilinear polygon.
On the torus, the curve γ is not injective, so it cannot be the boundary of a
curvilinear polygon.

Figure 94. Three surfaces containing piecewise smooth closed curves.

It will be useful to be able to talk about the different parts of the boundary of a
curvilinear polygon.

Definition 8.1.5. Let P be a curvilinear polygon with boundary γ, and with no-
tation as above, let a = a0 < a1 < a2 < · · · < an−1 < an = b be the minimal
partition of [a, b] such that γ is smooth on each (ai , ai+1 ). The edges of γ are the
sets ei = γ([ai , ai+1 ]). The vertices of γ are the points vi = γ(ai ), so v0 = vn .
The exterior angle of γ at vi = γ(ai ) is the signed angle θi ∈ [−π, π] between t− i
and t+i where, for 0 < i < n,

t−
i = lim γ (t) and ti = lim+ γ (t),
′ + ′
t→ai − t→ai

and for i = 0 or n,

t−
0 = tn = lim γ (t) and t0 = tn = lim+ γ (t).
− ′ + + ′
t→an − t→a0

Figure 95. The boundary of a curvilinear polygon with the vertices vi , edges ei , and exterior
angles θi labelled.

These definitions are illustrated in figure 95. To make sure that signs match up
in the statement of the Gauss-Bonnet Theorem, we need to reintroduce the idea of an
orientation on a surface.

97
Local versions The Gauss-Bonnet Theorem

Definition 8.1.6. Let S 2 be an orientable surface in R3 , and fix an orientation, ie

a smooth choice of unit normal Np at every point p ∈ S 2 . Let P be a curvilinear
polygon in S 2 with boundary γ. Then γ is positively oriented if γ follows the
right hand rule with respect to the orientation.

Finally, we are in a position to state a version of the Gauss-Bonnet Theorem.

Theorem 8.1.7 (Gauss-Bonnet Theorem for curvilinear polygons). Let S 2 be an ori-

ented surface, and P a curvilinear polygon in S 2 with positively oriented boundary
γ. For simplicity, assume that σ : U → S 2 is a regular local parametrisation with
P ⊂ σ(U ) such that the unit normal of σ agrees with the orientation of S 2 . Then,
ZZ Z n
X
KdAσ + κg ds + θi = 2π,
P γ i=1

√
where dAσ = EG − F 2 dudv is the area element of σ; the geodesic curvature is
integrated with respect to the arc length s on γ (which is well defined except at
the vertices); and the sum is of the exterior angles taken over all vertices.
Figure 96. A positively oriented
curve (top) and a negatively ori-
As a direct result of the proof of the Global Gauss-Bonnet Theorem in section 8.3, ented curve (bottom).
we will be able to drop the assumption that P ⊂ σ(U ), but, to do that, we need the
notion of a global surface.
The geodesic curvature is not defined at the vertices of γ, but this does not change the
value of the second integral. We can think of the exterior angles θi as a discrete version
of the geodesic curvature defined at the vertices, where the tangent γ ′ is discontinuous.
The way to think about this is

ZZ Z n
X
KdAσ + κg ds + θi = 2π,
P γ i=1
Total curvature
over P Total curvature over ∂P

We will use the following results in the proof. The first, proposition 8.1.8, gives
a way to compute the geodesic curvature along the curve in terms of the angle of the
tangent vector and the first fundamental form.

Proposition 8.1.8. Let γ be the boundary of a curvilinear polygon lying in the

image of an orthogonal28 local parametrisation σ, and write γ(s) = σ(u(s), v(s)). 28
Recall this means that F = 0, and
Let ϕ(s) be the angle between σ u and γ ′ (s), defined where γ is smooth. Then, σ u and σ v are perpendicular.

1 dϕ
κg = √ (Gu v ′ − Ev u′ ) +
2 EG ds

at the smooth points of γ.

σu σu σv σv
Proof. Let e1 = = √ and e2 = = √ so that Np = e1 × e2
∥σ u ∥ E ∥σ v ∥ G
and {e1 , e2 , Np } is a right handed orthonormal basis for R3 . We can assume that
where it is smooth, γ is parametrised with unit speed. By the definition of ϕ, we
can write
γ̇(s) = cos(ϕ)e1 + sin(ϕ)e2 .

We want to compute κg = γ̈ · (Np × γ̇), so first we can compute

Np × γ̇ = cos(ϕ)Np × e1 + sin(ϕ)Np × e2
= cos(ϕ)e2 − sin(ϕ)e1 ,

98
Local versions The Gauss-Bonnet Theorem

and

γ̈ = − sin(ϕ)ϕ̇e1 + cos(ϕ)ė1 + cos(ϕ)ϕ̇e2 + sin(ϕ)ė2 ,

d
where ėi = ei (u(s), v(s)). Since {e1 , e2 } is orthonormal, we have
ds

1 = e1 · e1 =⇒ 0 = e1 · ė1 ,
1 = e2 · e2 =⇒ 0 = e2 · ė2 ,
0 = e1 · e2 =⇒ 0 = ė1 · e2 + e1 · ė2 ,

so

κg = (− sin(ϕ)ϕ̇e1 + cos(ϕ)ė1 + cos(ϕ)ϕ̇e2 + sin(ϕ)ė2 ) · (cos(ϕ)e2 − sin(ϕ)e1 )

= 0 + cos2 (ϕ)ė1 · e2 + cos2 (ϕ)ϕ̇ + 0 + sin2 (ϕ)ϕ̇ + 0 + 0 − sin2 (ϕ) e1 · ė2
| {z }
=−ė1 ·e2

= (cos (ϕ) + sin (ϕ))(ϕ̇ + ė1 · e2 )

2 2

= ϕ̇ + ė1 · e2 .

We can compute

d σu σ u′ + σ v ′ 1 Eu u′ + Ev v ′
ė1 = √ = uu √ uv − √ σu ,
ds E E 2 E E

so, using lemma 7.1.5,

0
1
 = u − 2 Ev
F
* = 12 Gu
1 Eu u′ + Ev v ′

ė1 · e2 = √ 2u σ uu · σ v + 2v σ uv · σ v −
′ z }| { ′ z }| {
σu · σv
2 EG E
1 1 1
 
= √ −2u′ Ev + 2v ′ Gu
2 EG 2 2
1
= √ (Gu v ′ − Ev u′ ) ,
2 EG

hence the claim. ■

The next theorem we present without proof; it essentially says that as you go around
a piecewise smooth, positively oriented, smooth curve, the tangent vector makes one
whole revolution—turning through an angle of 2π.

Theorem 8.1.9 (Turning tangents). Let γ be the positively oriented boundary of a

curvilinear polygon in the image of a local parametrisation σ, then
n Z n
X ai
dϕ X
ds + θi = 2π
i=1 ai−1 ds i=1

More precisely, by the Fundamental Theorem of Calculus, we can write

dϕ
Z ai
ds = lim ϕ(s) − lim + ϕ(s) = ∠t− − +
i , σ u − ∠ti−1 , σ u = ∠ti , ti−1 ,
+
ai−1 ds s→ai − s→ai−1

so this integral computes the angle through which the tangent vector to γ turns as you
move along the edge from vi−1 to vi , see figure 97.
Finally, we use Green’s Theorem in the final step of the proof, and we recall it here
for completeness.

99
Local versions The Gauss-Bonnet Theorem

Z ai
dϕ
Figure 97. The quantity ds + θi represents the total angle the tangent γ ′ turns
ai−1
ds
through between the start of ei−1 and the start of ei .

Theorem 8.1.10 (Green’s Theorem). Let P be a curvilinear polygon with positively

oriented boundary γ(s), and let Q and R be functions defined on an open region
containing P which have continuous partial derivatives. Then,

du dv
Z   ZZ  
∂R ∂Q
Q +R ds = − dudv.
γ ds ds P ∂u ∂v

Now we can prove the Gauss-Bonnet Theorem for curvilinear polygons, theorem 8.1.7.
Proof. By theorem 4.4.5, we can choose to parametrise the surface using a
local parametrisation σ which is an orthogonal local parametrisation, in particular,
√
we can assume that F = 0. With this assumption dAσ = EGdudv, and by
corollary 7.1.9,

1
    
∂ Gu ∂ Ev
K=− √ √ + √ .
2 EG ∂u EG ∂v EG

We can further make the simplifying assumption that, where is is smooth, γ is

parametrised by arc length. Now we can compute the integral of the geodesic
curvature using each of the preliminary results. By proposition 8.1.8

1 dϕ
Z Z  
κg ds = √ (Gu v − Ev u ) +
′ ′
ds,
γ γ 2 EG ds

then the theorem of turning tangents implies

n
1
Z Z  
Gu ′ Ev ′ X
κg ds = √ v −√ u ds + 2π − θi ,
γ γ 2 EG EG i=1

Ev Gu
next, Green’s Theorem (with Q = − √ and R = √ ) allows us to rewrite
EG EG
this as
n
1 ∂ 1 √
Z ZZ     
G ∂ E
√ u √ v
X
κg ds = + √ EGdudv + 2π − θi ,
γ P 2 ∂u EG ∂v EG EG i=1

and finally the expression for the Gaussian curvature stated above gives
Z ZZ n
X
κg ds = −KdAσ + 2π − θi ,
γ P i=1

from which the claim follows. ■

100
Local versions The Gauss-Bonnet Theorem

Remark 8.1.11. Theorem 8.1.7 is examinable (you will be expected to be able

to state and apply it); however, the proof, including the statements of proposi-
tion 8.1.8, theorem 8.1.9, and theorem 8.1.10, are not examinable.
We can consider some special cases of this theorem.

Corollary 8.1.12 (Gauss-Bonnet Theorem for geodesic polygons). Let P ⊂ σ(u) be

a curvilinear polygon with n edges and whose boundary is positively oriented such
that each edge is a geodesic. Then
n
X ZZ
αi = (n − 2)π + KdAσ ,
i=1 P

where αi = π − θi is the interior angle at the vertex vi .

In particular, in the case of a geodesic triangle, when n = 3, we have
ZZ
KdAσ = α1 + α2 + α3 − π.
p

Proof. Since the edges are geodesics, κg = 0 by proposition 6.3.3. We can also
compute
Xn Xn Xn
θi = (π − αi ) = nπ − αi ,
i=1 i=1 i=1 Figure 98. A geodesic triangle in
so the result follows immediately from theorem 8.1.7. ■ S2 .

In particular, we can apply this to surfaces of constant curvature.

Example 8.1.13 (Geodesic polygons in S2 ). If K = 1, and P is a polygon in the

sphere bounded by arcs of great circles, then
n
X ZZ
αi = (n − 2)π + dAσ = (n − 2)π + Aσ (P ),
i=1 P

and for a spherical triangle, T , we get

Aσ (T ) = α1 + α2 + α3 − π.
Figure 99. A triangle in R2 .
In fact, you may already have proved this using Archimedes’ Tombstone Theorem
in exercise 11 on sheet 6. Since the area is always positive, this implies that the
the angle sum in a spherical triangle is always > π, see figure 98.

Example 8.1.14 (Geodesic polygons in R2 ). If K = 0 and P is a polygon in R2

bounded by line segments, then
n
X
αi = (n − 2)π,
i=1

in particular, when n = 3, ie a triangle, Figure 100. A geodesic triangle

in the pseudosphere.
α1 + α2 + α3 = π,

which is illustrated in figure 99.

Remark 8.1.15. This is probably the most complicated proof possible that the
angles in a triangle add to 180◦ !

101
Global surfaces The Gauss-Bonnet Theorem

Example 8.1.16 (Geodesics on the pseudosphere). If K = −1 and P is bounded by

geodesics, then P is called a hyperbolic polygon and
n
X ZZ
αi = (n − 2)π − dAσ = (n − 2)π − Aσ (P ),
i=1 P

and for a hyperbolic triangle, T , as shown in figure 100, we get

Aσ (T ) = π − (α1 + α2 + α3 ).

Since area is always positive, this implies that the angle sum is always < π.

Corollary 8.1.17 (Local Gauss-Bonnet Theorem). Let P be a curvilinear polygon in

the image of a local parametrisation σ whose boundary, γ, is smooth (including at
its start/end point γ(a) = γ(b)) and positively oriented. Then,
Z ZZ
κg ds = 2π − KdAσ ,
γ P

where κg is the geodesic curvature of γ, and K is the Gaussian curvature of σ.

This follows immediately from theorem 8.1.7. As an example of an application of

these local versions of the Gauss-Bonnet Theorem, we can prove that topologically simple
surfaces with non-positive Gaussian curvature have unique geodesics.

Corollary 8.1.18. Let S 2 be a surface which is homeomorphic to R2 and has Gaus-

sian curvature K ≤ 0 everywhere. Then, S 2 contains no closed geodesics.

Proof. Assume that γ is a closed geodesic in S 2 , by the Jordan Curve Theorem,29 29

Recall that this states that any
γ ∪ int(γ) is a curvilinear polygon with smooth boundary. Now, applying the Local simple closed curve in (a surface
homeomorphic to the) plane divides
Gauss-Bonnet Theorem, we have
the surface into two regions, exactly
one of which, the interior, is bounded
0
Z >
 ZZ and a topological disk,
κg ds = 2π −
 KdAσ ≥ 2π,
γ |
int(γ)
{z }
≤0

but 0 < 2π, a contradiction. ■

8.2 Global surfaces

Most surfaces cannot be completely covered with a single local parametrisation σ : U →
S 2 from a path connected domain U , for example, the sphere requires at least two such
local parametrisation (eg stereographic projection from the north and south poles), or
more (eg it can be covered by six hemispheres, see figure 101). To study a surface S 2 as
a whole, we need to be able to transition from one local parametrisation to another.

Definition 8.2.1. Two local parametrisations σ : U → S 2 and σ e: Ue → S 2 are

compatible if either σ(U ) ∩ σ
e (U
e ) = ∅, or for all p ∈ σ(U ) ∩ σ
e (U
e ), there exist
open subsets V ⊂ U and V ⊂ U such that
e e Figure 101. The sphere can be
covered using six hemispheres—
• p ∈ σ(V ), each one a local parametrisation.

• σ(V ) = σ
e (Ve ), and

• σ|V and σ
e |Ve are injective and the maps

e −1 ◦ σ : V → Ve
σ and σ −1 ◦ σ
e : Ve → V

102
Global surfaces The Gauss-Bonnet Theorem

are smooth.

This is illustrated in figure 102.

Figure 102. Two compatible local parametrisations.

Definition 8.2.2. Let S 2 be a subset of Rn . An atlas for S 2 is a collection of local

parametrisations {σ i : Ui → Rn }iI , where I is an indexing set, such that

• the σ i ’s cover S 2 , ie S 2 = i∈I σ i (Ui ),

S

• for all i, j ∈ I, σ i and σ j are compatible, and

• for all p ∈ S 2 , if p ∈ σ i (Ui ) then there exists an open subset Vi ⊂ Ui such

that p ∈ σ i (Vi ) and σ i (Vi ) − S 2 ∩ W for some W ⊂ Rn open.

Compare this definition with the original definition of a surface, definition 3.2.1.

Example 8.2.3. The atlas for S 2 ∈ R3 , using six hemispheres, can be written down
explicitly as
p 
σ+1 : U → R : (u, v) 7→ 1 − u2 − v 2 , u, v
3

 p 
σ−1 : U → R 3
: (u, v) →
7 − 1 − u2 − v 2 , u, v
 p 
σ+2 : U → R : (u, v) 7→ u, 1 − u2 − v 2 , v
3

 p 
σ−2 : U → R 3
: (u, v) →
7 u, − 1 − u2 − v2 , v
 p 
σ 3 : U → R3 : (u, v) 7→ u, v, 1 − u2 − v 2
+

 p 
σ−3 : U → R 3
: (u, v) →
7 u, v, − 1 − u 2 − v2

where U = {(u, v) ∈ R2 | u2 + v 2 < 1}. Then, for example,

p
1 (u, v) = σ 2 (e
σ+ u, ve) ⇐⇒ u
e= 1 − u2 − v 2 and ve = v.
+

So, on V = {(u, v) ∈ R2 | u2 + v 2 < 1 and u > 0} ⊂ U open,

p 
((σ +
2)
−1
1 )(u, v) =
◦ σ+ 1 − u2 − v 2 , v ,

103
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

which is smooth. Similarly, on Ve = V ⊂ U ,

p 
1)
((σ + −1
2 )(u, v) =
◦ σ+ 1 − u2 − v 2 , v ,

so σ +
1 and σ 2 are compatible. Checking all the other cases verifies that this is
+

indeed an atlas for S2 .

Definition 8.2.4. A global surface is a subset S 2 ∈ Rn together with an atlas for

S 2.

A topological notion which will be important is that of compactness, which you may
have seen before, but we’ll briefly revise here.

Definition 8.2.5. Let X ⊂ Rn be a set. An open cover of X is a collection of open

subsets of Rn , {Yi }i∈I , such that X ⊂ i∈I Yi . The set X is compact if for any
S

open cover of X, {Yi }i∈I , there is a finite subcover, ie J ⊂ I with the cardinality
of J , #J < ∞ such that X ⊂ j∈J Yj .
S

This definition can be extended to much more general topological settings, but, for
subsets of Rn , there is the following simple characterisation of compact sets.

Theorem 8.2.6 (Heine-Borel Theorem). A subset X ⊂ Rn is compact if and only if

it is closed and bounded.

Definition 8.2.7. The compact orientable (topological) surface of genus g,

where g ≥ 0 is an integer, is the surface Σ2g :
Figure 103. An example of a
topological sphere.

In particular, Σ20 = S2 is the (topological) sphere, see figure 103, and Σ21 = T2 is the
(topological) torus, see figure 104. Note, this only defines Σ2g as a topological surface, so
any surface which is homeomorphic to Σ2g is also Σ2g .
Figure 104. An example of a
Theorem 8.2.8. Each surface Σ2g can be embedded smoothly into R3 and given topological torus, knotted into a
trefoil.
an atlas, making it a global surface, for g ≥ 0. Moreover, every path connected
compact orientable global surface is homeomorphic to Σ2g for some g.

The proof of this theorem goes well beyond the scope of this course, but the second
part is proved in the module Geometric Topology.

Remark 8.2.9. By the definition of compactness,30 any compact global surface can 30
Recall a subset X ∈ Rn is com-
be given a finite atlas, ie one where #I < ∞. pact if any open cover has a finite
sub-cover.
8.3 The Gauss-Bonnet theorem for global surfaces
Although not strictly necessary for the statements to be true, we made the simplifying
assumption in all of the local versions of the Gauss-Bonnet Theorem that the curvilinear
polygon was contained in the image of a single (orthogonal) local parametrisation. By
using an atlas of orthogonal local parametrisations, this assumption could have been
dropped.

104
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

On the other hand, an assumption which was essential, was that the interior of
the curvilinear polygon is a topological disk. This is because disks are, topologically
speaking, very simple, so we could ignore the topology in some sense, and focus on the
geometry.
In the context of global surfaces, we lose this control on the topology. In order to
recover this, the idea will be to divide the surface up into curvilinear polygons, and then
we can apply the local versions to each of these.

Definition 8.3.1. A triangulation of a subset R ⊂ S 2 of a surface is a collection

of curvilinear triangles,31 T = {Tj }j∈J such that 31
That is, curvilinear polygons with
three edges.
• R = j∈J Tj , ie the triangles cover R, and
S

• if Ti ∩ Tj ̸= ∅ then one of the following holds

– Ti = Tj ,
– Ti ∩ Tj is exactly one edge, or
– Ti ∩ Tj is exactly one vertex.

These are illustrated in figure 105.

A triangulation is finite if it contains finitely many curvilinear triangles.

It is not true that every subset R ⊂ S 2 admits even a single triangulation.

Definition 8.3.2. If a subset R ⊂ S 2 has a triangulation, then it is called triangu-

lable.

Question 8.1. Is an open disk triangulable? Figure 105. Valid and invalid ar-
rangements of triangles in a trian-
If R is triangulable and non-empty, then it has infinitely many triangulations. gulation.

Lemma 8.3.3. A triangulable set R ⊂ S 2 is compact if and only if it has a finite

triangulation.

Question 8.2. Can you write down a rigorous proof of this fact?

Example 8.3.4. The tiling of the plane by equilateral triangles gives a triangulation
of the whole plane, see figure 106.

Example 8.3.5. The sphere can be triangulated by 4, 8, or 20 equilateral triangles Figure 106. Tiling of R2 by equi-
(corresponding to the tetrahedron, octahedron, and icosahedron respectively), see lateral triangles.
figure 107.

Definition 8.3.6. Let R be a compact triangulable subset of S 2 with finitetrian-

gulation T . Let v be the number of vertices in T , e the number of edges, and f
the number of triangles (if the same edge or vertex appears in several of the tri-
angles, then it is only counted once). Then, the Euler characteristic of T , χ, is
χ(T ) = v − e + f .

Example 8.3.7. Consider the three triangulations of the sphere given in exam-
ple 8.3.5, we can compute their Euler characteristics in table 1.
Figure 107. Tilings of S2 by equi-
You may notice that each of these triangulations, even though they have completely lateral triangles.
different values for v, e, and f , all have the same Euler characteristic, and that is not a
coincidence.

105
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

Table 1. Exler characteristics for three triangulations of the sphere.

Triangulation T v e f χ(T )
Tetrahedron 4 6 4 2
Octahedron 6 12 8 2
Icosahedron 12 30 20 2

Theorem 8.3.8. Let R ⊂ S 2 be a compact triangulable subset, and T and Te two

finite triangulations of R, then χ(T ) = χ(Te ).

This allows us to make the following definition.

Definition 8.3.9. The Euler characteristic of a compact triangulable subset R ⊂

S 2 is χ(R) = χ(T ) where T is any finite triangulation of R.

One final observation before we state and prove the global Gauss-Bonnet Theorem.

Proposition 8.3.10. If R and R′ are diffeomorphic32 , then they have the same Euler 32
The definition of Euler character-
characteristic. istic given here requires the bound-
aries of the triangles to be piece-
wise smooth, and this property is pre-
Proof. A diffeomorphism from R → R′ sends curvilinear triangles to curvilin-
served only by diffeomorphisms. In
ear triangles, and is invertible, so sets up a one-to-one correspondence between a more general topological setting,
triangles, edges and vertices. ■ triangles need only have continuous
boundary, and so in that case, the
Theorem 8.3.11 (Global Gauss-Bonnet Theorem). Let S 2 be an orientable global Euler characteristic is preserved by
homeomorphisms.
surface, and R ⊂ S 2 a compact triangulable subset with positively oriented bound-
ary ∂R = γ. Then,
ZZ Z n
X
KdAσ + κg ds + θi = 2πχ(R).
R γ i=1

Note some subtle differences in the global setting. As we are no longer working
in a single local parametrisation, the area element, in local coordinates, will change as
we transition from one local parametrisation to another. The first fundamental form
does not depend on the choice of local parametrisation (only its coefficients do), so we
can write dAσ = det(Ip )dudv, where (u, v) are the coordinates on the surface which
p

change from one local parametrisation to another.

Note, also, that since R can be much more complicated, its boundary is no longer
necessarily connected, γ could now be a disjoint collection of piecewise smooth, con-
Figure 108. An example of a
tinuous and injective closed curves in S 2 , see figure 108. We can still insist that each complicated region R in a surface.
component is positively oriented, and these components still have vertices and exterior R is not path connected, and its
angles, as defined before. boundary consists of five disjoint
loops.
Proof. By lemma 8.3.3, we can fix a finite triangulation T for R, and γ will be a
finite union of edges and vertices from this triangulation.
Let {σ i : Ui → S 2 } be the atlas for S 2 . After subdividing the triangulation, if
necessary, we can assume that every curvilinear triangle Tj ∈ T lies in the image
of exactly one of the local parametrisations Tj ⊂ σ ij (Uij ).

106
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

In this example, we can first subdivide any triangles which do not lie in the image of
a single local parametrisation, and then subdivide any triangles which have become
quadrilaterals, or other polygons.
By theorem 8.3.8, the Euler characteristic of R, computed with this new triangu-
lation, stays the same. Write T = {Tj }fj=1 , then we can apply the Gauss-Bonnet
theorem for curvilinear polygons, theorem 8.1.7, to each triangle. This gives
ZZ f ZZ
X
KdAσ = KdAσij
R j=1 Tj
exterior angles of Tj !
f
X Z z }| {
= 2π − κg ds − (θj1 + θj2 + θj3 )
j=1 ∂Tj

f
X XZ f X
X 3
= 2πf − κg ds − θjℓ . (∗)
j=1 edges e j=1 ℓ=1
of Tj
| {z } | {z }
P Q

First, we will compute P . Suppose first that e is

an edge in the interior of R, so that it is where
two triangles meet. In the sum, we integrate κg
along e twice, once in each direction because the
boundary of both triangles is positively oriented.
These integrals have the same size, but opposite
signs, and hence they cancel out. The only terms
which remain are the integrals along the edges
which form the boundary of R, in other words,
X Z Z
P = κg ds = κg ds,
exterior e γ
edges

where γ = ∂R is the boundary of R.

In order to compute Q, let us translate from using exterior angles to interior angles.
Recall, from corollary 8.1.12, that the interior angle at a vertex with exterior angle
θ is defined to be α = π − θ. Then Q becomes

f X
X 3 f X
X 3 f X
X 3
Q= θjℓ = (π − αjℓ ) = 3πf − αjℓ ,
j=1 ℓ=1 j=1 ℓ=1 j=1 ℓ=1

where the final sum runs over all interior angles of all triangles in T . In order to
simplify notation, let

107
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

v i = number of interior vertices,

v e = number of exterior vertices,
ei = number of interior edges,
ee = number of exterior edges.

On the right, v i = 3 (light blue), v e = 8

(dark blue), ei = 14 (pink), and ee = 8
(purple).
Notice that v e = ee since ∂R is always a (disjoint union of) piecewise smooth loops,
and the edges and vertices on a loop alternate, so there is the same number. We
also have e = ei + ee and v = v i + v e . Now we can compute

f X
X 3
Q = 3πf − αjℓ
j=1 ℓ=1
=2π =π−θi
 
z X}| { z X}| {
 X X
= 3πf −  αjℓ + αjℓ 

interior triangles exterior triangles
vertices v meeting v vertices vi meeting vi
ve
!
X
= 3πf − 2πv + πv −
i e
θi
i=1
e
v
X
= π(3f − 2v − v ) +
i e
θi .
i=1

Here we have used the facts that the sum of the

angles around an interior vertex must add up to
2π, the sum of the interior angles at an exte-
rior vertex is the the interior angle of that vertex,
thought of as a vertex of γ, and, hence, π minus
the exterior angle or that vertex of γ.
Now, note that every triangle has three edges, and
each internal edge meets exactly two triangles, so

3f = e + ei = ee + 2ei ;

Substituting this into our expression of Q gives

e
v
X
Q = π(e + 2e − 2v − v ) +
e i i e
θi
i=1
=v e v e
z}|{ X
= π(2e − e + 2e − 2v − v ) +
e e i i e
θi
i=1
ve
X
= 2π(ee + ei − v i − v e ) + θi
i=1
e
v
X
= 2π(e − v) + θi .
i=1

108
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

Substituting these expressions for P and Q into (∗), gives

e
ZZ Z v
X
KdAσ = 2πf − κg ds − 2π(e − v) − θi
R γ i=1
e
Z v
X
= 2πχ(R) − κg ds − θi ,
γ i=1

which proves the theorem. ■

Notice that if R is homeomorphic to a disk, then it is a curvilinear polygon, say with
e edges. It can be triangulated with e − 2 triangles, and so has Euler characteristic

χ(R) = v − e + f = e − (e + (e − 3)) + (e − 2) = 1,

and we recover the Gauss-Bonnet Theorem for curvilinear polygons, but without the
assumption that R lies in the image of a single local parametrisation.

Remark 8.3.12. The assumption that S 2 is orientable is essential for this proof, we
used it to simplify the term P by arguing that the orientation meant that the the
integrals along interior edges of the geodesic curvature cancelled out. Somewhat
remarkably, the Gauss-Bonnet Theorem also holds for non-orientable surfaces, and
this fact can be deduced from the version of the Gauss-Bonnet Theorem we just
proved for orientable surfaces. This generalisation uses the theory of covering
spaces which goes well beyond the scope of this course.
Remark 8.3.13. If R is not compact, all sorts of things can go wrong. The integral
of the curvature may not converge without this assumption, and even if it does,
the integral along the boundary is not really defined. On the face of it, the Euler
characteristic is also a problem, since it is defined as a sum of finite numbers, and
if R is not compact, it will only admit infinite triangulations. This turns out to be
less of a problem, because it is possible to define the Euler characteristic of spaces
which are not triangulable using tools from algebraic topology (in particular, using
(co-)homology).
Remark 8.3.14. Theorem 8.3.11 is examinable, but its proof is not.

Corollary 8.3.15 (Gauss-Bonnet Theorem for compact orientable surfaces). Let S 2 be

a compact orientable global surface, then
ZZ
KdAσ = 2πχ(S 2 ).
S2

Proof. Such a surface has empty boundary, so the other terms in the global
Gauss-Bonnet theorem vanish. ■
Example 8.3.16. We saw in example 8.3.7 that χ(S 2 ) = 2, where S 2 is any surface
diffeomorphic to the sphere S2 . It follows that
ZZ
KdAσ = 4π.
S2

In the case of the standard unit sphere, K = 1 and this recovers the usual area
formula we proved in example 4.2.6. But now consider any deformed sphere, say
the one in figure 103. The curvature may well not be constant, and may not even
be positive everywhere, and yet the integral of K over the whole sphere will always
be 4π.
We can generalise this to all compact orientable surfaces.

109
The Gauss-Bonnet theorem for global surfaces The Gauss-Bonnet Theorem

Corollary 8.3.17. Let Σ2g be the compact orientable surface of genus g ≥ 0,

smoothly embedded in Rn , then
ZZ
KdAσ = 4π(1 − g).
Σ2g

Proof. All of these quantities have been defined for surfaces in Rn , and would
not change under an isometry into R3 . On the Exercise sheet, you have the chance
to prove that χ(Σ2g ) = 2 − 2g (or take the Geometric Topology module), and so the Figure 109. A genus 6 surface,
result follows from the Gauss-Bonnet Theorem for compact orientable surfaces. ■ Σ26 whose total curvature is −20π.

What corollary 8.3.15 says is that the total curvature of a compact orientable surface
depends only on the topology of the surface, not on its geometry. In particular, the total
curvature of the sphere is always positive (g = 0), the total curvature of the torus is
always 0 (g = 1), and the total curvature for higher genus surfaces is always negative
(g > 1). We know that it is possible to find a constant positive curvature embedding of
the sphere, just take the unit sphere. It turns out also to be possible to embed the torus
in R4 so that it has constant zero curvature—such a torus is called a flat torus, and such
an embedding is given by the Clifford torus.
It turns out also to be possible to give the higher genus surfaces geometric structures
(Riemannian metrics), each with constant Gaussian curvature K = −1. It is known to
be impossible to smoothly33 and isometrically embed these constant curvature higher 33
If you drop the smooth hypothesis,
genus surfaces in R3 . this can be done in R3 by the Nash
Embedding Theorem.
Question 8.3 (Open problem). Is it possible to smoothly embed Σ2g into Rn with
constant Gaussian curvature K = −1 for some n > 3 when g ≥ 2?

110