Chapter 5: Surfaces and their parametrisations

Daniel V. Mathews
April 19, 2020

Contents
1 Introduction 2

2 Surfaces 2
2.1 What is a surface? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Embedded surfaces and parametrisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.3 Examples of surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Graphs of continuous functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 A non-example of a surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.6 Transition maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.7 Examples of transition maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 Defining smooth surfaces 13

3.1 Examples of non-smooth surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Smooth functions and open sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Directional derivatives with coordinate charts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Smooth coordinate charts are not enough . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.5 Nonzero partial derivatives are not enough . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.6 The definition of a smooth surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.7 Characterising regular surface patches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4 Verifying that you have a smooth surface 20

4.1 Many things to verify . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Examples from first principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Trick #1: Cross products verify linear independence . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Trick #2: Graphs of smooth functions are smooth surfaces . . . . . . . . . . . . . . . . . . . . . . . 22
4.5 Trick #3: Level sets are often smooth surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.6 Trick #4: From continuous to homeomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.7 Linearly independent partial derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.8 Proving that level sets are smooth surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5 Features of smooth surfaces 27

5.1 Transition maps of smooth surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2 Tangent vectors and tangent spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.3 Tangent plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4 Normal vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.5 Normals of level sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.6 Unit normals on different coordinate charts and orientations . . . . . . . . . . . . . . . . . . . . . . . 33
5.7 Orientability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

1
1 Introduction
We now step up from curves to surfaces: from one-dimensional objects, to two-dimensional objects. Indeed, the
rest of this course is about surfaces — we’re moving up a dimension!
We mainly consider surfaces in R3 — that’s essentially because we live in space that is (apparently) 3-dimensional,
and higher dimensions are hard to visualise. But much of what we say generalises to Rn .
We’ll take an approach which is analogous to our previous study of curves. We’ll consider how to define surfaces,
how to parametrise them, and what it means for a surface to be “smooth” — this is a little trickier than for curves!
We’ll also look at plenty of examples of curves, think about their tangents and normals, and think about maps from
one surface to another.
There’s lots to see and do, so roll up!

2 Surfaces
2.1 What is a surface?
In studying curves, we saw they could be imagined in many ways: via subsets, as embedded curves; via parametri-
sations; or via equations (or more precisely, as level sets of functions). All these approaches also apply to surfaces.
We will define a parametrisation of a surface, by analogy with a parametrisation of a curve. A parametrised
curve (in R3 ) is a map γ : I −→ R3 , where I is an interval. A parametrised surface in R3 is a map σ : U −→ R3 .
But before making the definition, we must discuss two questions: What sort of set should U be? And what sort of
map should σ be?
We will also define an embedded surface in R3 , by analogy with an embedded curve. And as with curves, the
definition of embedded surface builds upon the notion of parametrisation.
Later, we will look at equations defining surfaces, and level sets.

2.2 Embedded surfaces and parametrisations

If a parametrisation of a surface S is a map σ : U −→ R3 , then the set U should provide the “parameters”. For
a parametrised curve γ : I −→ R2 , a point t in the interval I provides a parameter value. Similarly, a point of U
provides parameters for S. On a two-dimensional surface, we need two coordinates to locate a point. So we will
take U to be a subset of R2 ; a point (u, v) ∈ U provides two parameters u, v to keep track of points on S.
It makes sense for U to be an open set in R2 . We’ve seen that open sets in R2 are 2-dimensional in a strong
sense: the definition of an open set requires that around every point of U , we can fit a 2-dimensional ball contained
in U . Just as a curve should look locally like an interval, a surface should look locally like an open set in R2 .
So a parametrisation of a surface will be given by a function σ : U −→ R3 , where U ⊆ R2 is open. As the
parameters (u, v) vary over U , the points σ(u, v) trace out a surface. But what sort of function should σ be?
It should certainly be continuous. But for σ to be continuous alone is not enough: a continuous function can
crush a 2-dimensional set down to a 1-dimensional set. It should also be injective, since different parameter values
(u, v) should correspond to different points σ(u, v) on the surface.
In fact, by analogy with embedded curves, σ should be a homeomorphism. As we argued in section 4.1 of chapter
2, an embedded curve is one that near any point looks like an interval. We made the idea of “looking like” an
interval precise by using homeomorphisms — an embedded curve is a subset of Rn which is locally homeomorphic
to an interval. Similarly, we will define an embedded surface to be a subset which is locally homeomorphic to an
open set in R2 .
More precisely, when defining an embedded curve C, we require every point p ∈ C to lie in an open ball B so
that there is a homeomorphism γ from an interval I to C ∩ B. The patch C ∩ B of C near p looks like an interval,
in the sense that it is homeomorphic to I.
Similarly, to define an embedded surface S, we require every point p ∈ S to lie in an open ball B so that there
is a homeomorphism σ from an open set U in R2 , to S ∩ B. The patch S ∩ B of S near p then looks 2-dimensional,
in the sense that it is homeomorphic to U , the open set in R2 .
We will however make one slight technological upgrade in the definition. Having discussed open sets in chapter
4, we will use open sets rather than balls. This leads us to the following definition.

Definition 2.1. A set S ⊆ Rn is an embedded surface (or just surface) if every point p ∈ S has an open neigh-
bourhood W in R3 such that S ∩ W is homeomorphic to an open set in R2 .

2
 That is, S is a surface if for every p ∈ S, there is an open neighbourhood W of p in R3 , an open set U ⊆ R2 ,
and a homeomorphism σ : U −→ S ∩ W .
Such a homeomorphism σ : U −→ S ∩ W is called a surface patch or parametrisation or coordinate chart or
just chart of S.

Remark 2.2. We haven’t used the word “neighbourhood” before. An open neighbourhood of a point p ∈ Rn is an
open set containing p. See Appendix A section 3.3 for further discussion.
Remark 2.3. This definition is not just for surfaces in R3 , but in any Rn ! Note how the definition works just the
same regardless. Of course, we are primarily concerned with n = 3.
Remark 2.4. We could have replaced the words “open neighbourhood” with “open ball containing p” in the definition
above. It would then be an exact analogy of the definition of an embedded curve (definition 4.5 of chapter 2). The
resulting definition would be equivalent to the one written. (In exercise 1 below you can try to prove this.) But the
definition written is more flexible and easier to verify, as we’ll see, because we can take any open set containing the
point p, not just a ball.
Remark 2.5. This definition is “local” in the sense that it only requires you to verify something in the neighbourhood
of any given point. There are no “global” requirements. A set is a surface if it locally looks like a surface!
Remark 2.6. This definition contains nothing about smoothness or differentiability. It has no differential geometry
in it! But that will come in due course.
Remark 2.7. The many different names are given to σ reflect different perspectives on it.
• If you think of σ as describing a portion of S near p, then it makes sense to think of it as a “patch” of the
surface.
• If you think of σ(u, v) as a 2-dimensional analogy of a parametrised curve γ(t), with parameters u, v, it makes
sense to call σ a “parametrisation”.
• If you think of the parameters (u, v) ∈ U as “coordinates” of points on S, then σ and U provide a “map” for
part of S — in the cartographic, or navigational sense! — and it makes sense to call σ a “chart”.
Remark 2.8. The image of the chart σ is S ∩ W , the intersection of the surface S with an open set W in R3 . So
S ∩ W is a subset of the surface S, but being the intersection with an open set W , it’s a portion of S that doesn’t
contain its boundary points. We can think of S ∩ W roughly as an “open” subset of S — although it’s certainly
not an open subset of R3 . This type of set has a name.

Definition 2.9. A relatively open subset (or just open subset) of a surface S is a set Y of the form Y = S ∩ W ,
where W ⊂ R3 is an open set.

So each chart σ is a homeomorphism from an open subset U of R2 , to a relatively open subset Y = S ∩ W of S.

Remark 2.10. In general, a surface requires more than one patch σ to cover it. Some surfaces are impossible to cover
with one surface patch. If there is only one surface patch σ, then σ provides a homeomorphism from an open set
U in R2 to the whole surface S. As it turns out, this means that U and S have the same topology, or “shape”. But
not all surfaces are homeomorphic to a plane, or a subset of the plane. For instance, a sphere is not homeomorphic
to a subset of R2 , nor is a torus — so you can’t cover them with a single patch. You’ll learn more about this in a
course on topology.
Since several charts are often required, the following cartographic terminology is used.

3
Definition 2.11. A collection of surface patches for S whose images cover S is called an atlas of S.

(Remember atlases? They are what people used for navigation before google maps.) Each chart is like a page
of the atlas, giving you a map of a certain region on the surface. If you have enough charts to give you a map of
every region of the surface, then you have a complete atlas.
As mentioned above, definition 2.1 of an embedded surface is not the only way to define this concept; there are
several equivalent alternatives, as you can investigate in the following exercise.

Exercise 1. Show that the following alternative definitions of a surface are equivalent to definition 2.1.
(i) A set S ⊆ R3 is a surface if at each point p ∈ S there is an open ball B containing p such that S ∩ B is
homeomorphic to an open set in R2 .a

(ii) A set S ⊆ R3 is a surface if every p ∈ S has an open neighbourhood W in R3 such that S ∩W is homeomorphic
to a product of intervalsb in R2 .
(iii) A set S ⊆ R3 is a surface if every p ∈ S has an open neighbourhood W in R3 such that S ∩W is homeomorphic
to an open ball in R2 .
a Hint:an open ball is already an open set; and conversely, if you have an open set about p, you can restrict it to a ball.
b Arguably, the most natural generalisation of an open interval (a, b) ⊆ R to R2 is a product of open intervals (a1 , b1 ) × (a2 , b2 ) ⊆ R2 .
Such sets are open in R2 . (See section 3.1 of Appendix A.)

2.3 Examples of surfaces

Having made so many definitions, we now give some examples.

Example 2.12. A plane. Let Π ⊂ R3 be the plane which passes through p = (1, 2, 3), with tangent vectors
a = (1, 0, 0) and b = (0, 3, 2).

Define σ : R2 −→ R3 as follows: we will show it is a chart for Π.

σ(u, v) = p + ua + vb = (1 + u, 2 + 3v, 3 + 2v)

The image of σ is Π. The map σ is certainly continuous. Injectivity of σ follows from the fact a and b are linearly
independent: if σ(u1 , v2 ) = σ(u2 , v2 ), then u1 a + v1 b = u2 a + v2 b, so (u1 − u2 )a + (v1 − v2 )b = 0, and linear
independence then implies u1 − u2 = v1 − v2 = 0, so (u1 , v1 ) = (u2 , v2 ).
So σ is a continuous bijection from R2 to S; to show it’s a homeomorphism onto its image, we must show further
that σ −1 : S −→ R2 is continuous. We can calculate σ −1 explicitly: given (x, y, z) ∈ S we find (u, v) such that
σ(u, v) = (x, y, z). From x = 1 + u and y = 2 + 3v we have σ −1 (x, y, z) = (u, v) = (x − 1, y−2 3 ). As the restriction
to S of a linear function R3 −→ R2 , the function σ −1 is continuous.a Thus σ is a homeomorphism.
For any point p ∈ S we can simply take the open neighbourhood W = R3 , so S ∩ W = S, and take U = R2 .
Then σ provides a homeomorphism from U = R2 to S ∩ W = S. Thus σ is a chart which covers all of S and which
hence forms an atlas. We conclude S is an embedded surface.
a There are many other expressions for σ −1 as a linear function in x, y, z.

Remark 2.13. We often use (u, v) for coordinates in the domain U of a surface patch, and in this course, we’ll use
them by default.
Remark 2.14. Of course, there’s nothing special about the particular point p and the vectors a, b in this example.
The same argument shows that any plane is a surface, and it has an atlas consisting of a single linear coordinate

4
chart.

Exercise 2. Prove this. Let Π be the plane through a point p with linearly independent tangent vectors a and b.
Show that σ : U −→ Π defined by σ(u, v) = p + ua + vb provides an atlas for Π.

Example 2.15. A cylinder. Let S be given as follows; we’ll show it’s a surface.

S = (x, y, z) ∈ R3 | x2 + y 2 = 1


We need need a function σ(u, v) with image on S. A point (x, y, z) ∈ S has (x, y) on the unit circle, hence
(x, y) = (cos u, sin u) for some u. So we define

σ : R2 −→ S, σ(u, v) = (cos u, sin u, v).

Clearly σ is continuous (indeed smooth). As u varies, we move around the cylinder at a constant height. As v
varies, you move up and down. So u can be regarded as an angular coordinate, and v as height.
However, σ is not injective: σ is periodic in u, σ(u, v) = σ(u + 2π, v). As varies over R, σ(u, v) wraps around
the cylinder infinitely many times. So σ is not a homeomorphism and we cannot use it as a surface patch.
Restricting σ to a smaller domain makes it injective. To this end, define

U1 = (u, v) ∈ R2 | 0 < u < 2π = (0, 2π) × R, and let σ1 = σ|U1 .



(The notation σ|U1 means the restriction of the function σ to U1 .) Now the product of open intervals is open,a so
U1 is open, and on this domain σ1 is injective.
The image of σ1 is all of S, except the line L1 that would be covered by u = 0 or u = 2π; that is, L1 =
{(1, 0, z) | z ∈ R}. So σ1 is a bijection U1 −→ S \ L1 . One can also show that the inverse σ1−1 is continuous, so σ1
is a homeomorphism. (This is an exercise below.)
Thus, σ1 provides a chart covering almost all of S. A very similar construction can cover the rest of the surface
— and in fact, almost all of S twice. Let

U2 = (u, v) ∈ R2 | −π < 0 < π = (−π, π) × R, and let σ2 = σ|U2 .



The image of σ2 is all of S except the line L2 defined by u = −π or u = π; that is, the line L2 = {(−1, 0, z) | z ∈ R}.
By a similar argument as for σ1 , we obtain a homeomorphism σ2 : U2 −→ S \ L2 .
For a point p ∈ S \ L1 , we can take the open neighbourhood

W1 = R3 \{(x, 0, z) | x ≥ 0}.

This set W1 is the complement of a half-plane F1 defined by y = 0 and x ≥ 0. It is an open set. (You can prove
this in an exercise.) Then S ∩ W1 = S \ L1 , so we obtain a homeomorphism σ1 : U1 −→ S ∩ W1 .
Similarly for p ∈ S \ L2 , we take the open neighbourhood

W2 = R3 \{(x, 0, z) | x ≤ 0},

5
the complement of the half-plane F2 defined by y = 0 and x ≤ 0. Then S ∩ W2 = S \ L2 , and we have a
homeomorphism σ2 : U2 −→ S ∩ W2 .
Any point of S lies in at least one of S \ L1 or S \ L2 , and we can take the homeomorphism σ1 or σ2 accordingly.
Thus σ1 , σ2 are coordinate charts providing an atlas for S.
a See section 3.1 of Appendix A.

In the above example, we did not show σ1−1 was continuous. This uses a fact which will be useful throughout
our examples, and so we’ll state it as a lemma.

Lemma 2.16. Let S 1 = {(x, y) | x2 + y 2 = 1} be the unit circle in R2 . Consider the function f : R −→ S 1 defined
by f (t) = (cos t, sin t). For any constant a ∈ R, f restricts to a homeomorphism

f : (a, a + 2π) −→ S 1 \{(cos a, sin a)}.

Thus, by restricting f to the interval of length 2π, it covers the circle exactly once — except for the point f (a) =
(cos a, sin a) — and gives a homeomorphism onto the circle with this point removed.
Proof. Perhaps this is easiest if we use complex numbers: regard the point (x, y) ∈ R2 as the complex number
x + yi. Then f is the function f (t) = cos t + i sin t = eit , which is clearly smooth. On any interval of length 2π,
f is a bijection between the desired sets. The inverse function is given by f −1 (z) = −i log z, where we choose the
branch of the logarithm function with argument lying in (a, a + 2π). The logarithm function is continuous, so f −1
is continuous, and f is a homeomorphism.

Exercise 3. Complete the gaps in the above example of the cylinder:

(i) Show that σ1−1 is continuous, and hence show σ1 a homeomorphism.
(ii) Prove that the sets W1 , W2 are open sets in R3 . Equivalently, prove that F1 , F2 are closed. The discussion in
Appendix A may be useful if you’re not familiar with open and closed sets.

Example 2.17. The unit sphere. This is a standard surface, with a standard name:a

S 2 = (x, y, z) ∈ R2 : x2 + y 2 + z 2 = 1 .


We tend to regard the positive and negative z directions as north and south, so (0, 0, 1) is the north pole and
(0, 0, −1) is the south pole.
Two standard parameters to navigate on a sphere are latitude and longitude. Let u be the latitude and v the
longitude. The latitude u determines the z-coordinate, by z = sin u. The distance from the z-axis is then cos v;
see below. So x2 + y 2 = cos2 v. The longitude v then determines the x and y coordinates: (x, y) is proportional to
(cos v, sin v). So we obtain
σ(u, v) = (cos u cos v, cos u sin v, sin u) .
As the latitude u varies (with v constant), we walk from one pole to the other along a semi-circular path of constant
longitude, known as a meridian. As the longitude v varies (with u constant) we walk along a circle at constant
height, known as a parallel.

6
 This function σ is continuous, indeed smooth, but not injective. To cover the full set of latitudes we only need
u between −π/2 and π/2; and to cover the full set of longitudes we only need v from, say, 0 to 2π. So define
n π π o  π π
U1 = (u, v) ∈ R2 | − < u < , 0 < v < 2π = − , × (0, 2π) and let σ1 = σ|U1 .
2 2 2 2
This set U1 is open: a rectangle in the (u, v)-plane without its boundary. Restricting σ to U1 yields the injective
map σ1 . Its image is almost all of the sphere: it just misses the meridian M1 given by v = 0 or 2π, or equivalently,
given by y = 0, x2 + z 2 = 1 and x ≥ 0 (the “Greenwich meridian”).
We can define another open set
n π π o  π π
U2 = (u, v) ∈ R2 | − < u < , −π < v < π = − , × (−π, π) and let σ2 = σ|U2 .
2 2 2 2
This gives another injective function σ2 whose image is again all of the sphere, except for a meridian M2 corre-
sponding to v = ±π (the “international date line”).
Both σ1 : U1 −→ S 2 \ M1 and σ2 : U2 −→ S 2 \ M2 are continuous (in fact smooth) and bijective. To find
an inverse of σ1 and σ2 , we need to write u, v√in terms of (x, y, z) ∈ S 2 . For both σ1 , σ2 we have z = sin u and
u ∈ (−π/2, π/2), so u = arcsin z (and cos u = 1 − z 2 ). From (x, y) = (cos u cos v, cos u sin v) we then have
 
x y
(cos v, sin v) = √ , √ ,
1 − z2 1 − z2
which is a point on the unit circle. By lemma 2.16 then v is a continuous function of (x, y, z), taking v ∈ (0, 2π)
for σ1 , and v ∈ (−π, π) for σ2 . Since u and v are continuous functions of (x, y, z) in both cases, σ1−1 and σ2−1 are
continuous. Thus σ1 , σ2 are both homeomorphisms.
For a point p ∈ S 2 \ M1 , we can take the neighbourhood W1 = R3 \ {(x, 0, z) | x ≥ 0} = R3 \ F1 (the same
W1 , F1 as in the previous example 2.15). Then S 2 ∩ W1 = S 2 \ M1 , which we have seen is homeomorphic to U1 via
σ1 .
Similarly, a point p ∈ S 2 \ M2 has a neighbourhood W2 = R3 \F2 where F2 = {(x, 0, z) | x ≤ 0} (the same W2 ,
F2 as in example 2.15) and then S 2 ∩ W2 = S 2 \ M2 is homeomorphic to U2 via σ2 .
Unfortunately, σ1 , σ2 do still not quite form an atlas. Their images cover all of S 2 except the north and south
poles. We still need another chart or two to cover these points.
For this, we can adjust the map σ a bit so that it covers the north and south poles and wraps around in a
slightly different direction. I’ll leave it to you to consider the map

ρ(u, v) = (− cos u cos v, − sin u, − cos u sin v)

and show that, restricting it to appropriate domains and co-domains, you can obtain two more coordinate charts
which cover the poles, yielding an atlas for S 2 .
a In general S n = (x0 , . . . , xn ) ∈ Rn+1 : x20 + · · · + x2n = 1 .


Exercise 4. Verify the missing details in this proof.

(i) Show that the sets W1 , W2 are open sets which contain all points of S 2 \ M1 , S 2 \ M2 respectively. (You could
show their complements F1 , F2 are closed.)
(ii) Define ρ1 = ρ|U1 and ρ2 = ρ|U2 , and Y1 = R3 \ H1 , Y2 = R3 \ H2 where H1 = {(x, y, 0) | x ≤ 0} and
H2 = {(x, y, 0) | x ≥ 0} are half-planes. Show that ρ1 : U1 −→ S 2 ∩ H1 and ρ2 : U2 −→ S 2 ∩ H2 are
homeomorphisms.
(iii) Show that σ1 , σ2 , ρ1 , ρ2 form an atlas for S 2 .

2.4 Graphs of continuous functions

As you can see from the examples of section 2.3, it can be a pain to verify that a given subset of R3 is a surface.
You need to find charts, you need to define them explicitly, and you need to worry about open and relatively open
sets, and you need to verify all the details that they are homeomorphisms.

7
 Thankfully, there are several tricks and shortcuts to verify that you have a surface. The following proposition
gives a first shortcut. It applies when you have a graph of a function z = f (x, y).

Proposition 2.18. Let U be an open set in R2 , and let f : U −→ R be a continuous function. Then its graph

S = (x, y, z) ∈ R3 : z = f (x, y) ,


is an embedded surface.

Proof. We use f to define a chart. To this end, define σ : U −→ S to take a point (u, v) ∈ U and, thinking of R2
as the xy-plane in R3 , send (u, v) to the point of S “above” it in R3 . In other words,

σ(u, v) = (u, v, f (u, v)) .

As f is continuous, so is σ. By construction, σ is surjective onto S. Moreover σ is injective: if σ(u, v) = σ(u0 , v 0 ),

comparing x and y coordinates gives u = u0 and v = v 0 . So σ is bijective.
We can verify σ −1 : S → R2 is continuous too. Since σ sends a point in the plane to the point of S “above it”,
−1
σ projects back to the xy-plane, i.e. σ −1 (x, y, z) = (x, y). Thus σ −1 is the restriction to S of the continuous
function R3 −→ R2 which sends (x, y, z) 7→ (x, y). Hence σ −1 is continuous, and σ is a homeomorphism.
For any p ∈ S we can take a neighbourhood W = R3 , and the open set U in the plane, then σ provides a
homeomorphism U −→ S ∩ W = S. So σ is a coordinate chart for S, and as its image is all of S, it forms an atlas.

This proposition now gives us many more examples of surfaces.

Example 2.19. Planes again. Any non-vertical plane is the graph of a function, and so this gives another method
to verify that it’s a surface. For instance, the plane with equation 2y + 3z = 5 can also be given by z = 5−2y
3 hence
is the graph of the function f : R2 −→ R, f (x, y) = 5−2y3 .

Example 2.20. The paraboloid. The equation z = x2 + y 2 describes a surface, since f (x, y) = x2 + y 2 is a
continuous function. This surface is called a paraboloid. Its bowl-like shape is used by satellite dishes and car
headlights. It’s an example of a quadric surface, which we will discuss later.

Exercise 5. Show that the hemisphere given by x2 + y 2 + z 2 = 1 and z > 0 is a surface by applying proposition
2.18.

8
Example 2.21. Wild surfaces. As we’ve discussed previously a few times, continuous functions can be quite
bizarre. We can use these to build some equally bizarre surfaces.
For instance, we saw in example 1.6 of chapter 4 the Weierstrass function W : R −→ R, a continuous function
which is nowhere differentiable, and not increasing or decreasing on any interval. From it, we can define a function
f : R2 −→ R simply by f (x, y) = W (x). It’s again a continuous function, so by proposition 2.18 the graph
z = f (x, y) is a surface. Being given by the equation z = W (x), this surface is simply the graph of the Weierstrass
function (a curve in the plane), stretched out in the y-direction.
We also saw in example 1.7 of chapter 4 the Devil’s staircase, a function c : [0, 1] −→ [0, 1] which is continuous,
surjective, and constant almost everywhere. From it we may again construct a function f : R2 −→ R by f (x, y) =
c(x). The resulting graph z = f (x, y), or z = c(x), is just the Devil’s staircase (a curve in the plane), stretched out
in the y-direction.

Exercise 6. Let K ⊆ R2 be the Koch snowflake, as described in example 1.9 of chapter 4. Show that the “stretched
out Koch snowflake”
S = K × R = (x, y, z) ∈ R3 : (x, y) ∈ K


is a surface.

p
Example 2.22. The cone givenpby z 2 = x2 + y 2 and z ≥ 0 is a surface. Being given by z = x2 + y 2 , it’s the
graph of f : R2 −→ R, f (x, y) = x2 + y 2 .

2.5 A non-example of a surface

We now consider a non-example of a surface.
In general, it’s difficult to show that a subset S of R3 is not a surface. To show S is a surface, you need to
find an atlas of coordinate charts satisfying various properties — and in general there are infinitely many possible
choices for such charts. To show S is not a surface, you need to show that no such atlas of charts exists, despite
infinitely many possibilities. So you need to find some property of S that could not possibly be satisfied by any
atlas of charts.
Example 2.23. The double cone
S = {(x, y, z) ∈ R3 : x2 + y 2 = z 2 }
is not a surface.

9
 A sketch of S reveals that S looks like a surface, except near the origin 0 = (0, 0, 0). To show S is not a surface,
we will show that there cannot be any coordinate chart near 0; i.e. there is no neighbourhood W of 0 in R3 such
that S ∩ W is homeomorphic to an open set in R2 .
We prove this by contradiction. Suppose there were a coordinate chart for S near 0. Then we would have an
open set U ⊆ R2 , an open set W ⊆ R3 containing 0, and a homeomorphism σ : U −→ S ∩ W . Restricting σ if
necessary, we can assume W is an open ball about the origin, so that S ∩ W is a small portion of the double cone
near the origin.
To find a contradiction, we’ll find a property of S ∩ W which cannot be shared by U . Since S ∩ W is a
neighbourhood of 0, it contains points with positive z-coordinates, and points with negative z-coordinates. The
key property of S ∩ W is that removing the origin splits it into two disconnected pieces. Connectedness and
disconnectedness are preserved under homeomorphisms, so this means that the open set U in R2 must contain a
point p such that removing this point splits U \{p} into two disconnected pieces. But open sets in R2 cannot be
split this way: as p is open, there is an open ball about p contained in U , and removing p does not disconnect this
ball, hence does not disconnect U .

Remark 2.24. We saw above in example 2.22 that a “single” cone is a surface. As the above proof shows, the
problem is not that a cone has a vertex, it’s that two cones are joined at the vertex.
Remark 2.25. The double cone, like the paraboloid, is described by a quadratic equation, hence is an example of a
quadric. We’ll discuss these in more detail later.

Exercise 7. Verify the main idea of this proof: show that if f : A −→ B is a homeomorphism and A is path-
connected, then B is path-connected.

2.6 Transition maps

Each chart of a surface S provides a (cartographic) “map” to part of the surface.
Suppose we wander around on S, in the image of a chart σ. The image of this chart is a relatively open subset
Y of S. But then we wander far afield, and we go off the edge of the map, outside the image of σ. (Into un-chart-ed
territory!) To navigate, we must then switch to a different chart.
A transition map is a function that tells you what happens when you switch charts.
Suppose S has two coordinate charts

σ : U −→ Y = S ∩ W and σ 0 : U 0 −→ Y 0 = S ∩ W 0 ,

where U, U 0 are open sets in R2 , and W, W 0 are open sets in R3 , so that Y, Y 0 are relatively open subsets of S.
Suppose that the images Y, Y 0 overlap, so that it’s possible to switch from the chart σ to the chart σ 0 . Then the
intersection Y ∩ Y 0 is also a relatively open subset of S, since Y ∩ Y 0 = S ∩ W ∩ W 0 , and W ∩ W 0 is an open set in
R3 .

10
 In this case, both σ and σ 0 provide homeomorphisms from subsets of R2 , to the set Y ∩ Y 0 = S ∩ W ∩ W 0 . If
we let V = σ −1 (Y ∩ Y 0 ) and V 0 = σ 0−1 (Y ∩ Y 0 ), then restricting σ to V yields a homeomorphism

σ|V : V −→ Y ∩ Y 0 .

Similarly, letting V 0 = σ 0−1 (Y ∩ Y 0 ), the restriction of σ 0 to V 0 yields a homeomorphism

σ 0 |V 0 : V 0 −→ Y ∩ Y 0 .

Now V, V 0 are subsets of U, U 0 respectively, and they are also open. If you think of U, U 0 as pages of the map you
are reading as you navigate the surface, then V ⊆ U is the portion of the page U which overlaps with the page U 0 ,
and V 0 ⊆ U 0 is the portion of the page U 0 which overlaps with the page U .
If we put the two homeomorphisms V −→ Y ∩ Y 0 and V 0 −→ Y ∩ Y 0 together, we obtain a homeomorphism
V −→ V 0 . This is what a transition map is, as we now define formally.

Definition 2.26. The composition

Φ = σ 0−1 ◦ σ : V −→ V 0
is a map between open subsets of R2 called a transition map from σ to σ 0 , or a reparametrisation.

Thus, Φ provides the relationship between the overlap regions V and V 0 . A point (u, v) ∈ V corresponds to
Φ(u, v) ∈ V 0 . They both describe the same point on the surface S. In other words,

σ(u, v) = σ 0 (Φ(u, v)).

Remark 2.27. There is a lot of notation in defining transition maps. Take the time to digest it!
In fact, strictly speaking, we should have defined Φ with even more notation! We can’t actually compose σ with
σ 0−1 ; the domains and ranges don’t match. Only the restrictions σ|V and σ 0 |V are composable, so strictly speaking,
−1
Φ = (σ 0 |V 0 ) ◦ (σ|V ) : V −→ V 0 .

It’s not unusual to write compositions in the “slack” way we did in definition 2.26; compositions are then taken to
be defined where they make sense.
Remark 2.28. A transition map is in some ways easier to deal with than a coordinate chart, because it maps from a
subset of R2 , to another subset of R2 . It bypasses the surface, and goes straight from one chart to the other, never
going into three dimensions!
In fact, the “modern” viewpoint of differential geometry is based precisely on dealing only with points in R2
and transition maps. In this perspective, we don’t really care how a surface is embedded in R3 (or even if it can
be), as long as we can understand the geometry from looking in 2 dimensions. Riemannian manifolds are based
on this idea. We’ll discuss this viewpoint a little towards the end of the course; and it will feature heavily in any
further course you take in differential geometry.
Remark 2.29. Transition maps generalise reparametrisations of curves. Recall that a reparametrisation of γ(t) :
(a, b) −→ R3 is given by δ(u) : (c, d) −→ R3 via a φ : (c, d) −→ (a, b), and then δ(u) = γ(φ(u)). The reparametri-
sation φ works just like the transition map Φ.1 With curves, the reparametrisation sends the points of one interval
(c, d) to the corresponding points of the other (a, b); with surfaces, the transition map sends the points of one open
set V to the corresponding points of the other open set V 0 . With curves we have δ(u) = γ(φ(u)); with surfaces we
have σ(u, v) = σ 0 (Φ(u, v)).
1 However, we required φ to be a diffeomorphism, while Φ here is only a homeomorphism. We will shortly consider differentiability

of all these functions.

11
2.7 Examples of transition maps

Example 2.30. Let’s return to the cylinder of example 2.15 and compute a transition map! The cylinder is

S = (x, y, z) ∈ R3 | x2 + y 2 = 1


and has an atlas of two charts σ1 , σ2 given by σ1 = σ|U1 , σ2 = σ|U2 , where

σ : R2 −→ R3 , σ(u, v) = (cos u, sin u, v)

U1 = (u, v) ∈ R2 | 0 < u < 2π = (0, 2π) × R


U2 = (u, v) ∈ R2 | −π < u < π = (−π, π) × R.



The first chart σ1 maps U1 homeomorphically onto the whole cylinder except for the line L1 = {(1, 0, z) | z ∈ R}.
The second chart σ2 maps U2 homeomorphically onto the whole cylinder except the line L2 = {(−1, 0, z) | z ∈ R}.
Let these images be Y1 = S\L1 and Y2 = S\L2

Now the overlap region Y1 ∩ Y2 = S\(L1 ∪ L2 ) is the complement of both lines L1 and L2 . It’s a disconnected
subset of S, consisting of the “front” and “back” halves of the cylinder. The “front” is defined by y > 0; the “back”
by y < 0.
Applying σ1−1 to Y1 ∩ Y2 yields a subset V1 of U1 which consists of U1 with the line u = π (which corresponds
to L2 ) removed. That is,

V1 = σ −1 (Y1 ∩ Y2 ) = {(u, v) | 0 < u < π or π < u < 2π} .

Similarly, applying σ2−1 to Y1 ∩ Y2 gives V2 ⊂ U2 consisting of all of U2 , except the line u = 0, which corresponds
to the line L1 . So
V2 = σ −1 (Y1 ∩ Y2 ) = {(u, v) | −π < u < 0 or 0 < u < π} .
Note that V1 , V2 are both disconnected — and their two halves correspond to the front and back halves of Y1 ∩ Y2 .
The transition map Φ : V1 −→ V2 simply takes points in V1 to the corresponding maps in V2 , via σ1 and then
σ2−1 . Let’s now see precisely what the transition maps do. We’ll take two cases, corresponding to the two connected
components of these sets.
First, let’s take a point (u, v) ∈ V1 with 0 < u < π. This is mapped via σ1 to the point (cos u, sin u, v) ∈ Y1 ∩ Y2 .
Since 0 < u < π, we have sin u > 0, so σ1 (u, v) lies in the “front” y > 0 half of Y1 ∩ Y2 . Now to apply σ2−1 , we need
to find a point (u0 , v 0 ) ∈ V2 such that σ2 (u0 , v 0 ) = (cos u, sin u, v), i.e. (cos u0 , sin u0 , v 0 ) = (cos u, sin u, v).
Now there is an obvious (u0 , v 0 ) which satisfies this equation, namely u = u0 and v = v 0 ! Since 0 < u < π, taking
(u , v ) = (u, v) gives 0 < u0 < π, hence (u0 , v 0 ) is a point in V2 such that σ2 (u0 , v 0 ) = σ1 (u, v), or equivalently,
0 0

σ2−1 (σ1 (u, v)) = (u0 , v 0 ). Thus the transition map is given simply by Φ(u, v) = (u, v).
Things are slightly trickier in the second case, where (u, v) ∈ V1 with π < u < 2π. Again σ1 (u, v) =
(cos u, sin u, v) ∈ Y1 ∩ Y2 . But now as π < u < 2π we have sin u < 0, and hence σ1 (u, v) lies in the y < 0 half of
Y1 ∩ Y2 . Again, to apply σ2−1 to this point, we need to find a (u0 , v 0 ) ∈ V2 such that σ2 (u0 , v 0 ) = (cos u, sin u, v), i.e.
(cos u0 , sin u0 , v 0 ) = (cos u, sin u, v).

12
 It follows that v 0 = v, but this time we cannot take u0 = u, because π < u < 2π, while V2 is defined by
u ∈ (−π, 0) ∪ (0, π). However, the fact that (cos u0 , sin u0 ) = (cos u, sin u) means that u and u0 must differ by an
0

integer multiple of 2π. Since π < u < 2π and −π < u0 < π, it follows that we must take u0 = u − 2π. Taking
(u0 , v 0 ) = (u − 2π, v), we have (u0 , v 0 ) ∈ V2 and σ1 (u, v) = σ2 (u0 , v 0 ), or equivalently, σ2−1 (σ1 (u, v)) = (u0 , v 0 ). So
the transition map is given by Φ(u, v) = (u − 2π, v).
To summarise, the transition map is a map from V1 = {(u, v) | u ∈ (0, π) ∪ (π, 2π)} to V2 = {(u, v) | u ∈
(−π, 0) ∪ (0, π)} defined by 
(u, v) u ∈ (0, π),
Φ(u, v) =
(u − 2π, v) u ∈ (π, 2π).

Exercise 8. Recall we defined two surface patches σ1 , σ2 on the sphere from the function

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

by taking the sets

n  π π o n  π π o
U1 = (u, v) ∈ R2 | u ∈ − , , v ∈ (0, 2π) , U2 = (u, v) ∈ R2 | u ∈ − , , v ∈ (−π, π)
2 2 2 2
and defining σ1 = σ|U1 , σ2 = σ|U2 . Let Y1 , Y2 be the images of σ1 , σ2 .
(i) Determine Y1 ∩ Y2 .

(ii) Determine V1 = σ −1 (Y1 ∩ Y2 ) and V2 = σ2−1 (Y1 ∩ Y2 ).

(iii) Determine the transition map Φ : V1 −→ V2 .

3 Defining smooth surfaces

This is a course on differential geometry, and so we want to be able to do things involving derivatives on surfaces.
We want to consider tangent lines, planes, and various notions of curvature.
But so far, our definition of a surface has nothing “differential” about it. The definition involves homeomor-
phisms. A homeomorphism is a continuous function, with continuous inverse, but need not be differentiable. So
the definition involves continuity but not differentiability or smoothness.
In fact, in example 2.21, we saw some very non-smooth examples of surfaces!
So, we now discuss smooth surfaces. We may have a fairly clear intuitive idea of what would constitute a smooth
surface: it shouldn’t have corners, folds, creases, and so on. We will try to make these ideas mathematically precise.

3.1 Examples of non-smooth surfaces

To see how a surface can fail to be smooth, let’s consider some examples.

Example 3.1. The cone again. In examplep2.22, we saw that the cone C defined by z 2 = x2 + y 2 and z ≥ 0 is a
surface. Indeed, it’s the graph of f (x, y) = x2 + y 2 . It has an atlas given by a single coordinate chart
 p 
σ : R2 −→ C, σ(u, v) = u, v, u2 + v 2 .

A cone has a point where it is not very smooth at all: the tip √ of the cone, also known as its vertex or apex,
where (u, v) = (0, 0). This corresponds to the fact that f (u, v) = u2 + v 2 is not differentiable there.

Example 3.2. A folded plane. Consider the graph of the function f : R2 −→ R given by f (x, y) = |y|. This is a
continuous function, so by proposition 2.18, the graph forms a surface S: it’s defined by the equation z = |y|, and
an atlas is given by a single coordinate chart

σ : R2 −→ S, σ(u, v) = (u, v, |v|).

13
 This surface is not smooth along v = 0. It has a “fold” there. This corresponds to the fact that the function
f (u, v) = |v| is not differentiable along the line v = 0.

3.2 Smooth functions and open sets

Since surfaces are defined via coordinate charts σ : U −→ R3 , it makes sense that coordinate charts for a smooth
surface ought to be smooth functions. So let’s discuss what “smooth” means for multivariable functions.
Recall the definition of a smooth multivariable function.2 A function F : Rn −→ Rm is C k smooth if all partial
derivatives of all components of F up to order k exist and are continuous. This F is C ∞ smooth if all partial
derivatives of all components of F of any order exist and are continuous. When we say smooth in this course, by
default we mean C ∞ smooth.
However, there is a technicality to consider about domains. When we defined smoothness for single-variable
functions, we considered functions f : (a, b) −→ R defined on an interval (a, b), not necessarily all of R. But when
we defined smoothness for multivariable functions, our function was defined on all of Rm . Can we give a version of
smoothness for a function only defined on a subset of Rm ?
Yes we can, using open sets. In order to take a partial derivative of a function f at a point x ∈ Rm , you need
to look at values of f near x. To take partial derivatives of f with respect to different variables, you need to look
at values of f near x in different directions from x. And to take directional derivatives in different directions, you
need to look at values of f near x in all possible directions. So f needs to be defined at points around x in all
possible directions. That is, f needs to be defined in a ball around x0 .
Therefore, f should be defined on a set U where every around every x ∈ U there is a ball Br (x) contained in U .
This type of set is nothing but an open set.
So open sets give the right type of set on which to define differentiable functions. They are the right type of set
for many things. Open sets are good. Let’s make a formal definition.

Definition 3.3. Let U ⊆ Rm be an open set. A function f : U −→ Rn is (C ∞ ) smooth if each component of f

has partial derivatives of all orders.

3.3 Directional derivatives with coordinate charts

Directional derivatives should be something you understand from multivariable calculus; we also discussed them
at length in the Introduction and Background chapter 1, in sections 3.2.2–3.2.3. We’re now going to see how they
work for the coordinate charts of a surface.
Recall that given a smooth function F : Rn −→ Rm , with components f1 , . . . , fm , its total derivative or matrix
derivative consists of its partial derivatives,
 ∂f ∂f1 ∂f1

∂x1
1
∂x2 · · · ∂x n
 ∂f2 ∂f2 ∂f2 
 ∂x1 ∂x2 · · · ∂x 
n 
DF =  .
 . . .
 .. .. .. .. 
∂fm ∂fm ∂fm
∂x1 ∂x2 ··· ∂xn ,
2 This is essentially stuff you are supposed to know from your previous multivariable calculus course; I also covered it in the

Introduction and Background chapter 1, in sections 3.2.4–3.2.5.

14
and the directional derivative of F at x ∈ Rn in the direction v ∈ Rn is given by multiplying DF by v.
So, let σ(u, v) : U −→ R3 be a surface patch defined on an open set U ⊆ R2 . We write the x, y, z components of
σ as x(u, v), y(u, v), z(u, v) respectively. This will be our default notation for surface patches, as it nicely indicates
how the x, y, z coordinates of a surface patch depend on the u, v coordinates of U ⊂ R2 .
 
σ(u, v) = x(u, v), y(u, v), z(u, v)

A partial derivative with respect to, say, u, can be denoted as

∂σ
or σu .
∂u
We generally prefer the subscript notation, as it uses fewer symbols. Thus
   
σu (u, v) = xu (u, v), yu (u, v), zu (u, v) , σv (u, v) = xv (u, v), yv (u, v), zv (u, v) .

We can also higher order derivatives using either notation. As σ is smooth, the order of differentiation does not
matter, so for instance
∂2σ ∂2σ
= σuv = σvu =
∂u∂v ∂v∂u
and

σuu (u, v) = (xuu (u, v), yuu (u, v), zuu (u, v))
σuv (u, v) = (xuv (u, v), yuv (u, v), zuv (u, v))
σvv (u, v) = (xvv (u, v), yvv (u, v), zvv (u, v)) .

The total derivative of σ is thus the 3 × 2 matrix

 
xu xv
Dσ =  yu yv  ,
zu zv

and the directional derivative of σ in the direction w = (a, b) in R2 is then given by

       
xu xv   axu + bxv xu xv
a
Dσ w =  yu yv  =  ayu + byv  = a  yu  + b  yv  = aσu + bσv .
b
zu zv azu + bzv zu zv

As w = (a, b) varies over all possible vectors in R2 , the directional derivatives Dσ v = aσu + bσv vary over all
possible linear combinations of σu and σv . Thus we have the following proposition.

Proposition 3.4. The set of all directional derivatives of σ at (u, v) is equal to the span of σu (u, v) and σv (u, v).

Exercise 9. If you’re not familiar with any of the above, please revise your multivariable calculus! We’ll be needing
all the above ideas and notation throughout the rest of the course.

3.4 Smooth coordinate charts are not enough

As it turns out, even if each surface patch σ of a surface is a C ∞ -smooth function, the surface may not be smooth
in any reasonable sense.
We saw something similar for curves in section 2 of chapter 2. Even when the parametrisation γ(t) was a smooth
function, the curve might have corners or bumps. For instance, recall the curve γ(t) = (t2 , t3 ). Although γ is a
perfectly smooth function, its image C (also given by y = ±x3/2 ) has a cusp at the origin.

15
 1.0

0.5

0.2 0.4 0.6 0.8 1.0

-0.5

-1.0

To see something similar for surfaces, we can adapt this γ into a surface patch, as in the following example.

Example 3.5. A cusped surface. Define

σ : R2 −→ R3 by σ(u, v) = (u, v 2 , v 3 ),

and let S ⊂ R3 be its image. This σ is a C ∞ smooth function, and it defines a coordinate chart giving an atlas for
S. However, the surface S has a “cusp fold” along the line (u, 0, 0), u ∈ R, corresponding to v = 0.

This surface S is defined by the equation z = ±y 3/2 . Each cross-section u = constant (or x = constant) is gives
the cusp curve just discussed. Indeed, we can describe S as R × C.

What is it about σ that causes S to fail to be smooth?

When discussing the cusp curve γ, we found that γ̇(0) was zero, and this made the cusp possible. Indeed, we
defined regular curves to be those curves such that γ̇(t) is never zero.
Similarly, if we look at the partial derivatives of σ here, we have

σu (u, v) = (1, 0, 0), σv (u, v) = (0, 2v, 3v 2 ),

so at any point where v = 0 we have σv = (0, 0, 0).

Exercise 10. Prove the assertions made above about the cusped surface S.
(i) Show that σ(u, v) = (u, v 2 , v 3 ) defines a coordinate chart giving an atlas for S.

16
 (ii) Show that S consists of the points (x, y, z) ∈ R3 such that z = ±y 3/2 .

3.5 Nonzero partial derivatives are not enough

So, what do we need for a smooth surface?
The condition for a parametrised curve γ(t) to be regular was that γ̇(t) 6= 0.
We might guess, by analogy, that the partial derivatives of σ should never be zero: σu 6= 0, σv 6= 0. However,
this isn’t right either, as we see in the following example.

Example 3.6. Let S be the image of the function

ρ : R2 −→ R3 , ρ(u, v) = u − v, (u + v)2 , (u + v)3 .

One can show that ρ is a homeomorphism R2 −→ S, and hence yields a coordinate chart forming an atlas for S.
So this S is a surface.
Moreover, we can compute the partial derivatives

ρu (u, v) = 1, 2(u + v), 3(u + v)2 , and ρv (u, v) = −1, 2(u + v), 3(u + v)2 ,

and observe that ρu is never zero (its x-coordinate is 1), nor is ρv (its x-coordinate is −1).
But what is this surface S? It turns out to be precisely the same surface as in the previous section with a “cusp
fold” along the line (x, 0, 0), x ∈ R, or equivalently, u = −v.
To see that the surface defined by ρ, is the same one as the surface defined by σ from the previous section, you
can check that ρ is a σ with a change of coordinates, or reparametrisation. If we define

F : R2 −→ R2 by F (u, v) = (u − v, u + v),

then
ρ(u, v) = u − v, (u + v)2 , (u + v)3 = σ(u − v, u + v) = σ(F (u, v)),

so σ and ρ are equal after the change of coordinates provided by F .

You can verify in the exercise below that F is a diffeomorphism R2 to R2 — in fact, it’s an invertible linear
map. So ρ and σ have precisely the same image, and define precisely the same surface.

So, if ρ is smooth and its partial derivatives are never zero, yet it does not define a smooth surface, what is
wrong with ρ?
With the previous parametrisation σ, we found that σv = 0 when v = 0. As ρ is the same as σ after a change of
coordinates, some derivative of ρ must be zero — but it’s not a partial derivative, rather it’s a directional derivative.
It turns out that the directional derivative of ρ in the direction (1, 1) is zero along the line u + v = 0. We
calculate
   
1 −1 1 −1
Dρ(u, v) =  2(u + v) 2(u + v)  so when u + v = 0, Dρ(u, v) = Dρ(u, −u) = 0 0  .
3(u + v)2 3(u + v)2 0 0

The directional derivative in the direction (1, 1) at such points (u, −u) is
   
  1 −1   0
1 1
Dρ(u, −u) = 0 0  = 0
1 1
0 0 0

Therefore, when u + v = 0, the directional derivative of ρ in the direction (1, −1) is zero.
So, to have a smooth surface, we don’t just want the partial derivatives of a coordinate charts to be nonzero.
We want all the directional derivatives of our coordinate charts to be nonzero.

Exercise 11. Verify the unproved assertions in the above example.

(i) Show that ρ is a homeomorphism R2 −→ S, and that the image S of ρ is a surface.

(ii) Show that F is a diffeomorphism (in fact, a linear map) R2 → R2 , and find its inverse.

17
 √
(iii) In fact, show that F is given by a rotation of π/4 about the origin, followed by a dilation of factor 2 from
the origin. (Such a transformation is known as a spiral symmetry.)

(iv) Using this description of F , explain why the singularity of σ(u, v) along v = 0 corresponds to the singularity
of ρ(u, v) along u = −v, and why the partial derivative of σ in the v-direction corresponds to the partial
derivative of ρ in the (1, 1) direction

3.6 The definition of a smooth surface

We now have the basic idea of a smooth surface.
As we saw in chapter 2, for a curve γ(t) to be regular, we require that its derivative γ̇(t) 6= 0 everywhere.
And as we saw just above, for a surface patch σ(u, v) : U −→ R3 , we should also require that its derivatives are
nonzero — by which we mean, its directional derivatives.
This makes sense. Since σ is defined on a 2-dimensional domain U , there are many directions in which we can
take derivatives, and there is no particular reason to privilege the partial derivatives, which are derivatives in the
u and v directions. The u and v directions are just an artifact of our choice of coordinates. The smoothness of a
surface should be independent of our choice of coordinates used to parametrise it!
So what we really want for a smooth surface is that all directional derivatives of σ should be nonzero.3 As we
saw in section 3.3, the directional derivative of σ at (u, v) in the direction w = (a, b) is given by

Dσ(u, v) w = aσu (u, v) + bσv (u, v).

So for the directional derivative of σ to be nonzero at (u, v) in all nonzero directions w, we require that all linear
combinations aσu (u, v) + bσv (u, v) of the vectors σu and σv are nonzero, for all (a, b) 6= (0, 0).
There is a name for a set of vectors whose linear combinations are never zero: linearly independent. There-
fore, we want σu (u, v) and σv (u, v) to be linearly independent vectors. Hence we make the following defini-
tions.
Definition 3.7. A smooth coordinate chart σ : U −→ Rn is regular at (u, v) ∈ U if σu (u, v), σv (u, v) are linearly
independent.
The chart σ is regular if it is regular at all points of U .

Definition 3.8. A smooth surface is a surface with an atlas of regular coordinate charts.

Remark 3.9. Although our discussion involved examples in R3 , this definition (like definition 2.1) works in any Rn .
In this case, σu (u, v) and σv (u, v) are n-dimensional vectors; we can still require them to be linearly independent.
Note the domain U of σ is always an open subset of R2 .
Remark 3.10. Although we came to it via a circuitous route, this definition is a direct generalisation of the definition
of a regular curve. For a curve γ(t), requiring that γ̇(t) 6= 0 ensures that all “directional derivatives” of γ are nonzero.
So we could have defined a regular curve as one whose “directional derivatives” are all nonzero.
Moreover, a single vector forms a linearly independent set if and only if it is nonzero! So we could have equally
defined a regular curve γ(t) to be one whose derivative γ̇(t) always forms a linearly independent set.
These would have been strange ways to define a regular curve, but they are equally valid.

3 Precisely, what we mean is that every directional derivative of σ, in the direction of a vector v 6= 0, should be nonzero. The

“directional derivative in the (0, 0) direction” is not much of a directional derivative anyway!

18
Exercise 12. Prove the claim in the discussion above, that a single vector v ∈ R2 forms a linearly independent set
if and only if it is nonzero.

3.7 Characterising regular surface patches

We defined a coordinate chart σ : U −→ R3 to be regular at (u, v) if σu , σv are linearly independent. We can give
some equivalent characterisations of this condition, just using linear algebra.
Recall Dσ is a 3 × 2 matrix, and its columns are σu , σv : if σ(u, v) = (x(u, v), y(u, v), z(u, v)) then
   
xu xv | |
Dσ =  yu yv  = σu σv  .
zu zv | |

In other words, the columns of Dσ are σu and σv ; and the regularity condition is that these columns are linearly
independent.
With some linear algebra, we can reformulate this condition a few different ways. Let’s state our result for 3 × 2
matrices in general.

Lemma 3.11. Let A be a 3 × 2 matrix with real entries. The following are equivalent:
(i) the column vectors are linearly independent vectors in R3 ;

(ii) A has rank 2;

(iii) the linear transformation TA : R2 −→ R3 corresponding to A is injective.

Proof. Let the column vectors of A be C1 , C2 .

To see (i) and (ii) are equivalent, we note that C1 and C2 span a 2-dimensional subspace of R3 if and only if
they are linearly independent. Since the rank of the column space of A is equal to the rank, we conclude that
C1 , C2 are linearly independent if and only if A has rank 2.
To see (i) and (iii) are equivalent, suppose TA is not injective. Then TA (w) = 0 for some nonzero w = (a, b) ∈ R2 .
So  
| |  
a
0 = TA (w) = C1 C2  = aC1 + bC2 ,
b
| |
hence C1 , C2 are linearly dependent. Conversely, if C1 , C2 are linearly dependent then aC1 + bC2 for some w =
(a, b) 6= 0, and the calculation above shows TA (w) = 0, so TA is not injective. Thus, TA is not injective if and only
if C1 , C2 are linearly dependent; and hence, TA is injective if and only if C1 , C2 are linearly independent.

Applying lemma 3.11 to the derivative Dσ of a surface patch, we immediately have the following.

Proposition 3.12. Let σ : U −→ R3 be a surface patch. The following are equivalent:

(i) σ is regular at (u, v)
(ii) Dσ(u, v) has rank 2

(iii) the linear transformation of Dσ(u, v) is injective.

In general, a map Rm −→ Rn whose derivative defines an injective linear transformation is called an immersion.
In more advanced courses on geometry and topology — and in active research — immersions are studied in detail.

19
4 Verifying that you have a smooth surface
4.1 Many things to verify
Suppose you’re given a subset S of R3 , perhaps given by an equation, perhaps given as the image of a map
σ : R2 −→ R3 . How do you verify that S is a smooth surface?
Starting from first principles, you need to find an atlas of coordinate charts. So you need to find homeomorphisms
σ : U −→ Y from open sets U ⊆ R2 to relatively open subsets Y = S ∩ W of S, i.e. where W ⊆ R3 is open. You
need enough such maps σ to form an atlas, covering the whole surface.
Then, to show that S is a smooth surface, you then need to show that each chart is regular, by showing it is a
smooth function, and that its partial derivatives are linearly independent at each point.
Let’s list out all the things you need to verify: it’s quite a checklist.
(i) Find open sets Ui ⊆ R2 , relatively open sets Yi ⊆ S which cover all of S, and maps σi : Ui −→ Yi which you
will show are surface patches.
(ii) Show that each such map σ : U −→ Y is a homeomorphism. This in turn requires showing that
(a) σ is injective,
(b) σ is surjective,
(c) σ is continuous,
(d) σ −1 is continuous.
(iii) Show that σ is regular. This in turn requires showing that
(a) σ is a smooth function,
(b) the partial derivatives σu and σv are linearly independent at all points in U .
This can take some effort. Step (i) sometimes requires some insight into the geometry of the surface. Some
of step (ii) is usually straightforward but part (d) can be painful. Then (iii) requires some algebra, which can be
tricky to verify at all points of U simultaneously.
It is character-building and instructive to make this effort a few times. But there are some shortcuts available.
Some of these tricks can dramatically simplify the process. We’ll discuss some of these in the following sections,
exploring some examples along the way.

4.2 Examples from first principles

We begin with some examples which can be handled from first principles.

Example 4.1. The cylinder again. We again take the unit cylinder S, defined by x2 + y 2 = 1. In example 2.15
we showed S is a surface, with two coordinate charts σ1 , σ2 , both restrictions of the map

σ : R2 −→ R3 , σ(u, v) = (cos u, sin u, v).

(It took some effort to show that σ1 , σ2 are homeomorphisms!) In example 2.30 we computed the transition map.
Let’s now show that S is a smooth surface. Clearly σ is a smooth function, with

σu = (− sin u, cos u, 0) σv = (0, 0, 1).

We observe that σu , σv are always linearly independent: if aσu + bσv = 0 then we have (−a sin u, a cos u, b) = 0, so
a = b = 0.
Hence both coordinate charts, being restrictions of σ, are regular, and S is a smooth surface.

Example 4.2. The cusped surface again. Recall the cusped surface z = ±y 3/2 of example 3.5. We saw it’s a
surface with an at atlas consisting of a single chart

σ : R2 −→ R3 , σ(u, v) = (u, v 2 , v 3 ).

Moreover, σ is smooth, with partial derivatives σu = (1, 0, 0) and σv = (0, 2v, 3v 2 ).

20
 When v = 0, we have σv = 0, hence σu and σv are linearly dependent. So σ is not regular when v = 0.
In example 3.6 We also saw another smooth chart

ρ : R2 −→ R3 , ρ(u, v) = u − v, (u + v)2 , (u + v)3 ,

with partial derivatives ρu = 1, 2(u + v), 3(u + v)2 and ρv = −1, 2(u + v), 3(u + v)2 .
When u + v = 0 we have ρu = (1, 0, 0) and ρv = (−1, 0, 0), which are linearly dependent vectors. So σ is not
regular at such points.
Thus, neither σ nor ρ provides a regular chart for S.

4.3 Trick #1: Cross products verify linear independence

To check a coordinate chart is regular, we must check that the two 3-dimensional vectors σu , σv are linearly
independent.
It’s not difficult to check the linear independence of two vectors: they are linearly dependent if and only if one
is a scalar multiple of the other. But when σu and σv are given by complicated expressions in u and v, it can be
difficult to check this for all u and v simultaneously.
The following proposition gives a trick, specific to three dimensions, to verify linear independence.

Proposition 4.3. Two vectors v, w ∈ R3 are linearly independent if and only if their cross product is nonzero,
i.e. v × w 6= 0.

Proof. First, we note that if one or both of v, w is zero, then v, w are linearly dependent, and v × w = 0. So we
may assume v, w are both nonzero.
We will show that v × w = 0 if and only if v, w are linearly dependent.
If v × w = 0, then since ||v × w|| = ||v|| ||w|| sin θ and v, w are both nonzero, we have sin θ = 0, where θ is
the angle between v and w. Hence θ = 0 or π, so v and w are parallel or anti-parallel, and v is a scalar multiple
of w. So v, w are linearly dependent.
Conversely, if v, w are linearly dependent, then av + bw = 0 for some (a, b) 6= (0, 0). Without loss of generality
suppose b 6= 0, so w = − ab v. Then, as desired,
 a  a
v × w = v × − v = − (v × v) = 0.
b b

Exercise 13. The cylinder once more. Return to the cylinder x2 + y 2 = 1 of example 2.15 and the charts σ1 , σ2
which are restrictions of σ(u, v) = (cos u, sin u, v), which we know form an atlas.
Let’s now verify we these charts are regular using a cross product. In example 4.1 we found that σu =
(− sin u, cos u, 0) and σv = (0, 0, 1). From these we can compute

σu × σv = (cos u, sin u, 0),

which is never zero. So at all points (u, v) then σu , σv are linearly independent, and σ is a regular surface patch.

Exercise 14. The cusped surface once more. We return to the cusped surface z = ±y 3/2 . In example 3.5, we found
an atlas consisting of the single chart σ(u, v) = (u, v 2 , v 3 ), and calculated σu = (1, 0, 0) and σv = (0, 2v, 3v 2 ).
In example 3.6 we found another atlas with the chart ρ(u, v) = (u − v, (u + v)2 , (u + v)3 ), and calculated
ρu = (1, 2(u + v), 3(u + v)2 ) and ρv = (−1, 2(u + v), 3(u + v)2 ).
In example 4.2 we found σu , σv are linearly dependent when v = 0, and ρu , ρv are linearly dependent when
u + v = 0. Verify the same result using cross products.

Exercise 15. The sphere again. In example 2.17 we
defined two charts σ1 , σ2 on
the sphere as restrictions of
σ(u, v) = (cos u cos v, cos u sin v, sin u) to U1 = − π2 , π2 × (0, 2π) and U2 = − π2 , π2 × (−π, π) respectively.
Show that these charts are regular. You might recall these patches covered the entire sphere except the poles.

21
If we extend the patches to the poles, would the patches be regular?

4.4 Trick #2: Graphs of smooth functions are smooth surfaces

We saw in proposition 2.18 that the graph z = f (x, y) of a continuous function f : R2 −→ R is a surface. Similarly,
the graph of a smooth function is a smooth surface, as we prove now.

Proposition 4.4. The graph S = {(x, y, z) | z = f (x, y)} of a smooth function f : R2 −→ R is a smooth surface.

Proof. Since smooth functions are continuous, by proposition 2.18, S is certainly a surface. Moreover, an
atlas is given by a single chart σ : R2 −→ S where σ(u, v) = (u, v, f (u, v)). To check σ is regular, we compute
σu × σv = (1, 0, fu ) × (0, 1, fv ) = (−fu , −fv , 1), which is never zero. So S is a smooth surface.

Let’s apply this trick a few times.

Example 4.5. Planes once more. As in example 2.19, any non-vertical plane is the graph of a linear function.
Linear functions are smooth, so proposition 4.4 tells us these planes are smooth surfaces.

Example 4.6. The paraboloid again. As in example 2.20, the graph of f (x, y) = x2 + y 2 is a paraboloid. As f is
smooth, the paraboloid is a smooth surface.

Exercise 16. Show that the hemisphere given by x2 + y 2 + z 2 = 1 and z > 0 is a smooth surface by applying
proposition 4.4.

Exercise 17. Show that the punctured cone given by z 2 = x2 + y 2 and z > 0 is a smooth surface by applying
proposition 4.4.

4.5 Trick #3: Level sets are often smooth surfaces

Another useful way to show a set is a smooth surface is to present it as the level set of a function.
In our earlier discussion of curves (section 4.2 of chapter 2), we saw that curves could sometimes be represented
as the level set of a function. Recall a level set f −1 (c) of a function f : R2 −→ R is the set of all (x, y) ∈ R2 such
that f (x, y) = c, for some constant c. We saw (theorem 4.27) that, essentially, an equation f (x, y) = c describes a
curve provided ∇f 6= 0.
Similarly, if we have a function f : R3 −→ R, we can consider the level set f −1 (c), i.e. the set of all (x, y, z) ∈ R3
such that f (x, y, z) is equal to some constant c. A surface given as a level set f −1 (c) can also be given by an
equation f (x, y, z) = c.
Some level sets of functions f : R3 −→ R are surfaces; some are not. Which level sets are surfaces, and which
aren’t? The answer is similar to the answer for curves: we consider the critical points of the function. For instance:
• If f (x, y, z) = 2x + 3y − 7z then all its level sets are parallel planes. Its gradient ∇f = (2, 3, −7) is never zero.
• If f (x, y, z) = x2 + y 2 then its level sets are cylinders, except that f −1 (0) is the z-axis. The gradient
∇f = (2x, 2y, 0) is zero precisely along this line.
• If f (x, y, z) = x2 + y 2 + z 2 then its level sets are spheres, except that f −1 (0) is a point, the origin. The
gradient ∇f = (2x, 2y, 2z) is zero only at the origin.
• If f (x, y, z) = x2 + y 2 − z 2 then the level set f −1 (0) is a double cone, with equation z 2 = x2 + y 2 ; as we’ve
seen previously, the double cone fails to be a surface at the origin. However, as we’ll see later, all other level
sets of f are surfaces called hyperboloids. The gradient ∇f = (2x, 2y, −2z) is zero only at the origin.
• If f (x, y, z) = xy, then f −1 (0) is given by two intersecting planes, with equation x = 0 and y = 0; but all
other level sets are surfaces. The gradient ∇f = (y, x, 0) is zero along the z-axis, which is where the planes of
f −1 (0) intersect.

22
 We can state the result — and a bit more cleanly than for curves, because we now know about open sets.

Theorem 4.7. Let S ⊂ R3 . Suppose that for all p ∈ S, there is an open set p ∈ W ⊆ R3 and a smooth f : W −→ R
such that
S ∩ W = {(x, y, z) : f (x, y, z) = 0},
and ∇f (p) 6= 0. Then S is a smooth surface.

We’ll prove this theorem a little later, using the inverse function theorem.
Remark 4.8. The theorem allows that different parts of the surface can be described as different level sets. It only
requires that at each point, we can find a neighbourhood W such that S ∩ W is the level set of a function. They
might be different functions at different neighbourhoods!
Often, though, we only need one equation, and we can take W to just be the whole of R3 . Then we immediately
have the following corollary. Usually this corollary is all you need.
Corollary 4.9. Let S ⊂ R3 . Suppose there is a smooth function f : R3 −→ R such that S = f −1 (0), and ∇f (p) 6= 0
for every p ∈ S. Then S is a smooth surface.
Remark 4.10. Although in the statement of the theorem and corollary S = f −1 (0), the same result applies to any
level set f −1 (c); one can simply adjust the function by a constant. For any constant c, the level set f −1 (c) of f is
equal to the level set g −1 (0) of the function g(x, y, z) = f (x, y, z) − c.
Let’s see some applications of this theorem.

Example 4.11. The unit sphere once more. S = {(x, y, z) ∈ R3 : x2 + y 2 + z 2 = 1} is the level set f −1 (1) of the
function f (x, y, z) = x2 + y 2 + z 2 . We compute ∇f = (2x, 2y, 2z), which is zero only at the origin, not on S. Hence
∇f 6= 0 at every point on S, and S is a smooth surface.

Remark 4.12. Note how much faster it is to verify the sphere is a smooth surface using theorem 4.7, than using
explicit charts (example 2.17) and taking cross products of partial derivatives (exercise 15)!

Exercise 18. Show using theorem 4.7 that the cylinder defined by x2 + y 2 = 1 is a smooth surface.
How does this compare to the effort of constructing charts (example 2.15), and checking linear independence of
partial derivatives directly (example 4.1) or using cross products (exercise 13) ?

Exercise 19. Let f (x, y, z) = x2 + y 2 − z 2 . Show that every level set f −1 (c) for c 6= 0 is a smooth surface, and the
level set f −1 (0) with the origin removed is also a smooth surface.

4.6 Trick #4: From continuous to homeomorphism

Our final trick is the deepest and most subtle.
Suppose you have your map σ : U −→ Y , defined on an open set U in R2 , with range Y ⊂ R3 , and you want to
show σ is a surface patch. (This trick doesn’t concern regularity.)
Suppose, further, that you can verify σ is injective, surjective and continuous. Then to show σ is a surface
patch, all that remains is to show σ −1 is continuous. As we’ve mentioned, this can be a pain. But the following
theorem can offer some pain relief.
This theorem however requires you to know what a closed set is. For a discussion of closed sets, see Appendix
A.
Theorem 4.13. Suppose f : A −→ B is a continuous bijection, where A is a closed and bounded subset of Rm ,
and B ⊆ Rn , for some m and n. Then f is a homeomorphism.

Remark 4.14. Although this theorem applies in any Rm and Rn , for surface patches we’re interested in m = 2 and
n = 3. Then A is a closed and bounded subset of R2 , and B is any subset of R3 .
This is not what we usually want when we have a surface chart! The domain of a chart σ should be an open set
U , not a closed set. But if we can extend the map σ to a slightly larger closed set, for instance by extending to the
boundary of U , then the theorem does apply — and the slightly extended map is a homeomorphism. A restriction
of a homeomorphism is still a homeomorphism, so we can use the theorem to show σ is a homeomorphism.

23
Remark 4.15. Sets which are closed and bounded are called compact. If you’ve studied analysis or topology, you
may have seen this concept already. We’ll discuss compact sets further later in the course.
Remark 4.16. This theorem is actually a very special case of a theorem in topology, that a continuous bijection
from a compact space to a Hausdorff space is a homeomorphism. Take a topology course and you can find a proof
of this very general theorem!
Let’s see an example using this theorem to save us the pain of proving σ −1 is continuous.

Example 4.17. A slightly different cylinder. Take the unit cylinder S defined by x2 + y 2 = 1 and −1 < z < 1.
We again consider the function σ(u, v) = (cos u, sin u, v), and define two charts covering S as restrictions of σ. But
let’s do it slightly differently. Let
n π o  π
U1 = (u, v) : − π < u < , −1 < v < 1 = −π, × (−1, 1)
 2   2

3π 3π
U2 = (u, v) : 0 < u < , −1 < v < 1 = 0, × (−1, 1)
2 2

and let σ1 = σ|U1 and σ2 = σ|U2 .

Again, σ1 , σ2 are both injective, continuous, and they cover the entire cylinder. However they cover the 360
degrees around the cylinder slightly differently: both patches cover a 270 degree region. (In example 2.15, each σi
covered the entire 360 degrees around the cylinder, except for a single line.)

The good thing now is that we can extend σ1 , σ2 over the boundaries of U1 , U2 in a continuous injective fashion.
To this end define  
h πi 3π
U1 = −π, × [−1, 1], U2 = 0, × [−1, 1]
2 2
and σ1 = σ|U1 , σ2 = σ|U2 . The sets U1 and U2 are now closed and bounded.
The functions σ1 and σ2 are now continuous (as restrictions of the continuous function σ) and injective (since
they only cover 270◦ of the cylinder — if we did the same thing for the charts of example 2.15, they would
not be injective) — hence bijections onto their images. So theorem 4.13 applies and we conclude σ1 , σ2 are
homeomorphisms.
In particular, σ1 −1 and σ2 −1 are continuous, hence so are their restrictions σ1−1 and σ2−1 . Hence we can conclude
σ1 , σ2 are homeomorphisms.

Exercise 20. The argument in the example above can be shortened by applying the following fact: a restriction
of a homeomorphism is a homeomorphism. Prove this fact.
In other words, suppose f : A −→ B is a homeomorphism. Let C be a subset of A, and let D = f (C). Show
that the restriction f |C : C −→ D is a homeomorphism.
(Hint: Use the fact that a restriction of a continuous function is continuous.)

4.7 Linearly independent partial derivatives

At some point, in checking through all the requirements for a surface, you might feel like some of what you’re doing
is redundant. And in some sense, some of it is.
For instance, suppose you have a map σ : U −→ Y defined on an open set U in R2 , with range Y ⊂ R3 , and
you want to show σ is a regular surface patch.
Suppose you can verify that, at a point (u, v) ∈ U , σu (u, v) and σv (u, v) are linearly independent. Then it seems
that σ can’t be too badly behaved. Directional derivatives of σ at (u, v) can always be found, and they span a

24
2-dimensional plane. So what could possibly go wrong with σ — at least, near the point (u, v)? If so, why would
we need to check everything else required of σ?
The next theorem says that, indeed, in these circumstances σ cannot behave too badly — at least, near that
particular point (u, v).

Theorem 4.18. Let U ⊆ R2 be open and Y ⊂ R3 . Suppose σ : U → Y is a smooth function such that σu (u, v)
and σv (u, v) are linearly independent at (u, v) ∈ U . Then σ is a regular surface patch near (u, v).
More precisely, there is an open neighbourhood V of (u, v), contained in U , such that the restriction of σ to V
is a regular surface patch.

We will not prove this theorem; perhaps the most natural proof uses a theorem called the constant rank theorem,
which is a generalisation of the inverse function theorem.
Remark 4.19. This theorem has a hidden subtlety. When the theorem applies to σ at a point (u, v) ∈ U , you can
conclude σ is a regular surface patch near (u, v). And σ might apply to σ at every point (u, v) ∈ U . You can then
conclude that σ is a regular surface patch near every point (u, v).
From there, you might think that σ must itself be a regular surface patch. Not so fast! Just because σ is a
regular surface patch near every point does not imply it’s a regular surface patch! Taken over all of U , σ might loop
back around and intersect itself! In this case σ is not even injective, so there’s no way it can be a surface patch.

But even worse, it can happen that σ is injective, and verify the hypotheses of the theorem at every point of U ,
yet not be a surface patch — as in the example below, where there is “almost” a self-intersection, and the image of
σ has a “T-intersection”.

So, you need to be a bit careful applying this result.

Remark 4.20. A surface which is allowed to have self-intersections like those shown above is called an immersed
surface. Immersed surfaces are important in geometry and topology, but we do not focus on them in this course.

4.8 Proving that level sets are smooth surfaces

The proof of theorem 4.7 relies on the inverse function theorem from multivariable calculus. We’ll state the inverse
function theorem, in the version we need. We’ve not going to prove it — you can find a proof in good multivariable
calculus texts.4 Not everyone learns the inverse function theorem properly when studying multivariable calculus,
but it’s a good theorem to know.

Theorem 4.21 (Inverse function Theorem). Let n ≥ 1. Let U ⊆ Rn be an open set, and let x0 ∈ U . Let
f : U −→ Rn be a smooth map and let f (x0 ) = y0 . Suppose that the matrix derivative DF (x0 ) at x0 is an
invertible matrix. Then there exists a local inverse g of f near y0 .
That is, there exist open sets x0 ∈ V ⊆ U and y0 ∈ W ⊆ Rn such that f restricts to a bijection f : V −→ W ,
with smooth inverse f −1 : W −→ V .

(The text has an equivalent but slightly more awkward statement.)

Now let’s see how to use the inverse function theorem, to prove that level sets are surfaces.

4 You can also watch some videos discussing it on the MLC moodle site.

25
Proof. [Proof of theorem 4.7] We have a set S ⊂ R3 , a point p ∈ S, an open set p ∈ W ⊆ R3 , and a smooth function
f : W −→ R such that ∇f (p) 6= 0 and S ∩ W is given by f = 0.
We will find a regular coordinate chart σ : U −→ R3 which forms a regular coordinate chart for S near p. Since
such a coordinate chart can be found for all points p on S, this will show that S is a smooth surface.
We are given that ∇f (p) 6= 0, so at least one of fx (p), fy (p), fz (p) 6= 0 at p. Without loss of generality suppose
fz (p) 6= 0; the proofs in the cases fx 6= 0 or fy 6= 0 are similar.
Now fz (p) 6= 0 means that, as you go from p in the z direction, f (p) does not stay constant: it changes with
some rate of change. But S is the level set f = 0, hence as you go in the z-direction from p, you do not stay on
S. Thus, S is transverse to the vertical line in the z-direction. The idea then, is that S looks like the graph of a
function in x and y.
Now there’s a bit of a trick. Consider the function F : W −→ R3 defined by
 
F (x, y, z) = x, y, f (x, y, z) .

What does this function do? For a point (x, y, z), it maintains the x, y coordinates, but adjusts the height and sets
the z-coordinate to f (x, y, z). Since S is defined by f (x, y, z), this function F will send S to lie on the plane z = 0.
The idea is that F flattens out S to lie on the plane, and then its inverse — or rather, a local inverse (we can’t be
sure there even is an inverse) — will essentially be the coordinate chart we want.

So we want to show F has a local inverse. And to do that, we need the inverse function theorem. For this, we
need to check the derivative DF .  
1 0 0
DF =  0 1 0
fx fy fz
So the Jacobian determinant det DF = fz 6= 0 at p.
Hence the inverse function theorem applies to F near p, and F has a local inverse G. Since F leaves x and y
coordinates as they are, G must do the same. So G must be of the form
 
G(x, y, z) = x, y, g(x, y, z) ,

where g is a real-valued smooth function.

Recall that S is described near p by f = 0; so F (S) is described by z = 0. And hence G sends the plane z = 0
back to the surface S. But this is almost exactly what a coordinate chart should do. The only difference is that
coordinate charts should be defined in the plane.
So define σ on the xy-plane by
σ(x, y) = (x, y, g(x, y, 0)).
Now σ is clearly injective. It has an inverse given by projection down to the xy-plane, (x, y, z) 7→ (x, y). So σ and
its inverse are continuous bijections, hence σ is a coordinate chart. So f is a surface. Indeed, σ shows that S is
locally described as a graph of the smooth function in two variables g(x, y, 0).
We can check σx = (1, 0, gx ), σy = (0, 1, gy ) so

σx × σy = (−gx , −gy , 1) 6= (0, 0, 0).

So σ is a regular surface patch near p.

Since this can be done near every point, S is a smooth surface.

26
Exercise 21. In the above proof, there are a few details missing. See if you can fill them in, including the following.
(i) We assumed without loss of generality that fz 6= 0. How does the proof go if fx 6= 0 or fy 6= 0?
(ii) When applying the inverse function theorem, we just said there was a local inverse G of F “near” p. To be
thorough, we ought to fill in the details of the domains and ranges of F and G.

(iii) We said that G sends the plane z = 0 back to the surface S, but it’s not the whole plane, because G is only
a local inverse and hence only defined in a neighbourhood of F (p). Fix this detail.
(iv) Similarly, σ is not defined on the whole xy-plane. It’s defined on an open subset of R2 though. Can you define
an open set in which σ can be defined to provide a regular coordinate chart near p?

The inverse function theorem is a crucial tool in proving many of the more technical results in this subject.

5 Features of smooth surfaces

We’ve now seen a definition of surface, and smooth surface. A surface has an atlas of coordinate charts, which are
homeomorphisms. A smooth surface has a little bit more: the coordinate charts must be regular.
This regularity produces many nice features of smooth surfaces: we can speak of tangent vectors, tangent planes,
and normal vectors. We can also speak of orientations. In this section we discuss these concepts.

5.1 Transition maps of smooth surfaces

But first, let’s return to transition maps. Recall from section 2.6 that these allow us to translate from one chart σ
to another chart ρ.
Specifically, a transition map is given by Φ = ρ−1 ◦ σ, where this composition is defined — on the overlap of the
images of σ and ρ.
On a surface in general, transition maps are continuous, but need not be smooth. (Indeed, the charts need not
be smooth.)
However on a smooth surface, we might ask — must the transition maps be smooth?
As it turns out, the answer is yes — and more, as we see in the following theorem. There is a lot of no-
tation, following the standard setup for a transition map. We won’t cover the proof here; one is given in the
text.
Theorem 5.1. Let σ : U −→ Y and ρ : U 0 −→ Y 0 be coordinate charts for a surface S, where U, U 0 are open
sets in R2 and Y, Y 0 are relatively open subsets of S. Suppose Y ∩ Y 0 is nonempty, and let V = σ −1 (Y ∩ Y 0 ),
V 0 = ρ−1 (Y ∩ Y 0 ), so that there is a transition map Φ : V → V 0 given by Φ = ρ−1 ◦ σ.
(i) If σ is regular on V , and ρ is regular on V 0 , then Φ is a diffeomorphism.

(ii) If ρ is regular on V 0 , and Φ is a diffeomorphism, then σ is regular on V .

This theorem may seem quite technical and obscure. But for a more advanced study of differential geometry
these ideas are crucial.
In this course, a surface S always lies in R3 , and an atlas is a collection of coordinate charts σi : Ui → R3 ,
where each Ui ⊆ R2 is open. But in fact, there is enough structure in the open sets Ui , and the transition maps,
to reconstruct the whole surface. Each point of S lies in the image of at least one chart σi — and if it lies in the
image of several coordinate charts, then the corresponding points in different open sets Ui are related by transition
maps.
Surfaces, in this sense, can be built purely out of open subsets Ui of R2 , glued together by transition maps. You
take open patches of the plane and glue them together. The open sets are like patches of a patchwork quilt; the
transition maps are the sewing instructions.
When one studies surfaces — and more generally manifolds — in more advanced courses, they are actually
defined along these lines. We haven’t described key details, such as how to describe the geometry of the surface
from the open sets Ui and transition maps alone. The first fundamental form or metric, which we describe later in
the course, allows this to be done.

27
5.2 Tangent vectors and tangent spaces
Just as for curves, the notion of a tangent vector to a surface is a very useful one.
Given a curve C and a point p on it, there are many tangent vectors to C at p, but they all point in the same
direction — they are all scalar multiples of each other. In other words, the set of all vectors tangent to C at p forms
a 1-dimensional space.
On the other hand, if you have a smooth surface S and a point p on it, the tangent vectors to S at p point in
many different directions. The set of all vectors tangent to S at p forms a 2-dimensional space, or plane.

We formally define tangent vectors to a surface using tangent vectors to curves.

Definition 5.2. Let S be a smooth surface and p ∈ S. A tangent vector to S at p is a tangent vector at p to a
curve on S passing through p. The tangent space to S at p, denoted Tp S, is the collection of all tangent vectors to
S at p.

Remark 5.3. If you’re a stickler for details, you might note that there are curves on S passing through p which don’t
have tangent vectors there. That’s fine — the definition just doesn’t include them. Despite there being non-smooth
curves through p, there are also smooth and regular ones, and the tangent vectors to them are the ones that count
as tangent vectors to S.
In the following example, we find the tangent space to a surface at a point. The calculations here are important
and standard — these are calculations you need to know how to do!

Example 5.4. Tangent space to the unit sphere. Let S be the unit sphere, with equation x2 + y 2 + z 2 = 1. Let’s
find the tangent space to S at p = (1, 0, 0). The diagram below suggests that the y- and z-directions should be
tangent; let’s verify this.
Let σ : U −→ S be a standard coordinate chart, where U = (−π/2, π/2) × (−π, π), defined by

σ(u, v) = (cos u cos v, cos u sin v, sin u) .

Since σ(0, 0) = (1, 0, 0) = p, the origin in U corresponds to p ∈ S.

To find tangent vectors to S at p, we consider curves γ(t) on S passing through p, in fact with γ(0) = p. These
correspond to curves δ(t) in U , in the uv-plane, related by γ(t) = σ(δ(t)), where δ(0) = (0, 0).

For instance, consider the curve δ(t) = (t, 0) in U . This curve has δ(0) = (0, 0), and proceeds in the positive
u-direction. We obtain a curve γ = σ ◦ δ in S such that γ(0) = p, given by

γ(t) = σ(δ(t)) = σ(t, 0) = (cos t, 0, sin t) .

Its derivative is γ̇(t) = (− sin t, 0, cos t), so the tangent vector to γ at p is γ̇(0) = (0, 0, 1). This is equal to
the partial derivative σu of σ at the origin. Indeed, we can verify σu = (− sin u cos v, − sin u sin v, cos u) and
σu (0, 0) = (0, 0, 1).

28
 Next, consider the δ(t) = (0, t), which also satisfies δ(0) = (0, 0), and then γ = σ ◦ δ satisfies γ(0) = p, given by

γ(t) = σ(δ(t)) = σ(0, t) = (cos t, sin t, 0) .

Its derivative is α̇(t) = (− sin t, cos t, 0) so α̇(0) = (0, 1, 0) is a tangent vector to S at p. This is equal to the partial
derivative σv at the origin: we can verify that σv = (− cos u sin v, cos u cos v, 0), so σv (0, 0) = (0, 1, 0).
More generally, consider a curve δ(t) in U passing through the origin, δ(0) = (0, 0), with derivative δ̇(0) = (a, b).
Let δ(t) = (u(t), v(t)), so δ̇(t) = (u̇(t), v̇(t)). Then (u(0), v(0)) = (0, 0) and (u̇(0), v̇(0)) = (a, b). The corresponding
curve on S is γ = σ ◦ δ, and the multivariable chain rule yields
d d
γ̇(t) = σ (δ(t)) = σ (u(t), v(t))
dt dt
∂σ du ∂σ dv
= +
∂u dt ∂v dt
= σu (u(t), v(t)) u̇(t) + σv (u(t), v(t)) v̇(t).

At t = 0, γ passes through p, with tangent vector

γ̇(0) = σu (0, 0) u̇(0) + σv (0, 0) v̇(0) = aσu (0, 0) + bσv (0, 0).

We calculate σu = (− sin u cos v, − sin u cos v cos u) and σv = (− cos u sin v, cos u cos v, 0), so σu (0, 0) = (0, 0, 1),
σv (0, 0) = (0, 1, 0); hence
γ̇(0) = a(0, 0, 1) + b(0, 1, 0) = (0, b, a).
Thus, the tangent vectors to S at p are indeed those in the y and z directions; Tp S is spanned by (0, 0, 1) and
(0, 1, 0).

In the above example, we found that a curve δ(t) in R2 with tangent vector (a, b) at (0, 0), translates to a curve
on S through p with tangent vector aσu (0, 0) + bσv (0, 0). In fact, we just computed the directional derivative of σ
in the direction (a, b) at the origin; recall the calculations of section 3.3.
In other words, we found that the tangent vectors to S at p are linear combinations of the two partial derivative
vectors σu and σv there. This is true for any surface.

Proposition 5.5. Let S be a smooth surface and σ : U −→ S a regular surface patch with σ(u0 , v0 ) = p. Then
the tangent vectors to S at p are precisely the vectors of the form

aσu (u0 , v0 ) + bσv (u0 , v0 ).

where a, b ∈ R. More precisely, a curve δ : I −→ U with δ(t0 ) = (u0 , v0 ) and δ̇(t0 ) = (a, b) maps under σ yields a
curve γ = σ ◦ δ : I −→ S such that γ(t0 ) = p and γ̇(t0 ) = aσu (u0 , v0 ) + vσu (u0 , v0 ).

This proposition says that the tangent vectors to S at p are precisely the linear combinations of σu (u0 , v0 ) and
σv (u0 , v0 ). In other words, Tp S is the span of σu (u0 , v0 ) and σv (u0 , v0 ).
Moreover, σu and σv are linearly independent: this is the defining property of a regular surface patch. So σu , σv
are linearly independent vectors which span Tp S — in other words, a basis. Thus proposition 5.5 immediately
implies the following.

Proposition 5.6. Let S be a smooth surface and σ : U −→ S a regular chart with σ(u0 , v0 ) = p. Then Tp S is a
2-dimensional vector subspace of R3 , with basis σu (u0 , v0 ), σv (u0 , v0 ).

We now prove proposition 5.5. The proof follows the same chain rule calculation as in the example.
Proof. Let δ(t) = (u(t), v(t)), so γ(t) = σ(δ(t)). We then have

d
γ̇(t) = σ(δ(t)) = σu (δ(t))u̇(t) + σv (δ(t))v̇(t)
dt

29
When t = t0 , we have γ(t0 ) = p, δ(t0 ) = (u0 , v0 ) and δ̇(t0 ) = (u̇(t0 ), v̇(t0 )) = (a, b), hence

γ̇(t0 ) = aσu (u0 , v0 ) + bσv (u0 , v0 )

as claimed.
Now we prove the first statement. Let v be a tangent vector to S at p; we show v has the desired form. By
definition v = γ̇(t0 ) for some curve γ(t) on S with γ(t0 ) = p. Such a γ can be expressed as γ(t) = σ(δ(t)) for some
curve δ : I −→ U where δ(t0 ) = (u0 , v0 ). Letting δ̇(t0 ) = (a, b) the above calculation shows that v = aσu + bσv .
Conversely, we show any aσu (u0 , v0 ) + bσv (u0 , v0 ) is a tangent vector. Let δ(t) be a curve in U with δ(t0 ) =
(u0 , v0 ) and δ̇(t0 ) = (a, b). Letting γ = σ ◦ δ be the corresponding curve on S, γ(t0 ) = p and the above calculation
shows γ̇(t0 ) = aσu + bσv , which is thus a tangent vector.
As you move around in the u, v plane, σ maps you to different points on the surface S. The partial derivative σu
gives you the tangent vector to S in the direction of increasing u. The partial derivative σv gives you the tangent
vector in the direction of increasing v. And the various linear combinations of these partial derivatives — which
are the directional derivatives of σ in the various possible directions — give you the full set of tangent vectors to S.
Remark 5.7. The same principle for a 2-dimensional surface S in R3 also holds for a k-dimensional manifold M in
Rn . A coordinate chart σ maps an open set U ⊂ Rk into Rn , and the chart is regular if the partial derivatives of σ
are linearly independent. The tangent space to M is k-dimensional, with basis given by the partial derivatives of σ.

Exercise 22. Let S be a smooth surface and p ∈ S.

(i) If v1 and v2 are vectors in Tp S, show that v1 + v2 is a vector in Tp S.
(Hint: if v1 , v2 ∈ Tp S then let v1 = a1 σu + b1 σv and v2 = a2 σu + b2 σv .)
(ii) If v is a vector in Tp S and c is a real number, show that cv is also a vector in Tp S.

5.3 Tangent plane

Just as there is a tangent space Tp S to a smooth surface S at a point p, there is also a tangent plane. They are not
the same thing, and they are sometimes confused.
We defined the tangent space Tp S in definition 5.2 as the set of all tangent vectors to S at p. We saw in
proposition 5.6 that Tp S is a 2-dimensional vector subspace of R3 with basis σu , σv . As a subspace, it contains the
zero vector; it need not contain (the position vector of) p.
On the other hand, the tangent plane is defined as follows.

Definition 5.8. The tangent plane to a smooth surface S at p is the plane parallel to Tp S, which passes through
p.

Since Tp S is a 2-dimensional vector subspace of R3 , there is a unique plane parallel to Tp S through p — so the
tangent plane is well-defined.
Indeed, the tangent space and the tangent plane are translates of each other: Tp S passes through the origin,
and the tangent plane passes through p.

30
5.4 Normal vectors
As with curves, we can consider normal vectors to surfaces.

Definition 5.9. A normal vector to a smooth surface S at a point p is a nonzero vector normal to Tp S.

Since Tp S is a 2-dimensional subspace of R3 , the set of normal vectors forms a 1-dimensional subspace of R3 —
a line normal to Tp S. In particular, if you have one normal N to S at p, then any nonzero scalar multiple of N will
also be a normal vector.
We often consider unit normal vectors. There are fewer of these. In fact, at any point p on a smooth surface S,
there are precisely two unit normal vectors — one pointing to each side of S.
A normal vector can be calculated from a regular chart σ : U −→ R3 . Suppose that σ(u0 , v0 ) = p. As we’ve
seen, the partial derivatives σu (u0 , v0 ) and σv (u0 , v0 ) form a basis for Tp S, and a normal vector to Tp S must be
perpendicular to both. Such a vector can be found from the cross product: so

σu (u0 , v0 ) × σv (u0 , v0 ).

is a normal vector to S at p. As σ is regular, the two vectors σu , σv are linearly independent, and the cross product
is not zero.
If we want a unit normal, we divide this vector by its magnitude to obtain the following proposition.

Proposition 5.10. Let S be a smooth surface, σ : U −→ S a regular chart, and p = σ(u0 , v0 ). A unit normal
vector to S at p is given by
σu (u0 , v0 ) × σv (u0 , v0 )
N= .
||σu (u0 , v0 ) × σv (u0 , v0 )||

Once you have a normal vector N , you can write down an equation for the tangent plane to S at p. This plane
consists of points x such that x − p is perpendicular to N , i.e. N · (x − p) = 0. Hence we have the following
proposition.

Proposition 5.11. Let S be a smooth surface, σ : U −→ S a regular chart, and p = σ(u0 , v0 ). The tangent plane
to S at p is given by the equation
N · (x − p) = 0,
where N is a unit normal to S at p, and x = (x, y, z).

2 2 2
Example 5.12. Let’s find  a unit normal, and tangent plane, to the unit sphere S given by x + y + z = 1, at
the point p = √12 , √12 , 0 .
Take the usual chart σ(u, v) = (cos u cos v, cos u sin v, sin u), defined on U = (−π/2, π/2) × (−π, π). Then p =
σ(0, π4 ). We can compute σu (u, v) = (− sin u cos v, − sin u sin v, cos u) and σv (u, v) = (− cos u sin v, cos u cos v, 0),
so at the point (u, v) = (0, π4 ) corresponding to p we have
 π  π  1 1

σu 0, = (0, 0, 1) , σv 0, = −√ , √ , 0 .
4 4 2 2
A normal vector is given by  
 π  π 1 1
σu 0, × σv 0, −√ , −√ , 0 ,
=
4 4 2 2
 
This turns out to be a unit vector, so we have our unit normal N = − √12 , − √12 , 0 . The tangent plane at p is
then given by N · (x − p) = 0, i.e.
   
1 1 1 1 1 1
− √ , − √ , 0 · x − √ , y − √ , a = 0, or simply √ x + √ y = 1.
2 2 2 2 2 2

31
Exercise 23. Let S be a smooth surface and σ a regular chart. Let p = σ(u0 , v0 ). Show that the two unit normal
vectors to S at p = σ(u0 , v0 ) are precisely

σu (u0 , v0 ) × σv (u0 , v0 ) σv (u0 , v0 ) × σu (u0 , v0 )

and .
||σu (u0 , v0 ) × σv (u0 , v0 )|| ||σv (u0 , v0 ) × σu (u0 , v0 )||

Exercise 24. For each point p on each smooth surface S below, find a regular coordinate chart σ for S around p,
a unit normal to S at p, and an equation for the tangent plane of S at p.
√
3
(i) The point p = ( 12 , 2 , 5) on the cylinder x2 + y 2 = 1, .
(ii) The point p = (1, 2, 5) on the paraboloid z = x2 + y 2 .
(iii) The point p = (−3, 4, 5) on the cone z 2 = x2 + y 2 .
(iv) The point p = (1, 1, 1) on the hyperboloid of one sheet x2 + y 2 − z 2 = 1.

Exercise 25. Let S be the graph of the smooth function f : R2 −→ R. Show that a unit normal to S at the point
(x, y, f (x, y)) is given by
1
N=q (−fx , −fy , 1)
fx2 + fy2 + 1

5.5 Normals of level sets

When a smooth surface S is a level set f −1 (c) of a smooth function f : R3 −→ R, there is an easier way to find a
normal vector.
Recall section 4.5: if f : R3 −→ R is smooth and ∇f 6= 0 along the level set f −1 (c), then f −1 (c) is a smooth
surface.
The idea is to consider ∇f along S. As you move along S, f is constant, always taking the value c. So the
directional derivative of f in any direction along S is zero.
On the other hand, ∇f gives the direction of fastest increase of f . So ∇f will not be tangent to S. In fact, it
turns out ∇f is actually normal to S.
More generally, f need not be defined on all of R3 . It could just be defined on an open set U . Then we have
the following proposition.

Proposition 5.13. Let U ⊂ R3 be open, f : U −→ R a smooth function, and let S = f −1 (c) for some constant c.
Suppose that for all p ∈ S, ∇f (p) 6= 0. Then
∇f
N=
||∇f ||
is a unit normal on S.

Exercise 26. Prove this proposition. (It’s on your assignment!)

Exercise 27. Using the above proposition, find a unit normal at a general point (x0 , y0 , z0 ) on each of the surfaces
defined by the following equations. Hence find an equation for the tangent plane to each surface at

(i) The cylinder x2 + y 2 = 1.

(ii) The paraboloid z = x2 + y 2 .
(iii) The cone z 2 = x2 + y 2 , away from the origin.

32
Exercise 28. Let S be the hyperboloid of one sheet given by x2 + y 2 − z 2 = 1. Find a regular chart σ for S near
the point p = (1, 1, 1). Find Tp S and given the equation for the tangent plane to S at p.

Exercise 29. Let S be the graph of the smooth function f : R2 −→ R. Show that a unit normal to S at the point
(x, y, f (x, y)) is given by
1
N=q (−fx , −fy , 1)
fx2 + fy2 + 1

The following exercise concerns paraboloids, which we saw in example 2.20. Paraboloids are used in headlights
and satellite dishes because they have an interesting focusing property we investigate here.

Exercise 30. Let Π be a plane with unit normal vector N . Show that the following linear transformations
R3 −→ R3 have the matrices claimed.

(i) Projection of vectors in the direction of N is given by the matrix N N T .

(ii) Projects of vectors onto Π is given by the matrix I − N N T .
(iii) Reflection of vectors in Π is given by I − 2N N T .

Exercise 31. Let S be the paraboloid defined by z = x2 + y 2 . Consider a line L in the z-direction, i.e. in the
direction (0, 0, 1). This line intersects S at a point p. Suppose the line is reflected off the tangent plane to S at p.
Show that the reflected line passes through the point (0, 0, 1/4).

5.6 Unit normals on different coordinate charts and orientations

As discussed above, a regular chart σ(u, v) for a smooth surface S yields a unit normal vector
σu × σv
N= .
||σu × σv ||

Since σ is smooth, the normal vectors given by the above expression vary smoothly over the surface.
At any point of S, there are precisely two unit normals, one on each side of S. They differ by a sign. If we
consider a different coordinate chart σ e(e
u, ve), which overlaps with σ(u, v), it yields another unit normal N e which
might agree with N , or might be −N .
The direction of the normal σu × σv is determined from σu , σv by the right hand rule. Similarly, the partial
derivatives σ
eue , σ
eve of σ
e determine the direction of the normal σ eue × σeve. Whether these normals are the same, or
opposite, depends on the relative orientation of σu , σv and σ eue , σ
eve.
So the question of when two overlapping coordinate charts give the same unit normal is closely related to the
idea of orientation.
Let’s calculate these normals explicitly. Let σ : U −→ S and σ e: Ue −→ S be two overlapping regular surface
patches for the smooth surface S. Let the coordinates on U be (u, v) and on U e as (e
u, ve). We have a transition map

Φ : V −→ Ṽ , e−1 ◦ σ;
Φ=σ Φ(u, v) = (e
u(u, v), ve(u, v))

where V ⊆ U and Ve ⊆ U e correspond to the overlap region, i.e. σ(V ) = σ

e(Ve ) = σ(U )∩e
σ (U
e ). We write u
e(u, v), ve(u, v)
for the coordinates of Φ; this indicates how u e, ve depend on u, v. The transition map Φ is smooth, by theorem 5.1
(in fact, it’s a diffeomorphism).

33
 We’ll compare σu ×σv and σ eue × σ
eve by writing the partial derivatives σu , σv of σ in terms of the partial derivatives
σ
eue , σ
eve of σ
e via the transition map Φ. Since Φ = σ e−1 ◦ σ we have σ = σ e ◦ Φ, so the matrix derivatives are related
by Dσ = De σ DΦ.
Now Dσ and De σ are 3 × 2 matrices whose columns are the partial derivatives of σ and σ e; and DΦ is a 2 × 2
matrix whose columns are the partial derivatives of u e(u, v) and ve(u, v). That is,
   
| | | |  
u u
eve , DΦ = u
ev
Dσ = σu σv  , De σ = σ eue σ .
e
veu vev
| | | |

Now since Dσ = De
σ DΦ, we have
     
| | | |   | |
σu σv  = σ u
eu u
ev
eue σ
eve = ueu σ
eue + veu σ
eve u
ev σ
eue + vev σ
eve .
veu vev
| | | | | |

The last equality holds since, when we expand out the matrix product, the first column of the resulting matrix is
given by multiplying the entries of σ
eue with u
eu , and the entries of σ
eve with veu ; and similarly for the second column.
In other words,
σu = u
eu σ
eue + veu σ
eve, σv = u
ev σ
eue + vev σ
eve,
which are really just expressions of the multivariable chain rule,

∂σ ∂σ ∂e
u ∂σ ∂e v ∂σ ∂σ ∂e
u ∂σ ∂e v
= + , = + .
∂u ∂e
u ∂u ∂e
v ∂u ∂v ∂e
u ∂v ∂e
v ∂v
Having expressed σu , σv in terms of σ eve we can express σu × σv in terms of σ
eue , σ eue × σ
eue . We compute

σu × σv = (αu σ eve) × (αv σ

eue + βu σ eue + βv σ
eve) .

Recall that for any vectors a, b, a × a = 0 and a × b = −b × a. Thus

 
u u
eve det u
ev
σu × σv = σeue × σ uu vev − u
eve (e eue × σ
ev veu ) = σ eue × σ
=σ eve det DΦ.
e
veu vev

So, σu × σv and σ
eue × σ
eve are scalar multiples of each other. This is as we should expect, since they have to both
be normal to the tangent plane to S!
Whether the unit vectors N and Ñ are the same, or opposite, depends on whether det DΦ is positive or negative.
If det DΦ > 0 then N = Ñ ; if det DΦ < 0 then N = −N e .5
We summarise our discussion so far in the following proposition.

5 Note that det DΦ can vary smoothly from point to point, but is never zero. We’ve discussed previously that a transition map of

between regular coordinate charts is a diffeomorphism, so Φ is a diffeomorphism. This means it has an inverse Φ−1 , which is smooth.
Since Φ ◦ Φ−1 is the identity, the product of their derivatives must be the identity matrix, DΦ DΦ−1 = I. Thus DΦ is an invertible
matrix. Matrices are invertible if and only if they have nonzero determinant.

34
Proposition 5.14. Two regular coordinate charts for a smooth surface S whose images overlap determine the
same unit normal if and only if the transition map Φ between them satisfies det DΦ > 0.

Let’s see what the sign of det DΦ means geometrically.

If you remember your linear algebra (also discussed in section 3.1.8 of chapter 1), the sign of the determinant
of a matrix tells you about whether the corresponding linear transformation preserves or reverses orientation. The
linear transformation TA : Rn −→ Rn of a matrix A is orientation preserving if det A > 0 and orientation reversing
if det A < 0. An orientation preserving linear transformation preserves the “handedness” of a basis. It sends a
left-handed basis to another left-handed basis, and a right-handed basis to a right-handed basis. On the other, erm,
hand, an orientation reversing linear transformation reverses the “handedness” of a basis.
Now given a smooth map F : Rn −→ Rn , its matrix derivative Dp F at a point p ∈ Rn is an n × n matrix. It is
the matrix of a linear transformation Rn −→ Rn describing how F acts on tangent vectors at p. It sends a vector
v at p, to a vector Dp F (v) at F (p), which gives the directional derivative of F at p in the direction v.
When det Dp F 6= 0, this determinant is positive or negative. When det Dp F > 0, it sends a basis of vectors at p
to a basis of vectors at F (p) of the same orientation. When det Dp F < 0, it sends a basis of vectors at p to a basis
of vectors at F (p) of the opposite orientation. In other words, the bases at p and F (p) have the same handedness
if Dp F is orientation preserving, and the opposite handedness if Dp F is orientation reversing.
In fact, we define a function F to be orientation preserving or reversing, based on this behaviour of its deriva-
tive.
Definition 5.15. Let V be an open subset of Rn . A function F : V −→ Rn is
(i) orientation preserving a a point p ∈ V if det Dp F > 0, and
(ii) orientation reversing if det Dp F < 0.

As Φ is a diffeomorphism, det DΦ is never zero, so at each point of V , det Φ is either positive or negative, and
Φ is orientation preserving or reversing accordingly.
We can now summarise the above discussion in the following proposition.

Proposition 5.16. Two regular coordinate charts for a smooth surface whose images overlap determine the same
unit normal if and only if the transition map Φ between them is orientation preserving.

Thus, when Φ is orientation-preserving, i.e. det DΦ > 0, then σ and σ e give equal unit normal vectors, N = N
e.
e give opposite unit normal vectors, N = −N .
When Φ is an orientation-reversing, i.e. det DΦ < 0, then σ and σ e

5.7 Orientability
So, some coordinate charts may give the same unit normals, and some may give opposite unit normals, depending
on whether the transition map is orientation preserving or reversing.
Is it always possible to cover a surface with coordinate charts which always give the same unit normals? As it
turns out, this is often possible, but not always.
If you can get a full atlas of such coordinate charts, with all the transition maps orientation-preserving, then all
their unit normals agree, and we obtain a unit normal vector at every point on S, which varies continuously.
The converse is also true. If you have a smooth surface, and you have a smooth unit normal vector field over
the entire surface S, then you can find an atlas of regular coordinate charts so that all the transition maps are
orientation-preserving. (One way to see this: the “bad” coordinate charts which give the “wrong” normals can
be replaced by other coordinate charts giving the “right” normals, for instance by pre-composing them with a
reflection.)
Thus we have shown the following proposition. You can find further details in the text.

Proposition 5.17. Let S be a smooth surface. The following are equivalent:

(i) S has a smooth unit normal vector field.
(ii) S has an atlas of regular charts for which all transition maps are orientation preserving.

Such a surface has a name.

35
Definition 5.18. A smooth surface which has a unit normal vector field — or equivalently, has an atlas with
orientation-preserving transition maps — is called orientable.
A smooth unit normal vector field on S is called an orientation. A surface with an orientation is called an
oriented surface.

Thus, an orientable surface has two orientations. There are two unit normal vector fields — pointing to either
side of the surface.

Example 5.19. Consider the unit sphere x2 + y 2 + z 2 = 1. It is an orientable surface, and an orientation (smooth
unit vector field) is given by taking the outward vector field N = (x, y, z) at each point (x, y, z).
The other orientation is the inward vector field N = (−x, −y, −z).

However, there are surfaces which are not orientable. If you try to find a unit normal vector field on such a
surface, you’ll go around and come back to where you started but find that your unit normal is now pointing the
wrong way. So for a non-orientable surface, “both sides” of the surface are the “same side”.
That’s definitely confusing, so let’s just look at an example.

Example 5.20. The Möbius strip is the smooth surface depicted below. It’s obtained by taking a ribbon, putting
a half-twist in it, and gluing the ends together.

36