Chapter 3: How curves curve

Daniel V. Mathews
April 1, 2022

How much does a curve curve?

The curvature of curves is one of the main topics of this course, and in this chapter we discuss it at length.
You probably have an intuitive notion of what sort of thing curvature is. As you proceed along a curve, you
turn and twist and curve around. At some times you may curve a lot; at other times you may not curve at all.
For instance, a straight line does not curve at all. But a circle does curve. Travelling along a small circle, you turn
faster than along a larger circle.
There are several different perspectives on curvature, each providing its own insights. As with many topics in
mathematics, deep understanding comes from unifying these different, but consistent, perspectives.
As we’ll see, one way to think about the curvature of a curve is as a measure of its deviation from being a straight
line, as measured by its second derivative. Another way is to think of curvature as the rate at which a curve turns.
A third way is that curvature is related to the radius of the best approximating circle to the curve at a point.
We’ll develop all these perspectives as we proceed.

What’s in this chapter.

Contents
1 The curvature of a curve 2
1.1 Curvature of a unit speed curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Curvature of a regular curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Two formulas for the curvature of a regular curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Calculating curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Curvature: technicalities and proofs 7

2.1 Is curvature well defined? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Invariance under isometries? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Proofs of curvature formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 A fact from 3-dimensional geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.2 Proving the formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Signed curvature 12
3.1 Turning angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Defining the signed curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Reparametrisations and orientations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Signed curvature of regular curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.5 The planar principal normal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.5.1 A moving frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5.2 The planar principal normal and signed curvature . . . . . . . . . . . . . . . . . . . . . . . . 18
3.6 Calculating signed curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.6.1 For a unit speed curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.6.2 For a regular curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1
4 Integrating curvature 23
4.1 Total curvatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 Total signed curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Total curvature vs. total signed curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.4 Reconstructing a curve from the signed curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

5 Curves in 3-dimensional space 28

5.1 The moving frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2 How does the moving frame move? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3 Frame equations in 2 dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.4 What does the torsion mean? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.5 Planes associated to a curve in space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.6 Reconstructing a 3D curve from the curvature and torsion . . . . . . . . . . . . . . . . . . . . . . . . 35
5.7 The moving frame for a regular curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.8 Torsion for a regular curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

1 The curvature of a curve

1.1 Curvature of a unit speed curve
Roughly,1 at each point on a curve γ(t), there is a curvature, which is a non-negative real number, often denoted
κ(t).2 So if γ is a function I −→ Rn , then its curvature κ is a function I −→ [0, ∞).
Here’s a precise first definition.

Definition 1.1. Let γ : I −→ Rn be a unit speed curve. The curvature function (or just curvature) of γ is the
function
κ : I −→ [0, ∞), κ(t) = ||γ̈(t)||.
We also say that κ(t) is the curvature of γ at γ(t).

Remark 1.2. As with much of our discussion in chapter 2, this definition works in Rn for any n, but for us in this
course we usually n is 2 or 3.
Remark 1.3. Beware! This definition only applies to unit speed curves! The curvature of a non-unit-speed curve
will be defined later. If you use this formula on a non-unit-speed curve, you’ll usually get the wrong answer for the
curvature.
Let’s think a little about the geometry of this formula. We’ve seen previously that for a unit speed curve, its
velocity γ̇(t) is always perpendicular to its acceleration γ̈(t):
γ̇(t) · γ̈(t) = 0.
As you travel along the curve, at unit speed in the direction γ̇(t), your acceleration is always perpendicular to the
direction you’re going. This acceleration is what makes you turn. The strength of this acceleration is precisely
||γ̈(t)||, the curvature.

Example 1.4. Let’s compute the curvature of a circle of radius R. To apply definition 1.1, we need a unit
speed parametrisation of such a circle. You did this in an exercise in chapter 2. Here’s such a parametrisation:
 
t t
γ : R → R2 , γ(t) = R cos , R sin .
R R

We can calculate γ̇(t) = (− sin(t/R), cos(t/R)) and verify that ||γ̇(t)|| = 1, so γ is in fact unit speed. The second
derivative is then  
1 t t
γ̈(t) = cos , sin ,
R R R

1 Provided the curve is nice enough! Here we work with unit speed curves; later, with regular curves. If the curve is not regular,

curvature won’t be defined.

2 The symbol κ is the Greek letter kappa.

2
 so the curvature is
1
κ(t) = ||γ̈(t)|| = .
R

We conclude from this example that the curvature of a circle is the reciprocal of its radius.3
This statement makes sense. A circle with very large radius “locally” looks very straight, and so its curvature
should be very small. And a very tiny circle curves around very tightly, so its curvature should be very large. These
intuitions agree with κ = 1/R.

Example 1.5. Let’s compute the curvature of a unit speed straight line,
 
3 2 1 2
γ : R −→ R , γ(t) = t − 1, t − 4, − t − 3
3 3 3

We have γ̇(t) = 23 , 13 , −2


3 and can verify ||γ̇(t)||2 = 1, so γ is indeed unit speed. The second derivative is then
γ̈(t) = 0, so the line has zero curvature, as expected.

1.2 Curvature of a regular curve

Having defined the curvature of unit speed curves, we now extend the definition of curvature to cover regular curves.
This extension is not particularly clever. In chapter 2, we found that regular curves are precisely the ones with
unit speed reparametrisations. So the idea is just to reparametrise and use the previous definition.

Definition 1.6. Let γ : I −→ Rn be a regular curve. The curvature of γ is the curvature of a unit speed
reparametrisation of γ.
More precisely, let δ = γ ◦ ϕ is a unit speed reparametrisation of γ, so that each point δ(s) of δ corresponds
to the point γ(t) of γ, where ϕ(s) = t. Then the curvature κ(t) of γ at γ(t) is the curvature of δ at δ(s). The
curvature function of γ is the function κ : I −→ [0, ∞) given by κ(t).

Remark 1.7. In practice, as discussed in the previous chapter, we often abuse notation and just write γ(t) for a
given regular curve, and γ(s) for its unit speed reparametrisation.
Remark 1.8. If this definition leaves you a little queasy, your may have some well-founded doubts; we’ll clear up
some subtle issues with this definition shortly. For now, let’s go with it.

Example 1.9. Let’s compute the curvature of the line

γ : R −→ R3 , γ(t) = (2t − 1, t − 4, −2t − 3) .

It has velocity γ̇(t) = (2, 1, −2) and speed ||γ̇(t)|| = 3, so γ(t) is not a unit speed curve. A unit speed parameter
is given s = 3t (or t = ϕ(s) = s/3), and a unit speed reparametrisation δ(s) slows γ down by a factor of 3:
   
1 2 1 −2
δ(s) = γ s = s − 1, s − 4, s−3 .
3 3 3 3

This is in fact exactly the unit speed line of example 1.5 above. Its curvature is zero. This is, again, what we
expect for the curvature of a straight line!

The particular coefficients in the previous example didn’t matter. This leads to the next exercise.

Exercise 1. (Useful, instructive.)

(i) Suppose γ : R −→ Rn is a regular curve defined by a linear function. Show γ has zero curvature.
(ii) More generally, show that any regular curve whose image is a line has zero curvature.

3 Strictly speaking, the above example only covers circles centred at the origin. If a circle is centred somewhere else, can you see

what this statement will still be true?

3
(iii) Conversely, show that a regular curve with zero curvature is a straight line.

Example 1.10. Consider again a circle of radius R, but now a non-unit-speed parametrisation:

γ : R −→ R2 , γ(t) = (R cos t, R sin t).

Note γ̇(t) = (−R sin t, R cos t), so ||γ̇(t)|| = R. Rescaling with s = Rt, we obtain the unit-speed reparametrisation
s  s s
δ(s) = γ = R cos , R sin .
R R R
This is the unit speed circle of example 1.4, where we calculated its curvature to be to be 1/R.
Again, the curvature of a circle is the reciprocal of its radius — even if it’s not traversed at unit speed.

Exercise 2. (Non-essential; reflection.) Do you think there are any other regular curves in R2 with constant
curvature, beside lines and circles? What about in R3 ? No need to prove anything at this stage; we’ll discuss the
answers to these questions as we proceed.

We saw in the previous chapter that it can be hard to find a unit speed reparametrisation of a regular curve
γ(t) explicitly.4 Because of this, it’s difficult in general to calculate curvature directly from definition 1.6 above. In
fact, in many cases it’s useless.
No fear! Shortly we’ll see some explicit formulas for curvature of a regular curve. But if you’re the sort of person
who likes working things out from first principls, you can try the following exercises in our present naive state.

Exercise 3. (Non-essential; low-tech; method will soon become obsolete) Find the curvature of the cusp curve
γ(t) = (t2 , t3 ) at γ(1) = (1, 1) directly. Use the unit speed reparametrisation you found in chapter 2, in an exercise.

Exercise 4. (Non-essential; low-tech; obsolescent) Find the curvature of the logarithmic spiral γ(t) = et (cos t, sin t)
at γ(0) = (1, 0) directly. Use the unit speed reparametrisation δ(s), from chapter 2:
     
s s s
δ(s) = √ + 1 cos ln √ + 1 , sin ln √ + 1 .
2 2 2

1.3 Two formulas for the curvature of a regular curve

We have now defined the curvature of a regular curve, in definition 1.6. However, applying this definition directly
involves finding a unit speed reparametrisation of a regular curve, which in general is difficult, and impossible to
write explicitly in an elementary way.
Luckily, there are results which provide us easier methods. Here I’ll present two of them.

Theorem 1.11. Let γ : I −→ Rn be a regular curve in Rn . Its curvature is given by

|γ̈(t) (γ̇(t) · γ̇(t)) − γ̇(t) (γ̇(t) · γ̈(t))|

κ(t) = 4 .
|γ̇(t)|

Theorem 1.12. Let γ : I −→ R3 be a regular curve in R3 . Its curvature is given by

|γ̇(t) × γ̈(t)|
κ(t) = 3 .
|γ̇(t)|

4 At least, if you want a nice formula for it.

4
Remark 1.13. Clearly these formulas are more complicated than the original formula κ = ||γ̈|| in definition 1.6! We
often leave it implicit that κ, γ̇ and γ̈ are functions of t, and so we can write these two formulas respectively as
|γ̈ (γ̇ · γ̇) − γ̇ (γ̇ · γ̈)| |γ̇ × γ̈|
κ= 4 . and κ = 3 .
|γ̇| |γ̇|
In any case, there are lots of dots. As usual, dots above γ̇ indicate derivatives with respect to the parameter t, and
dots in between them indicate the dot product.
Remark 1.14. Let’s briefly compare the two theorems / formulas. They’re both very useful formulas; 1.11 works in
general Rn and is maybe a little more complicated; 1.12 is specific to R3 and maybe a bit simpler. Indeed, theorem
1.12 involves a cross product, which is very specific to three dimensions!5
Remark 1.15. When γ is unit speed, both formulas simplify back to the original definition 1.1. Let’s see why.
4
First consider theorem 1.11. If γ is unit speed then the denominator |γ̇| = 1, and we have γ̇ · γ̇ = |γ̇|2 = 1.
Further, for unit speed curves we have γ̇ · γ̈ = 0. So the formula simplifies to κ(t) = |γ̈(t)|.
Second consider theorem 1.12. Since γ̇(t) and γ̈(t) are perpendicular, we have
|γ̇(t) × γ̈(t)| = |γ̇(t)| |γ̈(t)| sin 90◦ = |γ̈(t)| .
Then the formula simplifies to κ(t) = |γ̈(t)| again.
So the formulas all give the same answer for unit speed curves. This is reassuring!
Remark 1.16. Although the theorem is stated only for curves in R3 , you can use it on curves in R2 . Think of R2 as
(say) the xy-plane in R3 , and think of a point (x, y) in the plane as lying in the xy-plane at (x, y, 0). In this way,
you can regard a curve in R2 as lying in R3 , and apply the formula.

1.4 Calculating curvature

Let’s put the formula of theorem 1.12 into action, and calculate the curvature of some curves. Before we had this
formula, calculating curvature required explicitly finding a unit speed reparametrisation. But now the calculation
is as easy as substitution into a formula.
We’ll revisit some old examples and also compute some new ones. As mentioned earlier, although the formula
is stated for curves in R3 , but it also works for curves in R2 , since the points (x, y) of the plane can be embedded
into 3-dimensional space as (x, y, 0).

Example 1.17. A line. A non-unit-speed line, as in example 1.9:

 
1
γ : R −→ R2 , γ(t) = 2t − , t − 4, −2t − 3 .
3

We compute γ̇(t) = (2, 1, −2), so γ is regular, and γ̈(t) = 0. Hence γ̈(t) × γ̇(t) = 0, so

||γ̈(t) × γ̇(t)||
κ(t) = = 0.
||γ̇(t)||3

In fact, the same proof works for any linear γ.

Exercise 5.
(i) Let γ(t) = (at + b, ct + d, et + f ) for some constants a, b, c, d, e, f , with (a, c, e) ̸= (0, 0, 0). Show that γ has
curvature zero.
(ii) Let γ(t) be a regular curve whose image is a line (not necessarily a linear function of t). Show that its
curvature is zero.

Example 1.18. Let’s revisit the non-unit-speed circle of radius R > 0, as in example 1.10:

γ : R −→ R2 , γ(t) = (R cos t, R sin t) .

5 There is a similar type of operation in R7 , but that’s well beyond the scope of this course!

5
 In order to use the formula, we regard γ as a curve in R3 , lying in the xy-plane. This amounts to adding an
extra coordinate, z = 0:
γ : R −→ R3 , γ(t) = (R cos t, R sin t, 0) .
We can then compute γ̇ = R(− sin t, cos t, 0), so γ is regular. The second derivative is γ̈ = R(− cos t, − sin t, 0),
and their dot product is  
i j k
γ̈(t) × γ̇(t) = −R cos t −R sin t 0 = (0, 0, −R2 ).
−R sin t R cos t 0
Thus curvature is given by
||γ̈(t) × γ̇(t)|| R2 1
κ(t) = 3
= 3 = .
||γ̇(t)|| R R
Once more, we conclude that the curvature of a circle of radius R is 1/R.

Exercise 6. Revisit the logarithmic spiral from exercise 4 and previously, given by γ(t) = et (cos t, sin t). Calculate
its curvature. Compare your answer with exercise 4. Which method do you prefer?

Example 1.19. Let’s compute the curvature of the helix

γ(t) = (2 cos t, 2 sin t, 3t).

We first compute
γ̇(t) = (−2 sin t, 2 cos t, 3).
This is never zero (the z-coordinate is always 3), so γ is regular and the formula applies. We next compute

γ̈(t) = (−2 cos t, −2 sin t, 0)

and then
 
i j k
γ̈(t) × γ̇(t) = det −2 cos t −2 sin t 0 = (−6 sin t, 6 cos t, −4).
−2 sin t 2 cos t 3
√ √
We can compute the lengths ||γ̈ × γ̇|| = 52 and ||γ̇|| = 13, so that
√
||γ̈ × γ̇|| 52 2
κ(t) = = √
3 = .
||γ̇||3 13 13

Note that in general the curvature κ(t) will depend on t. For the helix we just found that κ(t) does not depend
on t; the helix has constant curvature. Can you see why this might be the case? A helix is quite symmetric.
Now it’s your turn to calculate the curvature!

Exercise 7. Calculate the curvature of the helix γ(t) = (cos t, sin t, 5t).

Exercise 8. Let a, b be nonzero real numbers. Consider the general helix γ(t) = (a cos t, a sin t, bt). Show that it
has constant curvature given by
|a|
κ= 2 .
a + b2
If b = 0 then a helix degenerates to a circle of radius |a|, which has radius 1/|a|. Indeed, when b = 0 in the above
formula, we obtain κ = 1/|a|.

6
2 Curvature: technicalities and proofs
2.1 Is curvature well defined?
As you were reading definition 1.6 of the definition of a regular curve above, you might have been confused by all
the reparametrisation stuff in the second paragraph. But that’s just an artifact of going through the motions of
reparametrisation.
There’s a larger problem with definition 1.6. It doesn’t actually make it clear that the curvature of a regular
curve γ(t) is properly defined!
What definition 1.6 says to do is to find a unit speed reparametrisation β(s) of γ(t), and then apply the first
definition 1.1 to it.
But here’s the problem: γ(t) has many different unit speed reparametrisations! If you take different unit speed
reparametrisations, will you come up with the same answer for the curvature of γ(t)?
In order for curvature to be well defined, the answer must be yes. And fortunately, the answer is yes. Let’s
think about why.
Suppose that β(s) and δ(s) are both unit speed reparametrisations of γ(t). Then β(s) and δ(s) are unit speed
reparametrisations of each other. So as you proceed along β(s) or δ(s), you go through exactly the same points, at
the same unit speed. This means that β(s) and δ(s) must be very similar. So similar, in fact, that they give the
same result for the curvature of γ.
But how similar? If β and δ both travel along the curve in the same direction, then wherever β goes, δ also
follows. Intuitively, both β, δ drive along the same road at the same speed in the same direction, and hence,
although they may be separated, their separation remains constant throughout their journey. One just trails the
other, without gaining on it or falling further behind. If β arrives at a point on the curve 10 seconds after δ, then
β will arrive at every point on the curve 10 seconds after δ. This means that β(s) = δ(s + c) for some constant c.
On the other hand, if β and δ travel in opposite directions, then they will pass each other at some point,
β(s0 ) = δ(s0 ); and wherever β is t seconds after that point, δ will have been there t seconds earlier. In other words,
β(s0 + t) = δ(s0 − t), which we can rearrange as β(s) = δ(−s + 2s0 ) (letting s = s0 + t). So β(s) = δ(−s + c), for
some constant c.
Let’s make a precise statement of what happens.

Proposition 2.1. Suppose β, δ : R −→ Rn are unit speed reparametrisations of a curve γ. Let β, δ have
curvature functions κβ , κδ : R −→ [0, ∞) respectively.
(i) If β and δ travel along γ in the same direction, then there exists a constant c such that β(s) = δ(s + c) for
all s ∈ R. Moreover, κβ (s) = κδ (s + c) for all s ∈ R

(ii) If β and δ travel along γ in opposite directions, then there exists a constant c such that β(s) = δ(−s + c)
for all s ∈ R. Moreover, κβ (s) = κδ (−s + c) for all s ∈ R.

In other words, β and δ are related by a very simple linear reparametrisation, and their curvature functions are
exactly equal at corresponding points: so we obtain the same result for the curvature of γ at each point.
In the next example we’ll see how this works in a specific case, and then you’ll prove this in an exercise.

Example 2.2. Consider the circle γ(t) = (R cos t, R sin t), and the two unit speed reparametrisations
 s s  u u
β(s) = R sin , R cos and δ(u) = R cos , R sin .
R R R R
You can verify that these are unit speed reparametrisations of γ, in opposite directions, and that
π  
 s π s  πR
β(s) = R cos − , R sin − =δ −s .
2 R 2 R 2

We can directly check that β and δ both have curvature 1/R, but more usefully for proving proposition 2.1, we

7
 can differentiate the equation β(s) = δ πR


2 − s twice with respect to s:
     
πR d πR πR
β̇(s) = δ̇ −s − s = −δ̇ −s
2 ds 2 2
     
πR d πR πR
β̈(s) = −δ̈ −s − s = δ̈ −s .
2 ds 2 2

Now κβ (s) = ||β̈(s)||, and κδ (u) = ||δ̈(u)||. So the above equation implies that
 
πR
κβ (s) = κδ −s .
2

Thus β and δ have equal curvature at corresponding points.

Having seen an example, let’s now give a proof in general. But you might find it less illuminating than the
intuitive argument at the start of this section.

Proof. Write s for the parameter in β, and u for the parameter in δ. As β and δ are unit speed reparametrisations
of each other, we have β = δ ◦ Φ for some diffeomorphism Φ.a In other words, β(s) = δ(Φ(s)) = δ(u), where
u = Φ(s). Each value of s corresponds to a unique value of u via the reparametrisation Φ.
We can now compare the velocities of β and δ:

dβ(s) dδ(u) dδ(u) du

= = .
ds ds du ds
dβ(s) dδ(u)
As β(s) and δ(u) are both unit speed, we have ds = du = 1. Hence, taking lengths of both sides of
du
the above equation yields ds = 1. Since u = Φ(s), this says that the derivative of Φ everywhere has absolute
value 1, i.e. |Φ′ (s)| = 1 for all s ∈ R.
Thus, Φ′ (s) = ±1 for each s. As Φ is a diffeomorphism, it is a smooth function, and hence Φ′ cannot jump
between values of 1 and −1. So either Φ′ (s) = 1 for all s, or Φ′ (s) = −1 for all s. We consider the two cases
separately.
In the first case, integrating the equation Φ′ (s) = 1 yields Φ(s) = s + c for some constant c. Then β(s) =
δ(Φ(s)) = δ(s + c) as claimed. Differentiating this equation, we observe that β̇(s) = δ̇(s + c), so β, δ have the
same velocity at corresponding points: in other words, they travel along the curve γ in the same direction.
In the second case, integrating Φ′ (s) = −1 yields Φ(s) = −s + c for some constant c, so β(s) = δ(−s + c) as
claimed. Differentiating this equation gives β̇(s) = −δ̇(−s + c), so β, δ have opposite velocities at corresponding
points, and travel in opposite directions.
It remains to prove that the curvature functions κβ , κδ are equal at corresponding points. We’ll leave this
for you in the next exercise. As a hint, we already differentiated the equations β = δ(±s + c) once!
a Since β, δ are reparametrisations of γ, we are really only given that β = γ ◦ ϕ and δ = γ ◦ ψ, for some diffeomorphisms ϕ, ψ. The

Φ in this argument is equal to ψ −1 ◦ ϕ, and we are really using the fact that taking inverses and compositions of diffeomorphisms
yield more diffeomorphisms.

Exercise 9. Let β, δ be unit speed curves.

(i) If β(s) = δ(s + c) for some constant c, show that the curvature of β at β(s) is equal to the curvature of δ at
δ(s + c).
(ii) If β(s) = δ(−s + c) for some constant c, show that the curvature of β at β(s) is equal to the curvature of δ
at δ(−s + c).
Conclude that the curvature of a regular curve is well-defined at each point.

Our discussion of curvature has been a little circuitous. First, in definition 1.1, we defined it for unit speed
curves. Then, in definition 1.6, we “defined” it for regular curves, by saying to take a unit speed reparametristaion
and using definition 1.1. Then, we were worried whether the definition gave the same answer for different unit speed
reparametrisations, but now these have been laid to rest.

8
 As a result, we now see that you can take any unit speed parametrisation — or, the same thing, any arclength
parameter — and the curvature you calculate at any point on the curve will be correct. If γ(t) is a regular curve,
we can simply write that its curvature is given by

d2 γ
κ=
ds2
2
d
where s is an arclength parameter. (Note we write ds 2 , rather than γ̈, becuase the original parameter may not be

a unit speed parameter.) By the above discussion, this is a well defined quantity.

2.2 Invariance under isometries?

2.3 Proofs of curvature formulas
In the next two sections, we’ll prove theorem 1.12.

2.3.1 A fact from 3-dimensional geometry

First of all, we need a random fact about 3-dimensional geometry.
This fact concerns the cross product. As an algebraic operation, the cross product is not particularly nice: it
is neither commutative nor associative! Rather, for any vectors v, w ∈ R3 , we have v × w = −w × v, so the cross
product is anticommutative. And in general u × (v × w) ̸= (u × v) × w, which is a little disturbing.
We’re interested in what happens when you take the cross product of three vectors u, v, w. The cross product
being non-associative, we’ll obtain different answers for u × (v × w) and (u × v) × w.
Consider u × (v × w). The term v × w is perpendicular to both v and w. Assuming v, w are linearly independent,
the span of v and w is a plane, and v × w is normal to this plane. Then, u × (v × w) is perpendicular to this normal
vector. So the result u × (v × w) must lie in the plane spanned by v and w. This means that it must be a linear
combination of v and w.
In other words, u × (v × w) should be a scalar multiple of v, plus a scalar multiple of w. But what are those
scalar multiples? The following lemma says what they are.

Lemma 2.3. For any vectors u, v, w ∈ R3 ,

u × (v × w) = (u · w)v − (u · v)w.

Exercise 10. Prove this lemma.

(i) One way to prove this is just to compute it. So let u = (u1 , u2 , u3 ), v = (v1 , v2 , v3 ) and w = (w! , w2 , w3 ), and
work out the left and right hand sides directly. Show that they are equal.
(ii) Can you find a direct geometric proof of this result? It will surely be nicer than direct computation, but
might be harder to find!

2.3.2 Proving the formula

Let’s now prove the formula for curvature in theorem 1.12.
The strategy is simply to use definition 1.6. Starting from the regular curve γ(t), we take a unit speed
reparametrisation of γ, with an arclength parameter s, and then the curvature is given by

d2 γ
κ= .
ds2
2
Work is required to figure out how to express ddsγ2 in terms of γ and its derivatives γ̇, γ̈ with respect to t. Eventually
we will express κ in terms of γ, γ̇, γ̈, and obtain the formula in theorem 1.12.
Throughout this section, dots refer to derivatives with respect to t. When we mean a derivative with respect to
d
s, we write it out as a ds .

9
 The proof is an extended exercise in multivariable differentiation. It’s a hotbed of chain, product and quotient
rule activity.
We’ll do it in two steps. First, we’ll express the first derivative dγ
ds in terms of t. Second, we’ll express the second
2
derivative ddsγ2 in terms of t. Then we’ll finish.
Here is the first step.

Lemma 2.4.
dγ γ̇
= (1)
ds ||γ̇||

Proof. This is a slightly-disguised versionRof something we’ve seen in the previous chapter: it’s just the chain
t
rule. Recall arclength is defined by s(t) = t0 ||γ̇(u)|| du, so

ds
= ||γ̇||.
dt
Hence, by the chain rule
dγ dγ dt 1 γ̇
= = γ̇ · = .
ds dt ds ||γ̇|| ||γ̇||

d2 γ
Now for the second step, we consider ds2 .Differentiating equation (1) with respect to s gives

d2 γ
   
d dγ d γ̇
= = .
ds2 ds ds ds ||γ̇||
γ̇
We now consider differentiating ||γ̇|| . Since γ̇ is a vector and ||γ̇|| is a scalar, but we use the (multivariable) quotient
rule — it was mentioned in the first chapter, if you missed it. In the end, we only want derivatives with respect to
γ̇
t, so it’s better to differentiate ||γ̇|| with respect to t rather than s; for this, we’ll need the chain rule again, and to
ds
remember that dt = ||γ̇||. We obtain
!
d d
d2 γ γ̈ ||γ̇|| − γ̇ dt ||γ̇|| γ̈ ||γ̇|| − γ̇ dt ||γ̇||
 
dt d γ̇ 1
2
= = 2
= 3
(2)
ds ds dt ||γ̇|| ||γ̇|| ||γ̇|| ||γ̇||

Unfortunately, this expression contains a tricky term with the derivative of the speed
d
||γ̇||.
dt
Unfortunately, although the derivative of γ̇ is γ̈, the derivative of ||γ̇|| is not ||γ̈|| ! For instance, on any unit speed
curve, ||γ̇|| is constant, so its derivative is zero — but ||γ̈|| is the curvature, very much not zero!
There’s a trick to calculate this derivative: instead of differentiating ||γ̇||, we’ll differentiate ||γ̇||2 . Length squared
is much better for differentiation than length. Why? Because length squared can be given by a dot product, and
dot products are no problem to differentiate: you just use the product rule. By contrast, lengths have unpleasant
square roots in them.
Applying this trick, we differentiate ||γ̇||2 = γ̇ · γ̇ and obtain
d d
||γ̇||2 = (γ̇ · γ̇) = γ̈ · γ̇ + γ̇ · γ̈ = 2γ̇ · γ̈. (3)
dt dt
Now once we know the derivative of ||γ̇||2 , it’s not hard to relate this to the derivative of ||γ̇||, because one is the
square of the other! We can apply the product rule to ||γ̇||2 = ||γ̇|| · ||γ̇|| and obtain
     
d d d d
||γ̇||2 = ||γ̇|| ||γ̇|| + ||γ̇|| ||γ̇|| = 2||γ̇|| ||γ̇|| . (4)
dt dt dt dt
Putting together equations (3) and (4), we now find an expression for the derivative of speed:
 
d d γ̇ · γ̈
γ̇ · γ̈ = ||γ̇|| ||γ̇|| , i.e. ||γ̇|| = . (5)
dt dt ||γ̇||

10
 d d2 γ
This is great, because we can now substitute this expression for dt ||γ̇|| into our formula for ds2 (equation (2)) and
obtain something nicer:

d2 γ γ̈ ||γ̇|| − γ̇ ||γ̇·γ̈
γ̇|| γ̈ ||γ̇||2 − γ̇(γ̇ · γ̈) γ̈(γ̇ · γ̇) − γ̇(γ̇ · γ̈)
= = = . (6)
ds2 ||γ̇||3 ||γ̇||4 ||γ̇||4
In the last equality we just replaced a squared length with a dot product.
Now this last expression is exactly the type of expression which occurred in our random fact about triple cross
products, lemma 2.3. That lemma expresses the numerator as a triple cross product:
γ̇ × (γ̈ × γ̇) = (γ̇ · γ̇)γ̈ − (γ̇ · γ̈)γ̇.
Hence we have proved the following lemma, completing our second step.

Lemma 2.5.
d2 γ γ̇ × (γ̈ × γ̇)
2
= .
ds ||γ̇||4

The length of this vector is the curvature, and so we can now prove our curvature formula.

Proof. [Proof of theorem 1.12] Using lemma 2.5, the curvature is

d2 γ ||γ̇ × (γ̈ × γ̇)||

κ= = .
ds2 ||γ̇||4

Now, the length of γ̇ × (γ̈ × γ̇) is given by ||γ̇|| ||γ̈ × γ̇|| sin θ, where θ is the angle between γ̇ and γ̈ × γ̇. But this
is a right angle, because the cross product of γ̇ with anything is perpendicular to γ̇! Hence

||γ̇|| ||γ̈ × γ̇|| ||γ̈ × γ̇||

κ= =
||γ̇||4 ||γ̇||3

as desired.

This was our longest proof to date. It involves several steps, and tricks from calculus and geometry. But
hopefully you can follow all the steps, even if the whole thing would be difficult to find yourself. Well done for
getting this far!
In the course of the proof, we found some equations which are interesting in their own right. Let’s pause to
make note of them.
Remark 2.6. In equation (5) we found an expression for the derivative of speed with respect to time on a regular
curve:
d γ̇ · γ̈
||γ̇|| = .
dt ||γ̇||
The expression ||γ̇·γ̈
γ̇|| is the length of the component of γ̈ in the direction of γ̇. As you proceed along the curve,
consider your acceleration: it has a component in the direction you are going (i.e. parallel to your velocity γ̇), and
it has a component perpendicular to your velocity. The component parallel to your velocity changes your speed,
and is reflected in this equation; the component perpendicular to your velocity changes your direction.
d2 γ
Remark 2.7. In lemma 2.5, we found an expression for ds2 ,

d2 γ γ̇ × (γ̈ × γ̇)
2
= .
ds ||γ̇||4
2
The second derivative ddsγ2 is the acceleration of γ with respect to the arclength coordinate s — the acceleration
when you proceed in a steady, unit-speed way. This acceleration is perpendicular to γ̇, being the cross product of
γ̇ with another vector; we should expect this, from the previous chapter.
The cross product γ̇ × (γ̈ × γ̇) might seem like it should be zero, as two of the vectors are the same. And indeed,
while for any vector v ∈ R3 it’s true that v × v = 0, the cross product is neither commutative nor associative, so
there’s no reason to expect an answer of zero here.

11
3 Signed curvature
When you’re driving a car along a road, the road has some curvature, so you turn. But you always turn left or
right. At some points you turn more left or less left, or more right or less right, depending on the curvature of the
road, but in some sense there are only two ways to turn: left or right. The car’s steering wheel only moves in a
1-dimensional way — from left to right.
On the other hand, when you move along in 3-dimensional space, you can turn in many more ways. You can’t
drive a rocket with a steering wheel alone! Any “steering wheel” on a rocket should actually move in a 2-dimensional
way. A joystick, as used in an aeroplane, moves in a 2-dimensional way. So for the time being, we’ll stick to curves
in R2 .
Thus, the curvature of plane curves, i.e. curves in R2 , is quite different from, and simpler than, the curvature
of curves in 3 (or more) dimensions. In this section, we take advantage of the special nature of curvature in
2 dimensions to refine it into something called signed curvature — so called because it has a sign, positive or
negative. Turning left or right corresponds to positive or negative signed curvature.
Given a regular curve in the plane γ : I −→ R2 , we’ve seen that its curvature is a function κ : I −→ [0, ∞).
Being originally defined as the length of a vector, κ ≥ 0 everywhere; curvature is never negative.
The signed curvature of γ will be another function κs : I −→ R, which can be positive or negative. We’ll see
that |κs | = κ, so at each point on the curve, κs = ±κ. The signed curvature will be equal to κ or −κ accordingly
as γ is turning left or right.
In order to define signed curvature, we first need a concept called turning angle, which tells you “which way”
the curve is going.

3.1 Turning angle

Let γ : I −→ R2 be a unit speed curve. Then for any s ∈ I, γ̇(s) is a unit vector in the plane, and hence is of the
form (cos φ, sin φ) for some angle φ.

y
(cos φ, sin φ)

φ
x

More generally, if γ : I −→ R2 is regular, then for any t ∈ I, γ̇(t) is a nonzero vector, and we may express it in
polar form as γ̇(t) = |γ̇(t)|(cos φ, sin φ) for some angle φ.
We get different angles φ at different points on the curve, so we can think of the angle φ as a function of the
parameter. This is what a turning angle is.

Definition 3.1. A turning angle for a regular curve γ : I −→ R2 is a continuous function φ : I −→ R such that,
for all t ∈ I,
γ̇(t) = |γ̇(t)| (cos φ(t), sin φ(t)) .

The turning angle φ(s) tells you the direction in which γ̇(s) points. As you move along the curve, you go through
the points γ(s) on the curve, with velocity γ̇(s), and the velocity is a unit vector pointing in the direction φ(s).
When we write (cos φ, sin φ) for a point on the unit circle, the angle φ is measured anticlockwise from the
positive real axis. So the angle φ increases as the velocity vector γ̇ turns anticlockwise, and the angle φ decreases
as the velocity vector γ̇ turns clockwise.
There is a subtlety with the turning angle — it is not uniquely defined. There is no one turning angle φ(s) which
we can say is the turning angle for φ(s). Given the curve γ(s), there are several possible choices for the turning
angle φ(s); in fact, infinitely many choices.
You might remember, from when you first learned about complex numbers, that the argument of a complex
number is not well defined; it’s ambiguous. The issue here is essentially the same.

12
 For instance, suppose at some point in time s = s0 the velocity vector is γ̇(s0 ) = (−1, 0). That is, at time
s = s0 , the curve is heading directly to the left (i.e. the negative x direction). What is the turning angle φ(s0 ) at
this point?
You could say the turning angle is π, and indeed, γ̇(s0 ) = (−1, 0) = (cos π, sin π). However, you could also say
the turning angle is 3π, since (−1, 0) = (cos 3π, sin 3π). If you turn through an angle of π or 3π anticlockwise from
the positive real axis, you end up facing in the same direction. Similarly, you could say the turning angle is −π;
the negative angle means that we turn clockwise from the positive x direction. Indeed, any turning angle π + 2kπ,
for any integer k, would be a valid turning angle.
So, the turning angle is ambiguous. But any two possible choices for the turning angle differ by an integer
multiple of 2π. We say that the turning angle φ(s) is defined up to multiples of 2π or modulo 2π. (If you recall
modular arithmetic, you should be able to see why we would use the word “modulo”.)
However, if you choose a turning angle φ(s0 ) at a particular point γ(s0 ) on the curve, and then proceed along the
curve from γ(s0 ), the tangent vector γ̇(s) varies in a continuous way.6 So the choice of φ(s0 ) can then be extended
to a continuous function φ(s). Once we have chosen the turning angle φ(s0 ) at one point γ(s0 ), then requiring φ
to be continuous, φ(s) is determined at all points on the curve.
The upshot of this is that although the turning angle of a curve at a point is ambiguous, it is not very ambiguous.
Once you fix the turning angle at a single point on the curve, it is fully determined and no longer ambiguous. If
your choice of turning angle and my choice of turning angle differ by 314π at one point on the curve, they will differ
by 314π at all points on the curve.
Let’s now state this fact precisely as a proposition.

Proposition 3.2. Let γ : I −→ R2 be a regular curve. Take s0 ∈ I and let φ0 ∈ R be a number such that

γ̇(s0 ) = |γ̇(s0 )| (cos φ0 , sin φ0 ) .

Then there exists a unique turning angle φ : I −→ R for γ such that φ(s0 ) = φ0 .

The discussion above gives an argument (hopefully a convincing one!) as to why this statement is true. But it’s
not quite a rigorous proof. The text has a rigorous proof (proposition 2.2.1).

3.2 Defining the signed curvature

Consider a unit speed plane curve γ(s) with turning angle φ(s).
If φ(s) does not change as you proceed along the curve — i.e. remains constant — then you must be going in
a straight line, because you are always proceeding in the same direction. Hence the curvature must be zero. But if
φ(s) increases as you proceed along the curve, then you must be turning. In fact, you must be turning left. The
curvature in this case should be nonzero. If φ(s) increases really fast as you proceed along the curve, then you are
turning sharply to the left and the curvature should be greater. Similarly, if φ(s) decreases, you must be turning
right, and if φ(s) decreases rapidly you should be turning more sharply to the right, with a greater curvature.
These considerations suggest that the derivative of φ is related to curvature. Indeed, by definition of the turning
angle
γ̇(s) = (cos φ(s), sin φ(s)) ,
and if we differentiate both sides of this equation with respect to s, we obtain
 
γ̈(s) = φ̇(s) − sin φ(s), cos φ(s) .

Since γ is unit speed, its curvature is κ = ||γ̈(s)||. Taking lengths of the vectors above, we obtain

κ(s) = |φ̇(s)|.

So the the derivative of φ is related to curvature: its absolute value is the curvature. Indeed, it is the signed
curvature, as we define now.

Definition 3.3. Let γ : I −→ R2 be a unit speed curve, with turning angle φ. The signed curvature of γ is the

6 Indeed, as we assume all our curves are C ∞ -smooth, the tangent vector varies in a smooth way.

13
 function
dφ
κs : I −→ R given by κs = .
ds

The signed curvature κs contains more information that the curvature κ: its absolute value is κ, and its sign is
positive or negative accordingly as φ is increasing or decreasing, i.e. accordingly as the curve turns left or right.
Remark 3.4. The s in κs stands for “signed”, not the arclength parameter s. Unfortunately, s is used both in κs ,
and the arclength parameter, so we often write κs (s), with two different s’s! The notation is unfortunate but we
follow the text here.
Remark 3.5. This definition only applies to unit speed curves! We’ll extend it to regular curves in general shortly.
Remark 3.6. We could have just written φ̇ instead of dφ
ds in the definition. But as we saw previously in our discussion
of curvature, often we need to consider curves γ(t) with parameters t distinct from the arclength parameter s.
Writing dφ
ds makes it clear that we are taking a derivative with respect to the arclength parameter s.
Let’s see an example.

Example 3.7. Calculate the signed curvature of the unit speed curve
 
1 1
γ(s) = cos 2s, sin 2s ,
2 2

which is a circle of radius 1/2 traversed anticlockwise. We know a circle of radius r has curvature 1/r, so the
signed curvature should be ±2. Travelling around a circle anticlockwise, you always turn left, so we expect
κs = 2.
To compute κs , we first find the velocity

γ̇(s) = (− sin 2s, cos 2s)

(which has ||γ̇|| = 1; γ is indeed a unit speed curve), and then the turning angle. For this we need to find φ(s)
such that γ̇(s) = (cos φ(s), sin φ(s)), that is,

(− sin 2s, cos 2s) = (cos φ(s), sin φ(s)) .

Perhaps the best way to find φ(s) is to sketch γ. In polar coordinates (r, θ), γ(s) is (1/2, 2s), and the velocity
γ̇ is at right angles to a radial vector, π/2 anticlockwise from it. So the turning angle is φ(s) = 2s + π2 . Indeed,
we can verify that
  π  π 
(cos φ(s), sin φ(s)) = cos 2s + , sin 2s + = (− sin 2s, cos 2s) = γ̇(s).
2 2
Taking a derivative of φ gives the signed curvature:
dφ d  π
κs = = 2s + = 2.
ds ds 2

Now try it yourself!

Exercise 11. Calculate the signed curvature of a unit speed circle of radius R, given by γ : R −→ R2 , γ(s) =
R (cos(s/R), sin(s/R)).

Exercise 12. Calculate the signed curvature of a unit speed line γ : R −→ R2 , γ(s) = at + b, where a, b are
constant vectors in R2 , and ||a|| = 1.

3.3 Reparametrisations and orientations

One way in which signed curvature differs from curvature is that it depends on the direction in which you travel
along a curve.

14
 Suppose you travel along a stretch of road which curves to the left. Then you turn around and travel back down
the same stretch of road, in the opposite direction. Now it curves to the right. So the signed curvature of a curve
depends on the direction of travel.
When you travel up and down the road, the signed curvatures you find in each direction will in fact be negatives
of each other.

Example 3.8. Let’s consider a unit speed curve γ(s) : R −→ R2 , and another curve δ(s) : R −→ R2 given by
δ(s) = γ(−s). So δ and γ travel along the same curve, in opposite directions. We’ll show that they have signed
curvatures which are negatives of each other.
Write φγ , φδ for turning angles, and κγs , κδs for the signed curvatures of γ, δ respectively.
Differentiating δ(s) = γ(−s) using the chain rule gives δ̇(s) = −γ̇(−s), so the velocity vectors δ̇ and γ̇ at
δ(s) = γ(−s) are negatives of each other. In particular, δ is also unit speed.
Now by definition of turning angle, we have

γ̇(s) = (cos φγ (s), sin φγ (s)) and δ̇(s) = cos φδ (s), sin φδ (s) .

Let’s try to relate the turning angles at corresponding points φδ (s) and φγ (−s). We expect that, since γ, δ
travel in opposite directions, one turning angle should be π more than the other (modulo 2π). And indeed, since
δ̇(s) = −γ̇(−s), we have

cos φδ (s), sin φδ (s) = (− cos φγ (−s), − sin φγ (−s))

= (cos (φγ (−s) + π) , sin (φγ (−s) + π)) ,

so
φδ (s) = φγ (−s) + π
is a valid turning angle for δ. Differentiating this equation with respect to s then gives

dφδ (s) dφγ (−s)

=− , i.e. κδs (s) = −κγs (−s).
ds ds
In other words, signed curvatures of γ, δ at corresponding points are negative.

These considerations lead to an important conclusion: the signed curvature can change when you reparametrise
a curve!
However, as we’ll see, the signed curvature does not change when you reparametrise it, provided the reparametrised
curve goes in the same direction. To this end, let’s make precise the definition of a reparametrisation “in the same
direction”. Recall that reparametrisations are defined via diffeomorphisms: a curve δ : I −→ Rn is a reparametri-
sation of γ : J −→ Rn if there exists a diffeomorphism ϕ : I −→ J such that δ = γ ◦ ϕ. The key idea is that ϕ is
either an increasing or decreasing function: if ϕ is increasing, then γ, δ are traversed in the same direction; if ϕ is
decreasing, then γ, δ are traversed in opposite directions.

Definition 3.9. Let I, J be intervals. A diffeomorphism ϕ : I −→ J is called orientation preserving if ϕ is

increasing, and orientation reversing if ϕ is decreasing.

Remark 3.10. In an exercise in the previous chapter, you proved that either ϕ′ > 0 everywhere, or ϕ′ < 0 everywhere.
Thus ϕ is either increasing or decreasing; there’s no other possibility. Every diffeomorphism of intervals is orientation
preserving or reversing.

Definition 3.11. Let γ, δ be curves such that δ is a reparametrisation of γ, so δ = γ ◦ϕ for some diffeomorphism
of intervals ϕ.
(i) If ϕ is orientation preserving, we say δ is an orientation-preserving reparametrisation of γ.

(ii) If ϕ is orientation reversing, we say δ is an orientation-reversing reparametrisation of γ.

15
Exercise 13. Using the definition, prove the following (hopefully intuitively reasonable, possibly obvious) facts.
(i) If δ is an orientation-preserving (resp. reversing) reparametrisation of γ, then γ is an orientation-preserving
(resp. reversing) reparametrisation of δ.

(ii) If β is an orientation-preserving reparametrisation of γ, and γ is an orientation-preserving reparametrisation

of δ, then β is an orientation-preserving reparametrisation of δ.
(iii) If β is an orientation-preserving reparametrisation of γ, and γ is an orientation-reversing reparametrisation
of δ, then β is an orientation-reversing reparametrisation of δ.

(iv) If β is an orientation-reversing reparametrisation of γ, and γ is an orientation-reversing reparametrisation of

δ, then β is an orientation-preserving reparametrisation of δ.

In the next section, we’ll use these ideas to extend the definition of signed curvature to regular curves.

3.4 Signed curvature of regular curves

We can extend the definition of signed curvature to regular curves, just as we did for ordinary curvature. In the
light of the previous section, though, we should restrict our attention to orientation-preserving orientations.

Definition 3.12. Let γ : I −→ R2 be a regular curve. The signed curvature of γ is the curvature of an
orientation-preserving unit speed reparametrisation of γ(t).
More precisely, if δ = γ ◦ ϕ is an orientation-preserving unit speed reparametrisation of γ, then the signed
curvature of γ(t) is equal to the signed curvature of δ(S) at the corresponding point, i.e. where ϕ(s) = t.

Again, we have the worry that there are many orientation-preserving unit speed reparametrisations of γ. How-
ever, we found in proposition 2.1 that any two unit speed reparametrisations β, δ are closely related. Moreover, case
(i) of the proposition refers to the case where β, δ travel along γ in the same direction, while case (ii) refers to β, δ
travelling in opposite directions. Only case (i) applies here, where all reparametrisations are orientation-preserving,
and so the proposition tells us that β(s) = δ(s+c) for some constant c. We can then prove the following proposition.

Proposition 3.13. The signed curvature of a regular curve γ is well defined.

More precisely, let β, δ be orientation-preserving unit speed reparametrisations of γ, related by β(s) = δ(s+c)
for some constant c. Let κβs , κδs be their signed curvature functions. Then κβs (s) = κδs (s + c).

Proof. We differentiate β(s) = δ(s + c) with respect to s to obtain β̇(s) = δ̇(s + c). Hence if φδ (s) is a turning
angle function for δ, then φβ (s) = φδ (s + c) is a turning angle function for β. Differentiating this equation then
gives κβ (s) = κδs (s + c) as desired.

Exercise 14. Following the arguments of the proposition above, and using case (ii) of proposition 2.1, see if you
can prove the following, which generalises example 3.8.
Given a regular curve γ and an orientation-reversing reparametrisation β, show that the signed curvatures of β
and γ at corresponding points are negatives of each other.

Because of the above, we can simply write that the curvature of a regular curve γ as
dφ
,κs =
ds
where s is any orientation-preserving arclength parameter.

3.5 The planar principal normal

We now define a concept which is useful in discussing curves in the plane and their signed curvature: the planar
principal normal.7 We will see in due course that it also has an interesting analogy for curves in R3 .
7 The text just calls it the “principal normal”, but I want to distinguish it from the “principal normal” to a curve in R3 , which is

similar, but different.

16
 Here, as usual8 in geometry, “normal” means “perpendicular”. A normal vector to a curve is a vector perpen-
dicular to the curve. By a “principal” normal vector, we mean a particular favourite choice of normal vector.
Given a unit speed curve γ : I −→ R2 , the velocity vector γ̇(s) is tangent to the curve at γ(s). The planar
principal normal N (s) is a particular choice of unit vector perpendicular to γ̇(s).
In two dimensions, if I give you a unit vector v, and ask you to find a unit vector w perpendicular to v, there
are not many possible answers! There are precisely two answers. Namely, you can rotate v by 90 degrees (or by
π/2 if you’re a grown-up) clockwise or anticlockwise. These two vectors are the only unit vectors perpendicular to
v.
So, if we rotate γ̇(s) by π/2 clockwise or anticlockwise, we get a unit normal vector. The convention for the
planar principal normal is that we rotate anticlockwise. Mathematicians tend to prefer to rotate things anticlockwise.
Maybe this is why they’re bad at keeping track of time.
In this context, we often write T (s) instead of γ̇(s): the T stands for “tangent”. But it is the same thing:
T (s) = γ̇(s). And we write N (s) for the planar principal normal, as we define now.

Definition 3.14. The planar principal normal vector N (s) to a unit speed curve γ in R2 at γ(s) is obtained
by rotating the tangent vector T (s) by π/2 anticlockwise.

Note that, as T (s) is a unit vector, so too is N (s). As we travel along the curve γ(s), we now have the tangent
vector T (s), pointing straight ahead, and the palanar principal normal vector N (s), pointing to our left. As the
curve turns, so too do T (s) and N (s), so that T (s) always points straight ahead, and N (s) always points left.

The tangent vector T (s) can be written in terms of a turning angle, T (s) = (cos φ(s), sin φ(s)). So can the
planar principal normal, as in the following lemma.

Lemma 3.15. If φ(s) is a turning angle function for γ(s), then

N (s) = (− sin φ(s), cos φ(s)) .

Proof. Since T (s) = (cos φ(s), sin φ(s)), the angular polar coordinate of T (s) is φ(s). If we just rotate this by
a further π/2 anticlockwise, we arrive at N (s). So
  π  π   
N (s) = cos φ(s) + , sin φ(s) + = − sin φ(s), cos φ(s) .
2 2

To calculate N (s) in practice, however, you don’t need to work out the turning angle; the following lemma
provides a shortcut.

Lemma 3.16. Let the unit speed curve γ : I −→ R2 be given by γ(s) = (x(s), y(s)). Then

N (s) = (−ẏ(s), ẋ(s)) .

This proof uses a general trick for rotating a vector in the plane 90◦ anticlockwise. When you rotate a vector
(a, b) anticlockwise by 90◦ , you get (−b, a). One way to see this is to just draw a picture and use elementary
geometry. Another way is to note that the π/2 anticlockwise rotation about the origin is given by the linear
transformation with matrix
cos π2 − sin π2
        
0 −1 0 −1 a −b
= , so that rotating (a, b) yields = .
sin π2 cos π2 1 0 1 0 b a
8I was going to say “as normal”, but that could be confusing!

17
The expressions T (s) = (cos φ, sin φ) and N (s) = (− sin φ, cos φ) above are indeed related in this way.

Proof. Differentiating γ(s) = (x(s), y(s)) gives T (s) = γ̇(s) = (ẋ(s), ẏ(s)). Rotating T (s) by π/2 anticlockwise,
as discussed above, yields N (s) = (−ẏ(s), ẋ(s)).

We can now compute N (s) fast! So let’s do it.

Example 3.17. Let’s compute N (s) for the a unit speed circle of radius R, given by γ(s) = R (cos(s/R), sin(s/R)).
We have T (s) = γ̇(s) = (− sin(s/R), cos(s/R)). The principal normal N (s) is obtained by rotating T (s) by
π/2 anticlockwise, so (using the trick above) is given by
 s  s 
N (s) = − cos , − sin .
R R

In the above example, N (s) is a scalar multiple of the γ(s), namely N (s) = − R1 γ(s). At each point on the circle,
the principal normal points towards the centre at the origin.

Exercise 15. Prove a converse result. Suppose you have a unit speed curve γ : R −→ R2 where N (s) is always a
scalar multiple of γ(s). That is, N (s) always points radially towards, or away from, the origin, so that we can write

N (s) = λ(s)γ(s) for some real-valued function λ(s).

Show that γ is a circle centred at the origin.

3.5.1 A moving frame

At each point on the unit speed curve γ(s), we now have two vectors T (s), N (s). They are both unit length, and
they are orthogonal to each other. Recalling our linear algebra, such a pair of vectors forms an orthonormal basis
for R2 .
Thus, at each point γ(s) on the curve γ, we have an orthonormal basis T (s), N (s), and these bases twist and
turn along with the curve.

Exercise 16. Linear algebra refresher. Prove that two vectors v, w ∈ R2 which are orthonormal form a basis for
R2 . More generally, prove that n vectors v1 , . . . , vn in Rn which are orthonormal, i.e. vi · vj = δij , form a basis for
Rn .

An ordered basis is sometimes called a frame. So we can say that we have a frame (T (s), N (s)) at each point
on the curve: this frame moves with the curve, and so we can say we have a moving frame.
In the next topic, we’ll see that this idea of a basis moving along a curve extends to 3 dimensions in a very
useful way. At a point on a 3-dimensional unit speed curve, we will see that there is an orthonormal basis formed
by a unit tangent vector T , a principal normal N , and a binormal vector B. But we’ll stick to 2 dimensions for the
moment.

3.5.2 The planar principal normal and signed curvature

We have now seen that the planar principal normal N is a normal vector to a unit speed curve.
But we have previously seen an example of a normal vector to a unit speed curve γ: its acceleration γ̈. Recall
that for a unit speed curve γ, we saw γ̇ · γ̈ = 0.
Now, as previously noted, there are not so many vectors perpendicular to a given vector in R2 ! Therefore, N
and γ̈ must be scalar multiples of each other.
To see what the scalar is, recall that in section 3.2, we expressed γ̈ in terms of a turning angle φ(s). We
differentiated T (s) = γ̇(s) = (cos φ(s), sin φ(s)) to obtain

γ̈ = φ̇(s) (− sin φ(s), cos φ(s)) .

On the right hand side, the term φ̇(s) is, by definition, the signed curvature. And from lemma 3.15, we now
recognise the vector (− sin φ(s), cos φ(s)) as none other than the planar principal normal N (s). So we have proved
the following proposition.

18
 Proposition 3.18. Let γ be a unit speed curve in R2 . Then

γ̈(s) = κs (s) N (s),

where κs is the signed curvature and N is the principal normal.

Remark 3.19. In the text, the equation γ̈(s) = κs N is used as a definition of signed curvature, and then κs = dφ
ds
is deduced as a consequence. We’ve done it the other way round. But the two definitions of the signed curvature
are equivalent.
Remark 3.20. Since γ̇ is also known as T , we could also write the equation in the proposition as

Ṫ = κs N.

We’ll see more equations of this type later, when we consider moving frames in R3 .
Remark 3.21. This proposition gives you another way to compute signed curvature of a unit speed curve. Given a
unit speed γ(s), compute the derivatives γ̇(s) = T (s) and γ̈(s). Then compute N (s) by rotating T (s) anticlockwise
by π/2. You should find that γ̈(s) is a scalar multiple of N (s). (If not, you’ve made a mistake!) The scalar factor
relating them is then κs .
This method allows you to compute the signed curvature, without any need to find the turning angle. But it
still only applies to unit speed curves.

Example 3.22. As in example 3.7, let’s calculate the signed curvature of the unit speed circle of radius 1/2,
given by
1
γ(s) = (cos 2s, sin 2s) .
2
(We calculated it using κs = dφ ds previously.)
Differentiating yields immediately T (s) = γ̇(s) = (− sin 2s, cos 2s). Differentiating, and rotating, respec-
tively give
γ̈(s) = (−2 sin 2s, −2 cos 2s) and N (s) = (− cos 2s, − sin 2s) .
From γ̈ = κs N we thus obtain κs = 2.

Exercise 17. Let r > 0 be a real number. Show that the signed curvature of the unit speed curve
 s  s 
γ(s) = r cos , r sin
r r
is 1/r, but the signed curvature of the unit speed curve
 s  s 
γ(s) = r sin , r cos
r r
is −1/r. So a circle of radius r traversed anticlockwise has signed curvature 1/r, but when traversed clockwise the
signed curvature is −1/r.

3.6 Calculating signed curvature

So far, we’ve seen two methods to calculate signed curvature.
ˆ In section 3.2, we saw an example applying the definition of κs directly. Given unit speed γ, we find the
turning angle φ and then calculate κs = φ̇.
ˆ In section 3.5.2, we saw another method. Given unit speed γ, we find the acceleration γ̈ and planar principal
normal N and compare: κs is the scalar relating them.
However, both these methods are indirect. Moreover, they only apply to unit speed curves. In this section we’ll
see two formulas, which calculate κs more directly. The first applies to unit speed curves; the second generalises
the first, and applies to regular curves in general.

19
 Both these formulas involve looking at the coordinates of γ(s), i.e. writing
γ(s) = (x(s), y(s))
and considering the coordinate functions x(s), y(s).

3.6.1 For a unit speed curve

When γ is unit speed, γ̇(s) = T (s) is a unit vector, and
T (s) = γ̇(s) = (ẋ(s), ẏ(s)) .
On the other hand, T (s) = (cos φ(s), sin φ(s)), where φ is a turning angle, so we have
ẋ(s) = cos φ(s), ẏ(s) = sin φ(s).
As it turns out, there is a clever way to extract the signed curvature κs = φ̇ from these expressions for ẋ(s) and
ẏ(s).

Proposition 3.23. Let γ : I −→ R2 be a unit speed curve and let γ(s) = (x(s), y(s)). Then

κs = ẋÿ − ẍẏ.

Proof. As above, ẋ(s) = cos φ(s) and ẏ(s) = sin φ(s). Differentiating these equations yields

ẍ(s) = −φ̇ sin φ, ÿ(s) = φ̇ cos φ.

Now κs = φ̇, and we compute

ẋÿ − ẍẏ = φ̇ cos2 φ + φ̇ sin2 φ = φ̇ cos2 φ + sin2 φ = φ̇ = κs .

Remark 3.24. This method for calculating κs is the easiest so far: it’s just a formula into which you plug γ(s). It
does not require you to fine N (s) or φ(s).
Remark 3.25. The expression ẋÿ − ẍẏ is not only a clever way to extract κs — it’s also an expression which arises
in a cross product. If you embed γ(s) into R3 as γ(s) = (x(s), y(s), 0), then the cross product of first and second
derivatives is given by  
i j k
γ̇ × γ̈ = det ẋ ẏ 0 = (0, 0, ẋÿ − ẍẏ) = (0, 0, κs ) .
ẍ ÿ 0
Now we had a formula for the curvature which involved just this cross product! It’s theorem 1.12, and in the case
of a unit speed curve, we can drop the factors of ||γ̇||, to obtain
κ = ||γ̇ × γ̈||.
Thus you can obtain the curvature κ by looking at the length of γ̇ × γ̈; but if you look at the vector γ̇ × γ̈ itself,
you see the signed curvature κs .
Remark 3.26. The formula κs = ẋÿ − ẍẏ can be used to define the turning angle. We can define the turning angle
to be given by the integral of κs = ẋÿ − ẍẏ, and then show it has the properties we expect of a turning angle, like
γ̇ = (cos φ, sin φ) and so on. This is how turning angle is shown to exist in the text, as we mentioned earlier in
section 3.1.

Example 3.27. Let’s once more calculate the signed curvature of the unit speed curve
 
1 1
γ(s) = cos 2s, sin 2s ,
2 2

now using this method.

20
 We differentiate to obtain

γ̇(s) = (− sin 2s, cos 2s) , γ̈(s) = (−2 cos 2s, −2 sin 2s) .

Then

κs = ẋÿ − ẍẏ
= (− sin 2s)(−2 sin 2s) − (cos 2s)(−2 cos 2s)
= 2 sin2 (2s) + 2 cos2 (2s) = 2.

Equivalently, we could take the cross product

 
i j k
0 = 0, 0, 2 sin2 2s + 2 cos2 2s = (0, 0, 2)


γ̇ × γ̈ = det  − sin 2s cos 2s
−2 cos 2s −2 sin 2s 0

and then κs is given by the z-coordinate 2.

3.6.2 For a regular curve

So far, all our computations have been of unit speed curves γ(s). But, as we saw in section 3.4, κs is well defined
for a regular curve. As with κ, we can find κs by taking a unit speed reparametrisation — but this is a pain. It
would be nice if we could compute κs directly!
As it turns out, there is a formula, generalising previous results.

Proposition 3.28. Let γ : I −→ R2 be a regular curve, and let γ(t) = (x(t), y(t)). Then

ẋÿ − ẍẏ
κs = .
||γ̇||3

Remark 3.29. This is by far the most useful formula in practice for the signed curvature. You just plug any regular
γ into this formula and you’re done.
Remark 3.30. You could have guessed this formula in advance!
We saw in theorem 1.12 that the curvature of a regular curve is given by

||γ̈ × γ̇||
κ= , which for unit speed γ reduces to κ = ||γ̇ × γ̈||.
||γ̇||3

And we saw in proposition 3.23 that signed curvature of a unit speed curve is given by
 
z coordinate of γ̇ × γ̈,
κs = ẋÿ − ẍẏ = .
regarded as lying in the xy-plane

Thus, for a unit speed curve, κ is given by the length of γ̇ × γ̈, while κs is given by its z-coordinate. For a regular
curve, κ is given by the length of
γ̇ × γ̈
,
||γ̇||3
and you might surmise that κs should be closely related to this vector, or more precisely, its z-coordinate. The
z-coordinate of (γ̇ × γ̈)/||γ̇||3 is given by (ẋÿ − ẍẏ)/||γ̇||3 , which is exactly the formula in the proposition.
In other words, the proposition can be written as
 
γ̇ × γ̈
κs = z-coordinate of ,
||γ̇||3

and it generalises proposition 3.23, the formula κs = ẋÿ − ẍẏ in the unit speed case.
Let’s prove the proposition; the method is similar to the unit speed case, with a few extra details.

21
 Proof. Since γ is regular, γ̇ = (ẋ, ẏ) is nonzero, and its direction is given by the turning angle φ, so

γ̇ = ||γ̇||(cos φ, sin φ) = (ẋ, ẏ) .

Thus
ẋ = ||γ̇|| cos φ, ẏ = ||γ̇|| sin φ.
Differentiating ẍ, ÿ yields
 
d
ẍ = ||γ̇|| cos φ − ||γ̇|| φ̇ sin φ,
dt
 
d
ÿ = ||γ̇|| sin φ + ||γ̇|| φ̇ cos φ.
dt
d
Now it is possible to simplify dt ||γ̇|| (and we did so in the proof of theorem 1.12), but we will just compute
d
ẋÿ − ẍẏ and observe that the terms with dt ||γ̇|| cancel out:
     
d d
ẋÿ − ẍẏ = ||γ̇|| cos φ ||γ̇|| sin φ + ||γ̇|| φ̇ cos φ − ||γ̇|| sin φ ||γ̇|| cos φ − ||γ̇|| φ̇ sin φ
dt dt
= ||γ̇||2 φ̇ cos2 φ + ||γ̇||2 φ̇ sin2 φ
= ||γ̇||2 φ̇.

Now φ̇, the derivative of the turning angle, is very similar to signed curvature! We have
ẋÿ − ẍẏ
φ̇ = ,
||γ̇||2

but the signed curvature is κs = dφ ds , where s is an (orientation-preserving) arclength parameter. But if we

remember that ds
dt = ||γ̇|| (note ds/dt > 0 as it’s an orientation-preserving parameter), then we have

dφ dφ dt 1 ẋÿ − ẍẏ
κs = = = φ̇ = .
ds dt ds ||γ̇|| ||γ̇||3

Example 3.31. Compute the signed curvature of the ellipse γ(t) = (a cos t, b sin t).
Differentiating, we first calculate γ̇, ||γ̇|| and γ̈:

γ̇ = (ẋ, ẏ) = (−a sin t, b cos t)

p
||γ̇|| = a2 sin2 t + b2 cos2 t
γ̈ = (ẍ, ÿ) = (−a cos t, −b sin t).

Hence
ẋÿ − ẍẏ (−a sin t)(−b sin t) − (−a cos t)(b cos t)
κs = =
||γ̇ 3 || a2 sin2 t + b2 cos2 t

3/2

ab
=
3/2
2
a2 sin t + b2 cos2 t

Note that when a = b this ellipse reduces to a circle of radius a, and the formula for the signed curvature
becomes a1 .

Exercise 18. Find the signed curvature of lima con γ : R −→ R, γ(t) = (1 + 2 cos t) (cos t, sin t).

22
Exercise 19. Find the signed curvature of the cusp curve γ : R −→ R, γ(t) = (t2 , t3 ), at all its regular points.

Exercise 20. Find the signed curvature of the logarithmic spiral γ(t) = eat (cos bt, sin bt). (If we identify R2 with
the complex plane then γ = e(a+bi)t .)

Exercise 21. Find the signed curvature of the “exponential speed circle” γ(t) = (cos (et ) , sin (et )) using proposition
3.28 directly. Observe why you obtain the same curvature as the more standard circle γ(t) = (cos t, sin t).

4 Integrating curvature
4.1 Total curvatures
It turns out that when you integrate the curvature over a curve, you get a quantity which has some interesting
properties. So we’ll make the following definitions.

Definition 4.1.
(i) The total curvature of a unit speed curve γ : (a, b) −→ Rn is
Z b
K= κ(s) ds.
a

(ii) The total signed curvature of a unit speed curve γ : (a, b) −→ R2 is

Z b
Ks = κs (s) ds
a

Note that we’ve only defined total curvatures for unit speed curves. And note that while total curvature is
defined for curves in any Rn , total signed curvature is defined only for curves in R2 .
When we have a closed curve (again, unit speed, and so with a periodic parametrisation), we can integrate
around the whole loop — one period — to obtain total curvatures.

Definition 4.2.

(i) The total curvature of a closed unit speed curve γ : R −→ Rn of period L is

Z L
K= κ(s) ds.
0

(ii) The total signed curvature of a closed unit speed curve γ : R −→ R2 of period L is
Z L
Ks = κs (s) ds.
0

4.2 Total signed curvature

Total signed curvature has a very nice geometric meaning. By definition,
dφ
κs = ;
ds
signed curvature is the rate of change of a turning angle with respect to an (orientation-preserving) arclength
parameter. Thus, by the fundamental theorem of calculus, integrating the signed curvature gives the total change

23
in turning angle:
Z b Z b
dφ
Ks = κs ds = ds = φ(b) − φ(a).
a a ds
You might recall that turning angle is not actually a well-defined thing — there’s an ambiguity of multiples of
2π — but once it’s chosen at one point, the choices at other points are determined by continuity. So the difference
between two values of the turning angle, such as φ(b) − φ(a), is well defined.
The total change in turning angle is something you can see by looking at a curve, as in the following examples.

Example 4.3. Consider the following curve γ:

At γ(0) and γ(1), γ points exactly horizontally to the right. Taking φ(0) = 0, we also have φ(1) = 0, and
hence Ks = 0: Z 1
Ks = κs ds = φ(1) − φ(0) = 0.
0

Here we chose φ(0) = φ(1) = 0, but we could equally have chosen φ(0) = 100π. In that case, continuity of
φ would demand φ(1) = 100π also, so that Ks = φ(1) − φ(0) = 0 again.

Example 4.4. Consider the following closed curve. It’s a closed curve with period T .

We have labelled γ(0) = γ(T ). Pick a value for the turning angle φ(0), and then see what happens as you go
around the curve. The turning angle φ(t) increases as you turn anticlockwise, and after you’ve gone around the
whole curve you will have turned through anticlockwise through a full 2π. So
Z T
Ks = κs ds = φ(T ) − φ(0) = 2π.
0

Example 4.5. Consider the three closed curves below. They loop around different numbers of times.
In the left curve, the turning angle makes two full turns anticlockwise as you proceed around the curve, so
Ks = 4π.
In the centre curve, the turning angle makes a half turn anticlockwise, and a half turn clockwise; these cancel
out to make a total change in turning angle of 0, so KS = 0.
In the right curve, the turning angle swings around all over the place, but after one loop, the turning angle
makes a full clockwise turn. So Ks = −2π

24
 In all the closed curve examples above, the turning angle may swing about clockwise and anticlockwise, and
some of these may cancel out, but the total signed curvature was always an integer multiple of 2π. Let’s prove this.

Proposition 4.6. The total signed curvature of a closed curve is an integer multiple of 2π.

Proof. Let γ : R −→ R2 be a unit speed parametrisation with period L. The total signed curvature is
Z L
Ks = κs ds = φ(L) − φ(0).
0

As γ is periodic, the tangent vectors γ̇(0) and γ̇(L) are equal. In particular, they point in the same direction.
It does not follow that φ(0) and φ(L) are equal; but it does follow that they differ by a multiple of 2π. So
Ks = φ(L) − φ(0) = 2kπ for some integer k.

4.3 Total curvature vs. total signed curvature

Let’s now consider total curvature K, and how it compares to total signed curvature Ks .
Recall that κ = |κs |; the signed curvature κs is equal to κ when you’re turning left, and −κ when you’re turning
right.
If, as you proceed along the curve, you always turn left, then κs = κ everywhere, and the total curvature equals
the total signed curvature, K = Ks . Similarly, if you always turn right, then κs = −κ everywhere, and the total
curvature is the negative of the total signed curvature, K = −Ks .
Most curves, however, will have some parts where you turn left (i.e. κs > 0) and some parts where you turn
right (i.e. κs < 0). In the integral for Ks , these parts contribute positively and negatively, and may cancel out.
But in the integral for K, both parts contribute positively, and do not cancel.
Let’s make this precise. First, we introduce some notation.

Definition 4.7. Given a unit speed curve γ : I −→ R2 , let K+ , K− denote the total curvature over those
sections of γ where κs > 0, and κs < 0 respectively. In other words,
Z Z
K+ = κ(s) ds, K− = κ(s) ds.
κs >0 κs <0

Note that by this definition, both K+ and K− are non-negative.

Lemma 4.8. For a unit speed curve γ in R2 ,

Ks = K+ − K− and K = K+ + K− .

Thus, while positive and negative signed curvature contribute positively and negatively to Ks , both contribute
positively to K.

Proof. The proof is just a matter of writing the integrals for K and Ks in terms of the integrals for K+ , K− .
First, for K: Z Z Z
K = κ(s) ds = κ(s) ds + κ(s) ds = K+ + K− .
I κs >0 κs <0

25
 Then, for Ks : Z Z Z
Ks = κs (s) ds = κ(s) ds + (−κ(s)) ds = K+ − K− .
I κs >0 κs <0

Here we have used the fact that κs = κ when κs > 0, and κs = −κ when κs < 0.

For any closed curve then Ks = K+ − K− is a multiple of 2π; but there is no reason in general to expect
K = K+ + K− to come out to be anything particularly nice.

Example 4.9. Consider the simple closed curve γ below. Most of the curve has κs > 0, turning left, but there
is a small section where it turns right.

As in previous examples, Ks = 2π. But whereas the negatively curved portion (turning right) cancels against
the positively curved portion (turning left) to give Ks , both contributions add to give the total curvature K.
So K > 2π.

4.4 Reconstructing a curve from the signed curvature

Imagine yourself driving along a curve γ in R2 at unit speed. As you go along, you turn left when κs > 0, and
right when κs < 0. The value of κs essentially tells you where your steering wheel is. A large positive κs means
your steering wheel is very far to the left. A small negative κs means your steering wheel is a little to the right. If
κs = 0, your steering wheel is in the center and you are going straight ahead.
A question then arises: if you follow the same driving directions, will you always end up with the same curve?
That is, does the signed curvature κs determine the curve γ? Can γ be reconstructed from κs alone?
The answer is, basically, yes. The driving directions given by κs essentially determine a unit speed curve.
Provided you specify where to start, and which way to go as you begin, the rest of the curve is completely determined
by signed curvature. That is, with appropriate initial conditions — a starting point, and a starting direction —
the curve γ can be reconstructed from κs .
Let’s see how this reconstruction can be done in an example.

Example 4.10. Let’s determine the unit speed curve γ(s) : R −→ R2 with initial conditions γ(0) = (0, 0),
φ(0) = 0, and κs (s) = s.
At s = 0, you start at the origin, facing to the right (the positive x direction). You proceed in this direction,
and your initial curvature is κs (0) = 0. As s increases, κs steadily increases, so you turn left, and more left, and
ever more left, in an ever-tightening spiral. You can similarly analyse what happens for negative s.
This curve is known as a Cornu spiral or Euler spiral and is used in the design of roads and railways.

26
 The first step in reconstructing γ(s) is to reconstruct the turning angle φ(s). We’ve seen that φ(s) is obtained
by integrating κs : Z s Z s
s2
φ(s) − φ(0) = κs (t) dt = t dt = .
0 0 2
Using the initial condition φ(0) = 0 we now have the turning angle

s2
φ(s) = .
2
The second step is to reconstruct the velocity γ̇(s). This is a unit vector (as γ is unit speed), and it points in a
direction determined by the turning angle. Thus
  2  2 
s s
γ̇(s) = (cos φ(s), sin φ(s)) = cos , sin .
2 2

The final step is to reconstruct the curve γ(s) itself. Since we have γ̇(s), we can integrate, using the initial
condition γ(0) = 0, to obtain γ(s).
Z s Z s  2  2  Z s  2 Z s  2 
t t t t
γ(s) = γ(0) + γ̇(t) dt = cos , sin dt = cos dt, sin dt
0 0 2 2 0 2 0 2

These integrals
s s
t2 t2
Z   Z  
cos dt, sin ldt
0 2 0 2
are known as Fresnel integrals. They are named after Augustin-Jean Fresnel, a 19th century French engineer and
physicist, and useful in optics. They are not elementary functions — so you can’t simplify the integrals any further.
However, as s → ∞, the integrals are known.

Exercise 22. If you’ve studied some complex analysis, see if you can use methods of contour integrals in the
complex plane to show that Z ∞  2 Z ∞  2 √
t t π
cos dt = sin dt = .
0 2 0 2 2
√ √ √ √
Hence show that Cornu’s spiral spirals into the two points ( π/2, π/2) and (− π/2, − π/2).

Although the above example is quite nice, the same method works in general to reconstruct a unit speed curve
γ(s) from its signed curvature function κs (s), and initial conditions γ(s0 ), φ(s0 ) at a particular time s = s0 .
(i) Integrate κs (s) and use the initial condition φ(s0 ) to obtain the turning angle φ(s).

(ii) Obtain the velocity vector as γ̇ = (cos φ, sin φ).

(iii) Integrate γ̇ and use the initial condition γ(s0 ) to obtain γ(s).

27
 As in the above example, the integrals obtained may be difficult or impossible to write as elementary functions.
Nonetheless, they reconstruct γ(s) uniquely. We can state this fact as a theorem.

Theorem 4.11. Let κs : I −→ R be a smooth function, where I is an interval. Let s0 ∈ I, let φ0 be a real
number and let p0 ∈ R2 . There exists a unique unit speed curve γ : I −→ R2 such that

(i) γ(s0 ) = p0 ,
(ii) γ̇(s0 ) = (cos φ0 , sin φ0 ), and
(iii) the signed curvature of γ is κs .

Proving the theorem rigorously amounts to applying the method of the above example for a general function.
There’s a proof in the text, or you could consider it an exercise.
Remark 4.12. The text proves a slightly different formulation of the uniqueness of the curve. Given two curves α
and β with the same signed curvature, the version in the text asserts they are related by an isometry M : R2 −→ R2
of the plane: α(s) = M (β(s)).
An isometry is a function that preserves distances. For instance, rotating the plane about a point by some angle
preserves distances on the plane; so rotations are isometries. Reflections and translations are also isometries of R2 .9
Following this line of thought, consider two unit speed curves α(s) and β(s) with the same signed curvature κs .
They may have different starting points, but if you apply a translation to δ(s) then you can make their initial points
γ(s0 ) and δ(s0 ) coincide. And the two curves may have different starting directions, but if you apply a rotation to
δ(s) then you can make the initial turning angles coincide. Once initial points and initial turning angles are equal,
the curves must coincide. This argument allows us to conclude the following theorem.

Theorem 4.13. Let α, β : I −→ R2 be two unit speed curves in R2 with the same signed curvature function
κs : I −→ R. Then there exist a translation and rotation which, when applied to δ(s), result precisely in γ(s).
Hence there is an isometry which takes δ(s) to γ(s).

Exercise 23. Show that a unit speed curve γ : R −→ R2 whose signed curvature is zero is a line.

Exercise 24. Show that a unit speed curve γ : R −→ R2 whose signed curvature is a nonzero constant is a circle.

Exercise 25. Find as much as you can about the unit speed curve γ : R −→ R2 with γ(0) = 0, φ(0) = 0 and
κs (s) = 2s . Write a formula for γ(s). Relate γ(s) to the Cornu’s spiral example above.

Exercise 26. We’ve just seen that the signed curvature κs , together with an initial point γ(s0 ) and turning angle
φ(s0 ), determine the curve γ(s).
Show that the same is not true for the curvature κ. In particular, find two unit speed curves γ(s) and δ(s),
which both have initial position γ(0) = δ(0) = 0 and initial turning angle φ(0) = 0, and which have curvature given
by κ = 1.

5 Curves in 3-dimensional space

Curves in R3 are a more complicated affair than in R2 . Walking in the plane, you can only turn left and right.
Walking in space, you can turn in many more ways. So there is no natural way to generalise signed curvature from
R2 to R3 . Instead, we have other techniques.
9 In fact any isometry of R2 (or any Rn ) is a composition of reflections, rotations and translations. Even better, any isometry can be

expressed as a composition of reflections only.

28
5.1 The moving frame
One idea that carries over well from R2 to R3 is the idea of a moving frame. In section 3.5.1 we discussed how the
vectors T (s) and N (s) form an orthonormal basis, or moving frame, along a unit speed curve γ(s). Here T (s) = γ̇(s)
is the velocity vector, and N (s) is the planar principal normal.
Analogously, there is an orthonormal basis (T (s), N (s), B(s)) forming a moving frame along a unit speed curve
γ(s) in R3 . It’s also known as a Frenet frame.
Throughout this section, γ denotes a unit speed curve I −→ R3 . We’ll define T (s), N (s) and B(s) in turn.
We start by defining the tangent vector T (s) as in R2 :

T (s) = γ̇(s).

Next, we turn to the principal normal N (s). For a curve in the plane, we defined the planar principal normal N (s)
by rotating T (s) anticlockwise by π/2. In space, though, nobody can hear your ideas of clockwise or anticlockwise,
no matter how loud you scream them.
Instead, we use the fact that, for a unit speed curve γ, the acceleration γ̈ is always normal to the curve, since
γ̇ · γ̈ = 0. However, γ̈ is usually not a unit vector: its length is the curvature κ, by definition.

||γ̈|| = κ

So if we divide γ̈ by its length κ, we obtain a unit vector normal to γ.

Definition 5.1. Suppose κ(s) ̸= 0. Then the principal normal N (s) at γ(s) is

γ̈(s)
N (s) = .
κ(s)

Remark 5.2. Since γ̇ = T we have γ̈ = Ṫ , so the principal normal can equally be written as

Ṫ
N= .
κ
Remark 5.3. This definition relies on γ being unit speed! If γ is not unit speed, then γ̈ will not usually be orthogonal
to T .
Remark 5.4. The principal normal defined here is similar to, but different from, the planar principal normal of
definition 3.14. We write N (s) for both, but it should be clear from the context which one we mean.
Remark 5.5. Both the planar principal normal in R2 , and principal normal in R3 , are closely connected to the
acceleration γ̈.
In R2 , we saw that the planar principal normal N satisfies (proposition 3.18) the equation

γ̈ = κs N.

With the principal normal, we have a straightforward generalisation, using the curvature κ rather than the signed
curvature (which no longer exists in R3 ):
γ̈ = κN.
Remark 5.6. When κ(s) = 0, N (s) is not defined. In particular, this means that for very simple curves like a
straight line, no principal normal vectors are defined!
You can think about it this way. At each point, either a curve has zero curvature, in which case it’s pretty
simple there, or the curvature is nonzero, and in that case you have the principal normal to help you out.
The third vector completing our orthonormal basis at γ(s) is the binormal vector B(s). When you have two
orthogonal vectors in R3 , there’s a cheap way to get a third vector orthogonal to both of them: take their cross
product.
As it turns out, we are that cheap.

Definition 5.7. The binormal vector B(s) is

B(s) = T (s) × N (s).

29
 Since T and N are orthogonal, and B is their cross product, all three vectors T, N, B are mutually orthogonal.
Moreover, since T and N are unit vectors, the length of B is given by ||T × N || = ||T || ||N || sin π/2 = 1. So the
three vectors T, N, B form an orthonormal basis at each point, or moving frame. Moving along γ, T, N, B may
twist and turn, but they always remain a right-handed orthonormal basis.

Additionally, as B = T × N , the basis (T, N, B) is right-handed : if you you curl the fingers on your right hand
around from T to N , then your right thumb points in the direction of T × N = B.10

Definition 5.8. At each point γ(s) of γ where κ(s) ̸= 0, the moving frame or Frenet frame or Frenet-Serret
frame is the triple of vectors (T (s), N (s), B(s)).

Example 5.9. Let’s compute the Frenet frame for the unit speed circle in the yz plane given by γ(s) =
(0, cos s, sin s).
We first compute T (s) = γ̇(s) = (0, − sin s, cos s). Then to compute N (s), we differentiate to find γ̈(s) =
(0, − cos s, − sin s), which has length κ(s) = ||γ̈(s)|| = 1, and then N (s) = γ̈(s)/κ(s) = (0, − cos s, − sin s). The
binormal B(s) is given by the cross product B(s) = T (s) × N (s) = (1, 0, 0).
Hence the moving frame is given by

T (s) = (0, − sin s, cos s), N (s) = (0, − cos s, − sin s), B(s) = (1, 0, 0).

2 2
sin t, 3t . Find its Frenet frame.


Exercise 27. Consider the unit speed helix γ : R −→ R2 given by γ(s) = 3 cos t, 3

Exercise 28. Consider Cornu’s spiral, in the xy-plane, embedded in R3 , given by

Z s  2 Z s  2 
t t
γ : R −→ R3 , γ(s) = cos dt, cos dt, 0 .
0 2 0 2

Find its Frenet frame.

5.2 How does the moving frame move?

Now that we have the Frenet frame (T (s), N (s), B(s)), we ask: how does the frame move? In other words, what
are Ṫ , Ṅ and Ḃ?
As it turns out, we will be able to write Ṫ , Ṅ and Ḃ in terms of T, N, B, quite nicely. This isn’t so surprising:
since the vectors T, N, B form a basis, any vector in R3 can be written as a linear combination of T, N and B.
As a preliminary matter, let’s think about how to write an arbitrary vector v ∈ R3 in terms of T, N, B. As
T, N, B is a basis, there are unique real numbers a, b, c such that

v = aT + bN + cB.

To express v in terms of T, N, B, we must find the coefficients a, b, c — and we can find them by taking some dot
products, as the following lemma shows.

10 Alternatively, if you extend your thumb, index and middle fingers on your right hand, they can point in the directions T, N, B

respectively.

30
 Lemma 5.10. Let v be any vector in R3 . Then v = aT + bN + cB, where

a = v · T, b = v · N, c = v · B.

Before proving the lemma, let’s observe the result of dot products among the vectors T, N, B of the Frenet
frame. Since T, N, B are all unit length,

T · T = N · N = B · B = 1,

and since T, N, B are orthogonal,

T · N = N · B = B · T = 0.

Proof. We simply compute

v · T = (aT + bN + cB) · T = aT · T + bN · T + cB · T = a
v · N = (aT + bN + cB) · N = aT · N + bN · N + cB · N = b
v · B = (aT + bN + cB) · B = aT · B + bN · B + cB · B = c.

Thus, to find a vector in terms of T, N, B, we find its dot products with T, N, B. We will follow this mantra to
express Ṫ , Ṅ , Ḃ in terms of T, N, B, by finding their dot products with T, N, B.
We begin with Ṫ . We can actually write Ṫ in terms of T, N, B without any dot products, because the definition
of the principal normal N = Ṫκ tells us that
Ṫ = κN.
We next turn to Ḃ, finding its dot products with T, N, B.
What is the dot product Ḃ · B? We can find out by differentiating the equation B · B = 1:

Ḃ · B + B · Ḃ = 0, so 2Ḃ · B = 0,

and hence Ḃ · B = 0.
Next, let’s find Ḃ · T . We can again differentiate a dot product: starting from B · T = 0, differentiation yields

Ḃ · T + B · Ṫ = 0, hence Ḃ · T = −B · Ṫ .

However we’ve already calculated Ṫ = κN . Hence

Ḃ · T = −κB · N = 0.

We’ve now found that Ḃ · T = 0 and Ḃ · B = 0. So the coefficients of T and B in Ḃ are both zero. Hence the
only nonzero coefficient in Ḃ is the N coefficient. In other words, Ḃ is a scalar multiple of N . The scalar involved
turns out to be an important quantity.

Definition 5.11. Let γ : I −→ R3 be a unit speed curve with nonzero curvature. The torsion of γ is the
function τ : I −→ R defined by
Ḃ(s) = −τ (s)N (s).

The symbol τ is the Greek letter tau.

Having written Ṫ and Ḃ in terms of T, N, B, it remains to consider Ṅ . Our mantra says we need to find the
dot products Ṅ · T , Ṅ · N , and Ṅ · B. We gain information about these dot products by differentiating N · T = 0,
N · N = 1 and N · B = 0, obtaining respectively

Ṅ · T + N · Ṫ = 0
2Ṅ · N = 0
Ṅ · B + N · Ḃ = 0

31
The first equation tells us Ṅ · T = −N · Ṫ , and we have calculated Ṫ = κN above, so

Ṅ · T = −N · (κN ) = −κ(N · N ) = −κ.

The second equation gives Ṅ · N = 0. The third equation gives Ṅ · B = −N · Ḃ; using Ḃ = −τ N , we obtain

Ṅ · B = −N · (−τ N ) = τ (N · N ) = τ.

Having obtained all three dot products of Ṅ with T, N, B, we have fulfilled the mantra and we conclude

Ṅ = −κT + τ B.

We have now completed the task of writing the derivatives Ṫ , Ṅ , Ḃ in terms of T, N, B. The resulting equations,
which tell us how the Frenet frame moves, are known as the Frenet-Serret equations or frame equations.

Theorem 5.12 (Frenet-Serret / frame equations). For a unit speed curve γ : I −→ R3 with κ ̸= 0,

Ṫ = κN
Ṅ = −κT +τ B
Ḃ = −τ N

Remark 5.13. The frame equations have been written suggestively so as to suggest a matrix; and indeed the frame
equations can be written in matrix form as
    
T 0 κ 0 T
d   
N = −κ 0 τ  N  .
ds
B 0 −τ 0 B
Remark 5.14. The matrix  
0 κ 0
M = −κ 0 τ
0 −τ 0
is skew-symmetric, M T = −M . This is not a coincidence! In fact if you have any set of orthonormal vectors
changing over time U (t), V (t), W (t) then    
U U
d  
V = AV 
dt
W W
where A is a 3 × 3 skew-symmetric matrix, AT = −A.
Why is this? Differentiating U · U = V · V = W · W = 1 gives U · U̇ = V · V̇ = W · Ẇ = 0; this says that the
diagonal entries of A are 0. Differentiating U · V = V · W = W · U = 0 gives equations like U · V̇ = −U̇ · V , showing
that A21 = −A12 , showing that off-diagonal entries are skew-symmetric.
Remark 5.15. Advanced aside. (This comment is beyond the scope of the course, but it may be of interest if it
means anything to you.) When you have a set of orthonormal vectors (U (t), V (t), W (t)) changing over time, the
frames at any two times are related by a rotation. The frame at any time (U (t), V (t), W (t)) is given by a rotation
R(t) applied to an initial frame (U (t0 ), V (t0 ), W (t0 )). That is,
   
U (t) U (t0 )
 V (t)  = R(t)  V (t0 ) 
W (t) W (t0 )

where R(t) is a rotation matrix : a matrix such that RT R = 1. Rotation matrices form a group known as SO(3),
which is a type of group known as a Lie group. Then R(t) is a smoothly varying family of rotation matrices, or
equivalently, R(t) is a curve in SO(3).
If we differentiate the above equation with respect to t, we find
   
U (t) U (t0 )
d  dR(t)
V (t)  =  V (t0 )  .
dt dt
W (t) W (t0 )

32
Here dR(t)
dt is the tangent vector to a curve in SO(3), the space of rotation matrices!
The tangent vectors to SO(3) are a very important object in many fields of mathematics: they form what is
called a Lie algebra, denoted so(3). As it turns out, the Lie algebra so(3) of tangent vectors to SO(3) at the identity
consists of skew-symmetric matrices. This is a deep reason why the matrix in the frame equations is skew-symmetric.

5.3 Frame equations in 2 dimensions

Although the frame equations are stated for curves in R3 , they apply to curves in R2 as well — you can embed
R2 in R3 as the xy-plane, as we’ve noted previously. When you do this, since γ(s) lies in the xy-plane for all s,
so too do the derivatives γ̇, γ̈ — and hence T, N also lie in the xy-plane. The binormal B must then point in the
z direction. But there are not many unit vectors in the z direction: only (0, 0, ±1)! As B varies continuously, B
must in fact be constant, and hence Ḃ = −τ N = 0. So the torsion τ = 0, and in the 2-dimensional case, the frame
equations reduce to
Ṫ = κN
Ṅ = −κT
Ḃ = 0
or just     
d T 0 κ T
=
ds N −κ 0 N
We can think of these as the “2D frame equations”.
Thus, when a curve γ(s) lies in the xy-plane, its torsion is zero. This is no coincidence: we’ll see shortly that τ
is a measure of the “non-planarity” of γ(s).

Exercise 29. In this exercise we’ll compare two equations satisfied by the two different definitions of principal
normal.
Let γ : I −→ R2 be a unit speed plane curve with unit tangent T .
(i) If N is the principal normal defined by rotating T by π/2 anticlockwise, and κs is the signed curvature, show
that
Ṅ = −κs T.

(ii) If κ is the curvature and N is the principal normal as defined by Ṫ = κN , show that

Ṅ = −κT.

2 2
sin t, 3t .


Exercise 30. In exercise 27, you found the Frenet frame (T, N, B) of the unit speed helix γ(s) = 3 cos t, 3
Now differentiate T, N, B and verify the frame equations directly.

5.4 What does the torsion mean?

In the process of deriving the frame equations, we introduced a new quantity: the torsion τ . Let’s now investigate
the geometric meaning of τ .
Roughly, τ measures the “non-planarity” of γ. We saw above that when γ lies in the xy-plane, τ = 0. In fact,
the same proof works for any plane, and even more, the following proposition is true.

Proposition 5.16. Let γ : I −→ R3 be a unit speed curve with κ ̸= 0 everywhere. The curve γ lies in a plane
if and only if τ = 0 everywhere.

As this is an “if and only if” statement, we have to prove two statements: first, that if γ lies in a plane, then
τ = 0; and second, that if τ = 0 then γ lies in a plane. Proving this second statement requires some facts about 3D
geometry and planes, which we’ll revise before jumping into the proof.

Exercise 31.
(i) Show that a plane with normal vector n is given by an equation of the form n · (x, y, z) = c, where c is a

33
 constant.
(ii) Conversely, show that the set of points (x, y, z) ∈ R3 satisfying an equation of the form n · (x, y, z) = c , where
n is a constant vector and c is a real constant, is a plane normal to n.

Let’s now give the proof of proposition 5.16; I think it is quite nice.

Proof. First we’ll prove that if γ lies in a plane, then τ = 0, following the same argument as in section 5.3. If
γ lies in a plane, then the derivatives γ̇, γ̈ are tangent to this plane. Then T = γ̇ and N = γ̈/κ are also tangent
to this plane. The binormal B must then be a unit normal to the plane.
Now, a given plane does not have many unit normal vectors! It has precisely two of them, pointing in
opposite directions (“up” and “down”). So B(s) must be one of these two vectors. As B(s) varies continuously
with s, it can’t jump from one of these two vectors to the other. Hence B(s) is constant.
Since B is constant, Ḃ = 0; and then by the third frame equation Ḃ = −τ N , so τ = 0 as desired.
For the converse, suppose that τ = 0 everywhere along a curve; we will show γ lies in a plane. Reversing
the end of the previous argument, consider the third frame equation Ḃ = −τ N . As τ = 0 we have Ḃ = 0, and
hence B is constant.
We now claim γ(s) lies entirely on a plane normal to B. To prove this, we use the 3D geometry fact from
above: a plane normal to B is given by an equation B · (x, y, z) = c, for some constant c.
So to show that γ(s) always lies on a single plane, normal to B, it suffices to show that B · γ(s) is a constant.
And to show that B · γ(s) is constant, we show its derivative is zero. This derivative is

d
(B · γ(s)) = Ḃ · γ(s) + B · γ̇(s).
ds
d
We’ve seen Ḃ = 0. And B · γ̇ = B · T = 0, as B, T are perpendicular. Thus ds (B · γ(s)) = 0. We conclude that
B · γ(s) is a constant, and γ(s) lies on a plane normal to B.

5.5 Planes associated to a curve in space

If you take any two of the three vectors T, N, B of the Frenet frame, they span a plane. All these planes have
names.

Definition 5.17. Let γ be a unit speed curve with Frenet frame (T, N, B).
ˆ The osculating plane of γ is the plane spanned by T and N .

ˆ The normal plane of γ is the plane spanned by N and B.

ˆ The rectifying plane of γ is the plane spanned by T and B.

Remark 5.18. The osculating plane, spanned by T = γ̇ and N = γ̈/κ, is also the plane spanned by γ̇ and γ̈. In this
sense, it is the “plane in which γ lies to second order”.11
The osculating plane to γ at a point γ(s) also contains the osculating circle to γ at γ(s). The osculating circle
is discussed in the following exercise.

Exercise 32. Let γ(s) be a unit speed curve with κ(s0 ) ̸= 0. Show that there is a unit speed circle β(s) that agrees
with γ at s0 to second order, i.e. β(s0 ) = γ(s0 ), β̇(s0 ) = γ̇(s0 ), and β̈(s0 ) = γ̈(s0 ).
This circle is called the osculating circle to γ at γ(s0 ). Show that its radius is 1/κ(s0 ).

Remark 5.19. The normal plane, being spanned by N and B, is normal to T , and hence normal to the curve γ. So
the name “normal plane” makes sense.
The osculating plane helps us a little further to understand torsion. Consider the second frame equation

Ṅ = −κT + τ B.
11 Osculate is also a word in non-mathematical English. To osculate is to kiss. The plane, so nicely tangent to γ to second order, is

“kissing” it in some sense.

34
This equation says that the change of the normal vector N consists of two components: a T component and a B
component. The T component is essentially similar to what we saw in two dimensions: it says that N turns in the
T -N plane according to the curvature.
However, the second term is new. It says that the change in N has a component in the direction of B. This
makes N twist out of the osculating plane. And the τ says how big this effect is. In this sense, τ measures how
much N is twisting out of the osculating plane.

5.6 Reconstructing a 3D curve from the curvature and torsion

We saw that for a curve in 2 dimensions, the signed curvature (together with initial conditions) determines the unit
speed curve γ(s).
In three dimensions, there is no signed curvature; and the curvature is definitely not enough to reconstruct the
curve. There are very different curves in R3 with the same curvature, as the following exercise shows.

Exercise 33. Show that a circle of radius 1, and the helix γ(t) = ( 12 cos t, 12 sin t, 12 t), both have constant curvature
1.

However, curves in space can be reconstructed from initial conditions, together with curvature — and a little
something more. That little something more is the torsion.
In space, your driving directions are not just the curvature — the are the curvature and the torsion. Let’s state
this as a theorem.
Theorem 5.20. Let κ > 0 and τ be smooth functions I −→ R, where I is an interval. Then given a starting
time s0 ∈ (a, b), a starting point p0 ∈ R3 , and a starting right-handed orthonormal frame (T0 , N0 , B0 ), there is
a unique unit speed curve γ : I −→ R3 such that
(i) γ(s0 ) = p0 ;
(ii) the Frenet frame (T (s), N (s), B(s)) of γ satisfies (T (s0 ), N (s0 ), B(s0 )) = (T0 , N0 , B0 );
(iii) γ(s) has curvature κ(s) and torsion τ (s).

We won’t give a proof, but we’ll try to convey the idea. The idea is simply that the frame equations

Ṫ = κN
Ṅ = −κT +τ B
Ḃ = −τ N

form a system of differential equations for T, N, B. Since each of T, N, B has three coordinates, and we need to find
the coordinates of all of T, N, B as functions of s, this is really a system of nine differential equations. And the
differential equations are coupled : they depend on each other.
The system of differential equations may be hard to solve, but it turns out that it can be solved! In an advanced
course on differential equations you might see (or have seen) methods for solving such systems. At the very least,
there are theorems that say that solutions exist. Moreover, given initial conditions (T0 , N0 , B0 ), the solution is
unique.
So, by solving a system of differential equations, we can determine (T (s), N (s), B(s)) uniquely. And once we
have T (s) = γ̇(s), we can integrate, using the initial condition γ(s0 ) = p0 , in order to obtain γ(s).

5.7 The moving frame for a regular curve

When γ is not unit speed, you can still set up a moving frame by taking a unit speed reparametrisation. You can
do this whenever γ is regular. If the unit speed reparametrisation is orientation-preserving, then you will obtain a
well-defined T, N, B.
As discussed already in relation to curvature and signed curvature, any two orientation-preserving unit speed
reparametrisations β, δ of γ are very closely related, with β(s) = δ(s + c) for some constant c. It’s then not difficult
to show that we obtain the same frame T, N, B at each point.

35
Exercise 34. What happens to a Frenet frame when you traverse a curve in the opposite direction? It is perhaps
clear that T is replaced by −T , but what happens to N and B?

However, as we already found several times, finding an explicit unit speed reparametrisation can be painful.
We’d prefer to be able to find T, N, B directly from a regular curve. As it turns out, this is not too difficult.
Let’s see why. First, finding T is not so bad: T is tangent to the curve, so you can take a unit vector in the
direction of γ̇.
As it turns out, the next easiest vector to find is B, because of the following lemma.

Lemma 5.21. Let γ(t) be a regular curve with κ > 0 everywhere, and let β(s) be an orientation-preserving
unit speed reparametrisation, with Frenet frame (T (s), N (s), B(s)). Then at a point γ(t) corresponding to β(s),

γ̇(t) × γ̈(t)
B(s) = .
||γ̇(t) × γ̈(t)||

As has now become a standard abuse of notation, rather than writing β̇(s) and γ̇(t), we will instead write γ̇ = dγ
dt
and dγ
ds .
The key step in this proof is a calculation we did previously. In our proof of the formula for curvature of a
regular curve, theorem 1.12, we similarly had a regular curve γ(t), and we calculated the derivatives of γ with
respect to an arclength parameter s, in terms of the derivatives with respect to t. In particular, we found dγds and
d2 γ dγ d2 γ
ds2 in terms of γ̇ = dt and γ̈ = dt2 . We found (lemma 2.4 and equation (6)) that

dγ γ̇ d2 γ γ̈(γ̇ · γ̇) − γ̇(γ̇ · γ̈)

= , 2
= .
ds ||γ̇|| ds ||γ̇||4
d2 γ
The salient fact about this equation is that it expresses ds2 as a linear combination of γ̇ and γ̈.

Proof. By definition, B(s) = T (s) × N (s). Now T (s) is given by the velocity with respect to the (orientation-
preserving) arclength parameter s,
dγ γ̇
T (s) = = .
ds ||γ̇||
Similarly, N (s) is given by the acceleration with respect to s, divided by the curvature:

1 d2 γ γ̈(γ̇ · γ̇) − γ̇(γ̇ · γ̈)

N (s) = = .
κ ds2 κ||γ̇||4

Thus  
γ̇ γ̈(γ̇ · γ̇) − γ̇(γ̇ · γ̈)
B(s) = T (s) × N (s) = × ,
||γ̇|| κ||γ̇||4
which contains a γ̇ × γ̈ term and a γ̇ × γ̇ term. The latter is zero, and hence we have
γ̇ · γ̇ 1
B(s) = γ̇ × γ̈ 5
= γ̇ × γ̈ .
κ||γ̇|| κ||γ̇||3

The key point here is that B(s) is a positive scalar mutliple of γ̇ × γ̈. We also know B(s) is a unit vector; hence
it is the unit vector in the direction of γ̇ × γ̈.

Once you have T and B, you can find N easily as a cross product N = B × T . (Check that the order of this
cross product is correct!)

5.8 Torsion for a regular curve

Finally, we give a formula for τ for a general (not necessarily unit speed) regular curve. This formula makes it
straightforward to find torsion, without having to find a unit speed reparametrisation!

36
 Proposition 5.22. Let γ be a regular curve in R3 with κ ̸= 0 everywhere. Then
...
(γ̇ × γ̈) · γ
τ= .
||γ̇ × γ̈||2

We prove this, assuming γ is unit speed. For the general case, see the text (prop. 2.3.1, p. 48).

Proof. Begin by computing the first two derivatives of γ, in terms of T, N, B:

γ̇ = T, γ̈ = Ṫ = κN.

We can then compute the cross product

γ̇ × γ̈ = T × κN = κB.
...
Next, we obtain the third derivative γ by differentiating γ̈ = κN ; we use the product rule and the frame
equation Ṅ = −κT + τ B.
...
γ = κ̇N + κṄ = κ̇N + κ(−κT + τ B) = κ̇N − κ2 T + κτ B.

We now obtain the dot product in the numerator as

...
(γ̇ × γ̈) · γ = κB · (N − κ2 T + κτ B) = κ2 τ.

The denominator is ||γ̇ × γ̈||2 = ||κB||2 = κ2 ; dividing numerator by denominator gives τ as desired.

Exercise 35. Compute the curvature and torsion of the helix γ(t) = (a cos t, a sin t, bt), and show that they are
constant.

Exercise 36. Show that every curve in R3 of constant curvature and constant torsion is a helix.

Exercise 37. Consider the curve γ : R −→ R2 given by γ(t) = (t, t2 , t3 ). What is its curvature and torsion?

37