Why Do the Unit Quaternions Double-Cover

the Space of Rotations?

Neil Bickford

1. Computing with Quaternions

The unit quaternion
𝜃𝜃
cos � �
⎛ 2 ⎞
𝜃𝜃
⎜ 𝑥𝑥 sin � � ⎟
⎜ 2 ⎟
𝒒𝒒 = ⎜
𝜃𝜃 ⎟
⎜ 𝑦𝑦 sin � �⎟
⎜ 2 ⎟
𝜃𝜃
𝑧𝑧 sin � �
⎝ 2 ⎠
represents a counterclockwise rotation by the angle 𝜃𝜃 around the normalized axis 𝒏𝒏 = (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 .

We’ll sometimes use 𝑞𝑞𝑤𝑤 , 𝑞𝑞𝑥𝑥 , 𝑞𝑞𝑦𝑦 , and 𝑞𝑞𝑧𝑧 to refer to the four components of a quaternion.

We can compose quaternions in the same way we can compose rotations: the product r of
quaternions p and q

𝒓𝒓 = 𝒑𝒑𝒑𝒑
represents the rotation given by performing q, then by performing p. For instance, if q is a rotation by
the angle 𝜃𝜃 around the axis (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 , then the product of q with itself is twice the rotation:

2
2𝜃𝜃 2𝜃𝜃 2𝜃𝜃 2𝜃𝜃 𝑇𝑇
𝒒𝒒 = �cos � � , 𝑥𝑥 sin � � , 𝑦𝑦 sin � � , 𝑧𝑧 sin � �� .
2 2 2 2
Similarly,

3𝜃𝜃 3𝜃𝜃 3𝜃𝜃 3𝜃𝜃 𝑇𝑇

𝒒𝒒3 = �cos � � , 𝑥𝑥 sin � � , 𝑦𝑦 sin � � , 𝑧𝑧 sin � ��
2 2 2 2
and so on. The inverse of a unit quaternion1 is given by reversing its rotation:

θ θ 𝜃𝜃 θ 𝑇𝑇
𝒒𝒒−1 = �cos �− � , 𝑥𝑥 sin �− � , 𝑦𝑦 sin �− � , 𝑧𝑧 sin �− �� .
2 2 2 2
If we ever need to, we can write out the result of performing q, then p (like finding the rotation
that corresponds to a product of two other rotations), as another quaternion:
𝑝𝑝𝑤𝑤 𝑞𝑞𝑤𝑤 − 𝑝𝑝𝑥𝑥 𝑞𝑞𝑥𝑥 − 𝑝𝑝𝑦𝑦 𝑞𝑞𝑦𝑦 − 𝑝𝑝𝑧𝑧 𝑞𝑞𝑧𝑧
𝑝𝑝𝑥𝑥 𝑞𝑞𝑤𝑤 + 𝑝𝑝𝑤𝑤 𝑞𝑞𝑥𝑥 − 𝑝𝑝𝑧𝑧 𝑞𝑞𝑦𝑦 + 𝑝𝑝𝑦𝑦 𝑞𝑞𝑧𝑧
𝒑𝒑𝒑𝒑 = �𝑝𝑝 𝑞𝑞 + 𝑝𝑝 𝑞𝑞 + 𝑝𝑝 𝑞𝑞 − 𝑝𝑝 𝑞𝑞 �
𝑦𝑦 𝑤𝑤 𝑧𝑧 𝑥𝑥 𝑤𝑤 𝑦𝑦 𝑥𝑥 𝑧𝑧
𝑝𝑝𝑧𝑧 𝑞𝑞𝑤𝑤 − 𝑝𝑝𝑦𝑦 𝑞𝑞𝑥𝑥 + 𝑝𝑝𝑥𝑥 𝑞𝑞𝑦𝑦 + 𝑝𝑝𝑤𝑤 𝑞𝑞𝑧𝑧

This gives us a way to express the product of two quaternions. Note that we have to be careful about
the order in which we apply quaternions (and rotations); for instance, a 90° rotation around the x axis
followed by a 90° rotation around the y axis produces a different result than a 90° rotation around the y
axis followed by a 90° rotation around the x axis.

We can also express this as a matrix-vector product, which is useful for computer
implementation: if 𝒓𝒓 = 𝒑𝒑𝒑𝒑, then
𝑟𝑟𝑤𝑤 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑥𝑥 −𝑝𝑝𝑦𝑦 −𝑝𝑝𝑧𝑧 𝑞𝑞𝑤𝑤
𝑟𝑟𝑥𝑥 𝑝𝑝𝑥𝑥 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑧𝑧 𝑝𝑝𝑦𝑦 𝑞𝑞𝑥𝑥
� 𝑟𝑟 � = � 𝑝𝑝 𝑝𝑝𝑧𝑧 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑥𝑥 � � 𝑞𝑞𝑦𝑦 �
𝑦𝑦 𝑦𝑦
𝑟𝑟𝑧𝑧 𝑝𝑝𝑧𝑧 −𝑝𝑝𝑦𝑦 𝑝𝑝𝑥𝑥 𝑝𝑝𝑤𝑤 𝑞𝑞𝑧𝑧

We can even express quaternions themselves as 4x4 matrices and have all of the normal notation
carry over:
𝑟𝑟𝑤𝑤 −𝑟𝑟𝑥𝑥 −𝑟𝑟𝑦𝑦 −𝑟𝑟𝑧𝑧 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑥𝑥 −𝑝𝑝𝑦𝑦 −𝑝𝑝𝑧𝑧 𝑞𝑞𝑤𝑤 −𝑞𝑞𝑥𝑥 −𝑞𝑞𝑦𝑦 −𝑞𝑞𝑧𝑧
𝑟𝑟𝑥𝑥 𝑟𝑟𝑤𝑤 −𝑟𝑟𝑧𝑧 𝑟𝑟𝑦𝑦 𝑝𝑝𝑥𝑥 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑧𝑧 𝑝𝑝𝑦𝑦 𝑞𝑞𝑥𝑥 𝑞𝑞𝑤𝑤 −𝑞𝑞𝑧𝑧 𝑞𝑞𝑦𝑦
� 𝑟𝑟 𝑟𝑟𝑧𝑧 𝑟𝑟𝑤𝑤 −𝑟𝑟𝑥𝑥 � = � 𝑝𝑝𝑦𝑦 𝑝𝑝𝑧𝑧 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑥𝑥 � � 𝑞𝑞𝑦𝑦 𝑞𝑞𝑧𝑧 𝑞𝑞𝑤𝑤 −𝑞𝑞𝑥𝑥 �
𝑦𝑦
𝑟𝑟𝑧𝑧 −𝑟𝑟𝑦𝑦 𝑟𝑟𝑥𝑥 𝑟𝑟𝑤𝑤 𝑝𝑝𝑧𝑧 −𝑝𝑝𝑦𝑦 𝑝𝑝𝑥𝑥 𝑝𝑝𝑤𝑤 𝑞𝑞𝑧𝑧 −𝑞𝑞𝑦𝑦 𝑞𝑞𝑥𝑥 𝑞𝑞𝑤𝑤

(We’ll show how to derive this in section 5.)

Alternatively, here’s an easy way to remember the product of two quaternions: Imagine we
extend the real numbers with three symbols i, j, and k (in the same way that we can extend the real line
to get the complex numbers) with the properties that

𝑖𝑖 2 = −1, 𝑗𝑗 2 = −1, 𝑘𝑘 2 = −1, and 𝑖𝑖𝑖𝑖𝑖𝑖 = −1.

1
There are such things as non-unit quaternions, which we won’t talk about in this article outside of this footnote.
They can be thought of as combinations of a rotation and a scale by the length of the quaternion, the square root
of the norm given by 𝑁𝑁(𝒒𝒒) = |𝑞𝑞|2 = 𝑞𝑞𝑤𝑤 2
+ 𝑞𝑞𝑥𝑥2 + 𝑞𝑞𝑦𝑦2 + 𝑞𝑞𝑧𝑧2 . The unit quaternions are those quaternions with norm 1:
2
𝑞𝑞𝑤𝑤 + 𝑞𝑞𝑥𝑥2 + 𝑞𝑞𝑦𝑦2 + 𝑞𝑞𝑧𝑧2 = 1. For the inverse of a general quaternion, divide the coefficients above by 𝑁𝑁(𝒒𝒒).
This norm happens to satisfy 𝑁𝑁(𝒑𝒑𝒑𝒑) = 𝑁𝑁(𝒑𝒑)𝑁𝑁(𝒒𝒒) for any two quaternions p and q, which gives a quick
way to derive Euler’s four-square identity. As it turns out, the requirement for such a norm to exist is the main
reason why normed division algebras over the reals are only possible in dimensions 1, 2, 4, and 8.
 From these properties, we can derive the product of any two basis elements: 𝑖𝑖𝑖𝑖 = 𝑘𝑘, 𝑗𝑗𝑗𝑗 =
−𝑘𝑘, 𝑗𝑗𝑗𝑗 = 𝑖𝑖, 𝑘𝑘𝑘𝑘 = −𝑖𝑖, 𝑘𝑘𝑘𝑘 = 𝑗𝑗, and 𝑘𝑘𝑘𝑘 = 𝑗𝑗. We can then write the quaternion 𝒒𝒒 = �𝑞𝑞𝑤𝑤 , 𝑞𝑞𝑥𝑥 , 𝑞𝑞𝑦𝑦 , 𝑞𝑞𝑧𝑧 � as

𝑞𝑞𝑤𝑤 + 𝑞𝑞𝑥𝑥 𝑖𝑖 + 𝑞𝑞𝑦𝑦 𝑗𝑗 + 𝑞𝑞𝑧𝑧 𝑘𝑘

and have all the normal multiplication work:

�𝑝𝑝𝑤𝑤 + 𝑝𝑝𝑥𝑥 𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑘𝑘��𝑞𝑞𝑤𝑤 + 𝑞𝑞𝑥𝑥 𝑖𝑖 + 𝑞𝑞𝑦𝑦 𝑗𝑗 + 𝑞𝑞𝑧𝑧 𝑘𝑘�

= �𝑝𝑝𝑤𝑤 𝑞𝑞𝑤𝑤 − 𝑝𝑝𝑥𝑥 𝑞𝑞𝑥𝑥 − 𝑝𝑝𝑦𝑦 𝑞𝑞𝑦𝑦 − 𝑝𝑝𝑧𝑧 𝑞𝑞𝑧𝑧 � + (𝑝𝑝𝑤𝑤 𝑞𝑞𝑥𝑥 + 𝑝𝑝𝑥𝑥 𝑞𝑞𝑤𝑤 )𝑖𝑖 + �𝑝𝑝𝑤𝑤 𝑞𝑞𝑦𝑦 + 𝑝𝑝𝑦𝑦 𝑞𝑞𝑤𝑤 �𝑗𝑗 + (𝑝𝑝𝑤𝑤 𝑞𝑞𝑧𝑧 + 𝑝𝑝𝑧𝑧 𝑞𝑞𝑤𝑤 )𝑘𝑘 + 𝑝𝑝𝑥𝑥 𝑞𝑞𝑦𝑦 𝑖𝑖𝑖𝑖
+ 𝑝𝑝𝑥𝑥 𝑞𝑞𝑧𝑧 𝑖𝑖𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑞𝑞𝑥𝑥 𝑗𝑗𝑗𝑗 + 𝑝𝑝𝑦𝑦 𝑞𝑞𝑧𝑧 𝑗𝑗𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑞𝑞𝑥𝑥 𝑘𝑘𝑘𝑘 + 𝑝𝑝𝑧𝑧 𝑞𝑞𝑦𝑦 𝑘𝑘𝑘𝑘
= �𝑝𝑝𝑤𝑤 𝑞𝑞𝑤𝑤 − 𝑝𝑝𝑥𝑥 𝑞𝑞𝑥𝑥 − 𝑝𝑝𝑦𝑦 𝑞𝑞𝑦𝑦 − 𝑝𝑝𝑧𝑧 𝑞𝑞𝑧𝑧 � + �𝑝𝑝𝑥𝑥 𝑞𝑞𝑤𝑤 + 𝑝𝑝𝑤𝑤 𝑞𝑞𝑥𝑥 − 𝑝𝑝𝑧𝑧 𝑞𝑞𝑦𝑦 + 𝑝𝑝𝑦𝑦 𝑞𝑞𝑧𝑧 �𝑖𝑖
+ �𝑝𝑝𝑦𝑦 𝑞𝑞𝑤𝑤 + 𝑝𝑝𝑧𝑧 𝑞𝑞𝑥𝑥 + 𝑝𝑝𝑤𝑤 𝑞𝑞𝑦𝑦 − 𝑝𝑝𝑥𝑥 𝑞𝑞𝑧𝑧 �𝑗𝑗 + �𝑝𝑝𝑧𝑧 𝑞𝑞𝑤𝑤 − 𝑝𝑝𝑦𝑦 𝑞𝑞𝑥𝑥 + 𝑝𝑝𝑥𝑥 𝑞𝑞𝑦𝑦 + 𝑝𝑝𝑤𝑤 𝑞𝑞𝑧𝑧 �𝑘𝑘.

With the equations above, quaternions give us a fast and efficient way to store rotations,
express the composition of rotations, and, importantly, to smoothly blend between rotations (which
we’ll cover in section 5.)
𝑇𝑇
Now, here’s something interesting: Consider the rotation around an axis �𝑛𝑛𝑥𝑥 , 𝑛𝑛𝑦𝑦 , 𝑛𝑛𝑧𝑧 � by an
angle 𝜃𝜃, represented by a quaternion:

𝜃𝜃 𝜃𝜃 𝜃𝜃 𝜃𝜃 𝑇𝑇
𝒒𝒒 = �cos � � , 𝑛𝑛𝑥𝑥 sin � � , 𝑛𝑛𝑦𝑦 sin � � , 𝑛𝑛𝑧𝑧 sin � �� .
2 2 2 2
If we rotate around this axis by an additional 360°, we get

𝜃𝜃 𝜃𝜃 𝜃𝜃 𝜃𝜃 𝑇𝑇
𝒒𝒒′ = �−cos � � , −𝑛𝑛𝑥𝑥 sin � � , −𝑛𝑛𝑦𝑦 sin � � , −𝑛𝑛𝑧𝑧 sin � ��
2 2 2 2
This is a different quaternion, but it represents the same rotation; we’ve just rotated an extra 360°.

Rotating an additional 360°, for a total of 720°, brings us back to the first quaternion.

In fact, we see the following: Every rotation – an angle around some axis – is represented not by
one, but by two quaternions. In this way, we say that the unit quaternions double-cover the space of
rotations. But why do quaternions have to double-cover rotations? Could we – say – just do something
𝜃𝜃
like replacing by 𝜃𝜃? 2
2

2
Answer: No, because then all 180° rotations would be represented by the same four-vector, (−1, 0, 0, 0)𝑇𝑇 .
 This article is about this phenomenon of double-covering. In short: When we’re talking about
interpolating between paths in the space of rotations, it actually matters how many times our rotation
has completed a full circle.

In a wider mathematical frame, it turns out that the space of unit quaternions is actually slightly
nicer than the space of rotations: the space of quaternions maps nicely to a sphere in four dimensions,
while the space of rotations isn’t simply connected. Put another way, the unit quaternions cover the
space of rotations twice, because they cannot cover the space of rotations once and also provide a way
to interpolate between paths of rotations. By working with quaternions, we get to work with points on a
sphere, instead of points on a real projective plane.

2. An Orthogonal Basis Problem

Here’s a system of representing rotations that doesn’t work.

Suppose we want to rotate a point around an axis. One way to do this might be to extend the axis
into a full coordinate frame, by somehow finding two additional vectors which meet at right angles to
the axis and to each other.

Once we have such a coordinate frame 3, we can map our point to our coordinate system, rotate
around the axis using a two-dimensional rotation in the plane of our other two vectors, and then
transform back into the original coordinate system. 4

| | | 1 0 0 | | | −1
𝑅𝑅 = �𝒏𝒏 𝒃𝒃 𝒕𝒕� �0 cos 𝜃𝜃 − sin 𝜃𝜃 � �𝒏𝒏 𝒃𝒃 𝒕𝒕�
| | | 0 sin 𝜃𝜃 cos 𝜃𝜃 | | |
Ideally, we’d also like our two additional vectors to vary smoothly as we adjust our axis. Put
another way, we’d like to find some continuous function which takes as input a normalized vector and

3
Technical note: With a fixed handedness.
4
If n, b, and t are normalized, then this is slightly easier; since the basis transform matrix is orthogonal, we have
| | | −1 | | | 𝑇𝑇
�𝒏𝒏 𝒃𝒃 𝒕𝒕� = �𝒏𝒏 𝒃𝒃 𝒕𝒕� .
| | | | | |
outputs the rest of the coordinate system. (The reason we want this function to be continuous is
because we’d like to be able to nicely interpolate between rotations; otherwise, although the results
might be OK, our internal model of the system would suddenly change its state as we passed over the
discontinuity.)

Unfortunately, this is impossible: no matter how we try to construct such a function, we’ll
always have a discontinuity somewhere in the space of unit vectors. The easiest way to see this is
through the Hairy Ball Theorem: if we had such a function, then we’d be able to place a coordinate
frame at each point of a sphere, like this (excepting the north and south poles in this illustration):

If we could do such a thing, then the green or blue vectors would form a continuous vector field
on the surface of the sphere. But this would produce a smooth combing of the sphere, which is
impossible by the Hairy Ball Theorem.

3. Euler Angles

If you’ve heard of Euler angles before, these singular points that arise when trying to comb the
sphere might remind you of gimbal lock, a problem that arises in using this system to represent
rotations – where at particular choices of axis, you lose a degree of freedom and the entire system
freezes up until you rotate it out or through of this particular range of axes. If you haven’t heard of Euler
angles before, let’s rewind a bit.

Euler angles give us a way to represent orientations as a unique product of a yaw, a pitch, and a roll,
as follows:

Suppose we have an object which we want to rotate to a particular orientation, which we express
(as with quaternions) as an axis and an angle around that axis.
 ?

First, we match the lateral bearing (yaw) of the object:

Then, we pitch the object up or down to match the new axis:

Finally, we rotate (roll) the object around this axis to match the full orientation.

If we limit the yaw, pitch, and roll to some range of values (e.g. 0 ≤ 𝑦𝑦𝑦𝑦𝑦𝑦, 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 < 360° and −90° <
𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝ℎ < 90°), these three numbers can then be used to uniquely represent any orientation outside of
gimbal lock. (We can also uniquely specify a rotation instead of an orientation by listing the yaw, pitch,
and roll that a model undergoes as a result of that rotation.)
 However, we have a problem: Regardless of our choice of Euler angle system 5, we’ll always be able
to find some axis near which slightly different orientations lead to wildly different Euler angles. For
instance, consider these two orientations with regards to the above system:

In the above system, the first of these two orientations can be given by a simple 90° pitch upwards.
For the second, we need to turn the object 180°, pitch it upwards just less than 90°, and then roll it
another 180°.

In particular, this means that our mapping in reverse from orientations to Euler angles is
discontinuous – and that as a result, in order to interpolate between two orientations using Euler angles,
we might have to take a longer route than necessary:

Gimbal lock has been the cause of a variety of problems in real-world systems; for more information on
the ill effects of gimbal lock, see [Hanson, pg. 19-27].

4. The Connectedness of Rotations

Euler angles, in fact, have problems for reasons beyond the Hairy Ball Theorem – their geometry
is that of a four-dimensional torus (which doesn’t correspond to the geometry of the space of rotations),
and the fact that they have three parameters also prevents them from smoothly representing the space
of rotations (if we consider the sets of points we get as we vary the angle of rotation from 0° to 360°, we

5
We’ve described the ZYX system of Euler angles above, where we rotate around the Z, Y, and X axes in our local
coordinate frame in sequence. (For instance, at the second step, we rotated around the object’s local Y axis.) We
can also describe Euler angles as a triplet of rotations around each of three axes, at least one different from the
rest. We’ll always have gimbal lock somewhere, regardless of the Euler angle system we choose; however, in some
cases (e.g. ships), we might be able to choose a frame such that we should never rotate to an orientation that
would cause gimbal lock.
can see that the sets of spheres we get must at some point “turn back” on themselves, which prevents
us from having a nice mapping from a region in ℝ3 to the space of rotations.)

Even if we uniquely parameterized rotations as a pair of an axis and an angle (with some
constraints, so that we express each rotation exactly once), we’d still run into problems when talking
about interpolation – specifically, when talking about interpolation between paths of rotations.

Consider a space, and draw two continuous paths through the space which begin and end at the
same two points. We say that this space is simply connected if, no matter which two points or paths we
choose, we can always continuously transform the first path into the second. For instance, the sphere is
simply connected, while the torus is not simply connected.

As it turns out, the space of rotations isn’t simply connected – unless we allow ourselves to
represent each rotation twice, in which case we get the quaternions. To see this, consider the cycle
given by starting with a 180-degree rotation around the vertical axis, and then continuously turning the
axis until it points downward.

Since a 180-degree clockwise rotation is the same as a 180-degree counterclockwise rotation,

our path starts and ends at the same rotation. If the space of rotations were simply connected, we
would be able to smoothly adjust and contract this loop until we get a single point. However, no matter
how we adjust this path of rotations, our axis of rotation must at some point be horizontal. Therefore,
we cannot turn this path into a single point, and so our space is not simply connected.

The unit quaternions get around this by representing each rotation in two ways. As a result, we
need to rotate our axis by a full 360° in the space of quaternions to get back to the same point (in the
space of unit quaternions) we started with. Additionally, we can easily map the unit quaternions to the
2
surface of a four-dimensional sphere and back (since the unit quaternions satisfy 𝑞𝑞𝑤𝑤 + 𝑞𝑞𝑥𝑥2 + 𝑞𝑞𝑦𝑦2 + 𝑞𝑞𝑧𝑧2 =
1). Since the four-dimensional sphere, like the three-dimensional sphere 6 is simply connected, the unit
quaternions themselves are simply connected, which is nice.

6
Unlike the two-dimensional sphere (the edge of a circle), which essentially doesn’t have enough dimensions to
transform paths.
 5. Working with Quaternions: 9 Recipes and Tricks You Might Not Have Heard
About
𝜃𝜃 𝜃𝜃 𝜃𝜃 𝜃𝜃 𝑇𝑇
Rotating points: We’ve talked about how the unit quaternion �cos , 𝑛𝑛𝑥𝑥 sin , 𝑛𝑛𝑦𝑦 sin , 𝑛𝑛𝑧𝑧 sin �
2 2 2 2
represents a rotation, and how to find the result of performing two rotations in sequence, but we
haven’t actually talked about how to rotate a point around an axis.

𝜃𝜃 𝜃𝜃 𝜃𝜃 𝜃𝜃 𝑇𝑇
Let’s say we have a point (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 and a quaternion 𝒒𝒒 = �cos , 𝑛𝑛𝑥𝑥 sin , 𝑛𝑛𝑦𝑦 sin , 𝑛𝑛𝑧𝑧 sin �
2 2 2 2
representing the angle and axis around which we want to rotate. We can actually think of the point 𝒙𝒙 =
(𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 not as a point, but as a quaternion (0, 𝑥𝑥, 𝑦𝑦, 𝑧𝑧), which represents a 180° counterclockwise
rotation around the axis (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 .

Now, consider the following sequence of rotations:

𝑇𝑇
• Rotate by -𝜃𝜃 around �𝑛𝑛𝑥𝑥 , 𝑛𝑛𝑦𝑦 , 𝑛𝑛𝑧𝑧 � .
• Rotate by 180° around x.
𝑇𝑇
• Rotate by 𝜃𝜃 around �𝑛𝑛𝑥𝑥 , 𝑛𝑛𝑦𝑦 , 𝑛𝑛𝑧𝑧 � .

We’d like to figure out what rotation this corresponds to. If we apply this rotation to our desired
result – the rotation of 𝒙𝒙 by q, which we’ll call x’ – we wind up transforming x’ to x, performing a
rotation which leaves 𝒙𝒙 in place, and finally transforming x back to x’.

On the other hand, if we started with a vector perpendicular to x’ (let’s call it w’), we’d
transform w’ to a vector w perpendicular to 𝒙𝒙, negate it, and rotate it back, ultimately, to -w. Therefore,
our sequence of rotations is the same as a 180° rotation around x’.

That is, if we express points in ℝ3 as corresponding 180° unit quaternions, then the quaternion

𝒙𝒙’ = 𝒒𝒒𝒒𝒒𝒒𝒒−𝟏𝟏
is the result of rotating the point x by q.

Rotating between two points: Suppose we have two points a and b, and we want to find some rotation
which rotates the point a to the point b. Interpreting points as 180° rotations as before, consider the
quaternion

𝒃𝒃𝒂𝒂−𝟏𝟏 .
If we apply this rotation to (the point) a, we actually wind up on the far side of b – twice as far as we
meant to go!

𝒂𝒂−1 𝒃𝒃

Intuitively (we can formalize this without much difficulty), the solution is to rotate half as far:
the quaternion

�𝒃𝒃𝒂𝒂−𝟏𝟏

then gives a smooth and direct rotation from a to b.

We can think of the square root in much of the same way we thought of powers of quaternions
at the start: the square root of an axis-angle unit quaternion corresponds to dividing its angle by 2,
which we can compute either by expressing the quaternion in axis-angle form or by using the half-angle
formulas for sin and cos. (We also have a sign problem from the square root; in this case, we always
want to choose the quaternion with a nonnegative w component.)

Rotations are orientations. If we have some object, we can fix some initial orientation for the object,
and then describe its orientation by writing a rotation which transforms the object’s initial orientation
into its current orientation. Conversely, if we’ve fixed some initial orientation for the object, we can
uniquely describe a rotation by giving the orientation of the object at the end of the rotation.

Interpolating between quaternions: Quaternions can be directly embedded in four dimensions as the
set of points on the unit four-dimensional sphere. As a result, we can measure the distance between
quaternions by the distance in four-space, and we can interpolate between quaternions by interpolating
between points on the sphere.

In two dimensions, we can interpolate between points a and b on the unit circle (using a
parameter 𝑡𝑡 that ranges from 0 to 1) using the formula

sin�(1 − 𝑡𝑡)𝜃𝜃� sin(𝑡𝑡𝑡𝑡)

𝒑𝒑(𝑡𝑡) = 𝒂𝒂 + 𝒃𝒃
sin(𝜃𝜃) sin(𝜃𝜃)
where 𝜃𝜃 is the angle between a and b. It’s easiest to see this geometrically – for the component of b in
the above expression, for instance:
 Since we can always think of a plane containing 0, a, and b, the same formula also works in
higher dimensions. (This is more commonly known as the slerp formula for spherically interpolating
between two points.)

We have to be careful, though! Although this formula always gives the shortest way to
interpolate between two unit quaternions, since b and -b represent the same rotation, we might have
parity problems using this formula to interpolate between two rotations if we’re not careful. In
particular, the double-covering property of the unit quaternions combined with the above formula
means that we have two ways to rotate between two rotations – and if we don’t check in advance, we
can wind up traversing the longer of the two ways. The solution when using quaternions to represent
rotations is to choose whether to interpolate between a and b or between a and -b, depending on which
of the two pairs are closer together.

Visualizing quaternions on the sphere in four dimensions: We can lay out the surface of a four-
dimensional sphere by separating it into a solid unit ball, a hollow spherical ‘equator’, and a second solid
unit ball. The point (𝑤𝑤, 𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 on the four-dimensional sphere maps to (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 in the left ball if 𝑤𝑤 <
0, (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 on the sphere if 𝑤𝑤 = 0, and (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 on the right ball if 𝑤𝑤 > 0.

We can also draw paths between quaternions using this approach. For instance, if we were to
perform a full 360-degree rotation around (0, 0, 1)𝑇𝑇 in the space of quaternions, we’d start at the center
on the right (at (1, 0, 0, 0)𝑇𝑇 ), go upwards through the right ball, pass through the sphere at (0, 0, 0, 1)𝑇𝑇 ,
then go downwards through the left ball, finishing in the center of the left ball at (−1, 0, 0, 0)𝑇𝑇 .
The belt trick. Here’s a nice way to show that the space of rotations isn’t simply connected, while the
space of unit quaternions is: Take a long strand of cloth (a belt will also do), fix one end of it to a
stationary object, and give the end one full twist. The goal is now to find a way to bend and manipulate
the middle of the cloth (possibly passing it over the end) while keeping the ends stationary so as to
remove the twist in the cloth.

For a full twist, this is impossible – we can turn a full clockwise twist into a full counterclockwise
twist, for instance, but we can’t untwist the fabric. Surprisingly, though, if we start out with two full
twists, we can untwist the fabric!

Here’s the real trick: Instead of thinking about the fabric as a surface, we can think of the fabric
as a series of orientations along a curve. These then form a path through the space of unit quaternions,
which we can visualize!

A better way, in fact, is to imagine a series of nested glass spheres around the free endpoint of
the fabric, each of which contains some slice of the fabric as it leads inwards. Then we can specify the
orientation and position of the fabric at any point by specifying the orientation of the corresponding
glass sphere.

For a full turn, we have a series of rotations around a single axis, which gives us roughly the
same path in the space of unit quaternions as before:
 Since we start and end at different unit quaternions, though, we can’t transform this path into a
point while keeping the endpoints intact, so we cannot untwist the fabric.

However, when we have two full twists, we have a full loop through the space of unit quaternions:

We can then untwist the fabric by transforming the path on the four-dimensional sphere to a
single point:

But we have one more trick. Suppose instead of one twisted piece of fabric, we have an entire
sphere of twisted pieces of fabric:
If we think of these pieces of fabric as being embedded in nested glass spheres as before, then no
matter how we rotate the glass spheres, the strands of fabric will never intersect. As a result, if we start
from an untwisted configuration, twist the spheres to create a double twist in one of the pieces of fabric
(in fact, in all of the pieces of fabric), and then reverse the result, we’ll be able to show dozens of double
twists being untwisted at once without a single intersection in the entire configuration at any point.

This is quite a sight when animated; for more visualizations of this trick, see Andrew Hanson’s
Belt Trick demonstration at https://www.cs.indiana.edu/~hansona/quatvis/Belt-Trick/index.html.

Deriving quaternion composition. If we know that rotations are linear transformations, that an arbitrary
rotation can be expressed as a product of rotations about the x and y axes, and that our resulting
structure will require a 720-degree rotation to be returned to its initial state, we can sort of rederive the
rules for composing quaternions (and in particular, the matrix at the beginning of this paper) as follows:

Let’s denote our identity rotation by 1, a 180-degree rotation about the x axis by i, and a 180-
degree rotation about the y axis by j. The result of performing j followed by i is a 180-degree rotation
about a third axis, which we’ll denote by k. Since we treat a 360-degree rotation as a sort of alternate
form of the identity rotation 1, we have i2=j2=k2=-1. We already have ij=k; manipulating this expression
gives 𝑗𝑗𝑗𝑗 = 𝑖𝑖 and 𝑘𝑘𝑘𝑘 = 𝑗𝑗. Finally, we can see that 𝑗𝑗𝑗𝑗 = −𝑘𝑘, since 𝑗𝑗𝑗𝑗 = −(𝑗𝑗𝑗𝑗)(𝑘𝑘𝑘𝑘) = −𝑖𝑖𝑖𝑖 = −𝑘𝑘, and we can
similarly determine that 𝑘𝑘𝑘𝑘 = −𝑖𝑖 and 𝑖𝑖𝑖𝑖 = −𝑗𝑗.

Since rotations are linear, quaternion composition should be linear as well; therefore, we can
split the product pq into a linear sum of p*1, p*i, p*j, and p*k. We get

�𝑝𝑝𝑤𝑤 + 𝑝𝑝𝑥𝑥 𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑘𝑘�1 = 𝑝𝑝𝑤𝑤 + 𝑝𝑝𝑥𝑥 𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑘𝑘

�𝑝𝑝𝑤𝑤 + 𝑝𝑝𝑥𝑥 𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑘𝑘�𝑖𝑖 = −𝑝𝑝𝑥𝑥 + 𝑝𝑝𝑤𝑤 𝑖𝑖 + 𝑝𝑝𝑧𝑧 𝑗𝑗 − 𝑝𝑝𝑦𝑦 𝑘𝑘
�𝑝𝑝𝑥𝑥 𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑘𝑘 + 𝑝𝑝𝑤𝑤 �𝑗𝑗 = −𝑝𝑝𝑦𝑦 − 𝑝𝑝𝑧𝑧 𝑖𝑖 + 𝑝𝑝𝑤𝑤 𝑗𝑗 + 𝑝𝑝𝑥𝑥 𝑘𝑘
�𝑝𝑝𝑥𝑥 𝑖𝑖 + 𝑝𝑝𝑦𝑦 𝑗𝑗 + 𝑝𝑝𝑧𝑧 𝑘𝑘 + 𝑝𝑝𝑤𝑤 �𝑘𝑘 = −𝑝𝑝𝑧𝑧 + 𝑝𝑝𝑦𝑦 𝑖𝑖 − 𝑝𝑝𝑥𝑥 𝑗𝑗 + 𝑝𝑝𝑤𝑤 𝑘𝑘

so if 𝒓𝒓 = 𝒑𝒑𝒑𝒑, then
𝑟𝑟𝑤𝑤 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑥𝑥 −𝑝𝑝𝑦𝑦 −𝑝𝑝𝑧𝑧 𝑞𝑞𝑤𝑤
𝑟𝑟𝑥𝑥 𝑝𝑝𝑥𝑥 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑧𝑧 𝑝𝑝𝑦𝑦 𝑞𝑞𝑥𝑥
� 𝑟𝑟 � = � 𝑝𝑝 𝑝𝑝𝑧𝑧 𝑝𝑝𝑤𝑤 −𝑝𝑝𝑥𝑥 � � 𝑞𝑞𝑦𝑦 �.
𝑦𝑦 𝑦𝑦
𝑟𝑟𝑧𝑧 𝑝𝑝𝑧𝑧 −𝑝𝑝𝑦𝑦 𝑝𝑝𝑥𝑥 𝑝𝑝𝑤𝑤 𝑞𝑞𝑧𝑧

Almost combing a hairy sphere. Although we know it’s impossible to comb a hairy sphere, having some
way to construct an orthonormal basis in a mostly continuous way for the points of the sphere (ideally,
from the point alone) is useful in many situations. One such method comes from [Frisvad]; basically, we
can start with an orthonormal basis at the top of the sphere, and then rotate this downwards along each
line of longitude, covering the entire sphere except for the south pole.

orthonormal bases tangent vectors only

We can rotate the coordinate frame using a variety of methods; one would be to determine the
rotation needed using the formula for rotating between two points above, and then to compute the
result of transforming (1,0,0)𝑇𝑇 and (0,1,0)𝑇𝑇 by this rotation. If (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 is a unit vector with 𝑧𝑧 ≠ −1,
we then have that

𝑥𝑥 2 𝑥𝑥𝑥𝑥
𝒃𝒃 = (1 − ,− , −𝑥𝑥)
1 + 𝑧𝑧 1 + 𝑧𝑧
and

𝑥𝑥𝑥𝑥 𝑦𝑦 2
𝒕𝒕 = (− ,1 − , −𝑦𝑦)
1 + 𝑧𝑧 1 + 𝑧𝑧
are perpendicular to (𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 , to each other, and form the rest of a coordinate frame containing
(𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 as one of its axes. When z=-1, we just output the coordinate frame
(0,0, −1)𝑇𝑇 , (0,1,0)𝑇𝑇 , (1,0,0)𝑇𝑇 .

Unfortunately, as a result of the singularity in this method, we wind up having numerical issues
when z is close to -1. There are at least two approaches to solve this problem, both of which work just
fine in practice. [Max] finds the optimal cutoff point for determining when we’re at the south pole,
which works pretty well. [Duff et al.] and [Reynolds] extend the original approach in a nice way: instead
of propagating a coordinate frame downwards from the north pole, we can also propagate a coordinate
frame upwards from the south pole and meet at the equator! Although we now have a discontinuity
along the entire equator, we can define an appropriate coordinate frame at every point of the sphere
and arrange the frames so that the frames on the southern hemisphere are just a flipped version of the
frames on the northern hemisphere.

Converting to matrices and back. Even if your internal rotation representation is based on unit
quaternions, rotation matrices can come in quite handy, whether you need to express a rotation in a
matrix format which a computer renderer can easily read, or to constrain your rotation to satisfy some
set of constraints.
Since rotations are linear transformations (in particular, the action of a quaternion 𝒒𝒒 on ℝ3 by 𝒙𝒙 →
𝒒𝒒𝒒𝒒𝒒𝒒−𝟏𝟏 is linear), we can start out by computing 𝒒𝒒𝑖𝑖𝒒𝒒−1 , 𝒒𝒒𝑗𝑗𝒒𝒒−1, and 𝒒𝒒𝑘𝑘𝒒𝒒−1 :
𝑇𝑇
𝒒𝒒𝑖𝑖𝒒𝒒−1 = �𝑞𝑞𝑤𝑤
2
+ 𝑞𝑞𝑥𝑥2 − 𝑞𝑞𝑦𝑦2 − 𝑞𝑞𝑧𝑧2 , 2�𝑞𝑞𝑥𝑥 𝑞𝑞𝑦𝑦 + 𝑞𝑞𝑤𝑤 𝑞𝑞𝑧𝑧 �, 2�𝑞𝑞𝑥𝑥 𝑞𝑞𝑧𝑧 − 𝑞𝑞𝑤𝑤 𝑞𝑞𝑦𝑦 ��
𝑇𝑇
𝒒𝒒𝑗𝑗𝒒𝒒−1 = �2�𝑞𝑞𝑥𝑥 𝑞𝑞𝑦𝑦 − 𝑞𝑞𝑤𝑤 𝑞𝑞𝑧𝑧 �, 𝑞𝑞𝑤𝑤
2
− 𝑞𝑞𝑥𝑥2 + 𝑞𝑞𝑦𝑦2 − 𝑞𝑞𝑧𝑧2 , 2�𝑞𝑞𝑤𝑤 𝑞𝑞𝑥𝑥 + 𝑞𝑞𝑦𝑦 𝑞𝑞𝑧𝑧 ��
𝑇𝑇
𝒒𝒒𝑘𝑘𝒒𝒒−1 = �2�𝑞𝑞𝑤𝑤 𝑞𝑞𝑦𝑦 + 𝑞𝑞𝑥𝑥 𝑞𝑞𝑧𝑧 �, 2�𝑞𝑞𝑦𝑦 𝑞𝑞𝑧𝑧 − 𝑞𝑞𝑤𝑤 𝑞𝑞𝑥𝑥 �, 𝑞𝑞𝑤𝑤
2
− 𝑞𝑞𝑥𝑥2 − 𝑞𝑞𝑦𝑦2 + 𝑞𝑞𝑧𝑧2 �

Since 𝒒𝒒 is a unit quaternion, we then have that rotating a point by 𝒒𝒒 is equivalent to multiplying by the
3x3 matrix

1 − 2�𝑞𝑞𝑦𝑦2 + 𝑞𝑞𝑧𝑧2 � 2�𝑞𝑞𝑥𝑥 𝑞𝑞𝑦𝑦 − 𝑞𝑞𝑤𝑤 𝑞𝑞𝑧𝑧 � 2�𝑞𝑞𝑤𝑤 𝑞𝑞𝑦𝑦 + 𝑞𝑞𝑥𝑥 𝑞𝑞𝑧𝑧 �
�2�𝑞𝑞𝑥𝑥 𝑞𝑞𝑦𝑦 + 𝑞𝑞𝑤𝑤 𝑞𝑞𝑧𝑧 � 1 − 2(𝑞𝑞𝑥𝑥2 + 𝑞𝑞𝑧𝑧2 ) 2�𝑞𝑞𝑦𝑦 𝑞𝑞𝑧𝑧 − 𝑞𝑞𝑤𝑤 𝑞𝑞𝑥𝑥 ��.
2�𝑞𝑞𝑥𝑥 𝑞𝑞𝑧𝑧 − 𝑞𝑞𝑤𝑤 𝑞𝑞𝑦𝑦 � 2�𝑞𝑞𝑤𝑤 𝑞𝑞𝑥𝑥 + 𝑞𝑞𝑦𝑦 𝑞𝑞𝑧𝑧 � 1 − 2(𝑞𝑞𝑥𝑥2 + 𝑞𝑞𝑦𝑦2 )

Trying to convert back from a (numerically computed) rotation matrix to a quaternion usually results
in an overdetermined system; we have seven constraints and four parameters (with a choice of sign).
We can just look at the values from some subset of the matrix, for instance, or use a nonlinear least
squares method to try to find optimal values for 𝑞𝑞𝑤𝑤 , 𝑞𝑞𝑥𝑥 , 𝑞𝑞𝑦𝑦 , and 𝑞𝑞𝑧𝑧 .

6. This Paper, but in Higher Dimensions

In two dimensions, rotations have one degree of freedom; there are no axes to choose, and any
rotation is parameterized by its angle. In three dimensions, rotations are parameterized by an axis and
an angle, for a total of three degrees of freedom. In four dimensions and higher, we cannot rely upon
the axis-angle model; instead, we can decompose any rotation into a product of rotations in two-
dimensional subspaces.

If you need to compute with rotations in higher dimensions, it might make the most sense to work
with rotation matrices (which are the set of orthogonal matrices with determinant 1) directly. In general,
in n dimensions orthogonal matrices have 𝑛𝑛2 parameters and 𝑛𝑛(𝑛𝑛 + 1)/2 constraints, for a total of
𝑛𝑛(𝑛𝑛 − 1)/2 degrees of freedom.

In fact, since a single unit quaternion has three degrees of freedom and four-dimensional rotations
have six degrees of freedom, it turns out we can represent a rotation in four dimensions by a pair of
quaternions. In higher dimensions, we can talk about the octonions, an eight-dimensional algebraic
system with inverses and a norm similar to the quaternions, but without the associative property (that
is, (𝑎𝑎𝑎𝑎)𝑐𝑐 may no longer be equal to 𝑎𝑎(𝑏𝑏𝑏𝑏)). These can then be used to represent eight-dimensional
rotations, although things get complicated.

The Hairy Ball Theorem is true in all odd dimensions (where combing a sphere is well defined);
however, we can comb the sphere in any even dimension! One easy way to do so is the following: For
any vector (𝑥𝑥1 , 𝑥𝑥2 , … , 𝑥𝑥2𝑛𝑛−1 , 𝑥𝑥2𝑛𝑛 )𝑇𝑇 , the vector

(−𝑥𝑥2 , 𝑥𝑥1 , −𝑥𝑥4 , 𝑥𝑥3 , … , −𝑥𝑥2𝑛𝑛 , 𝑥𝑥2𝑛𝑛−1 )𝑇𝑇

is perpendicular to (𝑥𝑥1 , … , 𝑥𝑥2𝑛𝑛 )𝑇𝑇 . In fact, in four dimensions, we can do even better, and provide a full
coordinate frame: if (𝑤𝑤, 𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 is a unit four-vector, then the four vectors
(𝑤𝑤, 𝑥𝑥, 𝑦𝑦, 𝑧𝑧)𝑇𝑇 ,
(−𝑥𝑥, 𝑤𝑤, 𝑧𝑧, −𝑦𝑦)𝑇𝑇 ,
(−𝑦𝑦, −𝑧𝑧, 𝑤𝑤, 𝑥𝑥)𝑇𝑇 ,
(−𝑧𝑧, 𝑦𝑦, −𝑥𝑥, 𝑤𝑤)𝑇𝑇
are all of length 1 and are all orthogonal to each other. Surprisingly, these form exactly the matrix we
used to represent quaternions (and is in fact a reshuffled multiplication table on 1, I, j, and k). We can
also do this in two dimensions: the vectors (x, y) and (y, -x) have the same length and are perpendicular
to each other. Although this cannot be done in six dimensions, while the octonions give us a way to do
this in eight dimensions.

Perhaps the most surprising thing is that past eight dimensions, the space of normed division
algebras just stops; the real numbers, the complex numbers, the quaternions, and the octonions are the
only algebras over the reals which have a norm and allow division by nonzero numbers. This is Hurwitz’
theorem, a nice proof of which can be found in Conway and Smith’s On Quaternions and Octonions.

As it turns out, the space of rotations is never simply connected (for instance, in dimension 2,
the space of rotations is diffeomorphic to a disk, which has a hole in it), but in dimensions above 2 we
can always do something like we did with quaternions, double-covering the space of rotations in order
to get a space which is simply connected.

7. Sources and Further Reading

Quaternions in General:

John H. Conway and Derek Smith. 2003. On Quaternions and Octonions. A K Peters, Ltd., Natick,
MA, USA.

Andrew J. Hanson. 2006. Visualizing Quaternions. Morgan Kaufmann Publishers Inc., San
Francisco, CA, USA.

Eugene Salamin. 1979. Application of Quaternions to Computation with Rotations. Stanford AI

Lab, Stanford, CA, USA.

Finding Orthonormal Bases:

Tom Duff, James Burgess, Per Christensen, Christophe Hery, Andrew Kensler, Max Liani, and
Ryusuke Villemin. 2017. Building an Orthonormal Basis, Revisited. Journal of Computer Graphics
Techniques (JCGT) 6, 1 (March), 1-8.

J. R. Frisvad. 2012. Building an orthonormal basis from a 3D unit vector without normalization.
Journal of Graphics Tools 16, 3, 151-159.

Nelson Max. 2017. Improved Accuracy when Building an Orthonormal Basis. Journal of
Computer Graphics Techniques (JCGT), 6, 1 (March), 9-16.

Marc B. Reynolds. 2016. Orthonormal Basis from Normal via Quaternion Similarity. URL:
http://marc-b-reynolds.github.io/quaternions/2016/07/06/Orthonormal.html