arXiv:1411.6502v4 [math.

GM] 23 May 2016

Geometric algebras for euclidean geometry

Charles Gunn

Keywords. metric geometry, euclidean geometry, Cayley-Klein construc-

tion, dual exterior algebra, projective geometry, degenerate metric, pro-
jective geometric algebra, conformal geometric algebra, duality, homo-
geneous model, biquaternions, dual quaternions, kinematics, rigid body
motion.

Abstract. The discussion of how to apply geometric algebra to euclidean

n-space has been clouded by a number of conceptual misunderstandings
which we first identify and resolve, based on a thorough review of cru-
cial but largely forgotten themes from 19th century mathematics. We
then introduce the dual projectivized Clifford algebra P(R∗n,0,1 ) (eu-
clidean PGA) as the most promising homogeneous (1-up) candidate for
euclidean geometry. We compare euclidean PGA and the popular 2-up
model CGA (conformal geometric algebra), restricting attention to flat
geometric primitives, and show that on this domain they exhibit the
same formal feature set. We thereby establish that euclidean PGA is
the smallest structure-preserving euclidean GA. We compare the two
algebras in more detail, with respect to a number of practical criteria,
including implementation of kinematics and rigid body mechanics. We
then extend the comparison to include euclidean sphere primitives. We
conclude that euclidean PGA provides a natural transition, both scien-
tifically and pedagogically, between vector space models and the more
complex and powerful CGA.

1. Introduction
Although noneuclidean geometry of various sorts plays a fundamental role
in theoretical physics and cosmology, the overwhelming volume of practical
science and engineering takes place within classical euclidean space En . For
this reason it is of no small interest to establish the best computational model
for this space. In particular, this article explores the question, which form of
geometric algebra is best-suited for computing in euclidean space? In order to

This article has been published as [Gun16]. The final publication is available at
link.springer.com.
2 Charles Gunn

have a well-defined domain for comparison, we restrict ourselves for most of

the article to flat geometric primitives (points, lines, planes, and their higher-
dimensional analogs), along with kinematics and rigid body mechanics. We
refer to this domain as “the chosen task”. In Sect. 8 we widen the primitive
set to include spheres.
1.1. Overview
As the later results of the article are based upon crucial but often overlooked
19th century mathematics, the article begins with the latter. In Sect. 2 we
review the quaternions and the biquaternions, focusing on their use to model
euclidean kinematics and rigid body mechanics, a central part of the cho-
sen task. In Sect. 3 we trace the path from exterior algebra to geometric
algebra in the context of projective geometry, paying special attention to
the dual exterior algebra and the Cayley-Klein construction of metric spaces
within projective space. This culminates with the introduction of projective
geometric algebra (PGA), a homogeneous model for euclidean (and other
constant-curvature metric) geometry. In Sect. 4 we demonstrate that PGA is
superior to other proposed homogeneous models. Sect. 5 discusses and dis-
poses of a variety of misconceptions appearing in the geometric algebra liter-
ature regarding geometric algebras with degenerate metrics (such as PGA).
In Sect. 6 we identify a feature set for a euclidean geometric algebra for the
chosen task, and verify that both PGA and CGA fulfil all features. Sect. 7
turns to a comparison based on a variety of practical criteria, such as the
implementation of kinematics and rigid body mechanics. Sect. 8 extends the
comparison to include spheres, and concludes that the “roundness” of CGA
has both positive and negative aspects. Finally, Sect. 9 positions PGA as a
natural stepping stone, both scientifically and pedagogically, between vector
space geometric algebra and CGA.

2. Quaternions, Biquaternions, and Rigid Body Mechanics

The quaternions (Hamilton, 1844) and the biquaternions ([Cli73]) are impor-
tant forerunners of geometric algebra. They exhibit most of the important
features of a geometric algebra, such as an associative geometric product con-
sisting of a symmetric (“inner”) part and an anti-symmetric (“outer”) part,
and the ability to represent isometries as sandwich operators. However, they
lack the graded algebra structure possessed by geometric algebra.
They have a special importance in our context, since the even subalge-
bra of PGA (see below, Sect. 3.8.1) is, for n = 3 (the case of most practical
interest), isomorphic to the biquaternions. All the desirable features of the
biquaternions are then inherited by PGA. In particular, the biquaternions
contain a model of kinematics and rigid body mechanics, an important com-
ponent of the chosen task, that compares favorably with modern alternatives
(see Sect. 7.3). The biquaternions also reappear below in Sect. 5.1. Because
of this close connection with the themes of this article, and the absence of a
comparable treatment in the literature, we give a brief formulation of relevant
 Geometric algebras for euclidean geometry2 3

results here. We assume the reader has an introductory acquaintance with

quaternions and biquaternions.
2.1. Quaternions
We first show how the Euler top can be advantageously represented using
quaternions. A more detailed treatment is available in Ch. 1 of [Gun11a]. Let
UH represent the unit quaternions, and IH represent the purely imaginary
quaternions. Every unit quaternion can be written as the exponential of an
imaginary quaternion. A rotation R around the unit vector (b, c, d) through
an angle α can be represented by a “sandwich operator” as follows. Define
the imaginary quaternion m := bi + cj + dk which then gives rise to the unit
α
quaternion g := e 2 m by exponentiation. Let x := xi + yj + zk represent an
arbitrary vector (x, y, z) ∈ R3 . Then the rotation R applied to (x, y, z) is
given in the quaternion product as x0 = gxg−1 (= gxg).
2.1.1. Quaternion model of Euler top. Recall that the Euler top is a rigid
body constrained to move around its centre of gravity. We assume there are
no external forces. Let g(t), a path in UH, be the motion of the rigid body. Let
Mc ∈ IH and Vc ∈ IH represent the instantaneous momentum and velocity,
resp., in body coordinates. Considered as vectors in R3 , they are related by
the inertia tensor A via Mc = A(Vc ). Then the Euler equations of motion
can be written using the quaternion product:
ġ = gVc
1
Ṁc = (Vc Mc − Mc Vc )
2
Notice that this representation has practical advantages over the traditional
linear algebra approach using matrices: normalizing a quaternion brings it
directly onto the 3D solution space of the unit quaternions, while the matrix
group SO(3) is a 3-dimensional subspace of the 9-dimensional space of 3x3
matrices. Numerical integration proceeds much more efficiently in the former
case since, after normalising, there are no chances for “wandering off”, while
the latter has a 6D space of invalid directions that lead away from SO(3).
We meet the same problematic below in Sect. 7.2.
2.2. Biquaternions
The biquaternions, introduced in [Cli73], consist of two copies of the quater-
nions, with the eight units {1, i, j, k, , i, j, k} where  is a new unit com-
muting with everything and satisfying 2 = 0.1
When one removes the constraint on the Euler top that its center of
gravity remains fixed, one obtains the free top, which, as the name implies,
is free to move in space. Here the allowable isometry group expands to the
orientation-preserving euclidean motions E(3), a six-dimensional group which
1 The cases 2 = ±1 were also considered by Clifford and lead to noneuclidean geome-

tries (and their kinematics and rigid body mechanics), but lie outside the scope of this
article. Eduard Study also made significant contributions to this field, under the name of
dual quaternions, which for similar reasons remain outside the scope of this article.
4 Charles Gunn

is a semi-direct product of the rotation and translation subgroups. Just as

SO(3) can be represented faithfully by the unit quaternions, E(3) can be
faithfully represented by the unit biquaternions.

2.2.1. Imaginary biquaternions and lines. Analogous to the way imaginary

quaternions represent vectors in R3 , imaginary biquaternions represent ar-
bitrary lines in euclidean space. An imaginary biquaternion of the form
ai + bj + ck (a standard vector) represents a line through the origin; one
of the form (ai + bj + ck) (a dual vector) represents an ideal line (aka “line
at infinity”). An imaginary biquaternion whose standard and dual vectors
are perpendicular (as ordinary vectors in R3 ) corresponds to a line in RP 3 ;
the general imaginary biquaternion represents a linear line complex, the fun-
damental object of 3D kinematics and dynamics in this context.

2.2.2. Screw motions via biquaternion sandwiches. For example, a general

euclidean motion R is a screw motion with a unique invariant line, called its
axis. R rotates around this axis by an angle α while translating along the
axis through a distance d. The axis can be represented by a unit imaginary
biquaternion m, as indicated above; then v := α+d 2 m, a linear line com-
plex, is the infinitesimal generator of the screw motion and the exponential
α+d
e 2 m yields a unit biquaternion g. The sandwich operation gxg−1 gives
the action of the screw motion on an arbitrary line in space, represented by
an imaginary biquaternion x. One can also provide a representation for its
action on the points and planes of space: a point (x, y, z) is represented as
1+(xi+yj+zk) and a plane ax+by +cz +d = 0 maps to d+ai+bj+ck, but
then the sandwich operators have slight irregularities ([Bla42]). These irreg-
ularities are the necessary consequence of introducing ad hoc representations
for elements (points and planes) which do not naturally have a representa-
tion in the algebra. We return to this point in Sect. 5.1 since it has generated
some confusion related to our main theme.

2.2.3. Seamless integration via ideal elements. Note that the biquaternion
representation seamlessly handles cases which the traditional linear algebra
approach has to handle separately. For example, when the generating bivec-
tor m is an ideal line (a pure dual vector), then the isometry is a translation;
hence one sometimes says, “a translation is a rotation around a line at infin-
ity”; in dynamics, a force couple (resp., angular momentum) is a force (resp.,
momentum) carried by an ideal line. Hence, one obtains the Euler top from
the free top by constraining all momenta to be carried by ideal lines.

2.2.4. Euler equations of motion. The Euler equations of motion for the free
top are then given by the same pair of ODE’s given above, except that the
symbols are to be interpreted in the biquaternion rather than quaternion
context (where one uses unit (resp., imaginary) biquaternions instead of unit
(resp., imaginary) quaternions). These equations, as inherited by PGA, will
be shown below in Sect. 7.3 to compare favorably with modern alternatives
for kinematics and rigid body mechanics.
 Geometric algebras for euclidean geometry4 5

2.2.5. Historical notes. The biquaternions, as developed in [Stu03], were re-

formulated by von Mises ([vM24]) using tensor and matrix methods; he called
the result the motor algebra. The motor algebra has been developed fur-
ther and applied in robotics by modern researchers, for example, [BCDK00].
[Zie85] is an excellent historical monograph, recounting how the combined ef-
forts of eminent mathematicians including Möbius, Plücker, Klein, Clifford,
Study, and others, led to the discovery of the biquaternions as a model for
euclidean (and non-euclidean!) kinematics and rigid body mechanics.

3. From exterior algebra to geometric algebras

To understand how to go beyond biquaternions to obtain a geometric algebra
for euclidean geometry, we have need of other mathematical innovations of the
19th century: projective geometry and exterior algebra. We first review some
fundamental facts from projective geometry that are crucial to understanding
the following treatment. The remaining discussion in this section focuses on
two topics important to this exposition, not well-represented in the current
literature:
• the use of the dual exterior algebra to construct geometric algebras not
available otherwise, and
• the Cayley-Klein construction of metric spaces atop projective space,
particularly the delicate subject of degenerate signatures.
3.1. Preliminary remarks on vector space and projective space
First, recall that real projective space RP n can be derived from Rn+1 by in-
troducing the equivalence relationship x ≡ y ↔ ∃λ 6= 0 with x = λy, that is,
the points of RP n are the lines through the origin of Rn+1 ; this construction is
sometimes called the “projectivization” of the vector space, and plays a large
role in what follows. We can either interpret an n-vector as being a vector in
a vector space, or as representing a point in a projective space; the former we
refer to as the vector space setting; the latter, as the projective setting. Also
recall that to every real vector space V there is associated the dual vector
space V ∗ , consisting of the linear functionals V → R. The dual space in turn
can be (and, in the following, is) canonically identified with the hyperplanes
of V by associating f ∈ V ∗ with its kernel Kf := {x ∈ V | f (x) = 0}, a
hyperplane. λf for λ 6= 0 is associated to the weighted hyperplane λKf .
3.2. Projectivized exterior algebra
Begin with the standard real Grassmann or exterior algebra Rn+1 which
V
encapsulates the subspace structure of the real vector space Rn+1 (without
inner product). It is a graded associative algebra. The 1-vectors represent the
vectors of Rn+1 . The higher grades are constructed via the wedge product,
an anti-symmetric, associative product which is additive on the grade of
its operands, and represents the join operator on subspaces. V Projectivize
this algebra to obtain the projectivized exterior algebra P( Rn+1 ) which
in a natural way represents the subspace structure of real projective n-space
6 Charles Gunn

RP n , as built up out of points by joining them to form higher-dimensional

subspaces.3
3.3. Dual projectivized exterior algebra
The above process can also be carried out with the dual
V vector space (Rn+1 )∗
n+1 ∗
to produce the dual projectivized exterior algebra P( (R ) ). It also mod-
els the subspace structure of RP n , but dually, so that the 1-vectors repre-
sent hyperplanes (by the canonical identity mentioned above between the
dual space (Rn+1 )∗ and the hyperplanes of Rn+1 ). The wedge product corre-
sponds to the meet or intersection of subspaces. To avoid confusion we write
theVwedge operator in P( (Rn+1 )∗ ) as ∧ (meet) and the wedge operator in
V
P( Rn+1 ) as ∨ (join).
3.4. Duality
For practical applications, it is necessary to be able to carry out both meet
and join inVa given exterior algebra. For example, consider the meet oper-
ator in P( Rn+1 ). This is often, in the context of a standard Grassmann
algebra,
V n+1 called the regressive product.5 To implement the meet operator in
P( R ), we take advantage of Poincaré duality ([Gre67b], Sec. 6.8). The
basic idea is this: the exterior algebra and its dual provide two views on the
same projective space. Any geometric entity in RP n appears once in each
algebra. The PoincaréV isomorphism V is n+1
then a grade-reversing vector-space
isomorphism J : P( R n+1
) ↔ P( (R )∗ ) that V
maps a geometric entity
) to the same geometric entity in P( (Rn+1 )∗ ). In this sense
V n+1
of P( R
it is an identity map; sometimes called theVdual coordinate map. Equipped
with J we define a meet operation ∧ in P( Rn+1 ) by
X ∧ Y := J(J(X) ∧ J(Y))
and similarly, a join operator ∨ for P( (Rn+1 )∗ ). An alternative non-metric
V
method for calculating the join operator is provided by the shuffle product,
see [Sel05], Ch. 10.
3.5. Cayley-Klein construction of metric spaces
Recall that the Sylvester Inertia Theorem asserts that a symmetric bilinear
form Q can be characterised by an integer triple (p, n, z), its signature, de-
scribing the number of basis elements ei such that Q(ei , ei ) = {1, −1, 0},
resp. If one attaches such a Q to RP n then, for many choices of Q, Cay-
ley and Klein showed it is possible to define a distance function on a subset
M ⊂ RP n which makes M into a constant-curvature metric space ([Kle26],
Ch. 6, [Gun11b], §3.1, or [Gun11a], Ch. 4). When z = 0, the construction is

3 One can also begin with begin with projective space and construct the Grassmann

algebra in the obvious way; one obtains the same algebra P( Rn+1 ).
V
5 In a dual Grassmann algebra, the regressive product is the join operation. In general,

it’s the “other” subspace operation not implemented by the wedge product of the algebra.
 Geometric algebras for euclidean geometry9 7

based on the projective invariance of the cross ratio of four collinear points.
The distance between two points A and B is defined as
d(P, Q) := k ln(CR(A, B; F1 , F2 ))
where k is an appropriate real (or complex) constant, and F1 and F2 are
the two intersection points (real or imaginary) of their joining line with the
absolute quadric (the null vectors of Q). Since the cross ratio is a multiplica-
tive function, d is additive, and satisfies the other properties of a distance
function.
For example the signature (n + 1, 0, 0) leads to elliptic space Elln , while
(n, 1, 0) leads to hyperbolic space Hn . Note: in the vector space setting, the
signature (n + 1, 0, 0) produces the euclidean metric; in the projective setting
(via the Cayley-Klein construction), however, the same signature produces
the elliptic (or spherical 6 ) metric. This ambiguity has led to misunderstand-
ings in the literature which we discuss below in Sect. 4.1.
3.5.1. Cayley-Klein construction of euclidean space. The signature for eu-
clidean geometry is degenerate, that is, z 6= 0. Since this is a crucial point,
we motivate the correct choice using the example of the euclidean plane E2 .
Consider two lines m1 : a1 x + b1 y + c1 = 0 and m2 : a2 x + b2 y + c2 = 0.
Assuming WLOG a2i + b2i = 1, then cos α = a1 a2 + b1 b2 , where α is the angle
between the two lines: changing the c coefficient translates the line but does
not change the angle it makes to other lines. This generalizes to the angle
between two hyperplanes in En . One coordinate plays no role in the angle
calculation, hence the signature has z = 1. Thus, the Cayley-Klein construc-
tion applies the signature (n, 0, 1) to dual projective space to obtain a model
for euclidean geometry in n dimensions.
3.5.2. Euclidean distance between points. The discussion above takes its
starting point so that the angle between lines (hyperplanes) can be calcu-
lated. What about the distance between points? Already Klein ([Kle26], Ch.
4 §3) provided an answer to this question (or in English see [Gun11a], §4.3.1).
The inner product given above on lines (hyperplanes) induces the (very de-
generate) signature (1, 0, n) on points, so that one cannot measure the dis-
tance between points via the inner product. However, using a sequence of
non-euclidean signatures that converge to (n, 0, 1) ([Gun11a], §3.2), one can
show that the distance function between normalized homogeneous points P
and Q converges, up to an arbitrary positive constant7 , to the familiar eu-
clidean formula kP − Qk, the length of the euclidean vector P − Q (here we
apply euclidean in the vector space setting).

6 Two copies of elliptic space Elln can be glued together to obtain spherical space Sn .
7 This positive constant determines the length scale of euclidean space, such as feet or
meters.
8 Charles Gunn

3.6. Geometric algebras from Cayley-Klein

It is straightforward to obtain geometric algebras from the Cayley-Klein con-
struction.8 We use the signature of the Cayley-Klein construction to define
the inner product for our geometric algebra. To remind us that we are oper-
ating within the projective space setting rather than the vector space one, we
call such a GA a projective geometric algebra or PGA for short. The name
reflects the fact that in such a geometric algebra, the metric is based (either
directly or via a limiting process) on the projective cross ratio (as explained
in §3.5 above). Hence, a projective geometric algebra is a special case of a
homogeneous geometric algebra ([DFM07], Ch. 11), which is also sometimes
called a “1-up” model, since it requires (n+1)-dimensional coordinates to rep-
resent the an n-dimensional metric space. So, P(Rn+1,1,0 ) (resp., P(Rn,1,0 ))
provides a model for n-dimensional elliptic (resp., hyperbolic) space, and is
called elliptic (resp., hyperbolic) PGA.
3.6.1. DualVgeometric algebras. It is also possible to use the dual exterior
algebra P( (Rn+1 )∗ ) as the basis for a geometric algebra. Thus, the inner
product is defined on the hyperplanes of the projective space (the 1-vectors).
We call such a geometric algebra a dual geometric algebra; a geometric al-
gebra built atop P( Rn+1 ) we call standard to distinguish it from the dual
V
case. We sometimes call the former a plane-based algebra and the latter a
point-based one, emphasising the very different meaning of the 1-vectors in
the respective cases. One can compare the standard and dual GA with the
same signature by calculating the induced metric on n-vectors (which corre-
spond to the 1-vectors in the dual algebra). One finds that the dual algebra
P(R∗n+1,0,0 ) yields elliptic space again. P(R∗n,1,0 ), on the other hand, yields
dual hyperbolic space, built up of the hyperplanes lying outside the unit
sphere (rather than points inside the unit sphere). The induced signature
(n, 1, 0) on n-vectors (calculated by writing the basis n-vectors as products
of 1-vectors and squaring the results) is, however, the same as in the stan-
dard algebra P(Rn,1,0 ). When the metric is non-degenerate, as here, the dual
geometric algebra can be obtained by multiplying the original algebra by the
pseudoscalar I and then reversing the grades. Hence, a non-degenerate sig-
nature applied to the dual exterior algebra yields nothing new; every metric
relationship in the dual algebra is mirrored in the standard algebra via pseu-
doscalar multiplication. This is not true for degenerate signatures, as we see
in the next section.
3.7. The degenerate signature (n, 0, 1)
We established above that the degenerate signature (n, 0, 1) applied to the
dual Grassmann algebra leads to the euclidean algebra P(R∗n,0,1 ). The stan-
dard algebra P(Rn,0,1 ), however, represents a different metric space, dual
euclidean space. These cannot be obtained from one another by pseudoscalar
8 In fact, given the personal and scientific friendship of Klein and Clifford in the 1870’s,
it is likely that the Cayley-Klein construction influenced both Clifford’s discovery of bi-
quaternions (1873) and geometric algebra (1878) ([Zie85], Ch. 7).
 Geometric algebras for euclidean geometry10 9

multiplication since the pseudoscalar is not invertible. For example, for two
normalized k-blades A and B, A and B are parallel ↔ AI = BI. The in-
duced signature on n-vectors, (1, 0, n), is very degenerate, and not equivalent
to the signature on 1-vectors. As a result, euclidean and dual euclidean space
exhibit an asymmetry not present in the non-degenerate case: the absolute
quadric of euclidean space is a single ideal plane, while that of dual euclidean
space is a single ideal point. This reflects the fact that euclidean space arises
by letting the curvature of a non-euclidean space go to 0, while dual euclidean
space arises when the curvature goes to ∞.
3.7.1. Dual euclidean space. Because the distinction between euclidean and
dual euclidean space is crucial to the theme of this article, and is not well-
known, we discuss it briefly here. The simplest example of a dual euclidean
space occurs within the hyperbolic algebra P(Rn,1,0 ) (which forms the basis
of conformal geometric algebra, §3.8.2 below). A hyperplane tangent to the
null sphere Q at a point P has induced signature (n, 0, 1). P provides the
degenerate basis vector satisfying P2 = 0, all other points have non-zero
square since they do not lie on Q. Furthermore, no standard geometric algebra
can contain euclidean space as a flat subspace in this way. Why? We saw above
that the induced signature (1, 0, n) on points is more degenerate than the
signature (n, 0, 1) on hyperplanes. This asymmetry is incompatible with an
algebra in which 1-vectors represent points; only a dual geometric algebra can
provide both the required signature (n, 0, 1) on hyperplanes and (1, 0, n) on
points. Dual euclidean space shows promise as a tool for effectively modeling
some aspects of the natural world, see [Kow09] and [Gun11a], Ch. 10.
3.8. Geometric algebras for euclidean geometry
In this section we give an overview of the field of candidates of geometric
algebras for doing euclidean geometry. We have already met one of the can-
didates, P(R∗n,0,1 ), above in Sect. 3.7; we describe in it more detail now. We
will see below in 4.1 that homogeneous models with non-degenerate metrics
are inferior to P(R∗n,0,1 ) for euclidean geometry. The other remaining candi-
date for the chosen task is a 2-up model, conformal geometric algebra (CGA),
which we introduce next.
3.8.1. Projective geometric algebra. The geometric algebra P(R∗n,0,1 ) intro-
duced above for euclidean geometry we call euclidean PGA. When the context
makes it clear, as generally in the remainder of this article, we refer to it sim-
ply as PGA. Other examples of PGA’s are elliptic PGA (P(Rn+1,0,0 )) and
hyperbolic PGA (P(Rn,1,0 )).
The measurement of angles is given then by the inner product on the
1-vectors as described above in 3.5.1. The distance function between points,
also described there, appears in (at least) two different sub-products of the
algebra: d(P, Q) = kP ∨ Qk = kP × Qk (assuming that P and Q have been
normalized). Here, P∨Q is the joining line of the points, and P×Q := hPQi2
is the orthogonal complement of the joining line. Details of the first of these
formulas, and many other formulas, can be found in [Gun11a], Ch. 6 and
10 Charles Gunn

Ch. 7. The absolute quadric is the ideal plane; because of its importance we
introduce for it the notation ωP (’P’ stands for projective).
For n = 3, the case of most general interest, the even subalgebra
P(R∗+3,0,1 ) is isomorphic to the biquaternions. To construct the isomorphism,
map the imaginary biquaternions to the bivectors of P(R∗+ 3,0,1 ) in the obvi-
ous way (since both provide Plücker coordinates for line space), and  to the
pseudo-scalar −I of the geometric algebra. This isomorphism brings with it
the elegant representation of rigid body motion described above in Sect. 2.2.
The representation can be extended to include points and planes; details can
be found in [Gun11b], §15.6. Also note that PGA replaces the irregular trans-
formation formula for the sandwich operators of the biquaternions and of the
motor algebra (Sect. 2.2.2), with the uniform sandwich operators of the geo-
metric algebra. Warning: The biquaternions are also isomorphic to P(R+ 3,0,1 ),
the even subalgebra of dual euclidean space. We return to this later in the
article (Sect. 5.1) as it appears to have been a source of confusion relevant
to our theme.
P(R∗3,0,1 ) has the distinction of integrating two of William Clifford’s
most significant inventions, geometric algebra and biquaternions, into a single
algebra.11 Seen in this light, P(R∗3,0,1 ) stands in the confluence of two streams
of 19th century mathematics: on the one hand, that leading to the metric-
neutral biquaternion formulation of rigid body mechanics, and on the other
hand, the Cayley-Klein integration of metric geometry in projective geometry,
so that it has close connections to the genesis of geometric algebra itself. The
algebra first appeared in the modern literature in [Sel00] and [Sel05], and
was then extended and embedded in the metric-neutral toolkit described in
[Gun11a]. A compact, self-contained treatment is given in [Gun11b] (extended
version [Gun11c]).
3.8.2. Conformal geometric algebra. If one begins with the (n+1)-dimensional
PGA P(Rn+1,1,0 ) for hyperbolic geometry, one can obtain another model for
euclidean geometry as follows. Identify the points of the absolute quadric Q
(the null sphere) with En by stereographic projection. Then one can normal-
ize the coordinates of these points so that the inner product between two
points yields the square of the euclidean distance between the two points.
Points outside (inside) the sphere can be identified with spheres in Rn of
positive (negative) radius. The points of the null sphere itself can be identi-
fied with the points of En itself; and are sometimes called zero-radius spheres.
Projectivities which preserve Q correspond to conformal maps of Rn , hence
this model is called the conformal model of euclidean geometry, and the as-
sociated geometric algebra is called conformal geometric algebra (CGA). It
was introduced in its present form in [HLR01] and has developed rapidly

11 That Clifford himself appears to have overlooked this algebra is not surprising, con-

sidering the tentative nature of his research into both of these objects (unavoidable due
to his early death); the presence of a degenerate signature and the use of a dual exterior
algebra are both features of geometric algebra which were not known during his lifetime.
 Geometric algebras for euclidean geometry12 11

since then ([DL03], [DFM07], [Per09], and [DL11]) . In light of the prolific
literature available, we omit a more detailed description here.
The flat representation in CGA. CGA contains a sub-algebra closely related
to PGA. Since it will play a role in the sequel we describe it here. As noted
above, the tangent plane TP at a point P of the null sphere of CGA is a
sub-algebra isometric to dual euclidean space P(Rn,0,1 ). Letting P = n∞ ,
and polarizing TP by multiplying by the pseudoscalar I ∈ P(Rn+1,1,0 ) yields
another sub-algebra isometric to euclidean space P(R∗n,0,1 ). We call this sub-
algebra S. It consists of all flat subspaces containing n∞ . The relationship
to PGA is this: in PGA, the k-dimensional subspaces of En are represented
by (n − k)-vectors of the algebra; in S, the k-dimensional subspaces of En
are represented by the (k + 1)-vectors containing the “star” point n∞ . The
name flat representation comes from the fact that one can also obtain it
by taking the standard representation of CGA (as zero-radius spheres) and
wedge it with n∞ . For the purposes of this article, we content ourselves with
the observation that S and P(R∗n,0,1 ) are isometric, hence are essentially
identical (except that S has an extra, irrelevant dimension). Further work to
establish the exact relationship between these two representations needs to
be done.

4. Clarification work
We have identified the two algebras PGA and CGA as candidates the chosen
task. There exists some controversy in the literature whether there might be
other candidates, as well as questions regarding the suitability of PGA. We
now turn to examine these issues in more detail. In this section we dispose
of other homogeneous models which appear in the literature.

4.1. Which homogeneous model?

Here we want to discuss other homogeneous (i. e., 1-up) models for the chosen
task besides P(R∗n,0,1 ). For example, Chapter 11 of [DFM07], entitled The
homogeneous model, describes one such model (which appears in several other
textbooks ([DL03], [Per09]). In §11.1 one reads:
...[the homogeneous model of euclidean geometry] embeds Rn in a
space Rn+1 with one more dimension and then uses the algebra of
Rn+1 to represent those elements of Rn in a structured manner.
The authors then reject the use of a degenerate metric as “ inconvenient”, and
therefore propose using any non-degenerate metric, for example, (n + 1, 0, 0)
or (n, 1, 0), yielding the geometric algebras P(Rn+1,0,0 ) and P(Rn,1,0 ). That
means the basis element e0 satisfies e20 = ±1. We mentioned above in Sect. 3.5
that these two algebras yield a elliptic, resp. hyperbolic, metric on projective
space. Here we apparently encounter the widespread confusion between the
meaning of “euclidean” in the vector space versus the projective space setting.
We return to the consequences of this confusion below in Sect. 4.1.3.
12 Charles Gunn

P P P

Π. P Π. P Π. P ((Π . P)ΛΠ)VP

(Π . P)ΛΠ (Π . P)ΛΠ
Π Π Π

Figure 1. Solution of the exercise in PGA.

4.1.1. Comparison based on worked-out example. We compare this non-

degenerate homogeneous model with P(R∗n,0,1 ) on a simple geometric con-
struction, taken from §11.9 of [DFM07]:
Given a point P and a non-incident line Π in E3 , find the unique
line Σ passing through P which meets Π orthogonally.
P(R∗3,0,1 ) yields directly the compact solution ((Π · P) ∧ Π) ∨ P; Fig. 1 de-
composes the solution in three easy-to-understand steps. This PGA solution
is coordinate-free and metric-neutral, hence valid for hyperbolic and elliptic
space also. The first step is the most important: Π · P = hΠPi1 is the plane
p through P perpendicular to Π.
We adopt the solution from [DFM07], p. 310, to conform to the notation
used here (so ∨ is join, ∧ is meet, · is contraction, A⊥ = AI is the polar of A).
The reasoning is similar. Once the perpendicular plane p has been produced,
the desired line Σ can be obtained, as in PGA, using (p ∧ Π) ∨ P. The
difference lies in the definition
p := (P · ((e0 · Π) ∨ e0 ))⊥
(We have chosen e20 = 1 to simplify the expressions.) The reader can verify
that (e0 · Π) ∨ e0 is the line through the origin e0 parallel to Π. Hence when
Π passes through e0 , the expression obtained is the same as that in PGA
(modulo the presence of the polar operator ⊥ , which reflects the fact that
we are working in a standard rather than dual GA). When Π doesn’t pass
through e0 , then by translating Π there, one obtains the desired answer,
since a plane through P perpendicular to a line through e0 will also be
perpendicular to any translated copy of this line.
Here we see why e0 appears in every expression obtained in the discus-
sion in [DFM07]: exactly at e0 , the elliptic metric and the euclidean metric
agree. This is equivalent to the fact, that only at e0 does the polar plane in
the elliptic metric agree with the euclidean polar plane (which is always the
ideal plane). This allows one to translate geometric entities to the origin, op-
erate on them in a metric-neutral way (for example to obtain their euclidean
directions), and translate them back when necessary. In comparison to the
PGA solution, however, the one obtained in this way is neither coordinate-
free nor metric-neutral– in addition to involving an extra pair of operations
to translate the line to the origin.
Things become even less satisfactory when one attempts to implement
euclidean isometries. The authors acknowledge that using the non-degenerate
 Geometric algebras for euclidean geometry14 13

metric, it is impossible to express euclidean translations as sandwich opera-

tors (§11.8 of [DFM07]). This leads to the conclusion:
The main problem with using the metric of Rn+1 is that you can-
not use it directly to do Euclidean geometry, for it has no clear
Euclidean interpretation.
The foregoing quote is a good motivation for the next section, where we
attempt to clarify the situation by differentiating various meanings of Rn
and “euclidean”.
4.1.2. Three meanings of Rn . The symbol Rn occurs 5 times in the two short
quotes of the above section, grounds for asking what exactly it means. We
can in fact distinguish at least three different meanings:
1) Vector space. In this form, Rn represents the vector space used to de-
fine real projective space RP n−1 . It is an n-dimensional linear space
with an addition operation, real scalar multiplication, and distributive
law, but without inner product. One can develop a theory of linear map-
pings between such spaces, and from this, the dual vector space V∗ . The
evaluation map Rn ⊗ (Rn )∗ → R of a vector and a dual vector (linear
functional), often written hv, µi := µ(v) is sometimes confused with an
inner product. We recommend using the terminology (n-dimensional) V
for this meaning of Rn , whenever possible.13 See [Gre67a], Chapter 1-2
for details.
2) Inner product space. One begins with a vector space and adds an inner
product between pairs of vectors, which is a symmetric bilinear form
on the vector space. This produces an inner product space. When the
form is positive definite, it’s called a euclidean inner product space. We
recommend retaining the use of Rn for this meaning. Consult [Gre67a],
Chapter 7 for details on inner product spaces.
3) Euclidean space. This is a simply-connected metric space, of constant
curvature 0, homeomorphic to Rn but equipped with the Euclidean dis-
tance function (discussed for example in [Gun11a], Chapter 4) between
its points. We recommend using the notation En for this space. The
points of En are in a 1:1 correspondence to the vectors of Rn (the origin
of En maps to the zero vector of Rn ), but En is not a vector space, and
the inner product discussed in the previous item has, a priori, nothing
to do with the measurement of distances in En .
Armed with these three different meanings which sometimes are at-
tached to the same symbol Rn , along with two meanings of euclidean (de-
pending on whether one is in the vector space or the projective setting), let’s
return to the discussion of the homogeneous model.

13 It is not always convenient, see for example 3.2 above, where Rn is traditionally used

to define the exterior algebra, even though the inner product plays no role thereby.
14 Charles Gunn

4.1.3. Rephrasing using the differentiated notation. When we apply this dif-
ferentiated terminology and what we learned about the Cayley-Klein con-
struction of metric spaces to the initial quote from [DFM07], we arrive at the
following:
...[the homogeneous model] embeds En in a real vector space V of
dimension n + 1 and then uses the algebra of V to represent the
elements of En in a structured manner.
In this form, Rn no longer occurs: there is no longer a given real vector space
nor inner product, implied by the original definition. Consequently, one can
use this modified description as a starting point for the search for the correct
choice of Cayley-Klein space; we have sketched above how one arrives at the
dual vector space V = (Rn+1 )∗ with signature (n, 0, 1), yielding the algebra
P(R∗n,0,1 ).
To sum up: this confusion of the three meanings of Rn and two meanings
of euclidean means that many of the objections to “the” homogeneous model
appear in the light of the foregoing discussion as legitimate complaints against
using the wrong vector space or the wrong signature to model euclidean ge-
ometry. In the next section we turn to consider if there are other choices
which yield better results.

5. Homogeneous models using a degenerate metric

Faced with the difficulties ensuing on the use of the non-degenerate metric,
[DFM07], p. 314, states:
We emphasize that the problem is not geometric algebra itself,
but the homogeneous model and our desire to use it for euclidean
geometry. It will be replaced by a much better model for that
purpose in Chapter 13 [the conformal model - cg].
In fact, what the authors of [DFM07] have shown is that a homogeneous
model with non-degenerate metric is “the problem” – recall that the use
of degenerate metrics was rejected as inconvenient. Hence it remains to be
seen whether a PGA based on a degenerate metric, such as P(R∗n,0,1 ), could
provide a faithful model for euclidean geometry. We now turn to an analysis
of three common objections to the use of such a degenerate metric.

5.1. Objection 1: lack of covariance

Covariance has a variety of meanings related to the behavior of maps and
coordinate systems; in our context, it is equivalent to the existence of sand-
wich (or versor ) implementations of the euclidean group E(n) ([DFM07], p.
369). That such versors exist for the even subalgebra of P(R∗3,0,1 ) (and the
associated Spin group) follows from its isomorphism with the biquaternions
mentioned above in Sect. 3.8.1. The extension to the full algebra (and the as-
sociated Pin group) is straightforward and is described in [Gun11b], §15.4.2,
§15.5.4.
 Geometric algebras for euclidean geometry17 15

One can gain a different impression, however, from some of the current
literature. For example, [Li08], p. 11, identifies the Clifford algebra P(Rn,0,1 )
(in our notation) as the appropriate algebra for euclidean geometry. It focuses
on the case of n = 3 and remarks on the isomorphism of the dual quaternions
with the even subalgebra. This leads to the remark:
However, the dual quaternion representations of primitive geomet-
ric objects such as points, lines, and planes in space are not co-
variant. More accurately, the representations are not tensors, they
depend upon the position of the origin of the coordinate system
irregularly.
In the first place, the proper algebra for euclidean geometry is not
P(Rn,0,1 ) but the dual version P(R∗n,0,1 ) (see above, Sect. 3.7.1). This con-
fusion is perhaps due to the fact that the dual quaternions are isomorphic
to the even sub-algebra of both these geometric algebras (Sect. 3.8.1 above).
Furthermore, as discussed above in Sect. 2.2.2, the dual quaternion represen-
tation of points and planes is not the same as the representation of points and
planes in these geometric algebras: the dual quaternion representation has ir-
regularities not exhibited by the geometric algebra representation due to the
fact there there is no natural representation for points and planes within the
algebra. These irregularities however have to do with the form the sandwich
operators take and do not effect the covariance of the representation. We
are aware of no grounds for the claim made here that the dual quaternion
representations are not covariant. It might be due to mixing up the two al-
gebras P(Rn,0,1 ) and P(R∗n,0,1 ) in the calculation, since each provides, taken
for itself, covariant representations for their respective isometry group.
Readers who would like to confirm the claimed covariance for them-
selves, and do not have access to [Gun11b] or [Gun11a] are referred to Ap-
pendix A, which provides a detailed discussion of the versors of P(R∗2,0,1 ) and
their associated sandwich operators.
5.2. Objection 2: lack of duality
Another common objection to the use of degenerate metrics is often expressed
in terms of a “lack of duality”. Consider the following quote from [HLR01],
an article often associated with the birth of modern CGA ([DFM07], §13.8):
Any degenerate algebra can be embedded in a non-degenerate al-
gebra of larger dimension, and it is almost always a good idea to
do so. Otherwise, there will be subspaces without a complete basis
of dual vectors, which will complicate algebraic manipulations.
As with the case of euclidean above, there are multiple meanings for the term
dual in the literature which must be carefully differentiated. Here there are
at least two distinct meanings:
1. The ability to calculate the regressive product. As shown in Sect. 3.4,
for this operation one only requires the dual coordinates of a geometric
entity, and this is provided in a non-metric way by Poincaré duality in
the exterior algebra, or by the shuffle product.
16 Charles Gunn

2. The effect of multiplication by the pseudoscalar: Π : Π(X) := XI. The

action of Π on the underlying projective space is known in classical
projective geometry as the polarity on the metric quadric, that is, a cor-
relation (maps points to hyperplanes and vice-versa) that maps a point
to its orthogonal hyperplane with respect to the inner product encoded
in the quadric, and vice-versa.15 When the metric is non-degenerate,
then I2 = ±1, and the polarity is a grade-reversing algebra bijection
which is a vector-space isomorphism on each grade, and whose square
is the identity16 . In this case, one can define the regressive product
(analogously to the use of Poincaré duality in Sect. 3.4):
X ∧ Y := Π(Π(X) ∨ Π(Y))
We recommend that the term polarity be adapted also in geometric
algebra for pseudoscalar multiplication, to distinguish it from the previous
non-metric meaning of duality. So that, in the above quote, one would speak
of the polar basis instead of the dual basis. The term dual basis would be
reserved for the result of the dual coordinate map J.
The implicit use of the metric to calculate the regressive product has
a long tradition, going all the way back to Grassmann and continuing up to
[HZ91], an influential modern article devoted to doing projective geometry
using geometric algebra; its continued use of Π for the regressive product –
despite the absence of a natural metric for projective geometry – appears
to have cemented the misunderstanding described here. As a result, in the
above quote as well as other popular texts ([DFM07], [Per09], [DL03]) one
might falsely gain the impression that a non-degenerate metric must be used
to implement the regressive product, as few or no details of an alternative are
provided. But in the absence of an invertible pseudo-scalar, one always has
access to Poincaré duality. Hence this objection to PGA cannot be sustained.
5.3. Objection 3: Absence of invertible pseudoscalar
[Li08], also p. 11, raises a further (related) objection to the use of a degenerate
metric:
Because the inner product in Rn,0,1 is degenerate, many important
invertibilities in non-degenerate Clifford algebras are lost.18
We have discussed the invertibility of I as a condition for duality above
in Objection 2, and shown that there are other means to implement duality.
It is true that many formulas in the GA literature tend to be given in terms
of I−1 rather than I. But in the cases we are familiar with, it is also possible

15 If one considers the inner product as a symmetric bilinear form B(u, v), then one

obtains a linear functional Bu by fixing u and defining Bu (v) := B(u, v). Then the kernel
of Bu is the indicated orthogonal hyperplane: the set of all vectors with vanishing inner
product with u. This is also sometimes referred to in GA as the inner product null space
(IPNS) of u.
16 Since even if Π2 = −1, X and −X represent, projectively, the same element.
18 Since it is irrelevant to this object, we overlook the fact, discussed above, that the

correct Clifford algebra here should be R∗n,0,1 .

 Geometric algebras for euclidean geometry19 17

to use I instead, perhaps at the cost of a more complicated expression for

the sign. For example, [Hes10] defines the dual (our polar !) of a multivector
A∗ := AI−1 and shows then that the inner and outer products obey the
relation: a · A∗ = (a ∧ A)∗ . If instead one defines A∗ := AI, the stated relation
remains true, and valid for any pseudoscalar.
Experience leads us, in fact, to a very different view of the non-invertible
pseudoscalar: it has proved to be an advantage, since it faithfully repre-
sents the metric relationships within euclidean geometry. The calculation
from Sect. 4.1.1 provides a good example. Consider the sub-expression Π · P.
Letting the point P move freely, one obtains a set of parallel planes Π · P
which all have the same polar point (Π · P)I, the ideal point of the line Π.
For invertible pseudo-scalars, however, the polar points of distinct planes are
distinct. Or, recall the discussion of the special role of e0 in the solution in
Sect. 4.1.1: it does not appear in the simpler PGA formula, since in PGA,
PI = ωP for all normalized P, while in the elliptic metric this is only true
for P = e0 .

6. Comparison: A feature-set for “doing geometry”

Having sketched its mathematical lineage dating back to Klein and Clifford,
and disposed of a series of modern objections which have been raised against
it, the reader is hopefully convinced that euclidean PGA deserves the title
of “standard” or “classical” homogeneous model of euclidean geometry. We
are now prepared to compare it to the conformal model, CGA. As a basis for
this comparison, we rely on a recent tutorial on the conformal model [Dor11].
This tutorial describes the challenge of “doing geometry” on a computer, a
challenge which matches well with the chosen task we set out at the beginning
of the article, so we use this tutorial as a basis for a first comparison of the
two algebras. The tutorial lists seven “tricks” and three “bonuses” which the
conformal model offers in this regard. We list them here:
1. Trick 1: Representing euclidean points in Minkowski space.
2. Trick 2: Orthogonal transformations as multiple reflections in a sand-
wiching representation.
3. Trick 3: Constructing elements by anti-symmetry.
4. Trick 4: Dual specifications of elements permits intersection.
5. Bonus: The elements of euclidean geometry as blades.
6. Bonus: Rigid body motions through sandwiching.
7. Bonus: Structure preservation and the transfer principle.
8. Trick 5: Exponential representation of versors.
9. Trick 6: Geometric calculus.
10. Trick 7: Sparse implementation at compiler level.
How does P(R∗n,0,1 ) stand with respect to these features? In fact, it offers
offers all the ten features listed. Some slight editing is required to “translate”
to PGA; for example, Trick 1 has to be rephrased as “Representing euclidean
points in projective space”. Duality (trick 4) is implemented in a non-metric
18 Charles Gunn

way in our homogeneous model, and is used to represent join, not intersection.
There are naturally some elements of euclidean geometry which cannot be
represented as blades in PGA (bonus 1), such as point pairs and spheres. But
the basic flat elements belonging to classical euclidean geometry are present:
points, lines, and planes; and these are the ones belonging to the chosen task.
We return to the richer class of primitives in CGA in Sect. 8 below.
One immediate corollary is that euclidean PGA (P(R∗3,0,1 )) is the small-
est known algebra that can model Euclidean transformations in a structure-
preserving manner, a distinction sometimes claimed for CGA ([DFM07], p.
364). The importance of this result will become more apparent in the next
section, which turns to a practical comparison of the two models.

7. Comparison: practical issues

Given the same feature set in these algebras, we shift our comparison to more
practical considerations for the chosen task.
7.1. Complexity
The point x = (x, y, z) ∈ E3 receives coordinates (x, y, z, 1) in PGA and
(1, x, y, z, 21 kxk2 ) in CGA.20 The last coordinate is clearly non-linear function
of the original ones. This standard representation is sometimes called the
zero radius sphere (ZRS) representation of points in CGA. Also note, that
as a result of having 1 more dimension, CGA also has twice the number of
dimensions as PGA.
A more serious effect of the non-linear embedding of this representa-
tion, is that flat euclidean geometric configurations have to be represented
and calculated as intersections of linear configurations with the null sphere
of P(Rn+1,1,0 ). For example, if you want to subdivide a polygon in PGA,
linear interpolation will preserve the flatness of the polygon; in the ZRS rep-
resentation of CGA, you have to follow linear interpolation with a projection
back onto the null sphere or devise other interpolation methods. As men-
tioned above in Sect. 3.8.2, one can alternatively use the flat representation
in CGA, which is essentially the same as PGA. But to access the distinctive
features of CGA (such as the distance function via the inner product) you
have then to convert from the flat back to the ZRS representation, leading
to the conclusion that one cannot in this way avoid the consequences of the
non-linear embedding.
In general, any computation applied to a geometric primitive in the
standard CGA representation risks moving off the null sphere, so potentially
each step has to be checked against an error tolerance and corrected. PGA

20 This parametrization produces a paraboloid of revolution as null quadric. To obtain

the unit sphere as null quadric, one can rotate by π2 in the plane of the two “extra”
dimensions to obtain the coordinates ( 12 (kxk2 + 1), x, y, z, 12 (kxk2 − 1)).
 Geometric algebras for euclidean geometry21 19

does not suffer from this difficulty, since all coordinates represent valid eu-
clidean or ideal elements. Related problems arise with differential equations,
as the next section shows.
7.2. Numerical analysis and differential equations
Consider the example of the euclidean equations of motion for a rigid body.
(In PGA, for n = 3, these are the biquaternion ODE’s given above in
Sect. 2.2.) As with any ODE, the difficulty of solving the ODE is directly
connected with how the space of valid solutions sits inside the full space of
possible solutions; the smaller the co-dimension of the former in the latter,
the easier the solution process is; see for example the discussion above in
Sect. 2.1.1 regarding the advantages of the quaternion representation of rigid
body motion over the matrix one. This difficulty is acknowledged in a recent
article on 3D euclidean rigid body motion in the conformal model ([LLD11],
§1.3.1):
... The idea here is to work in an overall space that is two dimen-
sions higher than the base space, using the usual conformal Eu-
clidean setup. The penalty for doing this, i. e., using a Euclidean
setup, is that the number of degrees of freedom is not properly
matched to the problem in hand, and we have to introduce addi-
tional Lagrange multipliers to cope with this. [our italics]
How do the two algebras compare in this regard? In both, the space of valid
solutions for the Euler equations of rigid body motion consists of a euclidean
bivector (in the Lie algebra e(3)) plus a euclidean rotor (in the Lie group
E(3)); each is 6D, so the space of valid solutions is 12-dimensional. In the
case of CGA, the full space of possible solutions is 26-dimensional (10D bivec-
tor and 16D even subalgebra). Here the co-dimension is 14, larger than the
solution space itself, forcing the use of Lagrange multipliers. In contrast, the
full solution space in PGA is 14-dimensional (6D bivector and 8D even subal-
gebra). Normalizing the rotor in PGA to have unit norm brings the solution
back onto the valid solution space and provided reliable results for the ex-
tensive simulations in [Gun11a], Ch. 12), although the use of a Lagrange
multiplier for the same purpose in the PGA case should be investigated.
Given the availability of a fast and reliable PGA solution with minimal need
for Lagrange multipliers, the question naturally arises, what advantages does
the CGA approach to rigid body mechanics offer in compensation?
7.3. Kinematics, rigid body mechanics, and classical screw theory
As we noted above, PGA contains within it the biquaternions and their el-
egant representation of 3D euclidean kinematics and rigid body mechanics.
This is essentially also the content of the screw theory of Robert Ball [Bal00].
As a result, all the features of these theories are included in PGA as a “na-
tive” element. [Hes10], on the other hand, envisions CGA as a means to
“rejuvenate” classical screw theory. It would be worthwhile to compare the
two approaches to screw theory with regard to such criteria as compactness
of expression, practicality, and comprehensibility. For example, screw theory
20 Charles Gunn

is built upon line geometry in RP 3 , which in turn is built upon Plücker co-
ordinates for lines (bivectors). As noted above, these coordinates are native
to PGA, but have to be extracted from the 10-dimensional coordinates of
bivectors in CGA using the so-called conformal split ([Hes10], §VII).

7.4. Learning curve

The simplicity of the PGA representation leads to considerable savings in
explaining the model, a significant advantage when considering the unfamil-
iarity of the underlying concepts. Since CGA is embedded in the hyperbolic
PGA (P(R∗n+1,1,0 )), learning the projective model is a natural step towards
understanding the conformal model, but not vice-versa. From a pedagogical
point of view, PGA forms a natural transition step between VGA and CGA.
(A vector geometric algebra, or VGA for short, is a geometric algebra whose
elements are interpreted as elements of a vector space, rather than projective
space. )

8. Roundness and CGA

We have established that for the chosen task, PGA exhibits a series of prac-
tical advantages. Most of these can be arise from the contrast between, on
the one hand, the flat embedding of En in RP n in PGA and, on the other,
the curved embedding of En within RP n+1 in CGA. We could say that the
“roundness” of CGA is a liability when one is restricted to flat primitives.
There is however another “roundness” in CGA that, for some euclidean ap-
plications, compensates for these liabilities: euclidean spheres are represented
as points in CGA and can be manipulated on the same level as traditional
flat primitives: points, lines, and planes. The restriction to spheres of radius
0 yields the curved model of En which has formed the basis of the compari-
son up to now. Removing the restriction to traditional flat primitives yields
a powerful geometric toolkit ideally suited for many euclidean tasks where
spheres (or conformal maps, see Sect. 3.8.2) play an intrinsic role ([DFM07],
Ch. 14). There are a number of application areas, from optimization to ro-
botics, whose problem settings do exhibit this close connection to sphere (or
conformal) geometry. See, for example, [Dor14].
Users choosing between PGA and CGA are therefore advised to care-
fully weigh the advantages and disadvantages of the “roundness” of CGA in
their decision. On the one side are the advantages of having a direct, powerful
representation of spheres; on the other hand, are the disadvantages (discussed
above) arising from the embedding of euclidean space itself as a sphere in a
higher-dimensional projective space (the null sphere of P(Rn+1,1,0 )). Hence,
if there is no intrinsic need for sphere geometry, as is the case for applications
based on flat geometry, classical kinematics and rigid body mechanics, then
the disadvantages listed in Sect. 7 can be expected to outweigh the advan-
tages; if spheres form an essential geometric primitive, CGA is probably the
right tool for the job. In between there’s much room for further research and
 Geometric algebras for euclidean geometry22 21

development on how the strengths and weaknesses of the two approaches can
be optimally combined.

9. Conclusion
We have reviewed important concepts from 19th century mathematics, and
clarified a set of fundamental terms, including homogeneous model, euclidean,
Rn , and duality, which are key to a correct understanding of how geomet-
ric algebra can be applied to doing euclidean geometry. On this basis, we
have established that the dual projective geometric algebra P(R∗n,0,1 ) de-
serves the title of “standard” homogeneous model for euclidean geometry.
We have shown that it exhibits all the attractive features with respect to
doing euclidean geometry which modern geometric algebra users expect, and
is the smallest such algebra. Furthermore, in regard to practical considera-
tions for the chosen task, we have found that it exhibits advantages over the
higher-dimensional CGA, which remains the tool of choice for applications
making essential use of spheres or conformal maps. For n = 3, the most pop-
ular case, the fact that P(R∗3,0,1 ) is built atop the biquaternions, William
Clifford’s “other” great discovery (besides geometric algebra), means that
users already familiar with the biquaternions are well-positioned to acquire
PGA skills quickly.

Euclide
an Rou
CO n ds
EuclideNFORMA (4,1)
P an Flat L
R
Euclide OJECTI
s )
an Dire VE (3,0,
1
VECTO ctions
R SPA 3,0)
CE (
Figure 2. Progression of geometric algebras for euclidean geometry.

One area of concern for all practitioners of geometric algebra, regard-

less of specialization, is the slow rate at which geometric algebra has been
adopted into the university curriculum (see [DL11], p. vi). We believe that
the foregoing comparison can make a significant contribution to a solution
of this challenge. PGA and CGA are not directly competitive, any more
than automobiles and airplanes are directly competitive. As the previous
section hopefully indicates, each has their proper, interdependent place in
the geometric algebra ecosystem. Pedagogically, P(R∗n,0,1 ) provides a natural
stepping-stone (the “automobile”) between the n-dimensional vector space
22 Charles Gunn

algebra Rn,0,0 (the “bicycle” of the GA world, aka VGA) and the (n + 2)-
dimensional CGA (the “airplane”). See Fig. 2. In light of Sect. 7.4, we can
expect that PGA will be accessible to a significantly larger pool of students
than currently is the case with CGA. This will also simplify the teaching of
CGA since, as remarked above, CGA is built on top of PGA. We suggest
that the PGA treatment of euclidean geometry, kinematics, and mechanics
is exactly what is needed to solve the GA adoption problem by providing a
non-trivial link between VGA and CGA, just as the automobile fits between
the bicycle and the airplane.
Thus, these two algebras exist not in a competitive but a complemen-
tary relationship. The nature of this complementarity was already expressed
by Johannes Kepler – perhaps the first scientist to apply mathematics in a
modern way to the study of the outer world – when he wrote ([Kep19]):
...God, in his measureless wisdom, selected at the very beginning
the straight and the curved, in order with them to imprint into
the world the divinity of the Creator... In this way the most Wise
devised the extensive world, whose whole being is encompassed
between the two contrary principles, the straight and the curved.
Rounding off our article in Kepler’s style, we might say that God has given
us PGA to understand and to master the world of the straight, while he has
given us CGA to do the same for the world of the round.

10. Acknowledgements
The author would like to thank the third reviewer for his detailed comments,
which led to numerous improvements in the article.

Appendix A. Versors for the euclidean plane.

Begin with two normalized 1-vectors a and b in P(R∗2,0,1 ), each representing
a line in the euclidean plane and assume that they meet in a euclidean point
P, i. e., a ∧ b = P 6= 0. We show that Ra (b), the reflection of the line b in
the line a, is given by aba.
Using basic facts of elementary geometry, it’s not hard to show that
x := Ra (b) can be uniquely characterized as the line which passes through
P, the intersection of a and b, and whose oriented angle to a is equal but
opposite to the angle b makes to a. The derivation of the signature (2, 0, 1)
in Sect. 3.5.1 allows us to translate these conditions into the language of the
geometric product in P(R∗2,0,1 ): x · a = b · a and x ∧ a = −b ∧ a. First note
that aba is a 1-vector (a line), since the three arguments are not linearly
independent, so their wedge is 0. Is it the desired line? Substituting aba for
x in the first condition and applying symmetry of (·) and the normalization
 Geometric algebras for euclidean geometry23 23

condition a2 = 1 yields
1
x·a= (aba2 + a2 ba)
2
1
= (ab + ba)
2
=a·b=b·a
And the second condition proceeds similarly, using the anti-symmetry of ∧:
1
x ∧ a = (aba2 − a2 ba)
2
1
= (ab − ba)
2
= −b ∧ a
Hence, x fulfills the conditions and is, therefore, the reflection of the line b in
the line a. We leave it as an exercise for the reader to verify that this argument
remains true when a and b are parallel (i. e., P is ideal). A further exercise:
aPa is the reflection of a euclidean point P in the line a. The tireless reader
can then extend this result to the full euclidean isometry group by applying
the well-known result that all isometries can be factored as a sequence of
reflections in euclidean lines. Such a sequence of reflections yields, in the
algebra, a versor R consisting of the product of the corresponding 1-vectors.
When the 1-vectors are normalized, then so is the resulting versor, which
then belongs to the rotor group, and one obtains the sandwich operation
associated to this rotor: RXR. e Nothing in this proof essentially depends on
the dimension n = 2, it generalizes directly and establishes the claim that the
versor representation of the isometry group in P(R∗n,0,1 ) works as advertised.

References
[Bal00] Robert Ball. A Treatise on the Theory of Screws. Cambridge University
Press, Cambridge, 1900.
[BCDK00] Eduardo Jose Bayro-Corrochano, D., and D. Khler. Motor algebra ap-
proach for computing the kinematics of robot manipulators. J. Robotic
Syst., 17:495–516, 2000.
[Bla42] Wilhelm Blaschke. Nicht-euklidische Geometrie und Mechanik. Teubner,
Leipzig, 1942.
[Cli73] William Clifford. A preliminary sketch of biquaternions. Proc. London
Math. Soc., 4:381–395, 1873.
[DFM07] Leo Dorst, Daniel Fontijne, and Stephen Mann. Geometric Algebra for
Computer Science. Morgan Kaufmann, San Francisco, 2007.
[DL03] Chris Doran and Anthony Lasenby. Geometric Algebra for Physicists.
Cambridge University Press, Cambridge, 2003.
[DL11] Leo Dorst and Joan Lasenby, editors. Guide to Geometric Algebra in
Practice. Springer London, 2011.
[Dor11] Leo Dorst. Tutorial appendix: Structure preserving representation of eu-
clidean motions through conformal geometric algebra. In Leo Dorst and
24 Charles Gunn

Joan Lasenby, editors, Guide to Geometric Algebra in Practice, pages

435–453. Springer London, 2011.
[Dor14] L. Dorst. Total least squares fitting of k-spheres in n-d euclidean space
using an (n+2)-d isometric representation. J of Mathematical Imaging
and Vision, pages 1–21, 2014.
[Gre67a] W. H. Greub. Linear Algebra. Springer, 1967.
[Gre67b] W. H. Greub. Multilinear Algebra. Springer, 1967.
[Gun11a] Charles Gunn. Geometry, Kinematics, and Rigid Body Mechanics in
Cayley-Klein Geometries. PhD thesis, Technical University Berlin, 2011.
http://opus.kobv.de/tuberlin/volltexte/2011/3322.
[Gun11b] Charles Gunn. On the homogeneous model of euclidean geometry. In
Leo Dorst and Joan Lasenby, editors, A Guide to Geometric Algebra in
Practice, chapter 15, pages 297–327. Springer, 2011.
[Gun11c] Charles Gunn. On the homogeneous model of euclidean geometry: Ex-
tended version. http://arxiv.org/abs/1101.4542, 2011.
[Gun16] Charles Gunn. Geometric algebras for euclidean geometry. Advances in
Applied Clifford Algebras, pages 1–24, 2016.
[Hes10] David Hestenes. New tools for computational geometry and rejuvenation
of screw theory. In Eduardo Jose Bayro-Corrochano and Gerik Scheuer-
mann, editors, Geometric Algebra Computing in Engineering and Com-
puter Science, pages 3–35. Springer, 2010.
[HLR01] David Hestenes, Hongbo Li, and Alyn Rockwood. A unified algebraic
approach for classical geometries. In Gerald Sommer, editor, Geometric
Computing with Clifford Algebra, pages 3–27. Springer, 2001.
[HZ91] David Hestenes and Renatus Ziegler. Projective geometry with clifford
algebra. Acta Applicandae Mathematicae, 23:25–63, 1991.
[Kep19] Johannes Kepler. Harmonices Mundi LIbri V. Linz, 1619.
[Kle26] Felix Klein. Vorlesungen Über Nicht-euklidische Geometrie. Chelsea,
New York, 1926. (Original 1926, Berlin).
[Kow09] Gerhard Kowol. Projektive Geometrie und Cayley-Klein Geometrien der
Ebene. Birkhauser, 2009.
[Li08] Hongbo Li. Invariant Algebras and Geometric Reasoning. World Scien-
tific, Singapore, 2008.
[LLD11] Anthony Lasenby, Robert Lasenby, and Chris Doran. Rigid body dy-
namics and conformal geometric algebra. In Leo Dorst and Joan
Lasenby, editors, Guide to Geometric Algebra in Practice, chapter 1,
pages 3–25. Springer, 2011.
[Per09] Christian Perwass. Geometric Algebra with Applications to Engineering.
Springer, 2009.
[Sel00] Jon Selig. Clifford algebra of points, lines, and planes. Robotica, 18:545–
556, 2000.
[Sel05] Jon Selig. Geometric Fundamentals of Robotics. Springer, 2005.
[Stu03] Eduard Study. Geometrie der Dynamen. Tuebner, Leibzig, 1903.
[vM24] Richard von Mises. Die Motorrechnung: Eine Neue Hilfsmittel in
der Mechanik. Zeitschrift für Rein und Angewandte Mathematik und
Mechanik, 4(2):155–181, 1924.
[Zie85] Renatus Ziegler. Die Geschichte Der Geometrischen Mechanik im 19.
Jahrhundert. Franz Steiner Verlag, Stuttgart, 1985.
 Geometric algebras for euclidean geometry24 25

Charles Gunn
Institüt fr Mathematik MA 8-3
Technische Universität Berlin
Str. des 17 Juni 136
10623 Berlin Germany

Charles Gunn
Raum+Gegenraum
Brieselanger Weg 1
14612 Falkensee
e-mail: [email protected]