Acta Applicandae Mathematicae 71: 207244, 2002.

2002 Kluwer Academic Publishers. Printed in the Netherlands.

207
Geometric Algebra in Linear Algebra and
Geometry
JOS MARA POZO
1
and GARRET SOBCZYK
2
1
Departament de Fsica Fonamental, Universitat de Barcelona, Diagonal 647, E-08028 Barcelona,
Spain. e-mail: jpozo@ffn.ub.es
2
Departamento de Fisica y Matematicas, Universidad de las Amricas-Puebla, Mexico,
72820 Cholula, Mxico. e-mail: sobczyk@mail.udlap.mx
(Received: 18 February 2000; in nal form: 17 July 2001)
Abstract. This article explores the use of geometric algebra in linear and multilinear algebra, and
in afne, projective and conformal geometries. Our principal objective is to show how the rich
algebraic tools of geometric algebra are fully compatible with and augment the more traditional
tools of matrix algebra. The novel concept of an h-twistor makes possible a simple new proof of the
striking relationship between conformal transformations in a pseudo-Euclidean space to isometries
in a pseudo-Euclidean space of two higher dimensions. The utility of the h-twistor concept, which is
a generalization of the idea of a Penrose twistor to a pseudo-Euclidean space of arbitrary signature,
is amply demonstrated in a new treatment of the Schwarzian derivative.
Mathematics Subject Classications (2000): 15A09, 15A66, 15A75, 17Bxx, 41A10, 51A05,
51A45.
Key words: afne geometry, Clifford algebra, conformal group, Euclidean geometry, geometric
algebra, Grassmann algebra, horosphere, Lie algebra, linear algebra, Mbius transformation, non-
Euclidean geometry, null cone, projective geometry, spectral decomposition, Schwarzian derivative,
twistor.
1. Introduction
Almost 125 years after the discovery of geometric algebra by William Kingdon
Clifford in 1878, the discipline still languishes off the centerstage of mathematics.
Whereas Cliffords geometric algebra has gained currency among an increasing
number scientists in different special interest groups, the authors of the present
work contend that geometric algebra should be known by all mathematicians and
other scientists for what it really is the natural algebraic completion of the real
number system to include the concept of direction. Whereas, evidently, most math-
ematicians and other scientists are either unfamiliar with or reject this point of
view, we will try to prevail by showing that Clifford algebra really has already
been universally recognized in the guise of linear algebra. Since linear algebra is
fully compatible with Clifford algebra, it follows that in learning linear algebra,
every scientist has really learned Clifford algebra but is generally unaware of this
208 JOS

E MAR

IA POZO AND GARRET SOBCZYK

fact! What is lacking in the standard treatments of linear algebra is the recognition
of the natural graded structure of linear algebra and, therefore, the geometric in-
terpretation that goes along with the denition of geometric algebra. As has been
often repeated by Hestenes and others, geometric algebra should be seen as a great
unier of the geometric ideas of mathematics (Hestenes, 1991).
The purpose of the present article is to develop the ideas of geometric algebra
alongside the more traditional tools of linear algebra by taking full advantage of
their fully compatible structures. There are many advantages to such an approach.
First, everybody knows matrix algebra, but not everybody is aware that exactly the
same algebraic rules apply to the multivectors in a geometric algebra. Because of
this fact, it is natural to consider matrices whose elements are taken from a geomet-
ric algebra. At the same time, by developing geometric algebra in such a way that
any problem can be easily changed into an equivalent problem in matrix algebra,
it becomes possible to utilize the powerful and extensive computer software that
has been developed for working with matrices. Whereas CLICAL has proven itself
to be a powerful computer aid in checking tedious Clifford algebra calculations,
it lacks symbolic capabilities (Lounesto, 1994). Geometric algebra offers not only
a comprehensive geometric interpretation but also a whole new set of algebraic
tools for dealing with problems in linear algebra. We show that matrices, which
are rectangular blocks of numbers, represent geometric numbers in a rather special
spinor basis of a geometric algebra with neutral signature.
This work consists of four main sections. This introductory section lays down
the rational for this article and gives a brief summary of its main ideas and content.
Section 2 is primarily concerned with the development of the basic ideas of linear
and multilinear algebra on an n-dimensional real vector space we call the null
space, since we are assuming that all vectors in N are null vectors (the square
of each vector is zero). Taking all linear combinations of sums of products of
vectors in N generates the 2
n
-dimensional associative Grassmann algebra G(N).
This stucture is sufciently rich to efciently develop many of the basic notions of
linear algebra, such as the matrix of a linear operator and the theory of determinants
and their properties.
Recently, there has been much interest in the application of geometric algebra
to afne, projective and other non-Euclidean geometries (Maks, 1989; Hestenes,
1991; Hestenes and Ziegler, 1991; Porteous, 1995; Havel, 1995). These non-
Euclidean models offer new computational tools for doing pseudo-Euclidean and
afne geometry using geometric algebra. Section 3 undertakes a systematic study
of some of these models, and shows how the tools of geometric algebra make it
possible to move freely between them, bringing a unication to the subject that
is otherwise impossible. One of the key ideas is to dene the meet and join op-
erations on equivalence classes of blades of a geometric algebra which represent
subspaces. Since a nonzero r-blade characterizes only the direction of a subspace,
the magnitude of the blade is unimportant. Basic formulas for incidence relation-
ships between points, lines, planes, and higher-dimensional objects are compactly
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 209
formulated. Examples of calculations are given in the afne plane which are just
plain fun!
Section 4 explores the deep relationships which exist between projective geom-
etry and the conformal group. The conformal geometry of a pseudo-Euclidean
space can be linearized by considering the horosphere in a pseudo-Euclidean space
of two dimensions higher. The introduction of the novel concept of an h-twistor
makes possible a simple new proof of the striking relationship between conformal
transformations in a pseudo-Euclidean space to isometries in a pseudo-Euclidean
space of two higher dimensions. The concept of an h-twistor greatly simplies
calculations and is in many ways a generalization of the successful spinor/twistor
formalisms to pseudo-Euclidean spaces of arbitrary signatures. The utility of the
h-twistor concept is amply demonstrated in a new derivation of the Schwarzian
derivative (Davis, 1974, p. 46; Nehari, 1952, p. 199).
2. Geometric Algebra and Matrices
Let N be an n-dimensional vector space over a given eld K, and let
{e} = ( e
1
e
2
. . . e
n
) (1)
be a basis of N. In this work we only consider real (K = R) or complex (K = C)
vector spaces although other elds could be chosen. By interpreting each of the
vectors in {e} to be the column vectors of the standard basis of the identity matrix
id(n) of the n n matrix algebra M(K) over the eld K, we are free to make the
identication {e} = id(n). We wish to emphasize that we are interpreting the basis
vectors e
i
to be elements of the 1 n row matrix (1), and not the elements of a set.
Thus, in what follows, we are assumming and often will apply the rules of matrix
multiplication when dealing with the (generalized) row vector of basis vectors {e}.
Now let N be the dual vector space of 1-forms over the the eld K, and let {e}
be the dual basis of N with respect to the basis {e} of N. If we now interpret each
of the vectors in {e} to be the row vectors of the standard basis of the identity matrix
id(n) of the n n matrix algebra M(K), we can again make the identication
{e} = id(n). Because we wish to be able to interpret the elements of {e} as row
vectors, we will always write the vectors in {e} in the column vector form
{e} =
_
_
_
_
e
1
e
2
.
.
.
e
n
_
_
_
_
. (2)
We also assume that the column vector {e} obeys all the rules of matrix addition
and multiplication of a n 1 column vector.
210 JOS

E MAR

IA POZO AND GARRET SOBCZYK

In terms of these bases, any vector or point x N can be written
x = {e}x
{e}
= ( e
1
e
2
. . . e
n
)
_
_
_
_
x
1
x
2
.
.
.
x
n
_
_
_
_
=
n

i=1
x
i
e
i
(3)
for x
i
R, where
x
{e}
=
_
_
_
_
x
1
x
2
.
.
.
x
n
_
_
_
_
are the column vector of components of the vector x with respect to the basis {e}.
Since vectors in N are represented by column vectors, and vectors y N by
row vectors, we dene the operation of transpose of the vector x by
x
t
= ({e}x
{e}
)
t
= x
t
{e}
{e} = ( x
1
x
2
. . . x
n
)
_
_
_
_
e
1
e
2
.
.
.
e
n
_
_
_
_
. (4)
In the case of the complex eld K = C, we have
x

{e}
= ( x
1
x
2
. . . x
n
) =
_
_
_
_
x
1
x
2
.
.
.
x
n
_
_
_
_
t
. (5)
The transpose and Hermitian transpose operations allows us to move between the
reciprocal vector spaces N and N. Clearly the operation of Hermitian transpose
reduces to the ordinary transpose for real vectors.
We now wish to weld together the structures of the matrix algebra M(K) and
the geometric algebras generated by the vectors in the dual null spaces N and N.
Following Doran et al. (1993), we rst consider the Grassmann algebra G(N),
generated by taking all linear combinations of sums and products of the elements
in the vector space N = span{e} subject to the condition that for each x N,
x
2
= xx = 0. It follows that
(x +y)
2
= x
2
+xy +yx +y
2
= xy +yx = 0 (6)
or xy = yx for all x, y in the null space N. The geometric algebra G(N)
generated by a null space N is called the Grassmann or exterior algebra for the
null space N.
As follows from (6), the Grassmann exterior product a
1
a
2
. . . a
k
of k vectors in
N is antisymmetric over the interchange of any two of its vectors;
a
1
. . . a
i
. . . a
j
. . . a
k
= a
1
. . . a
j
. . . a
i
. . . a
k
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 211
so that the exterior product of null vectors is equivalent to the outer product of
those vectors:
a
1
a
2
. . . a
k
= a
1
a
2
a
k
.
The 2
n
-dimensional standard basis SB{e} of G(N), is generated by taking all
products of the vectors in the standard basis {e} to get
SB{e} = {1; e
1
, . . . , e
n
; e
12
, . . . , e
(n1)n
; . . . ; e
1...k
, . . . , e
(nk+1)...n
; . . . ; e
12...n
}
= {{e
0
}, {e
1
}, {e
2
}, . . . , {e
n
}}, (7)
where
{e
k
} := ( e
1...k
. . . e
(nk+1)...n
)
is the
_
n
k
_
-dimensional standard basis of k-vectors
e
j
1
j
2
...j
k
e
j
1
e
j
2
. . . e
j
k
for the
_
n
k
_
sets of indices 1 j
1
< j
2
< < j
k
n. In particular, it is assumed
{e
0
} = (1) and {e
1
} = {e}. The unique component of {e
n
} is the pseudoscalar
or volume element I := e
12...n
. With respect to the standard basis SB(e) any
multivector X G(N) can be expressed in the matrix form
X = SB{e}X
{SB}
(8)
where X
{SB}
is the column vector of components
X
{SB}
=
_
_
_
_
_
_
x
{0}
x
{e}
x
{e
2
}
.
.
.
x
{e
n
}
_
_
_
_
_
_
.
Just as we used the tranpose operation (4) to move from the the null space
N = span{e} to the dual null space N, we can extend the denition of the transpose
to enable us to move from the Grassmann algebra G(N), to the Grassmann algebra
G(N) of the reciprocal null space N. Since multivectors in G(N) are represented
by column vectors, and multivectors Y G(N) by row vectors, we dene the
transpose X
t
G(N) by
X
t
= (SB{e}X
{SB}
)
t
= X
t
{SB}
SB{e}, (9)
where X
t
{SB}
is the row vector of components
X
t
{SB}
=
_
x
t
{0}
x
t
{e}
x
t
{e
2
}
. . . x
t
{e
n
}
_
.
The Hermitian transpose is similarly dened when we are dealing with complex
multivectors.
212 JOS

E MAR

IA POZO AND GARRET SOBCZYK

The dual basis of multivectors SB{e} for G(N) are arranged in a column and
are dened by
SB{e} = (SB{e})
t
=
_
_
_
_
_
_
{1}
{e}
{e
2
}
.
.
.
{e
n
}
_
_
_
_
_
_
,
where
{e
k
} :=
_
_
e
k...1
.
.
.
e
n...nk+1
_
_
(10)
is the
_
n
k
_
-dimensional basis of dual k-vectors dened by
e
j
1
j
2
...j
k
e
j
1
e
j
2
. . . e
j
k
for the
_
n
k
_
sets of indices n j
1
> j
2
> > j
k
1.
The dual space N of the space N, and more generally the dual Grassmann
algebra G(N) of the Grassmann algebra G(N), are dened to satisfy the usual
properties of the mathematical dual space. What Doran et al. (1993) observed
was that these same properties can be faithfully expressed in a larger neutral geo-
metric algebra G
n,n
(a fomal denition is given below) containing both of these
Grassmann algebras as subalgebras, by replacing the duality conditions with corre-
sponding reciprocal conditions. We accomplish all this by assuming the additional
properties
e
2
i
= 0 = e
2
i
, e
i
e
j
= e
j
e
i
,
e
i
e
j
= e
j
e
i
(for i = j), and e
i
e
j
= e
j
e
i
, (11)
together with the reciprocal relations
e
i
e
j
=
i,j
= e
j
e
i
(12)
for all i, j = 1, 2, . . . , n. With this denition, the Grassmann algebra G(N) of the
dual space N becomes the natural reciprocal of the Grassmann algebra G(N).
These relations imply that the reciprocal k-vectors and k-forms of Grassmann
algebras G(N) and G(N) satisfy the reciprocal relations
{e
k
} {e
k
} = id
__
n
k
_

_
n
k
__
.
The neutral pseudo-Euclidean space R
n,n
is dened as the linear space which
contains both the null spaces N and N. Thus,
R
n,n
= N N = {x +y | x N, y N}.
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 213
Likewise, the 2
2n
-dimensional associative geometric algebra G
n,n
is dened to be
the geometric algebra that contains both the Grassmann algebras G(N) and G(N).
We write
G
n,n
= G(N) G(N) = gen{e
1
, . . . , e
n
, e
1
, . . . , e
n
}, (13)
subjected to the relationships (11) and (12).
A simple example will serve to show the interplay between the well-known ma-
trix multiplication and the geometric product in the super matrix algebra M(G
n,n
).
Recalling the basic geometric product of two vectors x, y,
xy = x y +x y, (14)
we apply the same product to the row and column basis vectors {e} and {e}, and
simultaneously employ matrix multiplication, to get the expressions
{e}{e} = {e} {e} +{e} {e}
= id(n n) +
_
_
_
_
_
e
1
e
1
e
1
e
2
. . . e
1
e
n
e
2
e
1
e
2
e
2
. . . e
2
e
n
. . . . . . . . . . . .
. . . . . . . . . . . .
e
n
e
1
e
n
e
2
. . . e
n
e
n
_
_
_
_
_
,
where id(nn) is the nn identity matrix, computed by taking all inner products
e
i
e
j
between the basis vectors of {e} and {e}. Similarly,
{e}{e} = {e} {e} +{e} {e} =
n

i=1
e
i
e
i
+
n

i=1
e
i
e
i
= n +
n

i=1
e
i
e
i
,
giving the useful formulas
{e} {e} = n and {e} {e} =
n

i=1
e
i
e
i
. (15)
Because of the metrical structure induced by the reciprocal relationships (12),
we can express the components x
{e}
of the vector x N (3) in the form
x
{e}
=
_
_
_
_
x
1
x
2
.
.
.
x
n
_
_
_
_
=
_
_
_
_
e
1
x
e
2
x
.
.
.
e
n
x
_
_
_
_
= {e} x.
Similarly, the components of the reciprocal vector x
t
N can be found from
x
t
{e}
= ( x
1
x
2
. . . x
n
) = x
t
( e
1
e
2
. . . e
n
) . (16)
We call G
n,n
the universal geometric algebra of order 2
2n
. When n is countably
innite, we call G = G
,
the universal geometric algebra. The universal algebra
214 JOS

E MAR

IA POZO AND GARRET SOBCZYK

G contains all of the algebras G
n,n
as proper subalgebras. In Doran et al. (1993),
G
n,n
is called the mother algebra.
2.1. NONDEGENERATE GEOMETRIC ALGEBRAS
The standard bases {e} and {e} of the reciprocal null spaces N and N, taken to-
gether, are said to make up a Witt basis of null vectors (Ablamowicz and Salingaros,
1985) of the neutral pseudo-Euclidean space R
n,n
. From the Witt basis, we can
construct the standard orthonormal basis of R
n,n
{, } of G
n,n
,

i
= e
i
+
1
2
e
i
,
i
= e
i

1
2
e
i
, (17)
for i = 1, 2, . . . , n. Using the dening relationships (12) of the reciprocal frames
{e} and {e}, we nd that these basis vectors satisfy

2
i
= 1,
2
i
= 1,
i

j
=
j

i
, i, j = 1, . . . , n,

j
=
j

i
,
i

j
=
j

i
, i = j.
The basis {} spans a real Euclidean vector space R
n
and generates the geomet-
ric subalgebra G
n,0
, whereas {} spans an anti-Euclidean space R
0,n
and generates
the geometric subalgebra G
0,n
. The standard bases (7) of these geometric algebras
naturally take the forms SB{} and SB{}, so that a general multivector X G
n,0
can be written X = SB{}X
SB
and, similarly for X G
n,0
. We can now express
the geometric algebra G
n,n
as the product of these geometric subalgebras
G
n,n
= G
n,0
G
0,n
= gen{
1
, . . . ,
n
,
1
, . . . ,
n
}, (18)
again only as linear spaces, but not as algebras.
Notice that when we write down the relationship (17), we have given up the
possibility of interpreting the vectors in {e} and {e} as column and row vectors,
respectively. When working in the nondegenerate geometric algebras G
n,n
, G
n,0
or
G
0,n
, we use the operation of reversal. The reversal of any vector x G
n,n
is dened
by x

:= x, and for the k-vector A

k
= a
1
a
2
a
k
,
A

k
:= a
k
a
k1
a
1
= (1)
k(k1)/2
A
k
.
2.2. SPINOR BASIS
One nice application of the above formalism is that it allows us to simply express
a natural isomorphism that exists between the neutral geometric algebra G
n,n
and
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 215
the algebra of all real 2
n
2
n
matrices M
R
(2
n
). To express this isomorphism, we
rst dene 2
n
mutually commuting idempotents
u
i
() =
1
2
(1
i

i
) (19)
for i = 1, 2, . . . , n.
We can now dene 2
n
mutually annihiliating primitive idempotents for the
algebra G
n,n
,
u
signs
=

signs
u
i
(sign s
i
), (20)
where signs is a particular sequence of n signs, and sign s
i
is the ith sign in the
sequence. For example,
u
++++
=
n

i=1
u
i
(+) and u

=
n

i=1
u
i
().
The primitive idempotents satisfy the following basic properties:

2
n
signs
u
signs
= 1,

i
u
++++
= u
++
i
++

i
,
u
sign
1
u
sign
2
=
sign
1
sign
2
u
sign
1
, where
sign
1
sign
2
= 0,
except when sign
1
= sign
2
for which
sign
1
sign
2
= 1.
The above properties are easily veried.
In contrast to the standard basis SB{e} of the neutral geometric algebra G
n,n
, the
spinor basis of G
n,n
is dened to be the 2
n
2
n
multivectors in the matrix
SNB(n, n) = SB{}
t
u
++++
SB{}. (21)
The simplest example is the spinor basis for the geometric algebra G
1,1
. The 2
1
primitive idempotents for this geometric algebra are u

=
1
2
(1 ). Using (21),
the spinor basis SNB(1, 1) is found to be
SB()
t
u
+
SB() =
_
1

_
u
+
( 1 ) =
_
u
+
u

u
+
u

_
.
The signicance of the position of each multivector in the spinor basis, i.e. its
matrix representation corresponds to 1 in the same position (with zeros everywhere
else).
In terms of the spinor basis, any 2
n
2
n
matrix Arepresents the corresponding
element A G
n,n
given by
A = SB{}u
++++
ASB{}
t
.
The matrix A associated with the multivector A G
n,n
is denoted by A = [A].
This association constitutes an algebra isomorphism, since [A + B] = [A] + [B]
and [AB] = [A][B]. Noting that
u
++
SB{}
t
SB{}u
++
= u
++
id(2
n
2
n
),
216 JOS

E MAR

IA POZO AND GARRET SOBCZYK

it easily follows that
AB = SB{}u
++
[A]SB{}
t
SB{}u
++
[B]SB{}
t
= SB{}u
++
[A][B]SB{}
t
. (22)
We will use the spinor basis SNB(1, 1) for studying conformal transformation
in Section 4.
2.3. SYMMETRIC AND HERMITIAN INNER PRODUCTS
Until now we have only considered real geometric algebras and their corresponding
real matrices. Any pseudoscalar of the geometric algebra G
n,n
will always have
a positive square and will anticommute with the vectors in R
n,n
. If we insisted
on dealing only with real geometric algebras, we might consider working in the
geometric algebra G
n,n+1
where the pseudoscalar element i has the desired property
that i
2
= 1 and is in the center of the algebra (commutes with all multivectors). A
complex vector x+iy in G
n,n+1
consists of the real vector part x and a pseudovector
or (2n)-blade part iy.
Instead, we choose to directly complexify the geometric algebra G
n,n
to get the
complex geometric algebra G
2n
(C) (Sobczyk, 1996). Whereas this algebra is iso-
morphic to G
n,n+1
, it is somewhat easier to work with than the former. A complex
vector z C
2n
has the form z = x + iy where x, y R
2n
. The imaginary unit i,
where i
2
= 1, is dened to commute with all elements in the geometric algebra
G
2n
(C).
Consider an orthonormal basis {} C
2n
:
i

j
=
ij
. The complexied null
space N(C) and its reciprocal null space N(C) are the subspaces spanned by the
complex null vectors
e
j
=
1
2
(
j
+i
n+j
) and e
j
=
j
i
n+j
for j = 1, 2, . . . , n. This denition is consistent with (17) if we consider
j
=
i
n+j
. Thus, a null vector x N(C) has the form x = {e}x
{e}
for x
i
C.
Previously we have dened the transposition (4). This operation can be extended
to complex vectors in two different ways. The rst way is a linear extension. We
use the term transposition for the linear extension, so that denition (4) is still valid
when x
i
C. The second extension is antilinear and is equivalent to Hermitian
conjugation: x

{e}
{e}. Both operations, Hermitian conjugation and transposi-
tion, take us from the complex null space N(C) to the dual null space N(C), and
if the components of x are all real, both reduce to the real transposition. Applied
to the components x
{e}
, x

{e}
is the usual Hermitian transpose of the column vector
x
{e}
,
x

{e}
= ( x
1
x
2
. . . x
n
) =
_
_
_
_
x
1
x
2
.
.
.
x
n
_
_
_
_
t
. (23)
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 217
We now dene the symmetric inner product (x, y), and the Hermitian inner
product x, y, on N(C). For all x, y N(C), the two products are dened,
respectively, by using transposition and Hermitian conjugation:
(x, y) := x
t
y = x
t
{e}
y
{e}
and x, y := x

y = x

{e}
y
{e}
. (24)
The Hermitian inner product will be used in the next subsection.
2.4. LINEAR TRANSFORMATIONS
Let NN

and NN

be (n+n

)-dimensional reciprocal null spaces in R

n+n

,n+n

with the dual bases {e}{e

} and {e}{e

}. Let f : N N

be a linear transforma-
tion from the null space N into the null space N

. In light of the previous section,

we can consider the null spaces N and N

to be over the real or complex numbers.

Let
Hom(N, N

) = {f : N N

| f is a linear transformation}
denote the linear space of all homomorphisms from N to N

, with the usual opera-

tion of addition of transformations. Of course, only when N = N

is the operation
of multiplication (composition) dened.
Given an operator f Hom(N, N

), y

= f (x) f x, the matrix F of f with

respect to the bases {e} and {e

} is dened by
f {e} (f e
1
. . . f e
n
) = (e

1
. . . e

n
)F = {e

}F . (25)
Of course, the matrix F = (f
ij
) is dened by its n

n components f
ij
= e
i

f (e
j
) C for i = 1, 2, . . . , n

and j = 1, 2, . . . , n. It follows that f (e

j
) =

i=1
e

i
f
ij
. By dotting both sides of the above equation on the left by {e

}, we nd
the explicit expression
F = {e

} {e

}F = {e

} f {e}.
Equation (15) can be used to dene the bivector F of the linear operator f . It
is dened by F = f {e} {e

} and satises the property that f x = F x for all

x N. The bivector of a linear operator makes possible a new theory of linear
operators, and is particularly useful in dening the general linear group as a Lie
group of bivectors with the commutator product (Fulton and Harris, 1991; Bayro
and Sobczyk, 2001, p. 32).
Given the Hermitian inner product (24), the transpose (or Hermitian transpose
(23)) f

: N

N of the mapping f : N N

is dened by the requirement that

for all x N and y

,
x, f

(y

) = f (x), y

{e} f

{e

} = [f {e}]

{e

}.
Likewise, we can dene the transpose relative to the symmetric inner product
(x, f
t
(y

)) = (f (x), y

) F
t
{e} f
t
{e

} = [f {e}]
t
{e

}.
218 JOS

E MAR

IA POZO AND GARRET SOBCZYK

2.5. OUTERMORPHISM AND GENERALIZED TRACES
A linear transformation f naturally extends multilinearly to act on k-blades,
f (x y z) f (x) f (y) f (z) x, y, . . . , z N,
and where f (1) 1. Thus extended, f : G(N) G(N

) is called the outermor-

phism of the linear transformation f : N N

, since it preserves the structure of

the outer product:
f (A B +C D) = f (A) f (B) +f (C) f (D)
A, B, C, D G(N).
Geometrically, the outermorphism f maps directed areas into directed areas, and
more generally, directed k-vectors into directed k-vectors.
A linear transformation from N into itself is called an endomorphism. Let
End(N) = {f : N N | f is a linear operator}
denote the algebra of all endomorphisms on N. The operations of addition and
composition of linear operators is well dened for endomorphisms.
The determinant det f of the endomorphism f is dened to be the eigenvalue
of the pseudoscalar element I = e
12...n
:
f (I) = det f I det f = f (I) I.
Thus, det f is the factor by which volume is scaled by f . The trace of f is dened
by tr f := f {e} {e}. Given the outermorphism of f , we dene the generalized
traces of f by
tr
i
f := f {e
i
} {e
i
}.
Particular cases are tr
0
f = f (1) 1 = 1 and tr
1
f = tr f . The generalized trace of
degree n coincides with the determinant: tr
n
f = f (I) I = det f .
A second basis {a} of N is related to the standard basis {e} by the application
of some endomorphism a
{a} = a{e} = {e}A = ( e
1
e
2
. . . e
n
) A, (26)
where Ais called the matrix of transition from the basis {e} to the basis {a}. Taking
the outer product
_
n
i=1
{a} of the basis vectors {a}, we get
n

i=1
{a} a
1
a
2
a
n
= a(e
1
e
2
e
n
) = det aI. (27)
We see from (27) that the determinant of the matrix of transition, det A det a,
between two bases cannot be zero.
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 219
We can now easily construct a dual or reciprocal basis {a} for the basis {a}:
a
i
= (1)
i+1
(a
1
i

a
n
) I
det a
, (28)
where i

means that a
i
is omitted from the product. More compactly, using our
matrix notation,
{a} =
a({e} I) I
a(I) I
.
Checking, we nd that
{a} {a} =
[a({e} I) I] {a}
a(I) I
=
[a(I {e}) a{e}] I
a(I) I
=
[a((I {e}) {e})] I
a(I) I
=
a(I ({e} {e})) I
a(I) I
= {e} {e} = id.
We have actually found the inverse of the transition matrix A, given by A
1
=
{a} {e} (Bayro and Sobczyk, 2001, p. 25).
2.6. CHARACTERISTIC POLYNOMIAL
The characteristic polynomial of f : N N is dened by

f
() = det( f ) = ( f )(I) I.
The well-known CaleyHamilton theorem, which says that every linear operator
satises its characteristic equation, is a consequence of the identity
f [x {e
n1
}] {e
n1
} = (x {e
n1
}) {e
n1
} det f = x det f. (29)
When the left side of this identity is expanded, we get
f [x {e
n1
}] {e
n1
}
=
n

i=1
(1)
i+1
f {e
ni
} {e
ni
}f
i
(x)
=
n

i=1
(1)
i+1
tr
ni
ff
i
(x). (30)
Expressed in terms of the generalized traces of f , the characteristic polynomial is

f
() = ( f )(e
12...n
) e
n...21
=
n

i=0
(1)
i
f {e
i
} {e
i
}
ni
.
220 JOS

E MAR

IA POZO AND GARRET SOBCZYK

Thus, from (29) and (30), we have
f
(f ) = 0, i.e. f satises its characteristic
polynomial.
Equation (29) can also be used to derive a formula for the inverse of f . We get
x = f
1
(y) =
(y f {e
n1
}) {e
n1
}
det f
.
The minimal polynomial
f
() of f is the polynomial of least degree that has
the property that
f
(f ) = 0. Taken over the complex numbers C, we can express

f
and
f
in the factored form

f
() =
r

i=1
(
i
)
n
i
and
f
() =
r

i=1
(
i
)
m
i
,
where 1 m
i
n
i
n for i = 1, 2, . . . , r, and the roots
i
are all distinct.
The minimal polynomial uniquely determines, up to an ordering of the idem-
potents, the following spectral decomposition theorem of the linear operator f
(Sobczyk, 2001).
THEOREM 1. If f has the minimal polynomial (), then a set of commuting
mutually annihilating idempotents and corresponding nilpotents {(p
i
, q
i
) | i =
1, . . . , r} can be found such that f =

r
i=1
(
i
+q
i
)p
i
, where rank(p
i
) = n
i
, and
the of nilpotency index(q
i
) = m
i
, for i = 1, 2, . . . , r. Furthermore, when m
i
= 1,
q
i
= 0.
Clearly, the operator f is diagonalizable if and only if it has the spectral form
f =
r

i=1

i
p
i
.
The spectral decomposition theorem has many different uses and applies equally
well to a linear operator or a geometric number (Sobczyk, 1993, pp. 357364,
1997a, 1997b). For example, we can immediately dene a generalized inverse of
the operator f by
f
inv
=

i
=0
1

i
_
p
i

q
i

i
+ +
_
q
i

i
_
m
i
1
_
satisfying the conditions ff
inv
= f
inv
f =

i
=0
p
i
(Rao and Mitra, 1971, p. 20).
3. Geometric Algebra and Non-Euclidean Geometry
Leonardo da Vinci (14521519) was one of the rst to consider the problems of
projective geometry. However, projective geometry was not formally developed
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 221
until the work Trait des propris projectives des gure of the French mathe-
matician Poncelet (17881867), published in 1822. The extraordinary generality
and simplicity of projective geometry led the English mathematician Cayley to
exclaim: Projective geometry is all of geometry (Young, 1930).
Let R
n+1
be an (n+1)-dimensional Euclidean space and let G
n+1,0
be the corre-
sponding geometric algebra. The directions or rays of nonzero vectors in R
n+1
are
identied with the points of the n-dimensional projective plane
n
(Hestenes and
Ziegler, 1991). More precisely, we write
n
R
n+1
/R

, where R

= R{0}. We
thus identify points, lines, planes, and higher-dimensional k-planes in
n
with 1,
2, 3, and (k + 1)-dimensional subspaces S
r
of R
n+1
, where k n. To effectively
apply the tools of geometric algebra, we need to introduce the new basic operations
of meet and join (Bayro and Sobczyk, 2001, p. 27).
3.1. THE MEET AND JOIN OPERATIONS
The meet and join operations of projective geometry are most easily dened in
terms of the intersection and direct sum of the subspaces which name the objects
in
n
. On the other hand, each r-dimensional subspace A
r
can be described by
a nonzero r-blade A
r
G(R
n+1
). We say that an r-blade A
r
represents, or is a
representant of an r-subspace A
r
of R
n+1
if and only if
A
r
= {x R
n+1
| x A
r
= 0}. (31)
We denote the equivalence class of all nonzero r-blades A
r
G(R
n+1
) which
dene the subspace A
r
by
{A
r
}
ray
:= {t A
r
| t R, t = 0}. (32)
Evidently, every r-blade in {A
r
}
ray
is a representant of the subspace A
r
. With these
denitions, the problem of nding the meet and join is reduced to a problem in
geometric algebra of nding the corresponding meet and join of the (r + 1)- and
(s +1)-blades in the geometric algebra G(R
n+1
) which represent these subspaces.
Let A
r
, B
s
and C
t
be nonzero blades representing the three subspaces A
r
, B
s
and C
t
, respectively. We say that
DEFINITION 1. The t -blade C
t
= A
r
B
s
is the meet of A
r
and B
s
if there exists
a complementary (r t )-blade A
c
and a complementary (s t )-blade B
c
with the
property that A
r
= A
c
C
t
, B
s
= C
t
B
c
, and A
c
B
c
= 0.
It is important to note that the t -blade C
t
{C
t
}
ray
is not unique and is dened
only up to a nonzero scalar factor, which we choose at our own convenience. The
existence of the t -blade C
t
(and the corresponding complementary blades A
c
and
B
c
) is an expression of the basic relationships that exists between subspaces.
DEFINITION 1. 1. The (r +s t )-blade D = A
r
B
s
, called the join of A
r
and
B
s
is dened by D = A
r
B
s
= A
r
B
c
.
222 JOS

E MAR

IA POZO AND GARRET SOBCZYK

Alternatively, since the join A
r
B
s
is dened only up to a nonzero scalar factor,
we could equally well dene D by D = A
c
B
s
. We use the symbols intersection
and direct sum from set theory to mark this unusual state of affairs. The problem
of meet and join has thus been solved by nding the direct sum and intersection
of linear subspaces and their (r +s t )-blade and t -blade representants.
Note that it is only in the special case when A
r
B
s
= 0 that the join can be
considered to reduce to the outer product, i.e.
A
r
B
s
= 0 A
r
B
s
= A
r
B
s
.
However, after the join I
A
r
B
s
A
r
B
s
has been found, it can be used to nd the
meet A
r
B
s
,
A
r
B
s
= A
r
[B
s
I
A
r
B
s
] = [I
A
r
B
s
A
r
] B
s
. (33)
While the positive denite metric of R
n+1
is irrelevant to the denition of the meet
and join of subspaces, formula (33) holds only in R
n+1
.
A slightly modied version of this formula will hold in any nondegenerate
pseudo-Euclidean space R
p,q
, where p + q = n + 1. In this case, after we have
found the join I
A
r
B
s
, which is a (r +k)-blade, we nd a reciprocal (r +k)-blade
I
A
r
B
s
with the property that I
A
r
B
s
I
A
r
B
s
= 0. The meet A
r
B
s
may then be
dened by
A
r
B
s
= A
r
[B
s
I
A
r
B
s
] = [I
A
r
B
s
A
r
] B
s
. (34)
3.2. AFFINE AND PROJECTIVE GEOMETRIES
We have seen in the previous section how the meet and join of the n-dimensional
projective space
n
can be dened in an (n + 1)-dimensional Euclidean space
R
n+1
. There is a very close connection between afne and projective geometries.
A projective space can be considered to be an afne space with idealized points
at innity (Young, 1930). Since all the formulas for meet and join remain valid
in the pseudo-Euclidean space R
p,q
, subject only to (34), we will dene the n =
(p +q)-dimensional afne plane A
e
(R
p,q
) of the null vector e =
1
2
( +) in the
larger pseudo-Euclidean space R
p+1,q+1
= R
p,q
R
1,1
, where R
1,1
= span{, }
for
2
= 1 =
2
. Whereas, effectively, we are only extending the Euclidean
space R
p,q
by the null vector e, it is advantageous to work in the geometric algebra
G
p+1,q+1
of the nondegenerate pseudo-Euclidean space R
p+1,q+1
.
The afne plane A
p,q
e
:= A
e
(R
p,q
) is dened by
A
e
(R
p,q
) = {x
h
= x +e | x R
p,q
} R
p+1,q+1
, (35)
for the null vector e R
1,1
. The afne plane A
e
(R
p,q
) has the nice property that
x
2
h
= x
2
for all x
h
A
e
(R
p,q
), thus preserving the metric structure of R
p,q
. By
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 223
employing the reciprocal null vector e = with the property that e e = 1, we
can restate denition (35) of A
e
(R
p,q
) in the form
A
e
(R
p,q
) = {y | y R
p+1,q+1
, y e = 1 and y e = 0} R
p+1,q+1
.
This form of the denition is interesting because it brings us closer to the denition
of the n = (p +q)-dimensional projective plane.
We summarize here the important properties of the reciprocal null vectors e =
1
2
( + ) and e = that will be needed later, and their relationship to the
hyperbolic unit bivector u := .
e
2
= e
2
= 0, e e = 1, u = e e = , u
2
= 1. (36)
The projective n-plane
n
can be dened to be the set of all points of the afne
plane A
e
(R
p,q
), taken together with idealized points at innity. Each point x
h

A
e
(R
p,q
) is called a homogeneous representant of the corresponding point in
n
because it satises the property that x
h
e = 1. To bring these different viewpoints
closer together, points in the afne plane A
e
(R
p,q
) will also be represented by rays
in the space
A
rays
e
(R
p,q
)= {{y}
ray
| y R
p+1,q+1
, y e = 0, y e = 0}
R
p+1,q+1
. (37)
The set of rays A
rays
e
(R
p,q
) gives another denition of the afne n-plane, because
each ray {y}
ray
A
rays
e
(R
p,q
) determines the unique homogeneous point
y
h
=
y
y e
A
e
(R
p,q
).
Conversely, each point y A
e
(R
p,q
) determines a unique ray {y}
ray
in A
rays
e
(R
p,q
).
Thus, the afne plane of homogeneous points A
e
(R
p,q
) is equivalent to the afne
plane of rays A
rays
e
(R
p,q
).
Suppose that we are given k-points a
h
1
, a
h
2
, . . . , a
h
k
A
e
(R
p,q
) where each a
h
i
=
a
i
+ e for a
i
R
p,q
. Taking the outer product or join of these points gives the
projective (k 1)-plane A
h

n
. Expanding the outer product gives
A
h
= a
h
1
a
h
2
a
h
k
= a
h
1
(a
h
2
a
h
1
) a
h
3
a
h
k
= a
h
1
(a
h
2
a
h
1
) (a
h
3
a
h
2
) a
h
4
a
h
k
=
= a
h
1
(a
2
a
1
) (a
3
a
2
) (a
k
a
k1
),
or
A
h
= a
h
1
a
h
2
a
h
k
= a
1
a
2
a
k
+
+e (a
2
a
1
) (a
3
a
2
) (a
k
a
k1
). (38)
224 JOS

E MAR

IA POZO AND GARRET SOBCZYK

Whereas (38) represents a (k 1)-plane in
n
, it also belongs to the afne
(p, q)-plane A
p,q
e
, and thus contains important metrical information. Dotting this
equation with e, we nd that
e A
h
= e (a
h
1
a
h
2
a
h
k
)
= (a
2
a
1
) (a
3
a
2
) (a
k
a
k1
).
This result motivates the following
DEFINITION 1.1. 1. The directed content of the (k 1)-simplex A
h
= a
h
1
a
h
2

a
h
k
in the afne (p, q)-plane is given by
e A
h
(k 1)!
=
e (a
h
1
a
h
2
a
h
k
)
(k 1)!
=
(a
2
a
1
) (a
3
a
2
) (a
k
a
k1
)
(k 1)!
.
3.3. EXAMPLES
Many incidence relations can be expressed in the afne plane A
e
(R
p,q
) which are
also valid in the projective plane
n
(Bayro and Sobczyk, 2001, p. 263). A few
examples are provided below.
Given 4 coplanar points a
h
, b
h
, c
h
, d
h
A
e
(R
2
). The join and meet of the lines
a
h
b
h
and c
h
d
h
are given, respectively, by (a
h
b
h
) (c
h
d
h
) = a
h
b
h
c
h
,
and using (34)
(a
h
b
h
) (c
h
d
h
) = [I (a
h
b
h
)] (c
h
d
h
),
where I =
2

1
e. Carrying out the calculations for the meet and join, we nd
that
(a
h
b
h
) (c
h
d
h
) = det{a
h
, b
h
, c
h
}I = det{a, b}I, (39)
where I =
1

2
e, and
(a
h
b
h
) (c
h
d
h
) = det{c d, b c}a
h
+det{c d, c a}b
h
. (40)
Note that the meet (40) is not, in general, a homogeneous point. Normaliz-
ing (40), we nd the homogeneous point p
h
A
e
(R
2
)
p
h
=
det{c d, b c}a
h
+det{c d, c a}b
h
det{c d, b a}
,
which is the intersection of the lines a
h
b
h
and c
h
d
h
, see Figure 1. The meet
can also be solved for directly in the afne plane by noting that
p
h
=
p
a
h
+(1
p
)b
h
=
p
c
h
+(1
p
)d
h
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 225
Figure 1. Incidence relationships in the afne plane.
and solving to get
p
= det{b
h
, c
h
, d
h
}/ det{b
h
a
h
, c
h
, d
h
}.
Given the line a
h
b
h
A
e
(R
2
) and a third point d
h
A
e
(R
2
), as in Figure 1,
the point f
h
on the line a
h
b
h
which is closest to the point d
h
is called the foot
of the point d
h
on the line a
h
b
h
. Since f
h
a
h
b
h
= 0, it follows that f
h
=

f
a
h
+(1
f
)b
h
and f
h
b
h
=
f
a
h
b
h
. We can solve this last equation for
f
by dotting it with e, and invoking the auxilliary condition that (bf )(df ) = 0.
We get

f
=
(a b) (d b)
(a b)
2
. (41)
It should be carefully noted that a
h
b
h
= a b R
2
for any two homogeneous
points a
h
, b
h
A
2
e
. It follows that the foot f
h
on the line a
h
b
h
is given by
f
h
=
(b d) (b a)a
h
+(a d) (a b)b
h
(a b)
2
. (42)
Saying that a
h
, b
h
, c
h
A
2
e
are noncollinear points is equivalent to the condi-
tion a
h
b
h
c
h
= 0. If d
h
is any other point in A
2
e
, then d
h
a
h
b
h
c
h
= 0 so
that
d
h
=
d
a
h
+
d
b
h
+(1
d

d
)c
h
.
By wedging this last equation by b
h
c
h
and a
h
c
h
, respectively, we can easily
solve for
d
and
d
, getting

d
=
det{d
h
, b
h
, c
h
}
det{a
h
, b
h
, c
h
}
and
d
=
det{d
h
, c
h
, a
h
}
det{a
h
, b
h
, c
h
}
. (43)
Three noncollinear points a
h
, b
h
, c
h
A
2
e
determine a unique circle with center
r
h
=
r
a
h
+
r
b
h
+ (1
r

r
)c
h
. To nd the center, note that r
h
lies on the
intersection of the perpendicular bisectors of the cords a
h
b
h
and a
h
c
h
, and
therefore satises
r
h
=
1
2
(a
h
+b
h
) +s(c
h
w
h
) =
1
2
(a
h
+c
h
) +t (b
h
q
h
), (44)
226 JOS

E MAR

IA POZO AND GARRET SOBCZYK

where
w
h
= f
w
a
h
+(1 f
w
)b
h
and q
h
= f
q
a
h
+(1 f
q
)c
h
are the feet (42) of c
h
and b
h
along the lines a
h
b
h
and a
h
c
h
, respectively, for
f
w
=
(a b) (c b)
(a b)
2
and f
q
=
(c a) (b a)
(c a)
2
.
From (44), it follows that
(f
w
s f
q
t )a
h
+[t +(1 f
w
)s
1
2
]b
h
+[
1
2
(1 f
q
)t s]c
h
0,
which gives
s =
f
q
2f
q
(1 f
w
) +2f
w
and t =
f
w
2f
q
(1 f
w
) +2f
w
.
After simplication, the center r
h
is found to be
r
h
=
[f
q
+f
w
2f
b
f
w
]a
h
+f
w
b
h
+f
b
c
h
2[f
q
+f
w
f
b
f
w
]
. (45)
Another theorem of interest is Simpsons theorem for the circle. We have assem-
bled all of the tools necessary for a proof of this venerable theorem in the afne
plane A
e
(R
2
), but we will not prove it here (Bayro and Sobczyk, 2001, p. 39).
Simpsons theorem has also been proven in the nonlinear horosphere (Li et al.,
2000), but the proof is not trivial. It remains to be seen if there are any real ad-
vantages to proving such theorems on the horosphere and not in the simpler afne
plane. The issue at hand is how to best represent problems in distance geometry
(Dress and Havel, 1993).
Hestenes and Zigler have also given a proof of Desargues theorem in the pro-
jective plane
2
(Hestenes and Ziegler, 1991), by using its representation in the
Euclidean space R
3
. A proof of Desargues theorem can also be given in the afne
plane of rays A
rays
e
(R
p,q
), (Bayro and Sobczyk, 2001, p. 37). The importance of
such proofs is that even though geometric algebra is endowed with a metric, there
is no reason why we cannot use the tools of Euclidean space to give a proof of this
metric independent result. Indeed, as has been emphasized by Hestenes and others
(Barnabei et al., 1985), all the results of linear algebra can be supplied with such
a projective interpretation.
4. Conformal Geometry
The conformal geometry of a pseudo-Euclidean space can be linearized by con-
sidering the horosphere in a pseudo-Euclidean space of two dimensions higher.
Because it is so easy to introduce extra orthogonal anticommuting vectors into a
geometric algebra, without altering the structure of the geometric algebra in any
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 227
other way, the framework of geometric algebra offers a unication to the subject
that is impossible in other formalisms. The horosphere has recently attracted the
attention of many workers, see for example (Dress and Havel, 1993; Porteous,
1995; Havel, 1995).
The horosphere and null cone are formally introduced in Subsections 4.1 and 4.2.
In Subsection 4.3, the concept of an h-twistor is introduced which will greatly sim-
plify computations. An h-twistor is a generalization of the Penrose twistor concept.
In Subsection 4.4, we give a simple proof, using only basic concepts from differ-
ential geometry developed in (Hestenes and Sobczyk, 1984), of an intriging result
that relates conformal transformations in a pseudo-Euclidean space to isometries
in a pseudo-Euclidean space of two higher dimensions. The original proof of this
striking relationship was given by Haantjes (1937). In Subsection 4.5, we show
that for any dimension greater than two, that any isometry on the null cone can be
extended to all of the pseudo-Euclidean space.
In Subsections 4.6 and 4.7, we show the beautiful relationships that exists be-
tween Mobius transformations (linear fractional transformations) and their 2 2
matrix representation over a suitable geometric algebra. In a nal subsection, we
explore how all of the formalism developed in the previous sections can be utilized
in the characterization of conformal transformations of the pseudo-Euclidean space
R
p,q
. We develop the theory in a novel way which suggests a nontrivial general-
ization of the theory of two-component spinors and 4-component twistors. Recall
that a conformal transformation preserves angles between tangent vectors at each
point (Lounesto and Springer, 1989; Porteous, 1995). The utility of the h-twistor
concept is amply demonstrated in a new derivation of the Schwarzian derivative.
We begin by dening the horosphere H
p,q
e
in R
p+1,q+1
by moving up from the
afne plane A
p,q
e
:= A
e
(R
p,q
).
4.1. THE HOROSPHERE
Let G
p+1,q+1
= gen(R
p+1,q+1
) be the geometric algebra of R
p+1,q+1
, and recall
the denition (35) of the afne plane A
p,q
e
:= A
e
(R
p,q
) R
p+1,q+1
. Any point
y R
p+1,q+1
can be written in the form y = x + e + e, where x R
p,q
and
, R.
The horosphere H
p,q
e
is most directly dened by
H
p,q
e
:= {x
c
= x
h
+e | x
h
A
p,q
e
and x
2
c
= 0}. (46)
With the help of (36), the condition that
x
2
c
= (x
h
+e)
2
= x
2
+2 = 0
gives us immediately that := x
2
/2. Thus each point x
c
H
p,q
e
has the form
x
c
= x
h

x
2
h
2
e = x +e
x
2
2
e =
1
2
x
h
ex
h
. (47)
228 JOS

E MAR

IA POZO AND GARRET SOBCZYK

The last equality on the right follows from
1
2
x
h
ex
h
=
1
2
[(x
h
e)x
h
+(x
h
e)x
h
] = x
h

1
2
x
2
h
e.
Just as x
h
A
p,q
e
is called the homogeneous representant of x R
p,q
, the
point x
c
is called the conformal representant of both the points x
h
A
p,q
e
and
x R
p,q
. The set of all conformal representants H
p,q
:= c(R
p,q
) is called the
horosphere. The horosphere H
p,q
is a nonlinear model of both the afne plane
A
p,q
e
and the pseudo-Euclidean space R
p,q
. The horosphere H
n
for the Euclidean
space R
n
was rst introduced by F. A. Wachter, a student of Gauss (Havel, 1995),
and has been recently nding many diverse applications (Bayro and Sobczyk, 2001,
Chapters 1, 4, 6).
Dening the bivector K
x
:= e x
c
= e x
h
, it is easy to get back x
h
by the
simple projection,
x
h
= e K
x
(48)
and to x R
p,q
, by
x = u (u x
c
) = e (e x
h
), (49)
using the bivector u dened in (36).
The set of all null vectors y R
p+1,q+1
make up the null cone
N := {y R
p+1,q+1
| y
2
= 0}.
The subset of N containing all the representants y {x
c
}
ray
for any x R
p,q
is
dened to be the set
N
0
= {y N | y e = 0} =
_
xR
p,q
{x
c
}
ray
,
and is called the restricted null cone. The conformal representant of a null ray {z}
ray
is the representant y {z}
ray
which satises y e = 1. The horosphere H
p,q
is the
parabolic section of the restricted null cone,
H
p,q
= {y N
0
| y e = 1},
see Figure 2. Thus H
p,q
has dimension n = p +q.
The null cone N is determined by the condition y
2
= 0, which taking differen-
cials gives
y dy = 0 x
c
dy = 0, (50)
where {y}
ray
= {x
c
}
ray
. Since N
0
is an (n + 1)-dimensional surface, then (50) is a
condition necessary and sufcient for a vector v to belong to the tangent space to
the restricted null cone T (N
0
) at the point y
v T (N
0
) x
c
v = 0. (51)
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 229
Figure 2. The restricted null cone and representants of the point x in afne space and on the
horosphere.
It follows that the (n +1)-pseudoscalar I
y
of the tangent space to N
0
at the point y
can be dened by I
y
= Ix
c
where I is the pseudoscalar of R
p+1,q+1
. We have
x
c
v = 0 0 = I (x
c
v) = (Ix
c
) v = I
y
v. (52)
4.2. THE NULL CONE
The mapping
c: R
p,q
N
0
R
p+1,q+1
, x c(x) x
c
(53)
is continuous and innitely differentiable (indeed, its third differential vanishes),
and it is also an isometric embedding.
dx
c
= dx x dxe ( dx
c
)
2
= ( dx)
2
. (54)
The mapping c(x) (53) constitutes a vectorial chart for the horosphere. The pseudo-
scalar I
x
c
of the tangent space to H
p,q
at the point x
c
is given by
I
x
c
= IK
x
= Ie x
c
. (55)
We can extend the mapping c(x) to give a scalar-vector chart for the whole N
0
.
y: R
p,q
R

N
0
, (x, t ) y(x, t ) t c(x) = t x
c
. (56)
4.3. H-TWISTORS
Let us dene the h-twistor to be a rotor S
x
Spin
p+1,q+1
230 JOS

E MAR

IA POZO AND GARRET SOBCZYK

S
x
:= 1 +
1
2
xe = exp(
1
2
xe). (57)
Noting that S
x
S

x
= 1, we dene its angular velocity by

S
:= 2S

x
dS
x
= dxe or equivalently
S
(a) = ae a R
p,q
. (58)
Later, in Section 4.7, we more carefully dene an h-twistor to be an equivalence
class of two twistor components from G
p,q
, that have many twistor-like proper-
ties.
The reason for these denitions are found in their properties. The point x
c
is
generated from 0
c
= e by
x
c
= S
x
eS

x
, (59)
and the tangent space to the horosphere at the point x
c
is generated from dx R
p,q
by
dx
c
= dS
x
eS

x
+S
x
e dS

x
= S
x
(
S
e)S

x
= S
x
dxS

x
(60)
or, equivalently, in terms of the argument of the differential
dx
c
(a) = S
x
aS

x
a R
p,q
.
It also keeps unchanged the point at innity e = S
x
eS

x
.
The motivation for the term h-twistor is that it generates both points and tan-
gent vectors on the horosphere from the corresponding objects in R
p,q
. We call the
h-twistor (60) nonrotational because tangent vectors coincide with the differen-
tial of points. More generally, the h-twistor T
x
:= S
x
R
x
, with R
x
Spin(R
p,q
)
generates
x
c
= T
x
eT

x
= S
x
eS

x
and dx
c
(R
x
aR

x
) = T
x
aT

x
.
The angular velocity
T
of the more general h-twistor T
x
is easily calculated

T
:= 2T

x
dT
x
= R

S
R
x
+
R
= R

x
dxR
x
e +
R
. (61)
The analogy with Penrose twistors is, of course, not complete. We will have more
to say about this later.
4.4. CONFORMAL TRANSFORMATIONS AND ISOMETRIES
In this subsection we show that every conformal transformation in R
p,q
corre-
sponds to two isometries on the null cone N
0
in R
p+1,q+1
.
DEFINITION 1. A conformal transformation in R
p,q
is any twicely differentiable
mapping between two connected open subsets U and V,
f : U V, x x

= f (x)
such that the metric changes by only a conformal factor
(df (x))
2
= (x)(dx)
2
, (x) = 0.
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 231
If p = q then (x) > 0. In the case p = q, there exists the posibility that
(x) < 0, when the conformal transformations belong to two disjoint subsets. We
will only consider the case when (x) > 0.
Recall that N
0
can be coordinized by the vector-scalar chart (56). Using the
h-twistor (57), (59) and (60), we obtain the expressions
y = S
x
t eS

x
and dy = dt x
c
+t dx
c
= dt S
x
eS

x
+t S
x
dxS

x
. (62)
It easily follows that
(dy)
2
= t
2
(dx
c
)
2
= t
2
(dx)
2
. (63)
DEFINITION 1. 1. An isometry F on N
0
is any twicely differentiable mapping
between two connected open subsets U
0
and V
0
in the relative topology of N
0
,
F: U
0
V
0
, y y

= F(y)
which satises (dF(y))
2
= (dy)
2
.
Using the scalar-vector chart y(x, t ) = t x
c
, any mapping in N
0
can be expressed
in the form
y

= F(y) = t

c
= (x, t )f (x, t )
c
,
where t

= (x, t ) and x

c
= f (x, t )
c
are dened implicitly by F. Using (63), we
obtain the result that y

= F(y) is an isometry if and only if

(dy

)
2
= (dy)
2
t

2
(dx

)
2
= t
2
(dx)
2
(df (x, t ))
2
=
t
2
(x, t )
2
(dx)
2
.
Since f (x, t ), x R
p,q
(nondegenerate metric), and the right-hand side of this
equation does not contain dt , it follows that f (x, t ) = f (x) is independent of t . It
then follows that (x) := (x, t )/t is also independent of t . Thus, we can express
any isometry y

= F(y) in the form y

= t (x)f (x)
c
, where f (x)
c
N
0
is the
conformal representant of f (x) R
p,q
. This implies that y

= F(y) is an isometry
iff
y

= t (x)f (x)
c
and (df (x))
2
= ((x))
2
(dx)
2
.
Therefore, f (x) is a conformal transformation with
(x) = (x)
2
> 0 (x) =
1

(x)
.
4.5. ISOMETRIES IN N
0
In this section we show that for any dimension greater than 2 any isometry in N
0
is
the restriction of an isometry in R
p+1,q+1
. The inverse of the statement is obvious.
232 JOS

E MAR

IA POZO AND GARRET SOBCZYK

From the denition of an isometry, (dF(x))
2
= (dy)
2
. Since dF(y) and dy are
vectors in R
p+1,q+1
, dF(y) can be obtained as the result of applying a eld of
orthogonal transformations to dy,
dF(y) = R(y) dyR(y)
1
(64)
expressed here through a eld of versors
R(y) Pin
p+1,q+1
{X = a
1
a
2
. . . a
n
G
p+1,q+1
| a
2
i
= 1}.
Note that R(y)
1
= R(y)

, where R

and R

denote the main involution and the

reversion respectively. Thus, the result that we must prove is that R(y) is constant,
i.e. independent of the point y. This shall guarantee that F(y) is a global rigid
isometry.
The fact that the tangent space T (N
0
) has dimension n + 1 and a metrically
degenerate null direction x
c
is sufcient to guarantee that the image of dF(y)
denes a unique orthogonal transformation in R
p+1,q+1
, which determines (up to a
sign) the versor R(y).
Previously, we found that any isometry F(y) = t (x)f (x)
c
in N
0
is linear in
the scalar coordinate t . Taking the exterior derivative, we get
dF(y) =
dt
t
F(y) +t d((x)f (x)
c
) =
dt
t
F(y) +t d
_
(x)S
f (x)
eS

f (x)
_
and using (64) and (62), we also have
R(y) dyR(y)
1
=
dt
t
R(y)yR(y)
1
+t R(y)S
x
dxS

x
R(y)
1
.
It follows that
R(y)S
x
dxS

x
R(y)
1
= d((x)S
f (x)
eS

f (x)
)
and F(y) = R(y)yR(y)
1
, so that R(y) is independent of t . We have now shown
that any isometry in N
0
satises
dF(y) = R(x) dyR(x)
1
and F(y) = R(x)yR(x)
1
, (65)
where R(x) Pin
p+1,q+1
is solely a function of x R
p,q
. It remains to be shown
that R(x) = R is also independent of x so that F(y) = RyR
1
is a global
orthogonal transformation in N
0
R
p+1,q+1
.
We now slightly generalize the denition of the h-twistor to apply to the rotor
R
x
:= R(x) Pin
p+1,q+1
. Letting T
x
:= R
x
S
x
, we can rewrite (65) in the form
dF(y) = T
x
(t dx + dt e)T
x
1
and F(y) = T
x
t eT
x
1
. (66)
Analogous to (58) and (61), we dene the three bivector valued forms:

R
:= 2R
1
x
dR
x
,
0
:= S

R
S
x
and
T
:= 2T
1
x
dT
x
. (67)
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 233
From the denition of T
x
, we obtain the relation
T
=
0
+
S
.
In order to prove that R
x
is constant, let us rst impose the integrability con-
dition that the second exterior differential d dF must vanish. Note that in the
calculations below we are taking into account both the antisymmetry of exterior
forms as well as the noncommutativity of multivectors. Using (65), we nd
0 = d dF = d(R
x
dyR
1
x
) = dR
x
dyR
1
x
R
x
dy dR
1
x
=
1
2
R
x
(
R
dy + dy
R
)R
1
x

R
dy = 0.
Using (62), this is equivalent to

R
dy = (S

R
S
x
) (S

x
dyS
x
) =
0
(t dx + dt e) = 0. (68)
Since R(x) and S
x
are independent of t , then
0
does not contain dt . Thus,
Equation (68) can be separated into two parts:

0
(t dx + dt e) = t
0
dx + dt
0
e = 0
_

0
dx = 0,

0
e = 0.
(69)
From
0
e = 0, it follows that the bivector-valued form
0
can be written as

0
(x, a) = v(x, a) e +B(x, a),
where v(x, a) is a vector in R
p,q
, and B(x, a) is a bivector in the geometric algebra
G
2
p,q
of R
p,q
.
Imposing the rst equation in (69) we get

0
(a) b
0
(b) a = 0

_
v(a) b v(b) a = 0,
B(a) b B(b) a = 0
B(a) (b c) = B(b) (a c) B(a) = 0 a R
p,q

0
(a) = v(a) e. (70)
The second integrability condition is found by taking the exterior derivative of

R
= 2R
1
x
dR
x
to nd
d
R
= 2 dR
x
1
dR
x
= 2 dR
1
x
R
x
R
x
1
dR
x
=
1
2

R
. (71)
But (70) implies
R

R
= S
x

0
S

x
= 0, from which it follows that d
R
= 0.
Next, we write this as an equation in
0
, getting
0 = d
R
= d(S
x

0
S

x
) = S
x
(d
0
+
S

0
)S

x
d
0
+
S

0
= 0.
With the help of (70) and (58), we now split this equation into its three multivector
parts:
d
0
+
S

0
= dv e +v dx +v dx e e = 0

_
_
_
dv = 0,
v dx = 0,
v dx = 0.
(72)
234 JOS

E MAR

IA POZO AND GARRET SOBCZYK

The bivector part
v dx = 0 v(a) b = v(b) a
differentiates drastically between the dimension d = 2, and for the dimensions
d > 2. When d > 2, we can wedge this last expression with the vector a R
p,q
to infer
v(a) b a = 0 b R
p,q
v(a) a = 0
v(a) = a, R,
from which follows the desired result
a b = b a = 0 v = 0
0
= 0.
Therefore R(x) is constant,

R
= 0 dR(x) = 0 R(y) = R = constant.
Thus, F(y) is a global orthogonal transformation in R
p+1,q+1
,
F(y) = RyR
1
, R Pin
p+1,q+1
. (73)
Since the group of isometries in N
0
is a double covering of the group of con-
formal transformations Con
p,q
in R
p,q
, and the group Pin
p+1,q+1
is a double cov-
ering of the group of orthogonal transformations O(p + 1, q + 1), it follows that
Pin
p+1,q+1
is a four-fold covering of Con
p,q
.
The case of d = 2 will be treated after introducing the matrix representation of
next section.
4.6. MATRIX REPRESENTATION
The algebra G
p+1,q+1
is isomorphic to G
p,q
G
1,1
. This isomorphism can be spec-
ied by means of the so-called conformal split (Hestenes, 1991). Evidently, once
this isomorphism of algebras is established, we can use the matrix representation
introduced in Subsection 2.2 for SNB
1,1
, taking into account that the 2 2 ma-
trices are dened over the module G
p,q
. This identication makes possible a very
elegant treatment of the so-called Vahlen matrices (Lounesto, 1997; Maks, 1989;
Cnops, 1996; Porteous, 1995).
The conformal split does not identify the algebra G
p,q
appearing in the isomor-
phism G
p,q
G
1,1
directly with G
p,q
:= gen{R
p,q
}. Instead, the conformal split
identies G
p,q
with a subalgebra of G
p+1,q+1
generated by a subset of trivectors:
G
p,q
:= gen{R
p,q
u}, where u = is the unit bivector orthogonal to R
p,q
, as
introduced in (36) and Subsection 4.1. This subalgebra has the property that it
commutes with G
1,1
= gen{, } so that G
p+1,q+1
= G
p,q
G
1,1
. The multivectors
belonging to the subalgebra G
p,q
are characterized by
A G
p,q
A = A
+
+uA

, A
+
G
+
p,q
and A

p,q
,
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 235
where A
+
and A

are, respectively, the even and odd multivectors parts of the

multivector A = A
+
+A

G
p,q
. Thus, we have a direct correspondence between
the multivector A G
p,q
and the multivector A A
+
+A

G
p,q
.
Recall that the idempotents u

=
1
2
(1 u) of the algebra G
1,1
, rst dened
in (20), satisfy the properties given in Subsection 2.2:
u
+
+u

= 1, u
+
u

= u,
u
+
u

= 0 = u

u
+
, u
+
= u

,
and
u = e e, u
+
=
1
2
ee, u

=
1
2
ee,
ue = e = eu, eu = e = ue, u
+
= e, 2u

= e.
The representation of G
1,1
, introduced in Subsection 2.2 using the spinor basis,
enables us to write any multivector G G
p+1,q+1
= G
p,q
G
1,1
in the form
G = ( 1 ) u
+
_
A B
C D
__
1

_
,
where the entries of the 2 2 matrix are in G
p,q
. Noting that
u
+
A = u
+
(A
+
+uA

) = u
+
(A
+
+A

) = u
+
A,
makes it possible to work directly with the proper subalgebra G
p,q
, instead of
having to deal with the extra complexity introduced by using the subalgebra G
p,q
.
It follows that each multivector G G
p+1,q+1
can be written in the form
G = ( 1 ) u
+
[G]
_
1

_
= Au
+
+Bu
+
+C

+D

, (74)
where
[G]
_
A B
C D
_
for A, B, C, D G
p,q
.
The matrix [G] denotes the matrix corresponding to the multivector G, and as a
consequence of the general argument given in (22), we have the algebra isomor-
phism
[G
1
+G
2
] = [G
1
] +[G
2
] and [G
1
G
2
] = [G
1
][G
2
],
for all G
1
, G
2
G
p+1,q+1
. This result is an example of the unusual fact that a
matrix representation is sometimes possible even when the module of components
G
p,q
does not commute with the subalgebra G
1,1
. Note, also, the relationships
u
+
[G] = [u
+
G] =
_
A B
0 0
_
and [G]u
+
= [Gu
+
] =
_
A 0
C 0
_
.
236 JOS

E MAR

IA POZO AND GARRET SOBCZYK

The operation of reversion of multivectors translates into the following trans-
pose-like matrix operation:
if [G] =
_
A B
C D
_
then [G]

:= [G

] =
_
D B
C A
_
,
where A = A

is the Clifford conjugation.

4.7. h-TWISTORS AND MOBIUS TRANSFORMATIONS
As seen in Section 4.3, the point x
c
H
p,q
can be written in the form (59),
x
c
= S
x
eS

x
. More generally, in Subsection 4.5, we saw that any conformal trans-
formation F(x
c
) must be of the form
sT
x
eT

x
= F(x
c
) = (x)f (x)
c
= (x)S
f (x)
eS

f (x)
, (75)
where s := T
x
T
x
= 1.
Using the matrix representation of the previous section, for a general multivec-
tor G G
p+1,q+1
, we nd that
[GeG

] =
_
A B
C D
__
0 0
1 0
__
D B
C A
_
=
_
B
D
_
( D B ) , (76)
where
[e] =
_
0 0
1 0
_
, [G]
_
A B
C D
_
, [G]

=
_
D B
C A
_
.
The relationship (76) suggests dening the conformal h-twistor of the multi-
vector G G
p+1,q+1
to be [G]
c
:=
_
B
D
_
, which may also be identied with the
multivector G
c
:= Ge = Bu
+
+ D

e. The conjugate of the conformal h-twistor

is then naturally dened by [G]

c
:= ( D B ). Conformal h-twistors give us a
powerful tool for manipulating the conformal representant and conformal transfor-
mations much more efciently. For example, since x
c
is generated by the conformal
h-twistor [S
x
]
c
, it follows that
[x
c
] = [S
x
]
c
[S
x
]

c
=
_
x
1
_
(1 x) =
_
x x
2
1 x
_
.
Two conformal h-twistors [G
1
]
c
and [G
2
]
c
will be said to be equivalent if they
generate the same multivector, i.e., if
[G
1
]
c
[G
1
]

c
= [G
2
]
c
[G
2
]

c
.
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 237
This is equivalent to the condition G
1
eG

1
= G
2
eG

2
. Two conformal h-twistors
[G
1
]
c
and [G
2
]
c
will be said to be projectively equivalent if they generate the same
direction, i.e., if
[G
1
]
c
[G
1
]

c
= [G
2
]
c
[G
2
]

c
with R

.
This is equivalent to the condition {G
1
eG

1
}
ray
= {G
2
eG

2
}
ray
.
A sufcient condition for two spinor to be projectively equivalent is the follow-
ing:
If H G
p,q
such that HH R and [G
2
]
c
= [G
1
]
c
H,
then [G
2
]
c
[G
2
]

c
= HH[G
1
]
c
[G
1
]

c
. (77)
Moreover, it is not difcult to show that if any component A, B, C or D of the two
conformal h-twistors
_
A
B
_
and
_
C
D
_
is invertible, then this condition is necessary and
sufcient.
We can now write the conformal transformation (75) in its spinorial form
[F(x
c
)] = (x)[S
f (x)
]
c
[S
f (x)
]

c
= s[T
x
]
c
[T
x
]

c
,
from which it follows that [T
x
]
c
and [S
f (x)
]
c
are projectively equivalent spinors.
Since the bottom component of [S
f (x)
]
c
=
_
f (x)
1
_
is trivially invertible, the two
spinors are equivalent by (77). Letting [T
x
]
c
=
_
M
N
_
, it follows that
_
M
N
_
=
_
f (x)
1
_
H H = N and f (x) = MN
1
, (78)
and also that (x) = sNN.
The beautiful linear fractional expression for the conformal transformation f (x),
f (x) = (Ax +B)(Cx +D)
1
(79)
and
(x) = s(Cx +D)(D xC)
is a direct consequence of (78). Since T
x
= RS
x
for the constant versor (73),
R Pin
p+1,q+1
, its spinorial form is given by
[T
x
]
c
= [R][S
x
]
c
=
_
A B
C D
__
x
1
_
=
_
Ax +B
Cx +D
_
=
_
M
N
_
,
where
[R] =
_
A B
C D
_
, for constants A, B, C, D G
p,q
.
The linear fractional expression (79) extends to any dimension and signature the
well-known Mobius transformations in the complex plane. The components A, B,
C, D of [R] are, of course, subject to the condition that R Pin
p+1,q+1
.
238 JOS

E MAR

IA POZO AND GARRET SOBCZYK

Although more difcult to manipulate, our conformal h-twistors are a gener-
alization to any dimension and any signature of the familiar 2-component spinors
over the complex numbers, and the 4-component twistors. Penroses twistor the-
ory (Penrose and MacCallum, 1972) has been discussed in the framework of Clif-
ford algebra by a number of authors, for example see (Ablamowicz and Salingaros,
1985; Ablamowicz and Fauser, 2000, pp. 7592). In the language of spinors, any
null vector y N is the null pole of a conformal h-twistor, [y] = [G]
c
[G]

c
.
Also, two h-twistors will dene the same null pole if they differ only by a phase,
[G
2
]
c
= [G
1
]
c
H, where HH = 1. To complete the analogy, note that each confor-
mal h-twistor also denes a null ag, i.e. a null bivector, tangent to the null cone N.
It easily follows from the expressions (59) and (60) that
x
c
dx
c
= S
x
e dxS

x
[x
c
dx
c
] = [S
x
]
c
dx[S
x
]

c
.
Finally, any h-twistor differing only by a rotor H Spin
p,q
will give the same null
pole but with a different null ag, the null ag rotated by the rotor H:
[S
x
]
c
H dxH[S
x
]

c
.
4.8. THE RELATIVE MATRIX REPRESENTATION
In the two preceding subsections, we have introduced and used a matrix represen-
tation of G
p+1,q+1
, based on the isomorphism G
p+1,q+1
G
p,q
G
1,1
. This matrix
representation depends only upon the choice of a xed spin basis in G
1,1
, but not
on any basis of G
p,q
. We can introduce an alternative relative matrix representation
relative to a choosen nonnull direction a R
pq
, by a slight modication of the
former (74), namely,
G = ( 1 a
1
) u
+
[G]

_
1
a
_
, (80)
so that
[G]

=
_
1 0
0 a
_
[G]
_
1 0
0 a
1
_
.
Evidently, this relative representation has the disadvantage of depending on the
direction a that is chosen. However, it has the important advantage that the parity
of G G
p+1,q+1
is the same as the parity of the components A, B, C, D G
p,q
,
where
[G]

=
_
A B
C D
_
.
Moreover, it is more directly related to complex numbers and to the 4-component
twistors of Penrose and MacCallum (1972). This relative representation will enable
us to relate isometries on N
0
for d = 2 with analytic and antianalytic functions over
the complex numbers C or over the dual numbers D.
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 239
The vectorial representation of points is most directly related to the complex
representation of points via the paravector representant of x relative to a, de-
ned by
z
x
:= xa
1
x = z
x
a. (81)
Whereas this denition is valid in any dimension, we only consider here the di-
mension d = p + q = 2. The set of relative paravectors, in this case, is the even
subalgebra:
{z
x
| x R
p,q
} = G
0
p,q
G
2
p,q
= G
+
p,q
for p +q = 2.
Depending on the signature, the square of the pseudoscalar I G
p,q
can be
either negative (I
2
= 1) or positive (I
2
= 1). It follows that the algebra G
+
p,q
is
isomorphic to either the complex numbers C or to the dual numbers
G
+
2,0
G
+
0,2
C or G
+
1,1
D.
The two vectors {a, Ia} R
p,q
constitutes an orthonormal basis. Relative to this
basis, the vector x and its paravector z
x
have the coordinate forms
x = x
1
a +x
2
Ia and z
x
= x
1
+x
2
I, (82)
where x
1
, x
2
R.
For example, the relative matrix representation of the conformal representant
x
c
is
[x
c
]

=
_
z
x
z
x
z
x
1 z
x
_
a.
The relative matrix representation of the reversion of (80) is
[G]

:= [G

= a
1
_
D B
C A
_
a.
Conformal h-twistors can also be dened for the relative matrix representation in
the obvious way:
[G]

c
:=
_
B
D
_
and [G]

:= (D B),
and satisfy [GeG

= [G]

c
[G]

a.
4.9. CONFORMAL TRANSFORMATIONS IN DIMENSION 2
Before restricting ourselves in Subsection 4.5 to dimensions d > 2, we found the
expression (70)
0
= v(a)e, from which we derived the conditions (72). The ex-
pression for v can be derived from the versor T Pin
p+1,q+1
which generates (67)

T
=
0
+
S
.
240 JOS

E MAR

IA POZO AND GARRET SOBCZYK

Expressing the bivector (58)
S
= dxe in terms of the relative paravectors (81)
z
x
= xa
1
, we get
S
= dz
x
a e. From denition (67) of
T
, we nd
dT =
1
2
T
T
=
1
2
T (
0
+
S
) =
1
2
T (ve + dz
x
a e),
where the parity of T Pin
p+1,q+1
is even or odd. Let us dene
G :=
_
T, if T is even,
aT, if T is odd,
(83)
so that G is always even, and for which it is also true that dG =
1
2
G
T
.
Using the relative matrix representation introduced earlier, we have
[G]

_
A B
C D
_
and [
T
]

=
_
0 2 dz
x
av 0
_
,
where A, B, C, D G
+
p,q
. Note that since the matrix representation (80) is dened
in terms of constant vectors, the differential will commute with the representation
[dG]

= d[G]

. It follows that
_
dA dB
dC dD
_
=
_
A B
C D
__
0 dz
x

1
2
av 0
_
.
We can split this matrix into two columns, getting
_
dA
dC
_
=
1
2
av
_
B
D
_
and
_
dB
dD
_
=
_
A
C
_
dz
x
. (84)
Equation (84) implies that, considered as functions over G
+
p,q
(isomorphic to C
or to D), the two components B and D are analytic, since their differentials are
proportional to dz
x
. Therefore, the derivatives of these analytic functions are
_
B

_
=
_
A
C
_
, where B

:=
dB
dz
x
.
This implies, in turn, that the components A and C are also analytic
_
A
C
_
=
_
B

_
and
_
A

_
=
1
2
av
dz
x
_
B
D
_
. (85)
An immediate consequence of the above equations is that the 1-form av is propor-
tional to dz
x
, so that
0
takes the form
av = g(z
x
) dz
x

0
= a
1
eg(z
x
) dz
x
, (86)
where g(z
x
) is also an analytic function over G
+
p,q
.
Taking into account the change of representation of the conformal h-twistor
[G]
c
to [G]

c
, formula (78) for the spinor [T ]

c
=
_
M
N
_
becomes f (x) = MN
1
a.
Dening the function
f: G
+
p,q
G
+
p,q
, z
x
f(z
x
) := z
f (x)
= f (x)a
1
,
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 241
we obtain f(z
x
) = MN
1
.
We must now consider the two cases when T in (83) is either odd or even. If T
is even then
T = G
_
M
N
_
=
_
B
D
_
f(z
x
) = BD
1
.
If T is odd then
T = a
1
G
_
M
N
_
= a
1
_
B
D
_
f(z
x
) = a
1
BD
1
a.
The function h(z
x
) dened by
h(z
x
) :=
_
f(z
x
), if T is even,
af(z
x
)a
1
= f(z
x
), if T is odd,
has the property that, regardless of whether T is even or odd,
h(z
x
) = BD
1
B = h(z
x
)D. (87)
Since B and D are analytic, it follows that h(z
x
) is also analytic. In the case that
T is even, it generates the analytic transformation f(z
x
) = h(z
x
) in G
+
p,q
. On the
other hand, in the case that T is odd, it generates the anti-analytic transformation
f(z
x
) = h(z
x
).
Using (87) and (85), we can express [G]

in terms of h(z
x
),
[G]

=
_
A B
C D
_
=
_
(hD)

hD
D

D
_
. (88)
The fact that G is in Spin
p,q
can be used to nd an explicit expression for D in
terms of h(z
x
). Using that GG

= 1,
[GG

=
_
A B
C D
__
D B
C A
_
=
_
AD BC 0
0 AD BC
_
,
it follows that det[G]

AD BC = 1. From (85) and (87), it directly follows

that
1 = AD BC = 2(B

D BD

) = 2D
2
(BD
1
)

= D
2
h

so that formally we have

D = (h

)
1/2
=
1

(89)
which, in general, represents four solutions.
For the complex case C G
+
2,0
G
+
0,2
, where I
2
= 1, the four solutions
of (89) are given as usual by
D =
k

, where k = 1, I. (90)
242 JOS

E MAR

IA POZO AND GARRET SOBCZYK

The inverse and square roots of the dual number h

D G
+
1,1
in (89), where I
2
=
1, are not always well dened. The inverse of a dual number z
x
= x
1
+x
2
I D is
given by
z
1
x
=
1
z
x
=
z

x
x
2
1
x
2
2
=
x
1
x
2
I
x
2
1
x
2
2
,
so will only exist when x
1
= x
2
. It can be shown that the dual number h

(except
in the degenerate case when h

= 0) has exactly one of the four hyperbolic Euler

forms (Sobczyk, 1995),
h

=
_
exp(I),
I exp(I),
where =
_
|h

| and is the hyperbolic angle dened by h

. Only in the case

when the sign of h

can be chosen such that h

= exp(I), will h

have four
well-dened square roots in D. For this case we have
D =
k

=
k

exp(
1
2
I), where k = 1, I. (91)
Once we have found D, we also have A, B and C (88)
B =
kh

, A = k
(h

)
2

1
2
hh

(h

)
3/2
, C = k
h

2(h

)
3/2
,
but it is not, in general, possible to solve for the transformation h(z
x
) which cor-
responds to a given
0
= a
1
eg(z
x
) dz
x
. However, we can nd g(z
x
) in terms
of the function h(z
x
): From (85) and (86), we obtain the second-order differential
equation for the conformal h-twistor of G,
_
B

_
=
1
2
g(z
x
)
_
B
D
_
. (92)
From (92) and (90) or (91), we have
g(z
x
) = 2
D

D
=
h

3
2
_
h

_
2
.
It is recognized that g(z
x
) is the Schwarzian derivative of h(z
x
), which vanishes
whenever h(z
x
) is a Mbius transformation. There are many possibilities for the
further study of the Schwarzian derivative and its generalizations (Kobayashi and
Wada, 2000).
Acknowledgements
Jos Pozo acknowledges the support of the Spanish Ministry of Education (MEC),
grant AP96-52209390, the project PB96-0384, and the Catalan Physics Society
(IEC). Garret Sobczyk gratefully acknowledges the support of INIP of the Univer-
sidad de Las Americas-Puebla, and CIMAT-Guanajuato during his Sabbatical, in
the Fall of 1999.
GEOMETRIC ALGEBRA IN LINEAR ALGEBRA AND GEOMETRY 243
References
Ablamowicz, R. and Salingaros, N. (1985) On the relationship between twistors and Clifford
algebras, Lett. Math. Phys. 9, 149155.
Ablamowicz, A. and Fauser, B. (eds) (2000) Clifford Algebras and their Applications in Mathemati-
cal Physics, Birkhuser, Boston.
Barnabei, M., Brini, A. and Rota, G. C. (1985) On the exterior calculus of invariant theory, J. Algebra
96, 120160.
Bayro Corrochano, E. and Sobczyk, G. (eds) (2001) Geometric Algebra with Applications in Science
and Engineering, Birkhuser, Boston.
Cnops, J. (1996) Vahlen matrices for non-denite metrics, In: R. Ablamowicz, P. Lounesto and
J. M. Parra (eds), Clifford Algebras with Numeric and Symbolic Computations, Birkhuser,
Boston.
Davis, P. J. (1974) The Schwarz Function and its Applications, Math. Assoc. Amer.
Doran, C., Hestenes, D., Sommen, F. and Van Acker, N. (1993) Lie groups as spin groups, J. Math.
Phys. 34, 36423669.
Dress, A. W. M. and Havel, T. F. (1993) Distance geometry and geometric algebra, Found. Phys.
23(10), 13571374.
Fulton, W. and Harris, J. (1991) Representation Theory: A First Course, Springer-Verlag.
Havel, T. (1995) Geometric algebra and Mobius sphere geometry as a basis for Euclidean invari-
ant theory, In: N. L. White (ed.), Invariant Methods in Discrete and Computational Geometry,
Kluwer Acad. Publ., Dordrecht, pp. 245256.
Haantjes, J. (1937) Conformal representations of an n-dimensional Euclidean space with a non-
denite fundamental form on itself, Proc. Ned. Akad. Wet. (Math.) 40, 700705.
Hestenes, D. (1991) The design of linear algebra and geometry, Acta Appl. Math. 23, 6593.
Hestenes, D. and Sobczyk, G. (1984) Clifford Algebra to Geometric Calculus: A Unied Language
for Mathematics and Physics, D. Reidel, Dordrecht.
Hestenes, D. and Ziegler, R. (1991) Projective geometry with Clifford algebra, Acta Appl. Math. 23,
2563.
Kobayashi, O. and Wada, M. (2000) The Schwarzian and Mbius transformations in higher dimen-
sions, 5th International Clifford Algebra Conference, Ixtapa 1999, In: J. Ryan and W. Sprsig
(eds), Clifford Algebras and their Applications in Mathematical Physics, Vol. 2, Birkhuser,
Boston.
Li, H., Hestenes, D. and Rockwood, A. Generalized Homogeneous Coordinates for Computational
Geometry, to be published.
Lounesto, P. (1994) CLICAL software paquet and user manual, Helsinki Univ. Technol. Mathemat-
ics, Research Report A248.
Lounesto, P. (1997) Clifford Algebras and Spinors, Cambridge Univ. Press, Cambridge.
Lounesto, P. and Springer, P. (1989) Mobius transformations and Clifford algebras of Euclidean
and anti-Euclidean spaces, In: J. Lawrynowicz (ed.), Deformations of Mathematical Structures,
Kluwer Acad. Publ., Dordrecht, pp. 7990.
Maks, J. G. (1989) Modulo (1,1) periodicity of Clifford algebras and generalized Mobius transfor-
mations, Technical Univ. Delft (Dissertation).
Nehari, Z. (1952) Conformal Mapping, McGraw-Hill, New York.
Penrose, R. and MacCallum, M. A. H. (1972) Twistor theory: An approach to quantization of elds
and space-time, Phys. Rep. 6C, 241316.
Porteous, I. R. (1995) Clifford Algebras and the Classical Groups, Cambridge Univ. Press.
Rao, C. R. and Mitra, S. K. (1971) Generalized Inverse of Matrices and its Applications, Wiley, New
York.
244 JOS

E MAR

IA POZO AND GARRET SOBCZYK

Sobczyk, G. (1993) Jordan form in associative algebras, In: F. Brackx (ed.), Clifford Algebras and
their Applications in Mathematical Physics, Proc. Third Internat. Clifford Workshop, Kluwer
Acad. Publ., Dordrecht, pp. 3341.
Sobczyk, G. (1995) The hyperbolic number plane, College Math. J. 26(4), 268280.
Sobczyk, G. (1996) Structure of factor algebras and Clifford algebra, Linear Algebra Appl. 241243,
803810.
Sobczyk, G. (1997a) The generalized spectral decomposition of a linear operator, College Math. J.
28(1), 2738.
Sobczyk, G. (1997b) Spectral integral domains in the classroom, Aportaciones Mat. Comun. 20,
169188.
Sobczyk, G. (2001) The missing spectral basis in algebra and number theory, Amer. Math. Monthly
108(4), 336346.
Young, J. W., (1930) Projective Geometry, The Open Court Publ. Co., Chicago.