Introduction to Vector Spaces, Vector

Algebras, and Vector Geometries

Richard A. Smith

February 14, 2018 via pdflatex

2

Preface
Vector spaces appear so frequently in both pure and applied mathematics
that I felt that a work that promotes self-study of some of their more ele-
mentary appearances in a manner that emphasizes some basic algebraic and
geometric concepts would be welcome. Undergraduate mathematics majors
in their junior or senior year are the intended primary readers, and I hope
that they and others sufficiently mathematically prepared will find its study
worthwhile.
.
Copyright 2005-2010. Permission granted to transmit electronically. May
not be sold by others for profit.
.
Richard A. Smith
[email protected]
Contents

1 Fundamentals of Structure 7
1.1 What is a Vector Space? . . . . . . . . . . . . . . . . . . . . . 7
1.2 Some Vector Space Examples . . . . . . . . . . . . . . . . . . 8
1.3 Subspace, Linear Combination and Span . . . . . . . . . . . . 9
1.4 Independent Set, Basis and Dimension . . . . . . . . . . . . . 9
1.5 Sum and Direct Sum of Subspaces . . . . . . . . . . . . . . . . 15
1.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2 Fundamentals of Maps 21
2.1 Structure Preservation and Isomorphism . . . . . . . . . . . . 21
2.2 Kernel, Level Sets and Quotient Space . . . . . . . . . . . . . 24
2.3 Short Exact Sequences . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Projections and Reflections on V = W ⊕ X . . . . . . . . . . . 30
2.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3 More on Maps and Structure 35

3.1 Function Spaces and Map Spaces . . . . . . . . . . . . . . . . 35
3.2 Dual of a Vector Space, Dual of a Basis . . . . . . . . . . . . . 35
3.3 Annihilators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Dual of a Map . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5 The Contragredient . . . . . . . . . . . . . . . . . . . . . . . . 39
3.6 Product Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.7 Maps Involving Products . . . . . . . . . . . . . . . . . . . . . 41
3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 Multilinearity and Tensoring 47

4.1 Multilinear Transformations . . . . . . . . . . . . . . . . . . . 47
4.2 Functional Product Spaces . . . . . . . . . . . . . . . . . . . . 48

3
4 CONTENTS

4.3 Functional Tensor Products . . . . . . . . . . . . . . . . . . . 49

4.4 Tensor Products in General . . . . . . . . . . . . . . . . . . . 51
4.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5 Vector Algebras 57
5.1 The New Element: Vector Multiplication . . . . . . . . . . . . 57
5.2 Quotient Algebras . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3 Algebra Tensor Product . . . . . . . . . . . . . . . . . . . . . 60
5.4 The Tensor Algebras of a Vector Space . . . . . . . . . . . . . 61
5.5 The Exterior Algebra of a Vector Space . . . . . . . . . . . . . 62
5.6 The Symmetric Algebra of a Vector Space . . . . . . . . . . . 68
5.7 Null Pairs and Cancellation . . . . . . . . . . . . . . . . . . . 70
5.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

6 Vector Affine Geometry 73

6.1 Basing a Geometry on a Vector Space . . . . . . . . . . . . . . 73
6.2 Affine Flats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.3 Parallelism in Affine Structures . . . . . . . . . . . . . . . . . 74
6.4 Translates and Other Expressions . . . . . . . . . . . . . . . . 75
6.5 Affine Span and Affine Sum . . . . . . . . . . . . . . . . . . . 76
6.6 Affine Independence and Affine Frames . . . . . . . . . . . . . 77
6.7 Affine Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.8 Some Affine Self-Maps . . . . . . . . . . . . . . . . . . . . . . 81
6.9 Congruence Under the Affine Group . . . . . . . . . . . . . . . 82
6.10 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

7 Basic Affine Results and Methods 83

7.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.2 Two Axiomatic Propositions . . . . . . . . . . . . . . . . . . . 83
7.3 A Basic Configuration Theorem . . . . . . . . . . . . . . . . . 84
7.4 Barycenters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.5 A Second Basic Configuration Theorem . . . . . . . . . . . . . 86
7.6 Vector Ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.7 A Vector Ratio Product and Collinearity . . . . . . . . . . . . 91
7.8 A Vector Ratio Product and Concurrency . . . . . . . . . . . 92
CONTENTS 5

8 An Enhanced Affine Environment 95

8.1 Inflating a Flat . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.2 Desargues Revisited . . . . . . . . . . . . . . . . . . . . . . . . 98
8.3 Representing (Generalized) Subflats . . . . . . . . . . . . . . . 100
8.4 Homogeneous and Plücker Coordinates . . . . . . . . . . . . . 101
8.5 Dual Description Via Annihilators . . . . . . . . . . . . . . . . 103
8.6 Direct and Dual Plücker Coordinates . . . . . . . . . . . . . . 104
8.7 A Dual Exterior Product . . . . . . . . . . . . . . . . . . . . . 109
8.8 Results Concerning Zero Coordinates . . . . . . . . . . . . . . 109
8.9 Some Examples in Space . . . . . . . . . . . . . . . . . . . . . 111
8.10 Factoring a Blade . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.11 Algorithm for Sum or Intersection . . . . . . . . . . . . . . . . 116
8.12 Adapted Coordinates and ∨ Products . . . . . . . . . . . . . . 117
8.13 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

9 Vector Projective Geometry 125

9.1 The Projective Structure on V . . . . . . . . . . . . . . . . . . 125
9.2 Projective Frames . . . . . . . . . . . . . . . . . . . . . . . . . 126
9.3 Pappus Revisited . . . . . . . . . . . . . . . . . . . . . . . . . 127
9.4 Projective Transformations . . . . . . . . . . . . . . . . . . . . 130
9.5 Projective Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 132
9.6 Central Projection . . . . . . . . . . . . . . . . . . . . . . . . 132
9.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

10 Scalar Product Spaces 137

10.1 Pairings and Their Included Maps . . . . . . . . . . . . . . . . 137
10.2 Nondegeneracy . . . . . . . . . . . . . . . . . . . . . . . . . . 137
10.3 Orthogonality, Scalar Products . . . . . . . . . . . . . . . . . 139
10.4 Reciprocal Basis . . . . . . . . . . . . . . . . . . . . . . . . . . 140
10.5 Related Scalar Product for Dual Space . . . . . . . . . . . . . 141
10.6 Some Notational Conventions . . . . . . . . . . . . . . . . . . 143
10.7 Standard Scalar Products . . . . . . . . . . . . . . . . . . . . 144
10.8 Orthogonal Subspaces and Hodge Star . . . . . . . . . . . . . 145
10.9 Scalar Product on Exterior Powers . . . . . . . . . . . . . . . 146
10.10Another Way to Define ∗ . . . . . . . . . . . . . . . . . . . . . 149
10.11Hodge Stars for the Dual Space . . . . . . . . . . . . . . . . . 152
10.12Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6 CONTENTS
Chapter 1

Fundamentals of Structure

1.1 What is a Vector Space?

A vector space is a structured set of elements called vectors. The structure
required for this set of vectors is that of an Abelian group with operators
from a specified field. (For us, multiplication in a field is commutative, and
a number of our results will depend on this in an essential way.) If the
specified field is F, we say that the vector space is over F. It is standard for
the Abelian group to be written additively, and for its identity element to be
denoted by 0. The symbol 0 is also used to denote the field’s zero element,
but using the same symbol for both things will not be confusing. The field
elements will operate on the vectors by multiplication on the left, the result
always being a vector. The field elements will be called scalars and be said
to scale the vectors that they multiply. The scalars are required to scale the
vectors in a harmonious fashion, by which is meant that the following three
rules always hold for any vectors and scalars. (Here a and b denote scalars,
and v and w denote vectors.)

1. Multiplication by the field’s unity element is the identity operation:

1 · v = v.

2. Multiplication is associative: a · (b · v) = (ab) · v.

3. Multiplication distributes over addition: (a + b) · v = a · v + b · v and

a · (v + w) = a · v + a · w.

Four things which one might reasonably expect to be true are true.

7
8 CHAPTER 1. FUNDAMENTALS OF STRUCTURE

Proposition 1 If a is a scalar and v a vector, a · 0 = 0, 0 · v = 0, (−a) · v =

−(a · v), and given that a · v = 0, if a 6= 0 then v = 0.

Proof: a · 0 = a · (0 + 0) = a · 0 + a · 0; adding − (a · 0) to both ends yields

the first result. 0 · v = (0 + 0) · v = 0 · v + 0 · v; adding − (0 · v) to both ends
yields the second result. Also, 0 = 0 · v = (a + (−a)) · v = a · v + (−a) · v;
adding − (a · v) on the front of both ends yields the third result. If a · v = 0
with a 6= 0, then a−1 · (a · v) = a−1 · 0, so that (a−1 a) · v = 0 and hence v = 0
as claimed. 

Exercise 1.1.1 Given that a · v = 0, if v 6= 0 then a = 0.

Exercise 1.1.2 −(a · v) = a · (−v).

Because of the harmonious fashion in which the scalars operate on the

vectors, the elementary arithmetic of vectors contains no surprises.

1.2 Some Vector Space Examples

There are many familiar spaces that are vector spaces, although in their
usual appearances they may have additional structure as well. A familiar
vector space over the real numbers R is Rn , the space of n-tuples of real
numbers with component-wise addition and scalar multiplication by elements
of R. Closely related is another vector space over R, the space ∆Rn of all
translations of Rn , the general element of which is the function from Rn to
itself that sends the general element (ξ1 , ..., ξn ) of Rn to (ξ1 + δ1 , ..., ξn + δn )
for a fixed real n-tuple (δ1 , ..., δn ). Thus an element of ∆Rn uniformly vectors
the elements of Rn to new positions, displacing each element of Rn by the
same vector, and in this sense ∆Rn is truly a space of vectors. In the same
fashion as with Rn over R, over any field F, the space F n of all n-tuples
of elements of F constitutes a vector space. The space Cn of all n-tuples
of complex numbers in addition to being considered to be a vector space
over C may also be considered to be a (different, because it is taken over a
different field) vector space over R. Another familiar algebraic structure that
is a vector space, but which also has additional structure, is the space of all
polynomials with coefficients in the field F. To conclude this initial collection
of examples, the set consisting of the single vector 0 may be considered to
be a vector space over any field; these are the trivial vector spaces.
1.3. SUBSPACE, LINEAR COMBINATION AND SPAN 9

1.3 Subspace, Linear Combination and Span

Given a vector space V over the field F, U is a subspace of V, denoted
U C V, if it is a subgroup of V that is itself a vector space over F. To show
that a subset U of a vector space is a subspace, it suffices to show that U
is closed under sum and under product by any scalar. We call V and {0}
improper subspaces of the vector space V, and we call all other subspaces
proper.

Exercise 1.3.1 As a vector space over itself, R has no proper subspace. The
set of all integers is a subgroup, but not a subspace.

Exercise 1.3.2 The intersection of any number of subspaces is a subspace.

P
A linear combination of a finite set S of vectors is any sum s∈S cs · s
obtained by adding together exactly one multiple, by some scalar coefficient,
of each vector of S. A linear combination of an infinite set S of vectors
is a linear combination of any finite subset of S. The empty sum which
results in forming the linear combination of the empty set is taken to be
0, by convention. A subspace must contain the value of each of its linear
combinations. If S is any subset of a vector space, by the span of S, denoted
hSi, is meant the set of the values of all linear combinations of S. hSi is a
subspace. If U = hSi, then we say that S spans U and that it is a spanning
set for U.

Exercise 1.3.3 For any set S of vectors, hSi is the intersection of all sub-
spaces that contain S, and therefore hSi itself is the only subspace of hSi that
contains S.

1.4 Independent Set, Basis and Dimension

Within a given vector space, a dependency is said to exist in a given set of
vectors if the zero vector is the value of a nontrivial (scalar coefficients not
all zero) linear combination of some finite nonempty subset of the given set.
A set in which a dependency exists is (linearly) dependent and otherwise
is (linearly) independent. The “linearly” can safely be omitted and we
will do this to simplify our writing.
10 CHAPTER 1. FUNDAMENTALS OF STRUCTURE

For example, if v 6= 0, then {v} is an independent set. The empty set is

an independent set. Any subset of an independent set is independent. On
the other hand, {0} is dependent, as is any set that contains 0.

Exercise 1.4.1 A set S of vectors is independent if and only if no member

is in the span of the other members: for all v ∈ S, v ∈ / hS r {v}i. What
would the similar statement need to be if the space were not a vector space
over a field, but only the similar type of space over the ring Z of ordinary
integers?

Similar ideas of dependency and independence may be introduced for

vector sequences. A vector sequence is dependent if the set of its terms is
a dependent set or if the sequence has any repeated terms. Otherwise the
sequence is independent.

Exercise 1.4.2 The terms of the finite vector sequence v1 , . . . , vn satisfy a

linear relation a1 · v1 + · · · + an · vn = 0 with at least one of the scalar
coefficients ai nonzero if and only if the sequence is dependent. If the sequence
is dependent, then it has a term that is in the span of the set of preceding
terms and this set of preceding terms is nonempty if v1 6= 0.

A maximal independent set in a vector space, i. e., an independent

set that is not contained in any other independent set, is said to be a basis
set, or, for short, a basis (plural: bases), for the vector space. For example,
{(1)} is a basis for the vector space F 1 , where F is any field. The empty set
is the unique basis for a trivial vector space.

Exercise 1.4.3 Let Φ be a family of independent sets which is linearly or-

dered by set inclusion. Then the union of all the sets in Φ is an independent
set.

From the result of the exercise above and Zorn’s Lemma applied to inde-
pendent sets which contain a given independent set, we obtain the following
important result.

Theorem 2 Every vector space has a basis, and, more generally, every in-
dependent set is contained in some basis. 

A basis may be characterized in various other ways besides its definition

as a maximal independent set.
1.4. INDEPENDENT SET, BASIS AND DIMENSION 11

Proposition 3 (Independent Spanning Set) A basis for the vector space

V is precisely an independent set that spans V.

Proof: In the vector space V, let B be an independent set such that

hBi = V. Suppose that B is contained in the larger independent set C. Then,
choosing any v ∈ C r B, the set B 0 = B ∪ {v} is independent because it is
a subset of the independent set C. But then v ∈ / hB 0 r {v}i, i. e., v ∈
/ hBi,
contradicting hBi = V. Hence B is a basis for V.
On the other hand, let B be a basis, i. e., a maximal independent set,
but suppose that B does not span V. There must then exist v ∈ V r hBi.
However, B ∪ {v} would then be independent, since a nontrivial dependency
relation in B ∪ {v} would have to involve v with a nonzero coefficient and
this would put v ∈ hBi. But the independence of B ∪ {v} contradicts the
hypothesis that B is a maximal independent set. Hence, the basis B is an
independent set that spans V. 

Corollary 4 An independent set is a basis for its own span. 

Proposition 5 (Minimal Spanning Set) B is a basis for the vector space

V if and only if B is a minimal spanning set for V, i. e., B spans V, but
no subset of B unequal to B also spans V.

Proof: Suppose B spans V and no subset of B unequal to B also spans V.

If B were not independent, there would exist b ∈ B for which b ∈ hB r {b}i.
But then the span of B r {b} would be the same as the span of B, contrary
to what we have supposed. Hence B is independent, and by the previous
proposition, B must therefore be a basis for V since B also spans V.
On the other hand, suppose B is a basis for the vector space V. By the
previous proposition, B spans V. Suppose that A is a subset of B that is
unequal to B. Then no element v ∈ B r A can be in hB r {v}i since B is
independent, and certainly, then, v ∈
/ hAi since A ⊂ B r {v}. Thus A does
not span V, and B is therefore a minimal spanning set for V. 

Exercise 1.4.4 Consider a finite spanning set. Among its subsets that also
span, there is at least one of smallest size. Thus a basis must exist for any
vector space that has a finite spanning set, independently verifying what we
already know from Theorem 2.
12 CHAPTER 1. FUNDAMENTALS OF STRUCTURE

Exercise 1.4.5 Over any field F, the set of n n-tuples that have a 1 in one
position and 0 in all other positions is a basis for F n . (This basis is called
the standard basis for F n over F.)

Proposition 6 (Unique Linear Representation) B is a basis for the vec-

tor space V if and only if B has the unique linear representation prop-
P vector of V has a unique linear representation as a linear
erty, i. e., each
combination x∈X ax · x where X is some finite subset of B and all of the
scalars ax are nonzero.

Proof: Let B be a basis for V, and let v ∈ V. Since V is the span of

B, v certainly is a linear combination of a finite subset of B with all scalar
coefficients nonzero. Suppose that v has two different expressions of this
kind. Subtracting, we find that 0 is equal to a nontrivial linear combination
of the union U of the subsets of B involved in the two expressions. But U
is an independent set since it is a subset of the independent set B. Hence v
has the unique representation as claimed.
On the other hand, suppose that B has the unique linear representation
property. Then B spans V. We complete the proof by showing that B must
be an independent set. Suppose not. Then there would exist b ∈ B for which
b ∈ hB r {b}i . But this b would then have two different representations with
nonzero coefficients, contrary to hypothesis, since we always have 1 · b as one
such representation and there would also be another one not involving b in
virtue of b ∈ hB r {b}i. 

Exercise 1.4.6 A finite set of vectors is a basis for a vector space if and
only if each vector in the vector space has a unique representation as a linear
combination of this set: {x1 , . . . , xn } (with distinct xi , of course) is a basis
if and only if each v = a1 · x1 + · · · + an · xn for unique scalars a1 , . . . , an .
P P
Exercise 1.4.7 If S is a finite independent set, s∈S cs · s = s∈S ds · s
implies cs = ds for all s.

Example 7 In this brief digression we now apply the preceding two propo-
sitions. Let v0 , v1 , . . . , vn be vectors in a vector space over the field F, and
suppose that v0 is in the span V of the other vi . Then the equation

ξ1 · v1 + · · · + ξn · vn = v0
1.4. INDEPENDENT SET, BASIS AND DIMENSION 13

for the ξi ∈ F has at least one solution. Renumbering the vi if necessary,

we may assume that the distinct vectors v1 , . . . , vm form a basis set for V. If
m = n the equation has a unique solution. Suppose, however, that 1 6 m <
n. Then the equation

ξ1 · v1 + · · · + ξm · vm = v0 − ξm+1 · vm+1 − · · · − ξn · vn

has a unique solution for ξ1 , . . . , ξm for each fixed set of values we give the
other ξi , since the right-hand side is always in V. Thus ξ1 , . . . , ξm are func-
tions of the variables ξm+1 , . . . , ξn , where each of these variables is allowed
to range freely over the entire field F. When F is the field R of real num-
bers, we have deduced a special case of the Implicit Function Theorem of
multivariable calculus. When the vi are d-tuples (i. e., elements of F d ), the
original vector equation is the same thing as the general set of d consistent
numerical linear equations in n unknowns, and our result describes how, in
the general solution, n − m of the ξi are arbitrary parameters upon which the
remaining m of the ξi depend.

We have now reached a point where we are able to give the following key
result.

Theorem 8 (Replacement Theorem) Let V be a vector space with basis

B, and let C be an independent set in V. Given any finite subset of C no
larger than B, its elements can replace an equal number of elements of B and
the resulting set will still be a basis for V.

Proof: The theorem holds in the case when no elements are replaced.
Suppose that the elements y1 , ..., yN of C have replaced N elements of B
and the resulting set BN is still a basis for V. An element yN +1 of C r
{y1 , ..., yN } has a unique linear representation as a linear combination with
nonzero coefficients of some finite nonempty subset X = {x1 , ..., xK } of BN .
There must be an element x∗ ∈ X r {y1 , ..., yN } because yN +1 cannot be a
linear combination of {y1 , ..., yN } since C is independent. In BN replace x∗
with yN +1 . Clearly, the result BN +1 still spans V.
The assumption that x ∈ BN +1 is a linear combination of the other el-
ements of BN +1 will be seen to always contradict the independence of BN ,
proving that BN +1 is independent. For if x = yN +1 , then we can immedi-
ately solve for x∗ as a linear combination of the other elements of BN . And if
x 6= yN +1 , writing x = a·yN +1 +y, where y is a linear combination of elements
14 CHAPTER 1. FUNDAMENTALS OF STRUCTURE

of BN r {x, x∗ }, we find that we can solve for x∗ as a linear combination of

the other elements of BN unless a = 0, but then x is a linear combination of
the other elements of BN .
Thus the theorem is shown to hold for N + 1 elements replaced given that
it holds for N . By induction, the theorem holds in general. 
Suppose that B is a finite basis of n vectors and C is an independent set at
least as large as B. By the theorem, we can replace all the vectors of B by any
n vectors from C, the set of which must be a basis, and therefore a maximal
independent set, so that C actually contains just n elements. Hence, when
a finite basis exists, no independent set can exceed it in size, and no basis
can then be larger than any other basis. We therefore have the following
important result.

Corollary 9 If a vector space has a finite basis, then each of its bases is
finite, and each has the same number of elements. 

A vector space is called finite-dimensional if it has a finite basis, and

otherwise is called infinite-dimensional. The dimension of a vector space
is the number of elements in any basis for it if it is finite-dimensional, and
∞ otherwise. We write dim V for the dimension of V. If a finite-dimensional
vector space has dimension n, then we say that it is n-dimensional. Over
any field F, F n is n-dimensional.

Exercise 1.4.8 In an n-dimensional vector space, any vector sequence with

n + 1 terms is dependent.

Exercise 1.4.9 Let the vector space V have the same finite dimension as its
subspace U. Then U = V.

The corollary above has the following analog for infinite-dimensional

spaces.

Theorem 10 All bases of an infinite-dimensional vector space have the same

cardinality.

Proof: Let B be an infinite basis for a vector space and let C be another
basis for the same space. For each y ∈ C let Xy be the finite nonempty
subset of B such that y is the linear combination of its elements with nonzero
1.5. SUM AND DIRECT SUM OF SUBSPACES 15

scalar coefficients. Then each x ∈ B must appear in some Xy , for if some

x ∈ BSappears in no Xy , then that x could not be in the span of C. Thus
B = y∈C Xy . Using |S| to denote cardinality of the set S, we then have
S 
|B| =  y∈C Xy  and C must be infinite for this to hold. Since C is infinite,
 

for each y we have |Xy | 6 |C| and therefore |B| 6 |C × C| = |C|, where the last
equality is a well-known, but not so easily proved, result found in books on
set theory and in online references such as the Wikipedia entry for Cardinal
number. Similarly, we find |C| 6 |B|. The Schroeder-Bernstein theorem then
gives |B| = |C|. 

1.5 Sum and Direct Sum of Subspaces

By the sum of any number of subspaces is meant the span of their union.
We use ordinary additive notation for subspace sums.

Exercise 1.5.1 For any sets of vectors, the sum of their spans is the span
of their union.

Exercise 1.5.2 Each sum of subspaces is equal to the set of all finite sums
of vectors selected from the various subspace summands, no two vector sum-
mands being selected from the same subspace summand.

The trivial subspace {0} acts as a neutral element in subspace sums.

Because of this, we will henceforth often write 0 for {0}, particularly when
subspace sums are involved. By convention, the empty subspace sum is 0.
With partial order ⊂, meet ∩, and join +, the subspaces of a given vector
space form a complete lattice, which is modular, but generally not distributive.

Exercise 1.5.3 For subspaces, (U1 ∩ U3 ) + (U2 ∩ U3 ) ⊂ (U1 + U2 ) ∩ U3 , and

(Modular Law) U2 C U3 ⇒ (U1 ∩ U3 ) + U2 = (U1 + U2 ) ∩ U3 .

Example 11 In R2 , let X denote the x-axis, let Y denote the y-axis, and
let D denote the line where y = x. Then (X + Y) ∩ D = D while (X ∩ D) +
(Y ∩ D) = 0.

A sum of subspaces is called direct if the intersection of each summandP

with the
L sum of the other summands is 0. We often replace + with ⊕, or
with , to indicate that a subspace sum is direct.
16 CHAPTER 1. FUNDAMENTALS OF STRUCTURE

The summands in a direct sum may be viewed as being independent, in

some sense. Notice, though, that 0 may appear as a summand any number of
times in a direct sum. Nonzero direct summands must be unique, however,
and more generally, we have the following result.

Lemma 12 If a subspace sum is direct, then, for each summand U, U ∩ Σ =

0, where Σ denotes the sum of any selection of the summands which does not
include U. In particular, two nonzero summands must be distinct.

Proof: Σ ⊂ Σ where Σ denotes the sum of all the summands with the
exception of U. By the definition of direct sum, U ∩ Σ = 0, and hence
U ∩ Σ = 0 as was to be shown. 
This lemma leads immediately to the following useful result.

Theorem 13 If a subspace sum is direct, then the sum of any selection of

the summands is direct. 

There are a number of alternative characterizations of directness of sums

of finitely many subspaces. (The interested reader will readily extend the
first of these, and its corollary, to similar results for sums of infinitely many
subspaces.)

Theorem 14 u ∈ U1 ⊕ · · · ⊕ Un ⇔ u = u1 + · · · + un for unique u1 ∈

U1 , . . . , un ∈ Un .

Proof: Suppose that u ∈ U1 ⊕· · ·⊕ UnP , and that u = ui +ui = vi +v i where

ui and vi are in Ui and ui P and v i are in j6=i Uj . Then w = ui − vi = v i − ui
must be both in Ui and in j6=i Uj so that w = 0 and therefore ui = vi . Hence
the representation of u as u = u1 + · · · + un , where u1 ∈ U1 , . . . , un ∈ Un , is
unique.
On the other hand, suppose that each u ∈ U1 + · · · + Un has the unique
representation
P u = u1 + · · · + un where u1 ∈ U1 , . . . , un ∈ Un . Let u ∈
Ui ∩ j6=i Uj . Then, since u ∈ Ui , it has the representation u = ui for some
ui ∈ Ui , and the uniqueness of the representation of u as u = u1 + · · · + un
where u1 ∈ PU1 , . . . , un ∈ Un then implies that ujP= 0 for all j 6= i. Also,
since u ∈ j6=i Uj , it has the representation u = j6=i uj for some uj ∈ Uj ,
and the uniqueness of the representation of u as u = u1 + · · · + un where
u1 ∈ U1P, . . . , un ∈ Un implies that ui = 0. Hence u = 0 and therefore for each
i, Ui ∩ j6=i Uj = 0. The subspace sum therefore is direct. 
1.5. SUM AND DIRECT SUM OF SUBSPACES 17

CorollaryP15 U1 + · · · + Un is direct if and only if 0 has a unique represen-

tation as 16j6n uj , where each uj ∈ Uj .

U1 + · · · + Un is direct, then 0 has a unique representation

Proof: IfP
of the form 16j6n uj , where each uj ∈ Uj , namely that one where each
uj = 0. On Pthe other hand,
P suppose that 0 has that unique representation.
0 00 0 00
Then given 16j6n uj = 16j6n uj , where each uj ∈ Uj and each uj ∈ Uj ,
P 0 00 0 00
0 = 16j6n uj , where each uj = uj − uj ∈ Uj . Hence uj − uj = 0 for each j,
and the subspace sum is direct. 

Proposition 16 U1 + · · · + Un is direct if and only if both U n = U1 +

· · · + Un−1 and U n + Un are direct, or, more informally, U1 ⊕ · · · ⊕ Un =
(U1 ⊕ · · · ⊕ Un−1 ) ⊕ Un .

Proof: Suppose that U1 + · · · + Un is direct. By Theorem 13, U n is direct,

and the definition of direct sum immediately implies that U n + Un is direct.
On the other hand, suppose that both U n = U1 + · · · + Un−1 and U n + Un
are direct. Then 0 has a unique representation of the form u1 + · · · + un−1 ,
where uj ∈ Uj , and 0 has a unique representation of the form un + un , where
0 0
un ∈ U n and un ∈ Un . Let 0 be represented in the form u1 + · · · + un ,
0 0 0 0 0 0
where uj ∈ Uj . Then 0 = un + un where un = u1 + · · · + un−1 ∈ U n . Hence
0 0
un = 0 and un = 0 by the corollary above, but then, by the same corollary,
0 0
u1 = · · · = un−1 = 0. Using the corollary once more, U1 + · · · + Un is proved
to be direct. 
Considering in turn (U1 ⊕ U2 ) ⊕ U3 , ((U1 ⊕ U2 ) ⊕ U3 ) ⊕ U4 , etc., we get
the following result.

Corollary 17 U1 + · · · + Un is direct if and only if (U1 + · · · + Ui ) ∩ Ui+1 = 0

for i = 1, . . . , n − 1. 

Exercise 1.5.4 Let Σk = Uik−1 +1 + · · · + Uik for k = 1, . . . , K, where 0 =

i0 < i1 < · · · < iK = n. Then U1 + · · · + Un is direct if and only if each Σk
sum is direct and Σ1 + · · · + Σn is direct.

Exercise 1.5.5 U1 ⊕ · · · ⊕ U6 = ((U1 ⊕ (U2 ⊕ U3 )) ⊕ (U4 ⊕ U5 )) ⊕ U6 in the

fullest sense.

In the finite-dimensional case, additivity of dimension is characteristic.

18 CHAPTER 1. FUNDAMENTALS OF STRUCTURE

Theorem 18 The finite-dimensional subspace sum U1 + · · · + Un is direct if

and only if
dim (U1 + · · · + Un ) = dim U1 + · · · + dim Un .

Proof: Let each Uj have the basis Bj . It is clear that B1 ∪ · · · ∪ Bn spans

the subspace sum.
Suppose now that U1 + · · · + Un is direct. B1 ∪ · · · ∪ Bn is independent, and
therefore a basis, because, applying the definition of direct sum, a dependency
in it immediately implies one in one of the Bj . And, of course, no element of
any Bi can be in any of the other Bj , or it would be in the sum of their spans.
Hence the Bj are disjoint and dim (U1 + · · · + Un ) = dim U1 + · · · + dim Un .
On the other hand, suppose that dim (U1 + · · · + Un ) = dim U1 + · · · +
dim Un . Because a minimal spanning set is a basis, B1 ∪ · · · ∪ Bn must contain
at least dim U1 + · · · + dim Un distinct elements, but clearly cannot contain
more, and must therefore be a basis for the subspace sum. A nonzero element
in Ui = hBi i and simultaneously in the span of the other Bj , would entail a
dependency in the basis B1 ∪ · · · ∪ Bn , which of course is not possible. Hence
no nonzero element of any Ui can be in the sum of the other Uj and therefore
the subspace sum is direct. 

Exercise 1.5.6 Let U be any subspace of the vector space V. As we already

know, U has a basis A which is part of a basis B for V. Let U = hB r Ai.
Then V = U ⊕ U.
L
Exercise 1.5.7 Let B be a basis for V. Then V = x∈B h{x}i.

1.6 Problems
1. Assuming only that + is a group operation, not necessarily that of an
Abelian group, deduce from the rules for scalars operating on vectors
that u + v = v + u for any vectors v and u.
2. Let T be a subspace that does not contain the vectors u and v. Then

v ∈ hT ∪ {u}i ⇔ u ∈ hT ∪ {v}i .

3. Let S and T be independent subsets of a vector space and suppose that

S is finite and T is larger than S. Then T contains a vector x such
that S ∪ {x} is independent. Deduce Corollary 9 from this.
1.6. PROBLEMS 19

4. Let B and B 0 be bases for the same vector space. Then for every x ∈ B
there exists x0 ∈ B 0 such that both (B r {x}) ∪ {x0 } and (B 0 r {x0 }) ∪
{x} are also bases.

5. Let W and X be finite-dimensional subspaces of a vector space. Then

dim (W ∩ X ) 6 dim X and dim (W + X ) > dim W.

When does equality hold?

6. Let W and X be subspaces of a vector space V. Then V has a basis B

such that each of W and X are spanned by subsets of B.

7. Instead of elements of a field F, let the set of scalars be the integers Z,

so the result is not a vector space, but what is known as a Z-module. Z3 ,
the integers modulo 3, is a Z-module that has no linearly independent
element. Thus Z3 has ∅ as its only maximal linearly independent set,
but h∅i 6= Z3 . Z itself is also a Z-module, and has {2} as a maximal
linearly independent set, but h{2}i = 6 Z. On the other hand, {1} is
also a maximal linearly independent set in Z, and h{1}i = Z, so we
are prone to declare that {1} is a basis for Z. Every Z-module, in
fact, has a maximal linearly independent set, but the span of such a
set may very well not be the whole Z-module. We evidently should
reject maximal linearly independent set as a definition of basis for Z-
modules, but linearly independent spanning set, equivalent for vector
spaces, does give us a reasonable definition for those Z-modules that
are indeed spanned by an independent set. Let us adopt the latter as
our definition for basis of a Z-module. Then, is it possible for a Z-
module to have finite bases with differing numbers of elements? Also,
is it then true that no finite Z-module has a basis?
20 CHAPTER 1. FUNDAMENTALS OF STRUCTURE
Chapter 2

Fundamentals of Maps

2.1 Structure Preservation and Isomorphism

A structure-preserving function from a structured set to another with the
same kind of structure is known in general as a homomorphism. The
shorter term map is often used instead, as we shall do here. When the
specific kind of structure is an issue, a qualifier may also be included, and so
one speaks of group maps, ring maps, etc. We shall be concerned here mainly
with vector space maps, so for us a map, unqualified, is a vector space map.
(A vector space map is also commonly known as a linear transformation.)
For two vector spaces to be said to have the same kind of structure, we
require them to be over the same field. Then for f to be a map between them,
we require f (x + y) = f (x) + f (y) and f (a · x) = a · f (x) for all vectors x, y
and any scalar a, or, equivalently, that f (a · x + b · y) = a · f (x) + b · f (y)
for all vectors x, y and all scalars a,b. It is easy to show that a map sends
0 to 0, and a constant map must then send everything to 0. Under a map,
the image of a subspace is always a subspace, and the inverse image of a
subspace is always a subspace. The identity function on a vector space is a
map. Maps which compose have a map as their composite.

Exercise 2.1.1 Let V and W be vector spaces over the same field and let S
be a spanning set for V. Then any map f : V → W is already completely
determined by its values on S.

The values of a map may be arbitrarily assigned on a minimal spanning

set, but not on any larger spanning set.

21
22 CHAPTER 2. FUNDAMENTALS OF MAPS

Theorem 19 Let V and W be vector spaces over the same field and let B be
a basis for V. Then given any function f0 : B → W, there is a unique map
f : V → W such that f agrees with f0 on B.

Proof: By the unique linear representation property of a basis, given

v ∈ V, thereP is a unique subset X of B and unique nonzero scalars ax such
that v = x∈X ax · x. Because a map preserves linear combinations,
P any map
fPthat agrees with f0 on B can only P have the value f (v) = x∈X ax · f (x) =
x∈X xa · f 0 (x). Setting f (v) = x∈X ax · f0 (x) does define a function
f : V → W and this function f clearly agrees with f0 on B. Moreover, one
immediately verifies that this f is a map. 
It is standard terminology to refer to a function that is both one-to-one
(injective) and onto (surjective) as bijective, or invertible. Each invertible
f : X → Y has a (unique) inverse f −1 such that f −1 ◦ f and f ◦ f −1 are
the respective identity functions on X and Y, and an f that has such an
f −1 is invertible. The composite of functions is bijective if and only if each
individual function is bijective.
For some types of structure, the inverse of a bijective map need not always
be a map itself. However, for vector spaces, inverting does always yield
another map.

Theorem 20 The inverse of a bijective map is a map.

Proof: f −1 (a · v + b · w) = a · f −1 (v) + b · f −1 (w) means precisely the

same as f (a · f −1 (v) + b · f −1 (w)) = a · v + b · w, which is clearly true. 

Corollary 21 A one-to-one map preserves independent sets, and conversely,

a map that preserves independent sets is one-to-one.

Proof: A one-to-one map sends its domain bijectively onto its image and
this image is a subspace of the codomain. Suppose a set in the one-to-one
map’s image is dependent. Then clearly the inverse image of this set is also
dependent. An independent set therefore cannot be sent to a dependent set
by a one-to-one map.
On the other hand, suppose that a map sends the distinct vectors u and
v to the same image vector. Then it sends the nonzero vector v − u to 0, and
hence it sends the independent set {v − u} to the dependent set {0}. 
It is easy to see that any map preserves dependent sequences, and one-
to-one maps preserve dependent sets.
2.1. STRUCTURE PRESERVATION AND ISOMORPHISM 23

Because their inverses are also maps, the bijective maps are the isomor-
phisms of vector spaces. If there is a bijective map from the vector space V
onto the vector space W, we say that W is isomorphic to V. The notation
V∼ = W is commonly used to indicate that V and W are isomorphic. Viewed
as a relation between vector spaces, isomorphism is reflexive, symmetric and
transitive, hence is an equivalence relation. If two spaces are isomorphic to
each other we say that each is an alias of the other.

Theorem 22 Let two vector spaces be over the same field. Then they are
isomorphic if and only if they have bases of the same cardinality.

Proof: Applying Theorem 19, the one-to-one correspondence between

their bases extends to an isomorphism. On the other hand, an isomorphism
restricts to a one-to-one correspondence between bases. 

Corollary 23 Any n-dimensional vector space over the field F is isomorphic

to F n . 

Corollary 24 (Fundamental Theorem of Finite-Dimensional Vector

Spaces) Two finite-dimensional vector spaces over the same field are iso-
morphic if and only if they have the same dimension. 

The following well-known and often useful result may seem somewhat
surprising on first encounter.

Theorem 25 A map of a finite-dimensional vector space into an alias of

itself is one-to-one if and only if it is onto.

Proof: Suppose the map is one-to-one. Since one-to-one maps preserve

independent sets, the image of a basis is an independent set with the same
number of elements. The image of a basis must then be a maximal indepen-
dent set, and hence a basis for the codomain, since the domain and codomain
have equal dimension. Because a basis and its image are in one-to-one cor-
respondence under the map, expressing elements in terms of these bases, it
is clear that each element has a preimage element.
On the other hand, suppose the map is onto. The image of a basis is a
spanning set for the codomain and must contain at least as many distinct
elements as the dimension of the codomain. Since the domain and codomain
24 CHAPTER 2. FUNDAMENTALS OF MAPS

have equal dimension, the image of a basis must in fact then be a mini-
mal spanning set for the codomain, and therefore a basis for it. Expressing
elements in terms of a basis and its image, we find that, due to unique rep-
resentation, if the map sends u and v to the same element, then u = v.

For infinite-dimensional spaces, the preceding theorem fails to hold. It
is a pleasant dividend of finite dimension, analogous to the result that a
function between equinumerous finite sets is one-to-one if and only if it is
onto.

2.2 Kernel, Level Sets and Quotient Space

The kernel of a map is the inverse image of {0}. The kernel of f is a subspace
of the domain of f . A level set of a map is a nonempty inverse image of a
singleton set. The kernel of a map, of course, is one of its level sets, the only
one that contains 0 and is a subspace. The level sets of a particular map
make up a partition of the domain of the map. The following proposition
internally characterizes the level sets as the cosets of the kernel.

Proposition 26 Let L be a level set and K be the kernel of the map f . Then
for any v in L, L = v + K = {v + x | x ∈ K}.

Proof: Let v ∈ L and x ∈ K. Then f (v + x) = f (v) + f (x) = f (v) + 0 so

that v + x ∈ L. Hence v + K ⊂ L. On the other hand, suppose that u ∈ L
and write u = v + (u − v). Then since both u and v are in L, f (u) = f (v),
and hence f (u − v) = 0, so that u − v ∈ K. Hence L ⊂ v + K. 

Corollary 27 A map is one-to-one if and only if its kernel is {0}. 

Given any vector v in the domain of f , it must be in some level set, and
now we know that this level set is v + K, independent of the f which has
domain V and kernel K.

Corollary 28 Maps with the same domain and the same kernel have iden-
tical level set families. For maps with the same domain, then, the kernel
uniquely determines the level sets. 
2.2. KERNEL, LEVEL SETS AND QUOTIENT SPACE 25

The level sets of a map f are in one-to-one correspondence with the

elements of the image of the map. There is a very simple way in which
the level sets can be made into a vector space isomorphic to the image of
the map. Just use the one-to-one correspondence between the level sets
and the image elements to give the image elements new labels according
to the correspondence. Thus z gets the new label f −1 ({z}). To see what
f −1 ({z})+f −1 ({w}) is, figure out what z +w is and then take f −1 ({z + w}),
and similarly to see what a · f −1 ({z}) is, figure out what a · z is and take
f −1 ({a · z}). All that has been done here is to give different names to the
elements of the image of f . This level-set alias of the image of f is the
vector space quotient (or quotient space) of the domain V of f modulo
the subspace K = Kernel(f ) (which is the 0 of the space), and is denoted by
V/K. The following result “internalizes” the operations in V/K.

Proposition 29 Let f be a map with domain V and kernel K. Let L = u+K

and M = v + K. Then in V/K, L + M = (u + v) + K, and for any scalar
a, a · L = a · u + K.

Proof: L = f −1 ({f (u)}) and M = f −1 ({f (v)}). Hence L + M =

f −1 ({f (u) + f (v)}) = f −1 ({f (u + v)}) = (u + v) + K. Also, a · L =
f −1 ({a · f (u)}) = f −1 ({f (a · u)}) = a · u + K. 
The following is immediate.

Corollary 30 V/K depends only on V and K, and not on the map f that
has domain V and kernel K. 

The next result tells us, for one thing, that kernel subspaces are not
special, and hence the quotient space exists for any subspace K. (In the vector
space V, the subspace K is a complementary subspace, or complement,
of the subspace K if K and K haveL disjoint bases that together form a basis
for V, or what is the same, V = K K. Every subspace K of V has at least
one complement K, because any basis A of K is contained in some basis B
of V, and we may then take K = hB r Ai.)

Proposition 31 Let K be a subspace of the vector space V and let K be

a complementary subspace of K. Then there is a map φ from V into itself
which has K as its kernel and has K as its image. Hence given any subspace
K of V, the quotient space V/K exists and is an alias of any complementary
subspace of K.
26 CHAPTER 2. FUNDAMENTALS OF MAPS

Proof: Let A be a basis for K, and let A be a basis for K, so that A ∪ A is

a basis for V. Define φ as the map of V into itself which sends each element
of A to 0, and each element of A to itself. Then φ has kernel K and image
K. 
Supposing that K is the kernel of some map f : V → W, the image of f is
an alias of the quotient space, as is the image of the map φ of the proposition
above. Hence any complementary subspace of the kernel of a map is an
alias of the map’s image. For maps with finite-dimensional domains, we then
immediately deduce the following important and useful result. (The rank
of a map is the dimension of its image, and the nullity of a map is the
dimension of its kernel.)

Theorem 32 (Rank Plus Nullity Theorem) If the domain of a map has

finite dimension, the sum of its rank and its nullity equals the dimension of
its domain. 

Exercise 2.2.1 Let f : R3 → R2 be the map that sends (x, y, z) to (x − y, 0).

Draw a picture that illustrates for this particular map the concepts of kernel
and level set, and how complementary subspaces of the kernel are aliases of
the image.

If we specify domain V and kernel K, we have the level sets turned into
the vector space V/K as an alias of the image of any map with that domain
and kernel. Thus V/K can be viewed as a generic image for maps with
domain V and kernel K. To round out the picture, we now single out a map
that sends V onto V/K.

Proposition 33 The function p : V → V/K which sends each v ∈ V to

v + K is a map with kernel K.

Proof: Supposing, as we may, that K is the kernel of some map f : V →

W, let the map f [ : V → Image(f ) be defined by f [ (v) = f (v) for all v ∈ V.
Let Θ : Image(f ) → V/K be the isomorphism from the image of f to V/K
obtained through the correspondence of image elements with their level sets.
Notice that Θ sends w = f (v) to v + K. Then p = Θ ◦ f [ and is obviously a
map with kernel K. 
We call p the natural projection. It serves as a generic map for ob-
taining the generic image. For any map f : V → W with kernel K, we always
have the composition
2.3. SHORT EXACT SEQUENCES 27

V → V/K ←→ Image(f ) ,→ W
where V → V/K is generic, denoting the natural projection p, and the re-
mainder is the nongeneric specialization to f via an isomorphism and an
inclusion map. Our next result details how the generic map p in fact serves
in a more general way as a universal factor of maps with kernel containing
K.

Theorem 34 Let p : V → V/K be the natural projection and let f : V → W.

If K ⊂ Kernel(f ) then there is a unique induced map fV/K : V/K → W
such that fV/K ◦ p = f .

Proof: The prescription fV/K ◦ p = f determines exactly what fV/K must

do: for each v ∈ V, fV/K (p(v)) = fV/K (v + K) = f (v). For this to unam-
biguously define the value of fV/K at each element of V/K, it must be the
case that if u + K = v + K, then f (u) = f (v) . But this is true because
K ⊂ Kernel(f ) implies that f (k) = 0 for each k ∈ K. The unique function
fV/K so determined is readily shown to be a map. 

Exercise 2.2.2 Referring to the theorem above, fV/K is one-to-one if and

only if Kernel(f ) = K, and fV/K is onto if and only if f is onto.

The rank of the natural map p : V → V/K is the dimension of V/K,

also known as the codimension of the subspace K and denoted by codim K.
Thus the Rank Plus Nullity Theorem may also be expressed as

dim K + codim K = dim V

when dim V is finite.

Exercise 2.2.3 Let K be a subspace of finite codimension in the infinite-

dimensional vector space V. Then K ∼
= V.

2.3 Short Exact Sequences

There is a special type of map sequence that is a way to view a quotient.
The map sequence X → Y → Z is said to be exact at Y if the image of the
map going into Y is exactly the kernel of the map coming out of Y. A short
28 CHAPTER 2. FUNDAMENTALS OF MAPS

exact sequence is a map sequence of the form 0 → K → V → W → 0

which is exact at K, V, and W. Notice that here the map K → V is one-to-
one, and the map V → W is onto, in virtue of the exactness requirements
and the restricted nature of maps to or from 0 = {0}. Hence if V → W
corresponds to the function f , then it is easy to see that K is an alias of
Kernel(f ), and W is an alias of V/ Kernel(f ). Thus the short exact sequence
captures the idea of quotient modulo any subspace. It is true that with the
same K, V, and W, there are maps such that the original sequence with
all its arrows reversed is also a short exact sequence, one complementary to
the original, it might be said. This is due to the fact that K and W are
always aliases of complementary subspaces of V, which is something special
that vector space structure makes possible. Thus it is said that, for vector
spaces, a short exact sequence always splits, as illustrated by the short exact
sequence 0 → K → K ⊕ K → K → 0.
An example of the use of the exact sequence view of quotients stems from
considering the following diagram where the rows and columns are short
exact sequences. Each second map in a row or column is assumed to be an
inclusion, and each third map is assumed to be a natural projection.
0 0
↓ ↓
0 → K → H → H/K → 0
↓ ↓
0 → K → V → V/K → 0
↓ ↓
V/H (V/K)/(H/K)
↓ ↓
0 0
The sequence 0 → H → V → V/K → (V/K)/(H/K) → 0 is a subdiagram.
If we compose the pair of onto maps of V → V/K → (V/K)/(H/K) to yield
the composite onto map V → (V/K)/(H/K), we then have the sequence 0 →
H → V → (V/K)/(H/K) → 0 which is exact at H and at (V/K)/(H/K),
and would also be exact at V if H were the kernel of the composite. But
it is precisely the h ∈ H that map to the h + K ∈ H/K ⊂ V/K which then
are precisely the elements that map to H/K which is the 0 of (V/K)/(H/K).
Thus we have obtained the following isomorphism theorem.
Theorem 35 Let K be a subspace of H, and let H be a subspace of V. Then
V/H is isomorphic to (V/K)/(H/K). 
2.3. SHORT EXACT SEQUENCES 29

Exercise 2.3.1 Let V be R3 , let H be the (x, y)-plane, and let K be the
x-axis. Interpret the theorem above for this case.

Now consider the diagram below where X and Y are subspaces and again
the rows and columns are short exact sequences and the second maps are
inclusions while the third maps are natural projections.

0 0
↓ ↓
0 → X ∩Y → Y → Y/(X ∩ Y) → 0
↓ ↓
0 → X → X +Y → (X + Y)/X → 0
↓ ↓
X /(X ∩ Y) (X + Y)/Y
↓ ↓
0 0

As a subdiagram we have the sequence 0 → X ∩ Y → X → X + Y →

(X + Y)/Y → 0. Replacing the sequence X → X + Y → (X + Y)/Y with
its composite, the result would be a short exact sequence if the composite
X → (X + Y)/Y were an onto map with kernel X ∩ Y.
To see if the composite is onto, we must see if each element of (X + Y)/Y
is the image of some element of X . Now each element of (X + Y)/Y is the
image of some element w ∈ X + Y where w = u + v for some u ∈ X and
some v ∈ Y. But clearly the element u by itself has the same image modulo
Y. Hence the composite is onto. Also, the elements of X that map to the 0
of (X + Y)/Y, namely Y, are precisely those that then are also in Y. Hence
the kernel of the composite is X ∩ Y.
Thus we have obtained another isomorphism theorem and an immediate
corollary.

Theorem 36 Let X and Y be any subspaces of some vector space. Then

X / (X ∩ Y) is isomorphic to (X + Y)/Y. 

Corollary 37 (Grassmann’s Relation) Let X and Y be finite-dimensional

subspaces of a vector space. Then

dim X + dim Y = dim(X + Y) + dim(X ∩ Y). 

30 CHAPTER 2. FUNDAMENTALS OF MAPS

Exercise 2.3.2 Two 2-dimensional subspaces of R3 are either identical or

intersect in a 1-dimensional subspace.
0 0
Exercise 2.3.3 Let N , N , P, and P be subspaces of the same vector space
0 0 0 0
and let N ⊂ P and N ⊂ P . Set X = P ∩ P and Y = (P ∩ N ) + N.
0 0 0

Verify that X ∩ Y = (P ∩ N ) + (P ∩ N ) and X + Y = P ∩ P + N , so
that
0
0 0 0

P ∩P +N ∼ P ∩P ∼ P ∩ P +N
0 0 = 0 0 = 0 .
(P ∩ N ) + N (P ∩ N ) + (P ∩ N ) (P ∩ N ) + N

2.4 Projections and Reflections on V = W ⊕X

Relative to the decomposition of the vector space V by a pair W, X of direct
summands, we may define some noteworthy self-maps on V. Expressing the
general vector v ∈ V = W ⊕ X uniquely as v = w + x where w ∈ W and
x ∈ X , two types of such are defined by

PW|X (v) = w and RW|X (v) = w − x.

That these really are maps is easy to establish. We call PW|X the projection
onto W along X and we call RW|X the reflection in W along X . (The
function φ in the proof of Proposition 31 was the projection onto K along K,
so we have already employed the projection type to advantage.) Denoting
the identity map on V by I, we have

PX |W = I − PW|X

and
RW|X = PW|X − PX |W = I − 2PX |W = 2PW|X − I.
It bears mention that a given W generally has many different complements,
and if X and Y are two such, PW|X and PW|Y will generally differ.
The image of PW|X is W and its kernel is X . Thus PW|X is a self-map
with the special property that its image and its kernel are complements. The
image and kernel of RW|X are also complements, but trivially, as the kernel
of RW|X is 0 = {0}. The kernel of RW|X is 0 because w and x are equal only
if they are both 0 since W ∩ X = 0 from the definition of direct sum.
A double application of PW|X has the same effect as a single application:
2
PW|X = PW|X ◦ PW|X = PW|X (PW|X is idempotent). We then have PW|X ◦
2.4. PROJECTIONS AND REFLECTIONS ON V = W ⊕ X 31

PX |W = 0 (the constant map on V that sends everything to 0). We also

2 2
find that PW|X = PW|X implies that RW|X = I (RW|X is involutary). An
idempotent map always turns out to be some PW|X , and an involutary map
is almost always some RW|X , as we now explain.
Proposition 38 Let P : V → V be a map such that P 2 = P. Then P is the
projection onto Image (P ) along Kernel (P ).
Proof: Let w be in Image (P ), say w = P (v). Then
w = P 2 (v) = P (P (v)) = P (w)
so that if w is also in Kernel (P ), it must be that w = 0. Hence
Image (P ) ∩ Kernel (P ) = 0.
Let Q = I − P so that P (v) + Q (v) = I (v) = v and thus Image (P ) +
Image (Q) = V. But Image (Q) = Kernel (P ) since if x = Q (v) = I (v) −
P (v), then P (x) = P (v) − P 2 (v) = 0. Therefore
V = Image (P ) ⊕ Kernel (P ) .
Now if v = w + x, w ∈ Image (P ), x ∈ Kernel (P ), then P (v) = P (w + x) =
P (w) + P (x) = w + 0 = w, and P is indeed the projection onto Image (P )
along Kernel (P ). 
When V is over a field in which 1 + 1 = 0, the only possibility for RW|X
is the identity I. However, when V is over such a field there can be an
involutary map on V which is not the identity, for example the map of the
following exercise.
Exercise 2.4.1 Over the 2-element field F = {0, 1} (wherein 1 + 1 = 0), let
V be the vector space F 2 of 2-tuples of elements of F. Let F be the map on
V that interchanges (0, 1) and (1, 0). Then F ◦ F is the identity map.
The F of the exercise can reasonably be called a reflection, but it is not
some RW|X . If we say that any involutary map is a reflection, then reflections
of the form RW|X do not quite cover all the possibilities. However, when
1+1 6= 0, they do. (For any function f : V → V, we denote by Fix (f ) the set
of all elements in V that are fixed by f , i. e., Fix (f ) = {v ∈ V | f (v) = v}.)
Exercise 2.4.2 Let V be over a field such that 1 + 1 6= 0. Let R : V → V
be a map such that R2 = I. Then R = 2P − I, where P = 21 (R + I) is the
projection onto Fix (R) along Fix (−R).
32 CHAPTER 2. FUNDAMENTALS OF MAPS

2.5 Problems

1. Let f : S → T be a function. Then for all subsets A, B of S,

f (A ∩ B) ⊂ f (A) ∩ f (B) ,
but
f (A ∩ B) = f (A) ∩ f (B)
for all subsets A, B of S if and only if f is one-to-one. What about for ∪
instead of ∩? And what about for ∩ and for ∪ when f is replaced by f −1 ?

2. (Enhancing Theorem 19 for nontrivial codomains) Let B be a subset

of the vector space V and let W =6 {0} be a vector space over the same field.
Then B is independent if and only if every function f0 : B → W extends to
at least one map f : V → W, and B spans V if and only if every function
f0 : B → W extends to at most one map f : V → W. Hence B is a basis
for V if and only if every function f0 : B → W extends uniquely to a map
f : V → W.

3. Let B be a basis for the vector space V and let f : V → W be a map.

Then f is one-to-one if and only if f (B) is an independent set, and f is onto
if and only if f (B) spans W.

4. Deduce Theorem 25 as a corollary of the Rank Plus Nullity Theorem.

5. Let V be a finite-dimensional vector space of dimension n over the

finite field of q elements. Then the number of vectors in V is q n , the number
of nonzero vectors that are not in a given m-dimensional subspace is
(q n − 1) − (q m − 1) = (q n − q m ) ,
and the number of different basis sets that can be extracted from V is
(q n − 1) (q n − q) · · · (q n − q n−1 )
n!
since there are
(q n − 1) (q n − q) · · · q n − q n−1


2.5. PROBLEMS 33

different sequences of n distinct basis vectors that can be drawn from V.

6. (Correspondence Theorem) Let U CV. The natural projection p : V →

V/U puts the subspaces of V that contain U in one-to-one correspondence
with the subspaces of V/U, and every subspace of V/U is of the form W/U
for some subspace W that contains U.

7. PW1 |X 1 + PW2 |X 2 = PW|X if and only if PW1 |X 1 ◦ PW2 |X 2 = PW2 |X 2 ◦

PW1 |X 1 = 0, in which case W = W1 ⊕ W2 and X = X1 ∩ X2 . What happens
when PW1 |X 1 − PW2 |X 2 = PW|X is considered instead?

8. There is a map on R2 that sends all three of (1,0), (1,1), and (1,2) to the
same nonzero vector and therefore sends a dependent set to an independent
one.
34 CHAPTER 2. FUNDAMENTALS OF MAPS
Chapter 3

More on Maps and Structure

3.1 Function Spaces and Map Spaces

The set W D of all functions from a set D into the vector space W may be
made into a vector space by defining the operations in a pointwise fashion.
That is, suppose that f and g are in W D , and define f + g by (f + g) (x) =
f (x) + g (x), and for any scalar a define a · f by (a · f ) (x) = a · (f (x)).
This makes W D into a vector space, and it is this vector space that we
mean whenever we refer to W D as a function space. Considering n-tuples
as functions with domain {1, . . . , n} and the field F as a vector space over
itself, we see that the familiar F n is an example of a function space.
Subspaces of a function space give us more examples of vector spaces. For
instance, consider the set {V → W} of all maps from the vector space V into
the vector space W. It is easy to verify that it is a subspace of the function
space W V . It is this subspace that we mean when we refer to {V → W} as
a map space.

3.2 Dual of a Vector Space, Dual of a Basis

By a linear functional on a vector space V over a field F is meant an
element of the map space {V → F 1 }. (Since F 1 is just F viewed as a vector
space over itself, a linear functional is also commonly described as a map
into F.) For a vector space V over a field F, the map space {V → F 1 } will
be called the dual of V and will be denoted by V > .
Let the vector space V have the basis B. For each x ∈ B, let x> be the

35
36 CHAPTER 3. MORE ON MAPS AND STRUCTURE

linear functional on V that sends x to 1 and sends all other elements of B to

0. We call x> the coordinate function corresponding to the basis vector
x. The set B # of all such coordinate
P functions is an independent set. For if
for some finite X ⊂ B we have x∈X ax · x> (v) = 0 for all v ∈ V, then for
any v ∈ X we find that av = 0 so there can be no dependency in B # .
When the elements of B # span V > , we say that B has B # as its dual
and we may then replace B # with the more descriptive notation B > . When
V is finite-dimensional, B # does span V > , we claim. For the general linear
functional f is defined by prescribing its value on each element of B, and given
any v ∈ B we have f (v) = x∈B f (x)·x> (v). Therefore any
P
P linear functional
on V may be expressed as the linear combination f = x∈B f (x) · x> , so
that B # spans V > as we claimed.
We thus have the following result.

Theorem 39 Let V be a finite-dimensional vector space. Then any basis of

V has a dual, and therefore V > has the same dimension as V. 

When V is infinite-dimensional, B # need not span V > .

Example 40 Let N denote the natural numbers {0, 1, . . .}. Let V be the
subspace of the function space RN consisting of only those real sequences
a0 , a1 , . . . that are ultimately zero, i. e., for which there exists a smallest
N ∈ N (called the length of the ultimately zero sequence), such that P an = 0
for all n > N . Then φ : V → R defined by φ(a0 , a1 , . . .) = n∈N αn an
is a linear functional given any (not necessarily ultimately zero) sequence
of real coefficients α0 , α1 , . . .. (Given α0 , α1 , . . . that is not ultimately zero,
it cannot be replaced by an ultimately zero sequence that produces the same
functional, since there would always be sequences a0 , a1 , . . . of greater length
for which the results would differ.) The B # derived from the
basis B =
{(1, 0, 0, . . .) , (0, 1, 0, 0, . . .) , . . .} for V does not span V > because B # only
contains elements corresponding to sequences α0 , α1 , . . . that are ultimately
zero. Also note that by Theorem 19, each element of V > is determined by an
unrestricted choice of value at each and every element of B.

Fixing an element v ∈ V, the scalar-valued function Fv defined on V >

by Fv (f ) = f (v) is a linear functional on V > , i. e., Fv is an element of

>
V >> = V > . The association of each such v ∈ V with the corresponding
Fv ∈ V >> gives a map Θ from V into V >> . Θ is one-to-one, we claim. For
3.3. ANNIHILATORS 37

suppose that v and w are in V and that Fv = Fw so that for all f ∈ V > ,
f (v) = f (w). Then f (v − w) = 0 for every f ∈ V > . Let B be a basis for V,
and for each x ∈ B let x> be the corresponding coordinate function. Let X
be a finite subset of B such that v − w = x∈X ax · x. Applying x> to v − w
P
for each x ∈ X , we conclude that there is no nonzero ax . Hence v = w, and
Θ is one-to-one as claimed.
Thus it has been shown that each vector v ∈ V defines a unique element
Fv ∈ V >> . But the question still remains as to whether Θ is onto, that is
whether every F ∈ V >> is an Fv for some v ∈ V. We claim that Θ is onto if
V is finite-dimensional. This isbecause if we have bases B and B > as above
and we form the basis B >> = x>> >>
1 , . . . , xn dual to B > , we readily verify
that x>>
i = Fxi for each i. Thus for F = a1 · x>> 1 + · · · + an · x>>
n , and
v = a1 · x1 + · · · + an · xn , we have F = Fv .
We have proved the following.

Theorem 41 The map Θ : V → V >> which sends v ∈ V to the linear func-

tional Fv on V > defined by Fv (f ) = f (v) is always one-to-one, and moreover
is onto if V is finite-dimensional. 

Θ does not depend on any choice of basis, so V and Θ (V) ⊂ V >> are
isomorphic in a “natural” basis-invariant manner. We say that Θ naturally
embeds V in V >> and we call Θ the natural injection of V into V >> ,
or, when it is onto, the natural correspondence between V and V >> .
Independent of basis choices, a vector in V can be considered to act directly
on V > as a linear functional. If one writes f (v), where f ∈ V > and v ∈ V,
either of f or v can be the variable. When f is the variable, then v is
really playing the rôle of a linear functional on V > , i. e., an element of
V >> . One often sees notation such as hf, vi used for f (v) in order to more
clearly indicate that f and v are to viewed as being on similar footing, with
either being a functional of the other. Thus, as convenience dictates, we may
identify V with Θ (V) ⊂ V >> , and in particular, we agree always to do so
whenever Θ is onto.

3.3 Annihilators
Let S be a subset of the vector space V. By the annihilator S 0 of S is
meant the subset of V > which contains all the linear functionals f such that
f (s) = 0 for all s ∈ S. It is easy to see that S 0 is actually a subspace of V >
38 CHAPTER 3. MORE ON MAPS AND STRUCTURE

and that S 0 = hSi0 . Obviously, {0}0 = V > . Also, since a linear functional
that is not the zero functional, i. e., which is not identically zero, must have
at least one nonzero value, V 0 = {0}.

Exercise 3.3.1 Because of the natural injection of V into V >> , the only
x ∈ V such that f (x) = 0 for all f ∈ V > is x = 0.

When V is finite-dimensional, the dimension of a subspace determines

that of its annihilator.

Theorem 42 Let V be an n-dimensional vector space and let U C V have

dimension m. Then U 0 has dimension n − m.

Proof: Let B = {x1 , . . . , xn } be a basis for V such that {x1 , . . . , xm }

is a basis for U. We claim that x> >
m+1 , . . . , xn ⊂ B > spans U 0 . Clearly,
x> >
⊂ U 0 . On the other hand, if f = b1 · x> >

  0
m+1 , . . . , xn 1 + · · · + bn · x n ∈ U
then f (x) = 0
 for each x ∈ {x 1 , . . . , xm } so that b1 = · · · = bm = 0 and
therefore U 0 ⊂ x> m+1 , . . . , x >
n . 

Exercise 3.3.2 Let V be a finite-dimensional vector space. For any U C V,

0
(U 0 ) = U, remembering that we identify V with V >> .

3.4 Dual of a Map

Let f : V → W be a map. Then for ϕ ∈ W > , the prescription f > (ϕ) = ϕ◦f ,
or (f > (ϕ)) (v) = ϕ(f (v)) for all v ∈ V, defines a function f > : W > → V > .
It is easy to show that f > is a map. f > is called the dual of f . The dual of
an identity map is an identity map. The dual of the composite map f ◦ g is
g > ◦ f > . Thus, if f is a bijective map, the dual of its inverse is the inverse of
its dual.

Exercise 3.4.1 Let V and W be finite-dimensional vector spaces P with re-

spective bases B and C. Let the map f : V → W satisfy f (x) = y∈C ay,x · y
for each x ∈ B. Then f > satisfies f > y > = x∈B ay,x · x> for each y ∈ C

P

(where x> ∈ B > and y > ∈ C > are coordinate functions corresponding respec-
tively to x ∈ B and y ∈ C, of course.)

The kernel of f > is found to have a special property.

3.5. THE CONTRAGREDIENT 39

Theorem 43 The kernel of f > is the annihilator of the image of the map
f : V → W.

Proof: ϕ is in the kernel of f > ⇔ (f > (ϕ)) (v) = 0 for all v ∈ V ⇔

ϕ (f (v)) = 0 for all v ∈ V ⇔ ϕ ∈ (f (V))0 . 

When the codomain of f has finite dimension m, say, then the image of f
has finite dimension r, say. Hence, the kernel of f > (which is the annihilator
of the image of f ) must have dimension m − r. By the Rank Plus Nullity
Theorem, the rank of f > must then be r. Thus we have the following result.

Theorem 44 (Rank Theorem) When the map f has a finite-dimensional

codomain, the rank of f > equals the rank of f . 

Exercise 3.4.2 Let f be a map from one finite-dimensional vector space

>
to another. Then, making the standard identifications, f > = f and the
kernel of f is the annihilator of the image of f > .

Exercise 3.4.3 For Φ ⊂ V > , let Φ denote the subset of V that contains all
the vectors v such that ϕ (v) = 0 for all ϕ ∈ Φ. Prove analogs of Theorems 42
and 43. Finally prove that when the map f has a finite-dimensional domain,
the rank of f > equals the rank of f (another Rank Theorem).

3.5 The Contragredient

Suppose that V is a finite-dimensional vector space with basis B and dual
basis B > , and suppose that W is an alias of V via the isomorphism f : V → W
which then sends the basis B to the basis f (B). Then the contragredient
>
−1
f −> = (f −1 ) = f > maps V > isomorphically onto W > and maps B >
onto the dual of the basis f (B).

Theorem 45 Let f : V → W be an isomorphism between finite-dimensional

vector spaces. Let B be a basis for V. Then the dual of the basis f (B) is
f −> B > .

Proof: Let x, y ∈ B. We then have f −> x> (f (y)) = x> ◦ f −1 (f (y)) =

x> (y). 
40 CHAPTER 3. MORE ON MAPS AND STRUCTURE

This result is most often applied in the case when V = W and the iso-
morphism f (now an automorphism) is specifically being used to effect a
change of basis and the contragredient is being used to effect the correspond-
ing change of the dual basis.

3.6 Product Spaces

The function space concept introduced at the beginning of this chapter can be
profitably generalized. Instead of considering functions each value of which
lies in the same vector space W, we are now going to allow different vector
spaces (all over the same field) for different elements of the domain D. Thus
for each x ∈ D there is to be a vector space Wx , with each of these Wx being
over the same field, and we form all functions f such that f (x) ∈ Wx . The set
of all such f is the familiar Cartesian product of the Wx . Any such Cartesian
product of vector spaces may be made into a vector space using pointwise
operations in the same way that we earlier used them to make W D into a
function space, and it is this vector space that we mean whenever we refer to
the Cartesian product of vector spaces as a product space. Sometimes we
also refer to a product space or a Cartesian product simply as a product,
Q
when there is no chance of confusion. We use the general notation x∈D Wx
for a product, but for the product of n vector spaces W1 , . . . , Wn we often
write W1 × · · · × Wn . For a field F, we recognize F × · · · × F (with n factors)
as the familiar F n .
When there are only a finite number of factors, a product space is the di-
rect sum of the “internalized” factors. For example, W1 ×W2 = (W1 × {0})⊕
({0} × W2 ).
Exercise 3.6.1 If U1 and U2 are subspaces of the same vector space and have
only 0 in common, then U1 × U2 ∼ = U1 ⊕ U2 in a manner not requiring any
choice of basis.
However, when there are an infinite number of factors, a product space
is not equal to the direct sum of “internalized” factors.
Exercise 3.6.2 Let N = {0, 1, 2, ...} be the set of natural numbers and let
F be a field. Then the function space F N , which is the same thing as the
product space with each factor equal to F andLwith one such factor for each
natural number, is not equal to its subspace i∈N hxi i where xi (j) is equal
to 1 if i = j and is equal to 0 otherwise.
3.7. MAPS INVOLVING PRODUCTS 41

Selecting from a product space only those functions that have finitely
many nonzero values gives a subspace called the weak product space. As
an example, the set of all formal power series with coefficients in a field F
is a product space (isomorphic to the F N of the exercise above) for which
the corresponding weak product space is the set of all polynomials with
coefficients in F. The weak product space always equals the direct sum of
the “internalized” factors, and
U for that reason is also called the external
direct sum. WeQwill write x∈D Wx for the weak product space derived
from the product x∈D Wx .

3.7 Maps Involving Products

We single out some basic families of maps Q that involveUproducts and their
factors. The function πx that sends f ∈ x∈D Wx (or x∈D Wx ) to f (x) ∈
Wx , is called the canonical Q projection U onto Wx . The function ηx that
sends v ∈ Wx to the f ∈ x∈D Wx (or x∈D Wx ) such that f (x) = v and
f (y) = 0 for y 6= x, is called the canonical injection from Wx . It is easy
to verify that πx and ηx are maps. Also, we introduce ξx = ηx ◦ πx , called the
component projection for component x. Note that πx ◦ ηx = idx (identity
P Wx ), πx ◦ ηy = 0 whenever
on P x 6= y, and ξx ◦ ξx = ξx . Note also that
x∈D ξx (f ) = f so that x∈D ξx is the identity function, where we have
made the conventionQthat the pointwise sum of any number of 0s is 0.
A map F : V → x∈D Wx from a vector space V into the product of the
vector spaces Wx is uniquely determined by the map projections πx ◦ F .
Q
Theorem 46 There exists exactly one map F : V → x∈D Wx such that
πx ◦ F = ϕx where for each x ∈ D, ϕx : V → Wx is a prescribed map.
Q
Proof: The function F that sends v ∈ V to the f ∈ x∈D Wx such that
Q seen to be a map such that πx ◦F = ϕx . Suppose that
f (x) = ϕx (v) is readily
the map G : V → x∈D Wx satisfies πx ◦ G = ϕx for each x ∈ D. For v ∈ V
let G (v) = g. Then g (x) = πx (g) = πx (G (v)) = (πx ◦G) (v) = ϕx (v) = f (x)
and hence g = f . But then G (v) = f = F (v) so that G = F , which shows
the uniqueness of F . 

For weak products, it is the maps from them, rather than the maps into
them, that have the special property.
42 CHAPTER 3. MORE ON MAPS AND STRUCTURE
U
Theorem 47 There exists exactly one map F : x∈D Wx → V such that
F ◦ ηx = ϕx where for each x ∈ D, ϕx : Wx → V is a prescribed map.
U P
Proof: The function F that sends f ∈ x∈D Wx to {x∈D|f (x)6=0} ϕx (f (x))
U readily seen to be a map such that F ◦ ηx = ϕx . Suppose that
is U the map G :
x∈DPWx → V satisfies G ◦ ηx = ϕx for each x ∈ PD. Let f ∈ x∈D Wx . Then
f = {x∈D|f (x)6=0} ηx (f (x)) so that G (f ) = {x∈D|f (x)6=0} G (ηx (f (x))) =
P
{x∈D|f (x)6=0} ϕx (f (x)) = F (f ), which shows the uniqueness of F . 
U L
Exercise 3.7.1 x∈D Wx = x∈D ηx (Wx ).
L
U ∼ U
Exercise 3.7.2 x∈D x = x∈D Ux in a manner not requiring any choice
of basis.

Exercise 3.7.3 Let W1 , . . . , Wn be finite dimensional vector spaces over the

same field. Then

dim (W1 × · · · × Wn ) = dim W1 + · · · + dim Wn .

The last two theorems lay a foundation for two important isomorphisms.
As before, with all vector spaces over the same field, let the vector space V
be given, and let a vector space Wx be given for each x in some set D. Then,
utilizing map spaces, we may form the pairs of vector spaces
( )
Y Y
M= {V → Wx } , N = V → Wx
x∈D x∈D

and ( )
0 0
Y ]
M = {Wx → V} , N = Wx → V .
x∈D x∈D

Theorem 48 M ∼
= N and M ∼
0 0
=N .

Proof:
f ∈ M means that for each x ∈ D, f (x) = ϕx for some map ϕx : V →
Wx , and hence there exists exactly one map F ∈ N such that πx ◦ F = f (x)
for each x ∈ D. This association of F with f therefore constitutes a well-
defined function Λ : M → N which is easily seen to be a map. The map
Λ is one-to-one. For if Λ sends both f and g to F , then for each x ∈ D,
3.7. MAPS INVOLVING PRODUCTS 43

πx ◦ F = f (x) = g (x), and f and g must therefore be equal. Also, each

F ∈ N is the image under Λ of the f ∈ M such that f (x) = πx ◦ F so that
Λ is also onto. Hence Λ is an isomorphism from M onto N .
A similar argument shows that M ∼
0 0
=N . 

>
U ∼
= x∈D Wx> . (Compare with Example 40 at
Q
Exercise 3.7.4 x∈D Wx
the beginning of this chapter.)

Exercise 3.7.5 (W1 × · · · × Wn )> ∼

= W1> × · · · × Wn> .

We note that the isomorphisms of the two exercises and theorem above
do not depend on any choice of basis. This will generally be the case for
the isomorphisms we will be establishing. From now on, we will usually skip
pointing out when an isomorphism does not depend on specific choices, but
will endeavor to point out any opposite cases.
Up to isomorphism, the order of the factors in a product does not matter.
This can be proved quite readily directly from the product definition, but the
theorem above on maps into a product gives an interesting approach from
another perspective. We also incorporate the “obvious” result that using
aliases of the factors does not affect the product, up to isomorphism.

Theorem 49 x∈D Wx ∼
Q Q 0
= x∈D Wσ(x) where σ is a bijection of D onto
0
itself, and for each x ∈ D there is an isomorphism θx : Wx → Wx .

Q
Proof: In Theorem 46, take the product to be x∈D Wx , take V =
Q 0
x∈D Wσ(x) , and take ϕx = θx ◦ πσ −1 (x) so that there exists a map Ψ from
Q 0 Q
x∈D Wσ(x) to x∈D Wx such that πx ◦ Ψ = θx ◦ πσ −1 (x) . Interchanging
the product spaces and
Q applying the theorem again, there exists a map
0
Φ from x∈D Wx to x∈D Wσ(x) such that πσ−1 (x) ◦ Φ = θx−1 ◦ πx . Then
Q

πx ◦ Ψ ◦ Φ = θx ◦ πσ−1 (x) ◦ Φ = πx , and πσ−1 (x) ◦ Φ ◦ Ψ = θx−1 ◦ πx ◦ Ψ = πσ−1 (x) .

Now applying the theorem to the two cases where the V and the product
space are the same, the unique map determined in each case must be the
identity. Hence Ψ ◦ Φ and Φ ◦ Ψ must each be identity Q maps, thus Qmaking0 Φ
and Ψ a pair of mutually inverse maps that link x∈D Wx and x∈D Wσ(x)
isomorphically. 
44 CHAPTER 3. MORE ON MAPS AND STRUCTURE

3.8 Problems

1. Let V be a vector space over the field F. Let f ∈ V > . Then f > is the
>
map that sends the element ϕ of (F 1 ) to (ϕ (1)) · f ∈ V > .

2. Let V C W. Let j : V → W be the inclusion map of V into W which

sends elements in V to themselves. Then j > is the corresponding “restriction
of domain” operation on the functionals of W > , restricting each of their
domains to V from the original W. (Thus, using a frequently-seen notation,
j > (ϕ) = ϕ|V for any ϕ ∈ W > .)

3. Let f : V → W be a map. Then for all w ∈ W the equation

f (x) = w

has a unique solution for x if and only if the equation

f (x) = 0

has only the trivial solution x = 0. (This result is sometimes called the
Alternative Theorem. However, see the next problem.)

4. (Alternative Theorem) Let V and W be finite-dimensional and let

f : V → W be a map. Then for a given w there exists x such that

f (x) = w

if and only if


0
w ∈ Kernel f > ,
i. e., if and only if ϕ (w) = 0 for all ϕ ∈ Kernel f > . (The reason this

is called the Alternative Theorem is because it is usually phrased in the

equivalent form: Either the equation f (x) = w can be solved for x, or there
is a ϕ ∈ Kernel f > such that ϕ (w) 6= 0.)

5. Let V and W be finite-dimensional and let f : V → W be a map.

Then f is onto if and only if f > is one-to-one and f > is onto if and only if f
is one-to-one.
3.8. PROBLEMS 45

6. Let V and W have the respective finite dimensions m and n over the
finite field F of q elements. How many maps V → W are there? If m > n
how many of these are onto, if m 6 n how many are one-to-one, and if m = n
how many are isomorphisms?
46 CHAPTER 3. MORE ON MAPS AND STRUCTURE
Chapter 4

Multilinearity and Tensoring

4.1 Multilinear Transformations

Maps, or linear transformations, are functions of a single vector variable.
Also important are the multilinear transformations. These are the functions
of several vector variables (i. e., functions on the product space V1 ×· · ·×Vn )
which are linear in each variable separately. In detail, we call the function
f : V1 × · · · × Vn → W, where the Vi and W are vector spaces over the same
field F, a multilinear transformation, if for each k

f (v1 , . . . , vk−1 , a · u + b · v, . . . , vn ) =

= a · f (v1 , . . . , vk−1 , u, . . . , vn ) + b · f (v1 , . . . , vk−1 , v, . . . , vn )

for all vectors u, v, v1 , . . . , vn and all scalars a, b. We also say that such an
f is n-linear (bilinear, trilinear when n = 2, 3). When W = F, we
call a multilinear transformation a multilinear functional. In general, a
multilinear transformation is not a map on V1 × · · · × Vn . However, we will
soon construct a vector space on which the linear transformations “are” the
multilinear transformations on V1 × · · · × Vn .
A unique multilinear transformation is determined by the arbitrary as-
signment of values on all tuples of basis vectors.

Theorem 50 Let V1 , . . . , Vn , W be vector spaces over the same field and for
each Vi let the basis Bi be given. Then given the function f0 : B1 × · · · × Bn →
W, there is a unique multilinear transformation f : V1 × · · · × Vn → W such
that f agrees with f0 on B1 × · · · × Bn .

47
48 CHAPTER 4. MULTILINEARITY AND TENSORING

Proof: Given vi ∈ Vi there is a uniqueP finite subset Xi of Bi and unique

nonzero scalars axi such that vi = xi ∈Xi axi · xi . The n-linearity of f :
V1 × · · · × Vn → W implies
X X
f (v1 , . . . , vn ) = ··· ax1 · · · axn · f (x1 , . . . , xn ) .
x1 ∈X1 xn ∈Xn

Hence for f to agree with f0 on B1 × · · · × Bn , it can only have the value

X X
f (v1 , . . . , vn ) = ··· ax1 · · · axn · f0 (x1 , . . . , xn ) .
x1 ∈X1 xn ∈Xn

On the other hand, setting

X X
f (v1 , . . . , vn ) = ··· ax1 · · · axn · f0 (x1 , . . . , xn )
x1 ∈X1 xn ∈Xn

does define a function f : V1 × · · · × Vn → W which clearly agrees with f0

on B1 × · · · × Bn , and this f is readily verified to be n-linear. 

4.2 Functional Product Spaces

Let S1 , . . . , Sn be sets of linear functionals on the respective vector spaces
V1 , . . . , Vn all over the same field F. By the functional product S1 · · · Sn
is meant the set of all products f1 · · · fn with each fi ∈ Si , where by such
a product of linear functionals is meant the function f1 · · · fn : V1 × · · · ×
Vn → F such that f1 · · · fn (v1 , . . . , vn ) = f1 (v1 ) · · · fn (vn ). The function-
space subspace hS1 · · · Sn i obtained by taking all linear combinations of finite
subsets of the functional product S1 · · · Sn is called the functional product
space generated by S1 · · · Sn .

Exercise 4.2.1 Each element of hS1 · · · Sn i is a multilinear functional on

V1 × · · · × Vn .

Exercise 4.2.2 hhS1 i · · · hSn ii = hS1 · · · Sn i.

Lemma 51 Let W1 , . . . , Wn be vector spaces of linear functionals, all over

the same field, with respective bases B1 , . . . , Bn . Then B1 · · · Bn is an inde-
pendent set.
4.3. FUNCTIONAL TENSOR PRODUCTS 49

Proof: Induction will be employed. The result is clearly true for n =

1, and suppose that it P is true for n = N − 1. Let X be a finite subset
of B1 · · · BN such that x∈X ax · x = 0. Each x ∈ X has the form yz
where
P y ∈ B · ·
1P P · BN −1 and z ∈ BN . Collecting terms on the distinct z,
x∈X axP· x = z ( x=yz ax · y)z. The independence of BN implies that for
each z, x=yz ax · y = 0, and the assumed independence of B1 · · · BN −1 then
implies that each ax is zero. Hence B1 · · · BN is an independent set, and by
induction, B1 · · · Bn is an independent set for each n. 
Proposition 52 Let W1 , . . . , Wn be vector spaces of linear functionals, all
over the same field, with respective bases B1 , . . . , Bn . Then B1 · · · Bn is a basis
for hW1 · · · Wn i.
Proof: By the exercise above, hW1 · · · Wn i = hB1 · · · Bn i and hence B1 · · · Bn
spans hW1 · · · Wn i. By the lemma above, B1 · · · Bn is an independent set.
Thus B1 · · · Bn is a basis for hW1 · · · Wn i. 

4.3 Functional Tensor Products

In the previous chapter, it was shown that the vector space V may be embed-
ded in its double dual V >> independently of any choice of basis, effectively
making each vector in V into a linear functional on V > . Thus given the vector
spaces V1 , . . . , Vn , all over the same field F, and the natural injection Θi that
embeds each Vi in its double dual, we may form the functional product space
hΘ1 (V1 ) · · · Θn (Vn )i which will be called the N functional
N tensor product of
the Vi and which will be denoted by V1 · · · Vn . Similarly, an element
Θ1 (v1 ) · · · Θn (vn ) of Θ1 (V1 ) · · · Θn (Vn ) will be called the functional tensor
product of the vi and will be denoted by v1 ⊗ · · · ⊗ vn .
If W is a vector space over the same field as N the VN
i , the “universal”
multilinear transformation Ξ : V1 × · · · × Vn → V1 · · · Vn which sends
(v1 , . . . , vn ) to v1 ⊗ · · · ⊗ vn exchanges the multilinear transformations
N N f :
V1 × · · · × Vn → W with the linear transformations ϕ : V1 · · · Vn → W
via the composition of functions f = ϕ ◦ Ξ.
Theorem 53 Let V1 , . . . , Vn , W be vector spaces over the same field. Then
for each multilinear transformation N f : V1N× · · · × Vn → W there exists a
unique linear transformation N ϕ : V1N · · · Vn → W such that f = ϕ ◦ Ξ,
where Ξ : V1 × · · · × Vn → V1 · · · Vn is the tensor product function that
sends (v1 , . . . , vn ) to v1 ⊗ · · · ⊗ vn .
50 CHAPTER 4. MULTILINEARITY AND TENSORING

Proof: For each Vi , choose some basis Bi . Then by the N proposition

N above,
B = {x1 ⊗ · · · ⊗ xn | x1 ∈ B1 , . . . , xn ∈ Bn } is a basis for V1 · · · Vn . Given
N N f : V1 × · · · × Vn → W, consider the linear
the multilinear transformation
transformation ϕ : V1 · · · Vn → W such that for x1 ∈ B1 , . . . , xn ∈ Bn
ϕ (x1 ⊗ · · · ⊗ xn ) = ϕ (Ξ(x1 , . . . , xn )) = f (x1 , . . . , xn ) .
Since ϕ ◦ Ξ is clearly n-linear, it must then equal f by the theorem above.
Every ϕ determines an n-linear f via f = ϕ ◦ Ξ, but since the values of ϕ on
B determine it completely, and the values of f on B1 × · · · × Bn determine it
completely, a different ϕ determines a different f . 
Thus, while multilinear transformations themselves are generally not maps
on V1 × · · · × VnN
, theyNdo correspond one-to-one to the maps on the related
vector space V1 · · · Vn = h{v1 ⊗ · · · ⊗ vn }i. In fact, we can interpret
this correspondence
n as an isomorphism
o if we introduce the function space
(n−linear)
subspace V1 × · · · × Vn → W of n-linear functions, which is then
N N
clearlynisomorphic to the map o space {V1 · · · Vn → W}. The vector
(n−linear)
space V1 × · · · × Vn → F of multilinear functionals is therefore iso-
>
N N
morphic to (V1 · · · Vn ) . If all the Vi are finite-dimensional, we then
n (n−linear)
o>
have V1 · · · Vn ∼
N N
= V1 × · · · × Vn → F .
When the Vi are finite-dimensional, we may discard any even number of
consecutive dualization operators, since
Nwe identify a finite-dimensional
 space
with its double dual. Therefore, V1> · · · Vn> is just V1> · · · Vn> . Thus
N

n (n−linear)
o
V1> · · · Vn> is a subspace of V1 × · · · × Vn
N N
→ F , and it is easy
to see that both have the same dimension, so they are actually equal. This
and our observations immediately above then give us the following result.
Theorem 54 Let V1 , . . . , Vn be vector spaces over the field F. Then
N N >∼n (n−linear)
o
(V1 · · · Vn ) = V1 × · · · × Vn → F .
If the Vi are all finite-dimensional, we also have
N N ∼n (n−linear)
o>
V1 · · · Vn = V1 × · · · × Vn → F
and
n (n−linear)
o
Vn )> ∼
= V1> · · · Vn> = V1 × · · · × Vn
N N N N
(V1 ··· → F .
4.4. TENSOR PRODUCTS IN GENERAL 51

4.4 Tensor Products in General

Theorem 53 contains the essence of the tensor product concept. By mak-
ingNa definition
N out of the universal property that this theorem ascribes to
V1 · · · Vn and Ξ, a general tensor product concept results. Let the vec-
tor spaces V1 , . . . , Vn each over the same field F be given. A tensor product
of V1 , . . . , Vn (in that order) is a vector space Π (a tensor product space)
over F along with an n-linear function Υ : V1 ×· · ·×Vn → Π (a tensor prod-
uct function) such that given any vector space W, every n-linear function
f : V1 × · · · × Vn → W is equal Nto ϕN ◦ Υ for a unique map ϕ : Π → W. The
functional tensor product V1 · · · Vn along with the functional tensor
product function Ξ are thus an example of a tensor product of V1 , . . . , Vn .
Another tensor product of the same factor spaces may be defined using
any alias of a tensor product space along with the appropriate tensor product
function.

Theorem 55 Let V1 , . . . , Vn be vector spaces over the same field. Then if Π

along with the n-linear function Υ : V1 × · · · × Vn → Π is a tensor product
of V1 , . . . , Vn , and Θ : Π → Π0 is an isomorphism, then Π0 along with Θ ◦ Υ
is also a tensor product of V1 , . . . , Vn .

Proof: Given the vector space W, the n-linear function f : V1 × · · · ×

Vn → W is equal to ϕ ◦ Υ for a unique map ϕ : Π → W. The map
ϕ ◦ Θ−1 therefore satisfies (ϕ ◦ Θ−1 ) ◦ (Θ ◦ Υ) = f . On the other hand, if
ϕ0 ◦ (Θ ◦ Υ) = (ϕ0 ◦ Θ) ◦ Υ = f , then by the uniqueness of ϕ, ϕ0 ◦ Θ = ϕ and
hence ϕ0 = ϕ ◦ Θ−1 . 

Exercise 4.4.1 Make sense of the tensor product concept in the case where
n = 1, so that we have 1-linearity and the tensor product of a single factor.

If in the manner of the previous theorem, we have isomorphic tensor prod-

uct spaces with the tensor product functions related by the isomorphism as
in the theorem, we say that the tensor products are linked by the isomor-
phism of their tensor product spaces. We now show that tensor products of
the same factors always have isomorphic tensor product spaces, and in fact
there is a unique isomorphism between them that links them.

Theorem 56 Let V1 , . . . , Vn be vector spaces over the same field. Then if

Π along with the n-linear function Υ : V1 × · · · × Vn → Π, and Π0 along
52 CHAPTER 4. MULTILINEARITY AND TENSORING

with the n-linear function Υ0 : V1 × · · · × Vn → Π0 , are each tensor products

of V1 , . . . , Vn , then there is a unique isomorphism Θ : Π → Π0 such that
Θ ◦ Υ = Υ0 .

Proof: In the tensor product definition above, take W = Π0 and f = Υ0 .

Then there is a unique map Θ : Π → Π0 such that Υ0 = Θ◦Υ. Similarly, there
is a unique map Θ0 : Π0 → Π such that Υ = Θ0 ◦ Υ0 . Hence Υ0 = Θ ◦ Θ0 ◦ Υ0
and Υ = Θ0 ◦ Θ ◦ Υ. Applying the tensor product definition again, the unique
map ϕ : Π → Π such that Υ = ϕ ◦ Υ must be the identity. We conclude that
Θ0 ◦ Θ is the identity and the same for Θ ◦ Θ0 . Hence the unique map Θ has
the inverse Θ0 and is therefore an isomorphism. 
N N
We will write any tensor product space as V1 · · · Vn just as we did
for the functional tensor product. Similarly we will write v1 ⊗ · · · ⊗ vn for
Υ (v1 , . . . , vn ), and call it the tensor product of the vectors v1 , . . . , vn , just
as we did for the functional tensor product Ξ (v1 , . . . , vn ). With isomorphic
tensor product spaces, we will always assume that the tensor products are
linked by the isomorphism. By assuming this, tensor products of the same
vectors will then correspond among all tensor products of the same factor
spaces: v1 ⊗· · ·⊗vn in one of these spaces is always an isomorph of v1 ⊗· · ·⊗vn
in any of the others.

Exercise 4.4.2 A tensor product space is spanned by the image of its as-
sociated
N tensor
N product function: h{v1 ⊗ · · · ⊗ vn | v1 ∈ V1 , . . . , vn ∈ Vn }i =
V1 · · · Vn .

Up to isomorphism, the order of the factors in a tensor product does not

matter. We also incorporate the “obvious” result that using aliases of the
factors does not affect the result, up to isomorphism.
0 0
Theorem 57 Let the vector spaces V1 , . . . , Vn , all over the N sameN field, be
(possibly reordered ) aliases of V1 , . . . , Vn , and let V⊗ = V1 · · · Vn along
with the tensor product function
0 0 N NΥ :0 V× = V1 × · · · × Vn → V⊗ be a tensor
product, and let V⊗ = V1 · · · Vn along with the tensor product function
Υ : V× = V1 × · · · × Vn → V⊗ be a tensor product. Then V⊗ ∼
0 0 0 0 0 0
= V⊗ .
0
Proof: Let Φ : V× → V× denote the isomorphism that exists from the
0 0
product space V× to the product space V× . It is easy to see that Υ ◦ Φ
0
and Υ ◦ Φ−1 are n-linear. Hence there is a (unique) map Θ : V⊗ → V⊗ such
4.4. TENSOR PRODUCTS IN GENERAL 53

0 0 0
that Θ ◦ Υ = Υ ◦ Φ and there is a (unique) map Θ : V⊗ → V⊗ such that
0 0 0 0 0
Θ ◦ Υ = Υ ◦ Φ−1 . From this we readily deduce that Υ = Θ ◦ Θ ◦ Υ and
0 0
Υ = Θ ◦ Θ ◦ Υ. Hence the map Θ has the inverse Θ and is therefore an
isomorphism. 
Up to isomorphism, tensor multiplication of vector spaces may be per-
formed iteratively.
Theorem 58 Let V1 , . . . , Vn be vector spaces over the same field. Then for
any integer k such that 1 6 k < n there is an isomorphism
N N N N N N N
Θ : (V1 · · · Vk ) (Vk+1 · · · Vn ) → V1 · · · Vn
such that
Θ ((v1 ⊗ · · · ⊗ vk ) ⊗ (vk+1 ⊗ · · · ⊗ vn )) = v1 ⊗ · · · ⊗ vn .
N N
Proof: Set V× = V1 × · · · × Vn , V⊗ = V1 · · · Vn , V × = V1 × · · · × Vk ,
N N N N
V ⊗ = V1 · · · Vk , V × = Vk+1 × · · · × Vn , and V ⊗ = Vk+1 · · · Vn .
For fixed vk+1 ⊗ · · · ⊗ vn ∈ V ⊗ we define the k-linear function f vk+1 ⊗···⊗vn :
V × → V⊗ by f vk+1 ⊗···⊗vn (v1 , . . . , vk ) = v1 ⊗ · · · ⊗ vn . Corresponding to
f vk+1 ⊗···⊗vn is the (unique) map Θvk+1 ⊗···⊗vn : V ⊗ → V⊗ such that

Θvk+1 ⊗···⊗vn (v1 ⊗ · · · ⊗ vk ) = v1 ⊗ · · · ⊗ vn .

Similarly, for fixed v1 ⊗ · · · ⊗ vk ∈ V ⊗ we define the (n − k)-linear function

f v1 ⊗···⊗vk : V × → V⊗ by f v1 ⊗···⊗vk (vk+1 , . . . , vn ) = v1 ⊗· · ·⊗vn . Corresponding
to f v1 ⊗···⊗vk is the (unique) map Θv1 ⊗···⊗vk : V ⊗ → V⊗ such that

Θv1 ⊗···⊗vk (vk+1 ⊗ · · · ⊗ vn ) = v1 ⊗ · · · ⊗ vn .

We then define the function f : V ⊗ × V ⊗ → V⊗ by the formula

X
f (x, y) = av1 ⊗···⊗vk Θv1 ⊗···⊗vk (y)
P
when x = av1 ⊗···⊗vk · v1 ⊗ · · · ⊗ vk . We claim this formula gives the same
result no matter how x is expressed as a linear combination of elements
P form v1 ⊗ · · · ⊗ vk . To aid in verifying this, let y be expressed as
of the
y = bvk+1 ⊗···⊗vn · vk+1 ⊗ · · · ⊗ vn . The formula may then be written
X X
f (x, y) = av1 ⊗···⊗vk · bvk+1 ⊗···⊗vn · Θv1 ⊗···⊗vk (vk+1 ⊗ · · · ⊗ vn ) .
54 CHAPTER 4. MULTILINEARITY AND TENSORING

But Θv1 ⊗···⊗vk (vk+1 ⊗ · · · ⊗ vn ) = Θvk+1 ⊗···⊗vn (v1 ⊗ · · · ⊗ vk ) so that the for-
mula is equivalent to
X X
f (x, y) = av1 ⊗···⊗vk · bvk+1 ⊗···⊗vn · Θvk+1 ⊗···⊗vn (v1 ⊗ · · · ⊗ vk ) .

But this is the same as

X
f (x, y) = bvk+1 ⊗···⊗vn · Θvk+1 ⊗···⊗vn (x)

which of course has no dependence on how x is expressed as a linear combi-

nation of elements of the form v1 ⊗ · · · ⊗ vk .
f is clearly linear in either argument when the other argument is held
N
fixed. Corresponding to f is the (unique) map Θ : V ⊗ V ⊗ → V⊗ such that
Θ (x ⊗ y) = f (x, y) so that, in particular,

Θ ((v1 ⊗ · · · ⊗ vk ) ⊗ (vk+1 ⊗ · · · ⊗ vn )) = v1 ⊗ · · · ⊗ vn .
0 N
We define an n-linear function f : V× → V ⊗ V ⊗ by setting
0
f (v1 , . . . , vn ) = (v1 ⊗ · · · ⊗ vk ) ⊗ (vk+1 ⊗ · · · ⊗ vn ) .
0 0 N
Corresponding to f is the (unique) map Θ : V⊗ → V ⊗ V ⊗ such that
0
Θ (v1 ⊗ · · · ⊗ vn ) = (v1 ⊗ · · · ⊗ vk ) ⊗ (vk+1 ⊗ · · · ⊗ vn ) .
0 0
Thus each of Θ ◦ Θ and Θ ◦ Θ coincides with the identity on a spanning
set for its domain, so each is in fact the identity map. Hence the map Θ has
0
the inverse Θ and therefore is an isomorphism. 

Vn ∼
N N N N N N N
Exercise 4.4.3 (· · · ((V1 V2 ) V3 ) ··· Vn−1 ) = V1 · · · Vn .

V3 ∼
= V1 (V2 V3 ) ∼
N N N N N N
Corollary 59 (V1 V2 ) = V1 V2 V3 . 

The two preceding theorems form the foundation for the following general
associativity result.

Theorem 60 Tensor products involving the same vector spaces are isomor-
phic no matter how the factors are grouped.
4.4. TENSOR PRODUCTS IN GENERAL 55

Proof: Complete induction will be employed. The result is trivially true

for one or two spaces involved. Suppose the result is true for r spaces for
every r < n. Consider a grouped (perhaps nested) tensor product Π of the
n > 2 vector spaces V1 , ..., Vn in that order. At the outermost
N N essential level
of grouping, we have an expression of the form Π = W1 · · · Wk , where
1 < k 6 n and the n factor spaces V1 , ..., Vn are distributed
N N in (perhaps
∼
nested) groupings among the Wi . Then Π = (W1 · · · Wk−1 ) Wk
N
andN by the
N induction hypothesis N we may assume that,
N forN some i < n,
W1 · · N · WN k−1
∼
= V
N
· · · V and W ∼
= V · · · Vn . Hence
∼ N1 N Ni k i+1
∼ N N
Π = (V1 · · · Vi ) (Vi+1 · · · Vn ) and therefore Π = V1 · · · Vn
no matter how the n factors V1 , ..., Vn are grouped. The theorem thus holds
by induction. 
Not surprisingly, when we use the field F (considered as a vector space
over itself) as a tensor product factor, it does not really do anything.

Theorem 61 Let V be a vector space over the field F, and let F be consid-
ered
Nalso to be a vector spaceN
over itself. Then there is an isomorphism from
F V (and another from V F) to V sending a ⊗ v (and the other sending
v ⊗ a) to a · v.

Proof: The function from F × V to V that sends N (a, v) to a · v is clearly

N a (unique) map Θ : F
bilinear. Hence there exists
0
V → V that sends a ⊗ v
0
to a · v. Let Θ : V → F V be the map that sends v to 1 ⊗ v. Then Θ ◦ Θ
0

N the identity on V. On the other hand, Θ ◦ Θ sends each

is clearly N element
in F V of the form a ⊗ v to itself, and so is the identity on F V.
The other part is proved similarly. 
56 CHAPTER 4. MULTILINEARITY AND TENSORING

4.5 Problems
N
1. Suppose that in V1 V2 we have u ⊗ v = u ⊗ w for some nonzero
u ∈ V1 . Is it then possible for the vectors v and w in V2 to be different?
>
N N >∼ N N >
2. When all Vi are finite-dimensional, V1 · · · Vn = (V 1 · · · Vn )
via the unique map Φ : V1> · · · Vn> → (V1 · · · Vn )> for which
N N N N

Φ (ϕ1 ⊗ · · · ⊗ ϕn ) (v1 ⊗ · · · ⊗ vn ) = ϕ1 (v1 ) · · · ϕn (vn ) ,

or, in other words,

Φ (ϕ1 ⊗ · · · ⊗ ϕn ) = ϕ1 · · · ϕn .

Thus, for example, if the
Vi are all equal to V, and V has the basis
>
B, then Φ x> > > >
1 ⊗ · · · ⊗ xn = x1 · · · xn = (x1 ⊗ · · · ⊗ xn ) , where each
xi ∈ B.
Chapter 5

Vector Algebras

5.1 The New Element: Vector Multiplication

A vector space per se has but one type of multiplication, namely multipli-
cation (scaling) of vectors by scalars. However, additional structure may be
added to a vector space by defining other multiplications. Defining a vector
multiplication such that any two vectors may be multiplied to yield another
vector, and such that this multiplication acts in a harmonious fashion with
respect to vector addition and scaling, gives a new kind of structure when
this vector multiplication is included as a structural element. The maps of
such an enhanced structure, that is, the functions that preserve the entire
structure, including the vector multiplication, are still vector space maps, but
some vector space maps may no longer qualify due to the requirement that
the vector multiplication must also be preserved. Adding a new structural
element, giving whatever benefits it may, also then may give new burdens in
the logical development of the theory involved if we are to exploit structural
aspects through use of function images. We will see, though, that at the cost
of a little added complication, some excellent constructs will be obtained as
a result of including various vector multiplications.
Thus we will say that the vector space V over the field F becomes a vector
algebra (over F) when there is defined on it a vector multiplication
µ : V × V → V, written µ (u, v) = u ∗ v, which is required to be bilinear so
that it satisfies both the distributive laws

(t + u) ∗ v = t ∗ v + u ∗ v, and u ∗ (v + w) = u ∗ v + u ∗ w

57
58 CHAPTER 5. VECTOR ALGEBRAS

and also the law

a · (u ∗ v) = (a · u) ∗ v = u ∗ (a · v)

for all a ∈ F and all t, u, v, w ∈ V. A vector algebra is associative or

commutative according as its vector multiplication is associative or com-
mutative. If there is a multiplicative neutral element (a unit element) the
vector algebra is called unital. Since vector algebras are the only algebras
treated here, we will often omit the “vector” qualifier and simply refer to a
vector algebra as an algebra.

Exercise 5.1.1 v ∗ 0 = 0 ∗ v = 0.

Exercise 5.1.2 In a unital algebra with unit element 1, (−1) ∗ v = −v.

A (vector) subalgebra of an algebra is a vector space subspace that is

closed under the vector multiplication of the algebra, so that it is an algebra
in its own right with a vector multiplication that has the same effect as the
vector multiplication of the original algebra, but is actually its restriction to
the subspace.
An algebra map (or sometimes simply map when the kind of map is
clear) is a function between algebras over the same field, which function is a
vector space map that also preserves vector multiplication. The image of a
map is a subalgebra of the codomain and is associative (resp. commutative)
when the domain algebra is. Under a map, the image of a unit element acts
as a unit element in the image algebra.

Exercise 5.1.3 The inverse of a bijective algebra map is an algebra map.

(The isomorphisms of algebras are thus the bijective algebra maps.)

5.2 Quotient Algebras

The kernel of an algebra map is defined to be the kernel of the same function
viewed as a vector space map, namely those vectors of the domain that map
to the zero vector of the codomain. The kernel of an algebra map is a
subalgebra and more. It is an ideal in the algebra, i. e., a subalgebra closed
under multiplication on the left, and on the right, by all the elements of the
whole original algebra. Thus for an ideal I in the algebra V we have x∗v ∈ I
5.2. QUOTIENT ALGEBRAS 59

and v ∗ x ∈ I for all x ∈ I and for all v ∈ V. Every ideal is a subalgebra,

but not every subalgebra is an ideal. Hence there can be subalgebras that
are not kernels of any algebra map. However every ideal does turn out to be
the kernel of some algebra map, as we shall soon see.
The level sets of an algebra map are the level sets of a vector space
map, and they may be made into a vector space, and indeed into a vector
algebra, in the same manner as was previously used in Section 2.2. This
vector algebra made by mirroring the image of an algebra map in the level
sets is the quotient algebra V/K of the vector algebra V by the algebra
map’s kernel K. The theme remains the same: a quotient “is” the image of a
map, up to isomorphism, and all maps with the same domain and the same
kernel have isomorphic images. In a quotient algebra, the product of cosets
is modular, as the following exercise describes.

Exercise 5.2.1 Let V be a vector algebra and let K be the kernel of an algebra
map from V. Then in V/K, (u + K) ∗ (v + K) = u ∗ v + K for all u, v ∈ V.

The cosets of any ideal make up a vector space quotient. It is always

possible to impose the modular product on the cosets and thereby make the
vector space quotient into an algebra.

Exercise 5.2.2 Let V be a vector algebra over F and let I be an ideal in V.

On the vector space V/I , prescribing the modular product

(u + I) ∗ (v + I) = (u ∗ v) + I

for all u, v ∈ V gives a well-defined multiplication that makes V/I into a

vector algebra over F.

For any vector algebra V and ideal I C V, the natural projection function
p : V → V/I that sends v to the coset v +I is a vector space map with kernel
I, according to Proposition 33. p is also an algebra map if we employ the
modular product, for then p (u)∗p (v) = (u + I)∗(v + I) = u∗v+I = p(u∗v).
Thus we have the following result.

Theorem 62 Let V be a vector algebra and let I be an ideal in V. Then the

natural map p : V → V/I that sends v to the coset v + I is an algebra map
with kernel I, if the modular product is used on V/I. Hence each ideal I in
V is the kernel of some algebra map, and thus a quotient algebra V/I exists
for every ideal I in V. 
60 CHAPTER 5. VECTOR ALGEBRAS

5.3 Algebra Tensor Product

Nvector algebras V and W over the field F, the vector space tensor product
For
V W may be made into an algebra by requiring

(v ⊗ w) ∗ (v 0 ⊗ w 0 ) = (v ∗ v 0 ) ⊗ (w ∗ w 0 )
N N
and extending as a bilinear function on (V W) × (V W) so that
! !
X X X
vi ⊗ wi ∗ vj 0 ⊗ wj 0 = (vi ∗ vj 0 ) ⊗ (wi ∗ wj 0 ) .
i j i,j

N
This does give a well-defined vector multiplication for V W.

Proposition 63 Given vectorsN algebras V N and W overNthe field F, there

exits a bilinear function µ : (V W) × (V W) → V W such that for
all v, v 0 ∈ V and all w, w 0 ∈ W, µ(v ⊗ w, v 0 ⊗ w 0 ) = (v ∗ v 0 ) ⊗ (w ∗ w 0 ).

N
Proof: Consider the function from V × W × V × W to V W that sends
0 0 0 0
(v, w, v , w ) to (v ∗ v ) ⊗ (w ∗ w ). This function is linear N in each
N variable
N
separately and therefore there is a vector space map from V W V W
0 0 0 0
N
to V W that sends v ⊗ Nw ⊗ vN⊗ wNto (v ∗ v )N⊗ (wN ∗wN ). Since there is
an isomorphism from (V W) (V W) to V W V W that sends
0
(v ⊗ w) ⊗ (v N ⊗ w 0 )Nto vN⊗ w ⊗ v 0 ⊗Nw 0 , there is therefore a vector space
map from (V W) (V W) to V W that sends (v ⊗ w) ⊗ (v 0 ⊗ w 0 ) to
0 0
(v ∗ v ) ⊗ (w ∗ wN), and corresponding to
Nthis vector space map is 0a bilinear
W that sends (v ⊗ w, v ⊗ w 0 ) to
N
function µ : (V W) × (V W) → V
0 0
(v ∗ v ) ⊗ (w ∗ w ). 

N
V W with this vector multiplication function is the algebra tensor
product of the vector algebras V and W over the field F.

Exercise 5.3.1 If the vector algebras V and W over the field F are both
commutative, or both associative, or both unital, then so is their algebra
tensor product.
5.4. THE TENSOR ALGEBRAS OF A VECTOR SPACE 61

5.4 The Tensor Algebras of a Vector Space

N
The contravariant tensor algebra NV of a vector space N V over Nthe N field
F is formed from the tensor powers 0 V = F, and k V = V U · · ·N V
(with k factors V, k = 1, 2, . . .), by taking the weak product k>0 k V
and defining on it a vector multiplication. To simplify N notation, we will not
Nk N k
distinguish V from its alias ξk ( V) = ηk V (using the functions
N
introduced in section 3.7),
L so N that, for example,N V may also be written as
N k Ni
the direct sum V = k>0 V. For u, v ∈ V such that u ∈ V and
Nj Ni+j
v∈ V, u ∗ v will be defined as the element of V that corresponds
to u ⊗N v underL the isomorphism of Theorem 58 or Theorem 61. In general,
V = k>0 k V can be written uniquely as finite sums u =
N P
u, v ∈ ui
P Ni Nj
and v = vj where uP i ∈ V and vj ∈ P V and their product will then
be defined as u ∗ v = N wk where wk = i+j=k ui ∗ vj . With this bilinear
vector multiplication, V is an associative, unital algebra. For most V, this
algebra is not commutative, however. As is customary for this algebra, we
will henceforth use ⊗ instead of ∗ as the vector multiplication symbol.
N >
V is the covariant tensor algebra of a vector space V. Taking
the algebra tensor product N Nof N the contravariant algebra with the covariant
algebra gives TV = ( V) ( V > ), the (full) tensor algebra of V. The
elementsNof TV are called Ntensor combinations. In TV, the algebra product Np
of r ∈ V andNs ∈ V > is denoted r ⊗ s and when the r is in V
q >
and the s is in V , the product r ⊗ s is said to be a homogeneous
tensor combination, or simply a tensor, of contravariant degree p, of
covariantL degree q, and of total Npdegree p + q. It is not hard to see that
p p
V) ( q V > ).
NN
TV = p,q>0 Tq V where Tq V = (
Np
The elements of the form v1 ⊗· · ·⊗vp in V, p > 0, where each factor
N is
in 1 V, are the e-products
N
N0 (elementary products) of degree p
N in V.
The nonzero elements of V are the e-products of degree 0. V thus
is the vector space of linear combinations of e-products of various degrees,
and TV is the vector space of linear combinations of mixed e-products N
rN ⊗ s where the various r and s are e-products of various degrees in V and
V > , respectively.
N1 This way of defining the tensor algebra of V always
N1 keeps
the elements of V all together on the left, and the elements of V > all
together on the right, in each mixed e-product.
N
V is a basic construct from which two important algebras may be
derived as quotients, as we do in the next two sections.
62 CHAPTER 5. VECTOR ALGEBRAS

5.5 The Exterior Algebra of a Vector Space

Given an ideal I in an algebra W, the quotient algebra W/I is a kind of
scaled-down image of W that suppresses I and glues together into one image
element all the elements of each coset v + I in W, effectively treating the
elements of W modulo I. Elements of a particular type in W that can
be expressed as the members of an ideal I in W can be “factored out” by
N to the quotient W/I. We now proceed along these lines by defining
passing
in V an ideal that when factored out will give us a very useful new algebra.
We say that the e-product v1 ⊗ · · · ⊗ vp is dependent if the sequence
v1 , . . . , vp is dependent, that is if the set {v1 , . . . , vp } is dependent or if
there are any equal vectors among v1 , . . . , vp . We similarly call the tuple
(v1 , . . . , vp ), dependent under the same circumstances. Such that are not
dependent are independent. The set Nof values of all linear combinations
of the set of dependentNe-products in V will be denoted
V ND. It is clear
by
that D is an ideal in V. The quotient algebra V = ( V) /D is the
exterior algebra of V. The exterior e-product v1 ∧ · · · ∧ vpN is the image
V
of
Vp the e-product v 1 ⊗ · · · ⊗ vp under the natural projection π ∧ : V → V.
V = π∧ ( p V), the pth exterior power of V, is the subspace of ele-
N
ments of degree V0 p, and is spanned by all V1the exterior e-products of degree p.
Noting
V0 that V is
V1an alias of F, V and V is an alias of V, we will identify
V with F and V with V. V is an associative algebra because it is
the algebra map image of an associative V algebra. It is customary to use ∧ as
the multiplication symbol in V.
Suppose that the set P {v1 , . . . , vp } of nonzero vectors in V is dependent.
Then for some i, vi = j6=i aj · vj and then

X
v1 ⊗ · · · ⊗ vp = aj · v1 ⊗ · · · ⊗ vi−1 ⊗ vj ⊗ vi+1 ⊗ · · · ⊗ vp
j6=i

so that such a dependent e-product is always a linear combination of e-

products
N each of which has equal vectors among its factors. Hence the ideal D
in V may be more primitively described as the set of all linear combinations
of e-products each of which has at least one pair of equal vectors among its
factors.
Let u, v ∈ 1 V. Then 0 = (u + v)∧(u + v) = u∧u+u∧v+v∧u+v∧v =
N
u ∧ v + v ∧ u and therefore u ∧ v = − (v ∧ u).
5.5. THE EXTERIOR ALGEBRA OF A VECTOR SPACE 63

Exercise 5.5.1 Let v1 ∧ · · · ∧ vp be an exterior e-product. Then

vσ(1) ∧ · · · ∧ vσ(p) = (−1)σ · (v1 ∧ · · · ∧ vp )

where (−1)σ = +1 or −1 according as the permutation σ of {1, . . . , p} is

even or odd.

Exercise 5.5.2 Let r ∈ p V and let s ∈ q V. Then s ∧ r = (−1)pq · r ∧ s.

V V
(The sign changes only when both p and q are odd.)
Vp
Exercise 5.5.3 Let r ∈ V. Then it follows at once from the above
exercise that r ∧ r = 0 if p is odd and if 1 + 1 6= 0 in F. What about when p
is odd and 1 + 1 = 0 in F?

Let B be a basis for V. Based on B, the set of basis monomials of

degree p (allNthe e-products of degree p with factors chosen only from B), is
a basis for p V. Let r denote the e-product v1 ⊗ · · · ⊗ vp . Then r has an
expansion of the form
X X
r= ··· a1,i1 · · · ap,ip · xi1 ⊗ · · · ⊗ xip
i1 ip

in terms of some x1 , . . . , xN ∈ B. Let us write r = r= + r6= , where r= is the

sum of those terms that have some equal subscripts and r6= is the sum of the
remaining terms that have no equal subscripts. We have
X X
r6= = a1,iσ(1) · · · ap,iσ(p) · xiσ(1) ⊗ · · · ⊗ xiσ(p)
i1 <···<ip σ∈Sp

where Sp denotes the set of all permutations of {1, . . . , p}. In the event that
two of the factors, say vl and vm , l < m, are equal in r, the terms in the
summation above occur in pairs with the same coefficient since

a1,iσ(1) · · · ap,iσ(p) = a1,iσ(1) · · · al,iσ(m) · · · am,iσ(l) · · · ap,iσ(p) .

When vl = vm , l < m, we may thus write

X X
r6= = a1,iσ(1) · · · ap,iσ(p) ·
i1 <···<ip σ∈Ap

·(xiσ(1) ⊗ · · · ⊗ xiσ(p) + xiσ(1) ⊗ · · · ⊗ xiσ(m) ⊗ · · · ⊗ xiσ(l) ⊗ · · · ⊗ xiσ(p) )

64 CHAPTER 5. VECTOR ALGEBRAS

where Ap is the set of all even permutations of {1, . . . , p}. We say that the
two e-products v1 ⊗ · · · ⊗ vp and vσ(1) ⊗ · · · ⊗ vσ(p) are of the same parity,
or of opposite parity, according as σ is an even, or an odd, permutation of
{1, . . . , p}. We see therefore that the dependent basis monomials, along with
the sums of pairs of independent basis monomials of opposite parity, span D.
Let X = {x1 , . . . , xp } be a subset of the basis B for V. From the ele-
ments of X , K = p! independent basis monomials may be formed by mul-
tiplying together the elements of X in the various possible orders. Let
T1 = {t1 , t3, . . . , tK−1 } be the independent basis monomials of degree p,
with factors from X , and of the same parity as t1 = x1 ⊗ · · · ⊗ xp , and
let T2 = {t2 , t4 , . . . , tK } be those of the opposite parity. Then
T = {t1 , t1 + t2 , t2 + t3 , . . . , tK−1 + tK }
is a set of K independent elements with the same span as T1 ∪ T2 . Moreover,
we claim that for any s ∈ T1 and any t ∈ T2 , s + t is in the span of T r {t1 }.
It suffices to show this for s+t of the form ti +tk where i < k and exactly one
of i, k is odd, or, what is the same, for s + t = ti + ti+2j+1 . But ti + ti+2j+1 =
(ti + ti+1 ) − (ti+1 + ti+2 ) + (ti+2 + ti+3 ) − · · · + (ti+2j + ti+2j+1 ), verifying the
claim. The following result is now clear.
Proposition 64 Let B be a basis for V. From each nonempty subset X =
{x1 , . . . , xp } of p elements of B, let the sets T1 = {t1 , t3, . . . , tK−1 } and T2 =
{t2 , t4 , . . . , tK } comprising in total the K = p! degree p independent basis
monomials be formed, T1 being those that are of the same parity as t1 =
x1 ⊗ · · · ⊗ xp , and T2 being those of parity opposite to t1 , and let
T = {t1 , t1 + t2 , t2 + t3 , . . . , tK−1 + tK } .
Then hT i = hT1 ∪ T2 i and if s ∈ T1 and t ∈ T2 , then s + t ∈ hT r {t1 }i.
Let A0 denote the set of all dependent basis monomials based on B, let A1
denote the union of all the sets T r {t1 } for all p, and let E denote the union
Nthe singleton sets {t1 } for all p, and {1}. Then A0 ∪ A1 ∪ E is a basis
of all
for V, A0 ∪ A1 is a N basis for the ideal D of all linear combinations of
dependent e-products in V, and E is a basis for a complementary subspace
of D. 
Inasmuch as any independent set of vectors is part of some basis, in the
light of Proposition 31 we then immediately infer the following useful result
which also assures us that in suppressing the dependent e-products we have
not suppressed any independent ones.
5.5. THE EXTERIOR ALGEBRA OF A VECTOR SPACE 65

Corollary 65 (Criterion for Independence) The sequence v1 , . . . , vp of

vectors of V is independent if and only if v1 ∧ · · · ∧ vp 6= 0. 

Exercise 5.5.4 Let V be an n-dimensional vector space with basis B =

{x1 , .
. . , xn }. Then for 1 6 Vp 6 n, the set xi1 ∧V· · · ∧ xip | i1 < · · · < ip
of np elements is a basis for p V, while for p > n, p V = 0. In particular,
then, the singleton set {x1 ∧ · · · ∧ xn } is a basis for n V.
V

N V
We may restrict the domain of the natural projection π∧ : V → V
to pN V and theVcodomain to p V and thereby obtain
N V
Np the vector space map
π∧p : V p V → p
V which has kernel D p = D ∩ V. It is easy
Np to see
p Np
that
Np V = V/Dp soNthat π∧p is the natural projection of V onto
V/Dp . Let Υp : V p → p V be a tensor product function, and let f be a
p-linear function from V p to a vector space W over the same field as V. Then
by the universal
Np property of the tensor product, there is a unique vector space
map f⊗ : V → W such that f = f⊗ ◦ Υp . Assume now that f is zero
on dependent p-tuples, so that the corresponding f⊗ is zero on dependent
e-products of degree Vp p. Applying Theorem 34, we infer the existence of a
unique map f∧ : V → W such that f⊗ = f∧ ◦ π∧p . Thus there is a
universal property for the pth exterior power, which we now officially record
as a theorem. (An alternating p-linear function is one that vanishes on
dependent p-tuples.)

Theorem 66 (Universal Property of Exterior Powers) Let V and W

be vector spaces over the same field, and let f : V p → W be an V alternating
p-linear function. Then there is a unique vector space map f∧ : p V → W
such that
f∧ (v1 ∧ · · · ∧ vp ) = f (v1 , . . . , vp ).


Exercise 5.5.5Vp LetVpf : V V → W be a vector space map. Then there is a

unique map f: V → p W that satisfies
Vp
f (v1 ∧ · · · ∧ vp ) = f (v1 ) ∧ · · · ∧ f (vp ) .

The map p f of the above exercise is commonly called the pth exterior
V
power of f .
Let f : V → V be a vector space map from the n-dimensional Vnvector
Vnspace
V to itself. Then by the above exercise there is a unique map f: V→
66 CHAPTER 5. VECTOR ALGEBRAS
Vn
V such thatV n f (v1 ∧ · · · ∧ vn ) = f (v1 ) ∧ · · · ∧ f (vn ). Since n V is
V V
1-dimensional, n f (t) = a · t for some scalar a = det f , the determinant
of f . Note that the determinant of f is independent of any basis choice.
However, the determinant is only defined for self-maps on finite-dimensional
spaces.

Theorem 67 (Product Theorem) Let f, g be vector space maps on an

n-dimensional vector space. Then det g ◦ f = (det g) (det f ).

Proof: n (g ◦ f ) (v1 ∧ · · · ∧ vn ) = g (f (v1 )) ∧ · · · ∧ g (f (vn )) = (det g) ·

V
f (v1 ) ∧ · · · ∧ f (vn ) = (det g) (det f ) · v1 ∧ · · · ∧ vn . 

Proposition 68 Let f : V → W be a vector space map from the n-dimensional

vector space V to its alias W. Let {x1 , . . . , xn } be a basis for V. Then f is
invertible if and only if f (x1 ) ∧ · · · ∧ f (xn ) 6= 0.

Proof: Suppose that f (x1 ) ∧ · · · ∧ f (xn ) = 0. Then, by Corollary 65, the

sequence f (x1 ) , . . . , f (xn ) is dependent, so there exist scalars a1 , . . . , an , not
all zero, such that

0 = a1 · f (x1 ) + . . . + an · f (xn ) = f (a1 · x1 + . . . + an · xn ) .

Hence f sends the nonzero vector a1 · x1 + . . . + an · xn to 0, and thus the

kernel of f is not {0}. Therefore f fails to be one-to-one and hence is not
invertible.
On the other hand, if f (x1 )∧· · ·∧f (xn ) 6= 0, the sequence f (x1 ) , . . . , f (xn )
is independent and makes up a basis of W. Thus f sends each nonzero vector
in V to a nonzero vector in W, and thus the kernel of f is {0}. Therefore f
is one-to-one and hence invertible. 

Corollary 69 Let f : V → V be a vector space map from the n-dimensional

vector space V to itself. Then f is invertible if and only if det f 6= 0. 

Suppose the n-dimensional vector space aliases V and W have the respec-
P {x1 , . . . , xn }, and {y1 , . . . , yn }. The map f : V → W then sends
tive bases
xj to i ai,j · yi for some scalars ai,j . We have
! !
X X
f (x1 ) ∧ · · · ∧ f (xn ) = ai,1 · yi ∧ · · · ∧ ai,n · yi =
i i
5.5. THE EXTERIOR ALGEBRA OF A VECTOR SPACE 67
X
= aσ(1),1 · · · aσ(n),n · yσ(1) ∧ · · · ∧ yσ(n) =
σ∈Sn
!
X σ
= (−1) aσ(1),1 · · · aσ(n),n · y1 ∧ · · · ∧ yn
σ∈Sn

where Sn is the set of all permutations of {1, . . . , n} and (−1)σ = +1 or −1

according as the permutation σ is even or odd. Therefore f is invertible if
and only if X
0 6= (−1)σ aσ(1),1 · · · aσ(n),n = det [ai,j ] ,
σ∈Sn

i. e., if and only if the familiar determinant of the n × n matrix [ai,j ] is

nonzero. We observe that when V = W and xi = yi for all i, det [ai,j ] = det f .
We record this result as follows.

Proposition 70 Let the n-dimensional vector space aliases V and W have

P {x1 , . . . , xn }, and {y1 , . . . , yn }. Let the map f : V → W
the respective bases
send each xj to i ai,j ·yi . Then f is invertible if and only if the determinant
of the matrix [ai,j ], denoted det [ai,j ], is nonzero. When V = W and xi = yi
for all i, det [ai,j ] = det f .

Exercise 5.5.6 The determinant of the n × n matrix [ai,j ] is an alternating

n-linear functional of its columns (and also of its rows, due to the familiar
result that a matrix and its transpose have the same determinant).

The universal property of exterior powers leads nat once to the conclusion o
(alt p−lin)
that if V is over the field F, the map space subspace V p → F of alter-
>
nating p-linear functionals is isomorphic to ( p V) . There is also an analog
V
of Theorem 54, Vp which we now start to develop by exploring the coordinate
functions on V V relative to a basis B for V. Let t = v1 ∧· · ·∧vp be V an exterior
e-product in p V. Let x1 ∧ · · · ∧ xp be a typical basis vector of p V, formed
from elements xi ∈ B, and let t be expanded in terms of such basis vectors.
Then the coordinate function (x1 ∧ · · · ∧ xp )> corresponding to x1 ∧ · · · ∧ xp
>
is the element of ( p V) that gives the coefficient of x 1 ∧ · · · ∧
V
 xp in this>ex-
pansion of t. This coefficient is readily seen to be det x> i (v j , where xi is
)
the coordinate function on V that corresponds to xi ∈ B. Thus we are led to
consider what are easily seen to be alternating p-linear functionals fφ1 ,...,φp of
the form fφ1 ,...,φp (v1 , . . . , vp ) = det [φi (vj )] where the φi are linear functionals
68 CHAPTER 5. VECTOR ALGEBRAS

on V. By the universal property, fφ1 ,...,φp (v1 , . . . , vp ) = fbφ1 ,...,φp (v1 ∧ · · · ∧ vp )

>
for a unique linear functional fbφ1 ,...,φp ∈ ( p V) . Hence there is a func-
V
p >
tion Φ : V > → ( p V) such that Φ (φ1 , .
. . , φp ) = fbφ1 ,...,φp . But Φ is

V
p
clearly an alternating p-linear function on V > and by the universal prop-
erty there is a unique vector space map Φ b : Vp V > → (Vp V)> such that
b (φ1 ∧ · · · ∧ φp ) = Φ (φ1 , . . . , φp ). Based on our study of coordinate func-
Φ
b > > >
tions above, we see that Φ(x 1 ∧ · · · ∧ xp ) = (x1 ∧ · · · ∧ xp ) . Therefore, for
finite-dimensional V, Φ b is an isomorphism because it sends a basis to a basis
which makes it onto, and it is therefore also one-to-one by Theorem 25. The
following statement summarizes our results.
Vp > ∼
Theorem 71
n (alt p−lin) o Let V be a vector space over the field F. Then ( V) =
V> ∼
p p >
Vp
V V
→ F . If V is finite-dimensional, we also have = ( V)
via the map Φ b : Vp V > → (Vp V)> for which Φb (φ1 ∧ · · · ∧ φp ) (v1 ∧ · · · ∧ vp ) =
det [φi (vj )]. 

Exercise 5.5.7 What is Φ

b in the case p = 1?

Exterior algebra is an important subject that will receive further attention

in later chapters. Now we consider another example of a vector N algebra that
is an algebra map image of the contravariant tensor algebra V.

5.6 The Symmetric Algebra of a Vector Space

N
Noncommutativity
N in V can be suppressed by passing to the quotient
SV = V/N where N is an ideal that expresses noncommutativity. The
result is a commutative algebra that is customarily called the symmetric
algebra of V. The noncommutativity ideal N may be taken to be the set of
all linear combinations of differences of pairs of e-products of the same degree
p, p > 2, which contain the same factors with the same multiplicities, but
in different orders. The effect of passing to the quotient then is to identify
all the e-products of the same degree which have the same factors with the
same multiplicities without regard to the order in which these factors are
being multiplied. The symmetric e-product v1 · · · vp is Nthe image ofp the
e-product
Np v1 ⊗ · · · ⊗ v p under the natural projection π S : V → SV. S V =
πS ( V), the pth symmetric power of V, is the subspace of elements of
5.6. THE SYMMETRIC ALGEBRA OF A VECTOR SPACE 69

degree p. Noting that S0 V is an alias of F, and S1 V is an alias of V, we

will identify S0 V with F and S1 V with V. We will write products in SV
with no sign to indicate the product operation, and we will indicate repeated
adjacent factors by the use of an exponent. Thus r3 r22 r1 , r2 r3 r1 r2 , and r1 r22 r3
are typical (and equal) product expressions in SV.
Following a plan similar to that used with the exterior algebra, we will
produce a basis for the ideal N and a basis for a complementary subspace of
N . Let B be a basis for V. Consider the difference of the pair of e-products
v1 ⊗ · · · ⊗ vp and vσ(1) ⊗ · · · ⊗ vσ(p) , where σ is some permutation of {1, . . . , p}.
In terms of some x1 , . . . , xN ∈ B we have for this difference
X X
··· (a1,i1 · · · ap,ip ·xi1 ⊗· · ·⊗xip −aσ(1),iσ(1) · · · aσ(p),iσ(p) ·xiσ(1) ⊗· · ·⊗xiσ(p) )
i1 ip
X X
= ··· a1,i1 · · · ap,ip · (xi1 ⊗ · · · ⊗ xip − xiσ(1) ⊗ · · · ⊗ xiσ(p) ),
i1 ip

where the multiplicative commutativity of the scalars has been exploited.

Thus N is spanned by the differences of pairs of basis monomials that contain
the same factors with the same multiplicities, but in different orders.
To each basis monomial t = x1 ⊗· · ·⊗xp , where the xi are (not necessarily
distinct) elements of B, there corresponds the multiset

M = {(ξ1 , ν1 ), . . . , (ξm , νm )}

where the ξj are the distinct elements of {x1 , . . . , xp } and νj is the multiplicity
with which ξj appears as a factor in t. Putting ν = (ν1 , . . . , νm ) and |ν| =
ν1 + · · · + νm , we then have |ν| = p. Given a particular multiset M, let
T0 = {t1 , . . . , t
K } be the set of all basis monomials to which
M corresponds.
Here K = |ν| ν
is the multinomial coefficient given by |ν|
ν
= |ν|!/ν! where
ν! = (ν1 !) · · · (νm !). From each such T0 we may form the related set T =
{t1 , t1 − t2 , . . . , tK−1 − tK } with the same span. Now ti − ti+j = (ti − ti+1 ) +
· · · + (ti+j−1 − ti+j ) so that the difference of any two elements of T0 is in
hT r {t1 }i. We therefore have the following result.

Proposition 72 Let B be a basis for V. From each nonempty multiset M =

{(ξ1 , ν1 ), . . . , (ξk , νm )} where the ξj are distinct elements of B, let the set T0 =
{t1 , . . . , tK } of basis monomials that correspond be formed, where K = |ν|


ν
.
Let
T = {t1 , t1 − t2 , t2 − t3 , . . . , tK−1 − tK } .
70 CHAPTER 5. VECTOR ALGEBRAS

Then hT i = hT0 i and if s, t ∈ T0 , then s − t ∈ hT r {t1 }i. Let A denote the

union of all the sets T r {t1 } for all ν, and let S denote the union
N of all the
singleton sets {t1 } for all ν, and {1}. Then A ∪ S is a basis for V, A is a
basis for the noncommutativity ideal N , and S is a basis for a complementary
subspace of N . 

Exercise 5.6.1 Let V be an n-dimensional vector space with basis B =

{x1 , . . . , xn }. Then for p > 0, the set xi1 · · · xip | i1 6 · · · 6 ip of n+p−1

p
elements is a basis for Sp V.

Exercise 5.6.2 (Universal Property of Symmetric Powers) Let V and

W be vector spaces over the same field, and let f : V p → W be a sym-
metric p-linear
function, i. e., a p-linear function such that f (v1 , . . . , vp ) =
f vσ(1) , . . . , vσ(p) for every permutation σ of {1, . . . , p}. Then there is a
unique vector space map fS : Sp V → W such that f (v1 , . . . , vp ) = fS (v1 · · · vp ).
n (sym p−lin)
o
Exercise 5.6.3 (Sp V)> is isomorphic to the vector space V p → F
of symmetric p-linear functionals.

5.7 Null Pairs and Cancellation

By a null pair in an algebra we mean a pair u, v of nonzero elements, not
necessarily distinct, such that u ∗ v = 0. (We avoid the rather misleading
term zero divisor, or the similar divisor of zero, usually used in describing
this notion.) It is the same for an algebra to have no null pair as it is for it to
support cancellation of nonzero factors, as the following proposition details.

Proposition 73 An algebra has no null pair if and only if it supports can-

cellation: the equations u ∗ x = u ∗ y and x ∗ v = y ∗ v, where u, v are any
nonzero elements, each imply x = y.

Proof: Suppose that the algebra has no null pair. The equation u∗x = u∗y
with u 6= 0 implies that u ∗ (x − y) = 0 which then implies that x − y = 0
since the algebra has no null pair. Similarly x ∗ v = y ∗ v with v 6= 0 implies
x − y = 0. Thus in each case, x = y as was to be shown.
On the other hand, assume that cancellation is supported. Suppose that
u ∗ v = 0 with u 6= 0. Then u ∗ v = u ∗ 0, and canceling u gives v = 0. Hence
5.8. PROBLEMS 71

u ∗ v = 0 with u 6= 0 implies that v = 0 and it is therefore not possible that

u ∗ v = 0 with both u and v nonzero. 
An element u is called left regular if there is no nonzero element v such
that u ∗ v = 0, and similarly v is called right regular if there is no nonzero
u such that u ∗ v = 0. An element is regular if it is both left regular and
right regular.

Exercise 5.7.1 u may be canceled from the left of u ∗ x = u ∗ y if and only

if it is left regular, and v may be canceled from the right of x ∗ v = y ∗ v if
and only if it is right regular.

5.8 Problems
1. Give an example of an algebra map f : A → B where A and B are both
unital with unit elements 1A and 1B respectively, but f (1A ) 6= 1B .

2. Use exterior algebra to derive the Laplace expansion of the determinant

of an n × n matrix.

3. Use exterior algebra to derive Cramer’s rule for solving a set of n

linear equations in n unknowns when the coefficient matrix has nonzero
determinant. Then derive the well-known formula for the inverse of the
coefficient matrix in terms of cofactors.

4. In terms of the original coordinates with respect to the ti , give a formula

for the coordinate function of the basis vector t1 relative to the new
basis that Proposition 64 provides for the tensor algebra of V.

5. If V is finite-dimensional over a field of characteristic 0, then Sp V > ∼ =

(Sp V)> via the vector space map Ψ b : Sp V > → (Sp V)> for which
b (ψ1 · · · ψp ) (r1 · · · rp ) = per [ψi (rj )] where per [ai,j ] = P
Ψ σ∈Sn aσ(1),1 · · · aσ(n),n
is the permanent of the matrix [ai,j ] .

6. If u ∗ v = 0 in an algebra, must then v ∗ u = 0 also?

7. An algebra supports left-cancellation if v ∗ x = v ∗ y implies x = y

whenever v 6= 0, and right-cancellation is defined similarly. An alge-
bra supports left-cancellation if and only if it supports right-cancellation.
72 CHAPTER 5. VECTOR ALGEBRAS
Chapter 6

Vector Affine Geometry

6.1 Basing a Geometry on a Vector Space

One may base a geometry on a vector space by defining the fundamental
geometric objects to be the members of some specified family of subsets of
vectors, and by declaring that two such objects are incident if one is included
in the other by set inclusion. It is natural to take as fundamental geometric
objects a collection of “flat” sets, such as subspaces or cosets of subspaces,
and this is what we shall do here. By basing a geometry in the same way
on each vector space of a family of vector spaces over the same field, we will,
upon also specifying suitable structure-preserving maps, be able to view the
entire collection of fundamental geometric objects of all the vector spaces
of the family as a category of geometric spaces having the same kind of
structure.
In this chapter we derive from an underlying vector space over a general
field a type of geometry called affine, which embodies the more primitive
notions of what is no doubt the most familiar type of geometry, namely
Euclidean geometry. In affine geometry, just like in Euclidean geometry,
there is uniformity in the sense that the view from each point is the same.
The points all have equal standing, and any one of them may be considered
to be the origin. In addition to incidence, we also have parallelism as a
fundamental concept in affine geometry. Working, as we will, over a general,
possibly unordered, field, the availability of geometrically useful quantities
that can be expressed as field values will be quite limited, and even over the
real numbers, far fewer such quantities are available to us than in ordinary

73
74 CHAPTER 6. VECTOR AFFINE GEOMETRY

Euclidean geometry. Nevertheless, many important concepts and results are

embraced within affine geometry, and it provides a suitable starting point for
our exploration of vector geometry.

6.2 Affine Flats

Let V be a vector space V over the field F. We denote by V (V) the set of all
subspaces of V. We utilize V (V) to obtain an affine geometry by taking as
fundamental geometric objects the set V +V (V) of all affine flats of V, which
are just the cosets of the vector space subspaces of V. Thus we introduce
A (V) = V + V (V) as the affine structure on V. In A (V), the points
are the cosets of {0} (i. e., the singleton subsets of V, or essentially the
vectors of V), the lines are the cosets of the one-dimensional subspaces, the
planes are the cosets of the two-dimensional subspaces, and the n-planes
are the cosets of the n-dimensional subspaces. The hyperplanes are the
cosets of the subspaces of codimension 1. The dimension of an affine flat
is the dimension of the subspace of which it is a coset. If Φ ∈ A (V) is given
by Φ = v + W where W ∈ V (V), we call W the directional subspace, or
underlying subspace, of Φ. The notation Φv will be used to signify the
unique directional subspace of the affine flat Φ.

6.3 Parallelism in Affine Structures

We say that two affine flats, distinct or not, in the same affine structure are
parallel if their directional subspaces are incident (one is contained entirely
within the other). They are strictly parallel if their directional subspaces
are equal. They are nontrivially parallel if they are parallel and neither is
a point. They are properly parallel if they are parallel and unequal. Any
of these terms applies to a set of affine flats if it applies pairwise within the
set. Thus, for example, to say that S is a set of strictly parallel affine flats
means that for every Φ, Ψ ∈ S, Φ and Ψ are strictly parallel.
Exercise 6.3.1 Two strictly parallel affine flats that intersect must be equal.
Hence, two parallel affine flats that intersect must be incident.
Exercise 6.3.2 The relation of being strictly parallel is an equivalence rela-
tion. However, in general, the relation of being parallel is not transitive.
6.4. TRANSLATES AND OTHER EXPRESSIONS 75

6.4 Translates and Other Expressions

For vectors v, w, x and scalar a, a translate of v (in the direction of x − w)
is an expression of the form v + a · (x − w). This is a fundamental affine
concept. If x, w ∈ S, we call v + a · (x − w) an S-translate of v.

Proposition 74 The nonempty subset S of the vector space V is an affine

flat if and only if it contains the value of each S-translate of each of its
elements.

Proof: Let S ⊂ V be such that it contains each S-translate of each of its

elements and let v ∈ S. Then we wish to show that S = v + W where W C V.
This is the same as showing that 1) v + a · (w − v) ∈ S for all w ∈ S and
all scalars a, and that 2) v + ((w − v) + (x − v)) ∈ S for all w, x ∈ S. The
expression in 1) is clearly an S-translate of v and hence is in S as required.
The expression in 2) may be written as v + ((w + (x − v)) − v) and is seen
to be in S because z = w + (x − v) is an S-translate of w and is therefore in
S so that our expression v + (z − v) is also an S-translate of v and must be
in S.
On the other hand, suppose that S is an affine flat in V and let v ∈ S.
Let x = v + xv and w = v + wv where xv , wv ∈ S v . Then a · (x − w) =
a · (xv − wv ) ∈ S v and hence v + a · (x − w) ∈ S. 

Exercise 6.4.1 The intersection of affine flats is an affine flat or empty.

Iterating the process of forming translates gives the multi-translate

v + b1 · (x1 − w1 ) + · · · + bn · (xn − wn )

of the vector v in the directions of the xi − wi based on the vectors wi , xi and

the scalars bi .
An expression of the form a1 · v1 + · · · + an · vn is a linear expression
in the vectors v1 , . . . , vn , and a1 + · · · + an is its weight. Two important
cases are specially designated: A linear expression of weight 1 is an affine
expression, and a linear expression of weight 0 is a directional expression.
76 CHAPTER 6. VECTOR AFFINE GEOMETRY

If we expand out each of the terms bi · (xi − wi ) as bi · xi + (−bi ) · wi , the

multi-translate above becomes an affine expression in the wi , the xi , and v,
and it becomes v plus a directional expression in the wi and the xi . On the
other hand, since

a1 · v1 + · · · + an · vn = (a1 + · · · + an ) · v1 + a2 · (v2 − v1 ) + · · · + an · (vn − v1 ) ,

our typical affine expression may be written as the multi-translate

v1 + a2 · (v2 − v1 ) + · · · + an · (vn − v1 ) ,

and our typical directional expression may be written as the linear expression

a2 · (v2 − v1 ) + · · · + an · (vn − v1 )

in the differences (v2 − v1 ) , . . . , (vn − v1 ).

Exercise 6.4.2 A directional expression in the vectors v1 , . . . , vn is nontriv-

ially equal to 0 (some scalar coefficient not 0) if and only if some vj is equal
to an affine expression in the other vi .

6.5 Affine Span and Affine Sum

An affine combination of a finite nonempty set of vectors is any linear
combination of the set such that the coefficients sum to 1, and thus is an
affine expression equal to a multi-translate of any one of the vectors based
on all of them. If X is a nonempty set of vectors, its affine span hX iA is
the set of all affine combinations of all its finite nonempty subsets. We say
that X affine-spans hX iA , and is an affine spanning set for hX iA .

Exercise 6.5.1 Let X be a nonempty subset of the vector space V. Then

hX iA is the smallest affine flat containing X , i. e., it is the intersection of
all affine flats containing X .

Let Φ, Ψ ∈ A (V). Then Φ +A Ψ = hΦ ∪ ΨiA is their affine sum, and

similarly for more summands. With partial order ⊂, meet ∩, and join +A ,
A (V) ∪ {Ø} is a complete lattice which is upper semi-modular but in general
is neither modular nor distributive (see, e. g., Hall’s The Theory of Groups
for definitions and characterizations).
6.6. AFFINE INDEPENDENCE AND AFFINE FRAMES 77

Exercise 6.5.2 Let Φ, Ψ ∈ A (V). Suppose first that Φ ∩ Ψ 6= Ø. Then

(Φ +A Ψ)v = Φv + Ψv and (Φ ∩ Ψ)v = Φv ∩ Ψv , so that

dim (Φ +A Ψ) + dim (Φ ∩ Ψ) = dim Φ + dim Ψ.

On the other hand, supposing that Φ ∩ Ψ = Ø, we have

dim (Φ +A Ψ) = dim Φ + dim Ψ − dim (Φv ∩ Ψv ) + 1.

(Assigning Ø the standard dimension of −1, the formula above may be written

dim (Φ +A Ψ) + dim (Φ ∩ Ψ) = dim Φ + dim Ψ − dim (Φv ∩ Ψv ) .

Hence, no matter whether Φ and Ψ intersect or not, we have

dim (Φ +A Ψ) + dim (Φ ∩ Ψ) 6 dim Φ + dim Ψ

which is a characterizing inequality for upper semi-modularity.) Note also

that we have dim (Φ +A Ψ) − dim (Φv + Ψv ) = 0 or 1 according as Φ and Ψ
intersect or not.

6.6 Affine Independence and Affine Frames

Using an affine structure on a vector space allows any vector to serve as
origin. The idea of a dependency in a set of vectors needs to be extended
in this light. Here the empty set plays no significant rôle, and it is tacitly
assumed henceforth that only nonempty sets and nonempty subsets are being
addressed relative to these considerations. A directional combination of
a finite (nonempty) set of vectors is any linear combination of the set such
that the coefficients sum to 0. An affine dependency is said to exist in
a set of vectors if the zero vector is a nontrivial (scalar coefficients not all
zero) directional combination of one of its finite subsets (see also Exercise
6.4.2). A set in which an affine dependency exists is affine dependent and
otherwise is affine independent. A point is affine independent.

Exercise 6.6.1 If X is a nonempty set of vectors and w ∈ X , then X is

independent relative to w if {v − w | v ∈ X , v 6= w} is an independent
set of vectors (in the original sense). A nonempty set of vectors is affine
independent if and only if it is independent relative to an arbitrarily selected
one of its elements.
78 CHAPTER 6. VECTOR AFFINE GEOMETRY

Exercise 6.6.2 In an n-dimensional vector space, the number m of elements

in an affine independent subset must satisfy 1 6 m 6 n + 1 and for each such
value of m there exists an affine independent subset having that value of m
as its number of elements.

An affine frame for the affine flat Φ is an affine independent subset that
affine-spans the flat.

Exercise 6.6.3 An affine frame for the affine flat Φ is a subset of the form
{v} ∪ (v + B) where B is a basis for the directional subspace Φv .

Exercise 6.6.4 Let A be an affine frame for the affine flat Φ.PThen each
vector of Φ has a unique expression as an affine combination x∈X ax · x
where X is some finite subset of A and all of the scalars ax are nonzero.

Exercise 6.6.5 Let the affine flat Φ have the finite affine framePX . Then
each vector of Φ has a unique expression as an affine combination x∈X ax ·x.
The scalars ax are the barycentric coordinates of the vector relative to
the frame X .

6.7 Affine Maps

Let V and W be vector spaces over the same field, which we will always
understand to be the case whenever there are to be maps between their
affine flats. For Φ ∈ A (V) and Ψ ∈ A (W), we call the function α : Φ → Ψ
an affine map if it preserves translates:

α(v + a · (x − w)) = α(v) + a · (α(x) − α(w))

for all v, w, x ∈ Φ and all scalars a.

In a series of exercises we now point out a number of important results
concerning affine maps. We lead off with two basic results.

Exercise 6.7.1 The composite of affine maps is an affine map.

Exercise 6.7.2 A vector space map is an affine map.

6.7. AFFINE MAPS 79

If Φ 0 , Φ are elements of the same affine structure and Φ 0 ⊂ Φ, we say

that Φ 0 is a subflat of Φ. Thus two elements of the same affine structure are
incident whenever one is a subflat of the other. The image of an affine map
should be an affine flat, if we are even to begin to say that affine structure
is preserved. This is true, as the following exercise points out, and as a
consequence subflats are also sent to subflats with the result that incidence
is completely preserved.
Exercise 6.7.3 The image of an affine flat under an affine map is an affine
flat. Since the restriction of an affine map to a subflat is clearly an affine
map, an affine map therefore maps subflats to subflats and thereby preserves
incidence.
Affine expressions are preserved as well, and the preservation of affine
spans and affine sums then follows.
Exercise 6.7.4 An affine map preserves all affine expressions of vectors of
its domain. Hence under an affine map, the image of the affine span of a
subset of the domain is the affine span of its image, and affine maps preserve
affine sums.
A one-to-one function clearly preserves all intersections. However, in
general it can only be said that the intersection of subset images contains
the image of the intersection of the subsets. Thus it should not be too
suprising that affine maps do not necessarily even preserve the intersection
of flats as the following exercise demonstrates.
Exercise 6.7.5 As usual, let F be a field. Let α : F 2 → F 2 be the vector
space map that sends (a, b) to (0, a + b) . Let Φ = {(a, 0) | a ∈ F} and Ψ =
{(0, a) | a ∈ F}. Then α (Φ ∩ Ψ) 6= α (Φ) ∩ α (Ψ).
A basic result about vector space maps generalizes to affine maps.
Exercise 6.7.6 The inverse of a bijective affine map is an affine map.
The bijective affine maps are therefore isomorphisms, so that Φ ∈ A (V)
and Ψ ∈ A (W) are isomorphic (as affine flats) when there is a bijective
affine map from one to the other. Isomorphic objects, such as isomorphic
affine flats, are always aliases of each other, with context serving to clarify
the type of object.
Affine maps and vector space maps are closely related, as the following
theorem details.
80 CHAPTER 6. VECTOR AFFINE GEOMETRY

Theorem 75 Let α be an affine map from Φ ∈ A (V) to Ψ ∈ A (W)and fix

an element v ∈ Φ. Let f : Φv → Ψv be the function defined by

f (h) = α (v + h) − α (v)

for all h ∈ Φv . Then f is a vector space map that is independent of the choice
of v.

Proof: Let x, w ∈ Φ be arbitrary, so that h = x − w is an arbitrary

element of Φv . Then

f (a · h) = α (v + a · (x − w)) − α (v) = a · α (x) − a · α (w) =

= a · (α (v) + α (x) − α (w)) − a · α (v) =
= a · (α (v + h) − α (v)) = a · f (h) .

Similarly, we find that f (h + k) = f (h) + f (k) for arbitrary h, k ∈ Φv .

Hence f is a vector space map. Replacing h with x − w, we find that
f (h) = α (v + x − w) − α (v) = α (x) − α (w), showing that f (h) is actually
independent of v. 

The vector space map f that corresponds to the affine map α as in the
theorem above is the underlying vector space map of α, which we will
denote by αv . Thus for any affine map α we have

α (v) = αv (v − u) + α (u)

for all u, v in the domain of α.

Exercise 6.7.7 Let Φ ∈ A (V) and Ψ ∈ A (W) and let f : Φv → Ψv be a

given vector space map. Fix an element v ∈ Φ and let α : Φ → Ψ be the
function such that for each h ∈ Φv , α (v + h) = α (v) + f (h). Then α is an
affine map such that αv = f .

Other nice properties of affine maps are now quite apparent, as the fol-
lowing exercises detail.

Exercise 6.7.8 Under an affine map, the images of strictly parallel affine
flats are strictly parallel affine flats.
6.8. SOME AFFINE SELF-MAPS 81

Exercise 6.7.9 An affine map is completely determined by its values on

any affine spanning set, and if that affine spanning set is an affine frame,
the values may be arbitrarily assigned.
Exercise 6.7.10 Let α : Φ → Ψ be an affine map from Φ ∈ A (V) to Ψ ∈
A (W). Then there exists the affine map α] : V → Ψ that agrees with α on
Φ.

6.8 Some Affine Self-Maps

An affine self-map δt,u;b : Φ → Φ of the flat Φ that sends v ∈ Φ to
δt,u;b (v) = t + b · (v − u) ,
where t, u are fixed in Φ and b is a fixed scalar, is known as a dilation
(or dilatation, a term we will not use, but which means the same). Every
−1
proper dilation (b 6= 0) is invertible: δt,u;b = δu,t;b−1 . The following result
tells us that the proper dilations are the direction-preserving automorphisms
of affine flats.
Proposition 76 A function δ : Φ → Φ on the affine flat Φ is a dilation δt,u;b
if and only if δ (v) − δ (w) = b · (v − w) for all v, w ∈ Φ.
Proof: A dilation clearly satisfies the condition. On the other hand,
suppose that the condition holds. Fix u, w ∈ Φ and define t ∈ Φ by t =
δ (w) − b · (w − u). Then, employing the condition, we find that for any
v ∈ Φ, δ (v) = δ (w) + b · (v − w) = t + b · (w − u) + b · (v − w) = t + b · (v − u).


Letting h = t − u and
τh (v) = v + h = δt,u;1 (v)
for all v ∈ Φ gives the special dilation τh : Φ → Φ which we call a translation
by h. Every translation is invertible, τh−1 = τ−h , and τ0 is the identity
map. The translations on Φ correspond one-to-one with the vectors of the
directional subspace Φv of which Φ is a coset. We have τh ◦ τk = τk ◦ τh =
τk+h and thus under the operation of composition, the translations on Φ are
isomorphic to the additive group of Φv . Since every translation is a dilation
δt,u;b with b = 1, from the proposition above we infer the following result.
82 CHAPTER 6. VECTOR AFFINE GEOMETRY

Corollary 77 A function τ : Φ → Φ on the affine flat Φ is a translation if

and only if τ (v) − τ (w) = v − w for all v, w ∈ Φ. 
Exercise 6.8.1 The translations are the dilations that lack a unique fixed
point (or “center”).

6.9 Congruence Under the Affine Group

Viewing the vector space V as an affine flat in A (V) and using the composition
of maps as group operation, the bijective affine self-maps α : V → V form a
group GA (V) called the affine group of V. The elements of GA (V) may be
described as the invertible vector space self-maps on V composed with the
translations on V. Similarly, there is an affine group GA (Φ) for any affine flat
Φ ∈ A (V). However, GA (V) already covers the groups GA (Φ) in the sense
that any such GA (Φ) may be viewed as the restriction to Φ of the subgroup of
GA (V) that fixes Φ (those maps for which the image of Φ remains contained
in Φ).
A figure in V is just geometric language for a subset of V. Two figures
in V are called affine congruent if one is the image of the other under an
element of GA (V). Affine congruence is an equivalence relation.
Exercise 6.9.1 Affine frames in a vector space are affine congruent if and
only if they have the same cardinality.
Exercise 6.9.2 Affine flats in a vector space are affine congruent if and only
if their directional subspaces have bases of the same cardinality.

6.10 Problems

1. When F is the smallest field, namely the field {0, 1} of just two elements,
any nonempty subset S of a vector space V over F contains the translate
v + a · (x − v) for every v, x ∈ S and every a ∈ F. Not all such S are affine
flats, however.
On the other hand, if 1 + 1 6= 0 in F, a nonempty subset S of a vector
space over F is an affine flat if it contains v + a · (x − v) for every v, x ∈ S
and every a ∈ F.
What about when F is a field of 4 elements?
Chapter 7

Basic Affine Results and

Methods

7.1 Notations
Throughout this chapter, V is our usual vector space over the field F. We
generally omit any extra assumptions regarding V or F (such as dim V = 6 0)
that each result might need to make its hypotheses realizable. Without fur-
ther special mention, we will frequently use the convenient abuse of notation
P = {P } so that the point P ∈ A (V) is the singleton set that contains the
vector P ∈ V.

7.2 Two Axiomatic Propositions

The two propositions presented in this section are of an extremely basic
geometric character, and are assumed as axioms in most developments of
affine geometry from the synthetic viewpoint. While certainly not universal,
the first of these does apply to many ordinary kinds of geometries. It is
commonly, but somewhat loosely, expressed as “two points determine a line.”

Proposition 78 There is one and only one line that contains the distinct
points P and Q.

Proof: The affine flat k = {P + a · (Q − P ) | a ∈ F} clearly contains both

P and Q, and is a line since k v = {a · (Q − P ) | a ∈ F} is one-dimensional.

83
84 CHAPTER 7. BASIC AFFINE RESULTS AND METHODS

On the other hand, suppose that l is a line that contains both P and Q.
Since l is a line that contains P , it has the form l = P + lv where lv is one-
dimensional. Since l is an affine flat that contains both P and Q, it contains
the translate P + (Q − P ) so that Q − P ∈ lv , and therefore lv = k v and
l = k. 
The line that contains the distinct points P and Q is P +A Q, of course.
We will write P Q to mean the line P +A Q for (necessarily distinct) points
P and Q.

Corollary 79 Two distinct intersecting lines intersect in a single point. 

Exercise 7.2.1 Dimensional considerations then imply that the affine sum
of distinct intersecting lines is a plane.

Our second proposition involves parallelism, and can be used in a syn-

thetic context to prove the transitivity of parallelism for coplanar lines.

Proposition 80 Given a point P and a line l, there is one and only one
line m that contains P and is parallel to l.

Proof: In order for m and l to be parallel, we must have mv = lv , since m

and l have equal finite dimension. Hence the unique line sought is m = P +lv .
(Recalling that the cosets of a subspace form a partition of the full space, P
must lie in exactly one coset of lv .) 

7.3 A Basic Configuration Theorem

The following is one of the affine versions of a configuration theorem of pro-
jective geometry attributed to the French geometer Girard Desargues (1591-
1661). The particular affine version treated here is the completely nonparallel
one, the one that looks just like the projective theorem as we shall later see.

Theorem 81 (Desargues) Let A, A 0 , B, B 0 , C, C 0 be distinct points and let

the distinct lines AA 0 , BB 0 , and CC 0 be concurrent in the point P . Let the
nonparallel lines AB and A 0 B 0 intersect in the point C 00 , the nonparallel
lines BC and B 0 C 0 intersect in the point A 00 , and the nonparallel lines CA
and C 0 A 0 intersect in the point B 00 . Then A 00 , B 00 , and C 00 are collinear.
7.3. A BASIC CONFIGURATION THEOREM 85

B ''

A '

A A ''
B '

P
C C '' C '

Illustrating the Theorem of Desargues

Proof: We have P = a · A + a 0 · A 0 = b · B + b 0 · B 0 = c · C + c 0 · C 0 for
scalars a, a 0 , b, b 0 c, c 0 which satisfy a + a 0 = b + b 0 = c + c 0 = 1. We then
derive the three equalities
a · A − b · B = b 0 · B 0 − a 0 · A 0,
b · B − c · C = c 0 · C 0 − b 0 · B 0,
c · C − a · A = a 0 · A 0 − c 0 · C 0.
Consider the first of these. Suppose that a = b. Then a 0 = b 0 too, since
a+a 0 = b+b 0 = 1. But then a 6= 0, for if a = 0, then a 0 = 1, which would give
A 0 = B 0 contrary to hypothesis. Similarly a 0 6= 0. Thus supposing that a = b
leads to the nonparallel lines AB and A 0 B 0 being parallel. Hence a − b =
b 0 − a 0 6= 0, and similarly, b − c = c 0 − b 0 6= 0 and c − a = a 0 − c 0 6= 0. Denote
these pairs of equal nonzero differences by c 00 , a 00 , and b 00 , respectively. Then
the equalities above relate to the double-primed intersection points by
a · A − b · B = b 0 · B 0 − a 0 · A 0 = c 00 · C 00 ,
b · B − c · C = c 0 · C 0 − b 0 · B 0 = a 00 · A 00 ,
c · C − a · A = a 0 · A 0 − c 0 · C 0 = b 00 · B 00 ,
since if we divide by the double-primed scalar, each double-primed intersec-
tion point is the equal value of two affine expressions that determine points
on the intersecting lines. Adding up the equalities gives 0 = a 00 · A 00 + b 00 ·
B 00 +c 00 ·C 00 and noting that a 00 +b 00 +c 00 = 0 we conclude in light of Exercise
6.4.2 that A 00 , B 00 , and C 00 do lie on one line as was to be shown. 
86 CHAPTER 7. BASIC AFFINE RESULTS AND METHODS

7.4 Barycenters
We now introduce a concept analogous to the physical concept of “center
of gravity,” which will help us visualize the result of computing a linear ex-
pression and give us some convenient notation. Given the points A1 , . . . , An ,
consider the linear expression a1 · A1 + · · · + an · An of weight a1 + · · · + an 6= 0
based on A1 , . . . , An . The barycenter based on (the factors of the terms of)
this linear expression is the unique point X such that

(a1 + · · · + an ) · X = a1 · A1 + · · · + an · An .

(If the weight were zero, there would be no such unique point X.) The
barycenter of a1 · A1 + · · · + an · An , which of course lies in the affine span of
{A1 , . . . , An }, will be denoted by £ [a1 · A1 + · · · + an · An ]. Clearly, for any
scalar m 6= 0, £ [(ma1 ) · A1 + · · · + (man ) · An ] = £ [a1 · A1 + · · · + an · An ],
and this homogeneity property allows us to use £ [a1 · A1 + · · · + an · An ] as
a convenient way to refer to an affine expression by referring to any one of its
nonzero multiples instead. Also, supposing that 1 6 k < n, a = a1 +· · ·+ak 6=
0, a 0 = ak+1 + · · · + an 6= 0,

Xk = £ [a1 · A1 + · · · + ak · Ak ] ,

and
Xk 0 = £ [ak+1 · Ak+1 + · · · + an · An ] ,
it is easy to see that if a + a 0 6= 0 then

£ [a1 · A1 + · · · + an · An ] = £ [a · Xk + a 0 · Xk 0 ] .

Hence we may compute barycenters in a piecemeal fashion, provided that no

zero weight is encountered along the way.

7.5 A Second Basic Configuration Theorem

The following is the affine version of a configuration theorem of projective
geometry which dates back much farther than the theorem of Desargues given
above. Pappus of Alexandria, a Greek who lived in the 4th century, long
before projective geometry was created, used concepts of Euclidean geometry
to prove this same affine version of the projective theorem that now bears
his name.
7.5. A SECOND BASIC CONFIGURATION THEOREM 87

Theorem 82 (Theorem of Pappus) Let l, l 0 be distinct coplanar lines with

distinct points A, B, C ⊂ l r l 0 and distinct points A 0 , B 0 , C 0 ⊂ l 0 r l. If
BC 0 meets B 0 C in A 00 , AC 0 meets A 0 C in B 00 , and AB 0 meets A 0 B in C 00 ,
then A 00 , B 00 , C 00 are collinear.

A ''
A

C ''
B ''

A ' B ' C '

Illustrating the Theorem of Pappus
Proof: We will express A 00 , B 00 , C 00 in terms of the elements of the affine
frame {A, B 0 , C}. First, we may write for some a, b 0 , c with a + b 0 + c = 1,
B 00 = £ [a · A + b 0 · B 0 + c · C] .
Because B is on AC, we may write
B = £ [pa · A + c · C] .
Because A 0 is on B 00 C, we may write
A 0 = £ [B 00 + (q − 1)c · C]
= £ [a · A + b 0 · B 0 + qc · C] .
Because C 0 is on both AB 00 and A 0 B 0 , we may write C 0 = £ [r · A + q · B 00 ] =
£ [A 0 + s · B 0 ] or
C 0 = £ [(r + qa) · A + qb 0 · B 0 + qc · C]
= £ [a · A + (b 0 + s) · B 0 + qc · C]
88 CHAPTER 7. BASIC AFFINE RESULTS AND METHODS

and selecting between equal coefficients permits us to write

C 0 = £ [a · A + qb 0 · B 0 + qc · C] .

Because A 00 is on both BC 0 and B 0 C, we may write A 00 = £ [p · C 0 − k · B] =

£ [w · B 0 + x · C] or

A 00 = £ [(1 − k)pa · A + pqb 0 · B 0 + (pq − 1)c · C]

= £ [w · B 0 + x · C]

so that k = 1 and

A 00 = £ [pqb 0 · B 0 + (pq − 1)c · C] .

Because C 00 is on both A 0 B and AB 0 , we may write C 00 = £ [l · A 0 − q · B] =

£ [y · A + z · B 0 ] or

C 00 = £ [(l − qp)a · A + lb 0 · B 0 + (l − 1)qc · C]

= £ [y · A + z · B 0 ]

so that l = 1 and

C 00 = £ [(1 − qp)a · A + b 0 · B 0 ] .

Letting a 00 , b 00 , c 00 be the respective weights of the expressions above of which

A 00 , B 00 , C 00 are barycenters in terms of A, B 0 , C, we have

a 00 = pqb 0 + (pq − 1)c,

b 00 = a + b 0 + c,
c 00 = (1 − qp)a + b 0 .

We then find that

a 00 · A 00 + (1 − pq)b 00 · B 00 + (−c 00 ) · C 00 = 0

and the left-hand side is a directional expression. However, the coefficients

of this directional expression are not all zero. For if we suppose they are all
zero, then 1 − pq = 0, and b 0 = 0, which puts B 00 on AC so that C 0 would
be on AC contrary to hypothesis. We conclude in light of Exercise 6.4.2 that
A 00 , B 00 , and C 00 are collinear as was to be shown. 
7.6. VECTOR RATIOS 89

7.6 Vector Ratios

If two nonzero vectors u, v satisfy a · u = b · v for two nonzero scalars a, b, it
makes sense to speak of the proportion u : v = b : a. Under these conditions
we say that u, v have the same direction and only then do we impute a
u u b
scalar value to the vector ratio and write = .
−→ v v a
Let us write P Q = Q − P whenever P and Q are points. Nonzero vectors
−→ −→
P Q and RS have the same direction precisely when P Q and RS are parallel
−→ −→
lines. Hence the ratio P Q/RS has meaning whenever P Q and RS are parallel
lines, and in particular, when they are the same line.
X is a point on the line AB, and distinct from A and B, if and only if there
are nonzero scalars a, b, such that a+b 6= 0 and such that X = £[a·A+b·B].
But this is the same as saying that there are nonzero scalars a and b with
−−→ −→ −−→ −→
nonzero sum, such that (a + b) · AX = b · AB and (a + b) · XB = a · AB, or
equivalently such that
−−→
AX b
= ·
−−→ a
XB
−→ −−→ −−→
We say that AB = AX + XB is divided by X in the ratio b : a.
The result of the following Subcenter Exercise will find useful employment
below as a lemma.
Exercise 7.6.1 (Subcenter Exercise) Given noncollinear points A, B, C
and nonzero scalars a, b, c such that X = £ [a · A + b · B + c · C], barycentric
considerations imply that when a+b 6= 0, CX and AB meet in £ [a · A + b · B].
On the other hand, when a + b = 0, CX and AB are parallel. Hence
CX k AB if and only if a + b = 0, and CX ∩ AB = £ [a · A + b · B] if
and only if a + b 6= 0.

C X

A B
£ [a ·A + b ·B ]
Figure for Above Exercise
90 CHAPTER 7. BASIC AFFINE RESULTS AND METHODS

The next result is the affine version of Proposition VI.2 from Euclid’s
Elements. By P QR we will always mean the plane that is the affine sum of
three given (necessarily noncollinear) points P, Q, R.
Theorem 83 (Similarity Theorem) Let X be a point in ABC and not
on AB, BC, or CA, and let W = BX ∩ AC. Then
−−→ −−→ −−→
WC WX CX
CX k AB ⇒ = =
−−→ −−→ −→
WA WB AB
and −−→ −−→
WC WX
= ⇒ CX k AB.
−−→ −−→
WA WB

W
C X C X
W

A B B A
Two Possible Figures for the Similarity Theorem
Proof: “Let X be a point in ABC and not on AB, BC, or CA ” is
just the geometric way to say that there exist nonzero scalars a, b, c such
that X = £ [a · A + b · B + c · C]. Suppose first that CX k AB. Then
by the Subcenter Exercise above, a + b = 0 and we may replace a with
−b to give X = £ [(−b) · A + b · B + c · C] or c · (X − C) = b · (B − A).
−−→ −→
Hence CX/AB = b/c. Employing the Subcenter Exercise again, we find
−→
that W = BX ∩ AC = £ [(−b) · A + c · C] so that W divides AC in the
−−→ −−→
ratio c : −b and therefore W C/W A = b/c. X may also be obtained by
the piecemeal calculation X = £ [(c − b) · W + b · B] which is the same as
−−→ −−→
c · (X − W ) = b · (B − W ) or W X/W B = b/c. This proves the first part.
−−→ −−→ −−→ −−→
On the other hand, assume that W C/W A = W X/W B. We may suppose
that these two equal vector ratios both equal b/c for two nonzero scalars b, c.
Then c · (C − W ) = b · (A − W ) and c · (X − W ) = b · (B − W ). Subtracting
the first of these two equations from the second, we obtain c · (X − C) =
b · (B − A), which shows that CX k AB as desired. 
7.7. A VECTOR RATIO PRODUCT AND COLLINEARITY 91

7.7 A Vector Ratio Product and Collinearity

The following theorem provides a vector ratio product criterion for three
points to be collinear. The Euclidean version is attributed to Menelaus of
Alexandria, a Greek geometer of the late 1st century.

Theorem 84 (Theorem of Menelaus) Let A, B, C be noncollinear points,

and let A 0 , B 0 , C 0 be points none of which is A, B, or C, such that C 0 is on
AB, A 0 is on BC, and B 0 is on CA. Then A 0 , B 0 , C 0 are collinear if and
only if
−−→0 −−→0 −−→0
BA CB AC
· · = −1.
−−0→ −−→ 0
−−0→
A C B A C B

|
C

A '
B '

B A C '
Figure for the Theorem of Menelaus

Proof: Let A 0 , B 0 , C 0 be collinear. B, C, C 0 are noncollinear and we

may write B 0 = £ [b · B + c · C + c 0 · C 0 ]. Then A 0 = £ [b · B + c · C] and
A = £ [b · B + c 0 · C 0 ] by the Subcenter Exercise above. B 0 may also be
obtained by the piecemeal calculation B 0 = £ [c · C + (b + c 0 ) · A]. Thus A 0
−−→ −→
divides BC in the ratio c : b, B 0 divides CA in the ratio (b + c 0 ) : c, and A
−−→
divides BC 0 in the ratio c 0 : b. Hence
−−→0 −−→0 −−→0
BA CB AC c b + c0 b
· · = · · = −1
−−0→ −−→ −−0→ b c −b − c 0
A C B 0A C B
as required.
92 CHAPTER 7. BASIC AFFINE RESULTS AND METHODS

On the other hand, suppose that the product of the ratios is −1. We may
suppose that
A 0 = £ [1 · B + x · C] , B 0 = £ [y · A + x · C] , C 0 = £ [y · A + z · B]
where none of x, y, z, 1 + x, y + x, y + z is zero. Thus
x y z
· · = −1
1 x y
and hence z = −1. Then we have
(1 + x) · A 0 + (−y − x) · B 0 + (y − 1) · C 0 = 0
and (1 + x) + (−y − x) + (y − 1) = 0. We conclude in light of Exercise 6.4.2
that A 0 , B 0 , C 0 are collinear as required. 

7.8 A Vector Ratio Product and Concurrency

The Italian geometer Giovanni Ceva (1647-1734) is credited with the Eu-
clidean version of the following theorem which is quite similar to the above
theorem of Menelaus but instead provides a criterion for three lines to be
either concurrent or parallel.
Theorem 85 (Theorem of Ceva) Let A, B, C be noncollinear points, and
let A 0 , B 0 , C 0 be points none of which is A, B, or C, such that C 0 is on AB,
A 0 is on BC, and B 0 is on CA. Then AA 0 , BB 0 , CC 0 are either concurrent
in a point not lying on AB, BC, or CA, or are parallel, if and only if
−−→0 −−→0 −−→0
BA CB AC
· · = 1.
−−0→ −−→0
−−0→
A C B A C B

|
C A '

B ' A ' B '

P C

A C ' B A C ' B
Figures for the Theorem of Ceva
7.8. A VECTOR RATIO PRODUCT AND CONCURRENCY 93

Proof: Suppose AA 0 , BB 0 , CC 0 meet in P = £ [a · A + b · B + c · C] with

nonzero scalars a, b, c. By the Subcenter Exercise above, A 0 = £ [b · B + c · C]
−−→ −−→ −−→ −−→
so that BA 0 /A 0 C = c/b. Similarly, we find that CB 0 /B 0 A = a/c and
−−→0 −−0→
AC /C B = b/a. The desired product result follows at once.
Suppose next that AA 0 , BB 0 , CC 0 are parallel. Applying the Similarity
Theorem above, AA 0 k CC 0 implies that C divides A 0 B in the same ratio
as C 0 divides AB, and BB 0 k CC 0 implies that C also divides AB 0 in the
same ratio as C 0 divides AB. The desired product result follows readily.
Now suppose that the product is as stated. According to the conditions
under which we allow vector ratios to be written, there are nonzero scalars
−−→ −−→ −−→ −−→ −−→ −−→
a, b, c such that BA 0 /A 0 C = c/b, CB 0 /B 0 A = a/c, and AC 0 /C 0 B = b/a. If
−−→
a + b + c 6= 0, define P = £ [a · A + b · B + c · C]. Then BC is divided by
A 0 in the ratio c : b, A 0 = £ [b · B + c · C], and barycentric considerations
place P on AA 0 . Similarly, P is on BB 0 and CC 0 . If, on the other hand,
a + b + c = 0, then c = −a − b. We see then that C divides A 0 B in the same
ratio as C 0 divides AB, and that C also divides AB 0 in the same ratio as
C 0 divides AB. The Similarity Theorem above then gives AA 0 k CC 0 and
BB 0 k CC 0 . 
94 CHAPTER 7. BASIC AFFINE RESULTS AND METHODS
Chapter 8

An Enhanced Affine
Environment

8.1 Inflating a Flat

Let V be our usual vector space over the field F. Any flat Φ ∈ A (V) may
be embedded in a vector space Φ+ as a hyperplane not containing 0, a con-
struction that we will call inflating Φ. Employing the directional subspace
Φv of Φ, we form Φ+ = F × Φv and embed Φ in it as {1} × Φv by the affine
isomorphism that first sends x ∈ Φ to xv = (x − o) ∈ Φv and then sends xv
to 1 × xv = (1, xv ), where o is a fixed but arbitrary vector in Φ. Forming
Φ+ begins a transition from affine to projective geometry. Benefits of this
enhancement of the affine environment include more freedom in representing
points (similar to that gained by using the £ operator) and the ability to
completely separate the vectors representing points from the vectors repre-
senting directions. Having then gained the separation of point-representing
vectors and direction-representing vectors, the directions can be added to the
point community as a new special class of generalized points which from the
affine viewpoint are thought of as lying at infinity. These points at infinity
can be used to deal in a uniform way with cases where an intersection point
of lines disappears to infinity as the lines become parallel. This leads us
into the viewpoint of projective geometry where what were points at infinity
become just ordinary points, and parallel lines cease to exist.
We now assume that any flat Φ which is of geometric interest to us has
been inflated, so that the rôle of Φ will be played by its alias {1} × Φv in

95
96 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

A (Φ+ ), and the rôle of Φv then will be played by its alias {0}×Φv . To simplify
the notation, we will refer to {a} × Φv as Φa , so that Φ1 is playing the rôle of
Φ and Φ0 is playing the rôle of Φv . The rôle of any subflat Ψ ⊂ Φ is played by
the subflat Ψ1 = {1} × Ψv ⊂ Φ1 , of course. No matter what holds for Φ and
its directional subspace, Φ1 is disjoint from its directional subspace Φ0 . Lines
through the point O = {0} in Φ+ are of two fundamentally different types
with regard to Φ1 : either such a line meets Φ1 or it does not. The nonzero
vectors of the lines through O that meet Φ1 will be called point vectors, and
the nonzero vectors of the remaining lines through O (those that lie in Φ0 )
will be called direction vectors. There is a unique line through O passing
through any given point P of Φ1 , and since any nonzero vector in that line
may be used as an identifier of P , that line and each of its point vectors will
be said to represent P . Point vectors are homogeneous representatives of
their points in the sense that multiplication by any nonzero scalar gives a
result that still represents exactly the same point. To complete our view,
each line through O lying in Φ0 will be called a direction, and it and each of
its nonzero vectors will be said to represent that direction. Thus the zero
vector stands alone in not representing either a point or a direction.

© +

P © 1

© 0
O
.
Illustrating Inflating Φ
8.1. INFLATING A FLAT 97

We are now close to the viewpoint of projective geometry where the lines
through 0 actually are the points, and the n-dimensional subspaces are the
(n − 1)-dimensional flats of a projective structure. However, the inflated-flat
viewpoint is still basically an affine viewpoint, but an enhanced one that is
close to the pure projective one. From now on, we assume that any affine
flat has been inflated and freely use the resulting enhanced affine viewpoint.
Suppose now that P, Q, R ∈ Φ1 are distinct points represented by the
point vectors p, q, r. Our geometric intuition indicates that P, Q, R are
collinear in Φ1 if and only if p, q, r are coplanar in Φ+ . The following propo-
sition confirms this.

Proposition 86 Let P, Q, R ∈ Φ1 be distinct points represented by the point

vectors p, q, r. Then P, Q, R are collinear if and only if {p, q, r} is dependent.

Proof: Let i denote the vector 1 × 0 in Φ+ = F × Φv . Then points in

Φ1 all have the form {i + x} where x ∈ Φ0 , and point vectors all have the
form d · (i + x) for some nonzero scalar d. We thus set P = {i + u} , Q =
{i + v} , R = {i + w}, and p = a · (i + u) , q = b · (i + v) , r = c · (i + w).
Supposing that P, Q, R are collinear, there are nonzero scalars k, l, m such
that k · (i + u) + l · (i + v) + m · (i + w) = 0. Therefore p, q, r are related by

k l m
·p+ ·q+ · r = 0.
a b c
On the other hand, suppose that there are scalars f, g, h, not all zero,
such that f · p + g · q + h · r = 0. Then

f a · (i + u) + gb · (i + v) + hc · (i + w) = 0.

Since i ∈
/ Φ0 , f a + gb + hc = 0 and P, Q, R are collinear by Exercise 6.4.2. 

Note that since we have complete freedom in scaling point vectors by

nonzero scalars, if we are given distinct collinear P, Q, R, there always is a
{p, q, r} such that the scalar coefficients of a dependency linear combina-
tion of it are any three nonzero scalars we choose. Moreover, since a depen-
dency linear combination is unaffected by a nonzero scaling, besides the three
nonzero scalar coefficients, one of p, q, r may also be picked in advance. This
is the freedom in representing points which we get from having insured that
98 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

the flat where all the points lie is completely separated from its directional
subspace.
If we are given two distinct points P, Q represented by point vectors p, q,
the above proposition implies that any point vector in the span of {p, q}
represents a point on the line P Q, and conversely, that every point of P Q
in Φ1 is represented by a point vector of h{p, q}i. But besides point vectors,
there are other nonzero vectors in h{p, q}i, namely direction vectors. It is
easy to see that the single direction contained in h{p, q}i is V = h{p, q}i ∩ Φ0
which we refer to as the direction of P Q. We will consider V to be a
generalized point of P Q, and we think of V as lying “at infinity” on P Q.
Doing this will allow us to treat all nonzero vectors of h{p, q}i in a uniform
manner and to say that h{p, q}i represents the line PQ. This will turn out to
have the benefit of eliminating the need for separate consideration of related
“parallel cases” for many of our results.
Exercise 8.1.1 Let P, Q be distinct points represented by point vectors p, q,
and let the direction V be represented by the direction vector v. Then the lines
h{p, v}i ∩ Φ1 and h{q, v}i ∩ Φ1 are parallel and h{p, v}i ∩ h{q, v}i = h{v}i.
(Thus we speak of the parallel lines P V and QV that meet at V .)
Besides lines like P V and QV of the above exercise, there can be lines of
the form V W where V and W are distinct directions. We speak of such lines
as lines at infinity. If V is represented by v and W by w, then the nonzero
vectors of h{v, w}i (all of which are direction vectors) are exactly the vectors
that represent the points of V W . Thus we view the “compound direction”
h{v, w}i as a generalized line. Each plane has its line at infinity, and planes
have the same line at infinity if and only if they are parallel.

8.2 Desargues Revisited

We now state and prove another version of the theorem of Desargues treated
previously. The theorem’s points and lines are now assumed to exist in our
new generalized framework.
Theorem 87 (Desargues) Let A, A 0 , B, B 0 , C, C 0 , P be distinct points and
let the lines AA 0 , BB 0 , and CC 0 be distinct and concurrent in P . Let AB
and A 0 B 0 meet in the point C 00 , BC and B 0 C 0 meet in the point A 00 ,and
CA and C 0 A 0 meet in the point B 00 . Then A 00 , B 00 , and C 00 all lie on the
same line.
8.2. DESARGUES REVISITED 99

Proof: The corresponding small letter will always stand for a representing
vector. We can find representing vectors such that

p = a + a0 = b + b0 = c + c0

so that

a − b = b 0 − a 0 = c 00
b − c = c 0 − b 0 = a 00
c − a = a 0 − c 0 = b 00 .

Hence a 00 + b 00 + c 00 = 0. 
While we were able to use simpler equations, this proof is not essentially
different from the previous one. What is more important is that new in-
terpretations now arise from letting various of the generalized points be at
infinity. We depict two of these possible new interpretations in the pair of
figures below.

B ''
A '
A '
A '' A

B B '
A P
B '

B C

P
C '
C C '

New Interpretations Arising from Generalized-Framework Version

.
In the left-hand figure, notice that A 00 B 00 is parallel to the parallel lines
AB and A 0 B 0 , as it contains their intersection point C 00 at infinity. In the
right-hand figure, A 00 , B 00 , and C 00 all lie on the same line at infinity, and we
100 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

can say that the fact that any two of A 00 , B 00 , C 00 lie at infinity means that
the third must also. Thus the right-hand figure depicts the result that if any
two of the three corresponding pairs AB, A 0 B 0 , etc., are pairs of parallels,
then the third is also.

8.3 Representing (Generalized) Subflats

Notice: The qualifier “generalized” will henceforth usually be omitted. From
now on, points of Φ or other subflats to which we refer may conceivably be “at
infinity” unless specifically stated otherwise. This includes intersections, so
that, for example, parallel lines will be deemed to intersect in a point just like
nonparallel coplanar lines. Similarly, the more or less obvious generalizations
of all encountered affine concepts (affine span, affine sum, etc.) are to be
assumed to be in use whether specifically defined or not.

Within the context of an inflated flat we can readily show that there is a
one-to-one correspondence between the (n − 1)-dimensional subflats of Φ and
the n-dimensional subspaces of Φ+ . Moreover, n points of Φ affine-span one
of its (n − 1)-dimensional subflats if and only if every set of vectors which
represents those points is an independent set in Φ+ . (The concept of affine
span of points is generalized through use of the span of the representing
vectors to include all points, not just “finite” ones.)
The criterion for independence provided by exterior algebra (Corollary 65)
leads to the conclusion that a nonzero exterior e-product (a blade) represents
a subspace of Φ+ and therefore also represents a subflat of Φ in the same
homogeneous fashion that a single nonzero vector represents a point. That
is, if {v1 , . . . , vn } is an independent n-element set of vectors in Φ+ , then
the n-blade v1 ∧ · · · ∧ vn homogeneously represents an (n − 1)-dimensional
subflat of Φ. This is based on the following result.

Proposition 88 If {v1 , . . . , vn } and {w1 , . . . , wn } are independent n-element

sets of vectors that span the same subspace, then the blades w1 ∧ · · · ∧ wn and
v1 ∧ · · · ∧ vn are proportional, and conversely, if w1 ∧ · · · ∧ wn and v1 ∧
· · · ∧ vn are proportional blades made up of the vectors of the independent
sets {w1 , . . . , wn } and {v1 , . . . , vn }, then those sets of vectors span the same
subspace.
8.4. HOMOGENEOUS AND PLÜCKER COORDINATES 101

Proof: Suppose that the independent sets {v1 , . . . , vn } and {w1 , . . . , wn }

of n vectors span the same subspace. Then each wi can be expressed as a
linear combination of {v1 , . . . , vn }. Putting these linear combinations in the
blade w1 ∧ · · · ∧ wn , after expanding and collecting the nonzero terms we
are left with a nonzero (since w1 ∧ · · · ∧ wn is nonzero) multiple of the blade
v1 ∧ · · · ∧ vn .
On the other hand, suppose that w1 ∧ · · · ∧ wn and v1 ∧ · · · ∧ vn are
proportional blades expressed as exterior products of vectors. Then
0 = wi ∧ w1 ∧ · · · ∧ wn = wi ∧ v1 ∧ · · · ∧ vn
so that each wi is a linear combination of {v1 , . . . , vn }. 

We will find it convenient to class the empty exterior product, which we

by convention set equal to 1, and all its nonzero multiples, as blades that
represents the empty flat and the subspace {0}. Thus, there isVa one-to-
one correspondence between the sets of proportional blades in Φ+ and
the finite-dimensional subflats of Φ. Given the two blades v1 ∧ · · · ∧ vk and
vk+1 ∧· · ·∧vn , then either v1 ∧· · ·∧vn = 0 and the two corresponding subflats
intersect, or v1 ∧ · · · ∧ vn represents the (generalized) affine sum of the two
non-intersecting subflats. By the same token, given the blade β, the subspace
of Φ+ which it represents is {v ∈ Φ+ | v ∧ β = 0}. A blade that represents
a hyperplane will be referred to as a hyperplane blade, and similarly, the
terms plane blade, line blade, and point blade will be used.

8.4 Homogeneous and Plücker Coordinates

We suppose throughout this section and for the remainder of this chapter
that Φ+ has finite dimension d. Fix a basis B for Φ+ . The coordinates of
a nonzero vector v of Φ+ are known as V homogenous coordinates for the
∧
point represented by v. Bases B for Φ+ may be obtained by choosing
blades that are products of elements of B. The coefficients of the expan-
sion for the n-blade β in terms of the n-blades of any such B ∧ are known,
in honor of German geometer Julius Plücker (1801-1868), as Plücker co-
ordinates for the (n − 1)-dimensional subflat of Φ represented by β, and
sometimes, by abuse of language, as Plücker coordinates for any blade pro-
portional to β. Homogenous coordinates are then Plücker coordinates for a
point. Plücker coordinates are sometimes called Grassmann coordinates
102 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

for German schoolteacher Hermann Grassmann (1809-1877), the first author

to systematically treat vector spaces and vector algebras.
The coefficients of any nonzero linear combination of B are homogenous
coordinates for some point of Φ, but an arbitrary nonzero linear combination
of the n-blades of a B ∧ cannot be guaranteed in general to yield Plücker
coordinates for some (n − 1)-dimensional subflat Vn of Φ. That is, it is not
true in general that every nonzero element of Φ+ is expressible
Vd−1as an n-
blade. We shall find, however, that every nonzero element of Φ+ is a
(d − 1)-blade.
Plücker coordinates of a blade may be expressed in terms of the coordi-
nates of the vectors that make up the blade by expanding out the exterior
product. If we suppose that we are given the n (possibly dependent) vectors
v1 , . . . , vn with each vj given in terms of basis vectors xi by
d
X
vj = ai,j · xi ,
i=1

then
d
X d
X
v1 ∧ · · · ∧ vn = ··· · ai1 ,1 · · · ain ,n · xi1 ∧ · · · ∧ xin
i1 =1 in =1

and collecting terms on a particular set of basis elements (those that have
subscripts that increase from left to right) we then get
!
X X σ
(−1) aσ(i1 ),1 · · · aσ(in ),n · xi1 ∧ · · · ∧ xin
16i1 <···<in 6d σ

where the inner sum is over all permutations

 
i1 ··· in
σ=
σ (i1 ) · · · σ (in )

of each subscript combination that the outer summation has selected. We

recognize the inner sum as a determinant and thus the expansion may be
written as
 
 ai ,1 · · · ai ,n 
 1 1 
X  .. ..  · x ∧ · · · ∧ x .
 .
 ··· .  i1 in

in ,1 · · ·
16i1 <···<in 6d  a ain ,n 
8.5. DUAL DESCRIPTION VIA ANNIHILATORS 103

We may phrase this result in terms of points and infer the results of the
following exercise which also contains a well-known result about the inde-
pendence of the columns of a d × n matrix and the determinants of its n × n
minors.

Exercise 8.4.1 Let P1 , . . . , Pn be n 6 d distinct points of Φ and for each

j let Aj = (a1,j , . . . , ad,j ) be a d-tuple of homogenous coordinates for Pj in
terms of some basis B for Φ+ . Let A = [ai,j ] be the d × n matrix formed
by using the Aj as its columns. Then if {P1 , . . . , Pn }
is affine independent,
there is a B ∧ such that in terms of its n-blades the nd determinants of n × n
submatrices of A formed in each possible way by deleting all but n rows of
A are Plücker coordinates for the affine span of {P1 , . . . , Pn }. On the other
hand, if {P1 , . . . , Pn } is affine dependent, these determinants all vanish.

8.5 Dual Description Via Annihilators

Remembering that we are assuming that Φ+ has finite dimension d, Φ> +
also has finite dimension d. Each subspace W of dimension n in Φ+ has the
annihilator W 0 (the set of elements of Φ>
+ which annihilate every vector of W)
that is a (d − n)-dimensional subspace of Φ> + . As in Chapter 3, we identify
> > 0


Φ+ with Φ+ , and thereby are entitled to write W 00 = (W 0 ) = W. The
assignment of W 0 to W clearly creates a one-to-one correspondence between
the subspaces of Φ+ and the subspaces of Φ> 0
+ , so that giving W is equivalent
to giving W.
W = W 00 describes W as the annihilator of W 0 , which amounts to saying
that W is the set of those vectors in Φ+ which are annihilated by every
element of W 0 . This dual description gives W as the intersection of all the
hyperplanes that contain it. Verification of this is provided by the following
exercise, once we observe that the hyperplanes through 0 in Φ+ are precisely
the kernels of the linear functionals on Φ+ .

Exercise 8.5.1 Let W be a subspace of Φ+ . Then W 0 is the set of all

elements of Φ> + whose kernel contains W. Let X be the intersection of all
the kernels of the elements of W 0 , i. e., the intersection of all hyperplanes
containing W. Show that W = X by verifying that W ⊂ X and X ⊂ W 00 .

We now point out and justify what is from a geometric standpoint perhaps
the most important result concerning the annihilator, namely that for any
104 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

subspaces W and X of Φ+ we have

(W + X )0 = W 0 ∩ X 0 .

This is the same as saying that f ∈ Φ> + satisfies f (w + x) = 0 for all w ∈ W

and for all x ∈ X if and only if f (w) = 0 for all w ∈ W and f (x) = 0
for all x ∈ X . The truth of the “if” part is obvious, and that of the “only
if” part is nearly so. For, since 0 is in any subspace, it follows at once that
f (w + 0) = 0 for all w ∈ W and f (0 + x) = 0 for all x ∈ X . Interchanging
Φ+ with Φ> + then gives us also

0
W0 + X 0 =W ∩X .

Exercise 8.5.2 Let f1 , . . . , fn be elements of Φ>

+ . Then

h{f1 , . . . , fn }i0 = (h{f1 }i + · · · + h{fn }i)0 = {f1 }0 ∩ · · · ∩ {fn }0 ,

and thus any subspace of Φ+ is the intersection of finitely many hyperplanes.

8.6 Direct and Dual Plücker Coordinates

Let {t1 , . . . , tn } be a basis for the subspace W of Φ+ , and let A = {t1 , . . . , td }
be any extension ofthis basis for W to a basis for Φ+ . Then the basis A
has the dual A> = t> 1 , . . . , t >
d , and t>
n+1 , . . . , t >
d is a basis for W 0
. Thus
we have the blades proportional to ω = t1 ∧ · · · ∧ tn representing W, and
the corresponding annihilator blades proportional to ω 0 = t> >
n+1 ∧ · · · ∧ td
representing W 0 (and therefore also representing W in a dual sense).

Exercise 8.6.1 Given ω and ω 0 as above, the subspace of Φ+ which ω rep-

resents is the set of all vectors v ∈ Φ+ that simultaneously satisfy the d − n
equations t> >
n+1 (v) = 0, . . . , td (v) = 0.

V purposes, fix a basis B = {x1 , . . . , xd } for Φ+ . Con-

For coordinatization
∧
struct a basis VB for Φ+ as we did previously, and similarly construct a
basis B for Φ>
>∧ >
+ based on the elements of the dual basis B . In terms
∧
of B , Plücker coordinates stemming from blades proportional to ω will now
be called direct (Plücker) coordinates, while coordinates in terms of B >∧
and stemming from blades proportional to ω 0 will be called dual (Plücker)
coordinates.
8.6. DIRECT AND DUAL PLÜCKER COORDINATES 105

Each n-blade χ of B ∧ has a matching annihilator blade χ0 ∈ B >∧ which

is the product of the d − n coordinate functions of the basis vectors not in
χ. This gives a one-to-one match between the basis elements of B ∧ and B >∧ .
We will find that, apart from an easily-determined sign factor, direct and
dual coordinates corresponding to matching basis elements may be taken to
be equal. Using the bases B ∧ and B >∧ therefore makes it easy to convert
back and forth between direct and dual coordinates since at worst we need
only multiply by −1. To see how this works, let us suppose, as we did just
above, that we have a blade ω = t1 ∧ · · · ∧ tn and its annihilator blade
ω 0 = t> >
n+1 ∧ · · · ∧ td where A = {t1 , . . . , td } is a basis for Φ+ . The bases
A and A are respectively related to the coordinatization bases B and B >
>
>
by an automorphism f and (Section 3.5) its contragredient f −> = (f −1 )
according to
tj = f (xj ) and t> −>
x>


j = f j .

Now suppose that f and f −1 are given in terms of the elements of the
basis B by
d
X d
X
−1
f (xj ) = ai,j · xi and f (xj ) = αi,j · xi
i=1 i=1

and thus (Exercise 3.4.1)

d

X
f −> x>
j = αj,i · x>
i .
i=1

Therefore
d
! d
!
X X
ω = t1 ∧ · · · ∧ tn = ai,1 · xi ∧ ··· ∧ ai,n · xi
i=1 i=1

and
d
! d
!
X X
ω 0 = t> >
n+1 ∧ · · · ∧ td = αn+1,i · x>
i ∧ ··· ∧ αd,i · x>
i .
i=1 i=1

For distinct i1 , . . . , in we wish to compare a coordinate pi1 ,...,in related to the

basis element xi1 ∧ · · · ∧ xin in ω to that of a dual coordinate πin+1 ,...,id related
to the basis element x> > 0
in+1 ∧ · · · ∧ xid in ω , where the two sets of subscripts
106 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

are complements in {1, .., d}. As our coordinates we take the coefficients
of the relevant blades in the expressions for ω and ω 0 above, namely the
determinants X
pi1 ,...,in = (−1)σ aσ(i1 ),1 · · · aσ(in ),n
σ

and X
πin+1 ,...,id = (−1)σ αn+1,σ(in+1 ) · · · αd,σ(id )
σ

where each sum is over all permutations σ of the indicated subscript set, and
as usual, (−1)σ = +1 or −1 according as the permutation σ is even or odd.
The following result attributed to the noted German mathematician Carl
G. J. Jacobi (1804-1851) provides the final key.

Lemma 89 (Jacobi’s Determinant Identity) Fix a basis B = {x1 , . . . , xd }

for Φ+ and let g be the automorphism of Φ+ such that for each j
d
X d
X
−1
g (xj ) = bi,j · xi and g (xj ) = βi,j · xi .
i=1 i=1

Then
det [ci,j ] = (det g) det [γi,j ]
where [ci,j ] is the n×n matrix with elements ci,j = bi,j , 1 6 i, j 6 n, and [γi,j ]
is the (d − n) × (d − n) matrix with elements γi,j = βi+n,j+n , 1 6 i, j 6 d − n.

Proof: Define the self map h of Φ+ by specifying its values on B as follows:


xj , j = 1, . . . , n,
h (xj ) = −1
g (xj ) , j = n + 1, . . . , d.

Then 
g (xj ) , j = 1, . . . , n,
g (h (xj )) =
xj , j = n + 1, . . . , d.
We have

x1 ∧ · · · ∧ xn ∧ g −1 (xn+1 ) ∧ · · · ∧ g −1 (xd ) = (det h) · x1 ∧ · · · ∧ xd

and it is readily established that det h = det [γi,j ]. Similarly, det g ◦ h =

det [ci,j ]. The lemma now follows from the Product Theorem (Theorem 67).

8.6. DIRECT AND DUAL PLÜCKER COORDINATES 107

We now construct a g that we will use in the lemma to produce the

result we seek. Let ρ be the permutation such that ρ(k) = ik . Define the
automorphism q of Φ+ by
q (xi ) = xρ−1 (i) .
We apply the lemma to g = q ◦ f so that
d
X
g (xj ) = q(f (xj )) = ai,j · q(xi )
i=1
d
X d
X
= ai,j · xρ−1 (i) = aρ(k),j · xk
i=1 k=1
d
X d
X
= aik ,j · xk = bk,j · xk
k=1 k=1

and
d
X
−1 −1 −1 −1



g (xj ) = f q (xj ) = f xρ(j) = αi,ρ(j) · xi
i=1
d
X d
X
= αi,ij · xi = βi,j · xi .
i=1 i=1

We observe that for this particular g

det [ci,j ] = pi1 ,...,in , det [γi,j ] = πin+1 ,...,id .
We also see that
det g = (det q) (det f ) = (−1)ρ det f .
Hence, by the lemma
pi1 ,...,in = (−1)ρ (det f )πin+1 ,...,id
and this is the result we have been seeking. Up to the factor (−1)ρ , ω and
ω 0 may then be taken to have the same coordinate with respect to matching
basis elements, since the factor det f may be ignored due to homogeneity.
Thus we may easily obtain a corresponding annihilator blade for any
blade that is expressed as a linear combination of basis blades, and we then
have
V the means to define a vector space isomorphism
V > that sends each blade
in Φ+ to a corresponding annihilator blade in Φ+ , as we now record.
108 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

Theorem
V 90 Let Φ+ have the basis B = {x1 , . . . , xd }. Choose a basis B ∧ for
Φ+ made up of exterior products (including the empty product which V we
take to equal 1) of the elements of B. Similarly choose a basis V B for V Φ>
>∧
+
made up of exterior products of the elements of B . Let H : Φ+ → Φ>
>
+
be the vector space map such that for each xi1 ∧ · · · ∧ xin ∈ B ∧

H (xi1 ∧ · · · ∧ xin ) = (−1)ρ · x> >

in+1 ∧ · · · ∧ xid

where x> >

in+1 ∧ · · · ∧ xid ∈ B
>∧
, {i1 , . . . , in } ∪ {in+1 , . . . , id } = {1, . . . , d}, and
 
1 ··· d
ρ= .
i1 · · · id

Then H is a vector space isomorphism that sends each blade to a correspond-

ing annihilator blade. 

Exercise 8.6.2 The H of the theorem above does not depend on the partic-
ular order of the factors used to make up each element of the chosen bases
B ∧ and B >∧ . That is, for any n and any permutation
 
1 ··· d
ρ=
i1 · · · id

we have
H (xi1 ∧ · · · ∧ xin ) = (−1)ρ · x> >
in+1 ∧ · · · ∧ xid

and therefore
ρ
H −1 (x> >
in+1 ∧ · · · ∧ xid ) = (−1) · xi1 ∧ · · · ∧ xin .

Henceforth, unless stated otherwise, we suppose that we are employing

the H based on a fixed underlying basis B = {x1 , . . . , xd } with a fixed as-
signment of subscript labels to basis vectors.

Exercise 8.6.3 Use H and H −1 to show that any linear combination of a

set of (d − 1)-blades equals a (d − 1)-blade or 0.

Exercise 8.6.4 The hyperplane blade x1 ∧ · · · ∧ xi−1 ∧ xi+1 ∧ · · · ∧ xd repre-

 0
sents the subspace x>
i and the corresponding subflat, called a coordinate
hyperplane, for which the coordinate value corresponding to the basis vector
xi is always 0.
8.7. A DUAL EXTERIOR PRODUCT 109

8.7 A Dual Exterior Product

V
A dual to our original “wedge” product ∧ on Φ+ is provided by the “vee”
product ∨ that we define by
χ ∨ ψ = H −1 (H (χ) ∧ H (ψ)) .
We may describe ∨ as being obtained by “pulling back” ∧ through H.
V
Exercise 8.7.1 Verify that Φ+ with theV∨ product is a legitimate vector
algebra by showing that itVis just an alias of Φ> + with the ∧ product, obtained
by relabeling ζ and η of Φ+ as H (ζ) and H −1 (η), V
> −1
and writing ∨ instead
of ∧. What is the unit element of this new algebra on Φ+ ?
V
The vector space Φ+ with the ∨ product is thus another exterior algebra
on the same space. In its dual relationship with the original exterior algebra,
hyperplane blades play the rôle that V1 point blades Vd−1 (nonzero vectors) play in
the original. That is, the Vnrôles of Φ
Vd−n+ and Φ+ become interchanged,
and in fact the rôles of Φ+ and Φ+ all become interchanged. The ∨
product of independent hyperplane blades represents the intersection of the
hyperplanes that those blades represent, just as the ∧ product of independent
vectors represents the subspace sum of the subspaces that those vectors rep-
resent. The ∨ product of dependent hyperplane blades is 0. The ∨ product
of two blades is 0 if and only if those blades represent subspaces that have
a subspace sum smaller than the whole space Φ+ , just as the ∧ product of
two blades is 0 if and only if those blades represent subspaces that have an
intersection larger than the 0 subspace. Following custom (and Grassmann’s
lead), we will call the original ∧ product progressive and the new ∨ product
regressive, terms that reflect the respective relationships of the products to
subspace sum and intersection. In expressions, the progressive product will
be given precedence so that by u ∧ v ∨ w ∧ x will be meant (u ∧ v) ∨ (w ∧ x).
 0
Exercise 8.7.2 Let x̄i denote a hyperplane blade that represents x> .
 > 0  > i 0
Then if {ik+1 , . . . , id } = {1, . . . , d} r {i1 , . . . , ik }, xi1 ∩ · · · ∩ xik is
represented by x̄i1 ∨ · · · ∨ x̄ik = ±xik+1 ∧ · · · ∧ xid .

8.8 Results Concerning Zero Coordinates

The basic significance of a zero coordinate is given in the following result.
110 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

Proposition 91 In the blade β, the Plücker coordinate corresponding to the

basis element ξ = xi1 ∧ · · · ∧ xin is zero if and only if the flat represented
by β intersects the flat represented by the blade ξ¯ = xin+1 ∧ · · · ∧ xid , where
I = {in+1 , . . . , id } is the complement of I = {i1 , . . . , in } in D = {1, .., d}.

Proof: ψ ∧ ξ¯ = 0 for each basis n-blade ψ = xk1 ∧ · · · ∧ xkn with the

exception of ξ. (Each n-element subset of D except I meets I since the only
n-element subset of I is I itself). Therefore β ∧ ξ¯ is zero or not according as
β’s Plücker coordinate corresponding to the basis element ξ is zero or not.


Exercise 8.8.1 The proposition above implies, as expected, that the point
represented by the basis vector x1 lies in the coordinate hyperplanes repre-
0
sented by x> i , i = 2, . . . , d.

The previous result serves as a prelude to the following generalization

that treats the case where there is a particular family of zero coordinates.

Proposition
 > 0 92 Let  >1 6 k 6 n 6 d. The flat represented by the subspace
0
X = xi1 ∩ · · · ∩ xik intersects the flat represented by the n-dimensional
subspace Y in a flat of dimension at least n − k if and only if the flat repre-
sented by Y has a zero Plücker coordinate corresponding to each of the basis
elements xj1 ∧ · · · ∧ xjn such that {i1 , . . . , ik } ⊂ {j1 , . . . , jn }.

Proof: Suppose first that the flat represented by Y has a zero Plücker
coordinate corresponding to each of the basis elements xj1 ∧ · · · ∧ xjn such
that {i1 , . . . , ik } ⊂ {j1 , . . . , jn }. The flats represented by X and Y do in-
tersect since by the previous proposition the flat represented by Y inter-
sects (several, perhaps, but one is enough) a flat represented by a subspace
 > 0  0
xj 1 ∩ · · · ∩ x> jn such that {i1 , . . . , ik } ⊂ {j1 , . . . , jn } and must therefore
intersect the flat represented by its factor  > 0X . Denote by x̄m the basis ele-
ment that represents the hyperplane xm . Then X is represented by the
regressive product ξ = x̄i1 ∨ · · · ∨ x̄ik , and each basis n-blade xj1 ∧ · · · ∧ xjn
is a regressive product of the d − n hyperplane blades x̄mn+1 , . . . , x̄md such
that mn+1 , . . . , md ∈ {1, . . . , d} r {j1 , . . . , jn }. Hence the regressive product
η ∨ ξ, where η is a blade that represents the flat that Y represents, consists
of a sum of terms of the form

pj1 ,...,jn · x̄mn+1 ∨ · · · ∨ x̄md ∨ x̄i1 ∨ · · · ∨ x̄ik

8.9. SOME EXAMPLES IN SPACE 111

and upon careful scrutiny of these terms we find that each is zero, but only
because the flat represented by Y has a zero Plücker coordinate correspond-
ing to each of the basis elements xj1 ∧ · · · ∧ xjn such that {i1 , . . . , ik } ⊂
{j1 , . . . , jn }. The terms such that {j1 , . . . , jn } excludes at least one element
of {i1 , . . . , ik } have each excluded i appearing as an m, while the terms
such that {i1 , . . . , ik } ⊂ {j1 , . . . , jn } have no i appearing as an m. Hence
η ∨ ξ = 0 if and only if the flat represented by Y has a zero Plücker coor-
dinate corresponding to each of the basis elements xj1 ∧ · · · ∧ xjn such that
{i1 , . . . , ik } ⊂ {j1 , . . . , jn }. But η ∨ ξ = 0 if and only if the dimension of
X + Y is strictly less than d. The dimension of X is d − k and the dimension
of Y is n so that by Grassmann’s Relation (Corollary 37)

(d − k) + n − dim (X ∩ Y) = dim (X + Y) < d (∗)

and hence
n − k < dim (X ∩ Y) .
The flat represented by X ∩ Y thus has dimension at least n − k.
Suppose on the other hand that X and Y intersect in a flat of dimension
at least n − k. Then reversing the steps above, we recover Equation (∗) so
that η ∨ ξ = 0 which can hold only if the flat represented by Y has a zero
Plücker coordinate corresponding to each of the basis elements xj1 ∧ · · · ∧ xjn
such that {i1 , . . . , ik } ⊂ {j1 , . . . , jn }. 

Exercise 8.8.2 From Grassmann’s Relation alone, show that the subspaces
X and Y of the proposition above always intersect in a subspace of dimension
at least n − k. Hence, without requiring any zero Plücker coordinates, we get
the result that the flats represented by X and Y must intersect in a flat of
dimension at least n − k − 1.

8.9 Some Examples in Space

Let our flat of interest be Φ = F 3 for some field F so that Φ+ is a vector space
of dimension d = 4. The basis B will be {x1 , x2 , x3 , x4 } where {x2 , x3 , x4 }
is a basis for Φ0 and Φ1 = x1 + Φ0 is Φ in its inflated context. In the
following table we show how the elements of B ∧ and B >∧ are connected by
the isomorphism H, where we have only shown the subscript sequences (with
112 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

∅ indicating the empty subscript sequence corresponding to the scalar 1).

∅ → +1234 12 → +34 123 → +4

13 → −24 124 → −3
1 → +234 14 → +23 134 → +2
2 → −134 23 → +14 234 → −1
3 → +124 24 → −13
4 → −123 34 → +12 1234 → +∅

We will generally use subscript sequences to describe the blades made up of

basis vectors. Thus 3 · 143 will indicate either 3 · x1 ∧ x4 ∧ x3 or 3 · x> > >
1 ∧ x4 ∧ x3
as dictated by context. In progressive products we will generally omit the
∧ sign, and in regressive products we will generally replace the ∨ sign with
a “.”, so that (1 + 2)(1 + 3) = 13 + 23 and 234.123 = H −1 (−14) = −23.
We indicate the plane blade obtained by omitting any individual factor from
1234 by using the subscript of the omitted factor with a bar over it as in
234 = 1, 134 = 2, etc. The elements of B ∧ may be expressed as regressive
products of such plane blades as follows, where the empty regressive product
of such plane blades is indicated by ∅.

1234 = ∅ 34 = −1.2 4 = −1.2.3

24 = −1.3 3 = −1.2.4
234 = 1 23 = −1.4 2 = −1.3.4
134 = 2 14 = −2.3 1 = −2.3.4
124 = 3 13 = −2.4
123 = 4 12 = −3.4 ∅ = 1.2.3.4

We shall often find it convenient to engage in a harmless abuse of language

by referring to blades as the flats they represent.
We now present our first example. Consider the plane

(1 + 2) (1 + 3) (1 + 4) = 134 + 214 + 231 + 234 = 2 − 3 + 4 + 1.

We wish to compute its intersection with the plane 2, which is then the line


2 − 3 + 4 + 1 .2 = −3.2 + 4.2 + 1.2 = −14 + 13 − 34.

We now have an answer in the form of a line blade, but we may wish also
to know at least two points that determine the line. A point on a line can
be obtained by taking its regressive product with any coordinate plane that
8.9. SOME EXAMPLES IN SPACE 113

does not contain it. The result of taking the regressive product of our line
with each coordinate plane is as follows.


−3 + 4 + 1
.2.1 = −3.2.1 + 4.2.1 = 4 − 3
−3 + 4 + 1
.2.2 = 0
−3 + 4 + 1
.2.3 = 4.2.3 + 1.2.3 = −1 − 4
−3 + 4 + 1 .2.4 = −3.2.4 + 1.2.4 = −1 − 3.

Thus the line pierces 1 (the plane at ∞) at 4 − 3, it pierces 3 at −1 − 4, and

it pierces 4 at −1 − 3, but it is wholly contained in 2.
Let us now consider two lines in space. Let one be λ = (1) (1+2+3+4) and
the other be µ = (1 + 2) (1 + 3 + 4) . They intersect because their progressive
product is

(12 + 13 + 14) (13 + 14 + 21 + 23 + 24) = 1324 + 1423 = 0.

We can compute their affine sum by considering progressive products of one

with the individual progressive point factors of the other. Multiplying 1 and
µ gives 123+124 = 4+3, the affine sum we seek. Multiplying 1+2+3+4 and
µ also gives 123+124. The regressive product of the two lines is 0, confirming
that their affine sum fails to be all of space. We can compute their intersection
point by considering regressive products of one with the individual regressive
plane factors of the other. We factor λ into plane blades:




λ = (12 + 13 + 14) = − 3.4 + 2.4 + 2.3 = − 3 + 2 . 3 + 4 .

µ may be written as 2.4+2.3−3.4+1.4+1.3 and taking its regressive product

with 3 + 2 gives

2.4.3 + 1.4.3 − 3.4.2 + 1.4.2 + 1.3.2 = 2 · 1 + 2 + 3 + 4,

whereas taking its regressive product with 3 + 4 gives

2.4.3 + 1.4.3 + 2.3.4 + 1.3.4 = 0.

µ is therefore contained in the plane 3 + 4 (as we already know since we

found above that this same plane is the affine sum of λ and µ), but pierces
the plane 3 + 2 at the point 2 · 1 + 2 + 3 + 4 which must also be the point of
intersection of the two lines.
114 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

Exercise 8.9.1 Find the intersection and the affine sum of the two lines
(1̄) .(1̄ + 2̄ + 3̄ + 4̄) and (1̄ + 2̄) . (1̄ + 3̄ + 4̄).
Exercise 8.9.2 Suppose that the regressive product of two blades equals a
nonzero scalar. What is the meaning of this both for the corresponding sub-
spaces of Φ+ and for the corresponding flats of Φ? Give an example where
the blades both are lines in space.

8.10 Factoring a Blade

As illustrated by the examples of the previous section, there are instances
where we have a blade expressed in expanded form and we instead want
to express it, up to a nonzero scalar factor, as the exterior product of vec-
tors. That is, we seek an independent set of vectors that span the subspace
represented by the blade. Such factorizations can be found lurking among
the Plücker coordinates, almost in plain sight, it turns out. A method for
revealing these factors will now be presented. We start by introducing the
extended coordinate array of a blade based on a given set of Plücker co-
ordinates for it. This is a full skew-symmetric array of values Pj1 ,...,jn which
contains the given set of Plücker coordinates along with either ± duplicates
of them or zeroes. This d × · · · × d (n factors of d) array is the generaliza-
tion of a skew-symmetric d × d matrix. Given the Plücker coordinate pi1 ,...,in
related to the basis n-blade xi1 ∧ · · · ∧ xin , the values Pj1 ,...,jn at the array
positions that correspond to this coordinate (the positions with subscripts
that are the permutations of the subscripts of the given coordinate) are given
by
Pσ(i1 ),...,σ(in ) = (−1)σ pi1 ,...,in
for each permutation σ of i1 , . . . , in . Entries in positions that have any equal
subscripts are set to 0. The d-tuple of array values obtained by varying a
given subscript in order from 1 to d, leaving all other subscripts fixed at
chosen constants, is called a file, generalizing the concept of a row or column
of a matrix. Any entry in the array is contained in exactly n files. Here now
is the factorization result.
Proposition 93 (Blade Factorization) Given a set of Plücker coordinates
for an n-blade, the n files that contain a given nonzero entry of the corre-
sponding extended coordinate array constitute a factorization when we regard
these files as vector coordinate d-tuples.
8.10. FACTORING A BLADE 115

Proof: We may assume that the given Plücker coordinates were obtained
as minor determinants in the manner described in Section 8.4 above. Using
the notation of that section, each extended coordinate array value may be
written in determinant form as
 
 aj ,1 · · · aj ,n 
 1 1 
 .. .
.
Pj1 ,...,jn =  .  .

 · · · . 
 ajn ,1 · · · ajn ,n 

We assume without loss of generality that the given nonzero entry is P1,...,n .
We will first show that each of the n files containing P1,...,n is in the span of
the columns Aj of the matrix A = [ai,j ], so that each is a coordinate d-tuple
of a vector in the subspace represented by the blade with the given Plücker
coordinates. Let the files containing P1,...,n form the columns Bj of the d × n
matrix B = [bi,j ] so that the file forming the column Bj has the elements

···
 
 a1,1 a1,n 
 .. .. 
. .
 
 
aj−1,1 ··· aj−1,n
 
  Xn
bi,j = P1,...,j−1,i,j+1,...,n = ai,1 ··· ai,n = cj,k ai,k ,
 
aj+1,1 ··· aj+1,n
 
  k=1
 .. .. 

 . . 

 an,1 ··· an,n 

where the last expression follows upon expanding the determinant about row
j containing the ai,k . Note that the coefficients cj,k are independent of i, and
hence
Xn
Bj = cj,k Ak
k=1

showing that the file forming the column Bj is in the span of the columns of
A.
To verify the independence of the vectors for which the Bj are coordinate
d-tuples, let us examine the matrix B. Observe that the top n rows of B are
P1,...,n times the n × n identity matrix. Hence B contains an n × n minor with
a nonzero determinant, and the vectors wj for which the Bj are coordinate
d-tuples must therefore be independent, since it follows from what we found
in Section 8.4 that w1 ∧ · · · ∧ wn is not zero. 
116 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

Exercise 8.10.1 For each nonzero extended coordinate, apply the method
of the proposition above to the Plücker coordinates of the line obtained in
the first example of the previous section. Characterize the sets of nonzero
extended coordinates that are sufficient to use to yield all essentially different
obtainable factorizations. Compare your results to the points that in the
example were obtained by intersecting with coordinate planes.

Exercise 8.10.2 Use the method of the proposition above to factor the line
λ of the second example of the previous section into
V the regressive product of
>
plane blades. Do
V this first by factoring H (λ) in Φ+ and then carrying the
result back to
V Φ+ . Then also do it by factoring λ directly in the exterior
algebra on Φ+ that the ∨ product induces (where the “vectors” are the
plane blades).

8.11 Algorithm for Sum or Intersection

Given the blades α and β, α ∧ β only represents their affine sum when the
flats represented have an empty intersection. Otherwise the non-blade 0 is
the result. Dually, α ∨ β only represents their intersection when the flats
represented have a full affine sum (all of Φ, that is). Otherwise the non-
blade 0 is again the result. Blades that yield a 0 result when multiplied
contain redundancy and pose for us the extra challenge of coping with that
redundancy. As we have just seen in the previous section, a factorization
for a blade is always readily obtainable, and we can exploit this to construct
a resultant blade that omits redundant factors. We give an algorithm for
doing this in the case where the affine sum is sought, and the same algorithm
can be used on H (α) and H (β) (or the similar ∨-based algorithm on α and
β expressed in terms of basic hyperplane blades) when the intersection is
sought. Suppose that β has been factored as

β = v1 ∧ · · · ∧ vn

and start off with γ0 = α. Then calculate successively γ1 , · · · , γn according

to 
γi−1 ∧ vi if this is 6= 0,
γi =
γi−1 otherwise.
Then γn is the desired blade that represents the affine sum. This algorithm
is the obvious one based on omitting vectors that can already be expressed in
8.12. ADAPTED COORDINATES AND ∨ PRODUCTS 117

terms of the previously included ones that couldn’t be so expressed. Which

vectors are omitted, but not their number, depends on their labeling, of
course. There are similarities here to the problems of finding the least com-
mon multiple and greatest common divisor of positive whole numbers, in-
cluding that an algorithm seems to be required.

8.12 Adapted Coordinates and ∨ Products

We continue to employ the fixed underlying V basis BV= {x
>
1 , . . . , xd } for the
purposes of defining the isomorphism H : Φ+ → Φ+ and the resulting
regressive product that it produces. However, simultaneously using other
bases can sometimes prove useful, particularly in the case where we are deal-
ing with a subspace of Φ+ and we use coordinates adapted to a basis for that
subspace. We will find that doing this will lead to a useful new view of the
regressive product.
We start by considering finding coordinates for an n-blade γ that repre-
sents a subspace W that is itself a subspace of the subspace U of Φ+ . We
want these coordinates adapted to the basis {u1 , . . . , ul } for the l-dimensional
subspace U. The way we will describe these coordinates involves some new
notation.
Let I = {i1 , . . . , in } ⊂ {1, . . . , l} and let I = {in+1 , . . . , il } denote its
complement in {1, . . . , l}. Note that this implies that ij = ik only if j = k.
Note also that the elements of both I and I are shown as having subscript
labels that imply choices have been made of an order for each in which their
elements may be presented by following subscript numerical order. Thus
we will treat I and I as labeled sets that split {1, . . . , l} into two pieces
(one of which could be empty, however). The order choices i1 < · · · < in
and in+1 < · · · < il could be specified if one wishes. However, it will only
be important that the relevant n-element subsets and their complements be
given labelings that remain fixed throughout a particular discussion. The
particular labelings will be unimportant. Based on this foundation, we now
describe our new notation. We will denote the two permutations
   
1 ··· l 1 ··· l − n l − n + 1 ··· l
and
i1 · · · il in+1 · · · il i1 · · · in

by II and II respectively. Also we will denote ui1 ∧ · · · ∧ uin by uI , and

uin+1 ∧ · · · ∧ uil by uI .
118 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

Since γ is an n-blade that represents a subspace of the vector space U

that has the basis {u1 , . . . , ul }, we can expand γ in terms of the coordinates
(−1)II aI adapted to {u1 , . . . , ul } as
X
γ= (−1)II aI · uI
I

where the sum is over all subsets I ⊂ {1, . . . , l} such that |I| = n, where |I|
denotes the number of elements of the set I. Each aI is determined by

aI · u 1 ∧ · · · ∧ u l = u I ∧ γ

since if J is a particular one of the Is in the sum above for γ,

X
uJ ∧ γ = (−1)II aI · uJ ∧ uI = (−1)JJ aJ · uJ ∧ uJ = aJ · u1 ∧ · · · ∧ ul .
I

A similar result is obtained by using (−1)II instead of (−1)II in the expansion

for γ.
We are now ready to begin to apply this to the intersection of a pair
of subspaces in the case where that intersection corresponds to a nonzero
regressive product. Thus let U = h{u1 , . . . , ul }i and V = h{v1 , . . . , vm }i be
subspaces of Φ+ of dimension l and m respectively. Let W = U ∩ V and
let U + V = Φ+ . Then by Grassmann’s Relation (Corollary 37), W has
dimension n = l + m − d (so that, of course, l + m > d since n > 0). Let γ
be a blade that represents W. Then the n-blade γ has the expansions
X X
γ= (−1)II aI · uI = (−1)II bI · vI
I⊂{1,...,l} I⊂{1,...,m}

where the one adapted to the vs is intentionally set up using (−1)II rather
than (−1)II , and the sums are over only those subsets I such that |I| = n.
The aI and bI are determined by

aI · u1 ∧ · · · ∧ ul = uI ∧ γ and bI · v1 ∧ · · · ∧ vm = γ ∧ vI .

We can get some useful expressions for the aI and bI by first creating some
blades that represent U and V and contain γ as a factor. To simplify our
writing, denote u1 ∧ · · · ∧ ul by α and v1 ∧ · · · ∧ vm by β. Supposing that
8.12. ADAPTED COORDINATES AND ∨ PRODUCTS 119

γ = w1 ∧ · · · ∧ wn , the Replacement Theorem (Theorem 8) guarantees us

e = uj1 ∧ · · · ∧ ujl−n and βe = vk1 ∧ · · · ∧ vkm−n such that for some
blades α
nonzero scalars a and b we have a · α = α e ∧ γ and b · β = γ ∧ β. e Note
e ∧ γ ∧ β is a d-blade because it is the product of vectors that span
that α e
U + V = Φ+ . We find then that

aI · α ∧ βe = b · uI ∧ β and bI · α
e ∧ β = a · α ∧ vI

and since
a · α ∧ βe = α
e ∧ γ ∧ βe = b · α
e∧β
we then have
aI bI
·α
e ∧ γ ∧ βe = uI ∧ β and ·α
e ∧ γ ∧ βe = α ∧ vI .
ab ab

e ∧ γ ∧ β,
α e u ∧ β, and α ∧ v are each the progressive product of d vectors and
I I
therefore are scalar multiples of the progressive product of all the vectors of
any basis for Φ+ such as the fixed underlying basis B V = {x1 , . . V
. , xd } that we
use for the purposes of defining the isomorphism H : Φ+ → Φ> + and the
resulting regressive product that it produces. We might as well also then use
B for extracting scalar multipliers in the above equations. We therefore define
[ ξ ], the bracket of the progressive product ξ of d vectors, via the equation
ξ =[ ξ ]·x1 ∧ · · · ∧ xd . [ξ ] is zero when ξ = 0 and is nonzero otherwise (when
ξ is a d-blade). Thus we have

ab ab
aI = · [ uI ∧ β ] and bI = · [ α ∧ vI ] .
e ∧ γ ∧ βe ]
[α e ∧ γ ∧ βe ]
[α

Dropping the factor that is independent of I, we find that we have the two
equal expressions
X X
(−1)II [ uI ∧ β ] · uI = (−1)II [ α ∧ vI ] · vI (∗∗)
I⊂{1,...,l} I⊂{1,...,m}
|I|=n |I|=n

for an n-blade that represents U ∩V and which therefore must be proportional

to α ∨ β. The above development of (∗∗) follows Garry Helzer’s online class
notes for Math 431: Geometry for Computer Graphics at the University of
Maryland (see www.math.umd.edu/˜gah/).
120 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

We note that, in agreement with what α ∨ β gives, when the subspaces

represented by α and β do not sum to all of Φ+ , the two expressions of (∗∗)
both give 0. This is because the totality of the vectors that make up α and β
then do not span Φ+ and therefore neither do any d of them. Let us denote
these equal expressions by µ(α, β) in any case where the sum of the degrees
of the blades α and β is at least d.
We know that µ(α, β) = c · α ∨ β but it remains to be seen how c depends
on α and β. Let us see what happens in a simple example where α and β
are made up of xs drawn from the underlying basis B. Let

α = xj 1 ∧ · · · ∧ xj l and β = xj1 ∧ · · · ∧ xjn ∧ xjl+1 ∧ · · · ∧ xjd

where {xj1 , . . . , xjd } = B. Using the right-hand expression of (∗∗), we find

that it gives the single nonzero term (−1)σ · xj1 ∧ · · · ∧ xjn for µ(α, β), where
 
1 ··· d
σ= .
j1 · · · jd

We also find that

H (α) = (−1)σ · x> >

jl+1 ∧ · · · ∧ xjd and H (β) = (−1)ρ · x> >
jn+1 ∧ · · · ∧ xjl

where
 
1 ··· n n+1 ··· n+d−l n+d−l+1 ··· d
ρ=
j1 ··· jn jl+1 ··· jd jn+1 ··· jl

so that

H (α) ∧ H (β) = (−1)σ (−1)ρ · x> > > >

jl+1 ∧ · · · ∧ xjd ∧ xjn+1 ∧ · · · ∧ xjl

= (−1)σ · H (xj1 ∧ · · · ∧ xjn )

and therefore α ∨ β = (−1)σ · xj1 ∧ · · · ∧ xjn = µ(α, β). This example leads
us to conjecture that it is always the case that µ(α, β) = α ∨ β, a result that
we will soon prove.

Exercise 8.12.1 Verify the results of the example above in the case where
d = 4, u1 = x3 , u2 = x2 , u3 = x1 , and v1 = x3 , v2 = x2 , v3 = x4 . (In
evaluating the second expression of (∗∗), be careful to base I and I on the
subscripts of the vs, not on the subscripts of the xs, and notice that σ comes
from the bracket.)
8.12. ADAPTED COORDINATES AND ∨ PRODUCTS 121
V V
While ∨ is a bilinear function defined on all of Φ+ × Φ+ , µ is only
defined for pairs of blades that have degreesVthat sum Vto at least d. Extending
µ as a bilinear function defined on all of Φ+ × Φ+ will allow us prove
equality with ∨ by considering only what happens
V Vto the elements
V of B ∧ ×B ∧ .
We thus define the bilinear function µ e : Φ+ × Φ+ → Φ+ by defining
it on B ∧ × B ∧ as

µ (xJ , xK ) if |J| + |K| > d,
µ
e (xJ , xK ) =
0 otherwise.

where J, K ⊂ {1, . . . , d}. We now show that when α and β are blades that
have degrees that sum to at least d, µ e (α, β) = µ (α, β), so that µ
e extends µ.
Suppose that
X X
α= aJ · x J and β= bK · x K .
J K

Then, since the bracket is clearly linear, using in turn each of the expressions
of (∗∗) which define µ, we find that
X X X
µ(α, β) = (−1)II [ uI ∧ bK · x K ] · u I = bK · µ(α, xK ) =
I K K
X X
= bK · aJ · µ(xJ , xK ) = µ
e (α, β)
K J

and µ
e indeed extends µ. We now are ready to formally state and prove the
conjectured result.
V
Theorem 94 For any η, ζ ∈ Φ+ , µ e(η, ζ) = η ∨ ζ. Hence for blades α =
u1 ∧ · · · ∧ ul and β = v1 ∧ · · · ∧ vm , of respective degrees l and m such that
l + m > d, we have
X X
α ∨β = (−1)II [ uI ∧ β ]·uI = (−1)II [ α ∧ vI ]·vI (∗ ∨ ∗)
I⊂{1,...,l} I⊂{1,...,m}
|I|=n |I|=n

where n = l + m − d.

Proof: We need only verify that µ e(η, ζ) = η ∨ ζ for (η, ζ) ∈ B ∧ × B ∧ .

Thus we assume that (η, ζ) ∈ B∧ × B ∧ . Since µ
e(η, ζ) = η ∨ ζ = 0 when the
122 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT

subspaces represented by η and ζ do not sum to Φ+ (including the case when

the degrees of η and ζ sum to less than d), we assume that the subspaces
represented by η and ζ do sum to Φ+ (and therefore that the degrees of η
and ζ sum to at least d). The cases where η = α and ζ = β where α and β
are given as

α = xj 1 ∧ · · · ∧ xj l and β = xj1 ∧ · · · ∧ xjn ∧ xjl+1 ∧ · · · ∧ xjd

have been verified in the example above. The forms of α and β are suffi-
ciently general that every case we need to verify is obtainable by separately
permuting factors inside each of α and β. This merely results in η = (−1)τ ·α
and ζ = (−1)υ · β so that η ∨ ζ = (−1)τ (−1)υ · α ∨ β = µ
e(η, ζ) by bilinearity.


The formulas (∗ ∨ ∗) of the theorem above express α ∨ β in all the non-

trivial cases, i. e., when the respective degrees l and m of the blades α and
β sum to at least d, and they do it in a coordinate-free manner. This is not
to say that the idea of coordinates (or bases) is missing from these formulas
or their derivation. But these formulas do not involve a priori coordinates
for α and β, or their factors, in terms of some overall basis. Coordinates
appear in these formulas as outputs, not as inputs. We have defined the
bracket in terms of the overall basis B (the same one used to define H),
and the bracket is a coordinate function. The derivation of these formulas
proceeded by adapting coordinates to the blade factors, and then bringing
in the bracket to express the adapted coordinates as functions of the blade
factors. The result is a formulation of the regressive product where only the
blade factors appear explicitly. These formulas are known to be important in
symbolic algebraic computations, as opposed to the numeric computations
that we have previously carried out using a priori coordinates based on B.
They also now render the following result readily apparent.

Corollary 95 Any two regressive products are proportional.

Proof: The bracket depends on the basis B only up to a nonzero factor.


8.13. PROBLEMS 123

8.13 Problems

1. Theorem 87 and its proof require that P be distinct from the other
points. Broaden the theorem by giving a similar proof for the case where P
coincides with A.
2. What is the effect on H of a change in the assignment of the subscript
labels 1, . . . , d to the vectors of B?
3. Determine the proportionality factor between two regressive products
as the determinant of a vector space map.

4. Notating as in Section 8.9, 12+34 is not a blade in the exterior algebra

of a 4-dimensional vector space.

5. Referring to Theorem 90, for arbitrary vectors v1 , . . . , vn of Φ+ , n 6 d,

 >
 x1 (v1 ) · · · x> x> ··· x>

1 (vn ) 1 (·1 ) 1 (·d−n ) 

 .
.. .
.. .
.. ..
.
 
 
 >
 x (v ) · · · x> n (vn ) x>
n (·1 ) ··· x> (· )

H (v1 ∧ · · · ∧ vn ) =  >n 1 n d−n 
> > >
 xn+1 (v1 ) · · · xn+1 (vn ) xn+1 (·1 ) · · · xn+1 (·d−n ) 

 .. .. .. .. 

 . . . . 

 x> (v ) · · · x> (v ) >
x (· ) · · · x> (· ) 
d 1 d n d 1 d d−n

where x> > >

i (·k ) and xj (·k ) denote the unevaluated functionals xi and xj ,
>

each of which must be evaluated at the same k-th vector of some sequence
of vectors. The resulting determinant is then a function of x> >
1 , · · · , xd each
potentiallyVevaluated once at each of d-n arguments, which we identify with
a blade in Φ> + via the isomorphism of Theorem 71.
124 CHAPTER 8. AN ENHANCED AFFINE ENVIRONMENT
Chapter 9

Vector Projective Geometry

9.1 The Projective Structure on V

Recall that in Chapter 6 we introduced the set V (V) of all subspaces of the
vector space V, defined the affine structure on V as A (V) = V + V (V), and
interpreted the result as an affine geometry based on V. We now take just
V (V) by itself, call it the projective structure on V, and interpret it as a
projective geometry based on V. The projective flats are thus the subspaces
of V. The points are the one-dimensional subspaces of V, the lines are
the two-dimensional subspaces of V, the planes are the three-dimensional
subspaces of V, the (n − 1)-planes are the n-dimensional subspaces of V,
and the hyperplanes are the subspaces of V of codimension one. When
dim V = 2, the projective geometry based on V is called a projective line,
when dim V = 3, it is called a projective plane, and in the general finite-
dimensional case when dim V = d, it is called a projective (d − 1)-space.
The n-dimensional subspaces of V are said to have projective dimension
n − 1. Thus we introduce the function pdim defined by
pdim (X ) = dim (X ) − 1
to give the projective dimension of the flat or subspace X . The trivial sub-
space 0 = {0} is called the null flat and is characterized by pdim (0) = −1.
Grassmann’s Relation (Corollary 37) may be reread as
pdim X + pdim Y = pdim(X + Y) + pdim(X ∩ Y).
We also refer to a subspace sum as the join of the corresponding projective
flats. We denote by P (V) the set of all the points of the geometry and we put

125
126 CHAPTER 9. VECTOR PROJECTIVE GEOMETRY

P+ (V) = P (V) ∪ 0. All flats are joins of points, with the null flat being the
empty join. If X is a subspace of V, then V (X ) ⊂ V (V) gives us a projective
geometry in its own right, a subgeometry of V (V), with P (X ) ⊂ P (V).
V is conceptually the same as our Φ+ of the previous chapter, except
that it is stripped of all references to the affine flat Φ. No points of P (V) are
special. The directions that were viewed as special “points at infinity” in the
viewpoint of the previous chapter become just ordinary points without any
designated Φ0 for them to inhabit. We will at times find it useful to designate
some hyperplane through 0 in V to serve as a Φ0 and thereby allow V to be
interpreted in a generalized affine manner. However, independent of any
generalized affine interpretation, we have the useful concept of homogeneous
representation of points by vectors which we continue to exploit.

9.2 Projective Frames

Notice: Unless otherwise stated, we assume for the rest of this chapter that
V is a nontrivial vector space of finite dimension d over the field F.

The points X0 , . . . , Xn of P (V) are said to be in general position if

they are distinct and have vector representatives that are as independent as
possible, i. e., either all of them are independent, or, if n > d, every d of
them are independent. The case n = d is of particular interest: any d + 1
points in general position are said to form a projective frame for V (V).

Exercise 9.2.1 Any 3 distinct points of a projective line form a projective

frame for it.

To say that the d + 1 distinct points X0 , . . . , Xd are a projective frame

for V (V) is the same as to say that there are respective representing vectors
x0 , . . . , xd such that {x1 , . . . , xd } is an independent set (and therefore a basis
for V) and x0 = a1 · x1 + · · · + ad · xd , where all of the scalar coefficients ai
are nonzero. Given that {x1 , . . . , xd } is a basis, then with the same nonzero
ai , {a1 · x1 , . . . , ad · xd } is also a basis. Hence the following result.

Proposition 96 Points X0 , . . . , Xd of a projective frame for V (V) can be

represented respectively by vectors x0 , . . . , xd such that {x1 , . . . , xd } is a basis
for V and x0 = x1 + · · · + xd . 
9.3. PAPPUS REVISITED 127

A representation of the points X0 , . . . , Xd of a projective frame by respec-

tive vectors x0 , . . . , xd such that x0 = x1 + · · · + xd is said to be standard-
ized. In such a standardized representation, X0 is called the unit point,
and X1 , . . . , Xd are known by various terms such as base points, vertices
of the simplex of reference, or fundamental points of the coordinate
system. In terms of coordinate d-tuples, any d + 1 points X0 , . . . , Xd in gen-
eral position can thus be assigned the respective coordinate d-tuples ε0 =
(1, 1, . . . , 1) , ε1 = (1, 0, . . . , 0) , ε2 = (0, 1, 0, . . . , 0) , . . . , εd = (0, . . . , 0, 1).
However, up to the unavoidable (and ignorable) nonzero overall scalar mul-
tiplier, there is only one basis for V that permits this assignment.
Proposition 97 The basis {x1 , . . . , xd } of the previous proposition is unique
up to a nonzero overall scalar multiplier.
Proof: Suppose that X1 , . . . , Xd are also represented respectively by the
vectors y1 , . . . , yd which together automatically form a basis for V due to the
general position hypothesis, and X0 is represented by y0 = y1 +· · ·+yd . Then
yi = ai · xi for each i with ai a nonzero scalar. Hence
y0 = y1 + · · · + yd = a1 · x1 + · · · + ad · xd = a0 · (x1 + · · · + xd )
and since {x1 , . . . , xd } is a basis for V we must have ai = a0 for all i > 0.
Hence yi = a0 · xi for all i > 0 which is what was to be shown. 

Designating a standardized representation of the points of a projective

frame thus uniquely specifies a system of homogeneous coordinates on P (V).
Given that homogeneous coordinate d-tuples for the unit point and the base
points are ε0 and ε1 , . . . , εd , coordinate d-tuples of all the points of P (V)
are completely determined in the homogeneous sense (i. e., up to a nonzero
scalar multiplier).

9.3 Pappus Revisited

Employing a projective frame can ease the proof of certain theorems of pro-
jective geometry. The Theorem of Pappus is one such. We previously stated
and proved one of its affine versions as Theorem 82. That proof employed
rather elaborate barycentric calculations. The projective version that we now
state and prove has essentially the same statement in words as Theorem 82
but now refers to points and lines in a projective plane over a suitable field.
128 CHAPTER 9. VECTOR PROJECTIVE GEOMETRY

Theorem 98 (Theorem of Pappus) Let l, l 0 be distinct coplanar lines with

distinct points A, B, C ⊂ l r l 0 and distinct points A 0 , B 0 , C 0 ⊂ l 0 r l. If
BC 0 meets B 0 C in A 00 , AC 0 meets A 0 C in B 00 , and AB 0 meets A 0 B in C 00 ,
then A 00 , B 00 , C 00 are distinct collinear points.

A ''
A

C ''
B ''

A ' B ' C '

Illustrating the Theorem of Pappus

Proof: We choose the points B 0 , A, B, C 0 as our projective frame. These

points are in general position since any three of them always contain one
that is not allowed to be on the line through the other two. We designate
a standardized representation by choosing vectors x1 , x2 , x3 to respectively
represent A, B, C 0 and x0 = x1 + x2 + x3 to represent the unit point B 0 .
AB = h{x1 , x2 }i and C on it can be represented as x1 + c · x2 , or what is
the same, as the coordinate tuple (1, c, 0).
B 0 C 0 = h{x1 + x2 + x3 , x3 }i and A0 on it can be represented by (1, 1, a).
a 6= 0 because letting a = 0 in the representation (1, 1, a) for A0 would put
A0 on AC. This is so because AC = h{x1 , x1 + c · x2 }i contains c · (x1 + x2 ) =
(c − 1) · x1 + (x1 + c · x2 ) and c 6= 0 since C 6= A. Similarly, c 6= 1 because
letting c = 1 in the representation (1, c, 0) of C would put C on A0 B 0 . This is
so because A0 B 0 = h{x1 + x2 + a · x3 , x1 + x2 + x3 }i which contains (1 − a) ·
(x1 + x2 ) = (x1 + x2 + a · x3 ) − a · (x1 + x2 + x3 ) and a 6= 1 since A0 6= B 0 .
9.3. PAPPUS REVISITED 129

BC 0 = h{x2 , x3 }i, B 0 C = h{x1 + x2 + x3 , x1 + c · x2 }i and (1 − c) · x2 +

x3 = (x1 + x2 + x3 ) − (x1 + c · x2 ) lies in both so that A00 is represented by
(0, 1 − c, 1).

C 0 A = h{x1 , x3 }i, A0 C = h{x1 + x2 + a · x3 , x1 + c · x2 }i and (1 − c)·x1 −

ca · x3 = x1 + c · x2 − c · (x1 + x2 + a · x3 ) lies in both so that B 00 is represented
by (1 − c, 0, −ca).

AB 0 = h{x1 , x1 + x2 + x3 }i, A0 B = h{x1 + x2 + a · x3 , x2 }i and x1 + a ·

x2 + a · x3 = (1 − a) · x1 + a · (x1 + x2 + x3 ) = (x1 + x2 + a · x3 ) + (a − 1) · x2
lies in both so that C 00 is represented by (1, a, a).

The dependency relation

a · (0, 1 − c, 1) + (1 − c, 0, −ca) + (c − 1) · (1, a, a) = 0

shows that A00 , B 00 , C 00 are collinear. A00 , B 00 , C 00 are distinct since a 6= 0

and c 6= 1 (shown earlier in this proof) make it impossible for any one of
(0, 1 − c, 1), (1 − c, 0, −ca), (1, a, a) to be a multiple of any other. 
What we have just done in proving the Theorem of Pappus becomes even
more transparent when viewed in the light of a generalized affine interpreta-
tion. Using the same basis as for our chosen standardized representation, and
letting x1 = 1 be our Φ1 just as we did in the previous chapter, we get the
generalized affine interpretation of a rather simple affine theorem. The flat
Φ that we view as having been inflated to Φ1 has point A at its origin, and
the points B and C 0 have been “sent to infinity” in two different directions.
As a result, A00 is at infinity, and the conclusion of the theorem is that B 00 C 00
is parallel to B 0 C, as depicted in the figure below. The proof above is just a
proof of this affine theorem using the inflated flat method. We prove exactly
the same thing when we give an ordinary proof of the affine theorem, as we
now do. In coordinate form, the finite points are then A =
(0, 0), C = (c, 0),
B 0 = (1, 1), A0 = (1, a), C 00 = (a, a), B 00 = 0, ac (c − 1)−1 . Thus

C 00 − B 00 = a, a (1 − c)−1 = a (1 − c)−1 · (1 − c, 1) = a (1 − c)−1 · (B 0 − C)

so that B 00 C 00 is parallel to B 0 C as was to be shown.

130 CHAPTER 9. VECTOR PROJECTIVE GEOMETRY

C '

B ''

C ''
A '
A ''

B '

B
A C

Generalized Affine Interpretation

9.4 Projective Transformations

Besides bases and frames, certain functions play an important rôle in pro-
jective geometry. Let W be an alias of V. Any vector space isomorphism
f : V → W induces a bijection from P (V) to P (W). By a projective
transformation we will mean any such bijection from P (V) to P (W) in-
duced by a vector space isomorphism from V to W. The terms projec-
tivity and homography are also commonly used. For any nonzero scalar
a, f and a · f clearly induce the same projective transformation. On the
other hand, if g : V → W is another vector space isomorphism that in-
duces the same projective transformation as f, then g = a · f for some
nonzero scalar a, as we now show. Let {x1 , . . . , xd } be a basis for V. For
9.4. PROJECTIVE TRANSFORMATIONS 131

each i we must have g (xi ) = ai · f (xi ) for some nonzero scalar ai . Also
g (x1 + · · · + xd ) = a · f (x1 + · · · + xd ) for some nonzero scalar a. But then
a1 · f (x1 ) + · · · + ad · f (xd ) = a · f (x1 ) + · · · + a · f (xd )
and since {f (x1 ) , . . . , f (xd )} is an independent set, ai = a for all i. Hence
g (xi ) = a · f (xi ) for all i and therefore g = a · f . Thus, similar to the
classes of proportional vectors as the representative classes for points, we
have the classes of proportional vector space isomorphisms as the represen-
tative classes for projective transformations. A vector space isomorphism
homogeneously represents a projective transformation in the same fashion
that a vector represents a point. For the record, we now formally state this
result as the following theorem.
Theorem 99 Two isomorphisms between finite-dimensional vector spaces
induce the same projective transformation if and only if these isomorphisms
are proportional. 
It is clear that projective transformations send projective frames to pro-
jective frames. Given an arbitrary projective frame for V (V) and another for
V (W), there is an obvious projective transformation that sends the one to
the other. This projective transformation, in fact, is uniquely determined.
Theorem 100 Let X0 , . . . , Xd and Y0 , . . . , Yd be projective frames for V (V)
and V (W), respectively. Then there is a unique projective transformation
from P (V) to P (W) which for each i sends Xi to Yi .
Proof: Let x1 , . . . , xd be representative vectors corresponding to X1 , . . . , Xd
and such that x0 = x1 +· · ·+xd represents X0 . Similarly, let y1 , . . . , yd be rep-
resentative vectors corresponding to Y1 , . . . , Yd and such that y0 = y1 +· · ·+yd
represents Y0 . Then the vector space map f : V → W that sends xi to yi
for i > 0 induces a projective transformation that for each i sends Xi to Yi .
Suppose that the vector space map g : V → W also induces a projective
transformation that for each i sends Xi to Yi . Then there are nonzero scalars
a0 , . . . , ad such that, for each i, g (xi ) = ai · yi so that then
g (x0 ) = g (x1 + · · · + xd ) = a1 · y1 + · · · + ad · yd = a0 · (y1 + · · · + yd )
and since {y1 , . . . , yd } is an independent set, it must be that all the ai are
equal to a0 . Hence g = a0 · f , so that f and g induce exactly the same
projective transformation. 
132 CHAPTER 9. VECTOR PROJECTIVE GEOMETRY

9.5 Projective Maps

Because P (W) does not contain the null flat, only a one-to-one vector space
map from V to some other vector space W will induce a function from P (V)
to P (W). However, any vector space map from V to W does induce a
function from P+ (V) to P+ (W). By a projective map we will mean any
function from P+ (V) to P+ (W) induced by a vector space map from V to W.
Technically, a projective transformation is not a projective map because it is
from P (V) to P (W), not from P+ (V) to P+ (W). However, each projective
transformation clearly extends to a unique bijective projective map from
P+ (V) to P+ (V).

Exercise 9.5.1 The composite of vector space maps induces the composite of
the separately induced projective maps, and the identity induces the identity.

A vector space map from V to W also induces a function from V (V) to

V (W) and we could just as well have used this as our projective map. A
function is also induced from P (V) r P (K) to P (W), where K is the kernel
of the vector space map, and some authors call this a projective map. These
various concepts of projective map all amount to essentially the same thing
because each has the same origin in terms of classes of vector space maps.
For general vector space maps, just as for the isomorphisms, these classes are
the proportional classes as the following exercise records.

Exercise 9.5.2 Let f, g : V → W be vector space maps from the finite-

dimensional vector space V and suppose that f and g induce the same pro-
jective map. Show that f and g are proportional by considering how they act
on the basis vectors x1 , . . . , xk of a complementary subspace of their common
kernel, and on x1 + · · · + xk .

9.6 Central Projection

In V (V), fix a hyperplane t (the target), and fix a point C (the center)
such that C ∈ / t. Given the point P ∈ P (V) r C, we can form the line CP
through C and P . CP will intersect t in a unique point Q. The assignment of
Q = CP ∩ t to P ∈ P (V) r C is called central projection from the point C
onto the hyperplane t. By also assigning the null flat of V (t) both to C and
to the null flat of V (V), we get a function from P+ (V) to P+ (t) which turns
9.6. CENTRAL PROJECTION 133

out to be a projective map. Also, by restricting the domain to the points of

any hyperplane not containing C, and restricting the codomain to the points
of t, we will see that a projective transformation between hyperplanes results,
one that derives from the vector space self-map of projection onto t along C
(Section 2.4).

C Q

t P

C P

w
x

Projecting P from C onto t

The figure above, not the usual schematic but rather a full depiction
of the actual subspaces in V, illustrates central projection in the simple case
where V is of dimension 3 and the target t is a projective line (a 2-dimensional
subspace of V). Let P be represented by the vector v. Then we can (uniquely)
express v as v = x + w where x ∈ C and w ∈ t since V = C ⊕ t. But
w = v − x ∈ CP , so w ∈ CP ∩ t and therefore w represents the point Q that
134 CHAPTER 9. VECTOR PROJECTIVE GEOMETRY

we seek. Thus a vector representing Q may always be obtained by applying

the projection onto t along C to any vector that represents P , thereby making
the assignment of Q to P a projective map. C is the kernel of the projection
onto t along C, so the points of any projective line s that does not contain
C will be placed in one-to-one correspondence with the points of t. Thus the
restriction of the projection onto t along C to a projective line s that does
not contain C is a vector space isomorphism from s onto t, and the induced
map from the points of s onto the points of t is a projective transformation,
an example of a perspectivity between hyperplanes.
Although we have only illustrated here the case where V has dimension 3,
the same considerations clearly apply in any V of finite dimension d. Also, as
detailed in the following exercise, we may readily generalize to where center
and target are any complementary subspaces in V and obtain a concept of
central projection from a center of any positive dimension n to a target of
complementary dimension d − n.

Exercise 9.6.1 Let C be a subspace of dimension n > 0 in the vector space

V of dimension d, and let T be a complement of C. Let P be a 1-dimensional
subspace of V such that P ∩ C = 0. Then the join CP of C and P intersects
T in a 1-dimensional subspace Q. If v is a nonzero vector in P, then the
projection of v onto T along C is a nonzero vector in Q.

9.7 Problems

1. Let f and g be isomorphisms between possibly infinite-dimensional

vector spaces V and W, and let f and g be such that for all vectors v

g (v) = a (v) · f (v) for a nonzero scalar a (v) .

Then for all vectors u and v, a (u) = a (v). Thus Theorem 99 holds in the
infinite-dimensional case as well.
(Consider separately the cases where u and v represent the same point
in P (V) and when they do not. Write g (u) two different ways in the former
case and write g (u + v) in two different ways in the latter case.)
9.7. PROBLEMS 135

2. Suppose that V is over the finite field of q elements. Then if d > q,

d + 1 distinct points of P (V), but no more, can be in general position. But
if 2 6 d < q, then at least d + 2 points of P (V) can be in general position.
136 CHAPTER 9. VECTOR PROJECTIVE GEOMETRY
Chapter 10

Scalar Product Spaces

10.1 Pairings and Their Included Maps

Let V and W be vector spaces over the same field F. We will refer to a
bilinear functional g : V × W → F as a pairing of V and W. Consider the
pairing g : V × W → F. With the second argument of g held fixed at w,
allowing the first argument to vary produces the values of a linear functional
g w ∈ V > . Similarly, the values of a linear functional gv ∈ W > are produced
when the first argument is held fixed at v and the second argument is allowed
to vary. Now, if we take gv and let v vary, we get a map g1 : V → W > , and
if we take g w and let w vary, we get a map g2 : W → V > . Thus g includes
within itself the two maps g1 and g2 which we will refer to as the included
maps belonging to g. The various gs are related by

gv (w) = (g1 (v))(w) = g(v, w) = (g2 (w))(v) = g w (v).

For any map f : V → W > , (f (v))(w) is the value of a bilinear functional of

v and w, and hence any such f is the g1 included map of some pairing of V
and W. By the same token, any map from W to V > can be a g2 . However,
each included map clearly determines the other, so only one of them may be
specified arbitrarily.

10.2 Nondegeneracy
A pairing g is called nondegenerate if for each nonzero v, there is some
w for which g(v, w) 6= 0, and for each nonzero w, there is some v for which

137
138 CHAPTER 10. SCALAR PRODUCT SPACES

g(v, w) 6= 0. A nondegenerate pairing g : V × W → F is sometimes referred

to as a perfect (or duality) pairing, and is said to put V in duality with
W.

Exercise 10.2.1 For any vector space V over F, the natural evaluation
pairing e : V > × V → F, defined by e(f, v) = f (v) for f ∈ V > and v ∈ V,
puts V > in duality with V. (Each v 6= 0 is part of some basis and therefore
has a coordinate function v > .)

The nice thing about nondegeneracy is that it makes the included maps
one-to-one, so they map different vectors to different functionals. For, if
the distinct vectors s and t of V both map to the same functional (so that
gt = gs ), then gu , where u = t − s 6= 0, is the zero functional on W. Hence
for the nonzero vector u, gu (w) = g(u, w) = 0 for all w ∈ W, and g therefore
fails to be nondegenerate.

Exercise 10.2.2 If the included maps are both one-to-one, the pairing is
nondegenerate. (The one-to-one included maps send only the zero vector to
the zero functional.)

If g is nondegenerate and one of the spaces is finite-dimensional, then so is

the other. (W is isomorphic to a subspace of V > , so if V is finite-dimensional,
so is W.) Supposing then that a nondegenerate g is a pairing of finite-
dimensional spaces, we have dim W 6 dim V > = dim V and also dim V 6
dim W > = dim W. Hence the spaces all have the same finite dimension and
the included maps are therefore isomorphisms, since we know (Theorem 25)
that a one-to-one map into an alias must also be onto.
The included maps belonging to a nondegenerate pairing g of a finite-
dimensional V with itself (which puts V in duality with itself) are isomor-
phisms between V and V > , each of which allows any functional in V > to be
represented by a vector of V in a basis-independent manner. Any ϕ ∈ V > is
represented by vϕ = g1−1 (ϕ), and also by wϕ = g2−1 (ϕ), so that for any w ∈ V

ϕ(w) = (g1 (g1−1 (ϕ))(w) = g(g1−1 (ϕ), w) = g(vϕ , w)

and for any v ∈ V

g(v, wϕ ) = g(v, g2−1 (ϕ)) = (g2 (g2−1 (ϕ))(v) = ϕ(v).

10.3. ORTHOGONALITY, SCALAR PRODUCTS 139

10.3 Orthogonality, Scalar Products

Designating an appropriate pairing g of V with itself will give us the means
to define orthogonality on a vector space by saying that the vectors u and v
are orthogonal (written u⊥v) if g(u, v) = 0. Requiring nondegeneracy will
insure that no nonzero vector is orthogonal to all nonzero vectors. However,
we also want to rule out the undesirable possibility that g(u, v) = 0 while
g(v, u) 6= 0. Thus we will also insist that g be reflexive: g(u, v) = 0 if
and only if g(v, u) = 0. g will be reflexive if it is symmetric (g(u, v) =
g(v, u) for all u, v), or if it is alternating (g(v, v) = 0 for all v, which implies
g(u, v) = −g(v, u) for all u, v). Here we will be confining our attention to
the case where g is symmetric and nondegenerate. By designating a specific
symmetric nondegenerate self-pairing as the scalar product on a vector
space, the vector space with its designated scalar product becomes a scalar
product space. The scalar product g : V × V → F then provides the
orthogonality relation ⊥ on V.

Exercise 10.3.1 A symmetric pairing g of a vector space with itself is non-

degenerate if and only if g(v, w) = 0 for all w implies v = 0.

Notice that the two included maps of a scalar product are identical, and
abusing notation somewhat, we will use the same letter g to refer both to
the included map and to the scalar product itself. g : V × V → F and
g : V → V > will not be readily confused. When V is finite-dimensional, the
scalar product’s included map g is an isomorphism that allows each ϕ ∈ V >
to be represented by the vector g −1 (ϕ) via

ϕ(w) = g(g −1 (ϕ), w),

so that the evaluation of the functional ϕ at any vector w can be replaced

by the evaluation of the scalar product at (g −1 (ϕ), w). In slightly different
notation, with e being the evaluation pairing introduced in Exercise 10.2.1
above, the formula reads

e(ϕ, w) = g(g −1 (ϕ), w),

which shows how the two seemingly different pairings, e and g, are essen-
tially the same when V is a finite-dimensional scalar product space. The
natural nondegenerate pairing e between V > and V thus takes on various
140 CHAPTER 10. SCALAR PRODUCT SPACES

interpretations on V depending on the choice of scalar product on the finite-

dimensional space. For instance, given a particular vector w, the functionals
ϕ such that ϕ(w) = 0 determine by v = g −1 (ϕ) the vectors v for which v⊥w.
Another example is the interpretation of the dual of a basis for V as another
basis for V.

10.4 Reciprocal Basis

Notice: We assume for the remainder of the chapter that we are treating a
scalar product space V of finite dimension d over the field F, and that V has
the scalar product g.
Let B = {x1 , . . . , xd } be a basis for V and B > = {x> >
1 , . . . , xd } its dual on
V > . For each i, let x⊥ −1 > ⊥ ⊥ ⊥
i = g (xi ). Then B = {x1 , . . . , xd } is another basis
for V, known as the reciprocal of B. Acting through the scalar product,
the elements of B ⊥ behave the same as the coordinate functions from which
they derive via g −1 and therefore satisfy the biorthogonality conditions

⊥ > 1, if i = j,
g(xi , xj ) = xi (xj ) = δi,j =
0, if i 6= j,

for all i and j. On the other hand, if some basis {y1 , . . . , yd } satisfies
g(yi , xj ) = (g(yi ))(xj ) = δi,j for all i and j, then g(yi ) must by definition
be the coordinate function x> −1 > ⊥
i , which then makes yi = g (xi ) = xi . We
therefore conclude that biorthogonality characterizes the reciprocal. Hence,
if we replace B with B ⊥ , the original biorthogonality formula displayed above

⊥
tells us that B ⊥ = B.
For each element xj of the above basis B, g(xj ) ∈ V > has a (unique) dual
basis representation of the form
d
X
g(xj ) = gi,j · x>
i .
i=1

Applying this representation to xk yields gk,j = g(xj , xk ), so we have the

explicit formula
d
X
g(xj ) = g(xi , xj ) · x>
i .
i=1
10.5. RELATED SCALAR PRODUCT FOR DUAL SPACE 141

Applying g −1 to this result, we get the following formula that expresses each
vector of B in terms of the vectors of B ⊥ as
d
X
xj = g(xi , xj ) · x⊥
i .
i=1

Each element of B ⊥ thus is given in terms of the elements of B by

d
X
x⊥
j = g i,j · xi ,
i=1

where g i,j is the element in the ith row and jth column of the inverse of
 
g(x1 , x1 ) · · · g(x1 , xd )
[g(xi , xj )] = [gi,j ] =  .. ..
.
 
. ··· .
g(xd , x1 ) · · · g(xd , xd )

This we may verify by noting that since the g i,j satisfy

d 
X 1 if i = j,
gi,k g k,j
= δij =
0 otherwise,
k=1

we have
d
X d
X X d X
X d d
X
i,j
g · xi = g i,j ⊥
gk,i · xk = i,j ⊥
gk,i g · xk = δkj · x⊥ ⊥
k = xj .
i=1 i=1 k=1 k=1 i=1 k=1

Exercise 10.4.1 x1 ∧ · · · ∧ xd = G · x⊥ ⊥
1 ∧ · · · ∧ xd where G = det[g(xi , xj )].

10.5 Related Scalar Product for Dual Space

Given the scalar product g on V, we define the related scalar product ge on
V > . This we do in a natural way by mapping the elements of V > back to
V via g −1 , and then declaring the scalar product of the mapped elements to
be the scalar product of the elements themselves. Thus for ϕ, ψ ∈ V > , we
define ge : V > × V > → F by ge(ϕ, ψ) = g(g −1 (ϕ), g −1 (ψ)). It is easy to see
that this ge is indeed a scalar product. Note that g(v, w) = ge(g(v), g(w)), so
142 CHAPTER 10. SCALAR PRODUCT SPACES

the included map g : V → V > is more than just an isomorphism − it also

preserves scalar product. We call such a map an isometry. It makes V and
V > into isometric scalar product spaces which are then scalar product space
aliases, and we have justification to view them as two versions of the same
scalar product space by identifying each v ∈ V with g(v) ∈ V > .
For the elements of the dual basis B > = {x> >
1 , . . . , xd } of the basis B =
{x1 , . . . , xd } for V, we then have

ge(x> > −1 > −1 > ⊥ ⊥

i , xj ) = g g (xi ), g (xj ) = g xi , xj ,

and using the result we found in the previous section, the new included map
ge : V > → V has the (unique) basis representation
d
X
ge(x> g x⊥ ⊥


j ) = i , xj · xi .
i=1

Swapping B and B ⊥ , a formula derived in the previous section becomes the

equally valid formula
X d
⊥
xj = g(x⊥ ⊥
i , xj ) · xi
i=1

which has the same right hand side as the preceding basis representation
formula, so we see that ge(x> ⊥
j ) = xj . The included map ge must therefore be
−1 −1
g . Applying g = ge to both sides of the basis representation formula for
ge(x>j ), we get
Xd
>
g x⊥ ⊥


xj = i , xj · g(xi ),
i=1
or
d
> ⊥
X
x> g x⊥ ⊥



j = i , xj · xi .
i=1

As we already know from the previous section, the inverse relationship is

d

⊥ X
x>
i
−1
= ge (xj ) = g(xj ) = g (xi , xj ) · x>
i .
i=1

Exercise 10.5.1 g x⊥ ⊥
= [g i,j ], i. e., g x⊥ ⊥
= [g (xi , xj )]−1 =

 

i , xj i , xj
[gi,j ]−1 .
10.6. SOME NOTATIONAL CONVENTIONS 143

10.6 Some Notational Conventions

Notational conventions abound in practice, and we note some of them now.

It is not unreasonable to view ge as extending g to V > , and it would not be
unreasonable to just write g for both (remembering that the included map of
the g on V > is the inverse of the included map of the original g on V). Unless
explicitly stated otherwise, we will assume henceforth that the related scalar
product, no tilde required, is being used on V > . In the literature, either
scalar product applied to a pair of vector or functional arguments is often
seen written (v, w) or (ϕ, ψ), entirely omitting g at the front, but we will
avoid this practice here.
Relative to a fixed basis B = {x1 , . . . , xd } for V and its dual for V > , the
matrix [g(xi , xj )] of the included map of the original g may be abbreviated
as [gij ] (without any comma between the separate subscripts i and j), and
the matrix [g(xi , xj )]−1 of the related scalar product on V > may similarly be
abbreviated as [g ij ]. The typical element gij of the matrix [gij ] is often used
to designate the g on V, and similarly g ij to designate the related g on V > .
By the same token, v i designates the vector v = v 1 · x1 + · · · + v d · xd and ϕi
designates the functional ϕ = ϕ1 ·x> >
1 +· · ·+ϕd ·xd . The detailed evaluation of
i j
g(v, w) is usually then written as gij v w which, since i appears twice (only),
as does j, is taken to mean that summation over both i and j has tacitly
been performed over their known ranges of 1 to d. This convention, called the
summation convention P (attributed to Albert Einstein), can save a lot of ink
by omitting many signs. Similarly, the detailed evaluation of the scalar
product g(ϕ, ψ) = ge(ϕ, ψ) of the two functionals ϕ and ψ appears as g ij ϕi ψj .
Other examples are the evaluation of ϕ = g(v) as vi = gij v j (the vector v
is converted to a functional ϕ usually designated vi rather than ϕi ) and the
reverse v = g −1 (ϕ) as v i = g ij vj . Note that vi wi = gij v j wi = g(v)(w) =
g(v, w) and g ij gjk v k = δki v k = v i .
This component-style notation, including the summation convention, is
usually referred to as tensor notation. It has been a very popular notation
in the past, and is still widely used today in those applications where it has
gained a traditional foothold. Some concepts are easier to express using it,
but it can also bury concepts in forests of subscripts and superscripts. The
above samples of tensor notation used in scalar product calculations were
included to show how scalar product concepts are often handled notationally
in the literature, and not because we are going to adopt this notation here.
144 CHAPTER 10. SCALAR PRODUCT SPACES

10.7 Standard Scalar Products

We will now define standard scalar products on the finite-dimensional vector
space V over the field F. As we know, we may determine a bilinear function
by giving its value on each pair (xi , xj ) of basis vectors of whatever basis
{x1 , . . . , xd } we might choose. The bilinear functional g : V × V → F will
be called a standard scalar product if for the chosen basis {x1 , . . . , xd }
we have g(xi , xj ) = 0 whenever i 6= j and each g(xi , xi ) = ηi is either +1
or −1. A standard scalar product will be called definite if the ηi are all
equal, and any other standard scalar product will be called indefinite. The
standard scalar product for which ηi = +1 for all i will be referred to as
the positive standard scalar product or as the usual standard scalar prod-
uct. With the standard basis {(1, 0, . . . , 0), . . . , (0, . . . , 0, 1)} as the chosen
basis, the usual standard scalar product on the real d-dimensional space Rd
is the well-known Euclidean inner product, or dot product. A standard scalar
product with chosen basis {x1 , . . . , xd } always makes xi ⊥xj whenever i 6= j,
so the chosen basis is always an orthogonal basis under a standard scalar
product, and because the scalar product of a chosen basis vector with itself
is ±1, the chosen basis for a standard scalar product is more specifically an
orthonormal basis.

Exercise 10.7.1 A standard scalar product is indeed a scalar product. (For

the nondegeneracy, Exercise 10.3.1 above may be applied using each vector
from the chosen basis as a w.)

Exercise 10.7.2 On real 2-dimensional space R2 , there is a standard scalar

product that makes some nonzero vector orthogonal to itself. However, Rd
with the Euclidean inner product as its scalar product has no nonzero vector
that is orthogonal to itself.

With a standard scalar product, the vectors of the reciprocal of the chosen
basis can be found in terms of those of the chosen basis without computation.

Exercise 10.7.3 The included map g for a standard scalar product satisfies
g(xi ) = ηi · x>
i for each basis vector xi of the chosen basis, and also then
⊥ −1 >
xi = g (xi ) = ηi · xi for each i.

A scalar product that makes a particular chosen basis orthonormal can

always be imposed on a vector space, of course. However, it is not necessarily
10.8. ORTHOGONAL SUBSPACES AND HODGE STAR 145

true that an orthonormal basis exists for each possible scalar product that
can be specified on a vector space. As long as 1 + 1 6= 0 in F, an orthogonal
basis can always be found, but due to a lack of square roots in F, it may be
impossible to find any orthogonal basis that can be normalized. However,
over the real numbers, normalization is always possible, and in fact, over
the reals, the number of g(xi , xi ) that equal −1 is always the same for every
orthonormal basis of a given scalar product space due to the well-known
Sylvester’s Law of Inertia (named for James Joseph Sylvester who published
a proof in 1852).

Exercise 10.7.4 Let F = {0, 1} and let V = F2 . Let x1 and x2 be the
0 1
standard basis vectors and let [g(xi , xj )] = . Then for this g, V has
1 0
no orthogonal basis.

10.8 Orthogonal Subspaces and Hodge Star

Corresponding to each subspace X C V is its orthogonal subspace X ⊥
consisting of all the vectors from V that are orthogonal to every vector of X .
A vector v is thus in X ⊥ if and only if g(v, x) = g(v)(x) = 0 for all x ∈ X
or hence if and only if g(v) is in the annihilator of X (Section 3.3). Thus
X ⊥ = g −1 (X 0 ), where X 0 C V > is the annihilator of X C V, and X ⊥ is
therefore just X 0 being interpreted in V by using the scalar product’s ability
to represent linear functionals as vectors.
Suppose now that we fix a basis BH = {x V1 , . . . , xVd } and use it, as we
did in Theorem 90, in defining the map H : V → V > that maps each
blade to a corresponding annihilatorV blade. Then replacing each x> i in the
> ⊥
V in V by xi changes it into a correspond-
corresponding annihilator blade
ing orthogonal blade in V. The Vprocess of replacing each x> i by xi
⊥

may be viewed as extending g −1 to V > and the resulting isomorphism,

(which, incidentally,
V −1 V >is easily V seen to be independent
V −1 of the basis choice) is
denoted
V gV : V → V. The composite g ◦ H will be denoted by
∗ : V → V and will be called the Hodge star, for Scottish geometer
William Vallance Douglas Hodge (1903 – 1975), although it is essentially
the same as an operator that Grassmann used. Thus the “annihilator blade
map” H is reinterpreted as the “orthogonal blade map” ∗ for blades and the
subspaces they represent. Applying ∗ to exterior products of elements of BH
146 CHAPTER 10. SCALAR PRODUCT SPACES

gives
∗ (xi1 ∧ · · · ∧ xin ) = (−1)ρ · x⊥ ⊥
in+1 ∧ · · · ∧ xid ,

where ρ is the permutation i1 , . . . , id of 1, . . . , d and (−1)ρ = +1 or −1

according as ρ is even or odd. Employing the usual standard scalar product
with BH as its chosen basis, we then have, for example, ∗(x1 ∧ · · · ∧ xn ) =
xn+1 ∧ · · · ∧ xd .

Exercise 10.8.1 In R2 , with BH the standard basis x1 = (1, 0), x2 = (0, 1),
compute ∗((a, b))
 and check the orthogonality
 for each of these scalar products
g(x1 , x1 ) g(x1 , x2 )
g with matrix =
g(x2 , x1 ) g(x2 , x2 )
5 4
       
1 0 1 0 0 1 3 3
a) , b) , c) , d) 4 5 .
0 1 0 −1 1 0 3 3

Exercise 10.8.2 For X C V, dim X ⊥ = dim V − dim X .

Exercise 10.8.3 For X C V, V = X ⊕ X ⊥ if and only if no nonzero vector

of X is orthogonal to itself.

⊥
Exercise 10.8.4 For X C V, X ⊥⊥ = X ⊥ = X since X C X ⊥⊥ and
dim X ⊥⊥ = dim X .

The Hodge star was defined above by the simple formula g −1 ◦ H. This
V
formula could additionally be scaled, as some authors do, in order to meet
some particular normalization criterion, such as making a blade and V its star
−1
in some sense represent the same geometric
p content. For example, g ◦ H
is sometimes scaled up by the factor |det [g(xi , xj )]| when F is the field of
real numbers. However, for the time being at least, F will be kept general,
and the unscaled simple formula will continue to be our definition.

10.9 Scalar Product on Exterior Powers

Vp will now show that the scalar product g on V may be extended to each
We
V as the bilinear function g that on p-blades has the value

g(v1 ∧ · · · ∧ vp , w1 ∧ · · · ∧ wp ) = det [g(vi , wj )] .

10.9. SCALAR PRODUCT ON EXTERIOR POWERS 147
Vp
Thus
Vp we need to show that there is a unique bilinear function g : V×
V → F that satisfies the above specification, and that this g is symmetric
and nondegenerate. (For p = 0, by convention we put g(1, 1) = 1 so that
g(a · 1, b · 1) = ab. The cases p = 0 or 1 need no further attention, so we
assume p > 1 in the following treatment.)
For fixed (v1 , . . . , vp ), det [g(vi , wj )] is an alternating p-linear function
>
of (w1 , . . . , wp ) and by Theorem 66 there is a unique gv1 ,...,vp ∈ ( p V)
V
such that gv1 ,...,vp (w1 ∧ · · · ∧ wp ) = det [g(vi , wj )]. Now the assignment of
gv1 ,...,vp to (v1 , . . . , vp ) is an alternating p-linear function and hence there is
>
a unique map g∧ : p V → ( p V) such that g∧ (v1 ∧ · · · ∧ vp ) = gv1 ,...,vp and
V V
therefore such that g∧ (v1 ∧ · · · ∧ vp )(w1 ∧ · · · ∧ wp ) = det [g(vi , wj )]. Evidently
this Vg∧ is anVincluded map of the uniquely determined bilinear functional
g : p V × p V → F given by g(s, t) = g∧ (s)(t) for any elements s, t of
V p
V.
The symmetry of the g we have just defined on all of p V follows readily
V
from the symmetry of g on V, and the equality of the determinant of a matrix
with that of its transpose.
We will show nondegeneracy by showing that the included map is one-to-
p
one.VLet V have the basis B = {x1 , . . . , xd } from which we make a basis B ∧
for p V in the usual manner by taking exterior products of the elements
p
of each subset of p elements of B. Elements of B ∧ will be denoted xI =
xi1 ∧ · · · ∧ xip , xJ = xj1 ∧ · · · ∧ xjp , etc. using capital-letter subscripts that
denote the multi-index subscript sets I = {i1 , . . . , ip } and J = {j1 , . . . , jp } of
the p elements of B that are involved. Indexing now with these multi-indices,
our familiar explicit dual basis representation formula for the effect of the
p
included map (now also called g) on basis vectors from B ∧ appears as
X
g(xJ ) = g(xI , xJ ) · (xI )> ,
|I|=p

with the scalar product on basis vector pairs being

g(xI , xJ ) = g(xi1 ∧ · · · ∧ xip , xj1 ∧ · · · ∧ xjp ) = det [g(xik , xjl )]

of course.
The concept of the pth exterior power of a map, introduced by Exercise
Vp
5.5.5,
Vp will now
Vp > find employment. Consider the pth exterior power g :
V → V of the included map g of the original scalar product on V.
148 CHAPTER 10. SCALAR PRODUCT SPACES

We have
!  
d
X d
X
Vp
g(xJ ) = g(xj1 )∧· · ·∧g(xjp ) = g(xi1 , xj1 ) · x>
i1 ∧· · ·∧
 g(xip , xjp ) · x>
ip

i1 =1 ip =1

from which it follows that

Vp X
g(xJ ) = det [g(xik , xjl )] · x>
I .
|I|=p

So we see that the new included map has a basis representation

X
g(xJ ) = gI,J · (xI )> ,
|I|=p

and the pth exterior power of the original included map has a basis repre-
sentation X
Vp
g(xJ ) = gI,J · x>
I ,
|I|=p

both with the same coefficients gI,J , namely

gI,J = det [g(xik , xjl )] .

Therefore, applying Proposition 70, if one of these maps is invertible, then

so is the other one. Now, the invertibility of the original included map
g : V → V > makes its pth exteriorVpower invertible as V well, because the
g(xi ) being a basis for V > make the p g(xI ) a basis for p V > . Therefore
the new included map is also invertible and our new symmetric bilinear g
is indeed nondegenerate as we wished to show. What is going on here is
rather transparent: the pth exterior power of the included map g : V → V >
p p >
V > → ( V) of Theorem 71, evidently
V V
followed by the isomorphism: Φ b:
is the included map of the extension. That is, we have the formula g(s, t) =
((Φb ◦ Vp g)(s))(t), valid for all (s, t) ∈ Vp V × Vp V. If we agree to identify
>
( p V) with p V > through the isomorphism b then Vp g “is” the included
V V
Vp Φ,
map of our scalar product extension to V.
The Hodge star provides an isomorphic mapping of p V with d−p V that
V V
is also an isometry (except for scale), as the following proposition records.
10.10. ANOTHER WAY TO DEFINE ∗ 149

Proposition 101 Let s, r ∈ p V, let {x1 , . . . , xd } be the basis for V used

V
in defining ∗, and let G = det [g(xi , xj )]. Then
g(r, s) = Gg(∗r, ∗s).
Proof: It suffices to prove the result for r = xI and s = xJ . Now
 
gi1 ,j1 · · · gi1 ,jp
 .. .. 
g(xI , xJ ) = det  . ··· . 
gip ,j1 ··· gip ,jp
and in light of Exercise 10.5.1
 
g ip+1 ,jp+1 · · · g ip+1 ,jd
ρ σ
g(∗xI , ∗xJ ) = (−1) (−1) det 
 .. .. 
. ··· . 
g id ,jp+1 · · · g id ,jd

where ρ and σ are the respective permutations i1 , . . . , id and j1 , . . . , jd of

{1, . . . , d}. By Jacobi’s Determinant Identity (Lemma 89)
     
gi1 ,j1 · · · gi1 ,jp gi1 ,j1 · · · gi1 ,jd g ip+1 ,jp+1 · · · g ip+1 ,jd
det  ... ..  = det  .. ..  det  .. ..
 
··· .   . ··· .   . ··· . 
id ,jp+1 id ,jd
gip ,j1 · · · gip ,jp gid ,j1 · · · gid ,jd g ··· g

from which it follows that g(xI , xJ ) = Gg(∗xI , ∗xJ ) once we have attributed
the (−1)ρ (−1)σ to the rearranged rows and columns of the big determinant.


10.10 Another Way to Define ∗

A version V of the Hodge star may be defined using the scalar products defined
on the p V in the previous section. Let B = {x1 , . . . , xd } be a designated
Vp for V and consider the exterior product
basis Vd−p s ∧ t where s is a fixed element
of V and t is allowed to vary over V. Each such s ∧ t is of the form
V >
d−p
cs (t) · x1 ∧ · · · ∧ xd where cs ∈ V . Hence there is a unique element
V such that g (∗s, t) = cs (t) = (x1 ∧ · · · ∧ xd )> (s ∧ t) and this
Vd−p
∗s ∈
unique element ∗s for which
s ∧ t = g (∗s, t) · x1 ∧ · · · ∧ xd
150 CHAPTER 10. SCALAR PRODUCT SPACES

will be the new version of the Hodge star on p V that we will now scrutinize
V
and compare with the Vp original version. We will focus on elements s of degree p
(meaning those V in V), but we do intend this ∗ to be applicable to arbitrary
elements of V by applying it separately to the components of each degree
and summing the results.
IfVwe have an element r ∈ d−p V that we allege is our new ∗s for some
V
s ∈ p V, we can verify that by showing V that g (r, t) is the coefficient of
x1 ∧ · · · ∧ xd in s ∧ t for all t in a basis for d−p V. V If we show this for such
an r corresponding to each s that is V in a basis for p V, then we will have
completely determined the new ∗ on p V.
Let us see what happens when we choose the designated basis B to be the
basis BH used in defining the annihilator blade map H, with exactly the same
assignment of subscript labels to the basis vectors xi . With s and t equal to
the respective basis monomials xI = xi1 ∧ · · · ∧ xip and xJ = xjp+1 ∧ · · · ∧ xjd ,
we have
s ∧ t = xI ∧ xJ = εI,J · x1 ∧ · · · ∧ xd
where εI,J = 0 if {jp+1 , . . . , jd } is not complementary to {i1 , . . . , ip } in
{1, . . . , d}, and otherwise εI,J = (−1)σ with σ the permutation i1 , . . . , ip , jp+1 , . . . , jd
of 1, . . . , d. The original Hodge star gives

∗xI = ∗ xi1 ∧ · · · ∧ xip = (−1)ρ · x⊥ ⊥

ip+1 ∧ · · · ∧ xid

where ρ is the permutation i1 , . . . , id of 1, . . . , d, so for this original ∗xI ,

 
g(x⊥ ip+1 , xjp+1 ) · · · g(x ⊥
ip+1 , xjd
)
g (∗xI , xJ ) = (−1)ρ det 
 .. .. 
. ··· . 
⊥ ⊥
g(xi , xjp+1 ) · · · g(xid , xjd )
 > d 
xip+1 (xjp+1 ) · · · x> ip+1 (xjd )
= (−1)ρ det  .. ..
.
 
. ··· .
> >
xid (xjp+1 ) · · · xid (xjd )

If J is not complementary to I in {1, . . . , d} then jp+1 , . . . , jd is not a per-

mutation of ip+1 , . . . , id and some jp+k equals none of ip+1 , . . . , id so that the
kth column is all zeroes and the latter determinant vanishes. On the other
hand, if J is complementary to I in {1, . . . , d} then J = {ip+1 , . . . , id } and
there is no reason why we cannot assume that jp+1 = ip+1 , . . . , jd = id , which
then makes the determinant equal to 1 and makes σ equal to ρ. Hence using
10.10. ANOTHER WAY TO DEFINE ∗ 151

the same basis ordered the same way, our new version of the Hodge star is
exactly the same as the original version.
Exercise 10.10.1 For a given scalar product space, using B 0 = {x01 , . . . , x0d }
instead of B = {x1 , . . . , xd } to define the Hodge star gives h · ∗ instead of ∗,
where h is the nonzero factor, independent of p, such that h · x01 ∧ · · · ∧ x0d =
x1 ∧ · · · ∧ xd . Thus, no matter which definition is used for either, or what
basis
V is used in defining either, for any two of our Hodge stars, over all of
V the values of one are the same constant scalar multiple of the values of
the other. That is, V ignoring scale, for a given scalar product space V all of
our Hodge stars on V are identical, and the scale depends only on the basis
choice (including labeling) for V used in each definition.
Exercise 10.10.2 (Continuation) Suppose that for the scalar product g, the
bases  B 0 and B 0 have the same Gram determinant, i.e., det [g(xi , xj )] =
det g(xi , x0j ) , or equivalently, g(x1 ∧ · · · ∧ xd , x1 ∧ · · · ∧ xd ) = g(x01 ∧ · · · ∧


x0d , x01 ∧ · · · ∧ x0d ). Taking it as a given that in a field, 1 has only itself and
−1 as square roots, B and B 0 then produce the same Hodge star up to sign.
The following result is now apparent.
Proposition 102 Two bases of a given scalar product space yield the same
Hodge star up to sign if and only if they have the same Gram determinant.


Finally, we obtain some more formulas of interest, and some conclusions

based on them. Putting t = ∗r in the defining formula above gives
s ∧ ∗r = g (∗s, ∗r) · x1 ∧ · · · ∧ xd .
Similarly, we find
r ∧ ∗s = g (∗r, ∗s) · x1 ∧ · · · ∧ xd = g (∗s, ∗r) · x1 ∧ · · · ∧ xd ,
and therefore for all r, s ∈ p V we have
V

r ∧ ∗s = s ∧ ∗r.
Applying Proposition 101, we also get
s ∧ ∗r = G−1 g (r, s) · x1 ∧ · · · ∧ xd ,
where G = det [g(xi , xj )]. The results of the next two exercises then follow
readily.
152 CHAPTER 10. SCALAR PRODUCT SPACES
Vp
Exercise 10.10.3 For all r, s ∈ V,
g (r, s) = G · ∗ (r ∧ ∗s) = G · ∗ (s ∧ ∗r) .
Exercise 10.10.4 For any r ∈ p V,
V
X
∗r = G−1 (−1)ρ g(r, xI ) · x I ,
|I|=p

where I = {i1 , . . . , ip }, I = {ip+1 , . . . , id }, I ∪ I = {1, . . . , d}, and

 
1 ··· d
ρ= .
i1 · · · id
Thus, not only do we have a nice expansion formula for ∗r, but we may also
Vd−p
conclude that ∗r has yet another definition as the V unique t ∈ V for
which s ∧ t = G−1 g (r, s) · x1 ∧ · · · ∧ xd for all s ∈ p V.

10.11 Hodge Stars for the Dual Space

∗ on V > in a way
V
We will now define a related Hodge star e VsimilarV
to how we
defined ge on V > . We will use g −1 to map a blade from V >Vto V, use ∗
V
to get an orthogonal blade for the result V and map that back to V > using
V the
V −1 > > −1
as ( g ) ◦ ∗ ◦ g −1 ,
−1
V V
inverse of g . So we willVdefine e ∗: V → V V
which works out to be H ◦ g when the unscaled g −1 ◦ H is substituted
−1

for ∗. Thus, applying e∗ to an exterior product of vectors from the dual basis
>
BH gives
∗ x> > ⊥ ⊥



e i 1 ∧ · · · ∧ xi n = H xi 1 ∧ · · · ∧ xi n .

Employing the usual standard scalar product with BH as its chosen basis, we
∗(x>
then have, for example, e > > >
1 ∧ · · · ∧ xn ) = xn+1 ∧ · · · ∧ xd .

Exercise 10.11.1 Using ge toV define orthogonality for V > , e

∗β is an orthog-
>
onal blade of the blade β in V .
As an alternative to the related
V Hodge star e ∗ that we have just defined,
we could define a Hodge star on V > in the manner of the V previous section,
using the related scalar product ge on V extended to each p V > .
>

∗ with the Hodge star defined on V > in the

V
Exercise 10.11.2 Compare e
manner of the previous section, using the related scalar product ge on V > .
10.12. PROBLEMS 153

10.12 Problems
1. For finite-dimensional spaces identified with their double duals, the
included maps belonging to a pairing are the duals of each other.
2. A pairing of a vector space with itself is reflexive if and only if it is
either symmetric or alternating.
3. Give an example of vectors u, v, w such that u⊥v and v⊥w but it is
not the case that u⊥w.
4. The pairing g of two d-dimensional vector spaces with respective bases
{x1 , . . . , xd } and {y1 , . . . , yd } is nondegenerate if and only if
 
g(x1 , y1 ) · · · g(x1 , yd )
.. ..
det   6= 0.
 
. ··· .
g(xd , y1 ) · · · g(xd , yd )

5. A pairing g of two finite-dimensional vector spaces V and W with re-

spective bases {x1 , . . . , xd } and {y1 , . . . , ye } is an element
Pof the func-
tional product space V > W > and may be expressed as g = i,j g(xi , yj ) x> >
i yj .

6. When F is the field with just two elements, so that 1+1 = 0, any scalar
product on F 2 will make some nonzero vector orthogonal to itself.
7. For a scalar product space over a field where 1 + 1 6= 0, the scalar
product g satisfies g(w − v, w − v) = g(v, v) + g(w, w) if and only if
v⊥w.
V
8. What, if anything, is the Hodge star on V when dim V = 1?
9. With the same setup as Exercise 10.10.4,
X X
r= (−1)ρ g(∗r, xI ) · xI = (−1)JJ g(∗r, xJ ) · xJ ,
|I|=p |J|=d−p

while X
∗(∗r) = G−1 (−1)JJ g(∗r, xJ ) · xJ .
|J|=d−p

where J = {j1 , . . . , jd−p }, J = {jd−p+1 , . . . , jd }, and J ∪J = {1, . . . , d}.

Hence
∗(∗r) = G−1 (−1)p(d−1) · r = G−1 (−1)p(d−p) · r.
154 CHAPTER 10. SCALAR PRODUCT SPACES

10. ∗−1 (∗s ∧∗t) = H −1 (H (s) ∧ H (t)) = s ∨ t for s, t ∈ V. On the other
V

hand, e∗ e∗−1 σ ∧ e∗−1 τ = H (H −1 (σ) ∧ H −1 (τ )) = σ ∨ e τ for σ, τ ∈

V >
e compare to the regressive product defined on V >
V
V . How does ∨ V
using the dual of the basis used for V? How do ∗ (∗s ∧ ∗t) and
∗ (e
e ∗σ ∧ e
∗τ ) relate to the same respective regressive products?
11. In R3 with the standard basis used for defining H and as the chosen
basis for the usual standard scalar product, the familiar cross product
u × v of two vectors u and v is the same as ∗(u ∧ v).
12. Using multi-index notation as in Section 10.9 above,
xI ∧ ∗xI = g (xI , xI ) · x⊥ ⊥
1 ∧ · · · ∧ xd
= g (xI , xI ) · G−1 · x1 ∧ · · · ∧ xd

where {x1 , . . . , xd } is the basis used in defining ∗, and G = det [g(xi , xj )].
13. Using {x1 , . . . , xd } as the basis in defining ∗, and using multi-index
notation as in Section 10.9 above, then
X
∗xJ = G−1 (−1)ρ g (xI , xJ ) · x I ,
|I|=|J|

but X
∗e x>
J = (−1)ρ ge (xI , xJ ) · x>I ,
|I|=|J|

where I = {i1 , . . . , ip }, J= {j1 , . . . , jp },

 I = {ip+1 , . . . , id } is such that
1 ··· d
I ∪ I = {1, . . . , d}, ρ = and G = det [g(xi , xj )].
i1 · · · id
14. Let {x1 , . . . , xd } be the basis used in defining ∗, let F be the real
numbers, and let d = dim V > 1. The symmetric bilinear functional g
on V such that 
1, when i 6= j
g(xi , xj ) =
0, otherwise
is a scalar product that makes xj ∧ ∗xj = 0 for each j. (The nondegen-
eracy of g follows readily upon observing that multiplying the matrix
[g(xi , xj )] on the right by the matrix [hi,j ], where

−1, when i 6= j
hi,j = ,
d − 2, otherwise
10.12. PROBLEMS 155

gives −(d − 1) times the identity matrix.)

15. If two bases have the same Gram determinant with respect to a given
scalar product, then they have the same Gram determinant with respect
to any scalar product.

16. All chosen bases that produce the same positive standard scalar product
also produce the same Hodge star up to sign.
⊥
17. x⊥
J = (xJ ) .