Notes On Tensor Products and The Exterior Algebra: For Math 245
Notes On Tensor Products and The Exterior Algebra: For Math 245
Notes On Tensor Products and The Exterior Algebra: For Math 245
K. Purbhoo
July 16, 2012
1 Tensor Products
1.1 Axiomatic definition of the tensor product
In linear algebra we have many types of products. For example,
• The scalar product: V × F → V
• The dot product: Rn × Rn → R
• The cross product: R3 × R3 → R3
• Matrix products: Mm×k × Mk×n → Mm×n
Note that the three vector spaces involved aren’t necessarily the same. What these examples
have in common is that in each case, the product is a bilinear map. The tensor product is
just another example of a product like this. If V1 and V2 are any two vector spaces over a
field F, the tensor product is a bilinear map:
V1 × V2 → V1 ⊗ V2 ,
where V1 ⊗ V2 is a vector space over F. The tricky part is that in order to define this map,
we first need to construct this vector space V1 ⊗ V2 .
We give two definitions. The first is an axiomatic definition, in which we specify the
properties that V1 ⊗ V2 and the bilinear map must have. In some sense, this is all we need
to work with tensor products in a practical way. Later we’ll show that such a space actually
exists, by constructing it.
Definition 1.1. Let V1 , V2 be vector spaces over a field F. A pair (Y, µ), where Y is a vector
space over F and µ : V1 × V2 → Y is a bilinear map, is called the tensor product of V1 and
V2 if the following condition holds:
(*) Whenever β1 is a basis for V1 and β2 is a basis for V2 , then
µ(β1 × β2 ) := {µ(x1 , x2 ) | x1 ∈ β1 , x2 ∈ β2 }
is a basis for Y .
1
Notation: We write V1 ⊗ V2 for the vector space Y , and x1 ⊗ x2 for µ(x1 , x2 ).
The condition (*) does not actually need to be checked for every possible pair of bases
β1 , β2 : it is enough to check it for any single pair of bases.
Theorem 1.2. Let Y be a vector space, and µ : V1 × V2 → Y a bilinear map. Suppose there
exist bases γ1 for V1 , γ2 for V2 such that µ(γ1 × γ2 ) is a basis for Y . Then condition (*)
holds (for any choice of basis).
The proof is replete with quadruply-indexed summations and represents every sensible
person’s worst idea of what tensor products are all about. Luckily, we only need to do this
once. You may wish to skip over this proof, particularly if this is your first time through
these notes.
Proof. Let β1 , β2 be bases for V1 and V2 respectively. We first show that µ(β1 × β2 ) spans
Y . Let y ∈ Y . Since µ(γ1 × γ2 ) spans Y , we can write
X
y= ajk µ(z1j , z2k )
j,k
P
where z1j ∈ γ1 , P
z2k ∈ γ2 . But since β1 is a basis for V1 z1j = l bjl x1l , where x1l ∈ β1 , and
similarly z2k = m ckm x2m , where x2m ∈ β2 . Thus
X X X
y= ajk µ( bjl x1l , ckm x2m )
j,k l m
X
= ajk bjl ckm µ(x1l , x2j )
j,k,l,m
so y ∈ span(µ(β1 × β2 )).
Now we need to show that µ(β1 × β2 ) is linearly independent. If V1 and V2 are both finite
dimensional, this follows from the fact that |µ(β1 × β2 )| = |µ(γ1 × γ2 )| and both span Y . For
infinite dimensions, a more sophisticated change of basis argument is needed. Suppose
X
dlm µ(x1l , x2m ) = 0 ,
l,m
P
where xP1l ∈ β1 , x2m ∈ β2 . Then by change of basis, x1l = j elj z1j , where z1j ∈ γ1 , and
x2m = k fmkP z2k , where z2k ∈ γ2 . Note that P the elj form an inverse “matrix” to the bjl
above, in that j elj bjl0 = δll0 , and similarly k fmk ckm0 = δmm0 . Thus we have
X
0= dlm µ(x1l , x2m ) = 0
l,m
X X X
= dlm µ( elj z1j , fmk z2k )
l,m j k
XX
= dlm elj fmk µ(z1j , z2k )
j,k l,m
2
P
Since µ(γ1 × γ2 ) is linearly independent, l,m dlm elj fmk = 0 for all j, k. But now
X
dl0 m0 = dlm δll0 δmm0
l,m
! !
X X X
= dlm bjl0 elj ckm0 fmk
l,m j k
!
X X
= bjl0 ckm0 dlm elj fmk
j,k l,m
=0
for all l0 , m0 .
The tensor product can also be defined for more than two vector spaces.
Definition 1.3. Let V1 , V2 , . . . Vk be vector spaces over a field F. A pair (Y, µ), where Y is
a vector space over F and µ : V1 × V2 × · · · × Vk → Y is a k-linear map, is called the tensor
product of V1 , V2 , . . . , Vk if the following condition holds:
(*) Whenever βi is a basis for Vi , i = 1, . . . , k,
{µ(x1 , x2 , . . . , xk ) | xi ∈ βi }
is a basis for Y .
We write V1 ⊗ V2 ⊗ · · · ⊗ Vk for the vector space Y , and x1 ⊗ x2 ⊗ · · · ⊗ xk for µ(x1 , x2 , . . . , xk ).
where ai ∈ F, vi ∈ V , wi ∈ W .
Example 1.4. Show that
(1, 1) ⊗ (1, 4) + (1, −2) ⊗ (−1, 2) = 6(1, 0) ⊗ (0, 1) + 3(0, 1) ⊗ (1, 0)
in R2 ⊗ R2 .
Solution: Let x = (1, 0), y = (0, 1). We rewrite the left hand side in terms of x and y, and
use bilinearity to expand.
(1, 1) ⊗ (1, 4) + (1, −2) ⊗ (−1, 2) = (x + y) ⊗ (x + 4y) + (x − 2y) ⊗ (−x + 2y)
= (x ⊗ x + 4x ⊗ y + y ⊗ x + 4y ⊗ y)
+ (−x ⊗ x + 2x ⊗ y − 2y ⊗ x − 4y ⊗ y)
= 6x ⊗ y + 3y ⊗ x
= 6(1, 0) ⊗ (0, 1) + 3(0, 1) ⊗ (1, 0)
3
The other way is to actually identify the space V1 ⊗ V2 and the map V1 × V2 → V1 ⊗ V2
with some familiar object. There are many examples in which it is possible to make such
an identification naturally. Note, when doing this, it is crucial that we not only specify the
vector space we are identifying as V1 ⊗ V2 , but also the product (bilinear map) that we are
using to make the identification.
Example 1.5. Let V = Fnrow , and W = Fm col . Then V ⊗ W = Mm×n (F), with the product
defined to be v ⊗ w = wv, the matrix product of a column and a row vector.
To see this, we need to check condition (*). Let {e1 , . . . , em } be the standard basis for
W , and {f1 , . . . , fn } be the standard basis for V . Then {fj ⊗ ei } is the standard basis for
Mm×n (F). So condition (*) checks out.
Note: we can also say that W ⊗V = Mm×n (F), with the product defined to be w⊗v = wv.
From this, it looks like W ⊗ V and V ⊗ W are the same space. Actually there is a subtle
but important difference: to define the product, we had to switch the two factors. The
relationship between V ⊗ W and W ⊗ V is very closely analogous to that of the Cartesian
products A × B and B × A, where A and B are sets. There’s an obvious bijection between
them and from a practical point of view, they carry the same information. But if we started
conflating the two, and writing (b, a) and (a, b) interchangeably it would cause a lot of
problems. Similarly V ⊗ W and W ⊗ V are naturally isomorphic to each other, so in that
sense they are the same, but we would never write v ⊗ w when we mean w ⊗ v.
Example 1.6. Let V = F[x], vector space of polynomials in one variable. Then V ⊗
V = F[x1 , x2 ] is the space of polynomials in two variables, the product is defined to be
f (x) ⊗ g(x) = f (x1 )g(x2 ).
To check condition (*), note that β = {1, x, x2 , x3 , . . .} is a basis for V , and that
β ⊗ β = {xi ⊗ xj | i, j = 0, 1, 2, 3, . . . }
= {xi1 xj2 | i, j = 0, 1, 2, 3, . . . }
Example 1.7. If V is any vector space over F, then V ⊗ F = V . In this case, ⊗ is just
scalar multiplication.
Example 1.8. Let V, W be finite dimensional vector spaces over F. Let V ∗ = L(V, F) be the
dual space of V . Then V ∗ ⊗ W = L(V, W ), with multiplication defined as f ⊗ w ∈ L(V, W )
is the linear transformation (f ⊗ w)(v) = f (v) · w, for f ∈ V ∗ , w ∈ W .
This is just the abstract version of Example 1.5. If V = Fncol and W = Fm col , then
∗
V = Fnrow . Using Example 1.5, V ⊗ W is identified with Mm×n (F), which is in turn identified
with L(V, W ).
Exercise: Prove that condition (*) holds.
Note: If V and W are both infinite dimensional then V ∗ ⊗ W is a subspace of L(V, W )
but not equal to it. Specifically, V ∗ ⊗ W = {T ∈ L(V, W ) | dim R(T ) < ∞} is the set of finite
4
rank linear transformations in L(V, W ). This follows from the fact that f ⊗ w has rank 1,
for any f ∈ V ∗ , w ∈ W , and a linear combination of such transformations must have finite
rank.
Example 1.9. Let V be the space of velocity vectors in Newtonian 3-space. Let T be the
vector space of time measurements. Then V ⊗ T is the space of displacement vectors in
Newtonian 3-space. This is because velocity times time equals displacement.
The point of this example is that physical quantities have units associated with them.
Velocity is not a vector in R3 . It’s an element of a totally different 3-dimensional vector space
over R. To perform calculations, we identify it with R3 by choosing coordinate directions
such as up forward and right and units such as m/s, but these are artificial constructions.
Displacement again lives in a different vector space, and the tensor product allows us to
relate elements in these different physical spaces.
Definition 1.11. Let A be a set, and F a field. The free vector space over F generated by
A is the vector space Free(A) consisting of all formal finite linear combinations of elements
of A.
Example 1.12. Let A = {♠, ♥, ♣, ♦}, and F = R. Then the Free(A) is the four-dimensional
vector space consisting of all elements a♠ + b♥ + c♣ + d♦, where a, b, c, d ∈ R. Addition
and scalar multiplication are defined in the obvious way:
5
When the elements of the set A are numbers or vectors, the notation gets tricky, because
there is a danger of confusing the operations of addition and scalar multiplication and the
zero-element in the vector space Free(A), and the operations of addition and multiplication
and the zero element in A (which are irrelevant in the definition of Free(A)). To help keep
these straight in situations where there is a danger of confusion, we’ll write , and when
we mean the operations of addition and scalar multiplication in Free(A). We’ll denote the
zero vector of Free(A) by 0Free(A) .
Example 1.13. Let N = {0, 1, 2, 3, 4, . . . }, and F = R. Then Free(N) is the infinite dimen-
sional vector space whose elements are of the form
for some m ∈ N.
Note that the element 0 here is not the zero vector in Free(N). It’s called 0 because it
happens to be the zero element in N, but this is completely irrelevant in the construction of
the free vector space.
If we wanted we could write this a little differently by putting xi in place of i ∈ N. In
this new notation, the elements of Free(N) would look like
a0 x 0 + a1 x 1 + · · · + am x m ,
for some m ∈ N, in other words elements of the vector space of polynomials in a single
variable.
Definition 1.14. Let V and W be vector spaces over a field F. Let
P := Free(V × W ) ,
the free vector space over F generated by the set V × W . Let R ⊂ P be the subspace
spanned by all vectors of the form
(u + kv, w + lx) ((−1) (u, w)) ((−k) (v, w)) ((−l) (u, x)) ((−kl) (v, x))
with k, l ∈ F, u, v ∈ V , and w, x ∈ W .
Let πP/R : P → P/R be the quotient map. Let µ : V × W → P/R be the map
µ(v, w) = πP/R (v, w) .
The pair (P/R , µ) is the tensor product of V and W . We write V ⊗ W for P/R, and
v ⊗ w for µ(v, w).
We need to show that our two definitions agree, i.e.. that tensor product as defined in
Definition 1.14 satisfies the conditions of Definition 1.1. In particular, we need to show that
µ is bilinear, and that the pair (P/R , µ) satisfies condition (*).
We can show the bilinearity immediately. Essentially bilinearity is built into the defini-
tion. If the P is the space of all linear combinations of symbols (v, w), then R is the space of
all those linear combinations that can be simplified to the zero vector using bilinearity. Thus
P/R is the set of all expressions, where two expressions are equal iff one can be simplified
to the other using bilinearity.
6
Proposition 1.15. The map µ in Definition 1.14 is bilinear.
Lemma 1.16. Suppose V and W are vector spaces over a field F and T : V → W is a
linear transformation. Let S be a subspace of V . Then there exists a linear transformation
T : V /S → W such that T (x + S) = T (x) for all x ∈ V if and only if T (s) = 0 for all s ∈ S.
Moreover, if T exists it is unique, and very linear transformation V /S → W arises in
this way from a unique T .
Proof. Suppose that T exists. Then for all s ∈ S we have T (s) = T (s + S) = T (0V /S ) = 0.
Now suppose that T (s) = 0 for all s ∈ S. We must show that T (x + S) = T (x) makes T
well defined. In other words, we must show that if v, v 0 ∈ V are such that v + S = v 0 + S,
then T (v + S) = T (v) = T (v 0 ) = T (v 0 + S). Now, if v + S = v 0 + S, then v − v 0 ∈ S. Thus
T (v) − T (v 0 ) = T (v − v 0 ) = 0 and so T (v) = T (v 0 ) as required.
The final remarks are clear, since the statement T (x + S) = T (x) uniquely determines
either of T , T from the other.
Theorem 1.17 (Universal property of tensor products). Let V, W, M be vector spaces over
a field F. Let V ⊗ W = P/R be the tensor product, as defined in Definition 1.14. For any
bilinear map φ : V × W → M , there is a unique linear transformation φ : V ⊗ W → M ,
such that φ(v ⊗ w) = φ(v, w) for all v ∈ V , w ∈ W . Moreover, every linear transformation
in L(V ⊗ W, M ) arises in this way.
Proof. Since V × W is a basis for P, we can extend any map φ : V × W → M to a linear map
φ : P → M . By a similar argument to Proposition 1.15, one can see that φ is bilinear if and
only if φ(s) = 0 for all s ∈ R. Thus by Lemma 1.16, there exists a linear map φ : P/R → M
if and only if φ is bilinear, and every such linear transformation arises from a bilinear map
in this way.
More generally, every linear map V1 ⊗ V2 ⊗ · · · ⊗ Vk → M corresponds to a k-linear map
V1 × V2 × · · · × Vk → M .
7
Theorem 1.18. Condition (*) holds for the tensor product as defined in Definition 1.14.
Proof. Let β = be a basis for V , and γ a basis for W . We must show that µ(β × γ) is a
basis for P/R. First we show it is spanning. Let z ∈ P/R. Then we can write
z = a1 πP/R (u1 , x1 ) + · · · + am πP/R (um , xm )
= a1 µ(u1 , x1 ) + · · · + am µ(um , xm )
P
. am ∈ F, u1 , . . . um ∈ V , and x1 , . . . xm ∈ W . But now ui = j bij vj with vj ∈ β
with a1 , . .P
and xi = k cij wk with wk ∈ γ. Hence
z = a1 µ(u1 , x1 ) + · · · + am µ(um , xm )
Xm X X
= ai µ( bij vj , bik wk )
i=1 j k
Xm X
= ai bij cik µ(vj , wk ) .
i=1 j,k
with vi ∈ β, wj ∈ γ. Let fk ∈ V ∗ be the linear functional for which fk (vk ) = 1, and fk (v) = 0,
for v ∈ β \ {vk }. Define Fk : V × W → W to be Fk (v, w) = fk (v)w. Then Fk is bilinear,
so by Theorem 1.17 there is a induced linear transformation
P Fk : V ⊗ W → W such that
Fk (µ(u, k)) = fk (u)x. Applying Fk to the equation i,j dij µ(vi , wj ) = 0, we have
!
X
0 = Fk dij µ(vi , wj )
i,j
X
= dij fk (vi )wj
i,j
X
= dkj wj ,
j
8
Proposition 1.19. There is a natural linear transformation tr : V ∗ ⊗ V → F, such that
tr(f ⊗ v) = f (v) for all f ∈ V ∗ , v ∈ V .
Proof. The map (f, v) 7→ f (v) is bilinear, so the result follows by Theorem 1.17.
The map tr is called the trace on V . We can also define the trace from V ⊗ V ∗ → F,
using either the fact that V = V ∗∗ , or by switching the factors in the definition (both give
the same answer).
Example 1.20. Let V = Fncol . Let {e1 , . . . en } be the standard basis for V , and let
{f1 , . . . , fn } be the dual basis (standard basis for row vectors).
From Example 1.5, we can identify V ∗ ⊗ V with Mn×n . A matrix A ∈ Mn×n can be
written as
X
A= Aij ej fi
i,j
X
= Aij fi ⊗ ej .
i,j
Therefore
X
tr(A) = Aij tr(fi ⊗ ej )
i,j
X
= Aij fi (ej )
i,j
X
= Aij δij
i,j
Now, let V1 , V2 , . . . , Vk be finite dimensional vector spaces over F, and consider the tensor
product space V1 ⊗ V2 ⊗ · · · ⊗ Vk . Suppose that Vi = Vj∗ for some i, j. Using the trace, we
can define a linear transformation
satisfying
This transformation is sometimes called contraction. The most basic example of this is
something familiar in disguise.
Example 1.21. Let V, W, X be finite dimensional vector spaces. From Example 1.8 we have
V ∗ ⊗ W = L(V, W ), W ∗ ⊗ X = L(W, X) and V ∗ ⊗ X = L(V, X). The map
tr23 : V ∗ ⊗ W ⊗ W ∗ ⊗ X → V ∗ ⊗ X
9
can therefore be identified with a map from
This map is nothing other than the composition of linear transformations: (T, U ) 7→ U ◦ T .
Proof. Since composition is bilinear, and the space of all linear transformations is spanned
by rank 1 linear transformations, it is enough to prove this in the case where T ∈ L(V, W ),
and U ∈ L(W, X) have rank 1. If T ∈ L(V, W ) has rank 1, then T (v) = f (v)w, where w ∈ W ,
f ∈ V ∗ . As in Example 1.8 we identify this with f ⊗ w ∈ V ∗ ⊗ W . Similarly U (w) = g(w)x
where x ∈ X, g ∈ W ∗ , and U is identified with g ⊗ x ∈ W ∗ ⊗ X. Then
so U ◦ T is identified with
as required.
Theorem 1.17 can also be used to define the tensor product of two linear transformations.
Proposition 1.22. Let V, V 0 , W, W 0 be vector spaces over a field F. If T ∈ L(V, V 0 ), and
U ∈ L(W 0 , W 0 ), there a linear unique transformation T ⊗ U ∈ L(V ⊗ W, V 0 ⊗ W 0 ) satisfying
T ⊗ U (v ⊗ w) = (T v) ⊗ (U w)
for all v ∈ V , w ∈ W .
In particular, if T ∈ L(V ) and U ∈ L(W ) are linear operators, then T ⊗ U ∈ L(V ⊗ W )
is a linear operator on V ⊗ W .
Proof. The map (v, w) 7→ (T v) ⊗ (U w) is bilinear, so the result follows from Theorem 1.17.
10
Let C k be the subspace of Tk (V ) spanned by all vectors of the form
(x1 ⊗ · · · ⊗ xi ⊗ · · · ⊗ xj ⊗ · · · ⊗ xk ) − (x1 ⊗ · · · ⊗ xj ⊗ · · · ⊗ xi ⊗ · · · ⊗ xk ) ,
Definition 2.1. The k th symmetric power of V is the quotient space Symk (V ) = Tk (V )/C k .
We write x1 · x2 · · · xk (or x1 x2 · · · xk ) for πTk (V )/C k (x1 ⊗ x2 ⊗ · · · ⊗ xk ).
(x1 · x2 · · · xk , y1 · y2 · · · yl ) 7→ x1 · x2 · · · xk · y1 · y2 · · · yl .
In other words, Symk (V ) can be thought of as the space of polynomials of degree k in the
elements of V , and the operation “ · ” is just ordinary commutative polynomial multiplication.
When the elements are expressed in terms of a basis, these look exactly like polynomials.
Example 2.2. Consider the product (1, 1, 1) · (−1, 1, 1) · (0, 1, −1) ∈ Sym3 (R3 ) Let x =
(1, 0, 0), y = (0, 1, 0), z = (0, 0, 1). When we expand this product in terms of x, y, z we
obtain:
If we want to consider polynomials of mixed degrees, we can look at the direct sum of all
the symmetric products of V :
11
3 The exterior algebra
3.1 Definition and examples
The exterior algebra is constructed similarly to the polynomial algebra. Its properties are
analogous to those of the polynomial algebra, but the multiplication that arises is not com-
mutative. As before V will be a vector space over a field F.
Let Ak be the subspace of Tk (V ) spanned by all vectors of the form
x1 ⊗ x2 ⊗ · · · ⊗ xk
x1 ∧ · · · ∧ xi ∧ · · · ∧ xj ∧ · · · ∧ xk = −x1 ∧ · · · ∧ xj ∧ · · · ∧ xi ∧ · · · ∧ xk .
Proof.
0 = x1 ∧ · · · ∧ (xi + xj ) ∧ · · · ∧ (xi + xj ) ∧ · · · ∧ xk
= x1 ∧ · · · ∧ xi ∧ · · · ∧ xi ∧ · · · ∧ xk + x1 ∧ · · · ∧ xi ∧ · · · ∧ xj ∧ · · · ∧ xk
+ x1 ∧ · · · ∧ xj ∧ · · · ∧ xi ∧ · · · ∧ xk + x1 ∧ · · · ∧ xj ∧ · · · ∧ xj ∧ · · · ∧ xk
= x1 ∧ · · · ∧ xi ∧ · · · ∧ xj ∧ · · · ∧ xk + x1 ∧ · · · ∧ xj ∧ · · · ∧ x j ∧ · · · ∧ xk
12
On the other hand, if we do a similar calculation with three terms,
we get exactly zero. It looks the the dimensions of dim k (R2 ) are given by the table below.
V
k 0 1 2 3 4 5 ···
Vk 2
dim (R ) 1 2 1 0 0 0 ···
basis of k (R2 )
V
{1} {x, y} {x ∧ y} ∅ ∅ ∅ ···
However, at this point weVhaven’t completely proved this. We still need to show that x ∧ y
is not the zero vector in 2 (R2 ).
As with the symmetric products, there is a multiplication map between the exterior
products, which allows us to think of “∧00 as a multiplication operator. This product is
commonly known as the wedge product. It is associative, but not commutative.
Proposition 3.4. There is a natural bilinear map k (V ) × l (V ) → k+l (V ) such that
V V V
(v1 ∧ v2 ∧ · · · ∧ vk , w1 ∧ w2 ∧ · · · ∧ wl ) 7→ v1 ∧ v2 ∧ · · · ∧ vk ∧ w1 ∧ w2 ∧ · · · ∧ wl
is a ring, with multiplication defined by the wedge product. A vector space over F which
also has the structure of a ring is called an F-algebra.
Example 3.6. Let V = R3 , x = (1, 0, 0), y = (0, 1, 0), z = (0, 0, 1). We have
13
Here we’ve used y ∧ z ∧ x = −x ∧ z ∧ y = x ∧ y ∧ z. Note that the coefficient is equal to
a b c
det d e f
g h i
and moreover, the expression we obtained is exactly the cofactor expansion along the bottom
row.
Example 3.7. Let V = R4 with basis {x, y, z, w}. Let α = x ∧ y + z ∧ w. We have
α ∧ α = (x ∧ y + z ∧ w) ∧ (x ∧ y + z ∧ w)
=x∧y∧x∧y + x∧y∧z∧w
+ z∧w∧x∧y + z∧w∧z∧w
=x∧y∧z∧w − x∧w∧z∧y
=x∧y∧z∧w + x∧y∧z∧w
= 2x ∧ y ∧ z ∧ w
It may at first be surprising that the answer we obtained is not zero, but this is no mistake.
What this proves is that α = x ∧ y + z ∧ w is not a pure wedge product, i.e. α 6= u ∧ v for
any u, v ∈ R4 . Indeed, (u ∧ v) ∧ (u ∧ v) = 0 for all u, v. However, expressions that are not
pure wedge products can be multiplied with themselves to give a non-zero vector.
Proposition 3.8. If α ∈ k (V ), ω ∈ l (V ), then α ∧ ω = (−1)kl ω ∧ α.
V V
Thus although the wedge product is not commutative, it is reasonably close commutative.
A product with this property said to supercommutative.
Proof. If α = v1 ∧ v2 ∧ · · · ∧ vk , ω = w1 ∧ w2 ∧ · · · ∧ wl , then we can get from v1 ∧ v2 ∧ · · · ∧
vk ∧ w1 ∧ w2 ∧ · · · ∧ wl , to ±w1 ∧ w2 ∧ · · · ∧ wl ∧ v1 ∧ v2 ∧ · · · ∧ vk by swapping adjacent v’s
and w’s. Each swap contributes a factor of −1, and it takes exactly kl swaps.
14
We are now in a position to prove Proposition 3.4.
Proof. Bilinearity means linearity in each component, with the other held fixed. First, fix
w1 , . . . , wl ∈ V , and consider the map
(v1 , . . . , vk ) 7→ v1 ∧ v2 ∧ · · · ∧ vk ∧ w1 ∧ w2 ∧ · · · ∧ wl .
Vk
This map is k-linear and alternating, so by Theorem 3.9 we obtain a linear map (V ) →
V k+l
(V ) satisfying a linear map satisfying
v1 ∧ v2 ∧ · · · ∧ vk 7→ v1 ∧ v2 ∧ · · · ∧ vk ∧ w1 ∧ w2 ∧ · · · ∧ wl .
Denote this linear map α 7→ α ∧ (w1 ∧ w2 ∧ · · · ∧ wl ).
Next
V we apply a similar argument to the second factor. Consider an arbitrary element
α ∈ k (V ). The argument in the first paragraph shows that we have a well-defined map
(w1 , w2 , . . . , wl ) 7→ α ∧ (w1 ∧ w2 · · · ∧ wl )
P
If we write α = ai vi1 ∧ vi2 ∧ . . . vik where ai ∈ F and vij ∈ V , we can rewrite this as
i
X
(w1 , w2 , . . . , wl ) 7→ ai vi1 ∧ vi2 ∧ . . . vik ∧ w1 ∧ w2 ∧ · · · ∧ wl ,
i
from which V we see that this map is is l-linear and alternating. By Theorem 3.9 we obtain a
l Vk+l
linear map (V ) → (V ) which we denote ω 7→ α ∧ ω.
Finally, we need to check that the map (α, ω) 7→ α ∧ ω is bilinear. The argument of the
second paragraph shows that P it is linear in the second factor.
Vl It remains to show that it is
linear in the first. Fix ω = j bj wj1 ∧ wj2 ∧ · · · ∧ wjl ∈ (V ). Then
X
α 7→ α ∧ ω = bj α ∧ (wj1 ∧ wj2 ∧ · · · ∧ wjl ) .
j
From the first paragraph, we know that each map α 7→ α ∧ (wj1 ∧ wj2 ∧ · · · ∧ wjl ) is linear.
Hence α 7→ α ∧ ω is a linear combination of linear maps, and therefore linear.
15
Theorem 3.11. Let f1 , f2 , . . . , fk ∈VV ∗ , where V ∗ is the dual space to V . There is a well
defined linear functional Ff1 ,f2 ,...,fk : k (V ) → F, such that
f1 (v1 ) f1 (v2 ) ··· f1 (vk )
f2 (v1 ) f2 (v2 ) ··· f2 (vk )
Ff1 ,f2 ,...,fk (v1 ∧ v2 ∧ · · · ∧ vk ) = det .. .. .
..
. . .
fk (v1 ) fk (v2 ) ··· fk (vk )
Vk
Moreover, if f10 , f20 , . . . , fk0 ∈ V ∗ , and f1 ∧ f2 ∧ · · · ∧ fk = f10 ∧ f20 ∧ · · · ∧ fk0 ∈ (V ∗ ), then
Ff1 ,f2 ,...,fk = Ff10 ,f20 ,...,fk0 .
In light of this second part, we write f1 ∧ f2 ∧ · · · ∧ fk (v1 ∧ v2 ∧ · · · ∧ vk ) for Ff1 ,f2 ,...,fk (v1 ∧
v2 ∧ · · · ∧ vk ).
Proof. The first statement follows from Theorem 3.9 and the fact that determinant is k-
linear alternating in the columns. The second statement is equivalent V to saying that for
any given v1 , v2 , . . . vk , there is a well defined linear map from G : k (V ∗ ) → F such that
G(f1 ∧ f2 ∧ · · · ∧ fk ) = Ff1 ,f2 ,...,fk (v1 ∧ v2 ∧ · · · ∧ vk ). This follows from Theorem 3.9 and the
fact that the determinant is k-linear and alternating in the rows.
Proof of Theorem 3.10. First, suppose that {v1 , v2 , . . . , vk } is linearly dependent. Then we
can write one of these vectors as a linear combination of the others; without loss of generality,
say vk = a1 v1 + a2 v2 + · · · + ak−1 vk−1 . Then
Now, suppose that {v1 , v2 , . . . , vk } is linearly independent. Since this set can be extended
to a basis, we can define linear functionals f1 , . . . , fk ∈ V ∗ such that fi (vj ) = δij . But then
by Theorem 3.11 we have
1 0 ··· 0
0 1 · · · 0
f1 ∧ f2 ∧ · · · ∧ fk (v1 ∧ v2 ∧ · · · ∧ vk ) = det .. .. .. = 1 ,
. . .
0 0 ··· 1
so v1 ∧ v2 ∧ · · · ∧ vk 6= 0.
Vk
Corollary 3.12. If dim V < k then dim (V ) = 0.
Proof. If dim V < k then any k vectors will be linearly dependent, so any k-fold wedge
product is zero.
Using similar ideas, we can compute the dimensions of k (V ) when V is finite dimen-
V
sional, for all k < dim V .
16
Theorem 3.13. Let V be an n-dimensional vector space over a field F, and let β1 =
{e1 , . . . en } be a basis for V . For each k ≤ n, let
for all v1 , v2 , . . . , vk ∈ V . To see that T∗ exists, by Theorem 3.9 it is enough to check that
the map
(v1 , v2 , . . . , vk ) 7→ T (v1 ) ∧ T (v2 ) ∧ · · · ∧ T (vk )
is bilinear and alternating, which is straightforward.
Proposition 3.14. Let V and W be vector spaces over a field F, and let T ∈ L(V, W ). For
any non-negative integer k, the following are true.
(i) If T is surjective then T∗ ∈ L( k (V ), k (W )) is surjective.
V V
Proof. For (i) suppose T is surjective. It suffices to show that every vector w1 ∧ w2 ∧ . . . wk ∈
V k
(W ) is in the range of T∗ , which is clear.
For (ii) suppose T is injective. Let γ be a basis for V , and extend {T (w) | w ∈ γ} to a
basis β for W . If
!
X
0 = T∗ ai1 i2 ...ik wi1 ∧ wi2 ∧ · · · ∧ wik
1≤i1 <i2 <···<ik ≤n
X
= ai1 i2 ...ik T (wi1 ) ∧ T (wi2 ) ∧ · · · ∧ T (wik )
1≤i1 <i2 <···<ik ≤n
where {w1 , w2 , . . . , wn } is an ordered finite subset of γ then all ai1 i2 ...ik must be equal to zero.
(Exercise: Prove this.) Thus T∗ is injective.
Finally (iii) follows from (i) and (ii).
Vk
One consequence
Vk of this is that if W is a subspace of V , then (W ) can be regarded as
a subspace
Vk of Vk). More formally, let i : W → V be theVinclusion
(V
k
map. This Vis injective,
so i∗ : (W ) → (V ) is injective. Thus we can identify (W ) with R(i∗ ) ⊂ k (V ).
17
Proposition 3.15. Let V1 , V2 , V3 be vector spaces over F, and let T ∈ L(V1 , V2 ), U ∈
L(V2 , V3 ). For any non-negative integer k, we have:
Proof. LetV{e1 , e2 , . . . , en } be the standard basis for V ; let {f1 , f2 , . . . , fn } be the dual basis.
Since dim n (V ) = 1, it suffices to prove the result for ω = e1 ∧ e2 ∧ · · · ∧ en . We have
Vn 3.17. Let V be an n-dimensional vector space over F, and T ∈ L(V ). Then for
Corollary
all ω ∈ (V ), we have T∗ (ω) = det(T )ω.
Proof. Let Φ : V → Fncol be any isomorphism. Then by definition, det T = det(ΦT Φ−1 ).
Φ∗ (T∗ (ω))) = (ΦT Φ−1 )∗ (Φ∗ (ω)) = (det T )(Φ∗ (ω)) = Φ∗ ((det T )ω) .
x1 ∧ x2 ∧ · · · ∧ xk = c y1 ∧ y2 ∧ · · · ∧ yk
18
Exercise: Prove Theorem 3.18.
In other words, vectors in k (V ) which are pure wedges represent k-dimensional sub-
V
spaces of VV, up to a non-zero constant. It is therefore an important question to ask which
vectors in k (V ) are pure wedges.
Our first criterion for determining when a vector ω ∈ k (V ) is a pure wedge product is
V
the following.
Theorem 3.19. Let V vector space over F, and let ω ∈ k (V ) be a non-zero vector. Let
V
Wω = {x ∈ V | x ∧ ω = 0}. Then dim Wω ≤ k, and dim Wω = k if and only if ω is a pure
wedge.
LemmaVk 3.20.VLet V be a vector space over F. Let f ∈ V ∗ . For each k, there is a map
χf : (V ) → k−1 (V ) such that for all v1 , v2 , . . . , vk ∈ V ,
Proof. The right hand side is k-linear and alternating in (v1 , v2 , . . . vk ), so this follows by
Theorem 3.9.
More generally, we have the following important theorem.
Theorem 3.21. Let V be a vector space over F, and let k ≥ l be a positive integers. There
exists a bilinear map:
χ : l (V ∗ ) × k (V ) → k−l (V )
V V V
such that
for all f1 , f2 , . . . , fl ∈ V ∗ , v1 , v2 , . . . , vk ∈ V .
is l-linear and alternating. The l-linearity is clear. To see that this map is alternating, we
need to check that the right hand side is 0 if fi = fj for some i 6= j. To prove this, first one
checks that χf ◦ χf = 0 for all f ∈ V ∗ . This implies all that χf ◦ χg = −χg ◦ χf , so we can
swap adjacent pairs (picking up a factor of −1 with each swap) until fi and fj are next to
each other. But then we have
as required.
19
Exercise: If k = l, prove that
χ(f1 ∧ f2 ∧ · · · ∧ fk , v1 ∧ v2 ∧ · · · ∧ vk ) = f1 ∧ f2 ∧ · · · ∧ fk (v1 ∧ v2 ∧ · · · ∧ vk ) .
Proof. Let f ∈ V ∗ be a linear functional such that f (x) = 1, f (y) = 0 for all y ∈ V 0 . Then
we have 0 = χf (x ∧ θ) = f (x)θ + 0 = θ.
Proof. Let V 0 be any codimension one subspace of V which does not contain x. Let f ∈ V ∗
be the of Lemma 3.22. We can write ω = x ∧ α + β, uniquely, where
linear functional V
Vk−1
α ∈ (V ), and β ∈ k (V 0 ). But 0 = χf (x ∧ ω) = χf (x ∧ β) = β, so ω = x ∧ α, as
0
required.
Proof of Theorem 3.19. Suppose ω = v1 ∧ v2 ∧ · · · ∧ vk . By Theorem 3.10, since ω 6= 0,
{v1 , v2 , . . . , vk } is linearly independent. Again by Theorem 3.10, Wω = span{v1 , v2 , . . . , vk },
hence dim Wω = k.
Now suppose that dim Wω = k. We proceed by induction. The result is trivial for k = 1.
Assume the result is true for all smaller values of k.VLet x ∈ Wω be a non-zero vector. By
Lemma 3.23 we can write ω = x1 ∧ α, where α ∈ k−1 (V 0 ), and V 0 is a codimension one
subspace of V that does not contain x. Let W 0 = Wω ∩ V 0 . We claim that x0 ∧ α = 0 for all
x0 ∈ W 0 . To see this let f ∈ V ∗ be the linear functional of Lemma 3.22. Then we have
Since V 0 has codimension one, dim W 0 = dim W − 1 = k − 1. Thus the inductive hypothesis,
α = v1 ∧ v2 ∧ · · · ∧ vk−1 , so ω = x ∧ v1 ∧ v2 ∧ · · · ∧ vk−1 .
Finally, if dim Wω > k we could apply this second argument V with twoV different k-
dimensional subspaces W1 , W2 ⊂ Wω . But then we would have ω ∈ k (W1 )∩V k (W2 ), which
contradicts Theorem 3.18. Therefore we must have dim Wω ≤ k for all ω ∈ k (V ).
For some purposes,
Vk Theorem 3.19 is not the best answer to the problem of determining
which elements of (V ) are pure wedges. The following criterion has the advantage that
it can be used to give explicit necessary and sufficient equations that the coefficients of a
pure wedge product must satisfy (when everything is expanded in terms of a basis). These
equations are called the Plücker relations.
Theorem 3.24 (PlückerVk relations). Let V be a vector space over F, and let k V be a positive
integer. A vector ω ∈ (V ) is a pure wedge product if and only if, for all ξ ∈ k−1 (V ∗ ) we
have
χ(ξ, ω) ∧ ω = 0 .
20
Lemma 3.25. Let V be a vector space over F, and let k ≥ l be a positive integers. Let
W = span{v1 , . . . , vk }, where v1 , v2 , . . . vk ∈ V , and let f1 , f2 , . . . fl ∈ V ∗ . Then
χ(f1 ∧ f2 ∧ · · · ∧ fl , v1 ∧ v2 ∧ · · · ∧ vk ) ∈ k−l (W ) .
V
21
It is possible to simplify this further, but at the expense of messier notation. However, from
here we can see quite explicitly what happens for any specific value of k. For example, if
k = 1 we get.
5
X X
0= a1j alm ej ∧ el ∧ em
j=2 l<m
{l,m}∩{1,j}=∅
X X
= a12 alm e2 ∧ el ∧ em + a13 alm e3 ∧ el ∧ em
l<m l<m
{l,m}∩{1,2}=∅ {l,m}∩{1,3}=∅
X X
+ a14 alm e4 ∧ el ∧ em + a15 alm e5 ∧ el ∧ em
l<m l<m
{l,m}∩{1,4}=∅ {l,m}∩{1,5}=∅
Each other value of k produces four similar looking equations, but in fact there are only
5 equations in total, because each equation comes up four times in this process. The last
equation is (unsurprisingly):
By Theorem 3.24, ω is a pure wedge if an only if its coefficients satisfy these five equations.
22
2. For x, y ∈ V , hx, yi = hy, xi.
23
Proof. Let W be the subspace spanned by {v1 , . . . , vk }. Since dim W = k, any linear trans-
formation from W to Rk preserves the ratios of k-dimensional volumes. Thus it enough to
prove this for the case where the vertices are n = k, and vi = e1 + e2 + · · · + ei , where
{e1 , . . . , ek } is the basis for Rk . In this case, the volume of the parallelepiped is 1, and the
volume of the simplex is
Z 1 Z x1 Z x2 Z xk−1
··· 1 dxk dxk−1 . . . dx2 dx1 ,
x1 =0 x2 =0 x3 =0 xk =0
1
which one can easily show is equal to k!
.
24