Some multilinear algebra

Multilinear forms
Let V be a finite dimensional R-vector space, dim V “ n.
For k P N0 , we denote by T k V the vector space of k-linear forms
µ
V ˆ . . . ˆ V ÝÑ R.

Then T 1 V “ V ˚ and, by convention, T 0 V “ R.

Let tei u be a basis of V . Then a k-linear form µ P T k V is determined by the values
µpei1 , . . . , eik q for 1 ď i1 , . . . , ik ď n, and these values can be arbitrary. This follows from the
representation
ÿ k
ź
µpv1 , . . . , vk q “ µpei1 , . . . , eik q e˚ij pvj q (0.1)
1ďi1 ,...,ik ďn j“1

where te˚i u ˚
denotes the dual basis of V . The k-linear forms
k
ź
e˚i1 ...ik pv1 , . . . , vk q “ e˚ij pvj q
j“1

constitute a basis of T k V which we refer to as the basis induced by the basis tei u, and (0.1) can
be rewritten as: ÿ
µ“ µpei1 , . . . , eik q e˚i1 ...ik (0.2)
1ďi1 ,...,ik ďn
k k
In particular, dim T V “ pdim V q .
There is a natural product of multilinear forms, called the tensor product. Namely, for
k, l P N0 one has the natural bilinear map
b
T k V ˆ T l V ÝÑ T k`l V

given by
pµ b νqpv1 , . . . , vk`l q “ µpv1 , . . . , vk q ¨ νpvk`1 , . . . , vk`l q
for µ P T k V and ν P T l V . Equipped with the tensor product,
à k
T ˚V “ T V
kě0

becomes an associative graded R-algebra, called the tensor algebra of V ˚ . It can be characterized
by the universal property of being the largest associative R-algebra with unity containing V ˚ .
For elements of induced bases, it holds that

e˚i1 ...ik b e˚ik`1 ...ik`l “ e˚i1 ...ik`l

and, in particular,
e˚i1 ...ik “ e˚i1 b ¨ ¨ ¨ b e˚ik .

1
We therefore can rewrite (0.2) again as
ÿ
µ“ µpei1 , . . . , eik q e˚i1 b ¨ ¨ ¨ b e˚ik
1ďi1 ,...,ik ďn

One can pull back multilinear forms with linear maps, and the pull-back is compatible with the
A
tensor product: A vector space homomorphism V Ñ W induces a homomorphism of graded
algebras
A˚
T ˚ W ÝÑ T ˚ V (0.3)
by
pA˚ µqpv1 , . . . , vk q “ µpAv1 , . . . , Avk q
for µ P T k W .

A reminder on permutations
Let Sk denote the symmetric group of k symbols, realized as the group of bijective self maps of
the set t1, . . . , ku. The sign of a permutation π P Sk is defined as
ź πpiq ´ πpjq
sgnpπq “ P t˘1u.
1ďiăjďk
i´j

It counts the parity of the number of inversions of π, i.e. of pairs pi, jq such that i ă j and
πpiq ą πpjq. The sign is positive iff the number of inversions is even, and such permutations
are called even. Transpositions are odd, and a permutation is even iff it can be written as the
product of an even number of transpositions. The sign is multiplicative,

sgnpπ1 π2 q “ sgnpπ1 q sgnpπ2 q,

i.e. the sign map

sgn
Sk ÝÑ t˘1u
is a homomorphism.

Alternating multilinear forms

We return to our discussion of multilinear forms.
There is a natural action
Sk ñ T k V
by
pσµqpv1 , . . . , vk q “ µpvσp1q , . . . , vσpkq q
A k-linear form µ is called alternating or skew-symmetric if

σµ “ sgnpσq ¨ µ

2
for all σ P Sk , i.e.
µpvσp1q , . . . , vσpkq q “ sgnpσq ¨ µpv1 , . . . , vk q.
Note that µ then vanishes on k-tuples of linearily dependent vectors. In particular, all alter-
nating k-forms for k ą n vanish.
We denote by
Λk V Ă T k V
the linear subspace of alternating k-linear forms. Then Λ0 V “ R, Λ1 V “ V ˚ and Λk V “ 0 for
k ą n.
There is a natural projection
alt
T k V ÝÑ Λk V
by skew-symmetrization,
1 ÿ
alt µ “ sgnpσq σµ,
k! σPS
k

i.e.
1 ÿ
palt µqpv1 , . . . , vk q “ sgnpσq ¨ µpvσp1q , . . . , vσpkq q.
k! σPS
k

Fix now a basis tei u of V . Then an alternating k-linear form α P Λk V is already determined
by the values αpei1 , . . . , eik q for 1 ď i1 ă ¨ ¨ ¨ ă ik ď n, and these values can be arbitrary. To
see this, we consider the forms
1 ÿ
alt e˚i1 ...ik “ sgnpσq ¨ σe˚i1 ...ik .
k! σPS
k

Note that they are given by determinants:

1 ÿ
palt e˚i1 ...ik qpv1 , . . . , vk q “ ˚
sgnpσq ¨ eloooooooooooomoooooooooooon
i1 ...ik pvσp1q , . . . , vσpkq q
k! σPS
k śk
j“1 ij e˚ pvσpjq q
looooooooooooooooooooomooooooooooooooooooooon
detpe˚
i pvl qqj,l
j

The forms alt e˚i1 ...ik constitute a basis of Λk V . The form α can be represented as
ÿ
α “ k! ¨ αpei1 , . . . , eik q ¨ alt e˚i1 ...ik . (0.4)
1ďi1 㨨¨ăik ďn
`n˘
In particular, dim Λk V “ k
.
There is a natural product of alternating multilinear forms, called the wedge product, which
is derived from the tensor product by skew-symmetrization: There are natural bilinear maps
^
Λk V ˆ Λl V ÝÑ Λk`l V

given by
pk ` lq!
α^β “ altpα b βq (0.5)
k!l!

3
i.e.
1 ÿ
pα ^ βqpv1 , . . . , vk`l q “ sgnpσq ¨ αpvσp1q , . . . , vσpkq q ¨ βpvσpk`1q , . . . , vσpk`lq q
k!l! σPS
k`l

The wedge product is associative and (graded) anticommutative. The latter property means
that
β ^ α “ p´1qkl α ^ β.
Equipped with the wedge product,
n
à
Λ˚ V “ Λk V
k“0

becomes an anticommutative associative graded R-algebra, called the exterior algebra of V ˚ . It

can be characterized by the universal property of being the largest associative R-algebra with
unity containing V ˚ so that the elements in V ˚ anticommute. We have that dim Λ˚ V “ 2dim V .
For multiple products, (0.5) generalizes to
pk1 ` ¨ ¨ ¨ ` kr q!
α1 ^ ¨ ¨ ¨ ^ αr “ alt α1 b ¨ ¨ ¨ b αr ,
k1 ! ¨ ¨ ¨ kr !
equivalently,
1 ÿ
pα1 ^ ¨ ¨ ¨ ^ αr qpv1 , . . . , vk q “ sgnpσq ¨ α1 pvσp1q , . . . q ¨ ¨ ¨ αr p. . . , vσpkq q
k1 ! ¨ ¨ ¨ kr ! σPS
k

where k “ k1 ` ¨ ¨ ¨ ` kr . For 1-forms λ1 , . . . , λk P Λ1 V “ V ˚ this specializes to

pλ1 ^ ¨ ¨ ¨ ^ λk qpv1 , . . . , vk q “ detpλj pvl qqj,l . (0.6)

In particular,
pe˚i1 ^ ¨ ¨ ¨ ^ e˚ik qpv1 , . . . , vk q “ detpe˚ij pvl qqj,l
i.e.
e˚i1 ^ ¨ ¨ ¨ ^ e˚ik “ k! ¨ alt e˚i1 ...ik .
We call the set of forms e˚i1 ^ ¨ ¨ ¨ ^ e˚ik the basis of Λk V induced by the basis tei u of V .
The representation (0.4) can be rewritten as
ÿ
α“ αpei1 , . . . , eik q ¨ e˚i1 ^ ¨ ¨ ¨ ^ e˚ik .
1ďi1 㨨¨ăik ďn

For a vector v P V , the natural linear maps iv : Λk V Ñ Λk´1 V given by

iv α “ αpv, . . . q

are called interior multiplication or contraction with v. The combined linear map iv : Λ˚ V Ñ
Λ˚ V satisfies i2v “ 0 and is an antiderivation of degree ´1, i.e. one has the product rule

iv pα ^ βq “ piv αq ^ β ` p´1qk α ^ piv βq (0.7)

4
for α P Λk V and β P Λ˚ V . To verify this, we may restrict to decomposable forms α “ λ1 ^¨ ¨ ¨^λk
and β “ λ11 ^ ¨ ¨ ¨ ^ λ1l with λi , λ1i P Λ1 V , because both sides are bilinear. From (0.6), we obtain
that
ÿk
iv pλ1 ^ . . . ^ λk q “ p´1qi`1 lo λoimo
pv1onq ¨pλ1 ^ . . . λ̂i . . . ^ λk q
i“1

where “λ̂i ” indicates that the factor λi is omitted. We get an analogous expression for iv pλ11 ^
. . . ^ λ1l q and (0.7) follows.
A
If V Ñ W is a linear map of vector spaces, then the pull-back (0.3) of multilinear forms
restricts to the pull-back
A˚
Λ˚ W ÝÑ Λ˚ V (0.8)
of alternating forms, i.e. the pull-back of an alternating multilinear form is again alternating.
The pull-back is also compatible with the wedge product, i.e. (0.8) is a homomorphism of
graded algebras.
If A P EndpV q, then A˚ |Λn V “ det A ¨ idΛn V .

Orientation
We suppose first that n ě 1. Let BpV q denote the space of bases e “ pe1 , . . . , en q of V . There
is a natural simply transitive right action BpV q ð GLpn, Rq given by
ÿ
e ¨ A “ e1 with e1j “ aij ei
i

for A “ paij q.
Note that GLpn, Rq has two (path) connected components, namely the subgroup GL` pn, Rq
of matrices with positive determinant and the matrices with negative determinant. Accordingly,
BpV q has two path components, which are the two GL` pn, Rq-orbits.
Let e, e1 P BpV q be bases and let A P GLpn, Rq be the unique transition matrix with
e ¨ A “ e1 . Then e and e1 lie in the same component of BpV q iff det A ą 0. In this case one
says that e and e1 have the same orientation. One defines an orientation of V as a component
of BpV q, i.e. as an equivalence class of bases with the same orientation. One then calls these
bases positively oriented, and the others negatively oriented. A basis determines an orientation,
namely the component of BpV q containing it.
The standard orientation of Rn is determined by its standard basis pe1 , . . . , en q.
We have that dim Λn V “ 1. A non-zero element 0 ‰ ω P Λn V is called a volume form. For
bases e, e1 with e ¨ A “ e1 it holds that ωpe11 , . . . , e1n q “ det A ¨ ωpe1 , . . . , en q. The volume form
ω therefore determines an orientation by defining a basis e as positively oriented iff

ωpe1 , . . . , en q ą 0.

One can alternatively define an orientation as a (ray) component of the pointed line Λn V ´ t0u.

5
 In the case n “ 0, one defines an orientation of the trivial vector space V “ t0u as a choice
of sign ˘. In particular, there is the natural orientation `.

Complex vector spaces have natural orientations as real vector spaces, because the groups
GLpn, Cq are path connected. Indeed, if W is a C-vector space with dimC W “ n, then the
space BC pW q of complex bases (acted upon simply transitively by GLpn, Cq) is path connected,
and the image of the natural map

BC pW q ÝÑ BR pW q, pe1 , . . . , en q ÞÑ pe1 , ie1 , . . . , en , ien q

is contained in a component of BR pW q which one defines to be the natural orientation.

Symmetric multilinear forms and polynomials

We continue our discussion of multilinear forms.
A k-linear form µ is called symmetric if

σµ “ µ

for all σ P Sk , i.e.

µpvσp1q , . . . , vσpkq q “ µpv1 , . . . , vk q.
We denote by
SkV Ă T kV
the linear subspace of alternating k-linear forms. Then S 0 V “ R and S 1 V “ V ˚ .
There is a natural projection
sym
T k V ÝÑ S k V
by symmetrization,
1 ÿ
sym µ “ σµ,
k! σPS
k

i.e.
1 ÿ
psym µqpv1 , . . . , vk q “ µpvσp1q , . . . , vσpkq q.
k! σPS
k

Fix now a basis tei u of V . Then a symmetric k-linear form µ P S k V is already determined by
the values µpei1 , . . . , eik q for 1 ď i1 ď ¨ ¨ ¨ ď ik ď n, and these values can be arbitrary. To see
this, we consider the forms
1 ÿ ˚
sym e˚i1 ...ik “ σei1 ...ik .
k! σPS
k

where e˚i1 ...ik “ e˚i1 b ¨¨¨ b e˚ik . They are given by:

1 ÿ ˚
psym e˚i1 ...ik qpv1 , . . . , vk q “ ei1 ...ik pvσp1q , . . . , vσpkq q
k! σPS loooooooooooomoooooooooooon
k śk ˚
j“1 eij pvσpjq q

6
 `n`k´1˘
The forms sym e˚i1 ...ik constitute a basis of S k V . In particular, dim S k V “ k
.
There is a natural product of symmetric multilinear forms, called the symmetric product,
which is derived from the tensor product symmetrization: There are natural bilinear maps
¨
S k V ˆ S l V ÝÑ S k`l V

given by
µ d ν “ sympµ b νq (0.9)
i.e.
1 ÿ
pµ d νqpv1 , . . . , vk`l q “ µpvσp1q , . . . , vσpkq q ¨ νpvσpk`1q , . . . , vσpk`lq q
pk ` lq! σPS
k`l

The normalization is motivated by the isomorphism to the polynomial algebra constructed

below. The symmetric product is associative and commutative. Equipped with the symmetric
product,
àn
S ˚V “ SkV
k“0
becomes a commutative associative graded R-algebra, called the symmetric algebra of V ˚ . It
can be characterized by the universal property of being the largest commutative associative
R-algebra with unity containing V ˚ .
There are natural linear isomorphisms

S k V ÝÑ Polk pV q (0.10)

to the spaces of homogeneous polynomials of degree k on V by restriction to the diagonal, i.e.

the polynomial Pµ associated to a symmetric k-linear form µ is given by

Pµ pvq “ µpv, . . . , vq.

The basis element sym e˚i1 ...ik of S k V corresponds to the polynomial e˚i1 ¨ . . . ¨ e˚ik . The inverse of
(0.10) is called polarization. An explicit formula for it is given by:1
k
p´1qk ÿ ÿ
µpw1 , . . . , wk q “ p´1qj Pµ pwi1 ` . . . ` wij q
k! j“1 i 㨨¨ăi 1 j

Due to our choice of normalization, multiplication is preserved by the maps (0.10),

Pµdν “ Pµ ¨ Pν ,

and they constitute an isomorphism of graded algebras

à
S ˚ V ÝÑ PolpV q “ Polk pV q
kě0

to the algebra of polynomials on V .

1
Proof: Indeed, when replacing Pµ pwi1 ` . . . ` wij q by µpwi1 ` . . . ` wij , . . . , wi1 ` . . . ` wij q for every proper
subset of indices I Ĺ t1, . . . , ku, |I| “ l ă k, the terms µpwm1 , . . . , wmk q with m1 , . . . , mk P I occur in the sums
řk
for j ě l, and hence with the multiplicity j“l p´1qj k´l
` ˘
j´l “ 0. This leaves the terms µpwσp1q , . . . , wσpnq q for
the permutations σ P Sn , which occur with multiplicity k! in the sum for j “ k.

7
Tensor product
A bilinear map
β
U ˆ V ÝÑ W (0.11)
of R-vector spaces can be regarded as a product of vectors in U with vectors in V whose values
are vectors in W . There is a natural such product with the least possible amount of linear
relations between the values:

Definition 0.12 (Tensor product). The tensor product of two vector spaces U and V is a
vector space U b V together with a bilinear map
b
U ˆ V ÝÑ U b V

satisfying the following universal property: For every bilinear map (0.11) there exists a unique
linear map
λ
U b V ÝÑ W
such that β “ λ ˝ b.

In other words, the natural homomorphism

HompU b V, W q ÝÑ BilpU, V ; W q, λ ÞÑ λ ˝ b

is an isomorphism.

Theorem 0.13. A tensor product exists and is unique up to natural isomorphism.

We can therefore speak of the tensor product.

b
Proof. Uniqueness follows from the universal property: Given two tensor products U ˆ V ÝÑ
b
r λ λ
r
U bV and U ˆV ÝÑ U bV r , there exist unique linear maps U bV ÝÑ U bV
r and U bVr ÝÑ U bV
such that b “ λ ˝ b and b “ λ ˝ b. It follows that b “ pλ ˝ λq ˝ b and b “ pλ ˝ λq ˝ b. The
r r r r r r r
uniqueness part of the universal property then implies that λr ˝ λ “ idU bV and λ ˝ λ
r “ id r .
U bV
Hence, between any two tensor products there is a natural isomorphism.
Existence. We form a vector space with basis U ˆ V , namely let E be the vector space
consisting of all finite linear combinations
ÿ
ai pui , vi q
i

with ui P U , vi P V and ai P R. We denote by

ι
U ˆ V ÝÑ E

the inclusion. It is only a map of sets. To make it bilinear, we divide out the minimal necessary
amount of relations: Let R Ă E be the linear subspace generated by the elements

pu1 , vq ` pu2 , vq ´ pu1 ` u2 , vq, pu, v1 q ` pu, v2 q ´ pu, v1 ` v2 q,

8
 pau, vq ´ apu, vq, pu, avq ´ apu, vq
for u, u1 , u2 P U , v, v1 , v2 P V and a P R. Then ι descends to the bilinear map
b
U ˆ V ÝÑ E{R “: U b V, pu, vq ÞÑ pu, vq ` R “: u b v.
β
We must verify that it satisfies the universal property. Given a bilinear map U ˆ V Ñ W , there
λ̂
exists a unique linear map E Ñ W extending it, β “ λ̂ ˝ ι. The bilinearity of β is equivalent to
λ ` ˘
λ̂pRq “ 0. Hence, λ̂ descends to a linear map U b V Ñ W with λpu b vq “ λ̂ pu, vq “ βpu, vq.
The last property uniquely determines λ because the elements u b v generate U b V .

A more concrete idea of the tensor product is provided by the following fact.

Lemma 0.14. If tei : i P Iu and tfj : j P Ju are bases of U and V , then tei b fj : i P I, j P Ju
is a basis of U b V . In particular, in the case of finite dimensions, it holds that

dim U b V “ dim U ¨ dim V.

Proof. From the construction of the tensor product we know that the elements u b v generate
U b V . Consequently, also the elements ei b fj generate.
In order to show that they are linearily independent, we separate them by linear forms.
Namely, we use the linear forms
λ
kl
U b V ÝÑ R, u b v ÞÑ e˚k puqfl˚ pvq

which are induced by the bilinear forms

β
kl
U ˆ V ÝÑ R, pu, vq ÞÑ e˚k puqfl˚ pvq.

For a finite linear relation ÿ

cij ei b fj “ 0
i,j

it follows by applying
`ÿ ˘
ckl “ λkl cij ei b fj “ 0.
i,j

Thus, the elements ei b fj form a basis.

With respect to the bases, the tensor product is given by:

ÿ ÿ ÿ
p ai ei q b p bj fj q “ ai b j e i b f j
i j i,j

When changing the bases of the factors, ẽk “ i gki ei and f˜l “ j hlj fj , the induced change of
ř ř

basis for the tensor product is given by:

ÿ
ẽk b f˜l “ gki hlj ei b fj
i,j

9
Remark 0.15. It follows that if tei : i P Iu is a basis of U , then every element in U b V can
be written as ÿ
ei b vi
i
with unique vectors vi P V .

Remark 0.16. (i) There is a canonical isomorphism switching factors

U bV –V bU (0.17)

which identifies the elements u b v with the elements v b u. It is induced by the bilinear map
U ˆ V Ñ V b U sending pu, vq ÞÑ v b u.
(ii) There are canonical homomorphisms

U ˚ b V ÝÑ HompU, V q; u˚ b v ÞÑ u˚ p¨qv (0.18)

relating spaces of homomorphisms to tensor products. If dim U ă 8, then these are isomor-
phisms, i.e. the space of homomorphisms can then be represented as a tensor product,

HompU, V q – U ˚ b V.

Indeed, if tei u is a basis of U , then te˚i u is a basis of U ˚ (because of finite dimensionality),

ř
Elements of U ˚ b V have unique representations as sums i e˚i b vi and, correspondingly,
ř
elements of HompU, V q as sums i e˚i p¨qvi .
If also dim V ă 8 and tfj u is a basis of V , then the homomorphism
ÿ
aji e˚i b fj

is given with respect to the chosen bases by the matrix paji qj,i .
In particular, if U “ V , then we have the natural isomorphism EndpU q – U ˚ b U , and idU
ř
corresponds to the element i e˚i b ei .
(ii’) The homomorphism (0.18) is always injective, as one sees by restricting to finite di-
mensional subspaces of U .

By analogy with the twofold tensor product, the multiple tensor product of an arbitrary
number of vector spaces U1 , . . . , Un is a multilinear map
b
U1 ˆ . . . ˆ Un ÝÑ U1 b . . . b Un

with the universal property that every multilinear map U1 ˆ . . . ˆ Un Ñ W is the composition
of b with a unique linear map U1 b . . . b Un Ñ W . Existence and uniqueness of the multiple
tensor product are proven in the same way.

Lemma 0.19 (Associativity). There are natural isomorphisms

pU1 b . . . b Un q b pUn`1 b . . . b Un`m q ÝÑ U1 b . . . b Un`m

mapping elements pu1 b . . . b un q b pun`1 b . . . b un`m q to elements u1 b . . . b un`m .

10
Proof. The natural multilinear map

U1 ˆ . . . ˆ Un`m ÝÑ pU1 b . . . b Un q b pUn`1 b . . . b Un`m q

induces a linear map

U1 b . . . b Un`m ÝÑ pU1 b . . . b Un q b pUn`1 b . . . b Un`m q

mapping elements u1 b . . . b un`m to elements pu1 b . . . b un q b pun`1 b . . . b un`m q. That it

is an isomorphism, follows by choosing bases.

Remark 0.20. (i) Permutations of factors. Generalizing (0.17), for permutations σ P Sn there
are the natural isomorphisms

U1 b . . . b Un – Uσp1q b . . . b Uσpnq

mapping elements u1 b . . . b un to elements uσp1q b . . . b uσpnq .

(ii) Functoriality. Homomorphisms αi : Ui Ñ Vi induce a homomorphism

α1 b ¨ ¨ ¨ b αn : U1 b . . . b Un ÝÑ V1 b . . . b Vn

mapping elements u1 b . . . b un to elements α1 pu1 q b . . . b αn pun q. Indeed, it is induced by the

multilinear map U1 ˆ . . . ˆ Un ÝÑ V1 b . . . b Vn sending pu1 , . . . , un q ÞÑ α1 pu1 q b . . . b αn pun q
(iii) Multilinear forms. Generalizing (0.18), we have natural injective homomorphisms
` ˘
U1˚ b. . .bUn˚ bV ÝÑ MultpU1 , . . . , Un ; V q; u˚1 b. . .bu˚n bv ÞÑ pu1 , . . . , un q ÞÑ u˚1 pu1 q¨. . .¨u˚n pun qv

which are isomorphisms if dim Ui ă 8.

Tensor algebra. For a vector space U and m P N0 we denote by

Tm U :“ bm U :“ loooooomoooooon
U b ... b U
m

the m-fold tensor power of U . Then T1 U “ U . Convention: T0 U “ R.

Consider the graded vector space

T˚ U :“ ‘8
m“0 Tm U

Due to the associativity of the tensor product, there are natural bilinear maps
b
Tm U ˆ Tn U ÝÑ Tm`n U

which, by bilinear extension, yield a product

b
T˚ U ˆ T˚ U ÝÑ T˚ U.

Equipped with this product, T˚ U becomes a graded associative R-algebra with unity, the (co-
variant) tensor algebra of U . Covariant, because the functor U ÞÑ T˚ U from vector spaces to

11
algebras is covariant, i.e. a linear homomorphism U Ñ V induces a homomorphism of graded
algebras T˚ U Ñ T˚ V in the same direction.
The covariant tensor algebra can also be characterized by a universal property. The algebra
T˚ U is the “largest” associative algebra with unity “generated by U ” in the sense that every
linear homomorphism α : U Ñ A to an associative algebra with unity uniquely extends to an
algebra homomorphism T˚ U Ñ A; it maps elements u1 b . . . b un to αpu1 q ¨ . . . ¨ αpun q.
The contravariant tensor algebra of U ˚ is defined as

T ˚ U :“ T˚ pU ˚ q.

Contravariant, because a linear homomorphism U Ñ V induces a linear homomorphism V ˚ Ñ

U ˚ of dual spaces and hence an algebra homomorphism T ˚ V Ñ T ˚ U in the reverse direction.
Both types of tensors can be combined to mixed tensors with covariant and contravariant
components. Consider
Trs U :“ Tr U b T s U
for r, s P N0 , with the convention T00 U “ R, and put
s
T pU q :“ T˚˚ U :“ ‘8
r,s“0 Tr U.

Again, there is a natural product b on T pU q satisfying

pu1 b . . . b ur1 b u˚1 b . . . b u˚s1 q b pv1 b . . . b vr2 b v1˚ b . . . b vs˚2 q

“ u1 b . . . b ur1 b v1 b . . . b vr2 b u˚1 b . . . b u˚s1 b v1˚ b . . . b vs˚2

which makes T pU q into a bigraded associative algebra with unity, the tensor algebra of U . There
are natural inclusions T˚ U Ă T pU q and T ˚ U Ă T pU q such that Tr U “ Tr0 U and T s U “ T0s U .
The elements of T pU q are called tensors, and the elements of Trs U are called homogeneous
tensors of type pr, sq.

Example 0.21. Examples: Tensors of type p1, 0q are vectors, tensors of type p0, 1q are linear
forms, tensors of type p0, 2q are bilinear forms, tensors of type p1, 1q are endomorphisms.

Trace and contraction. Suppose now that dim U ă 8. Then the natural inclusion U ãÑ U ˚˚
is an isomorphism and induces natural isomorphisms

Trs U ˚ – Tsr U, T pU ˚ q – T pU q.

The natural bilinear pairing

U ˆ U ˚ ÝÑ R, pu, u˚ q ÞÑ u˚ puq

induces the linear form

tr
T11 U “ U b U ˚ – EndpU q ÝÑ R
which is nothing but the trace. Indeed, if tei u is a basis of U , then the endomorphism A “
ř ˚
ř ˚
ř
i,j aij ei b ej given by the matrix paij q is mapped to i,j aij ej pei q “ i aii “ trpAq.

12
 More generally, one can pair the i-th covariant factor of a homogeneous tensor with the j-th
contravariant factor and thus obtains the contraction homomorphisms

Cj
Trs U ÝÑ
i s´1
Tr´1 U

for 1 ď i ď r and 1 ď j ď s satisfying

u1 b . . . b ur b u˚1 b . . . b u˚s ÞÑ u˚j pui q ¨ u1 b . . . ûi . . . b ur b u˚1 b . . . û˚j . . . b u˚s .

By composing partial contractions one obtains (various) total contractions

Trr U ÝÑ R,

e.g. ź
C11 ˝ . . . ˝ Crr : u1 b . . . b u˚r ÞÑ u˚i pui q.
i

We apply contractions to construct pairings. A natural non-degenerate bilinear pairing

Trs U ˆ Tsr U ÝÑ R
r`s b
is obtained by composing the tensor product Trs U ˆ Tsr U Ñ Tr`s U with a total contraction.
The pairing induces a natural isomorphism

pTrs U q˚ – Tsr U – Trs U ˚ .

In particular, homogeneous contravariant tensors can be interpreted as multilinear forms:

Mults pU q – pTs U q˚ – T s U – Ts U ˚

To a tensor u˚1 b . . . b u˚s corresponds the multilinear form

ź
pu1 , . . . , us q ÞÑ u˚i pui q.
i

If tei u is a basis of U , then te˚i1 b . . . b e˚is u is a basis of Mults pU q. An s-linear form µ on U can
be written with respect to this basis as
ÿ
˚ ˚
µ“ µpe i1 , . . . , eis q ei1 b . . . b eis .
looooooomooooooon
i1 ,...,is
PR

13