Jump to content

Linear algebra (Osnabrück 2024-2025)/Part II/Lecture 55



Let K a field. We consider two finite index sets I and J, and the mapping spaces Map⁡(I,K) and Map⁡(J,K), which are both finite-dimensional K-vector spaces. Is there a relation with the mapping set

Map⁡(I×J,K)?

For mappings f∈Map⁡(I,K) and g∈Map⁡(J,K), we easily obtain a mapping on I×J, called f⊗g (f tensor g) that is defined by

(f⊗g)(i,j):=f(i)⋅g(j).

For the standard vectors, we have

ei⊗ej=e(i,j).

Every element h∈Map⁡(I×J,K) can be written as a linear combination of elements of the form f⊗g but not every h has this simple form. We have

(a1f1+a2f2)⊗g=a1(f1⊗g)+a2(f2⊗g),

and accordingly in the second component.

In this lecture, we introduce an important construction for vector spaces, the so-called tensor product. In the special case just considered, the tensor product yields the mapping space on the product set; that is, we have

Map⁡(I,K)⊗KMap⁡(I,K)≅Map⁡(I×J,K).

Here, the properties of the constructed object are more important than the construction itself. The construction is quite abstract, and rests on the construction of residue class spaces, and the following construction.


For an arbitrary set of symbols S, and a field K, we can construct the vector space K(S) that consists of all mappings

S⟶K

that have everywhere except for finitely many places the value 0. If es denotes the mapping that has the value 1 at the position s, and everywhere else the value 0, then K(S) consists of all finite sums

∑s∈Sases.

Hence, the es form a basis of this space.

This construction is a special case of the direct sum of (in general) infinitely many K-vector spaces; we take the direct sum of the vector spaces K with itself as often as the set S tells us.



The tensor product of vector spaces

Let K be a field, and let V1,…,Vn be K-vector spaces. We recall that a multilinear mapping in another K-vector space W is a mapping

ψ:V1×⋯×Vn⟶W

that is K-linear in every component, that is, when all the other components are fixed. We want to construct a vector space U, together with a multilinear mapping

φ:V1×⋯×Vn⟶U

such that for every multilinear mapping ψ as above, there exists a linear mapping

L:U⟶W

fulfilling ψ=L∘φ. By this construction, multilinear mappings are transformed to linear mappings on a new vector space.


Let K be a field, and let V1,…,Vn be K-vector spaces. Let F be the K-vector space generated by all symbols (v1,…,vn) (with vi∈Vi, we write the basis elements as e(v1,…,vn)). Let U⊆F be the K-linear subspace of F generated by all elements of the form

  1. re(v1,…,vi−1,vi,vi+1,…,vn)−e(v1,…,vi−1,rvi,vi+1,…,vn),
  2. e(v1,…,vi−1,u+w,vi+1,…,vn)−e(v1,…,vi−1,u,vi+1,…,vn)−e(v1,…,vi−1,w,vi+1,…,vn).

Then the residue class space F/U is called the tensor product of the Vi, i∈{1,…,n}. It is denoted by

V1⊗KV2⊗K⋯⊗KVn.

We usually just write V1⊗⋯⊗Vn. The images of e(v1,…,vn) in V1⊗KV2⊗K⋯⊗KVn are denoted by

v1⊗⋯⊗vn.

This is the equivalence class of e(v1,…,vn) for the equivalence relation given by the linear subspace U. Every element in V1⊗K⋯⊗KVn has a (not unique) representation of the form

a1v1,1⊗⋯⊗v1,n+⋯+amvm,1⊗⋯⊗vm,n

(with ai∈K, and vi,j∈Vj). In particular, the decomposable tensors v1⊗⋯⊗vn form a K-generating system of the tensor product. The defining generators of the linear subspace become equations in the tensor product; they express the multilinearity. In particular, we have

v1⊗⋯⊗vi−1⊗rvi⊗vi+1⊗⋯⊗vn=r(v1⊗⋯⊗vi−1⊗vi⊗vi+1⊗⋯⊗vn)

for arbitrary i, and

v1⊗⋯⊗vi−1⊗(u+w)⊗vi+1⊗⋯⊗vn=v1⊗⋯⊗vi−1⊗u⊗vi+1⊗⋯⊗vn+v1⊗⋯⊗vi−1⊗w⊗vi+1⊗⋯⊗vn

For K=ℝ and V1=ℝ2 and V2=ℝ3, the elements of F (in the sense of Definition) are linear combinations like

4e((32),(6−15))−5e((17),(334))+6e((24),(−478)).

Using the standard vectors of ℝ2 and of ℝ3, this is

4e(3e1+2e2,6e1−e2+5e3)−5ee1+7e2,3e1+3e2+4e3+6e(2e1+4e2,−4e1+7e2+8e3).

Because these tuples are different, we can not simplify this expression in F. The image of this element in F/U=ℝ2⊗ℝℝ3 is

4(3e1+2e2)⊗(6e1−e2+5e3)−5(e1+7e2)⊗(3e1+3e2+4e3)+6(2e1+4e2)⊗(−4e1+7e2+8e3).

This expression can be simplified as a linear combination of the family ei⊗ej, i=1,2, j=1,2,3.

More important than the construction of the tensor product is the following universal property.


Theorem

Let K be a field, and let V1,…,Vn denote vector spaces

over K.
  1. The mapping
    π:V1×⋯×Vn⟶V1⊗K⋯⊗KVn,(v1,…,vn)⟼v1⊗⋯⊗vn,

    is K-multilinear.

  2. Let W be another K-vector space, and let
    ψ:V1×⋯×Vn⟶W

    denote a multilinear mapping. Then there exists a uniquely determined K-linear mapping

    ψ¯:V1⊗K⋯⊗KVn⟶W

    fulfilling ψ=ψ¯∘π.

Proof  

  1. follows immediately from the definition of the tensor product.
  2. Since the v1⊗⋯⊗vn are a K-generating system of V1⊗K⋯⊗KVn, and because
    ψ¯(v1⊗⋯⊗vn)=ψ(v1,…,vn)

    must hold, there can exist at most one such a linear mapping. To show existence, we consider the K-vector space F from the construction of the tensor product. The e(v1,…,vn) form a basis of F; therefore, the assignment

    ψ~(e(v1,…,vn)):=ψ(v1,…,vn)

    defines a linear mapping

    ψ~:F⟶W.

    Because of the multilinearity of ψ, the linear subspace U is mapped by ψ~ to 0. Hence, according to the factorization theorem, this mapping induces a K-linear mapping

    ψ‾:F/U≅V1⊗K⋯⊗KVn⟶W.
◻



Corollary

Let K be a field, and let V1,…,Vn and W be vector spaces over K. Then there exists a natural isomorphism

HomK(V1⊗⋯⊗Vn,W)⟶Mult(V1,…,Vn;W),φ⟼φ∘π.

Proof  

From Theorem 55.4   (2), we may conclude immediately that we have a bijection.

◻



Corollary

Let K be a field, and let V1,…,Vn denote vector spaces over K. Then there exists a natural isomorphism

(V1⊗⋯⊗Vn)∗⟶Mult(V1,…,Vn;K),φ⟼φ∘π.

Proof  

This follows immediately from Corollary 55.5 , applied to W=K.

◻



For V=V1=V2, and W=K, the multilinear mappings from V1×V2 to W are just the bilinear forms on V. Corollary 55.6 says in this situation that the dual space of V⊗V represents all bilinear forms. For V=Kn, the standard inner product (this term is only for K=ℝ correct) corresponds to the linear form φ:V⊗V→K given by

φ(ei⊗ej)=δij.


The tensor product is determined up to (unique) isomorphy by this universal property; this means the following.


Theorem

Let K be a field, and let V1,…,Vn denote vector spaces over K. Let Z be a K-vector space, together with a multilinear mapping

ρ:V1×⋯×Vn⟶Z,

and suppose that (Z,ρ) fulfills the universal property from Theorem 55.4   (2). Then there exists a uniquely determined isomorphism

V1⊗⋯⊗Vn⟶Z.

Proof  

Because V1×⋯×Vn→Z is multilinear, there exists, due to the universal property of the tensor product, a uniquely determined linear mapping

V1⊗⋯⊗Vn⟶Z.

Because of the universal property of Z and the multilinearity of π:V1×⋯×Vn→V1⊗⋯⊗Vn, there is also a linear mapping

Z⟶V1⊗⋯⊗Vn.

Because of the universal property, they are inverse to each other.

◻

Therefore, this universal property is more important than the explicit construction of the tensor product.


Lemma

Let K be a field, and let V1,…,Vn denote vector spaces

over K. Then the following calculation rules hold.
  1. For vectors vj∈Vj, and c∈K, we have
    v1⊗⋯⊗vi−1⊗cvi⊗vi+1⊗⋯⊗vn=c(v1⊗⋯⊗vi−1⊗vi⊗vi+1⊗⋯⊗vn).
  2. For vectors vj∈Vj, we have
    v1⊗⋯⊗vi−1⊗0⊗vi+1⊗⋯⊗vn=0.
  3. Let vi1,…,vimi∈Vi, and aij∈K. Then
    (∑j=1m1a1jv1j)⊗⋯⊗(∑j=1mnanjvnj)=∑(j1,…,jn)∈{1,…,m1}×⋯×{1,…,mn}a1j1⋯anjnv1j1⊗⋯⊗vnjn
    holds.

Proof  

(1) follows immediately from the construction. (2) follows from (1). (3) follows from the distributive law for multilinear mappings.

◻



In ℝ2⊗ℝℝ2, we have

(5−7)⊗(−34)=(5e1−7e2)⊗(−3e1+4e2)=−15e1⊗e1+20e1⊗e2+21e2⊗e1−28e2⊗e2.

Lemma

Let K be a field, and let V denote a

K-vector space. Then the following properties hold.
  1. We have
    V⊗K0=0.
  2. We have
    V⊗KK=V,

    where v⊗s corresponds to the vector sv

Proof  

(1) follows from Lemma 55.9   (2).

(2). The scalar multiplication

V×K⟶V,(v,s)⟼sv,

is multilinear; therefore, there exists, according to Theorem 55.4 , a linear mapping

V⊗K⟶V,v⊗s⟼sv.

This mapping is surjective, because v⊗1 is mapped to v. An element in the tensor product has the form

∑i=1nai(vi⊗si)=∑i=1n(aisi)(vi⊗1)=(∑i=1naisivi)⊗1.

If this is mapped to 0, then we have

∑i=1naisivi=0;

but then also the tensor element is 0, and the mapping is injective as well.

◻



Theorem

Let K be a field, and let V1,…,Vn denote vector spaces over K. Let J1,…,Jn be index sets, and let

vij,j∈Ji,
be vectors in Vi. Then the following statements hold.
  1. If each family is a generating system of Vi, then the family
    v1j1⊗⋯⊗vnjn with (j1,…,jn)∈J1×⋯×Jn

    is a generating system of V1⊗⋯⊗Vn.

  2. If each family is linearly independent in Vi, then the family
    v1j1⊗⋯⊗vnjn with (j1,…,jn)∈J1×⋯×Jn

    is linearly independent in V1⊗⋯⊗Vn.

  3. If each family is a basis of Vi, then the family
    v1j1⊗⋯⊗vnjn with (j1,…,jn)∈J1×⋯×Jn

    is a basis of V1⊗⋯⊗Vn.

Proof  

(1). Due to the construction, the decomposable tensors v1⊗⋯⊗vn form a generating system of the tensor product. Hence, it is enough to show that they are linear combinations of the given family. But this follows from Lemma 55.9   (3).

(2). We can restrict to finite families. We want to apply Lemma 14.8 . Let (r1,…,rn)∈J1×⋯×Jn be fixed. Because of the linear independence of the families vij, j∈Ji, in Vi, there exist linear forms

φi:Vi⟶K

with φi(viri)≠0 and with φi(vij)=0 for j≠ri. Therefore,

V1×⋯×Vn⟶K,(w1,…,wn)⟼φ1(w1)⋯φn(wn),

is, according to Exercise 16.39 , a multilinear mapping. Due to Corollary 55.6 , we have a corresponding linear mapping

V1⊗⋯⊗Vn⟶K,

which sends v1r1⊗⋯⊗vnrn to

φ1(v1r1)⋯φn(vnrn)≠0,

and all other elements v1j1⊗⋯⊗vnjn of the family to 0.

(3) follows from (1) and (2).

◻



Corollary

Let K be a field, and let V1,…,Vn denote finite-dimensional vector spaces over K. Then the dimension of the tensor product is

dimK(V1⊗⋯⊗Vn)=dimK(V1)⋯dimK(Vn).

Proof  

This follows immediately from Theorem 55.12   (3).

◻

We relate the motivating Example 55.1 with the general construction of the tensor product. The mapping (with the direct meaning of f⊗g from the example)

Map⁡(I,K)×Map⁡(J,K)⟶Map⁡(I×J,K),(f,g)⟼f⊗g,

is, due to Exercise 55.5 , multilinear. Because of Theorem 55.4   (2), there exists a uniquely determined linear mapping

Map⁡(I,K)⊗KMap⁡(J,K)⟶Map⁡(I×J,K),

where the tensor products correspond to each other. The surjectivity follows from the fact that the basis elements e(i,j) are in the image. The injectivity follows from Corollary 55.13 .


<< | Linear algebra (Osnabrück 2024-2025)/Part II | >>
PDF-version of this lecture
Exercise sheet for this lecture (PDF)