Tensores 2
Tensores 2
Tensores 2
KEITH CONRAD
1. Introduction
Continuing our study of tensor products, we will see how to combine two linear maps
M −→ M 0 and N −→ N 0 into a linear map M ⊗R N → M 0 ⊗R N 0 . This leads to flat modules
and linear maps between base extensions. Then we will look at special features of tensor
products of vector spaces (including contraction), the tensor products of R-algebras, and
finally the tensor algebra of an R-module.
We call ϕ ⊗ ψ the tensor product of ϕ and ψ, but be careful to appreciate that ϕ ⊗ ψ is not
denoting an elementary tensor. This is just notation for a new linear map on M ⊗R N .
ϕ 1⊗ϕ ϕ⊗1
When M −−→ M 0 is linear, the linear maps N ⊗R M −−−−→ N ⊗R M 0 or M ⊗R N −−−−→
M 0 ⊗R N are called tensoring with N . The map on N is the identity, so (1 ⊗ ϕ)(n ⊗ m) =
n ⊗ ϕ(m) and (ϕ ⊗ 1)(m ⊗ n) = ϕ(m) ⊗ n. This construction will be particularly important
for base extensions in Section 4.
i i⊗1
Example 2.1. Tensoring inclusion aZ −−→ Z with Z/bZ is aZ ⊗Z Z/bZ −−−→ Z ⊗Z Z/bZ,
where (i ⊗ 1)(ax ⊗ y mod b) = ax ⊗ y mod b. Since Z ⊗Z Z/bZ ∼ = Z/bZ by multiplication,
we can regard i ⊗ 1 as a function aZ ⊗Z Z/bZ → Z/bZ where ax ⊗ y mod b 7→ axy mod b.
Its image is {az mod b : z ∈ Z/bZ}, which is dZ/bZ where d = (a, b); this is 0 if b | a and is
Z/bZ if (a, b) = 1.
0 0
Example 2.2. Let A = ( ac db ) and A0 = ( ac0 db 0 ) in M2 (R). Then A and A0 are both linear
maps R2 → R2 , so A ⊗ A0 is a linear map from (R2 )⊗2 = R2 ⊗R R2 back to itself. Writing e1
and e2 for the standard basis vectors of R2 , let’s compute the matrix for A ⊗ A0 on (R2 )⊗2
1
2 KEITH CONRAD
(A ⊗ A0 )(e1 ⊗ e1 ) = Ae1 ⊗ A0 e1
= (ae1 + ce2 ) ⊗ (a0 e1 + c0 e2 )
= aa0 e1 ⊗ e1 + ac0 e1 ⊗ e2 + ca0 e2 ⊗ e1 + cc0 e2 ⊗ e2 ,
0
(A ⊗ A )(e1 ⊗ e2 ) = Ae1 ⊗ A0 e2
= (ae1 + ce2 ) ⊗ (b0 e1 + d0 e2 )
= cb0 e1 ⊗ e1 + ad0 e1 ⊗ e2 + cb0 e2 ⊗ e2 + cd0 e2 ⊗ e2 ,
and similarly
The construction of tensor products (Kronecker products) of matrices has the following
application to finding polynomials with particular roots.
Theorem 2.3. Let K be a field and suppose A ∈ Mm (K) and B ∈ Mn (K) have eigenvalues
λ and µ in K. Then A ⊗ In + Im ⊗ B has eigenvalue λ + µ and A ⊗ B has eigenvalue λµ.
(A ⊗ In + Im ⊗ B)(v ⊗ w) = Av ⊗ w + v ⊗ Bw
= λv ⊗ w + v ⊗ µw
= (λ + µ)(v ⊗ w)
and
(A ⊗ B)(v ⊗ w) = Av ⊗ Bw = λv ⊗ µw = λµ(v ⊗ w),
TENSOR PRODUCTS II 3
√ √
Example 2.4. The numbers
√ √ 2 and 3 are eigenvalues of A = ( 01 20 ) and B = ( 01 30 ). A
matrix with eigenvalue 2 + 3 is
0 0 2 0 0 3 0 0
0 0 0 2 1 0
0 0
A ⊗ I2 + I2 ⊗ B = 1 +
0 0 0 0 0 0 3
0 1 0 0 0 0 1 0
0 3 2 0
1 0 0 2
= 1
,
0 0 3
0 1 1 0
√ √
whose characteristic polynomial is T 4 − 10T 2 + 1. So this is a polynomial with 2+ 3 as
a root.
Although we stressed that ϕ ⊗ ψ is not an elementary tensor, but rather is the notation
for a linear map, ϕ and ψ belong to the R-modules HomR (M, M 0 ) and HomR (N, N 0 ), so
one could ask if the actual elementary tensor ϕ ⊗ ψ in HomR (M, M 0 ) ⊗R HomR (N, N 0 ) is
related to the linear map ϕ ⊗ ψ : M ⊗R N → M 0 ⊗R N 0 .
that sends the elementary tensor ϕ ⊗ ψ to the linear map ϕ ⊗ ψ. When M, M 0 , N , and N 0
are finite free, this is an isomorphism.
Proof. We adopt the temporary notation T (ϕ, ψ) for the linear map we have previously
written as ϕ ⊗ ψ, so we can use ϕ ⊗ ψ to mean an elementary tensor in the tensor product
of Hom-modules. So T (ϕ, ψ) : M ⊗R N → M 0 ⊗R N 0 is the linear map sending every m ⊗ n
to ϕ(m) ⊗ ψ(n).
Define HomR (M, M 0 )×HomR (N, N 0 ) → HomR (M ⊗R N, M 0 ⊗R N 0 ) by (ϕ, ψ) 7→ T (ϕ, ψ).
This is R-bilinear. For example, to show T (rϕ, ψ) = rT (ϕ, ψ), both sides are linear maps
so to prove they are equal it suffices to check they are equal at the elementary tensors in
M ⊗R N :
T (rϕ, ψ)(m ⊗ n) = (rϕ)(m) ⊗ ψ(n) = rϕ(m) ⊗ ψ(n) = r(ϕ(m) ⊗ ψ(n)) = rT (ϕ, ψ)(m ⊗ n).
T (Ei0 i ⊗ E
ej 0 j )(eµ ⊗ fν ) = Ei0 i (eµ ) ⊗ E
ej 0 j (fν )
= δµi e0i0 ⊗ δνj fj0 0
(
e0i0 ⊗ fj0 0 , if µ = i and ν = j,
=
0, otherwise,
so T (Ei0 i ⊗Ej 0 j ) sends ei ⊗fj to e0i0 ⊗fj0 0 and sends other members of the basis of M ⊗R N to 0.
That means the linear map HomR (M, M 0 ) ⊗R HomR (N, N 0 ) → HomR (M ⊗R N, M 0 ⊗R N 0 )
sends a basis to a basis, so it is an isomorphism when the modules are finite free.
The upshot of Theorem 2.5 is that HomR (M, M 0 ) ⊗R HomR (N, N 0 ) naturally acts as
linear maps M ⊗R N → M 0 ⊗R N 0 and it turns the elementary tensor ϕ ⊗ ψ into the linear
map we’ve been writing as ϕ ⊗ ψ. This justifies our use of the notation ϕ ⊗ ψ for the linear
map, but it should be kept in mind that we will continue to write ϕ ⊗ ψ for the linear map
itself (on M ⊗R N ) and not for an elementary tensor in a tensor product of Hom-modules.
Properties of tensor products of modules carry over to properties of tensor products of
linear maps, by checking equality on all tensors. For example, if ϕ1 : M1 → N1 , ϕ2 : M2 →
N2 , and ϕ3 : M3 → N3 are linear maps, we have ϕ1 ⊗ (ϕ2 ⊕ ϕ3 ) = (ϕ1 ⊗ ϕ2 ) ⊕ (ϕ1 ⊗ ϕ3 )
and (ϕ1 ⊗ ϕ2 ) ⊗ ϕ3 = ϕ1 ⊗ (ϕ2 ⊗ ϕ3 ), in the sense that the diagrams
ϕ1 ⊗(ϕ2 ⊕ϕ3 )
M1 ⊗R (M2 ⊕ M3 ) / N1 ⊗R (N2 ⊕ N3 )
and
ϕ1 ⊗(ϕ2 ⊗ϕ3 )
M1 ⊗R (M2 ⊗R M3 ) / N1 ⊗R (N2 ⊗R N3 )
Example 2.7. The composition rule for tensor products of linear maps helps us compute
determinants of tensor products of linear operators. Let M and N be finite free R-modules
ϕ ψ
of respective ranks k and `. For linear operators M −−→ M and N −−→ N , we will compute
det(ϕ ⊗ ψ) by breaking up ϕ ⊗ ψ into a composite of two maps M ⊗R N → M ⊗R N :
ϕ ⊗ ψ = (ϕ ⊗ idN ) ◦ (idM ⊗ψ),
so the multiplicativity of the determinant implies det(ϕ ⊗ ψ) = det(ϕ ⊗ idN ) det(idM ⊗ψ)
and we are reduced to the case when one of the “factors” is an identity map. Moreover, the
isomorphism M ⊗R N → N ⊗R M where m ⊗ n 7→ n ⊗ m converts ϕ ⊗ idN into idN ⊗ϕ, so
det(ϕ ⊗ idN ) = det(idN ⊗ϕ), so
det(ϕ ⊗ ψ) = det(idN ⊗ϕ) det(idM ⊗ψ).
What are the determinants on the right side? Pick bases e1 , . . . , ek of M and e01 , . . . , e0` of
N . We will use the k` elementary tensors ei ⊗ e0j as a bases of M ⊗R N . Let [ϕ] be the
matrix of ϕ in the ordered basis e1 , . . . , ek . Since (ϕ ⊗ idN )(ei ⊗ e0j ) = ϕ(ei ) ⊗ e0j , let’s order
the basis of M ⊗R N as
e1 ⊗ e01 , . . . , ek ⊗ e01 , . . . , e1 ⊗ e0` , . . . , ek ⊗ e0` .
The k` × k` matrix for ϕ ⊗ idN in this ordered basis is the block diagonal matrix
[ϕ] O · · · O
O [ϕ] · · · O
.. ,
.. .. ..
. . . .
O O ··· [ϕ]
whose determinant is (det ϕ)` .
Thus
(2.1) det(ϕ ⊗ ψ) = (det ϕ)` (det ψ)k .
Note ` is the rank of the module on which ψ is defined and k is the rank of the module on
which ϕ is defined. In particular, in Example 2.2 we have det(A ⊗ A0 ) = (det A)2 (det A0 )2 .
Let’s review the idea in this proof. Since N ∼
= R` , M ⊗R N ∼ = M ⊗R R` ∼ = M ⊕` . Under
such an isomorphism, ϕ ⊗ idN becomes the `-fold direct sum ϕ ⊕ · · · ⊕ ϕ, which has a block
diagonal matrix representation in a suitable basis. So its determinant is (det ϕ)` .
Example 2.8. Taking M = N and ϕ = ψ, the tensor square ϕ⊗2 has determinant (det ϕ)2k .
Corollary 2.9. Let M be a free module of rank k ≥ 1 and ϕ : M → M be a linear map.
i−1
For every i ≥ 1, det(ϕ⊗i ) = (det ϕ)ik .
Proof. Use induction and associativity of the tensor product of linear maps.
Remark 2.10. Equation (2.1) in the setting of vector spaces and matrices says det(A⊗B) =
(det A)` (det B)k , where A is k × k, B is ` × `, and A ⊗ B = (aij B) is the matrix incarnation
of a tensor product of linear maps, called the Kronecker product of A and B at the end
of Example 2.2. While the label “Kronecker product” for the matrix A ⊗ B is completely
standard, it is not historically accurate. It is based on Hensel’s attribution of the formula
for det(A ⊗ B) to Kronecker, but the formula is due to Zehfuss. See [2].
Let’s see how the tensor product of linear maps behaves for isomorphisms, surjections,
and injections.
6 KEITH CONRAD
It is a fundamental feature of tensor products that if ϕ and ψ are both injective then
ϕ ⊗ ψ might not be injective. This can occur even if one of ϕ or ψ is the identity function.
Example 2.13. Taking R = Z, let α : Z/pZ → Z/p2 Z be multiplication by p: α(x) = px.
This is injective, and if we tensor with Z/pZ we get the linear map 1 ⊗ α : Z/pZ ⊗Z Z/pZ →
Z/pZ ⊗Z Z/p2 Z with the effect a ⊗ x 7→ a ⊗ px = pa ⊗ x = 0, so 1 ⊗ α is identically 0 and
its domain is Z/pZ ⊗Z Z/pZ ∼ = Z/pZ 6= 0, so 1 ⊗ α is not injective.
This provides an example where the natural linear map
HomR (M, M 0 ) ⊗R HomR (N, N 0 ) → HomR (M ⊗R N, M 0 ⊗R N 0 )
in Theorem 2.5 is not an isomorphism; R = Z, M = M 0 = N = Z/pZ, and N 0 = Z/p2 Z.
Because the tensor product of linear maps does not generally preserve injectivity, a tensor
has to be understood in context: it is a tensor in a specific tensor product M ⊗R N . If
M ⊂ M 0 and N ⊂ N 0 , it is generally false that M ⊗R N can be thought of as a submodule
of M 0 ⊗R N 0 since the natural map M ⊗R N → M 0 ⊗R N 0 might not be injective. We might
say it this way: a tensor product of submodules need not be a submodule.
Example 2.14. Since pZ ∼ = Z as abelian groups, by pn 7→ n, we have Z/pZ ⊗Z pZ ∼ =
∼
Z/pZ ⊗Z Z = Z/pZ as abelian groups by a ⊗ pn 7→ a ⊗ n 7→ na mod p. Therefore 1 ⊗ p in
Z/pZ ⊗Z pZ is nonzero, since the isomorphism identifies it with 1 in Z/pZ. However, 1 ⊗ p
in Z/pZ ⊗Z Z is 0, since 1 ⊗ p = p ⊗ 1 = 0 ⊗ 1 = 0. (This calculation with 1 ⊗ p doesn’t work
in Z/pZ ⊗Z pZ since we can’t bring p to the left side of ⊗ and leave 1 behind, as 1 6∈ pZ.)
It might seem weird that 1 ⊗ p is nonzero in Z/pZ ⊗Z pZ while 1 ⊗ p is zero in the “larger”
abelian group Z/pZ ⊗Z Z! The reason there isn’t a contradiction is that Z/pZ ⊗Z pZ is not
really a subgroup of Z/pZ ⊗Z Z even though pZ is a subgroup of Z. The inclusion mapping
i : pZ → Z gives us a natural mapping 1 ⊗ i : Z/pZ ⊗Z pZ → Z/pZ ⊗Z Z, with the effect
a ⊗ pn 7→ a ⊗ pn, but this is not an embedding. In fact its image is 0: in Z/pZ ⊗Z Z,
a ⊗ pn = pa ⊗ n = 0 ⊗ n = 0. The moral is that an elementary tensor a ⊗ pn means
something different in Z/pZ ⊗Z pZ and in Z/pZ ⊗Z Z.
ϕ⊗ψ
This example also shows the image of M ⊗R N −−−−→ M 0 ⊗R N 0 need not be isomorphic
to ϕ(M ) ⊗R ψ(N ), since 1 ⊗ i has image 0 and Z/pZ ⊗Z i(pZ) ∼
= Z/pZ ⊗Z Z ∼= Z/pZ.
Example 2.15. While Z/pZ ⊗Z Z ∼ = Z/pZ, if we enlarge the second tensor factor Z to Q
we get a huge collapse: Z/pZ ⊗Z Q = 0 since a ⊗ r = a ⊗ p(r/p) = pa ⊗ r/p = 0 ⊗ r/p = 0.
In particular, 1 ⊗ 1 is nonzero in Z/pZ ⊗Z Z but 1 ⊗ 1 = 0 in Z/pZ ⊗Z Q.
In terms of tensor products of linear mappings, this example says that tensoring the
inclusion i : Z ,→ Q with Z/pZ gives us a Z-linear map 1 ⊗ i : Z/pZ ⊗Z Z → Z/pZ ⊗Z Q
that is not injective: the domain is isomorphic to Z/pZ and the target is 0.
TENSOR PRODUCTS II 7
Example 2.16. Here is an example of a linear map f : M → N that is injective and its
tensor square f ⊗2 : M ⊗2 → N ⊗2 is not injective.
Let R = A[X, Y ] with A a nonzero commutative ring and I = (X, Y ). In R⊗2 , we have
(2.2) X ⊗ Y = XY (1 ⊗ 1) and Y ⊗ X = Y X(1 ⊗ 1) = XY (1 ⊗ 1),
so X ⊗Y = Y ⊗X. We will now show that in I ⊗2 , X ⊗Y 6= Y ⊗X. (The calculation in (2.2)
makes no sense in I ⊗2 since 1 is not an element of I.) To show two tensors are not equal, the
best approach is to construct a linear map from the tensor product space that has different
values at the two tensors. The function I × I → A given by (f, g) 7→ fX (0, 0)gY (0, 0), where
fX and gY are partial derivatives of f and g with respect to X and Y , is R-bilinear. (Treat
the target A as an R-module through multiplication by the constant term of polynomials in
R, or just view A as R/I with ordinary multiplication by R.) Thus there is an R-linear map
I ⊗2 → A sending any elementary tensor f ⊗ g to fX (0, 0)gY (0, 0). In particular, X ⊗ Y 7→ 1
and Y ⊗ X 7→ 0, so X ⊗ Y 6= Y ⊗ X in I ⊗2 .
It might seem weird that X ⊗ Y and Y ⊗ X are equal in R⊗2 but are not equal in I ⊗2 ,
even though I ⊂ R. The point is that we have to be careful when we want to think about
a tensor t ∈ I ⊗R I as a tensor in R ⊗R R. Letting i : I ,→ R be the inclusion map, thinking
about t in R⊗2 really means looking at i⊗2 (t), where i⊗2 : I ⊗2 → R⊗2 . For the tensor
t = X ⊗ Y − Y ⊗ X in I ⊗2 , we computed above that t 6= 0 but i⊗2 (t) = 0, so i⊗2 is not
injective even though i is injective. In other words, the natural way to think of I ⊗R I
“inside” R ⊗R R is actually not an embedding. For polynomials f and g in I, you have to
distinguish between the tensor f ⊗ g in I ⊗R I and the tensor f ⊗ g in R ⊗R R.
Generalizing this, let R = A[X1 , . . . , Xn ] where n ≥ 2 and I = (X1 , . . . , Xn ). The
inclusion i : I ,→ R is injective but the nth tensor power (as R-modules) i⊗n : I ⊗n → R⊗n
is not injective because the tensor
X
t := (sign σ)Xσ(1) ⊗ · · · ⊗ Xσ(n) ∈ I ⊗n
σ∈Sn
⊗n , which is 0, but t is not 0 in
P
gets sent to σ∈Sn (sign σ)X1 · · · Xn (1 ⊗ · · · ⊗ 1) in R
I ⊗n because there is an R-linear map I ⊗n → A sending t to 1: use a product of partial
derivatives at (0, 0, . . . , 0), as in the n = 2 case.
Remark 2.17. The ideal I = (X, Y ) in R = A[X, Y ] from Example 2.16 has another
interesting feature when A is a domain: it is a torsion-free R-module but I ⊗2 is not:
X(X ⊗ Y ) = X ⊗ XY = Y (X ⊗ X) and X(Y ⊗ X) = XY ⊗ X = Y (X ⊗ X), so in I ⊗2 we
have X(X ⊗ Y − Y ⊗ X) = 0, but X ⊗ Y − Y ⊗ X 6= 0. Similarly, Y (X ⊗ Y − Y ⊗ X) = 0.
Therefore a tensor product of torsion-free modules (even over a domain) need not be torsion-
free.
While we have just seen a tensor power of an injective linear map need not be injective,
here is a condition where injectivity holds.
Theorem 2.18. Let ϕ : M → N be injective and ϕ(M ) be a direct summand of N . For
k ≥ 0, ϕ⊗k : M ⊗k → N ⊗k is injective and the image is a direct summand of N ⊗k .
Proof. Write N = ϕ(M ) ⊕ P . Let π : N M by π(ϕ(m) + p) = m, so π is linear and
π ◦ ϕ = idM . Then ϕ⊗k : M ⊗k → N ⊗k and π ⊗k : N ⊗k M ⊗k are linear maps and
π ⊗k ◦ ϕ⊗k = (π ◦ ϕ)⊗k = id⊗k
M = idM ⊗k ,
8 KEITH CONRAD
so ϕ⊗k has a left inverse. That implies ϕ⊗k is injective and M ⊗k is isomorphic to a direct
summand of N ⊗k by criteria for when a short exact sequence of modules splits.
We can apply this to vector spaces: if V is a vector space and W is a subspace, there
is a direct sum decomposition V = W ⊕ U (U is non-canonical), so tensor powers of the
inclusion W → V are injective linear maps W ⊗k → V ⊗k .
Other criteria for a tensor power of an injective linear map to be injective will be met in
Corollary 3.13 and Theorem 4.9.
ϕ⊗ψ
We will now compute the kernel of M ⊗R N −−−−→ M 0 ⊗R N 0 in terms of the kernels of
ϕ and ψ, assuming ϕ and ψ are onto.
ϕ ψ
Theorem 2.19. Let M −−→ M 0 and N −−→ N 0 be R-linear and surjective. The kernel
ϕ⊗ψ
of M ⊗R N −−−−→ M 0 ⊗R N 0 is the submodule of M ⊗R N spanned by all m ⊗ n where
ϕ(m) = 0 or ψ(n) = 0. That is, intuitively
ker(ϕ ⊗ ψ) = (ker ϕ) ⊗R N + M ⊗R (ker ψ),
i j
while rigorously in terms of the inclusion maps ker ϕ −−→ M and ker ψ −−→ N ,
ker(ϕ ⊗ ψ) = (i ⊗ 1)((ker ϕ) ⊗R N ) + (1 ⊗ j)(M ⊗R (ker ψ)).
The reason (ker ϕ)⊗R N +M ⊗R (ker ψ) is only an intuitive formula for the kernel of ϕ⊗ψ
is that, strictly speaking, these tensor product modules are not submodules of M ⊗R N .
Only after applying i ⊗ 1 and 1 ⊗ j to them – and these might not be injective – do those
modules become submodules of M ⊗R N .
Proof. Both (i⊗1)((ker ϕ)⊗R N ) and (1⊗j)(M ⊗R (ker ψ)) are killed by ϕ⊗ψ: if m ∈ ker ϕ
and n ∈ N then (ϕ ⊗ ψ)((i ⊗ 1)(m ⊗ n)) = (ϕ ⊗ ψ)(m ⊗ n) = ϕ(m) ⊗ ψ(n) = 0 since1
ϕ(m) = 0. Similarly (1 ⊗ j)(m ⊗ n) is killed by ϕ ⊗ ψ if m ∈ M and n ∈ ker ψ. Set
U = (i ⊗ 1)((ker ϕ) ⊗R N ) + (1 ⊗ j)(M ⊗R (ker ψ)),
so U ⊂ ker(ϕ ⊗ ψ), which means ϕ ⊗ ψ induces a linear map
Φ : (M ⊗R N )/U → M 0 ⊗R N 0
where Φ(m ⊗ n mod U ) = (ϕ ⊗ ψ)(m ⊗ n) = ϕ(m) ⊗ ψ(n). We will now write down an
inverse map, which proves Φ is injective, so the kernel of ϕ ⊗ ψ is U .
Because ϕ and ψ are assumed to be onto, every elementary tensor in M 0 ⊗R N 0 has the
form ϕ(m) ⊗ ψ(n). Knowing ϕ(m) and ψ(n) only determines m and n up to addition by
elements of ker ϕ and ker ψ. For m0 ∈ ker ϕ and n0 ∈ ker ψ,
(m + m0 ) ⊗ (n + n0 ) = m ⊗ n + m0 ⊗ n + m ⊗ n0 + m0 ⊗ n0 ∈ m ⊗ n + U,
so the function M 0 × N 0 → (M ⊗R N )/U defined by (ϕ(m), ψ(n)) 7→ m ⊗ n mod U is well-
defined. It is R-bilinear, so we have an R-linear map Ψ : M 0 ⊗R N 0 → (M ⊗R N )/U where
Ψ(ϕ(m) ⊗ ψ(n)) = m ⊗ n mod U on elementary tensors.
Easily the linear maps Φ◦Ψ and Ψ◦Φ fix spanning sets, so they are both the identity.
Remark 2.20. If we remove the assumption that ϕ and ψ are onto, Theorem 2.19 does not
correctly compute the kernel. For example, if ϕ and ψ are both injective then the formula
for the kernel in Theorem 2.19 is 0, and we know ϕ ⊗ ψ need not be injective.
1The first m ⊗ n is in (ker ϕ) ⊗ N , while the second m ⊗ n is in M ⊗ N .
R R
TENSOR PRODUCTS II 9
Unlike the kernel computation in Theorem 2.19, it is not easy to describe the torsion
submodule of a tensor product in terms of the torsion submodules of the original modules.
While (M ⊗R N )tor contains (i⊗1)(Mtor ⊗R N )+(1⊗j)(M ⊗R Ntor ), with i : Mtor → M and
j : Ntor → N being the inclusions, it is not true that this is all of (M ⊗R N )tor , since M ⊗R N
can have nonzero torsion when M and N are torsion-free (so Mtor = 0 and Ntor = 0). We
saw this at the end of Example 2.16.
ϕ ψ
Corollary 2.21. If M −−→ M 0 is an isomorphism of R-modules and N −−→ N 0 is surjec-
ϕ⊗ψ
tive, then the linear map M ⊗R N −−−−→ M 0 ⊗R N 0 has kernel (1 ⊗ j)(M ⊗R (ker ψ)), where
j
ker ψ −−→ N is the inclusion.
Proof. This is immediate from Theorem 2.19 since ker ϕ = 0.
Corollary 2.22. Let f : R → S be a homomorphism of commutative rings and M ⊂ N as
i
R-modules, with M −−→ N the inclusion map. The following are equivalent:
1⊗i
(1) S ⊗R M −−−→ S ⊗R N is onto.
(2) S ⊗R (N/M ) = 0.
π 1⊗i
Proof. Let N −−→ N/M be the reduction map, so we have the sequence S ⊗R M −−−→
1⊗π
S ⊗R N −−−−→ S ⊗R (N/M ). The map 1 ⊗ π is onto, and ker π = M , so ker(1 ⊗ π) =
(1 ⊗ i)(S ⊗R M ). Therefore 1 ⊗ i is onto if and only if ker(1 ⊗ π) = S ⊗R N if and only if
1 ⊗ π = 0, and since 1 ⊗ π is onto we have 1 ⊗ π = 0 if and only if S ⊗R (N/M ) = 0.
Example 2.23. If M ⊂ N and N is finitely generated, we show M = N if and only if the
1⊗i
natural map R/m ⊗R M −−−→ R/m ⊗R N is onto for all maximal ideals m in R, where
i
M −−→ N is the inclusion map. The “only if” direction is clear. In the other direction, if
1⊗i
R/m ⊗R M −−−→ R/m ⊗R N is onto then R/m ⊗R (N/M ) = 0 by Corollary 2.22. Since N
is finitely generated, so is N/M , and we are reduced to showing R/m ⊗R (N/M ) = 0 for all
maximal ideals m if and only if N/M = 0. When P is a finitely generated module, P = 0
if and only if P/mP = 0 for all maximal ideals2 m in R, so we can apply this to P = N/M
since P/mP ∼ = R/m ⊗R P .
Corollary 2.24. Let f : R → S be a homomorphism of commutative rings and I be an
ideal in R[X1 , . . . , Xn ]. Write I · S[X1 , . . . , Xn ] for the ideal generated by the image of I in
S[X1 , . . . , Xn ]. Then
S ⊗R R[X1 , . . . , Xn ]/I ∼
= S[X1 , . . . , Xn ]/(I · S[X1 , . . . , Xn ]).
as S-modules by s ⊗ h mod I 7→ sh mod I · S[X1 , . . . , Xn ].
Proof. The identity S → S and the natural reduction R[X1 , . . . , Xn ] R[X1 , . . . , Xn ]/I
are both onto, so the tensor product of these R-linear maps is an R-linear surjection
(2.3) S ⊗R R[X1 , . . . , Xn ] S ⊗R (R[X1 , . . . , Xn ]/I)
and the kernel is (1 ⊗ j)(S ⊗R I) by Theorem 2.19, where j : I → R[X1 , . . . , Xn ] is the
inclusion. Under the natural R-module isomorphism
(2.4) S ⊗R R[X1 , . . . , Xn ] ∼
= S[X1 , . . . , Xn ],
2P/mP = 0 ⇒ P = mP ⇒ P = mP ⇒ P = 0 by Nakayama’s lemma. From P = 0 for all maximal
m m m m
ideals m, P = 0: for all x ∈ P , x = 0 in Pm implies ax = 0 in P for some a ∈ R − m. Thus AnnR (x) is not
in any maximal ideal of R, so AnnR (x) = R and thus x = 1 · x = 0.
10 KEITH CONRAD
(1 ⊗ j)(S ⊗R I) on the left side corresponds to I · S[X1 , . . . , Xn ] on the right side, so (2.3)
and (2.4) say
S[X1 , . . . , Xn ]/(I · S[X1 , . . . , Xn ]) ∼
= S ⊗R (R[X1 , . . . , Xn ]/I).
as R-modules. The left side is naturally an S-module and the right side is too using extension
of scalars. It is left to the reader to check the isomorphism is S-linear.
Example 2.25. For h(X) ∈ Z[X], Q ⊗Z Z[X]/(h(X)) ∼
= Q[X]/(h(X)) as Q-vector spaces
and Z/mZ ⊗Z Z[X]/(h(X)) = (Z/mZ)[X]/(h(X)) as Z/mZ-modules where m > 1.
3. Flat Modules
Because a tensor product of injective linear maps might not be injective, it is important
to give a name to those R-modules N that always preserve injectivity, in the sense that
ϕ 1⊗ϕ
M −−→ M 0 being injective implies N ⊗R M −−−−→ N ⊗R M 0 is injective. (Notice the map
on N is the identity.)
ϕ
Definition 3.1. An R-module N is called flat if for all injective linear maps M −−→ M 0
1⊗ϕ
the linear map N ⊗R M −−−−→ N ⊗R M 0 is injective.
The concept of a flat module is pointless unless one has some good examples. The next
two theorems provide some.
Theorem 3.2. Any free R-module F is flat: if the linear map ϕ : M → M 0 is injective,
then 1 ⊗ ϕ : F ⊗R M → F ⊗R M 0 is injective.
Proof. When F = 0 it is clear, so take F 6= 0 with basis {ei }i∈I . From ourPprevious
development of the tensor product, every element of F ⊗R M can be written as i ei ⊗ mi
for a unique choice of mi ∈ M , and similarly
P for F ⊗R M 0 .
For t ∈ ker(1 ⊗ ϕ), we can write t = i ei ⊗ mi with mi ∈ M . Then
X
0 = (1 ⊗ ϕ)(t) = ei ⊗ ϕ(mi ),
i
M 0,
in F ⊗RP which forces each ϕ(mi ) to be 0. So every mi is 0, since ϕ is injective, and we
get t = i ei ⊗ 0 = 0.
Note that in Theorem 3.2 we did not need to assume F has a finite basis.
Theorem 3.3. Let R be a domain and K be its fraction field. As an R-module, K is flat.
This is not a special case of the previous theorem: if K were a free R-module then3
K = R, so whenever R is a domain that is not a field (e.g., R = Z) the fraction field of R
is a flat R-module that is not a free R-module.
ϕ
Proof. Let M −−→ M 0 be an injective linear map of R-modules. Every tensor in K ⊗R M is
elementary (use common denominators in K) and an elementary tensor in K ⊗R M is 0 if
and only if its first factor is 0 or its second factor is torsion. (Here we are using properties
of K ⊗R M proved in part I.)
3Any two nonzero elements of K are R-linearly dependent, so if K were a free R-module then it would
have a basis of size 1: K = Rx for some x ∈ K. Therefore x2 = rx for some r ∈ R, so x = r ∈ R, which
implies K ⊂ R, so K = R.
TENSOR PRODUCTS II 11
M
ϕ
/ M0
K ⊗R M / K ⊗R M 0
1⊗ϕ
Example 3.4. The natural inclusion Z ,→ Z/3Z ⊕ Z is Z-linear and injective. Applying
Q⊗Z to both sides and using properties of tensor products turns this into the identity map
Q → Q, which is also injective.
Remark 3.5. Theorem 3.3 generalizes: for any commutative ring R and multiplicative set
D in R, the localization RD is a flat R-module.
Theorem 3.6. If M is a flat R-module and I is an ideal in R then I ⊗R M ∼
= IM by
i ⊗ m 7→ im.
Proof. The inclusion I → R is injective. Applying ⊗R M to this makes an injective R-linear
map I ⊗R M → R ⊗R M since M is flat, and composing with the isomorphism R ⊗R M ∼ =M
where r ⊗ m 7→ rm makes the injective map I ⊗R M → M where i ⊗ m 7→ im. The image
is IM , so I ⊗R M ∼= IM as R-modules with the desired effect on elementary tensors.
To say an R-module N is not flat means there is some example of an injective linear map
ϕ 1⊗ϕ
M −−→ M 0 whose induced linear map N ⊗R M −−−−→ N ⊗R M 0 is not injective.
Example 3.7. For a nonzero torsion abelian group A, the natural map Z ,→ Q is injective
but if we apply A⊗Z we get the map A → 0, which is not injective, so A is not a flat
Z-module. This includes nonzero finite abelian groups and the infinite abelian group Q/Z.
Remark 3.8. Since Q/Z is not flat as a Z-module, for a homomorphism of abelian groups
f 1⊗f
G −−→ G0 the kernel of Q/Z ⊗Z G −−−−→ Q/Z ⊗Z G0 need not be Q/Z ⊗ ker f but could be
larger. Therefore it is not easy to determine the kernel of a group homomorphism after base
extension by Q/Z. Failure to take this into account created a gap in a proof of a widely
used theorem in knot theory. See [1, p. 927].
Example 3.9. If R is a domain with fraction field K, any nonzero torsion R-module T
(meaning every element of T is killed by a nonzero element of R) is not a flat R-module
since tensoring the inclusion R ,→ K with T produces the R-linear map T → 0, which is
not injective. In particular, the quotient module K/R is not a flat R-module. The previous
example is the special case R = Z: Q/Z is not a flat Z-module.
12 KEITH CONRAD
ϕ
Theorem 3.10. If N is a flat R-module and M −−→ M 0 is R-linear then the kernel of
1⊗ϕ
N ⊗R M −−−−→ N ⊗R M 0 is N ⊗R ker ϕ, viewed as a submodule of N ⊗R M in a natural
way.
Proof. The diagram
ϕ(M )
<
m7→ϕ(m) i
#
M / M0
ϕ
The map 1 ⊗ i is injective since i is injective and N is flat. Therefore the two maps out
of N ⊗R M above have the same kernel. The kernel of N ⊗R M −→ N ⊗R ϕ(M ) can be
computed by Corollary 2.21 to be the natural image of N ⊗R ker ϕ inside N ⊗R M , and we
can identify the image with N ⊗R ker ϕ since N is flat.
Theorem 3.11. A tensor product of two flat modules is flat.
ϕ
Proof. Let N and N 0 be flat. For any injective linear map M −−→ M 0 , we want to show the
1⊗ϕ
induced linear map (N ⊗R N 0 ) ⊗R M −−−−→ (N ⊗R N 0 ) ⊗ M 0 is injective.
1⊗ϕ
Since N 0 is flat, N 0 ⊗R M −−−−→ N 0 ⊗R M 0 is injective. Tensoring now with N , N ⊗R
1⊗(1⊗ϕ)
(N 0 ⊗R M ) −−−−−−−→ N ⊗R (N 0 ⊗R M 0 ) is injective since N is flat. The diagram
1⊗(1⊗ϕ)
N ⊗R (N 0 ⊗R M ) / N ⊗R (N 0 ⊗ M 0 )
1⊗ϕ
(N ⊗R N 0 ) ⊗R M / (N ⊗R N 0 ) ⊗ M 0
commutes, where the vertical maps are the natural isomorphisms, so the bottom map is
injective. Thus N ⊗R N 0 is flat.
ϕ ψ
Theorem 3.12. Let M −−→ M 0 and N −−→ N 0 be injective linear maps. If the four
ϕ⊗ψ
modules are all flat then M ⊗R N −−−−→ M 0 ⊗R N 0 is injective. More precisely, if M and
N 0 are flat, or M 0 and N are flat, then ϕ ⊗ ψ is injective.
The precise hypotheses (M and N 0 flat, or M 0 and N flat) can be remembered using
dotted lines in the diagram below; if both modules connected by one of the dotted lines are
flat, then ϕ ⊗ ψ is injective.
ϕ
M / M0
N / N0
ψ
TENSOR PRODUCTS II 13
Proof. The trick is to break up the linear map ϕ ⊗ ψ into a composite of linear maps ϕ ⊗ 1
and 1 ⊗ ψ in the following commutative diagram.
M8 ⊗R N 0
1⊗ψ ϕ⊗1
'
M ⊗R N / M 0 ⊗R N 0
ϕ⊗ψ
Both ϕ ⊗ 1 and 1 ⊗ ψ are injective since N 0 and M are flat, so their composite ϕ ⊗ ψ
is injective. Alternatively, we can write ϕ ⊗ ψ as a composite fitting in the commutative
diagram
M8 0 ⊗R N
ϕ⊗1 1⊗ψ
'
M ⊗R N / M 0 ⊗R N 0
ϕ⊗ψ
and the two diagonal maps are injective from flatness of N and M 0 , so ϕ⊗ψ is injective.
Corollary 3.13. Let M1 , . . . , Mk , N1 , . . . , Nk be flat R-modules and ϕi : Mi → Ni be injec-
tive linear maps. Then the linear map
ϕ1 ⊗ · · · ⊗ ϕk : M1 ⊗R · · · ⊗R Mk → N1 ⊗ · · · ⊗ Nk
is injective. In particular, if ϕ : M → N is an injective linear map of flat modules then the
tensor powers ϕ⊗k : M ⊗k → N ⊗k are injective for all k ≥ 1.
Proof. We argue by induction on k. For k = 1 there is nothing to show. Suppose k ≥ 2 and
ϕ1 ⊗ · · · ⊗ ϕk−1 is injective. Then break up ϕ1 ⊗ · · · ⊗ ϕk into the composite
(N1 ⊗R · · · ⊗R Nk−1 ) ⊗R Mk
3
(ϕ1 ⊗···⊗ϕk−1 )⊗1 1⊗ϕk
*
M1 ⊗R · · · ⊗R Mk−1 ⊗R Mk / N1 ⊗R · · · ⊗R Nk .
ϕ1 ⊗···⊗ϕk
The first diagonal map is injective because Mk is flat, and the second diagonal map is
injective because N1 ⊗R · · · ⊗R Nk−1 is flat (Theorem 3.11 and induction).
Corollary 3.14. If M and N are free R-modules and ϕ : M → N is an injective linear
map, any tensor power ϕ⊗k : M ⊗k → N ⊗k is injective.
Proof. Free modules are flat by Theorem 3.2.
Note the free modules in Corollary 3.14 are completely arbitrary. We make no assump-
tions about finite bases.
Corollary 3.14 is not a special case of Theorem 2.18 because a free submodule of a free
module need not be a direct summand (e.g., 2Z is not a direct summand of Z).
Corollary 3.15. If M is a free module and {m1 , . . . , md } is a finite linearly independent
subset then for any k ≤ d the dk elementary tensors
(3.1) mi1 ⊗ · · · ⊗ mik where i1 , . . . , ik ∈ {1, 2, . . . , d}
are linearly independent in M ⊗k .
14 KEITH CONRAD
free, the kth tensor power (Rd )⊗k → M ⊗k is injective. This map sends the basis
ei1 ⊗ · · · ⊗ eik
of (Rd )⊗k ,
where i1 , . . . , ik ∈ {1, 2, . . . , d}, to the elementary tensors in (3.1), so they are
linearly independent in M ⊗k
Corollary 3.15 is not saying the elementary tensors in (3.1) can be extended to a basis of
M ⊗k , just as m1 , . . . , md usually can’t be extended to a basis of M .
Proof. Tensoring the displayed R-module isomorphism by K gives a K-vector space iso-
morphism K ⊗R M ∼ = K d since K ⊗R (R/(ai )) = 0. Thus d = dimK (K ⊗R M ).
Example 4.3. Let A = ( ac db ) ∈ M2 (R). Regarding A as a linear map R2 → R2 , its base
extension AS : S ⊗R R2 → S ⊗R R2 is S-linear and S ⊗R R2 ∼ = S 2 as S-modules.
Let {e1 , e2 } be the standard basis of R2 . An S-basis for S ⊗R R2 is {1 ⊗ e1 , 1 ⊗ e2 }. Using
this basis, we can compute a matrix for AS :
AS (1 ⊗ e1 ) = 1 ⊗ A(e1 ) = 1 ⊗ (ae1 + ce2 ) = a(1 ⊗ e1 ) + c(1 ⊗ e2 )
and
AS (1 ⊗ e2 ) = 1 ⊗ A(e2 ) = 1 ⊗ (be1 + de2 ) = b(1 ⊗ e1 ) + d(1 ⊗ e2 ).
Therefore the matrix for AS is ( ac db ) ∈ M2 (S). (Strictly speaking, we should have entries
f (a), f (b), and so on.)
The next theorem says base extension doesn’t change matrix representations, as in the
previous example.
ϕ
Theorem 4.4. Let M and M 0 be nonzero finite-free R-modules and M −−→ M 0 be an
R-linear map. For any bases {ej } and {e0i } of M and M 0 , the matrix for the S-linear map
ϕS
S ⊗R M −−− → S ⊗R M 0 with respect to the bases {1 ⊗ ej } and {1 ⊗ e0i } equals the matrix for
ϕ with respect to {ej } and {e0i }.
Proof. Say ϕ(ej ) = i aij e0i , so the matrix of ϕ is (aij ). Then
P
X X
ϕS (1 ⊗ ej ) = 1 ⊗ ϕ(ej ) = 1 ⊗ aij ei = aij (1 ⊗ ei ),
i i
such that L(s⊗ϕ) = sϕS . The map L is S-linear (check!). If M 0 = R and we identify S ⊗R R
with S as S-modules by multiplication, then L becomes an S-linear map S ⊗R (M ∨R ) →
(S ⊗R M )∨S .
Now suppose M and M 0 are both finite free. We want to show L is an isomorphism. If
M or M 0 is 0 it is clear, so we may take them both to be nonzero with respective R-bases
{ei } and {e0j }, say. Then S-bases of S ⊗R M and S ⊗R M 0 are {1 ⊗ ei } and {1 ⊗ e0j }. An
R-basis of HomR (M, M 0 ) is the functions ϕij sending ei to e0j and other basis vectors ek of
M to 0. An S-basis of S ⊗R HomR (M, M 0 ) is the tensors {1 ⊗ ϕij }, and
L(1 ⊗ ϕij )(1 ⊗ ei ) = (ϕij )S (1 ⊗ ei ) = (1 ⊗ ϕij )(1 ⊗ ei ) = 1 ⊗ ϕij (ei ) = 1 ⊗ e0j
while L(1 ⊗ ϕij )(1 ⊗ ek ) = 0 for k 6= i. That means L sends a basis to a basis, so L is an
isomorphism.
Example 4.7. Taking R = R, S = C, M = Rn , and M 0 = Rm , the theorem says
C ⊗R Mm,n (R) ∼ = Mm,n (C) as complex vector spaces by sending the elementary tensor
z ⊗ A for z ∈ C and A ∈ Mm,n (R) to the matrix zA. In particular, C ⊗R Mn (R) ∼
= Mn (C).
Example 4.8. Let R = Z/p2 Z and S = Z/pZ. Take M = Z/pZ as an R-module. The
linear map
S ⊗R M ∨R → (S ⊗R M )∨S
has image 0, but the right side is isomorphic to HomZ/pZ (Z/pZ, Z/pZ) ∼
= Z/pZ, so this
map is not an isomorphism.
In Corollary 3.14 we saw the tensor powers of an injective linear map between free
modules over any commutative ring are all injective. Using base extension, we can drop the
requirement that the target module be free provided we are working over a domain.
Theorem 4.9. Let R be a domain and ϕ : M ,→ N be an injective linear map where M is
free. Then ϕ⊗k : M ⊗k → N ⊗k is injective for any k ≥ 1.
Proof. We have a commutative diagram
ϕ⊗k
M ⊗k / N ⊗k
1⊗(ϕ⊗k )
K ⊗R M ⊗k / K ⊗R N ⊗k
(1⊗ϕ)⊗k
(K ⊗R M )⊗k / (K ⊗R N )⊗k
where the top vertical maps are the natural ones (t 7→ 1 ⊗ t) and the bottom vertical maps
are the base extension isomorphisms. (Tensor powers along the bottom are over K while
those on the first and second rows are over R.) From commutativity, to show ϕ⊗k along
the top is injective it suffices to show the composite map along the left side and the bottom
1⊗ϕ
is injective. The K-linear map map K ⊗R M −−−−→ K ⊗R N is injective since K is a
flat R-module, and therefore the map along the bottom is injective (tensor products of
injective linear maps of vector spaces are injective). The bottom vertical map on the left
is an isomorphism. The top vertical map on the left is injective since M ⊗k is free and thus
torsion-free (R is a domain).
This theorem may not be true if M isn’t free. Look at Example 2.16.
TENSOR PRODUCTS II 17
5. Vector Spaces
Because all (nonzero) vector spaces have bases, the results we have discussed for modules
assume a simpler form when we are working with vector spaces. We will review what we
have done in the setting of vector spaces and then discuss some further special properties
of this case.
Let K be a field. Tensor products of K-vector spaces involve no unexpected collapsing: if
V and W are nonzero K-vector spaces then V ⊗K W is nonzero and in fact dimK (V ⊗K W ) =
dimK (V ) dimK (W ) in the sense of cardinal numbers.
ϕ ψ
For any K-linear maps V −−→ V 0 and W −−→ W 0 , we have the tensor product linear
ϕ⊗ψ ϕ
map V ⊗K W −−−−→ V 0 ⊗K W 0 that sends v ⊗ w to ϕ(v) ⊗ ψ(w). When V −−→ V 0 and
ψ ϕ⊗ψ
W −−→ W 0 are isomorphisms or surjective, so is V ⊗K W −−−−→ V 0 ⊗K W 0 (Theorems
2.11 and 2.12). Moreover, because all K-vector spaces are free a tensor product of injective
K-linear maps is injective (Theorem 3.2).
ϕ
Example 5.1. If V −−→ W is an injective K-linear map and U is any K-vector space, the
1⊗ϕ
K-linear map U ⊗K V −−−−→ U ⊗K W is injective.
Example 5.2. A tensor product of subspaces “is” a subspace: if V ⊂ V 0 and W ⊂ W 0 the
natural linear map V ⊗K W → V 0 ⊗K W 0 is injective.
Because of this last example, we can treat a tensor product of subspaces as a subspace
ϕ ψ
of the tensor product. For example, if V −−→ V 0 and W −−→ W 0 are linear then ϕ(V ) ⊂ V 0
and ψ(W ) ⊂ W 0 , so we can regard ϕ(V ) ⊗K ψ(W ) as a subspace of V 0 ⊗K W 0 , which we
couldn’t do with modules in general. The following result gives us some practice with this
viewpoint.
Theorem 5.3. Let V ⊂ V 0 and W ⊂ W 0 where V 0 and W 0 are nonzero. Then V ⊗K W =
V 0 ⊗K W 0 if and only if V = V 0 and W = W 0 .
Proof. Since V ⊗K W is inside both V ⊗K W 0 and V 0 ⊗K W , which are inside V 0 ⊗K W 0 ,
by reasons of symmetry it suffices to assume V ( V 0 and show V ⊗K W 0 ( V 0 ⊗K W 0 .
Since V is a proper subspace of V 0 , there is a linear functional ϕ : V 0 → K that vanishes
on V and is not identically 0 on V 0 , so ϕ(v00 ) = 1 for some v00 ∈ V 0 . Pick nonzero ψ ∈ W 0∨ ,
and say ψ(w00 ) = 1. Then the linear function V 0 ⊗K W 0 → K where v 0 ⊗ w0 7→ ϕ(v 0 )ψ(w0 )
vanishes on all of V ⊗K W 0 by checking on elementary tensors but its value on v00 ⊗ w00 is
1. Therefore v00 ⊗ w00 6∈ V ⊗K W 0 , so V ⊗K W 0 ( V 0 ⊗K W 0 .
When V and W are finite-dimensional, the K-linear map
(5.1) HomK (V, V 0 ) ⊗K HomK (W, W 0 ) → HomK (V ⊗K W, V 0 ⊗K W 0 )
sending the elementary tensor ϕ ⊗ ψ to the linear map denoted ϕ ⊗ ψ is an isomorphism
(Theorem 2.5). So the two possible meanings of ϕ⊗ψ (elementary tensor in a tensor product
of Hom-spaces or linear map on a tensor product of vector spaces) really match up. Taking
V 0 = K and W 0 = K in (5.1) and identifying K ⊗K K with K by multiplication, (5.1) says
V ∨ ⊗K W ∨ ∼= (V ⊗K W )∨ using the obvious way of making a tensor ϕ ⊗ ψ in V ∨ ⊗K W ∨
act on V ⊗K W , namely through multiplication of the values: (ϕ ⊗ ψ)(v ⊗ w) = ϕ(v)ψ(w).
By induction on the numbers of terms,
V ∨ ⊗K · · · ⊗K V ∨ ∼
1 k= (V1 ⊗K · · · ⊗K Vk )∨
18 KEITH CONRAD
is an isomorphism (Theorem 4.6). When we choose K-bases for V and W and use the
corresponding L-bases for L ⊗K V and L ⊗K W , the matrix representations of a K-linear
map V → W and its base extension by L are the same (Theorem 4.4). Taking W = K, the
natural L-linear map
(5.3) L ⊗K V ∨ ∼
= (L ⊗K V )∨
is an isomorphism for finite-dimensional V , using K-duals on the left and L-duals on the
right.5
Remark 5.4. We don’t really need L to be a field; K-vector spaces are free and therefore
their base extensions to modules over any commutative ring containing K will be free
as modules over the larger ring. For example, the characteristic polynomial of a linear
ϕ
operator V −−→ V could be defined in a coordinate-free way using base extension of V from
K to K[T ]: the characteristic polynomial of ϕ is the determinant of the linear operator
T ⊗ idV −ϕK[T ] : K[T ] ⊗K V −→ K[T ] ⊗K V since det(T ⊗ idV −ϕK[T ] ) = det(T In − A),
where A is a matrix representation of ϕ.
We will make no finite-dimensionality assumptions in the rest of this section.
The next theorem tells us the image and kernel of a tensor product of linear maps of
vector spaces, with no surjectivity hypotheses as in Theorem 2.19.
ϕ1 ϕ2
Theorem 5.5. Let V1 −−−→ W1 and V2 −−−→ W2 be linear. Then
ker(ϕ1 ⊗ ϕ2 ) = ker ϕ1 ⊗K V2 + V1 ⊗K ker ϕ2 , Im(ϕ1 ⊗ ϕ2 ) = ϕ1 (V1 ) ⊗K ϕ2 (V2 ).
In particular, if V1 and V2 are nonzero then ϕ1 ⊗ ϕ2 is injective if and only if ϕ1 and ϕ2
are injective, and if W1 and W2 are nonzero then ϕ1 ⊗ ϕ2 is surjective if and only if ϕ1 and
ϕ2 are surjective.
Here we are taking advantage of the fact that in vector spaces a tensor product of sub-
spaces is naturally a subspace of the tensor product: ker ϕ1 ⊗K V2 can be identified with
its image in V1 ⊗K V2 and ϕ1 (V1 ) ⊗K ϕ2 (V2 ) can be identified with its image in W1 ⊗K W2
under the natural maps. Theorem 2.19 for modules has weaker conclusions (e.g., injectivity
of ϕ1 ⊗ ϕ2 doesn’t imply injectivity of ϕ1 and ϕ2 ).
4We allow infinite or even non-algebraic extensions, such as R/Q.
5If we drop finite-dimensionality assumptions, (5.1), (5.2), and (5.3) are all still injective but generally
not surjective.
TENSOR PRODUCTS II 19
# #
V1 / W1 V2 / W2
ϕ1 ϕ2
(
V1 ⊗K V2 / W1 ⊗K W2
ϕ1 ⊗ϕ2
commutes. As i1 ⊗ i2 is injective, both maps out of V1 ⊗K V2 have the same kernel. The
kernel of the map V1 ⊗K V2 −→ ϕ1 (V1 ) ⊗K ϕ2 (V2 ) can be computed by Theorem 2.19 to be
ker ϕ1 ⊗K V2 + V1 ⊗K ker ϕ2 , where we identify tensor products of subspaces with a subspace
of the tensor product.
If ϕ1 ⊗ ϕ2 is injective then its kernel is 0, so 0 = ker ϕ1 ⊗K V2 + V1 ⊗K ker ϕ2 from the
kernel formula. Therefore the subspaces ker ϕ1 ⊗K V2 and V1 ⊗K ker ϕ2 both vanish, so
ker ϕ1 and ker ϕ2 must vanish (because V2 and V1 are nonzero, respectively). Conversely,
if ϕ1 and ϕ2 are injective then we already knew ϕ1 ⊗ ϕ2 is injective, but the formula for
ker(ϕ1 ⊗ ϕ2 ) also shows us this kernel is 0.
If ϕ1 ⊗ϕ2 is surjective then the formula for its image shows ϕ1 (V1 )⊗K ϕ2 (V2 ) = W1 ⊗K W2 ,
so ϕ1 (V1 ) = W1 and ϕ2 (V2 ) = W2 by Theorem 5.3 (here we need W1 and W2 nonzero).
Conversely, if ϕ1 and ϕ2 are surjective then so is ϕ1 ⊗ ϕ2 because that’s true for all modules.
Corollary 5.6. Let V ⊂ V 0 and W ⊂ W 0 . Then
(V 0 ⊗K W 0 )/(V ⊗K W 0 + V 0 ⊗K W ) ∼
= (V 0 /V ) ⊗K (W 0 /W ).
π π
Proof. Tensor the natural projections V 0 −−−
1
→ V 0 /V and W 0 −−− → W 0 /W to get a linear
2
π1 ⊗π2
map V 0 ⊗K W 0 −−− −−→ (V 0 /V ) ⊗K (W 0 /W ) that is onto with ker(π1 ⊗ π2 ) = V ⊗K W 0 +
0
V ⊗K W by Theorem 5.5.
Remark 5.7. It is false that (V 0 ⊗K W 0 )/(V ⊗K W ) ∼
= (V 0 /V ) ⊗K (W 0 /W ). The subspace
6
V ⊗K W is generally too small to be the kernel. This is a distinction between tensor
products and direct sums (where (V 0 ⊕ W 0 )/(V ⊕ W ) ∼
= (V 0 /V ) ⊕ (W 0 /W )).
ϕ
Corollary 5.8. Let V −−→ W be a linear map and U be a K-vector space. The linear map
1⊗ϕ
U ⊗K V −−−−→ U ⊗K W has kernel and image
(5.4) ker(1 ⊗ ϕ) = U ⊗K ker ϕ and Im(1 ⊗ ϕ) = U ⊗K ϕ(V ).
In particular, for nonzero U the map ϕ is injective or surjective if and only if 1 ⊗ ϕ has
that property.
Proof. This is immediate from Theorem 5.5 since we’re using the identity map on U .
6Exception: V 0 = V or W = 0, and W 0 = W or V = 0.
20 KEITH CONRAD
ϕ
Example 5.9. Let V −−→ W be a linear map and L/K be a field extension. The base
ϕL
extension L ⊗K V −−−→ L ⊗K W has kernel and image
ker(ϕL ) = L ⊗K ker ϕ, Im(ϕL ) = L ⊗K Im(ϕ).
The map ϕ is injective if and only if ϕL is injective and ϕ is surjective if and only if ϕL is
surjective.
Let’s formulate this in the language of matrices. If V and W are finite-dimensional then
ϕ can be written as a matrix with entries in K once we pick bases of V and W . Then
ϕL has the same matrix representation relative to the corresponding bases of L ⊗K V and
L ⊗K W . Since the base extension of a free module to another ring doesn’t change the size
of a basis, dimL (L ⊗K Im(ϕ)) = dimK Im(ϕ) and dimL (L ⊗K ker(ϕ)) = dimK ker(ϕ). That
means ϕ and ϕL have the same rank and the same nullity: the rank and nullity of a matrix
in Mm×n (K) do not change when it is viewed in Mm×n (L) for any field extension L/K.
In the rest of this section we will look at tensor products of many vector spaces at once.
Lemma 5.10. For v ∈ V with v 6= 0, there is ϕ ∈ V ∨ such that ϕ(v) = 1.
Proof. The set {v} is linearly independent, so it extends to a basis {vi }i∈I of V . Let v = vi0
in this indexing. Define ϕ : V → K by
!
X
ϕ ci vi = ci0 .
i
Proof. Let K be the fraction field of R. The torsion elements of M ⊗R N are precisely the
elements that go to 0 under the map M ⊗R N → K ⊗R (M ⊗R N ) sending t to 1 ⊗ t. We
want to show 1 ⊗ (x ⊗ y) 6= 0.
The natural K-vector space isomorphism K ⊗R (M ⊗R N ) ∼ = (K ⊗R M ) ⊗K (K ⊗R N )
identifies 1 ⊗ (x ⊗ y) with (1 ⊗ x) ⊗ (1 ⊗ y). Since x and y are non-torsion in M and
N , 1 ⊗ x 6= 0 in K ⊗R M and 1 ⊗ y 6= 0 in K ⊗R N . An elementary tensor of nonzero
vectors in two K-vector spaces is nonzero (Theorem 5.11), so (1 ⊗ x) ⊗ (1 ⊗ y) 6= 0 in
(K ⊗R M ) ⊗K (K ⊗R N ). Therefore 1 ⊗ (x ⊗ y) 6= 0 in K ⊗R (M ⊗R N ), which is what we
wanted to show.
Remark 5.14. If M and N are torsion-free, Corollary 5.13 is not saying M ⊗R N is
torsion-free. It only says all (nonzero) elementary tensors have no torsion. There could be
non-elementary tensors with torsion, as we saw at the end of Example 2.16.
In Theorem 5.11 we saw an elementary tensor in a tensor product of vector spaces is
0 only under the obvious condition that one of the vectors appearing in the tensor is 0.
We now show two nonzero elementary tensors in vector spaces are equal only under the
“obvious” circumstances.
Theorem 5.15. Let V1 , . . . , Vk be K-vector spaces. Pick pairs of nonzero vectors vi , vi0 in
Vi for i = 1, . . . , k. Then v1 ⊗ · · · ⊗ vk = v10 ⊗ · · · ⊗ vk0 in V1 ⊗K · · · ⊗K Vk if and only if there
are nonzero constants c1 , . . . , ck in K such that vi = ci vi0 and c1 · · · ck = 1.
Proof. If vi = ci vi0 for all i and c1 · · · ck = 1 then v1 ⊗ · · · ⊗ vk = c1 v10 ⊗ · · · ⊗ ck vk0 =
(c1 · · · ck )v10 ⊗ · · · ⊗ vk0 = v10 ⊗ · · · ⊗ vk0 .
Now we want to go the other way. It is clear for k = 1, so we may take k ≥ 2.
By Theorem 5.11, v1 ⊗· · ·⊗vk is not 0 since each vi is not 0. Fix ϕi ∈ Vi∨ for 1 ≤ i ≤ k −1
such that ϕi (vi ) = 1. (Such ϕi exist since v1 , . . . , vk−1 are nonzero vectors.) For arbitrary
ϕ ∈ Vk∨ , let hϕ = ϕ1 ⊗ · · · ⊗ ϕk−1 ⊗ ϕ, so hϕ (v1 ⊗ · · · ⊗ vk−1 ⊗ vk ) = ϕ(vk ). Also
hϕ (v1 ⊗ · · · ⊗ vk−1 ⊗ vk ) = hϕ (v10 ⊗ · · · ⊗ vk−1
0
⊗ vk0 ) = ϕ1 (v10 ) · · · ϕk−1 (vk−1
0
)ϕ(vk0 ) = ϕ(ck vk0 ),
where ck := ϕ1 (v10 ) · · · ϕk−1 (vk−1
0 ) ∈ K. So we have
ϕ(vk ) = ϕ(ck vk0 )
for arbitrary ϕ ∈ Vk∨ . Therefore ϕ(vk − ck vk0 ) = 0 for all ϕ ∈ Vk∨ , so vk − ck vk0 = 0, which
says vk = ck vk0 .
In the same way, for every i = 1, 2, . . . , k there is ci in K such that vi = ci vi0 . Then
v1 ⊗· · ·⊗vk = c1 v10 ⊗· · ·⊗ck vk0 = (c1 · · · ck )(v10 ⊗· · ·⊗vk0 ). Since v1 ⊗· · ·⊗vk = v10 ⊗· · ·⊗vk0 6= 0,
we get c1 · · · ck = 1.
In quantum mechanics, the quantum states of a system are described by the nonzero
vectors in a complex Hilbert space H where a (nonzero) scalar multiple of a vector in
H determines the same quantum state as the original vector (this condition is motivated
by physics), so the states of a quantum system can be described by the 1-dimensional
subspaces of H instead of by the individual (nonzero) elements of H.7 When two quantum
systems with corresponding Hilbert spaces H1 and H2 are combined, the Hilbert space
for the combined system is the (completed) tensor product H1 ⊗C H2 . In H1 ⊗C H2 , a 1-
dimensional subspace C(v1 ⊗v2 ) spanned by an elementary tensor determines the individual
subspaces Cv1 and Cv2 of H1 and H2 by Theorem 5.15, but most 1-dimensional subspaces of
H1 ⊗C H2 are not spanned by an elementary tensor and thus do not “come from” particular
1-dimensional subspaces of H1 and H2 . The non-elementary tensors in H1 ⊗C H2 describe
states that are called “entangled” and Schrödinger [4, p. 555] called this phenomenon “the
characteristic trait of quantum mechanics, the one that enforces its entire departure from
classical lines of thought [. . .] the best possible knowledge of a whole does not necessarily
include the best possible knowledge of all its parts.” He did not use the terminology of
tensor products, but you can see it at work in what he did say [4, p. 556], where his Hilbert
spaces are L2 -spaces of functions on some Rn : “Let x and y stand for all the coordinates of
the first and second systems respectively and Ψ(x, y) [stand for the] state of the composed
system [. . .]. What constitutes the entanglement is that Ψ is not a product of a function of
x and a function of y.” That is analogous to most elements of R[x, y] not being of the form
f (x)g(y), e.g., x3 y − xy 2 + 7 is an “entangled” polynomial.
Here is the analogue of Theorem 5.15 for linear maps (compare to Corollary 5.12).
Theorem 5.16. Let ϕi : Vi → Wi and ϕ0i : Vi → Wi be nonzero linear maps between K-
vector spaces for 1 ≤ i ≤ k. Then ϕ1 ⊗ · · · ⊗ ϕk = ϕ01 ⊗ · · · ⊗ ϕ0k as linear maps V1 ⊗K
· · · ⊗K Vk → W1 ⊗K · · · ⊗K Wk if and only if there are c1 , . . . , ck in K such that ϕi = ci ϕ0i
and c1 c2 · · · ck = 1.
Proof. Since each ϕi : Vi → Wi is not identically 0, for i = 1, . . . , k − 1 there is vi ∈ Vi such
that ϕi (vi ) 6= 0 in Wi . Then there is fi ∈ Wi∨ such that fi (ϕi (vi )) = 1.
Pick any v ∈ Vk and f ∈ Wk∨ . Set hf = f1 ⊗ · · · ⊗ fk−1 ⊗ f ∈ (W1 ⊗K · · · ⊗K Wk )∨ where
hf (w1 ⊗ · · · ⊗ wk−1 ⊗ wk ) = f1 (w1 ) · · · fk−1 (wk−1 )f (wk ).
Then
hf ((ϕ1 ⊗ · · · ⊗ ϕk−1 ⊗ ϕk )(v1 ⊗ · · · ⊗ vk−1 ⊗ v)) = f1 (ϕ1 (v1 )) · · · fk−1 (ϕk−1 (vk−1 ))f (ϕk (v)),
and since fi (ϕi (vi )) = 1 for i 6= k, the value is f (ϕk (v)). Also
hf ((ϕ01 ⊗ · · · ⊗ ϕ0k−1 ⊗ ϕ0k )(v1 ⊗ · · · ⊗ vk−1 ⊗ v)) = f1 (ϕ01 (v1 )) · · · fk−1 (ϕ0k−1 (vk−1 ))f (ϕ0k (v)).
Set ck = f1 (ϕ01 (v1 )) · · · fk−1 (ϕ0k−1 (vk−1 )), so the value is ck f (ϕ0k (v)) = f (ck ϕ0k (v)). Since
ϕ1 ⊗ · · · ⊗ ϕk−1 ⊗ ϕk = ϕ01 ⊗ · · · ⊗ ϕ0k−1 ⊗ ϕ0k ,
f (ϕk (v)) = f (ck ϕ0k (v)).
This holds for all f ∈ Wk∨ , so ϕk (v) = ck ϕ0k (v). This holds for all v ∈ Vk , so ϕk = ck ϕ0k as
linear maps Vk → Wk .
7A nonzero vector in H, viewed as a quantum state, is often scaled to have length 1, but this still allows
ambiguity up to scaling by a complex number eiθ of absolute value 1, with θ called a phase-factor.
TENSOR PRODUCTS II 23
The set of all symmetric tensors and the set of all anti-symmetric tensors in V ⊗k each
form subspaces. If K does not have characteristic 2, every tensor in V ⊗2 is a unique sum
of a symmetric and anti-symmetric tensor:
t + P(12) (t) t − P(12) (t)
t= + .
2 2
(The map P(12) is the flip automorphism of V ⊗2 sending v ⊗ w to w ⊗ v.) Concretely, if
{e1 , . . . , en } is a basis of V then each tensor in V ⊗2 has the form i,j cij ei ⊗ ej , and it’s
P
symmetric if and only if cij = cji for all i and j, while it’s anti-symmetric if and only if
cij = −cji for all i and j (in particular, cii = 0 for all i).
For k > 2, the symmetric and anti-symmetric tensors in V ⊗k do not span the whole space.
There are additional subspaces of tensors in V ⊗k connected to the representation theory of
the group GL(V ). The appearance of representations of GL(V ) inside tensor powers of V
is an important role for tensor powers in algebra.
6. Tensor contraction
Continuing the theme of the previous section, let V be a finite-dimensional vector space
over a field K. The evaluation pairing V × V ∨ → K, where (v, ϕ) 7→ ϕ(v), is K-bilinear and
thus induces a K-linear map c : V ⊗ V ∨ → K called a contraction9, where c(v ⊗ ϕ) = ϕ(v).
This map is independent of the choice of a basis, but it’s worth seeing how this looks in a
basis. We adopt the notation of physicists and geometers for bases and coefficients in V and
V ∨ , as in the section about tensors in physics in the first handout about tensor products:
for a basis P{e1 , . . . , en } of V , let {e1 , . . . , en } be itsPdual basis in V ∨ , with elements of V
written as i v i ei and elements of V ∨ written as i vi ei . (Here v i , vi ∈ K.) The space
9Here and elsewhere in this section, tensor products are always over K: ⊗ = ⊗ .
K
TENSOR PRODUCTS II 25
The final sum is independent of the basis used on V , since contraction comes from evaluating
V ∨ on V and that does not depend on a choice of basis.
From the first tensor product handout, V ⊗ V ∨ ∼ = HomK (V, V ) by
v ⊗ ϕ 7→ [w 7→ ϕ(w)v]
How does cr,s look in a basis? For a basis {e1 , . . . , en } of V and dual basis {e1 , . . . , en } of
V a basis of V ⊗k ⊗ (V ∨ )⊗` is the nk+` elementary tensors ei1 ⊗ · · · ⊗ eik ⊗ ej1 ⊗ · · · ⊗ ej` ,
∨,
and
n
···m···ik
X i1 ...i X X O O
Tj1 ...j`k ei1 ⊗ · · · ⊗ eik ⊗ ej1 ⊗ · · · ⊗ ej` Tji11···m···j ejb ,
cr,s
= `
e ia ⊗
i1 ,...,ik i1 ,...,j` m=1 a6=r b6=s
j1 ,...,j` ir ,js missing
where the m appears in the rth slot of the superscript and the sth slot of the subscript of
the coefficient on the right. The outer sum on the right initially runs over all sets of indices
and the coefficient is multiplied by the numerical factor eis (eir ), which is nonzero only for
is = ir (call this common value m, and its runs from 1 to n), when the factor is 1.
The contraction cr,s depends on its domain V ⊗k ⊗ (V ∨ )⊗` , not just r and s (which are at
most k and `, respectively). For example, on V ⊗2 ⊗ V ∨ and V ⊗2 ⊗ (V ∨ )⊗2 , the contraction
c2,1 on elementary tensors is v ⊗ w ⊗ ϕ 7→ ϕ(w)v and v ⊗ w ⊗ ϕ ⊗ ψ 7→ ϕ(w)v ⊗ ψ. Using
26 KEITH CONRAD
Remark 6.1. In physics and engineering, tensors are often describedP i just by components
(coordinates), with summing over the basis being understood: i
i T ei is written as T and
i j
P
i,j Tij e ⊗ e is written as Tij . That components have indices in the opposite position
(up vs. down) to basis vectors needs to be known to reconstruct the tensor from how its
components are written. Another convention, which is widely used in geometry as well, is
that summation signs within a component (due to contraction or to applying the metric –
see below) are omitted and an implicit summation running from 1 to the dimension of V
is intended for each index repeated as both a superscript and subscript. This is called the
Einstein P summation convention. For example, the most basic contraction V ⊗ V ∨ → K,
i j
P i
where i,j Tj ei ⊗ e 7→ i Ti , is written in this convention as
Tji 7→ Tii .
The contractions c2,1 in (6.2) and (6.3), on V ⊗2 ⊗ V ∨ and V ⊗2 ⊗ (V ∨ )⊗2 respectively, are
denoted in the Einstein summation convention as
Tji11 i2 7→ Tii21 i2 , Tji11ji22 7→ Tii21ji22 .
Contraction only makes sense when we combine V and V ∨ . We can’t contract V and V
– it doesn’t make sense to evaluate elements of V on V . However, if we have a preferred
isomorphism g : V → V ∨ of vector spaces then we can use g to turn V into V ∨ . We have
V ⊗V ∼ = V ⊗ V ∨ using id. ⊗ g and we can contract the right side. So we can contract on
V ⊗ V after all, provided we use an isomorphism g from V to V ∨ and remember what g is.
Contraction on V ⊗ V (using g) means the composite map
id.⊗g c
(6.4) V ⊗ V −−−→ V ⊗ V ∨ −−−→ K.
This depends on the choice of g but not on a choice of basis of V .
There should be nothing special about using g in the second tensorand in (6.4). We want
g⊗id.
(6.5) V ⊗ V −−−→ V ∨ ⊗ V −−−→ K
to be the same overall mapping as (6.4), where the second mapping in (6.5) is induced by
evaluation of V ∨ on V . The agreement of (6.4) and (6.5) means g(v)(w) = g(w)(v) for all
v and w in V . This means that if we regard g : V → V ∨ as a bilinear map V × V → K
by (v, w) 7→ g(v)(w) or (v, w) 7→ g(w)(v) we want the same result: we want g to be
symmetric as a bilinear form on V . To check a given g is symmetric, it enough to check this
TENSOR PRODUCTS II 27
on a basis {e1 , . . . , en } of V : g(ei )(ej ) = g(ej )(ei ) in K for all i and j. Set gij = g(ei )(ej ), so
X
g(ei ) = gij ej for all i and gij = gji for all i, j from 1 to n: the matrix (gij ) is symmetric.
j
We can think of the isomorphism g : V → V ∨ as a bilinear map V × V → K in two ways:
either as (v, w) 7→ g(v)(w) or as (v, w) 7→ g(w)(v). The fact that g is an isomorphism is
equivalent to g as a bilinear form on V being nondegenerate and is also equivalent to the
matrix (gij ) coming from any basis of V (gij = g(ei )(ej )) being invertible. In practice we
always assume g is symmetric and nondegenerate as a bilinear form on V .
Let’s write g and then the contraction on V ⊗ V from (6.4) or (6.5) P in icoordinates. Pick
a basis {e1 , . . . , en } of V and dual basis {e1 , . . . , en } of V ∨ . For v = iPv ei in V , how does
g(v) look in the dual basis of V ∨ ? That gij = g(ei )(ej ) means g(ei ) = j gij ej , so
! !
X X X X X X
g(v) = g v i ei = v i g(ei ) = vi gij ej = gij v i ej .
i i i j j i
X X
Thus g( i v i ei ) = j gij v i = gji v i for all j (recall gij = gji ).
P P
j vj e , where vj =
i i
The passage from the numbers {v i }
(components of v) to the numbers {vj } (components
of g(v)) by multiplying each v i by gij and summing over all i is called lowering an index.
It is the coordinate version (using a basis of V and its dual basis in V ∨ ) of going from v in
V to g(v) in ∨
PV i, which depends
P ion g but ⊗2
not on the choice of basis of V .
For x = i x ei and y = i y ei in V , (6.4) has the following effect on the elementary
tensor x ⊗ y:
x ⊗ y 7→ x ⊗ g(y)
!
X X X X
= xi e i ⊗ yj ej where yj = gkj y k = gjk y k
i j k k
c
X
7→ xi yi by (6.1).
i
xi y i
P
This is usually not i unless gij = δij for all i and j. P More generally, the contraction of
ij ⊗2 using g is the scalar i,j gij T ij in K. This contraction
P
the tensor T = i,j T ei ⊗ej in V
is independent of the basis {e1 , . . . , en } of V but depends on the isomorphism g : V → V ∨ .
Example 6.2. Let g be an isomorphism V → V ∨ that is symmetric (and nondegenerate)
as a bilinear form V × V →P K. Applying g to the sth tensorand of V ⊗k , where k ≥ 2 and
1 ≤ s ≤ k, turns the tensor i1 ,...,ik T i1 ...ik ei1 ⊗ · · · ⊗ eik into
!
X X X
i1 ...ik i1 ...ik i
T ei1 ⊗ · · · ⊗ g(eis ) ⊗ · · · ⊗ eik = T e i1 ⊗ · · · ⊗ giis e ⊗ · · · ⊗ eik
i1 ,...,ik i1 ,...,ik i
| {z }
sth tensorand
28 KEITH CONRAD
and contracting the rth and sth tensorands10 where r 6= s gives us the following tensor in
V ⊗(k−2) :
X X
gir is T i1 ...ik ei1 ⊗ · · · ⊗ eik .
| {z }
i1 ,...,ik ir ,is
ir ,is missing ir ,is missing
Remark 6.4. Putting the Einstein summation convention (Remark 6.1) to work, the con-
traction V ⊗V ∨ → K sends v i ei ⊗wj ej to v i wi , which is the same as v j wj and is independent
of the basis since it is simply evaluating V ∨ on V . Using an isomorphism g : V → V ∨ that is
symmetric (and nondegenerate) as a bilinear form V × V → K, any v = v i ei in V turns into
g(v) = vi ei in V ∨ where vi = gki v k = gik v k since (gij ) is symmetric, and the contraction
V ⊗ V → K that depends on g has the effect xi ei ⊗ y j ej 7→ xi yi = xj y j = gij xi y j (remember
gij = gji !) and more generally T ij ei ⊗ ej 7→ gij T ij . The contraction on V ⊗2 depends on g
but not on the basis of V . Contraction on V ⊗k , for k ≥ 2, using the rth and sth tensorand
is a linear map to V ⊗(k−2) with the effect T i1 ...ik ei1 ⊗ · · · ⊗ eik 7→ gir is T i1 ···ik ei1 ⊗ · · · ⊗ eik .
| {z }
ir ,is missing
For Tji11 i2 in V ⊗ V ⊗ V ∨ , if we lower the first upper index we get gi1 j2 Tji11 i2 in V ∨ ⊗ V ⊗ V ∨
and if we lower both upper indices then we get gi1 j2 gi2 j3 Tji11 i2 in V ∨ ⊗ V ∨ ⊗ V ∨ .
10While V ⊗(s−1) ⊗ (V ∨ ) ⊗ V ⊗(k−s) is not in standard form, the meaning of contraction using the rth
tensorand V and the sth tensorand V ∨ should be obvious to the reader.
TENSOR PRODUCTS II 29
11Raising or lowering indices describes what happens to components in a basis. The basis undergoes the
opposite change (lowering or raising its indices).
30 KEITH CONRAD
for r ∈ R and a and b in B, which says R-scaling commutes with multiplication in the
R-algebra. We also want 1 · a = a for a ∈ A, where 1 is the identity element of R.
Examples of R-algebras include the matrix ring Mn (R), a quotient ring R/I, and the
polynomial ring R[X1 , . . . , Xn ]. We will assume, as in all these examples, that our algebras
have associative multiplication and a multiplicative identity, so they are genuinely rings
(perhaps not commutative) and being an R-algebra just means they have a little extra
structure related to scaling by R. When an R-algebra contains R, (7.1) is a special case of
associative multiplication in the algebra.
The difference between an R-algebra and a ring is exactly like that between an R-module
and an abelian group. An R-algebra is a ring on which we have a scaling operation by R
that behaves nicely with respect to the addition and multiplication in the R-algebra, in the
same way that an R-module is an abelian group on which we have a scaling operation by
R that behaves nicely with respect to the addition in the R-module. While Z-modules are
nothing other than abelian groups, Z-algebras in our lexicon are nothing other than rings
(possibly noncommutative).
Because of the universal mapping property of the tensor product, to give an R-bilinear
multiplication A × A → A in an R-algebra A is the same thing as giving an R-linear map
A ⊗R A → A. So we could define an R-algebra as an R-module A equipped with an R-
m
linear map A ⊗R A −−→ A, and declare the product of a and b in A to be ab := m(a ⊗ b).
Associativity of multiplication can be formulated in tensor language: the diagram
1⊗m
A ⊗R A ⊗R A / A ⊗R A
m⊗1 m
A ⊗R A
m /A
commutes.
Theorem 7.1. Let A and B be R-algebras. There is a unique multiplication on A ⊗R B
making it an R-algebra such that
(7.2) (a ⊗ b)(a0 ⊗ b0 ) = aa0 ⊗ bb0
for all elementary tensors. The multiplicative identity is 1 ⊗ 1.
Proof. If there is an R-algebra multiplication on A ⊗R B satisfying (7.2) then multiplication
between any two tensors is determined:
k
X `
X X X
ai ⊗ bi · a0j ⊗ b0j = (ai ⊗ bi )(a0j ⊗ b0j ) = ai a0j ⊗ bi b0j .
i=1 j=1 i,j i,j
four square roots of 1, namely (±1, ±1) using independent choices of signs, but in R ⊗Q R
there are at least six square roots of 1:
1 √ √ 1 √ √
±(1 ⊗ 1), ± ( 2 ⊗ 2) and ± ( 3 ⊗ 3).
2 3
The R-algebras A × B and A ⊗R B have dual universal mapping properties. For any
ϕ ψ
R-algebra C and R-algebra homomorphisms C −−→ A and C −−→ B, there is a unique
R-algebra homomorphism C −→ A × B making the diagram
ϕ ψ
A×B
π1 π2
{ #
A B
ϕ ψ
commute. For any R-algebra C and R-algebra homomorphisms A −−→ C and B −−→ C
such that the images of A and B in C commute (ϕ(a)ψ(b) = ψ(b)ϕ(a)), there is a unique
R-algebra homomorphism A ⊗R B −→ C making the diagram
A B
a7→a⊗1 b7→1⊗b
# {
A ⊗R B
ϕ ψ
% y
C
commute.
A practical criterion for showing an R-linear map of R-algebras is an R-algebra homo-
morphism is as follows. If ϕ : A → B is an
PR-linear map of R-algebras and {ai } is a spanning
set for A as an R-module (that is, A = i Rai ), then ϕ is multiplicative as long as it is so
on these module generators: ϕ(ai aj ) = ϕ(ai )ϕ(aj ) for all i and j. Indeed, if this equation
34 KEITH CONRAD
holds then
X X X
ϕ ri ai · rj0 aj = ϕ ri rj0 ai aj
i j i,j
X
= ri rj0 ϕ(ai aj )
i,j
X
= ri rj0 ϕ(ai )ϕ(aj )
i,j
X X
= ri ϕ(ai ) rj0 ϕ(aj )
i j
!
X X
= ϕ ri ai ϕ rj0 aj .
i j
This will let us bootstrap a lot of known R-module isomorphisms between tensor products
to R-algebra isomorphisms by checking the behavior only on products of elementary tensors
(and checking the multiplicative identity is preserved, which is always easy). We give some
concrete examples before stating some general theorems.
Example 7.6. For ideals I and J in R, there is an isomorphism ϕ : R/I ⊗R R/J −→
R/(I + J) of R-modules where ϕ(x ⊗ y) = xy. Then ϕ(1 ⊗ 1) = 1 and
ϕ((x ⊗ y)(x0 ⊗ y 0 )) = ϕ(xx0 ⊗ yy 0 ) = xx0 yy 0 = xy x0 y 0 = ϕ(x ⊗ y)ϕ(x0 ⊗ y 0 ).
So R/I ⊗R R/J = ∼ R/(I + J) as R-algebras, not just as R-modules. In particular, the
additive isomorphism Z/aZ ⊗Z Z/bZ ∼ = Z/(a, b)Z is in fact an isomorphism of rings.
Example 7.7. There is an R-module isomorphism ϕ : R[X] ⊗R R[Y ] −→ R[X, Y ] where
ϕ(f (X) ⊗ g(Y )) = f (X)g(Y ). Let’s show it’s an R-algebra isomorphism: ϕ(1 ⊗ 1) = 1 and
ϕ((f1 (X) ⊗ g1 (Y ))(f2 (X) ⊗ g2 (Y ))) = ϕ(f1 (X)f2 (X) ⊗ g1 (Y )g2 (Y ))
= f1 (X)f2 (X)g1 (Y )g2 (Y )
= f1 (X)g1 (Y )f2 (X)g2 (Y )
= ϕ(f1 (X) ⊗ g1 (Y ))ϕ(f2 (X) ⊗ g2 (Y )),
so R[X] ⊗R R[Y ] ∼ = R[X, Y ] as R-algebras, not just as R-modules. (It would have sufficed
to check ϕ is multiplicative on pairs of monomial tensors X i ⊗ Y j .)
In a similar way, the natural R-module isomorphism R[X]⊗k ∼ = R[X1 , . . . , Xk ], where the
indeterminate Xi on the right corresponds on the left to the tensor 1 ⊗ · · · ⊗ X ⊗ · · · ⊗ 1
with X in the ith position, is an isomorphism of R-algebras.
Example 7.8. When R is a domain with fraction field K, K ⊗R K ∼ = K as R-modules by
x ⊗ y 7→ xy. This sends 1 ⊗ 1 to 1 and preserves multiplication on elementary tensors, so it
is an isomorphism of R-algebras.
Example 7.9. Let F be a field. When x and y are independent indeterminates over F ,
F [x] ⊗F F [y] ∼
= F [x, y] as F -algebras by Example 7.7. It is natural to think that we should
also have F (x) ⊗F F (y) ∼ = F (x, y) as F -algebras, but this is always false! Why should it be
false, and is there a concrete way to think about F (x) ⊗F F (y)?
TENSOR PRODUCTS II 35
P
Every tensor t in F (x)⊗F F (y) is a finite sum i,j fi (x)⊗gj (y). We can give all the fi (x)’s
a common denominator and all the gj (y)’s a common denominator, say fi (x) = ai (x)/b(x)
and gj (y) = cj (y)/d(y) where ai (x) ∈ F [x] and cj (y) ∈ F [y]. Then
X
X ai (x) cj (y) 1 1
t= ⊗F = ⊗ ai (x) ⊗ cj (y).
b(x) d(y) b(x) d(y)
i,j i,j
The distributivity of ⊗ over × suggests denoting the direct product of algebras as a direct
sum with ⊕.
Proof. Exercise. Note the direct product of two R-algebras is the direct sum as R-modules
with componentwise multiplication, so first just treat the direct product as a direct sum.
Corollary 7.13. For R-algebras A and B, A ⊗R B n = ∼ (A ⊗R B)n as R-algebras.
Proof. Induct on n. Note B n here means the n-fold product ring, not B ⊗n .
We turn now to base extensions. Fix a homomorphism f : R → S of commutative rings.
We can restrict scalars from S-modules to R-modules and extend scalars from R-modules to
S-modules. What about between R-algebras and S-algebras? An example is the formation
of C ⊗R Mn (R), which ought to look like Mn (C) as rings (really, as C-algebras) and not
just as complex vector spaces.
If A is an S-algebra, then we make A into an R-module in the usual way by ra = f (r)a,
and this makes A into an R-algebra (restriction of scalars). More interesting is extension
of scalars. For this we need a lemma.
ϕ ψ
Lemma 7.14. If A, A0 , B, and B 0 are all R-algebras and A −−→ A0 and B −−→ B 0 are
ϕ⊗ψ
R-algebra homomorphisms then the R-linear map A ⊗R B −−−−→ A0 ⊗R B 0 is an R-algebra
homomorphism.
Proof. Exercise.
Theorem 7.15. Let A be an R-algebra.
(1) The base extension S ⊗R A, which is both an R-algebra and an S-module, is an
S-algebra by its S-scaling.
ϕ 1⊗ϕ
(2) If A −−→ B is an R-algebra homomorphism then S ⊗R A −−−−→ S ⊗R B is an
S-algebra homomorphism.
Proof. 1) We just need to check multiplication in S ⊗R A commutes with S-scaling (not
just R-scaling): s(tt0 ) = (st)t0 = t(st0 ). Since all three expressions are additive in t and t0 ,
it suffices to check this when t and t0 are elementary tensors:
? ?
s((s1 ⊗ a1 )(s2 ⊗ a2 )) = (s(s1 ⊗ a1 ))(s2 ⊗ a2 ) = (s1 ⊗ a1 )(s(s2 ⊗ a2 )).
From the way S-scaling on S ⊗R A is defined, all these products equal ss1 s2 ⊗ a1 a2 .
ϕ 1⊗ϕ
2) For an R-algebra homomorphism A −−→ B, the base extension S ⊗R A −−−−→ S ⊗R B is
S-linear and it is an R-algebra homomorphism by Lemma 7.14. Therefore it is an S-algebra
homomorphism.
We can also give A ⊗R S an S-algebra structure by S-scaling and the natural S-module
isomorphism S ⊗R A ∼ = A ⊗R S is an S-algebra isomorphism.
Example 7.16. Let I be an ideal in R[X1 , . . . , Xn ]. Check the S-module isomorphism
S ⊗R R[X1 , . . . , Xn ]/I ∼
= S[X1 , . . . , Xn ]/(I · S[X1 , . . . , Xn ]) from Corollary 2.24 is an S-
algebra isomorphism.
In one-variable, with I = (h(X)) a principal ideal in R[X],12 Example 7.16 gives us an
S-algebra isomorphism
S ⊗R R[X]/(h(X)) ∼ = S[X]/(hf (X)),
12Not all ideals in R[X] have to be principal, but this is just an example.
TENSOR PRODUCTS II 37
√ √
The Q-linear embedding Q( n 2) ,→ R extends to an R-linear embedding R ⊗Q Q( n 2) →
R ⊗Q R that is multiplicative (it suffices to check that on elementary tensors), so the ring
R ⊗Q R contains a subring isomorphic to R[X]/(X n − 2). When n is odd, X n − 2 has
one linear factor and (n − 1)/2 quadratic irreducible factors in R[X], so R[X]/(X n − 2) ∼ =
R×C (n−1)/2 n
as R-algebras. Therefore R[X]/(X −2) contains 2 1+(n−1)/2 =2(n+1)/2 square
roots of 1. Letting n → ∞ shows R ⊗Q R contains infinitely many square roots of 1.
Example 7.18. We revisit Example 7.3, using Example 7.16:
C ⊗R C ∼
= C ⊗R (R[X]/(X 2 + 1)) ∼
= C[X]/(X 2 + 1) = C[X]/(X − i)(X + i) ∼
=C×C
as R-algebras.
Let’s make the R-algebra isomorphism C ⊗R C ∼ = C × C explicit, as it is not z ⊗ w 7→
(z, w). Tracing the effect of the isomorphisms on elementary tensors,
z ⊗ (a + bi) 7→ z ⊗ (a + bX) 7→ za + zbX 7→ (za + zbi, za + ab(−i)) = (z(a + bi), z(a − bi)),
so z ⊗ w 7→ (zw, zw). Thus 1 ⊗ 1 7→ (1, 1), z ⊗ 1 7→ (z, z), and 1 ⊗ w 7→ (w, w).
In these examples, a tensor product of fields is not a field. But the tensor product of
fields can be a field (besides the trivial case K ⊗K K ∼
= K). Here is an example.
Example 7.19. We have
√ √ √ √
Q( 2) ⊗Q Q( 3) ∼ = Q( 2) ⊗Q Q[X]/(X 2 − 3) ∼ = Q( 2)[X]/(X 2 − 3),
√
which is a field because X 2 − 3 is irreducible in Q( 2)[X].
Example 7.20. As an example of a tensor product involving a finite field and a ring,
Z/5Z ⊗Z Z[i] ∼
= Z/5Z ⊗Z Z[X]/(X 2 + 1) ∼= (Z/5Z)[X]/(X 2 + 1) ∼ = Z/5Z × Z/5Z
since X 2 + 1 = (X − 2)(X − 3) in (Z/5Z)[X].
A general discussion of tensor products of fields is in [3, Sect. 8.18].
Theorem 7.21. If A is R-algebra and B is an S-algebra, then the S-module structure on
the R-algebra A ⊗R B makes it an S-algebra, and
A ⊗R B ∼= (A ⊗R S) ⊗S B
as S-algebras sending a ⊗ b to (a ⊗ 1) ⊗ b.
38 KEITH CONRAD
F E
K
Then there is a ring isomorphism
F ⊗K L ∼
= (F ⊗K E) ⊗E L
which is also an isomorphism as E-algebras, F -algebras (from the left factor) and L-algebras
(from the right factor).
Theorem 7.22. Let A and B be R-algebras. There is an S-algebra isomorphism
S ⊗R (A ⊗R B) → (S ⊗R A) ⊗S (S ⊗R B)
by s ⊗ (a ⊗ b) 7→ s((1 ⊗ a) ⊗ (1 ⊗ b)).
Proof. By part I, there is an S-module isomorphism with the indicated effect on tensors of
the form s ⊗ (a ⊗ b). This function preserves multiplicative identities and is multiplicative
on such tensors (which span S ⊗R (A ⊗R B)), so it is an S-algebra isomorphism.
where tk ⊗ t0` ∈ M ⊗n if k + ` = n. To show this multiplication makes the direct sum of all
M ⊗k an R-algebra we use the following construction theorem.
Theorem 8.1. Let {Mk }k≥0 be a sequence of R-modules with M0 = R and let there be
R-bilinear mappings (“multiplications”) µk,` : Mk × M` → Mk+` for all k and ` such that
1) (scaling by R) µk,0 : Mk × R → Mk and µ0,` : R × M` → M` are both scaling by R:
µk,0 (x, r) = rx and µ0,` (r, y) = ry for r ∈ R, x ∈ Mk , and y ∈ M` ,
2) (associativity) for k, `, n ≥ 0 we have µk,`+n (x, µ`,n (y, z)) = µk+`,n (µk,` (x, y), z) in
Mk+`+n for all x ∈ Mk , y ∈ M` , and z ∈ Mn .
L
The direct sum k≥0 Mk is an associative R-algebra with identity using the multiplication
rule k≥0 mk · `≥0 m0` = n≥0 0
P P P P
k+`=n µk,` (mk , m` ) .
L
Proof. The direct sum k≥0 Mk is automatically an R-module. It remains to check (i) the
multiplication defined on the direct sum is R-bilinear, (ii) 1 ∈ R = M0 is a multiplica-
tive identity, and (iii) multiplication is associative. The R-bilinearity of multiplication is a
bookkeeping exercise left to the reader. In particular, this includes distributivity of multi-
plication over addition. To prove multiplication has 1 as an identity and is associative, it
suffices by distributivity to consider only multiplication with factors from direct summands
Mk , in which case we can use the two properties of the maps µk,` in the theorem.
M ⊗k × M ⊗`
3
(⊗,⊗)
(M × · · · × M ) × (M × · · · × M ) βk,`
f
+
M ⊗(k+`)
commute. (This collapsing is analogous to the proof of associativity of the tensor product
in part I.) By commutativity of the diagram, on a pair of elementary tensors we have
⊗k
L
Theorem 8.3. The M -module k≥0 M is an R-algebra using the multiplication in (8.1).
Proof. Use Theorem 8.1 with Mk = M ⊗k and µk,` = βk,` from Lemma 8.2. The first prop-
erty of the maps µk,` in Theorem 8.1 is automatic from the definition of βk,0 and β0,` . To
prove the second property of the maps µk,` in Theorem 8.1, namely µk,`+n (t1 , µ`,n (t2 , t3 )) =
µk+`,n (µk,` (t1 , t2 ), t3 ) in M ⊗(k+`+n) for all t1 ∈ M ⊗k , t2 ∈ M ⊗` , and t3 ∈ M ⊗n , by multi-
linearity of each µk,` = βk,` it suffices to consider the case when each ti is an elementary
tensor, in which case the equality is a simple calculation.
⊗k
L
Definition 8.4. For an R-module M , its tensor algebra is T (M ) := k≥0 M with
multiplication defined by (8.1).
Example 8.6. If M is a finite free R-module with basis e1 , . . . , ed then T (M ) is the poly-
nomial ring over R in d noncommuting indeterminates e1 , . . . , ed : in M ⊗2 , ei ⊗ ej 6= ej ⊗ ei
when i 6= j, which says in T (M ) that ei ej 6= ej ei .
Theorem 8.7. Let M be an R-module. For each R-algebra A and R-linear map f : M → A,
there is a unique R-algebra map F : T (M ) → A such that the diagram
T (M )
7
i
M F
(
f
A
commutes.
This says R-linear maps from M to an R-algebra A turn into R-algebra homomorphisms
from T (M ) to A. It works by extending an R-linear map from M → A to an R-algebra map
T (M ) → A by forcing multiplicativity, and there are no relations to worry about keeping
track of because T (M ) is an R-algebra formed from M in the most general way possible.
Proof. First suppose there is an R-algebra map F that makes the indicated diagram com-
mute. For r ∈ R, F (r) = r since F is an R-algebra homomorphism. Since T (M ) is generated
as an R-algebra by M , an R-algebra homomorphism out of T (M ) is determined by its values
on M , which really means its values on i(M ). For m ∈ M , we have F (i(m)) = f (m), and
thus there is at most one R-algebra homomorphism F that fits into the above commutative
diagram.
To construct F , we will define it first on each M ⊗k in T (M ) and then extend by additivity.
For k ≥ 1, the R-linear map f : M → A leads to an R-linear map f ⊗k : M ⊗k → A⊗k that is
m1 ⊗· · ·⊗mk 7→ f (m1 )⊗· · ·⊗f (mk ) on elementary tensors. Multiplication on A gives us an
R-linear map A⊗k → A that is a1 ⊗ · · · ⊗ ak 7→ a1 · · · ak on elementary tensors. Composing
this with f ⊗k gives us an R-linear map Fk : M ⊗k → A whose value on elementary tensors
is
Fk (m1 ⊗ · · · ⊗ mk ) = f (m1 ) · · · f (mk ).
⊗0
Define F0 : M → A by F0 (r) = r · 1A for r ∈ R. Finally, define F : T (M ) → A by
X X
F tk = Fk (tk )
k≥0 k≥0
Tensor algebras are useful preliminary constructions for other structures that can be
defined as a quotient of them, such as the exterior algebra of a module, the Clifford algebra
of a quadratic form, and the universal enveloping algebra of a Lie algebra.
References
[1] S. Friedl, Eta invariants as sliceness obstructions and their relation to Casson–Gordon invariants, Algebr.
Geom. Topol. 4 (2004), 893–934.
[2] H. V. Henderson and F. Pukelsheim, On the History of the Kronecker Product, Linear and Multilinear
Algebra 14 (1983), 113–120.
[3] N. Jacobson,“Basic Algebra II,” 2nd ed., W. H. Freeman & Co., New York, 1989.
[4] E. Schrödinger, Discussion of Probability Relations between Separated Systems, Math. Proc.
Camb. Philos. Society 31 (1935), 555–563. Online at https://www.cambridge.org/core/
journals/mathematical-proceedings-of-the-cambridge-philosophical-society/article/discussion-
of-probability-relations-between-separated-systems/C1C71E1AA5BA56EBE6588AAACB9A222D.