For a ring R, a right R-module M, a left R-module N, and an abelian group G, a map φ: M × N → G is said to be R-balanced, R-middle-linear or an R-balanced product if for all m, m′ in M, n, n′ in N, and r in R the following hold:[1]: 126
The set of all such balanced products over R from M × N to G is denoted by LR(M, N; G).
If φ, ψ are balanced products, then each of the operations φ + ψ and −φ defined pointwise is a balanced product. This turns the set LR(M, N; G) into an abelian group.
For M and N fixed, the map G ↦ LR(M, N; G) is a functor from the category of abelian groups to itself. The morphism part is given by mapping a group homomorphism g : G → G′ to the function φ ↦ g ∘ φ, which goes from LR(M, N; G) to LR(M, N; G′).
Remarks
Properties (Dl) and (Dr) express biadditivity of φ, which may be regarded as distributivity of φ over addition.
For every abelian group G and every balanced product
there is a unique group homomorphism
such that
As with all universal properties, the above property defines the tensor product uniquely up to a unique isomorphism: any other abelian group and balanced product with the same properties will be isomorphic to M ⊗RN and ⊗. Indeed, the mapping ⊗ is called canonical, or more explicitly: the canonical mapping (or balanced product) of the tensor product.[3]
The definition does not prove the existence of M ⊗RN; see below for a construction.
This is a succinct way of stating the universal mapping property given above. (If a priori one is given this natural isomorphism, then can be recovered by taking and then mapping the identity map.)
Similarly, given the natural identification ,[4] one can also define M ⊗RN by the formula
for the image of (x, y) under the canonical map . It is often called a pure tensor. Strictly speaking, the correct notation would be x ⊗Ry but it is conventional to drop R here. Then, immediately from the definition, there are relations:
x ⊗ (y + y′) = x ⊗ y + x ⊗ y′
(Dl⊗)
(x + x′) ⊗ y = x ⊗ y + x′ ⊗ y
(Dr⊗)
(x ⋅ r) ⊗ y = x ⊗ (r ⋅ y)
(A⊗)
The universal property of a tensor product has the following important consequence:
Proposition — Every element of can be written, non-uniquely, as
In other words, the image of generates . Furthermore, if f is a function defined on elements with values in an abelian group G, then f extends uniquely to the homomorphism defined on the whole if and only if is -bilinear in x and y.
Proof: For the first statement, let L be the subgroup of generated by elements of the form in question, and q the quotient map to Q. We have: as well as . Hence, by the uniqueness part of the universal property, q = 0. The second statement is because to define a module homomorphism, it is enough to define it on the generating set of the module.
Application of the universal property of tensor products[edit]
Determining whether a tensor product of modules is zero[edit]
In practice, it is sometimes more difficult to show that a tensor product of R-modules is nonzero than it is to show that it is 0. The universal property gives a convenient way for checking this.
To check that a tensor product is nonzero, one can construct an R-bilinear map to an abelian group such that . This works because if , then .
For example, to see that , is nonzero, take to be and . This says that the pure tensors as long as is nonzero in .
The proposition says that one can work with explicit elements of the tensor products instead of invoking the universal property directly each time. This is very convenient in practice. For example, if R is commutative and the left and right actions by R on modules are considered to be equivalent, then can naturally be furnished with the R-scalar multiplication by extending
to the whole by the previous proposition (strictly speaking, what is needed is a bimodule structure not commutativity; see a paragraph below). Equipped with this R-module structure, satisfies a universal property similar to the above: for any R-module G, there is a natural isomorphism:
If R is not necessarily commutative but if M has a left action by a ring S (for example, R), then can be given the left S-module structure, like above, by the formula
Analogously, if N has a right action by a ring S, then becomes a right S-module.
Tensor product of linear maps and a change of base ring[edit]
Given linear maps of right modules over a ring R and of left modules, there is a unique group homomorphism
The construction has a consequence that tensoring is a functor: each right R-module M determines the functor
from the category of left modules to the category of abelian groups that sends N to M ⊗ N and a module homomorphism f to the group homomorphism 1 ⊗ f.
If is a ring homomorphism and if M is a right S-module and N a left S-module, then there is the canonical surjective homomorphism:
The resulting map is surjective since pure tensors x ⊗ y generate the whole module. In particular, taking R to be this shows every tensor product of modules is a quotient of a tensor product of abelian groups.
(This section need to be updated. For now, see § Properties for the more general discussion.)
It is possible to extend the definition to a tensor product of any number of modules over the same commutative ring. For example, the universal property of
M1 ⊗ M2 ⊗ M3
is that each trilinear map on
M1 × M2 × M3 → Z
corresponds to a unique linear map
M1 ⊗ M2 ⊗ M3 → Z.
The binary tensor product is associative: (M1 ⊗ M2) ⊗ M3 is naturally isomorphic to M1 ⊗ (M2 ⊗ M3). The tensor product of three modules defined by the universal property of trilinear maps is isomorphic to both of these iterated tensor products.
Let R be a commutative ring, and M, N and P be R-modules. Then
Identity
Associativity
The first three properties (plus identities on morphisms) say that the category of R-modules, with R commutative, forms a symmetric monoidal category. Thus is well-defined.
Symmetry
In fact, for any permutation σ of the set {1, ..., n}, there is a unique isomorphism:
which is an isomorphism if either or is a pair of finitely generated projective modules.
To give a practical example, suppose M, N are free modules with bases and . Then M is the direct sum
and the same for N. By the distributive property, one has:
i.e., are the R-basis of . Even if M is not free, a free presentation of M can be used to compute tensor products.
The tensor product, in general, does not commute with inverse limit: on the one hand,
If R is not commutative, the order of tensor products could matter in the following way: we "use up" the right action of M and the left action of N to form the tensor product ; in particular, would not even be defined. If M, N are bi-modules, then has the left action coming from the left action of M and the right action coming from the right action of N; those actions need not be the same as the left and right actions of .
The associativity holds more generally for non-commutative rings: if M is a right R-module, N a (R, S)-module and P a left S-module, then
as abelian group.
The general form of adjoint relation of tensor products says: if R is not necessarily commutative, M is a right R-module, N is a (R, S)-module, P is a right S-module, then as abelian group[9]
The adjoint relation in the general form has an important special case: for any R-algebra S, M a right R-module, P a right S-module, using , we have the natural isomorphism:
This says that the functor is a left adjoint to the forgetful functor , which restricts an S-action to an R-action. Because of this, is often called the extension of scalars from R to S. In the representation theory, when R, S are group algebras, the above relation becomes the Frobenius reciprocity.
The structure of a tensor product of quite ordinary modules may be unpredictable.
Let G be an abelian group in which every element has finite order (that is G is a torsion abelian group; for example G can be a finite abelian group or ). Then:[10]
Indeed, any is of the form
If is the order of , then we compute:
Similarly, one sees
Here are some identities useful for calculation: Let R be a commutative ring, I, J ideals, M, NR-modules. Then
Example: If G is an abelian group, ; this follows from 1.
Example:; this follows from 3. In particular, for distinct prime numbers p, q,
Tensor products can be applied to control the order of elements of groups. Let G be an abelian group. Then the multiples of 2 in
are zero.
Example: Let be the group of n-th roots of unity. It is a cyclic group and cyclic groups are classified by orders. Thus, non-canonically, and thus, when g is the gcd of n and m,
Example: Consider . Since is obtained from by imposing -linearity on the middle, we have the surjection
whose kernel is generated by elements of the form
where r, s, x, u are integers and s is nonzero. Since
the kernel actually vanishes; hence, .
However, consider and . As -vector space, has dimension 4, but has dimension 2.
Thus, and are not isomorphic.
Example: We propose to compare and . Like in the previous example, we have: as abelian group and thus as -vector space (any -linear map between -vector spaces is -linear). As -vector space, has dimension (cardinality of a basis) of continuum. Hence, has a -basis indexed by a product of continuums; thus its -dimension is continuum. Hence, for dimension reason, there is a non-canonical isomorphism of -vector spaces:
Consider the modules for irreducible polynomials such that . Then,
Another useful family of examples comes from changing the scalars. Notice that
Good examples of this phenomenon to look at are when .
The construction of M ⊗ N takes a quotient of a free abelian group with basis the symbols m ∗ n, used here to denote the ordered pair(m, n), for m in M and n in N by the subgroup generated by all elements of the form
−m ∗ (n + n′) + m ∗ n + m ∗ n′
−(m + m′) ∗ n + m ∗ n + m′ ∗ n
(m · r) ∗ n − m ∗ (r · n)
where m, m′ in M, n, n′ in N, and r in R. The quotient map which takes m ∗ n = (m, n) to the coset containing m ∗ n; that is,
is balanced, and the subgroup has been chosen minimally so that this map is balanced. The universal property of ⊗ follows from the universal properties of a free abelian group and a quotient.
If S is a subring of a ring R, then is the quotient group of by the subgroup generated by , where is the image of under . In particular, any tensor product of R-modules can be constructed, if so desired, as a quotient of a tensor product of abelian groups by imposing the R-balanced product property.
More category-theoretically, let σ be the given right action of R on M; i.e., σ(m, r) = m · r and τ the left action of R of N. Then, provided the tensor product of abelian groups is already defined, the tensor product of M and N over R can be defined as the coequalizer:
where without a subscript refers to the tensor product of abelian groups.
In the construction of the tensor product over a commutative ring R, the R-module structure can be built in from the start by forming the quotient of a free R-module by the submodule generated by the elements given above for the general construction, augmented by the elements r ⋅ (m ∗ n) − m ∗ (r ⋅ n). Alternately, the general construction can be given a Z(R)-module structure by defining the scalar action by r ⋅ (m ⊗ n) = m ⊗ (r ⋅ n) when this is well-defined, which is precisely when r ∈ Z(R), the centre of R.
The direct product of M and N is rarely isomorphic to the tensor product of M and N. When R is not commutative, then the tensor product requires that M and N be modules on opposite sides, while the direct product requires they be modules on the same side. In all cases the only function from M × N to G that is both linear and bilinear is the zero map.
The dual module of a right R-module E, is defined as HomR(E, R) with the canonical left R-module structure, and is denoted E∗.[11] The canonical structure is the pointwise operations of addition and scalar multiplication. Thus, E∗ is the set of all R-linear maps E → R (also called linear forms), with operations
The dual of a left R-module is defined analogously, with the same notation.
There is always a canonical homomorphism E → E∗∗ from E to its second dual. It is an isomorphism if E is a free module of finite rank. In general, E is called a reflexive module if the canonical homomorphism is an isomorphism.
In the general case, each element of the tensor product of modules gives rise to a left R-linear map, to a right R-linear map, and to an R-bilinear form. Unlike the commutative case, in the general case the tensor product is not an R-module, and thus does not support scalar multiplication.
Given right R-module E and right R-module F, there is a canonical homomorphism θ : F ⊗RE∗ → HomR(E, F) such that θ(f ⊗ e′) is the map e ↦ f ⋅ ⟨e′, e⟩.[12]
Given left R-module E and right R-module F, there is a canonical homomorphism θ : F ⊗RE → HomR(E∗, F) such that θ(f ⊗ e) is the map e′ ↦ f ⋅ ⟨e, e′⟩.[13]
Both cases hold for general modules, and become isomorphisms if the modules E and F are restricted to being finitely generated projective modules (in particular free modules of finite ranks). Thus, an element of a tensor product of modules over a ring R maps canonically onto an R-linear map, though as with vector spaces, constraints apply to the modules for this to be equivalent to the full space of such linear maps.
Given right R-module E and left R-module F, there is a canonical homomorphism θ : F∗ ⊗RE∗ → LR(F × E, R) such that θ(f′ ⊗ e′) is the map (f, e) ↦ ⟨f, f′⟩ ⋅ ⟨e′, e⟩.[citation needed] Thus, an element of a tensor product ξ ∈ F∗ ⊗RE∗ may be thought of giving rise to or acting as an R-bilinear map F × E → R.
Let R be a commutative ring and E an R-module. Then there is a canonical R-linear map:
induced through linearity by ; it is the unique R-linear map corresponding to the natural pairing.
If E is a finitely generated projective R-module, then one can identify through the canonical homomorphism mentioned above and then the above is the trace map:
When R is a field, this is the usual trace of a linear transformation.
Example from differential geometry: tensor field[edit]
The most prominent example of a tensor product of modules in differential geometry is the tensor product of the spaces of vector fields and differential forms. More precisely, if R is the (commutative) ring of smooth functions on a smooth manifold M, then one puts
where Γ means the space of sections and the superscript means tensoring p times over R. By definition, an element of is a tensor field of type (p, q).
To lighten the notation, put and so .[15] When p, q ≥ 1, for each (k, l) with 1 ≤ k ≤ p, 1 ≤ l ≤ q, there is an R-multilinear map:
where means and the hat means a term is omitted. By the universal property, it corresponds to a unique R-linear map:
It is called the contraction of tensors in the index (k, l). Unwinding what the universal property says one sees:
Remark: The preceding discussion is standard in textbooks on differential geometry (e.g., Helgason). In a way, the sheaf-theoretic construction (i.e., the language of sheaf of modules) is more natural and increasingly more common; for that, see the section § Tensor product of sheaves of modules.
It can be shown that and are always right exact functors, but not necessarily left exact (, where the first map is multiplication by , is exact but not after taking the tensor with ). By definition, a module T is a flat module if is an exact functor.
If and are generating sets for M and N, respectively, then will be a generating set for Because the tensor functor sometimes fails to be left exact, this may not be a minimal generating set, even if the original generating sets are minimal. If M is a flat module, the functor is exact by the very definition of a flat module. If the tensor products are taken over a field F, we are in the case of vector spaces as above. Since all F modules are flat, the bifunctor is exact in both positions, and the two given generating sets are bases, then indeed forms a basis for .
If S and T are commutative R-algebras, then, similar to #For equivalent modules, S ⊗RT will be a commutative R-algebra as well, with the multiplication map defined by (m1 ⊗ m2) (n1 ⊗ n2) = (m1n1 ⊗ m2n2) and extended by linearity. In this setting, the tensor product become a fibered coproduct in the category of commutative R-algebras. (But it is not a coproduct in the category of R-algebras.)
If M and N are both R-modules over a commutative ring, then their tensor product is again an R-module. If R is a ring, RM is a left R-module, and the commutator
rs − sr
of any two elements r and s of R is in the annihilator of M, then we can make M into a right R module by setting
mr = rm.
The action of R on M factors through an action of a quotient commutative ring. In this case the tensor product of M with itself over R is again an R-module. This is a very common technique in commutative algebra.
For example, if C is a chain complex of flat abelian groups and if G is an abelian group, then the homology group of is the homology group of C with coefficients in G (see also: universal coefficient theorem.)
The tensor product of sheaves of modules is the sheaf associated to the pre-sheaf of the tensor products of the modules of sections over open subsets.
In this setup, for example, one can define a tensor field on a smooth manifold M as a (global or local) section of the tensor product (called tensor bundle)
The exterior bundle on M is the subbundle of the tensor bundle consisting of all antisymmetric covariant tensors. Sections of the exterior bundle are differential forms on M.
One important case when one forms a tensor product over a sheaf of non-commutative rings appears in theory of D-modules; that is, tensor products over the sheaf of differential operators.
^First, if , then the claimed identification is given by with . In general, has the structure of a right R-module by . Thus, for any -bilinear map f, f′ is R-linear .
^This is actually the definition of differential one-forms, global sections of , in Helgason, but is equivalent to the usual definition that does not use module theory.