Let G1,G2 be two Z-modules; a mapping u of the set G=G1×G2 into a Z-module is called biadditive (or Z-bilinear) if u(x1,x2) is "additive with respect to x1 and with respect to x2"; to be precise, this means that, for x1,y1 in G1, x2,y2 in G2,
u(x1+y1,x2)=u(x1,x2)+u(y1,x2)
u(x1,x2+y2)=u(x1,x2)+u(x1,y2).
Note that this implies in particular that u(0,x2)=u(x1,0)=0 for all x1∈G1,x2∈G2.
Let A be a ring, E a right A-module and F a left A-module. We are going to consider the universal mapping problem (Set Theory, IV, § 3, no. 1) where Σ is the species of Z-module structure (the morphisms then being Z-linear mappings, in other words, additive group homomorphisms) and the α-mappings are the mappings f of E×F into a Z-module G which are Z-bilinear and further satisfy, for all x∈E,y∈F and λ∈A
(1)
f(xλ,y)=f(x,λy).
We show that this problem admits a solution. For this we consider the Z-module C=Z(E×F) of formal linear combinations of the elements of E×F with coefficients in Z (§ 1, no. 11), a basis of which can be considered to consist of the ordered pairs (x,y), where x∈E and y∈F. Let D be the sub-Z-module of C generated by the elements of one of the following types:
where x,x1,x2 are in E, y,y1,y2 are in F and λ is in A.
Definition 1
The tensor product of the right A-module E and the left A-module F*, denoted by E⊗AF or E⊗AF (or simply E⊗F if no confusion is to be feared) is the quotient Z-moduleC/D (the quotient of the Z-module C of formal linear combinations of elements of E×F with coefficients in Z, by the submodule D generated by the elements of one of the types (2)). *For x∈E and y∈F, the element of E⊗AF which is the canonical image of the element (x,y) of C=Z(E×F) is denoted by x⊗y and called the tensor product of x and y.
(b) Conversely, let f be a Z-bilinear mapping of E×F into a Z-module G satisfying conditions (1). Then there exists one and only one Z-linear mapping g of E⊗AF into G such that f(x,y)=g(x⊗y) for x∈E,y∈F.
If ϕ denotes the canonical mapping of E×F into E⊗AF, then f=g∘ϕ; whence (a). To show (b), we note that, in the notation of Definition 1, f extends to a Z-linear mapping fˉ of C into G (§ 1, no. 11, Proposition 17). By virtue of relations (1), fˉ is zero for all the elements of C of one of the types (2) and hence on D. There therefore exists a Z-linear mapping g of C/D=E⊗AF into G such that fˉ=g∘ψ, where ψ:C→C/D is the canonical homomorphism (§ 1, no. 8, Remark). The uniqueness of g is immediate since E⊗AF is generated, as a Z-module, by the elements of the form x⊗y.
Proposition 1 defines a canonical isomorphism of the Z-module of Z-bilinear mappings f of E×F into G, satisfying conditions (1), onto the Z-module HomZ(E⊗AF,G).
When A=Z, conditions (1) are automatically satisfied for everyZ-bilinear mapping f and the submodule D of C is already generated by the elements of the first two types in (2).
If now we return to the general case and E′ and F′ denote the underlying Z-modules of E and F respectively, the above remark and Definition 1 show immediately that the Z-module E⊗AF can be canonically identified with the quotient of the Z-module E′⊗ZF′ by the sub-Z-module generated by the elements of the form (xλ)⊗y−x⊗(λy), where x runs through E, y runs through F and λ runs through A.
Corollary 1
Let H be a Z-module and h:E×F→H a Z-bilinear mapping satisfying conditions (1) and such that H is generated by h(E×F). Suppose that for every Z-module G and every Z-bilinear mapping f of E×F into G satisfying (1)* there exists a Z-linear mapping g:H→G such that f=g∘h. Then, if ϕ denotes the canonical mapping of E×F into E⊗AF, there exists one and only one isomorphism θ of E⊗AF onto H such that h=θ∘ϕ.
The hypothesis that h(E×F) generates H implies the uniqueness of g; the corollary is then just the general uniqueness property of a solution of a universal mapping problem (Set Theory, IV, § 3, no. 1).
Corollary 2
Let* E0 (resp. F0) denote the moduleE (resp. F) considered as a left (resp. right) module over the opposite ringA0; then there exists one and only oneZ*-module isomorphism* σ:E⊗AF→F0⊗A0E0such thatσ(x⊗y)=y⊗xforx∈Eandy∈F ("commutativity" of tensor products).
By definition of the A0-module structures on E0 and F0, the mapping (x,y)↦y⊗x of E×F into F0⊗A0E0 is Z-bilinear and satisfies conditions (1), whence the existence and uniqueness of the Z-linear mapping σ. Similarly a Z-linear mapping τ:F0⊗A0E0→E⊗AF is defined such that τ(y⊗x)=x⊗y and clearly σ and τ are inverse isomorphisms.
Remark
The tensor product of non-zero modules may be zero: for example, taking the two Z-modules E=Z/2Z and F=Z/3Z, 2x=0 and 3y=0 for all x∈E and y∈F; therefore, in E⊗ZF,
Let A be a ring, E,E′ two right A-modules, F,F′ two left A-modules and u:E→E′ and v:F→F′ two A-linear mappings. It is easily verified that the mapping
(x,y)↦u(x)⊗v(y)
of E×F into E′⊕AF′ is Z-bilinear and satisfies conditions (1) of no. 1. By Proposition 1 of no. 1 there thus exists one and only one Z-linear mapping w:E⊗AF→E′⊗AF′ such that
w(x⊗y)=u(x)⊗v(y)
for x∈E,y∈F. This mapping is denoted by u⊗v (when no confusion can arise) and is called the tensor product of the linear mappings u and v.
It follows immediately from (3) that (u,v)↦u⊗v is a Z-bilinear mapping called canonical
HomA(E,E′)×HomA(F,F′)→HomZ(E⊗AF,E′⊗AF′).
There corresponds to it by Proposition 1 of no. 1 a Z-linear mapping called canonical
(4)
HomA(E,E′)⊗ZHomA(F,F′)→HomZ(E⊗AF,E′⊗AF′)
which associates with every element u⊗v of the tensor product the linear mapping u⊗v:E⊗AF→E′⊗AF′. Note that the canonical mapping (4) is not necessarily injective nor surjective. The notation u⊗v can therefore lead to confusion and it will be necessary for the context to indicate whether it denotes an element of the tensor product or a linear mapping.
Further, let E′′ be a right A-module, F′′ a left A-module and u′:E′→E′′, v′:F′→F′′A-linear mappings; it follows from (3) that (5)
(u′∘u)⊗(v′∘v)=(u′⊗v′)∘(u⊗v).
3. Change of ring
Proposition 2
Let A,B be two rings, ρ:B→A a ring homomorphism and E (resp. F) a right (resp. left) A-module. Then there exists one and only one Z-linear mapping* (6)
ϕ:ρ∗(E)⊗Bρ∗(F)→E⊗AF
*such that, for all x∈E and y∈F, the image under ϕ of the element x⊗y of ρ∗(E)⊗Bρ∗(F) is the element x⊗y of E⊗AF; this Z-linear mapping is surjective.
Cites 1 · Cited by 1
We consider the mapping (x,y)↦x⊗y of ρ∗(E)×ρ∗(F) into E⊗AF; it is Z-bilinear and, for all β∈B, by definition (xρ(β))⊗y=x⊗(ρ(β)y), hence conditions (1) of no. 1 hold, whence the existence and uniqueness of ϕ (no. 1, Proposition 1). The latter assertion follows from the fact that the elements x⊗y generate the Z-module E⊗AF.
The mapping (6) is called canonical.
Corollary
Let J be a two-sided ideal of A such that J is contained in the annihilator of E and in the annihilator of F, so that E (resp. F) has a canonical left (resp. right) (A/J)-module structure* (§1, no. 12). Then the canonical homomorphism (6)
ϕ:E⊗AF→E⊗A/JF
*corresponding to the canonical homomorphism ρ:A→A/J is the identity.
Cites 1 · Cited by 1
For all αˉ∈A/J, all x∈E and all y∈F, xαˉ=xα (resp. αˉy=αy) for all α such that ρ(α)=αˉ. If C=Z(E×F), the submodule of C generated by the elements (xα,y)−(x,αy) is then equal to the submodule generated by the elements (xαˉ,y)−(x,αˉy).
With the hypotheses and notation of Proposition 2, let E′ be a right B-module, F′ a left B-module and consider two semi-linear mappings u:E′→E, v:F' \to F relative to the homomorphism ρ:B→A; u (resp. v) can be considered as a B-linear mapping E′→ρ∗(E) (resp. F′→ρ∗(F′)), whence a Z-linear mapping w:E′⊗BF′→ρ∗(E)⊗Bρ∗(F) such that w(x′⊗y′)=u(x′)⊗v(y′) for x′∈E′,y′∈F′; composing the canonical mapping (6) with this mapping, a Z-linear mapping w′:E′⊗BF′→E⊗AF is hence obtained such that w′(x′⊗y′)=u(x′)⊗v(y′) for x′∈E′,y′∈F′; this is the mapping which will normally be denoted by u⊗v if no confusion can arise. Clearly (u,v)↦u⊗v is a Z-bilinear mapping
Moreover, if C is a third ring, σ:C→B a homomorphism, E′′ a right C-module, F′′ a left C-module, u′:E′′→E′ and v′:F′′→F′ semi-linear mappings relative to σ, then
(u∘u′)⊗(v∘v′)=(u⊗v)∘(u′⊗v′).
4. Operators on a tensor product; tensor products as multimodules
With the hypotheses and notation of no. 1, for every endomorphism u (resp. v) of the A-module E (resp. F), u⊗1F (resp. 1E⊗v) is an endomorphism of the Z-module E⊗AF; it follows immediately from (5) (no. 2) that the mapping u↦u⊗1F (resp. v↦1E⊗v) is a ring homomorphism
EndA(E)→EndZ(E⊗AF)
(resp. EndA(F)→EndZ(E⊗AF)); moreover,
(u⊗1F)∘(1E⊗v)=(1E⊗v)∘(u⊗1F)=u⊗v
and therefore (§ 1, no. 14) E⊗AF has a canonical left bimodule structure with respect to the rings EndA(E) and EndA(F).
This being so, suppose given on E a ((Bi′);A,(Cj′))-multimodule structure and on F a (A,(Bh′′);(Ckc′′))-multimodule structure (§ 1, no. 14); it amounts to the same to say that ring homomorphisms Bi′→EndA(E), Cj′0→EndA(E) are given with pairwise permutable images, and ring homomorphisms Bh′′→EndA(F), Ckc′′0→EndA(F) with pairwise permutable images. If these homomorphisms are composed respectively with the canonical homomorphisms EndA(E)→EndZ(E⊗AF) and EndA(F)→EndZ(E⊗AF) defined above, it is seen (taking account of (7)) that ring homomorphisms
are defined with pairwise permutable images; in other words, there has been defined on E⊗AF a ((Bi′),(Bh′′);(Cj′),(Ckc′′))-multimodule structure; it is this multimodule which is also called the tensor product (relative to A) of the
((B'_i); A, (C'_j))-multimodule E and the (A, (B''_h); (C''_k))-multimodule F. This multimodule is the solution of a universal mapping problem analogous to that considered in no. 1; to be precise:
Proposition 3
Let G be a ((B'_i), (B''_h); (C'_j), (C''_k))-multimodule.
(a) Let g be a linear mapping of the multimodule E ⊗A F into G. The mapping f:(x,y)↦g(x⊗y) of E × F into G is Z-bilinear and satisfies relations (1) of no. 1 and the conditions
for all x∈E,y∈F,μi′∈Bi′,vj′∈Cj′,μh′′∈Bh′′,vk′′∈Ck′′,i,j,h,k arbitrary.
(b) Conversely, let f be a Z-bilinear mapping of E × F into G satisfying conditions (1) (no. 1) and (8). Then there exists one and only one linear mapping g of the multimodule E ⊗A F into the multimodule G such that f(x,y)=g(x⊗y) for x∈E,y∈F.
Assertion (a) follows immediately from the definition of the multimodule structure on E ⊗A F, since for example (x⊗y)vj′=(xvj′)⊗y. To prove (b), we note first that Proposition 1 of no. 1 gives the existence and uniqueness of a Z-linear mapping g such that g(x⊗y)=f(x,y) for x∈E,y∈F; all that is needed is to verify that g is linear for the multimodule structures. As the elements x⊗y generate the Z-module E ⊗A F, it suffices to verify the relations g(μ′(x⊗y))=μi′g(x⊗y) and their analogues; but this follows immediately from the formula g(x⊗y)=f(x,y) and relations (8).
Scholium
An element of E ⊗A F may in general be written in several ways in the form ∑i(xi⊗yi), where xi∈E and yi∈F; but to define a linear mapping g of the multimodule E ⊗A F into a multimodule G, there is no need to verify that, if ∑i(xi⊗yi)=∑j(xj′⊗yj′), then ∑ig(xi⊗yi)=∑jg(xj′⊗yj′); it suffices to be given g(x⊗y) for x∈E and y∈F and to verify that (x,y)↦g(x⊗y) is Z-bilinear and satisfies conditions (1) (no. 1) and (8).
Let E' be a ((B'_i), A, (C'_j))-multimodule, F' an (A, (B''_h); (C''_k))-multimodule and u:E→E′,v:F→F′ linear mappings of multimodules; it follows immediately from the definitions (no. 2) that u⊗v is a linear mapping of the multimodule E ⊗A F into the multimodule E' ⊗A F'.
With E always denoting a right A-module, let sAd denote the ring A considered as an (A, A)-bimodule (§ 1, no. 14, Example 1); by the above, the tensor product E ⊗A(sAd) has a canonical right A-module structure such that (x⊗λ)μ=x⊗(λμ) for x∈E,λ∈A,μ∈A. The mapping (x,λ)↦xλ of E ×(sAd) into E is Z-bilinear and satisfies conditions (1) (no. 1) and (8) (where, in the latter, the Bi′,Cj′ and Bh′′ are absent and the family (Ck′′) reduces to A); hence (Proposition 3), there exists an A-linear mapping g (called canonical) of E⊗A(sAd) into E such that g(x⊗λ)=xλ for x∈E,λ∈A.
If E is a right A-module, the mapping h:x↦x⊗1 of E into E⊗A(sAd) is a right A-module isomorphism, whose inverse isomorphism g is such that g(x⊗λ)=xλ for x∈E,λ∈A.
If g is the canonical mapping, g∘h is the identity mapping 1E and h∘g coincides with the identity mapping of E⊗A(sAd) onto itself for elements of the form x⊗y, which generate the latter Z-module; hence the conclusion.
We shall normally write E⊗AA instead of E⊗A(sAd) and shall often identify E⊗AA with E by means of the above canonical isomorphisms. Observe that, if E also has a (left or right) B-module structure which is compatible with its right A-module structure, g and h are also isomorphisms for the B-module structures on E and E⊗AA (and hence multinode module isomorphisms).
Now let F be a left A-module; (sAd)⊗AF (also denoted by A⊗AF) then has a canonical left A-module structure and as in Proposition 4 a canonical isomorphism is defined of A⊗AF onto F mapping λ⊗x to λx, and its inverse isomorphism is x↦1⊗x.
In particular there exists a canonical isomorphism of the (A,A)-bimodule (sAd)⊗A(sAd) onto sAd which maps λ⊗μ to λμ.
5. Tensor product of two modules over a commutative ring
Let C be a commutative ring; for every C-module E, the module structure on E is compatible with itself (§ 1, no. 14). If E and F are two C-modules, the considerations of no. 4 then allow us to define twoC-module structures on the tensor product E⊗CF, respectively such that γ(x⊗y)=(γx)⊗y and such that γ(x⊗y)=x⊗(γy); but as, by Definition 1 of no. 1, in this case (γx)⊗y=x⊗(γy), these two structures are the same. Henceforth when we speak of E⊗CF as a C-*module*, we shall mean with the structure thus defined, unless otherwise stated. The canonical isomorphism.
It follows from this definition that, if (aλ)λ∈L (resp. (bμ)μ∈M) is a generating system of the C-module E (resp. F), (aλ⊗bμ) is a generating system of the C-module E⊗CF; in particular, if E and F are finitely generatedC-modules, so is E⊗CF.
For every C-module G, the Z-bilinear mappings f of E×F into G for which
f(γx,y)=f(x,γy)=γf(x,y)for x∈E,y∈F,γ∈C
are then called C-bilinear and form a C-module denoted by L2(E,F;G); Proposition 3 (no. 4) defines a canonical C-module isomorphism (cf. § 1, no. 14, Remark 1).
L2(E,F;G)→HomC(E⊗CF,G).
Let E', F' be two C-modules and u:E→E′, v:F→F′ two C-linear mappings; then (no. 4) u⊗v is a C-linear mapping of E⊗CF into E′⊗CF′. Further, it is immediate that (u,v)↦u⊗v is a C-bilinear mapping of HomC(E,E′)×HomC(F,F′) into HomC(E⊗CF,E′⊗CF′); hence there canonically corresponds to it a C-linear mapping, called canonical:
HomC(E,E′)⊗CHomC(F,F′)→HomC(E⊗CF,E′⊗CF′)
which associates with every element u⊗v of the tensor product
HomC(E,E′)⊗CHomC(F,F′)
the linear mapping u⊗v. Note that the canonical mapping (11) is not necessarily injective nor surjective (§ 4, Exercise 2).
Remark
(1) Let A, B be two commutative rings, ρ:B→A a ring homomorphism and E and F two A-modules; then the canonical mapping (6) of no. 3 is a B-linear mapping
ρ∗(E)⊗ρ∗(F)→ρ∗(E⊗AF).
(2) What has been said in this no. can be generalized to the following case: let E be a right A-module, F a left A-module, C a commutative ring and ρ:C→A a homomorphism of C into A such that ρ(C) is contained in the centre of A (cf. III, § 1, no. 3). We may then consider the C-modules ρ∗(E) and ρ∗(F) and the hypothesis on ρ implies that these C-module structures are compatible respectively with the A-module structures on E and F (§ 1, no. 14). The tensor product E⊗AF is thus (by virtue of no. 4) given two C-module structures such that γ(x⊗y)=(xρ(γ))⊗y and
γ(x⊗y)=x⊗(ρ(γ)y)
respectively for γ∈C, x∈E, y∈F and Definition 1 (no. 1) shows also that these two structures are identical. If E' (resp. F') is a right (resp. left) A-module and u:E→E′, v:F→F′ are two A-linear mappings, then u⊗v:E⊗AF→E′⊗AF′ is C-linear for the C-module structures just defined; the mapping (u,v)↦u⊗v:
HomA(E,E′)×HomA(F,F′)→HomC(E⊗AF,E′⊗AF′)
is C-bilinear (for the C-module structures on HomA(E,E′) and HomA(F,F′)) defined in § 1, no. 14, Remark 1), whence we also derive a C-linear mapping, called *canonical
(3) Let A be an integral domain and K its field of fractions. If E and F are two vector K-spaces, the canonical mapping
(E[A])⊗A(F[A])→E⊗KF
(no. 3 and § 1, no. 13) is bijective. It suffices (no. 4) to prove that if f is an A-bilinear mapping of E×F into a vector K-space G, f is also K-bilinear. Now, for all α=0 in A.
By virtue of no. 2, formula (5), vˉ∘uˉ=(v∘u)⊗1F=0; the image H=uˉ(E′⊗F) is contained in the kernel L=Ker(vˉ); by passing to the quotient, we therefore derive from vˉ a Z-linear mapping f of the cokernel M=(E⊗F)/H of uˉ into E′′⊗F; it must be proved that f is bijective and it will hence suffice to define a Z-linear mapping g:E′′⊗F→M such that g∘f and f∘g are the identity mappings.
Let x′′∈E′′, y∈F; by hypothesis there exists x∈E such that v(x)=x′′. We show that, if x1,x2 are two elements of E such that v(x1)=v(x2)=x′′ and ϕ:E⊗F→M is the canonical mapping, then ϕ(x1⊗y)=ϕ(x2⊗y). It suffices to prove that if v(x)=0 then ϕ(x⊗y)=0, which follows from the fact that x=u(x′) with x′∈E′, whence x⊗y=u(x′)⊗y=uˉ(x′⊗y)∈H. If (x′′,y) is mapped to the unique value of ϕ(x⊗y) for all x∈E such that v(x)=x′′, a mapping is defined of E′′×F into M; this mapping is Z-bilinear and satisfies conditions (1) (no. 1), since v(xλ)=x′′λ and (xλ)⊗y=x⊗(λy) for x∈E; hence there is a Z-linear mapping g of E′′⊗F into M such that g(x′′⊗y)=ϕ(x⊗y) for y∈F,x∈E and x′′=v(x). This definition further proves that f∘g coincides with the identity mapping for the elements of
E′′⊗F of the form x′′⊗y and hence f∘g is the identity mapping of E′′⊗F; on the other hand, for x∈E and y∈F, f(ϕ(x⊗y))=v(x)⊗y by definition, hence g(f(ϕ(x⊗y)))=ϕ(x⊗y) and, as the elements of the form ϕ(x⊗y) generate M, g∘f is the identity mapping of M.
Corollary
Let F, F', F'' be left A-modules, E a right A-module and*
F′sFtF′′⟶0
(16)an exact sequence of linear mappings. Writing \bar{s} = lE \otimes s, \bar{t} = l_E \otimes t, the sequence of \mathbf{Z}-homomorphisms*
When E (resp. F) is considered as a left (resp. right) A^0-module, F \otimes_{A^0} E is identified with E \otimes_A F and there are analogous identifications for F' \otimes_{A^0} E and F'' \otimes_{A^0} E (no. 1, Corollary 2 to Proposition 1); the corollary then follows immediately from Proposition 5.
Remark
Note that in general, if E' is a submodule of a right A-module E and j : E' \to E the canonical injection, the canonical mapping
j⊗1F:E′⊗F→E⊗F
is not necessarily injective. In other words, for an exact sequence
0⟶E′uEvE′′⟶0
(18) it cannot in general be concluded that the sequence
0⟶E′⊗FuE⊗FvE′′⊗F⟶0
(19) *is exact.
Take for example A = \mathbf{Z}, E = \mathbf{Z}, E' = 2\mathbf{Z}, F = \mathbf{Z}/2\mathbf{Z}. As E' is isomorphic to E, E' \otimes F is isomorphic to E \otimes F which is itself isomorphic to F (no. 4, Proposition 4). But for all x' = 2x \in E' (where x \in E) and all y \in F, j(x') \otimes y = (2x) \otimes y = x \otimes (2y) = 0, since 2y = 0, and the canonical image of E' \otimes F in E \otimes F reduces to 0.
In other words, care must be taken to distinguish, for a submodule E' of E and an element x \in E', between the element x \otimes y "calculated in E' \otimes F'" and the element x \otimes y "calculated in E \otimes F" (in other words, the element j(x) \otimes y).
We shall study later, under the name of flat modules, the modules F such that the sequence (19) is exact for every exact sequence (18) (Commutative Algebra, I, § 2).
Now v⊗t=(v⊗1F′′)∘(1E⊗t) (no. 2, formula (5)) and v⊗t is therefore surjective, being the composition of two surjective homomorphisms by virtue of Proposition 5 and its Corollary. On the other hand, for z∈E⊗F to be in the kernel of v⊗t, it is necessary and sufficient that (1E⊗t)(z) belong to the kernel of v⊗1F′′, that is, by virtue of (15) to the image of
u⊗1F′′:E′⊗F′′→E⊗F′′.
But as the homomorphism t:F→F′′ is surjective, so is
1E′⊗t:E′⊗F→E′⊗F′′
by the Corollary to Proposition 5, hence the condition on z reduces to the existence of an a∈E′⊗F such that
(lE⊗t)(z)=(u⊗t)(a).
Let b=z−(u⊗1F)(a); then (lE⊗t)(b)=0, and by virtue of (17), b belongs to the image of lE⊗s, which proves the proposition.
In other words:
Corollary 1
Let E′ be a submodule of a right A-module E, F′ a submodule of a left A-module F and Im(E′⊗AF) and Im(E⊗AF′) the sub-Z-modules of E⊗AF, the respective images of the canonical mappings E′⊗AF→E⊗AF,
*such that, for ξ∈E/E′, η∈F, π(ξ⊗η) is the class of all elements x⊗y∈E⊗AF such that x∈ξ and y∈η.
Note that when E is a ((Bi′);A,(Cj′))-multimodule, F a (A,(Bh′′);(Ck′′))-multimodule and E′ and F′submultimodules of E and F respectively, the isomorphism (20) is an isomorphism for the ((Bi′),(Bh′′);(Cj′),(Ck′′))-multimodule structures of the two sides (no. 3).
Corollary 2
Let a be a right ideal of A, F a left A-module and aF the sub-Z-module of F generated by the elements of the form λx, where λ∈a and x∈F. Then there is a canonical Z-module isomorphism
*such that, for all λˉ∈A/a and all x∈F, π(λˉ⊗x) is the class mod. aF of λx, where λ∈λˉ.
In particular, for A=Z, it is seen that for every integer n and every Z-module F, (Z/nZ)⊗ZF is canonically identified with the quotient Z-module F/nF.
Corollary 3
Let A be a commutative ring, a an ideal of A and E and F two A-modules such that a is contained in the annihilator of F. Then the (A/a)-modules E⊗AF and (E/aE)⊗A/aF are canonically isomorphic.
F and E⊗AF are annihilated by a and hence have canonical (A/a)-module structures (§1, no. 12) and if we write E′=aE, then Im(E′⊗AF)=0; then there is a canonical isomorphism (20) of E⊗AF onto (E/aE)⊗AF and the latter is itself identical with (E/aE)⊗A/aF (no. 3, Corollary to Proposition 2).
Corollary 4
Let a,b be two ideals in a commutative ring C; the C-module (C/a)⊗C(C/b) is then canonically isomorphic to C(a+b).
Let (Eλ)λ∈L be a family of right A-modules, (Fμ)μ∈M a family of left A-modules and consider the product modules C=∏λ∈LEλ,D=∏μ∈MFμ. The mapping ((xλ),(yμ))↦(xλ⊗yμ) of C×D into the product Z-module
(λ,μ)∈L×M∏(Eλ⊗AFμ)
is Z-bilinear and obviously satisfies conditions (1) (no. 1). Thus there exists (no. 1, Proposition 1) a Z-linear mapping, called canonical
f:(λ∈L∏Eλ)⊗Aμ∈M∏Fμ→(λ,μ)∈L×M∏(Eλ⊗AFμ)
such that f((xλ)⊗(yμ))=(xλ⊗yμ).
When C=RL,D=SM,R (resp. S) being a right (resp. left) A-module, the canonical mapping (22) associates with every tensor product u⊗v, where u is a mapping of L into R and v a mapping of M into S, the mapping (λ,μ)↦u(λ)⊗v(μ) of L×M into R⊗AS; even in this case the canonical mapping (22) is in general neither injective nor surjective (Exercise 3; cf. Corollary 3 to Proposition 7).
When the Eλ are ((Bi′);A,(Cj′))-multimodules and the Fμ(A,(Bh′′);(Ckc′′))-multimodules, the homomorphism (22) is also a homomorphism for the ((Bi′),(Bh′′);(Cj′),(Ckc′′))-multimodule structures of the two sides.
Consider now the submodule E=⨁λ∈LEλ (resp. ⨁μ∈MFμ) of C (resp. D); the canonical injections E→C,F→D define canonically a Z-linear mapping E⊗AF→C⊗AD which, composed with the mapping (22), gives a Z-linear mapping g of E⊗AF into ∏λ,μ(Eλ⊗AFμ) such that
g((xλ)⊗(yμ))=(xλ⊗yμ);
moreover, as the families (xλ) and (yμ) have finite support, so does (xλ⊗yμ) and hence finally g is a canonical homomorphism
g:(λ∈L⨁Eλ)⊗Aμ∈M⨁Fμ→(λ,μ)∈L×M⨁(Eλ⊗AFμ),
which is a multimodule homomorphism under the same conditions as (22).
To prove this it suffices to define a Z-linear mapping h of the direct sum G=⨁(λ,μ)∈L×M(Eλ⊗AFμ) into E⊗AF such that g∘h and h∘g are the identity mappings. But, to define a Z-linear mapping of G into E⊗AF, it suffices (§ 1, no. 6, Proposition 6) to define a Z-linear mapping
hλμ:Eλ⊗AFμ→E⊗AF
for every ordered pair (λ,μ), and we take hλμ=iλ⊗jμ, where iλ:Eλ→E and jμ:Fμ→F are the canonical injections. Then clearly h∘g coincides with the identity mapping for the elements of the form (∑λxλ)⊗(∑μyμ) which generate the Z-module E⊗AF; similarly g∘h coincides with the identity mapping for the elements of the form ∑λ,μ(xλ⊗yμ) which generate the Z-module G, since for each ordered pair (λ,μ) the products
xλ⊗yμ(xλ∈Eλ,yμ∈Fμ)
generate the Z-module Eλ⊗AFμ. Whence the proposition.
Let uλ:Eλ→Eλ′, vμ:Fμ→Fμ′ be A-homomorphisms; clearly the diagram
If the left A-module F admits a basis (bμ)μ∈M, every element of E⊗AF can be written uniquely in the form ∑μ(xμ⊗bμ), where xμ∈E and the family (xμ) has finite support. The Z-module E⊗AF is isomorphic to E(M) considered as a Z-module.
The basis (bμ) defines an isomorphism of F onto ⨁μ∈MAbμ, whence there is an isomorphism E⊗AF→⨁μ∈M(E⊗AAbμ) by virtue of Proposition 7; as ξ ↦ ξb_μ is an isomorphism of A_s onto Ab_μ, x ↦ x ⊗ b_μ is an isomorphism of E onto E ⊗_A (Ab_μ) by virtue of Proposition 4 of no. 4, whence the corollary.
If E is a ((B'_i); A, (C'_j))-multimodule, the canonical isomorphism E ⊗_A F → E^{(M)} is a ((B'_i); (C'_j))-multimodule isomorphism.
In particular, if also E admits a basis (a_λ)_{λ ∈ L}, every z ∈ E ⊗_A F may be written in one and only one way in the form ∑λ,μ(aλξλμ)⊗bμ, where the ξ_{λμ} belong to A (and form a family of finite support); the mapping z↦(ξλμ)(λ,μ)∈L×M is an isomorphism of E ⊗_A F onto A^{(L × M)} for the \mathbf{Z}-module structures (and even the module structures over the centre of A). More particularly:
Corollary 2
If E and F are two free modules over a commutative ring C and (a_λ) (resp. (b_μ)) is a basis of the C-module E (resp. F), then (a_λ ⊗ b_μ) is a basis of the C-module E ⊗_C F.
By an abuse of language, the basis (a_λ ⊗ b_μ) is sometimes called the tensor product of the bases (a_λ) and (b_μ).
Remark (1). Let E be a free right A-module, F a free left A-module, (a_λ)_{λ ∈ L} a basis of E and (b_μ)_{μ ∈ M} a basis of F. Every element z ∈ E ⊗_A F may be written uniquely as ∑λaλ⊗yλ, where y_λ ∈ F, and also uniquely as ∑μxμ⊗bμ, where x_μ ∈ E. If we write yλ=∑μηλμbμ, xμ=∑λaλξλμ, where the ξ_{λμ} and η_{λμ} belong to A, then ξλμ=ηλμ for all (λ, μ), for
If (b_μ) is a basis of F, every element of (∏λ∈LEλ)⊗AF can be written uniquely as z=∑μ((xλ(μ))⊗bμ) (Corollary 1); to say that its canonical image is zero means that, for all λ ∈ L, ∑μ(xλ(μ)⊗bμ)=0, hence xλ(μ)=0 for all λ ∈ L and all μ (Corollary 1) and therefore z=0.
Showing that the canonical mapping is bijective when F admits a finite basis is immediately reduced, by virtue of Proposition 7, to the case where
F = A_s; but then the two sides are canonically identified with ∏λ∈LEλ (no. 4, Proposition 4) and after these identifications the canonical mapping (22) becomes the identity.
Corollary 4
Let A be a ring with no divisor of zero, E a free right A-module and F a free left A-module. Then the relation x ⊗ y = 0 in E ⊗A F implies x = 0 or y = 0.
Let (aλ) be a basis of E, (bμ) a basis of F and let x=∑λaλξλ,y=∑μημbμ; then x⊗y=∑λ,μ((aλξλημ)⊗bμ and the relation x⊗y=0 implies ξλημ=0 for every ordered pairs of indices (λ,μ) (Corollary 1). Hence, if x=0, that is ξλ=0 for at least one λ, it follows that ημ=0 for all μ, whence y=0.
Corollary 5
Let E be a right A-module, F a left A-module, M a submodule of E and N a submodule of F. If M is a direct factor of E and N a direct factor of F, the canonical homomorphism M⊗AN→E⊗AF is injective and the image of M⊗AN under this homomorphism is a direct factor of the Z-module E⊗AF.
Note that if E is a ((Bi′);A,(Cj′))-multimodule and F an (A,(Bh′′);(Ck′′))-multimodule and M and N direct factors in these multimodules, M⊗N is a direct factor of the ((Bi′),(Bh′′);(Cj′),(Ck′′))-multimodule E⊗F.
Corollary 6
Let P be a projective left A-module and E, F two right A-modules. For every injective homomorphism u:E→F, the homomorphism*
This follows immediately from Corollary 5 and the fact that the tensor product of two free C-modules is a free C-module (Corollary 2).
Remark 2
Under the hypotheses of Proposition 7, let Eλ′ be a submodule of Eλ, Fμ′ a submodule of Fμ and let E′=⨁λ∈LEλ′, F′=⨁μ∈MFμ′. Let Im(E′⊗AF′) (resp. Im(E'_λ ⊗_A F'_μ)) denote the image of E' ⊗_A F' (resp. E'_λ ⊗_A F'_μ) in E ⊗_A F (resp. E_λ ⊗_A F_μ) under the canonical mapping; then the isomorphism (23) identifies the sub-Z-modules
Im(E′⊗AF′)and(λ,μ)∈I×M⨁Im(Eλ′⊗AFμ′);
this follows immediately from the commutativity of the diagram
Let A, B be two rings, E a right A-module, F an (A, B)-bimodule and G a left B-module. Then E ⊗_A F is a right B-module, F ⊗_B G a left A-module and there exists one and only one Z-linear mapping
such thatϕ((x⊗y)⊗z)=x⊗(y⊗z)forx∈E,y∈F,z∈G; moreover this Z-linear mapping is bijective ("associativity" of the tensor product).
The right B-module structure on E ⊗_A F and left A-module structure on F ⊗_B G have been defined in no. 4. The uniqueness of ϕ is obvious since the elements (x⊗y)⊗z generate the Z-module (E⊗AF)⊗BG. To show the existence of ϕ, we note that, for all z∈G,hz:y↦y⊗z is an A-linear mapping of the left A-module F into the left A-module F ⊗_B G. We write gz=1E⊗hz, which is therefore a Z-linear mapping of E ⊗_A F into E ⊗_A (F ⊗_B G) and consider the mapping (t,z)↦gz(t) from (E⊗AF×G into E⊗A(F⊗BG); as hz+z′=hz+hz′ for z∈G,z′∈G, it is immediate that the above mapping is Z-bilinear. Further, we show that for all μ∈B,gμz(t)=gz(tμ); it is obviously sufficient to do this for t=x⊗y where x∈E and y∈F; now and
gμz(x⊗y)=x⊗(y⊗μz)
gz((x⊗y)μ)=gz(x⊗yμ)=x⊗(yμ⊗z).
Proposition 1 (no. 1) then proves the existence of a Z-linear mapping
ϕ:(E⊗AF)⊗BG→E⊗A(F⊗BG)
such that ϕ(t⊗z)=gz(t), hence ϕ((x⊗y)⊗z)=x⊗(y⊗z). Similarly a Z-linear mapping
ψ:E⊗A(F⊗BG)→(E⊗AF)⊗BG
is defined such that ψ(x⊗(y⊗z))=(x⊗y)⊗z and clearly ψ∘ϕ and ϕ∘ψ are the identity mappings of (E⊗AF)⊗BG and E⊗A(F⊗BG) respectively, since they reduce to the identity mappings on generating systems of these Z-modules.
It is immediate that, if E is a ((Ci′);A,(Dj′))-multimodule, F an (A,(Ch′′);B,(Dk′′))-multimodule and G a (B,(Cl′′′);(Dm′′′))-multimodule, the canonical isomorphism defined in Proposition 8 is a ((Ci′),(Ch′′),(Cl′′′);(Dj′),(Dk′′),(Dm′′′))-multimodule isomorphism. In particular, if C is a commutative ring and E,F,G three C-modules, there is a canonical C-module isomorphism
(E⊗CF)⊗CG→E⊗C(F⊗CG).
We shall see below that, under certain conditions, the definition of tensor product can be generalized to a family of multimodules, which will in particular give us under the hypotheses of Proposition 8 a Z-module E⊗AF⊗BG, which is canonically isomorphic to each of the Z-modules (E⊗AF)⊗BG and E⊗A(F⊗BG) and with which the latter will be identified.
9. Tensor product of families of multimodules
Let (Gλ)λ∈L be a family of Z-modules; a mapping u of the set G=∏λ∈LGλ into a Z-module is said to be multiadditive (or Z-*multilinear*) if (xλ)↦u((xλ)) is additive with respect to each of the variables xλ; to be precise, this means that, for all μ∈L and every element (aλ)∈∏λ=μGλ, canonically identifying G with Gμ×∏λ=μGλ,
u(xμ+yμ,(aλ))=u(xμ,(aλ))+u(yμ,(aλ))for xμ,yμ in Gμ.
This implies in particular that u((xλ))=0 if one of the xλ is zero.
We consider also the universal mapping problem where Σ is the species of Z-module structure and the α-mappings are the multiadditive mappings of G into a Z-module. A solution is still obtained by considering the Z-module C=Z(G) of formal linear combinations of elements of G with coefficients in Z and the sub-Z-module D of C generated by the elements of the form
where μ∈L,xμ∈Gμ,yμ∈Gμ and the zλ∈Gλ (λ=μ) are arbitrary. The quotient Z-module C/D is called the tensor product (over Z) of the family(Gλ)λ∈L of Z-*modules* and is denoted by ⨂λ∈LGλ; for every element (xλ)λ∈L of G which is an element of the canonical basis of C, ⨂λ∈Lxλ denotes the canonical image of this element in C/D. It follows from the above definitions that the mapping ϕ:(xλ)↦⨂λ∈Lxλ of G into ⨂λ∈LGλ is Z-multilinear and that, for every Z-multilinear mapping f of G into a Z-module H, there exists one and only one Z-linear mapping g:⨂λ∈LGλ→H such that f=g∘ϕ; the ordered pair (⨂λ∈LGλ,ϕ) thus resolves the universal mapping problem in question.
Let (Gλ′)λ∈L be another family of Z-modules and, for all λ∈L, let vλ:Gλ→Gλ′ be a Z-linear mapping (in other words a homomorphism of commutative groups). Then the mapping
(xλ)↦λ∈L⨂vλ(xλ)
of G into ⨂λ∈LGλ′ is Z-multilinear and hence defines canonically a Z-linear mapping of ⨂λ∈LGλ into ⨂λ∈LGλ′ denoted by ⨂λ∈Lvλ and such that
(λ∈L⨂vλ)(λ∈L⨂xλ)=λ∈L⨂vλ(xλ).
In particular, we consider, for some μ∈L, an endomorphism θ of Gμ; we denote by θ~ the endomorphism of ⨂λ∈LGλ equal to ⨂λ∈Lvλ where vμ=θ and vλ=1Gλ for λ=μ.
Then we suppose given a set Ω, a mapping
c:ω↦(ρ(ω),σ(ω))
of Ω into L×L and, for all ω∈Ω, an endomorphism pω of Gρ(ω) and an endomorphism qω of Gσ(ω); there correspond to them two endomorphisms p~ω and q~ω of P=⨂λ∈LGλ. Let R be the sub-Z-module of P generated by the union of the images of the endomorphismsp~ω−q~ω when ω runs through Ω. The quotient Z-module P/R is called the tensor product of the family(Gλ)λ∈Lrelative to c, p, q and is denoted by ⨂(c,p,q)Gλ; composing the canonical homomorphism P→P/R with the mapping ϕ:G→⨂λ∈LGλ defined above, we obtain a Z-multilinear mapping ϕ(c,p,q):G→⨂(c,p,q)Gλ and write ϕ(c,p,q)((xλ))=⨂(c,p,q)xλ or simply ⨂(c)xλ. The ordered pair consisting of ⨂(c,p,q)xλ and ϕ(c,p,q) resolves the following universal mapping problem: let pˉω (resp. qˉω) denote the mapping
(resp. (xσ(ω),(xλ)λ=σ(ω))↦(qω(xσ(ω)),(xλ)λ=σ(ω))) of G into itself. Then Σ is taken to be the species of Z-module structure and the α-mappings the Z-multilinear mappings u of G into a Z-module which further satisfy the conditions
(26)
u∘pˉω=u∘qˉω
for all ω∈Ω. The proof is obvious from the above definitions.
This construction recovers in particular that of E⊗AF described in no. 1: in this case we take L={1,2}, G1=E, G2=F, Ω=A; further, for all ω∈A, we must have ρ(ω)=1, σ(ω)=2, ω is the endomorphism x↦xω of the Z-module E and qω the endomorphism y↦ωy of the Z-module F.
Let (Gλ′)λ∈L be a second family of Z-modules; keeping the mapping c the same, suppose given for all ω∈Ω, an endomorphism pω′ of Gρ(ω)′ and an endomorphism qω′ of Gσ(ω)′. For all λ∈L, then let vλ:Gλ→Gλ′ be a Z-linear mapping such that, for all ω∈Ω,
(27)
vρ(ω)∘pω=pω′∘vρ(ω)andvσ(ω)∘qω=qω′∘vσ(ω)
(in other words, for all λ∈L, vλ is a morphism for the laws of action on Gλ (resp. Gλ′) defined on the pξ and qη (resp. pξ′ and qη′) with ξ and η such that ρ(ξ)=λ and σ(η)=λ). Then the mapping
u:(xλ)↦(c,p′,q′)⨂vλ(xλ)
of G into ⨂(c,p′,q′)Gλ′ is Z-multilinear and satisfies conditions (26) and hence defines a Z-linear mapping of ⨂(c,p,q)Gλ into ⨂(c,p′,q′)Gλ′, which we shall denote simply by ⨂(c)vλ if no confusion can arise.
We shall now give an "associativity" property for the general tensor products thus defined. Let (Li)1≤i≤n be a finite partition of L; for every index i, let Ωi denote the subset of Ω consisting of the elements such that ρ(ω)∈Li and σ(ω)∈Li; clearly the Ωi are pairwise disjoint; we set Ω′=Ω−(⋃iΩi). For each index i, c(i) will denote the mapping ω↦(ρ(ω),σ(ω)) of Ωi into Li×Li; for ω∈Ωi, we write pω(i) and qω(i) instead of pω and qω. Then for each i there is a "partial" tensor product
Fi=(c(i),p(i),q(i))⨂Gλ.
We shall further make the following "permutability" hypothesis:
(P) If ω∈Ω′, pω (resp. qω) permutes with each of the endomorphisms pξ and qη of Gρ(ω) (resp. Gσ(ω)) such that ξ∈/Ω′, η∈/Ω′ and ρ(ω)=ρ(ξ)=σ(η) (resp. σ(ω)=ρ(ξ)=σ(η)).
For each ω∈Ω′, let i be the index such that ρ(ω)∈Li; then consider the family (vλ)λ∈Li where vρ(ω)=pω and vλ=1Gλ for λ=ρ(ω); hypothesis (P) implies that the family (vλ) satisfies conditions (27) (where p′ and p must be replaced by p(i), q′ and q by q(i), ω by an element ξ running through Ωi); thus an endomorphism ⨂(c(i))vλ=rω is derived of the Z-module Fi. Similarly, an endomorphism sω is defined of the Z-module Fj starting with qω, j being the index such that σ(ω)∈Lj; finally let d(ω)=(i,j). Then we can define the tensor product⨂(d,r,s)Fi and the corresponding canonical mapping
ϕ(d,r,s):i=1∏nFi→(d,r,s)⨂Fi.
On the other hand, for each i, the canonical mapping
ψi=ϕ(c(i),p(i),q(i)):λ∈Li∏Gλ→Fi;
using the associativity of the product of sets, a Z-multilinear mapping ψ=ϕ(d,r,s)∘(ψi) of G into ⨂(d,r,s)Fi is derived. We show that the ordered pair(⨂(d,r,s)Fi,ψ)is a solution of the same universal problem as(⨂(c,p,q)Gλ,ϕ(c,p,q)), whence will follow the existence of a uniqueZ-module isomorphism
By induction on n, the proof is reduced to the case n=2; for simplicity we write F1⊗(d)F2 and y1⊗(d)y2 instead of ⨂(d,r,s)Fi and ⨂(d,r,s)yi. Consider the mapping from G to F1⊗(d)F2
h:(xλ)→(c(1))⨂xλ⊗(d)(c(2))⨂xλ.
It is obviously Z-multilinear; we show that it satisfies conditions (26) for all ω∈Ω. This is obvious if ω∈Ω1 or ω∈Ω2; otherwise, supposing, in order to fix the ideas, that ρ(ω)∈L1 and σ(ω)∈L2, the values of h∘pˉω and h∘qˉω at (xλ) are respectively
which are also equal by definition of F1⊗(d)F2.
This being so, let u be a Z-multilinear mapping of G into a Z-module H, satisfying conditions (26); we shall define a Z-linear mapping v:F1⊗(d)F2→H such that u=v∘h and that will prove our assertion (repeating the argument of no. 1, Corollary 1 to Proposition 1). For all z2=(xλ)λ∈L2 we consider the "partial" linear mapping of ∏λ∈L1Gλ into H
Clearly it is Z-multilinear and satisfies conditions (26) for ω∈Ω1; by definition there thus exists a Z-linear mapping y1↦w1(y1,x2) of F1 into H such that
w1(c(1))⨂xλ,z2=u((xλ)λ∈L1,z2)
We next consider the mapping
u2:(xλ)λ∈L2↦w1(.,(xλ)λ∈L2)
of ∏λ∈L2Gλ into HomZ(F1,H); it is obviously Z-multilinear and satisfies conditions (26) for ω∈Ω2, by virtue of the hypothesis on u and relations (28) and (29) and taking account of the fact that the elements of the form ⨂(c(1))xλ generate the Z-module F1. Hence there is a Z-linear mapping
w2:F2→HomZ(F1,H)
such that
w2(c(2))⨂xλ=u2((xλ)λ∈L2)
or also
w2(c(2))⨂xλ(c(1))⨂xλ=u((xλ)λ∈L).
We now consider, for y1∈F1,y2∈F2, the element of H
w(y1,y2)=(w2(y2))(y1).
Clearly w is a Z-bilinear mapping of F1×F2 into H. We show further that, for all ω∈Ω′, (assuming, to fix the ideas, that ρ(ω)∈L1 and σ(ω)∈L2)
w(rω(y1),y2)=w(y1,sω(y2)).
It suffices to verify this relation when y1 (resp. y2) is of the form ⨂(c(1))xλ (resp. ⨂(c(2))xλ), since these elements generate the Z-module F1 (resp. F2). But by definition, rω(⨂(c(1))xλ)=⨂(c(1))xλ′, where xρ(ω)′=pω(xρ(ω)) and xλ′=xλ for λ ≠ ρ(ω) in L₁; similarly sω(⨂(c(2))xλ)=⨂(c(2))xλ′′, where xσ(ω)′′=qω(xσ(ω)′) and xλ′′=xλ for λ ≠ σ(ω) in L₂; using (30) and (31), relation (32) then follows from (26). Hence there exists a Z-linear mapping v of F₁ ⊗ F₂ into H such that v(y1⊗y2)=w(y1,y2) and it then follows from (30) and (31) that v∘h=u.
The most important special case of the general tensor product defined above is the following: we start with a family (Ai)1≤i≤n−1 of rings and a family (Ei)1≤i≤n, where E1 is a right A1-module, En is a left An−1 module and for 2≤i≤n−1, Ei is an (Ai−1,Ai)-bimodule. Then the above definition is applied as follows: L is the set [1,n], Gi=Ei, Ω is the set the sum of the Ai (1≤i≤n−1). For ω∈Ai (1≤i≤n−1), take ρ(ω)=i, σ(ω)=i+1, pω is the endomorphism x↦xω of the Z-module Ei and qω the endomorphism y↦ωy of the Z-module Ei+1; the corresponding tensor product is denoted by
E1⊗A1E2⊗A2E3⊗⋯⊗An−2En−1⊗An−1En
(a notation where the Ai may occasionally be suppressed) and the elements ⨂(c,p,q)xi of this tensor product, for a family (xi) such that xi∈Ei for 1≤i≤n, may be written x1⊗x2⊗⋯⊗xn if no confusion can arise; an analogous notation is used for a Z-linear mapping ⨂(c)vi. Hypothesis (P) holds for every partition of [1,n], since the Ei are bimodules for 2≤i≤n−1. When n=3, we have thus defined the Z-module E⊗AF⊗BG alluded to in no. 8, and recovered Proposition 8 (no. 8).
When each of the Ei is a multimodule (with, for 2≤i≤n−1, Ai−1 one of the rings operating on the left and Ai one of the rings operating on the right and analogous conditions for i=1 and i=n), as in no. 4, a multimodule structure is defined on E1⊗A1E2⊗⋯⊗An−1En with respect to all the rings except the Ai which operate on the Ei (1≤i≤n).
In particular, let C be a commutative ring, (Ei)1≤i≤n a family of C-modules. By giving E1 and En two C-module structures identical with the given structure and Ei for 2≤i≤n−1 three C-module structures identical with the given structure, we define on the tensor product
E1⊗CE2⊗CE3⊗⋯⊗CEn−1⊗CEn
n C-modules structures which are compatible with one another and which are in fact identical, since for all γ∈C and (xi)∈∏i=1nEi, by definition
When we speak of the tensor product (34) as a C-module, we shall always mean with this structure, unless otherwise mentioned, and the tensor product (34) is also denoted by ⨂1≤i≤nEi if no confusion can arise. For every C-module G, the Z-multilinear mappings of ∏i=1nEi into G which, for every index i, satisfy the relation
f(x1,…,xi−1,γxi,xi+1,…,xn)=γf(x1,…,xn)
for γ∈C and (xi)∈∏iEi are then called C-multilinear and form a C-module denoted by Ln(E1,…,En;G); the universal property of the tensor product (34) then allows us to define a canonical C-module isomorphism
Ln(E1,…,En;G)→HomC(E1⊗CE2⊗⋯⊗CEn,G)
which associates with every C-multilinear mapping f the C-linear mapping g such that
f(x1,…,xn)=g(x1⊗x2⊗⋯⊗xn).
A C-multilinear mapping of E1×⋯×En into C is also called an n-linear form.
Let (Fi)1≤i≤n be another family of C-modules; for every system of n C-linear mappings ui:Ei→Fi,u1⊗u2⊗⋯⊗un (also denoted by ⨂1≤i≤nui) is a C-linear mapping of
E1⊗CE2⊗⋯⊗CEn into F1⊗CF2⊗⋯⊗CFn.
Further, (u1,…,un)↦u1⊗u2⊗⋯⊗un is a C-multilinear mapping of ∏iHomC(Ei,Fi) into
HomC(E1⊗CE⊗⋯⊗CEn,F1⊗CF2⊗⋯⊗CFn).
Hence there corresponds canonically to the latter mapping a C-linear mapping called canonical
The general associativity property seen earlier can be specialized here as follows. Given a partition (Jk)1≤k≤m of the interval {1,n} of N, let, for each k, Fk be the tensor product Ei1⊗CEi2⊗⋯⊗CEir, where (i1,…,ir) is the strictly increasing sequence of elements of Jk. Then there is a canonical C-module isomorphism (called "associativity isomorphism")
F1⊗CF2⊗⋯⊗CFm→E1⊗CE2⊗⋯⊗CEn
which, in the above notation, maps the tensor product
y1⊗y2⊗⋯⊗ym,where yk=xi1⊗xi2⊗⋯⊗xir,
to the tensor product x1⊗x2⊗⋯⊗xn (where xi∈Ei for all i).
In particular, if π is a permutation of {1,n}, writing Jk={π(k)} for 1≤k≤n, we obtain a canonical ("commutativity") isomorphism
Eπ(1)⊗CEπ(2)⊗⋯⊗CEπ(n)→E1⊗CE2⊗⋯⊗CEn
which maps xπ(1)⊗xπ(2)⊗⋯⊗xπ(n) to x1⊗x2⊗⋯⊗xn. We shall often identify the various tensor products which correspond to one another under these canonical isomorphisms.
For 1≤i≤n, suppose that Ei admits a basis (bλi(i))λi∈Li; by induction on n, it follows from no. 7, Corollary 2 to Proposition 7 that the family (bλ1(1)⊗bλ2(2)⊗⋯⊗bλn(n)), where (λ1,…,λn) runs through ∏1≤i≤nLi, is a basis of ⨁1≤i≤nEi, sometimes called the tensor product of the bases (bλ1(1)) in question.
Remark
(1) The above remarks concerning the case of modules over a commutative ring generalize as in no. 5, Remark 2 when there is a tensor product E1⊗A1E2⊗⋯⊗An−1En where the rings Ai are not necessarily commutative and where, for each i, there is a homomorphism ρi:C→A of the same commutative ring C such that: (1) ρi(C) is contained in the centre of Ai; (2) for 2≤i≤n−1, the C-module structures on Ei obtained using the homomorphisms ρi−1 and ρicoincide. Then a C-*module* structure is obtained on E1⊗A1E2⊗⋯⊗An−1En and canonical mappings analogous to (13) (no. 5), which will be left to the reader to describe.
(2) Let A,B be two rings, E a right A-module, E′ a left A-module, F a right B-module and F′ a left B-module. The Z-bilinear mappings of (E⊗AE′)×(F⊗BF′) into a Z-moduleG are then in one-to-one correspondence with the Z-multilinear mappings f of E×E′×F×F′ into G satisfying the conditions
for λ∈A,μ∈B,x∈E,x′∈E′,y∈F,y′∈F′. The general constructions given in this no. reduce the proof of this to defining a canonical Z-module isomorphism between (E⊗AE′)⊗Z(F⊗BF′) and E⊗AE′⊗ZF⊗BF′, which follows from the associativity property of tensor products of the form (33).