§ 4. Relations between tensor products and homomorphism modules
1. The isomorphisms HomB(E⊗AF,G)→HomA(F,HomB(E,G)) and Homc(E⊗AF,G)→HomA(E,Homc(F,G))
Let E be a right A-module, F a left A-module, G a Z-module and H the Z-module of mappings f:E×F→G which are Z-bilinear and satisfy
(1)
f(xλ,y)=f(x,λy)for x∈E,y∈F,λ∈A.
It has been seen (§ 3, no. 1, Proposition 1) that there exists a canonical Z-module homomorphism
(2)
H→HomZ(E⊗AF,G).
On the other hand, a left A-module structure has been defined on HomZ(E,G) and a right A-module structure on HomZ(F,G) (§ 3, no. 3); we may therefore consider the Z-modules HomA(E,HomZ(F,G)) and HomA(F,HomZ(E,G)). A mapping f of E×F into G is canonically identified with a mapping of E into the set GF of mappings of F into G (Set Theory, II, § 5, no. 2); by expressing the fact that the latter mapping belongs to HomA(E,HomZ(F,G)), we obtain precisely the fact that f is biadditive and conditions (1); whence there is a canonical isomorphism
(3)
H→HomA(E,HomZ(F,G))
and similarly there is defined a canonical isomorphism
(4)
H→HomA(F,HomZ(E,G)).
Suppose now that E and G also have left (resp. right) B-module structures and that the A-module and B-module structures on E are compatible. Then E⊗AF has canonically a left (resp. right) B-module structure (§ 3, no. 4) and on the other hand HomB(E,G) has canonically a left A-module structure (§ 1, no. 14). We may therefore consider the Z-modules HomB(E⊗AF,G) and HomA(F,HomB(E,G)), which are submodules of HomZ(E⊗AF,G) and HomA(F,HomZ(E,G)) respectively (§ 2, no. 1, Theorem 2). We examine under what condition a mapping f∈H has as image under the isomorphisms (2) and (4) an element of HomB(E⊗AF,G) and an element of HomA(E,HomB(F,G)) respectively; in each of the two cases we find the same condition
f(βx,y)=βf(x,y)
(resp. f(xβ,y)=f(x,y)β)
for x∈E,y∈F,β∈B.
Similarly, suppose that F and G are left (resp. right) C-modules and that the A-module and C-module structures on F are compatible. Then, for a mapping f∈H to have as image under (2) or (3) an element of HomC(E⊗AF,G) or HomA(E,HomC(F,G)) respectively, it is necessary and sufficient that it satisfy the same condition
f(x,γy)=γf(x,y)
(resp. f(x,yγ)=f(x,y)γ)
for x∈E,y∈F,γ∈C.
We have therefore established the following result (in the notation introduced above):
Proposition 1
(a) *Let E be a (B, A)-bimodule, F a left A-module and G a left B-module. For every mapping g∈HomB(E⊗AF,G), let g′ be the mapping of F into HomB(E,G) defined by (g′(y))(x)=g(x⊗y) for x∈E,y∈F. The mapping g↦g′ is an isomorphism
(5)
β:HomB(E⊗AF,G)→HomA(F,HomB(E,G)).
(b) Let E be a right A-module, F an (A, C)-bimodule and G a right C-module. For every mapping h∈HomC(E⊗AF,G), let h′ be the mapping of E into HomC(F,G) defined by (h′(x))(y)=h(x⊗y) for x∈E,y∈F. The mapping h↦h′ is an isomorphism
(6)
γ:HomC(E⊗AF,G)→HomA(E,HomC(F,G)).
In particular B and C may be taken to be a subring Γ of the centre of the ring A; then for every Γ-module G, the three Γ-modules
are canonically isomorphic to the Γ-module of Γ-*bilinear* mappings of E×F into G which satisfy (1). More particularly:
Corollary
If C is a commutative ring and E, F, G three C-modules, then the C-modules*
HomC(E⊗CF,G),HomC(E,HomC(F,G)),
HomC(F,HomC(E,G)),L2(E,F;G)
*are canonically isomorphic.
2. The homomorphism** E∗⊗AF→HomA(E,F)
Let A, B be two rings, E a left A-module, F a left B-module and G an (A, B)-*bimodule*. The Z-module HomA(E,G) has canonically a right B-*module* structure (§ 1, no. 14) such that (uβ)(x)=u(x)β for β∈B, u∈HomA(E,G), x∈E. On the other hand, G⊗BF has canonically a left A-module structure (§ 3, no. 4). We shall define a canonical Z-homomorphism
ν:HomA(E,G)⊗BF→HomA(E,G⊗BF).
To this end, we consider, for all y∈F and all u∈HomA(E,G), the mapping ν′(u,y):x↦u(x)⊗y of E onto G⊗BF. It is immediately verified that ν′(u,y) is A-linear and that ν′ is a Z-bilinear mapping of HomA(E,G)×F into HomA(E,G⊗BF); moreover, for all β∈B, ν′(uβ,y) and ν′(u,βy) are equal, for (u(x)β)⊗y=u(x)⊗(βy). We conclude (§ 3, no. 1, Proposition 1) the existence of the desired homomorphism ν such that ν(u⊗y) is the A-linear mapping x↦u(x)⊗y.
It is immediately verified that, if E is an (A,(Ci′);(Dj′))-multimodule, F a (B,(Ch′′);(Dk′′))-multimodule and G an (A,(C1′′′);B,(Dm′′′))-multimodule, the mapping (7) is a ((Dj′),(Ch′′),(C1′′′);(Ci′),(Dk′′),(Dm′′′))-multimodule homomorphism.
Proposition 2
(i) When F is a projective (resp. finitely generated projective) B-module, the canonical homomorphism (7) is injective (resp. bijective).
(ii) When E is a finitely generated projective A-module, the canonical homomorphism (7) is bijective.
(i) Fixing E and G, for every left B-module F, we write
T(F)=HomA(E,G)⊗BF,T′(F)=HomA(E,G⊗BF);
for every left B-module homomorphism u:F→F′, we write T(u)=1⊗u (1 here denoting the identity mapping of HomA(E,G)),
T′(u)=Hom(1E,1G⊗u);
on the other hand we write νF instead of ν. Then we have the following lemmas:
Lemma 1
For every homomorphism u:F→F′, the diagram
T(F)↓T(u)T(F′)νFνF′T′(F)↓T′(u)T′(F′)
is commutative.
The verification is immediate.
Lemma 2
Let M,N be two supplementary submodules in F and i:M→F,j:N→F the canonical injections. The diagram
is commutative and the vertical arrows are bijective.
The commutativity follows from Lemma 1, the other assertions from § 1, no. 6, Corollary 2 to Proposition 6 and § 3, no. 7, Proposition 7.
Lemma 3
Under the hypotheses of Lemma 2, for νF to be injective (resp. surjective), it is necessary and sufficient that νM and νN be so.
This follows from Lemma 2 and § 1, no. 6, Corollary 1 to Proposition 6.
Then Lemma 3, together with § 2, no. 2, Proposition 4, shows that it suffices to consider the case where F is a free module. But, if (bμ) is a basis of F, every element of HomA(E,G)⊗BF may then be written uniquely as ∑μuμ⊗bμ, where uμ∈HomA(E,G) (§ 3, no. 7, Corollary 1 to Proposition 7); the image of this element under ν is the A-linear mapping x↦∑μuμ(x)⊗bμ; it cannot be zero for all x∈E unless uμ(x)=0 for all x∈E and all μ, which is equivalent to saying that uμ=0 for all μ; hence ν is injective. When also F admits a finite basis, Lemma 3 shows (by induction on the number of elements in the basis of F) that to prove that ν is surjective, it suffices to do so when F=Bs; but in this case the two sides of (7) are canonically identified with HomA(E,G) (§ 3, no. 4, Proposition 4) and ν becomes the identity.
(ii) To show the proposition when E is projective and finitely generated, this time we fix F and G and write, for every left A-module E,
T(E)=HomA(E,G)⊗BF,T′(E)=HomA(E,G⊗BF)
and, for every left A-module homomorphism v:E→E′,
T(v)=Hom(v,1G)⊗1F,T′(v)=Hom(v,1G⊗1F);
on the other hand, we write νE instead of ν. Then we have the two lemmas:
Lemma 4
For every homomorphism v:E→E′, the diagram
T(E′)↓T(v)T(E)νE′νET′(E′)↓T′(v)T′(E)
is commutative.
Lemma 5
Let M and N be two supplementary submodules in E and p:E→M,q:E→N the canonical projections. The diagram
is commutative and the vertical arrows are bijective.
They are proved as Lemmas 1 and 2, taking account of § 1, no. 6, Corollary 2 to Proposition 6, § 2, no. 1, Proposition 1 and § 3, no. 7, Proposition 7.
The remainder of the proof then proceeds as in (i) and is reduced to the case where E=As; the two sides of (7) are then canonically identified with G⊗BF and ν becomes the identity.
In particular take B=A and G the (A,A)-bimodule sAd (§3, no. 4), so that the right A-module HomA(E,sAd) is just the dual E∗ of E and (sAd)⊗AF is canonically identified with F (§3, no. 4, Proposition 4). Homomorphism (7) then becomes a canonical Z-homomorphism
(11)
θ:E∗⊗AF→HomA(E,F)
and θ(x∗⊗y) is the linear mapping of E into F
x↦⟨x,x∗⟩y.
Remark (1). The characterization of projectiveA-modules given in §2, no. 6, Proposition 12, can also be expressed as follows: for a left A-module E to be projective, it is necessary and sufficient that the canonical homomorphism
θE:E∗⊗AE→HomA(E,E)=EndA(E)
be such that 1E belongs to the image of θE.
Corollary
(i) When F is a projective (resp. finitely generated projective) module, the canonical homomorphism (11) is injective (resp. bijective). (ii) When E is a finitely generated projective module, the canonical homomorphism (11) is bijective.
Even when E and F are both finitely generated, θ is not necessarily surjective, as is shown by the example A=Z,E=F=Z/2Z; the right hand side of (11) is non-zero but E∗=0. On the other hand, examples can be given where E is free, but (11) is neither injective nor surjective (Exercise 3(b)).
When E admits a finite basis (ei), the inverse isomorphism θ−1 of θ can be found explicitly as follows. Let (ei∗) be the dual basis of (ei) (§ 2, no. 6); for all u∈Hom(E,F) and all x=∑iξiei with ξi∈A,
u(x)=i∑ξiu(ei)=i∑⟨x,ei∗⟩u(ei)
and therefore u=∑iθ(ei∗⊗u(ei)), in other words
θ−1(u)=i∑ei∗⊗u(ei).
In particular, if further F=E, it is seen that the image under θE−1 of the identity mapping 1E is the element ∑iei∗⊗ei, which is therefore independent of the basis (ei) considered in E.
Note on the other hand that when E is a finitely generated projective module the ring structure on EndA(E) can be transported by θE−1 to E∗⊗AE; it is immediately verified that, for x,y in E, x∗,y∗ in E∗, in the ring EndA(E),
θE(x∗⊗x)∘θE(y∗⊗y)=θE((y∗⟨y,x∗⟩)⊗x).
Remark (2). Let E be a right A-module; replacing E by E∗ in (11), we obtain a canonical Z-homomorphism
E∗∗⊗AF→HomA(E∗,F).
On the other hand, there is a canonical A-homomorphism cE:E→E∗∗, whence there is a Z-homomorphism cE⊗1F:E⊗AF→E∗∗⊗AF; composing the latter with the homomorphism (14), we hence obtain a canonical Z-homomorphism
θ′:E⊗AF→HomA(E∗,F)
such that θ′(x⊗y) is the linear mapping
x∗↦⟨x,x∗⟩y.
If E and F are projective modules, the mapping (15) is injective. For cE is then injective (§ 2, no. 7, Corollary 4 to Proposition 13) and as F is projective, the Z-homomorphism cE⊗1F:E⊗AF→E∗∗⊗AF is also injective (§ 3, no. 7, Corollary 6 to Proposition 7); finally, it has been seen (Proposition 2) that the homomorphism (14) is injective, whence the conclusion.
If E is projective and finitely generated, the mapping (15) is bijective for the two mappings of which it is composed are then bijective (§ 2, no. 7, Corollary 4 to Proposition 13 and Proposition 2 above).
3. Trace of an endomorphism
Let C be a commutative ring and E a C-module. The mapping (x∗,x)↦⟨x,x∗⟩ of E∗×E into C is then C*-bilinear*, since, for all γ∈C, ⟨γx,x∗⟩=γ⟨x,x∗⟩ and ⟨x,x∗γ⟩=⟨x,x∗⟩γ; we derive a canonical C*-linear* mapping
τ:E∗⊗CE→C
such that τ(x∗⊗x)=⟨x,x∗⟩ (§ 3, no. 5). Suppose now also that E is a finitely generated projectiveC-module; the canonical isomorphism (11) of no. 2 is then a C*-module* isomorphism and we can therefore define by transporting the structure a canonical linear formTr=τ∘θE−1 on the C-module EndC(E). For all u∈EndC(E) the scalar Tr(u) is called the trace of the endomorphism u; every u∈EndC(E) can be written (in general in an infinity of ways) in the form x↦∑i⟨x,xi∗⟩yi where xi∗∈E∗ and yi∈Ei by virtue of no. 2, Corollary to Proposition 2; then
Tr(u)=i∑⟨yi,xi∗⟩(cf. § 10, no. 11).
By definition,
Tr(u+v)=Tr(u)+Tr(v)
Tr(γu)=γTr(u)
for u,v in EndC(E) and γ∈C. Moreover:
Proposition 3
Let C be a commutative ring, E,F two finitely generated projective C*-modules* and u:E→F and v:F→E two linear mappings; then
Tr(v∘u)=Tr(u∘v).
The two mappings (u,v)↦Tr(u∘v),(u,v)↦Tr(v∘u) of
HomC(E,F)×HomC(F,E)
into C are C*-bilinear*; it therefore suffices to verify (20) when u is of the form x↦⟨x,a∗⟩b and v of the form y↦⟨y,b∗⟩a, with a∈E,a∗∈E∗,b∈F,b∗∈F∗. But then v∘u is the mapping x↦⟨x,a∗⟩⟨b,b∗⟩a and u∘v the mapping y↦⟨y,b∗⟩⟨a,a∗⟩b. Formula (17) shows that the values of the two sides of (20) are equal to ⟨a,a∗⟩⟨b,b∗⟩.
Corollary
If u1,…,up are endomorphisms of E, then
Tr(u1∘u2∘⋯∘up)=Tr(ui∘ui+1∘⋯∘up∘u1∘⋯∘ui−1)
for1≤i≤p ("invariance of the trace under cyclic permutation").
It suffices to apply (20) to the product
(u1∘u2∘⋯∘ui−1)∘(ui∘ui+1∘⋯∘up).
Note that on the other hand it is not necessarily true that
Tr(u∘v∘w)=Tr(u∘w∘v)
for three endomorphisms u,v,w of E.
4. The homomorphism Homc(E1,F1)⊗cHomc(E2,F2)→Homc(E1⊗cE2,F1⊗cF2)
Let C be a commutative ring and E1,E2,F1,F2 four C-modules; in § 3, no. 5, formula (13) we defined a canonical C-module homomorphism
When one of the ordered pairs* (E1,E2),(E1,F1),(E2,F2)consists of finitely generated projective C-modules, the canonical homomorphism (21) *is bijective.
It is obviously sufficient to perform the proof for the ordered pairs (E1,F1) and (E1,E2).
We consider first the case of the ordered pair (E1,F1); we fix E2,F1,F2 and write for every C-module T(E)=Hom(E,F1)⊗cHom(E2,F2) and T′(E)=Hom(E⊗E2,F1⊗F2) and, for every C-homomorphism v:E→E′,
T(v)=Hom(v,1F1)⊗1Hom(E2,F2)
and
T′(v)=Hom(v⊗1E2,1F1⊗F2).
Then Lemmas 4 and 5 (no. 2) (where v is replaced by λ) are valid and are proved by completely analogous methods.
We next fix E2 and F2 and this time write, for every C-module F,
T(F)=Hom(C,F)⊗cHom(E2,F2)
and T′(F)=Hom(C⊗E2,F⊗F2) and, for every C-homomorphism u:F→F′,
T(u)=Hom(1C,u)⊗1Hom(E2,F2)
and
T′(u)=Hom(1C⊗1E2,u⊗1F2).
This time it is immediately verified that Lemmas 1 and 2 (no. 2) (where λ always replaces v) are valid.
This being so, we show the proposition first when E1=C and F1 is projective and finitely generated. The argument of no. 2 (which rests on Lemmas 1 and 2), together with the above remarks, reduces this to proving the proposition when also F1=C; then Hom(E1,F1),E1⊗E2 and F1⊗F2 are identified with C, E2 and F2 respectively (§ 3, no. 4, Proposition 4); the two sides of (21)
THE HOMOMORPHISM Homc(E1,F1)⊗cHom(E2,F2)
are then both canonically identified with Hom(E2,F2) and, after these identifications, it is verified that λ becomes the identity.
We now suppose that F1 is projective and finitely generated; the argument of no. 2 (depending this time on Lemmas 4 and 5) reduces the proof for E1 any finitely generated projective module to the case where E1=C, that is to the first case dealt with.
For the ordered pair (E1,E2), the procedure is similar, this time applying Lemmas 4 and 5 twice; we leave the details to the reader.
Note that when E1=C(I), E2=C(J) are free (finitely generated or not), then Hom(E1,F1)=F1I, Hom(E2,F2=F2J and
Hom(E1⊗E2,F1⊗F2)=(F1⊗F2)I×J
to within canonical isomorphisms and (21) is then identical with a special case of the canonical homomorphism (22) of § 3, no. 7.
When E2=C, the canonical homomorphism (21) gives, after identifying Hom(E2,F2) with F2 and E1⊗E2 with E1, a canonical homomorphism
Hom(E,F)⊗G→Hom(E,F⊗G)
for any three C-modules E,F,G which is just the homomorphism (7) of no. 2 for A=B=C.
Note that when F=C the canonical homomorphism (22) again gives (11) (no. 2) for the case of a commutative ring.
Suppose now that F1=F2=C; as F1⊗F2 is identified with C, there is this time a canonical homomorphism
μ:E∗⊗F∗→(E⊗F)∗
for two C-modules E,F; for x∗∈E∗,y∗∈F∗, the image of x∗⊗y∗ under the canonical homomorphism (23) is the linear form u on E⊗F such that
u(x⊗y)=⟨x,x∗⟩⟨y,y∗⟩.
Moreover, if E1,E2,F1,F2 are four C-modules, f:E1→E2,g:F1→F2 two linear mappings, it follows immediately from (24) that the diagram
If one of the modules E, F is projective and finitely generated, the canonical homomorphism (23) is bijective.
Corollary 2
Let E₁, E₂ be two finitely generated projective C-modules, u₁ an endomorphism of E₁ and u₂ an endomorphism of E₂; then
Tr(u1⊗u2)=Tr(u1)Tr(u2).
By linearity, it suffices to consider the case where u₁ is of the form x1↦⟨x1,x1∗⟩y1 and u₂ of the form x2↦⟨x2,x2∗⟩y2; then the image of x1⊗x2 under u1⊗u2 is by definition
x1∗⊗x2∗ being canonically identified under μ with an element of (E1⊗E2)∗. As ⟨y1⊗y2,x1∗⊗x2∗⟩=⟨y1,x1∗⟩⟨y2,x2∗⟩, formula (26) follows in this case from (17).
Remark
If E, F, G are any three C-modules, it is immediately verified that the diagram
E∗⊗F∗⊗G∗1⊗μ↓E∗⊗(F⊗G)∗μ⊗1μ(E⊗F)∗⊗G∗↓μ(E⊗F⊗G)∗
is commutative, by virtue of formula (24).
We note also that, without any hypothesis on the C-modules E, F, there are canonical *isomorphisms
(28)
(E⊗F)∗→Hom(E,F∗)
(29)
(E⊗F)∗→Hom(F,E∗)
which are just the isomorphism (6) and (5) of no. 1 for G = C, A = B = C.
Thus a canonical one-to-one correspondence has been defined between the bilinear forms on E × F, the homomorphisms of E into F and the homomorphisms of F into E: if u (resp. v) is a homomorphism of E into F* (resp. of F into E*), the corresponding bilinear form is given by