Bourbaki

§ 5. Tensor algebra, tensors

1. Definition of the tensor algebra of a module

Let A be a commutative ring and M an A-module. For every integer n⩾0n \geqslant 0, the A-module the tensor product of nn modules equal to M (also called the n-th tensor power of M) is denoted by ⨂nM\bigotimes^n M, or M⊗nM^{\otimes n}, or Tn(M)T^n(M), or TAn(M)T^n_A(M), or Tensn(M)\mathrm{Tens}^n(M); then T1(M)=MT^1(M) = M; also we write T0(M)=AT^0(M) = A. The A-module the direct sum ⨁n⩾0Tn(M)\bigoplus_{n \geqslant 0} T^n(M) is denoted by T(M)T(M) or Tens(M)\mathrm{Tens}(M). We shall define a graded A-algebra structure of type N\mathbf{N} on T(M)T(M), by defining for every ordered pair of integers p⩾0,q⩾0p \geqslant 0, q \geqslant 0, an A-linear mapping

mpq:Tp(M)⊗ATq(M)→Tp+q(M) m_{pq} : T^p(M) \otimes_A T^q(M) \to T^{p+q}(M)
(§3§ 3, no. 1, Remark). For p>0p > 0 and q>0q > 0, mpqm_{pq} is the associativity isomorphism (II, §3§ 3, no. 9) and, when p=0p = 0 (resp. q=0q = 0), m0,qm_{0,q} is the canonical isomorphism of A⊗ATq(M)A \otimes_A T^q(M) onto Tq(M)T^q(M) (resp. mp,0m_{p,0} is the canonical isomorphism of Tp(M)⊗AAT^p(M) \otimes_A A onto Tp(M)T^p(M) (II, §3§ 3, no. 4, Proposition 4)). Then, for xi∈Mx_i \in M, α∈A\alpha \in A,
(x1⊗⋯⊗xp)⋅(xp+1⊗⋯⊗xp+q)=x1⊗⋯⊗xp⊗xp+1⊗⋯⊗xp+qα⋅(x1⊗⋯⊗xp)=α(x1⊗⋯⊗xp). \begin{align} (x_1 \otimes \cdots \otimes x_p) \cdot (x_{p+1} \otimes \cdots \otimes x_{p+q}) &= x_1 \otimes \cdots \otimes x_p \otimes x_{p+1} \otimes \cdots \otimes x_{p+q} \\ \alpha \cdot (x_1 \otimes \cdots \otimes x_p) &= \alpha(x_1 \otimes \cdots \otimes x_p). \end{align}
It is immediate that the multiplication thus defined on T(M)T(M) is associative and admits as unit element the unit element 1 of A=T0(M)A = T^0(M).

Definition 1

For every module MM over a commutative ring AA, the tensor algebra of MM, denoted by T(M)T(M), or Tens(M)\mathrm{Tens}(M), or TA(M)T_A(M), is the algebra ⨁n≥0Tn(M)\bigoplus_{n \geq 0} T^n(M) with the multiplication defined in (1). The canonical injection ϕ:T1(M)→T(M)\phi : T^1(M) \to T(M) (II, § 1, no. 12) (also denoted by ϕM\phi_M) is called the canonical injection of MM into T(M)T(M).

Proposition 1

Let EE be a (unital) AA-algebra and f:M→Ef : M \to E an AA-linear mapping. There exists one and only one AA-algebra homomorphism g:T(M)→Eg : T(M) \to E such that f=g∘ϕf = g \circ \phi.

In other words, (T(M),ϕ)(T(M), \phi) is a solution of the universal mapping problem (Set Theory, IV, § 3, no. 1), where Σ\Sigma is the species of AA-algebra structure, the α\alpha-mappings being the AA-linear mappings from the module MM to an AA-algebra. Observe that here there is no question of a graduation on T(M)T(M).

For every finite family (xi)1≤i≤n(x_i)_{1 \leq i \leq n} of nn elements of MM, by definition of the product in T(M)T(M), x1⊗x2⊗⋯⊗xn=ϕ(x1)ϕ(x2)…ϕ(xn)x_1 \otimes x_2 \otimes \cdots \otimes x_n = \phi(x_1)\phi(x_2)\ldots\phi(x_n); then necessarily g(x1⊗x2⊗⋯⊗xn)=f(x1)f(x2)…f(xn)g(x_1 \otimes x_2 \otimes \cdots \otimes x_n) = f(x_1)f(x_2)\ldots f(x_n) for n≥1n \geq 1 and g(α)=αeg(\alpha) = \alpha e (if ee is the unit element of EE) for α∈A\alpha \in A, which proves the uniqueness of gg. Conversely, note that, for all n>0n > 0, the mapping

(x1,…,xn)↦f(x1)f(x2)…f(xn) (x_1, \ldots, x_n) \mapsto f(x_1)f(x_2)\ldots f(x_n)
of MnM^n into EE is AA-multilinear; hence there corresponds to it an AA-linear mapping gn:Tn(M)→Eg_n : T^n(M) \to E such that
g(x1⊗x2⊗⋯⊗xn)=f(x1)f(x2)…f(xn). g(x_1 \otimes x_2 \otimes \cdots \otimes x_n) = f(x_1)f(x_2)\ldots f(x_n).
Also we define the mapping g0:T0(M)→Eg_0 : T^0(M) \to E as equal to ηE\eta_E (§1§ 1, no. 3), in other words g0(α)=αeg_0(\alpha) = \alpha e for α∈A\alpha \in A. Let gg be the unique AA-linear mapping of T(M)T(M) into EE whose restriction to Tn(M)T^n(M) is gng_n (n≥0n \geq 0); it is immediate that g∘ϕ=g1=fg \circ \phi = g_1 = f and it remains to verify that gg is an AA-algebra homomorphism. By construction g(1)=eg(1) = e and it suffices by linearity to show that g(uv)=g(u)g(v)g(uv) = g(u)g(v) for u∈Tp(M)u \in T^p(M) and v∈Tq(M)v \in T^q(M) (p>0,q>0p > 0, q > 0); now it follows from formulae (1) and (2) that this relation is true when u=x1⊗x2⊗⋯⊗xpu = x_1 \otimes x_2 \otimes \cdots \otimes x_p and v∈xp+1⊗⋯⊗xp+qv \in x_{p+1} \otimes \cdots \otimes x_{p+q} (where the xix_i belong to EE). It is therefore true for u∈Tp(M)u \in T^p(M) and v∈Tq(M)v \in T^q(M) by linearity.

Remark

Suppose that EE is a graded AA-algebra of type Z\mathbf{Z}, with graduation (En)(E_n), and suppose also that

f(M)⊂E1. f(M) \subset E_1.
Then it follows from (2) that g(Tp(M))⊂Epg(T^p(M)) \subset E_p for all p≥0p \geq 0 and hence gg is a graded algebra homomorphism.

2. Functorial properties of the tensor algebra

Proposition 2

Let AA be a commutative ring, MM and NN two AA-modules and

u:M→N u : M \to N

an A-linear mapping. There exists one and only one A-algebra homomorphism

u′:T(M)→T(N) u': \mathbf{T}(M) \to \mathbf{T}(N)

such that the diagram

M→uN↓ϕM↓ϕNT(M)→u′T(N) \begin{array}{ccc} M & \xrightarrow{u} & N \\ \downarrow \phi_M & & \downarrow \phi_N \\ \mathbf{T}(M) & \xrightarrow{u'} & \mathbf{T}(N) \end{array}

is commutative. Further, u′u' is a graded algebra homomorphism.

The existence and uniqueness of u′u' follow from no. 1, Proposition 1, applied to the algebra T(N)\mathbf{T}(N) and the linear mapping ϕN∘u:M→T(N)\phi_N \circ u : M \to \mathbf{T}(N); as

u(M)⊂T1(N)=N, u(M) \subset \mathbf{T}^1(N) = N,

the fact that u′u' is a graded algebra homomorphism follows from the Remark of no. 1.

The homomorphism u′u' of Proposition 2 will henceforth be denoted by T(u)\mathbf{T}(u). If P is an A-module and v:N→Pv : N \to P an A-linear mapping, then

T(v∘u)=T(v)∘T(u) \mathbf{T}(v \circ u) = \mathbf{T}(v) \circ \mathbf{T}(u)

for T(v)∘T(u)\mathbf{T}(v) \circ \mathbf{T}(u) is an algebra homomorphism rendering commutative the diagram

M→v∘uP↓ϕM↓ϕPT(M)→T(v)∘T(u)T(P) \begin{array}{ccc} M & \xrightarrow{v \circ u} & P \\ \downarrow \phi_M & & \downarrow \phi_P \\ \mathbf{T}(M) & \xrightarrow{\mathbf{T}(v) \circ \mathbf{T}(u)} & \mathbf{T}(P) \end{array}

T(u)\mathbf{T}(u) is sometimes called the canonical extension of uu to T(M)\mathbf{T}(M) (which contains M=T1(M)M = \mathbf{T}^1(M)). Note that the restriction Tn(u):Tn(M)→Tn(N)\mathbf{T}^n(u) : \mathbf{T}^n(M) \to \mathbf{T}^n(N) is just the linear mapping u⊗n=u⊗u⊗⋯⊗uu^{\otimes n} = u \otimes u \otimes \cdots \otimes u (n times), for

Tn(u)(x1⊗⋯⊗xn)=u(x1)⊗⋯⊗u(xn) \mathbf{T}^n(u)(x_1 \otimes \cdots \otimes x_n) = u(x_1) \otimes \cdots \otimes u(x_n)

since T(u)\mathbf{T}(u) is an algebra homomorphism and T1(u)=u\mathbf{T}^1(u) = u; the restriction T0(u)\mathbf{T}^0(u) to A is the identity mapping. Tn(u)\mathbf{T}^n(u) is called the nn-th tensor power of uu.

Proposition 3

If u:M→Nu : M \to N is a surjective A-linear mapping, the homomorphism T(u):T(M)→T(N)\mathbf{T}(u) : \mathbf{T}(M) \to \mathbf{T}(N) is surjective and its kernel is the two-sided ideal of T(M)\mathbf{T}(M) generated by the kernel P⊂M⊂T(M)P \subset M \subset \mathbf{T}(M) of uu.

T0(u):T0(M)→T0(N)\mathbf{T}^0(u) : \mathbf{T}^0(M) \to \mathbf{T}^0(N) is bijective and for every integer n>0n > 0,

Tn(u):Tn(M)→Tn(N) \mathbf{T}^n(u) : \mathbf{T}^n(M) \to \mathbf{T}^n(N)

is surjective, as is seen by induction on nn using II, § 3, no. 6, Proposition 6; the latter proposition also shows, by induction on nn, that the kernel In\mathfrak{I}_n of Tn(u)T^n(u) is the submodule of Tn(M)T^n(M) generated by the products

x1⊗x2⊗⋯⊗xn x_1 \otimes x_2 \otimes \cdots \otimes x_n
where at least one of the xix_i belongs to PP. This shows that the kernel I=⨁n≥1In\mathfrak{I} = \bigoplus_{n \geq 1} \mathfrak{I}_n of T(u)T(u) is the two-sided ideal generated by PP in T(M)T(M).

If u:M→Nu : M \to N is an injective linear mapping, it is not always true that T(u)T(u) is an injective mapping (Exercise 1). However, this is true when uu is an injection such that u(M)u(M) is a direct factor of NN, for then there exists a linear mapping v:N→Mv : N \to M such that v∘uv \circ u is the identity mapping on MM and therefore

T(v∘u)=T(v)∘T(u) T(v \circ u) = T(v) \circ T(u)
is the identity mapping on T(M)T(M), hence T(u)T(u) is injective and its image (isomorphic to T(M)T(M)) is a direct factor of T(N)T(N) (II, § 1, no. 9, Proposition 15). More precisely:

Proposition 4

Let NN and PP be two submodules of an AA-module MM such that their sum N+PN + P is a direct factor in MM and their intersection N∩PN \cap P is a direct factor in NN and in PP. Then the homomorphisms T(N)→T(M)T(N) \to T(M), T(P)→T(M)T(P) \to T(M) and*

T(N∩P)→T(M), T(N \cap P) \to T(M),
canonical extensions of the canonical injections, are injective; if T(N)T(N), T(P)T(P) and T(N∩P)T(N \cap P) are identified with subalgebras of T(M)T(M) by means of these homomorphisms, then (5)
T(N∩P)=T(N)∩T(P). T(N \cap P) = T(N) \cap T(P).

By hypothesis, there exist submodules N′⊂NN' \subset N and P′⊂PP' \subset P such that N=N′⊕(N∩P)N = N' \oplus (N \cap P), P=P′⊕(N∩P)P = P' \oplus (N \cap P); then

N+P=N′⊕P′⊕(N∩P) N + P = N' \oplus P' \oplus (N \cap P)
and there exists by hypothesis a submodule M′M' of MM such that
M=M′⊕(N+P)=M′⊕N′⊕P′⊕(N∩P)=M′⊕P′⊕N=M′⊕N′⊕P. M = M' \oplus (N + P) = M' \oplus N' \oplus P' \oplus (N \cap P) \\ = M' \oplus P' \oplus N = M' \oplus N' \oplus P.

In particular, N+PN + P, NN, PP and N∩PN \cap P are direct factors in MM, which implies, as has been seen above, that the canonical homomorphisms

T(N+P)→T(M),T(N)→T(M), T(N + P) \to T(M), \quad T(N) \to T(M),
T(P)→T(M),T(N∩P)→T(M) T(P) \to T(M), \quad T(N \cap P) \to T(M)
are injective. The three algebras T(N)T(N), T(P)T(P) and T(N∩P)T(N \cap P) are thus identified with subalgebras of T(N+P)T(N + P) and the latter with a subalgebra of T(M)T(M);

writing Q=N∩PQ = N \cap P, it remains to show that, if T(Q)T(Q), T(N′⊕Q)T(N' \oplus Q) and T(P′⊕Q)T(P' \oplus Q) are identified with subalgebras of T(N′⊕P′⊕Q)T(N' \oplus P' \oplus Q), then

T(N′⊕Q)∩T(P′⊕Q)=T(Q). T(N' \oplus Q) \cap T(P' \oplus Q) = T(Q).

Now, consider the commutative diagram

N′⊕Q⟶N′⊕P′⊕Q↓↓Q⟶P′⊕Q \begin{array}{ccc} N' \oplus Q & \longrightarrow & N' \oplus P' \oplus Q \\ \downarrow & & \downarrow \\ Q & \longrightarrow & P' \oplus Q \end{array}

where the horizontal arrows are the canonical injections and the vertical arrows the canonical projections. We derive a commutative diagram

T(N′⊕Q)⟶uT(N′⊕P′⊕Q)↓r↓sT(Q)⟶vT(P′⊕Q) \begin{array}{ccc} T(N' \oplus Q) & \overset{u}{\longrightarrow} & T(N' \oplus P' \oplus Q) \\ \downarrow r & & \downarrow s \\ T(Q) & \overset{v}{\longrightarrow} & T(P' \oplus Q) \end{array}

where rr and ss are surjective homomorphisms (Proposition 3) and uu and vv injective homomorphisms. Hence, to prove (6), note that the right hand side is obviously contained in the left; it therefore suffices to verify that if

x∈T(N′⊕Q)∩T(P′⊕Q), x \in T(N' \oplus Q) \cap T(P' \oplus Q),

then x∈T(Q)x \in T(Q). Now the definition of the homomorphism ss shows that its restriction to T(P′⊕Q)T(P' \oplus Q) (identified with a subalgebra of T(N′⊕P′⊕Q)T(N' \oplus P' \oplus Q)) is the identity mapping; the hypothesis on xx therefore implies that s(u(x))=xs(u(x)) = x. But then also v(r(x))=xv(r(x)) = x, in other words xx belongs to the image of T(Q)T(Q) in T(P′⊕Q)T(P' \oplus Q), which was to be proved.

Remark

Note in particular that the hypotheses of Proposition 4 always hold for arbitrary submodules N,PN, P of MM when AA is a field (II, § 7, no. 3, Proposition 4). Moreover, if N⊂PN \subset P and N≠PN \neq P, then Tn(N)≠Tn(P)T^n(N) \neq T^n(P) for all n≥1n \geq 1, since if RR is a complement of PP in NN, then Tn(P)∩Tn(R)={0}T^n(P) \cap T^n(R) = \{0\} by (4) and Tn(R)≠{0}T^n(R) \neq \{0\}.

Corollary

Let KK be a commutative field and MM a vector space over KK. For every element z∈T(M)z \in T(M), there exists a smallest vector space NN of MM such that z∈T(N)z \in T(N) and NN is of finite rank over KK.

It is understood in this statement that for every vector subspace PP of MM, T(P)T(P) is canonically identified with a subalgebra of T(M)T(M). Let z∈T(M)z \in T(M); zz can be expressed as a linear combination of elements each of which is a finite product of elements of M=T1(M)M = T^1(M); all the elements of MM which occur in these products generate a vector subspace QQ of finite rank and z∈T(Q)z \in T(Q).

Let F\mathscr{F} be the (non-empty) set of vector subspaces PP of finite rank such that z∈T(P)z \in T(P). Every decreasing sequence of elements of F\mathscr{F} is stationary, since they are vector spaces of finite rank. Hence F\mathscr{F} has a minimal element NN (Set Theory, III, § 6, no. 5). It remains to verify that every P∈FP \in \mathscr{F} contains NN; now, z∈T(P)∩T(N)=T(P∩N)z \in T(P) \cap T(N) = T(P \cap N) (Proposition 4); in view of the definition of NN, this implies N∩P=NN \cap P = N, that is P⊃NP \supset N.

The subspace NN of MM is said to be associated with zz.

3. Extension of the ring of scalars

Let A,A′A, A' be two commutative rings and ρ:A→A′\rho : A \to A' a ring homomorphism. Let MM be an AA-module, M′M' an A′A'-module and u:M→M′u : M \to M' an AA-homomorphism; as the canonical injection ϕM′:M′→TA′(M′)\phi_{M'} : M' \to T_{A'}(M') is also an AA-homomorphism (by restriction of scalars), so is the composition M→uM′→ϕM′TA′(M′)M \xrightarrow{u} M' \xrightarrow{\phi_{M'}} T_{A'}(M'). An AA-algebra homomorphism TA(M)→ρ∗(TA′(M′))T_A(M) \to \rho^*(T_{A'}(M')) is derived (no. 2), also denoted by T(u):TA(M)→TA′(M′)T(u) : T_A(M) \to T_{A'}(M'), which is the unique AA-homomorphism rendering commutative the diagram

M→uM′↓ϕM↓ϕM′TA(M)→T(u)TA′(M′) \begin{array}{ccc} M & \xrightarrow{u} & M' \\ \downarrow \phi_M & & \downarrow \phi_{M'} \\ T_A(M) & \xrightarrow{T(u)} & T_{A'}(M') \end{array}

If σ:A′→A′′\sigma : A' \to A'' is a commutative ring homomorphism, M′′M'' an A′′A''-module and v:M′→M′′v : M' \to M'' an A′A'-homomorphism, the above uniqueness property shows that

T(v∘u)=T(v)∘T(u). T(v \circ u) = T(v) \circ T(u).

Proposition 5

Let A,BA, B be two commutative rings, ρ:A→B\rho : A \to B a ring homomorphism and MM an AA-module. The canonical extension

ψ:TB(B⊗AM)→B⊗ATA(M) \psi : T_B(B \otimes_A M) \to B \otimes_A T_A(M)

of the B-linear mapping lB⊗ϕM:B⊗AM→B⊗ATA(M)l_B \otimes \phi_M : B \otimes_A M \to B \otimes_A T_A(M) is an isomorphism of graded B-algebras.

Consider the two AA-algebra homomorphisms: the canonical injection j:B=T0(B⊗AM)→T(B⊗AM)j : B = T^0(B \otimes_A M) \to T(B \otimes_A M) and the homomorphism

h=T(i):T(M)→T(B⊗AM) h = T(i) : T(M) \to T(B \otimes_A M)

derived (cf. formula (8)) from the canonical AA-linear mapping

i:M→B⊗AM. i : M \to B \otimes_A M.

As T0(B⊗AM)T^0(B \otimes_A M) is contained in the centre of T(B⊗AM)T(B \otimes_A M), Proposition 5 of § 4, no. 2 can be applied and an A-algebra homomorphism

ψ′:B⊗AT(M)→T(B⊗AM) \psi': B \otimes_A T(M) \to T(B \otimes_A M)

is obtained such that, for β∈B,xi∈M\beta \in B, x_i \in M for 1≤i≤n1 \leq i \leq n,

ψ′(β⊗(x1⊗x2⊗⋯⊗xn))=β((1⊗x1)⊗(1⊗x2)⊗⋯⊗(1⊗xn)), \psi'(\beta \otimes (x_1 \otimes x_2 \otimes \cdots \otimes x_n)) = \beta((1 \otimes x_1) \otimes (1 \otimes x_2) \otimes \cdots \otimes (1 \otimes x_n)),

which shows immediately that ψ′\psi' is also a B-algebra homomorphism. It suffices to prove that ψ∘ψ′\psi \circ \psi' and ψ′∘ψ\psi' \circ \psi are the identity mappings on B⊗AT(M)B \otimes_A T(M) and T(B⊗AM)T(B \otimes_A M) respectively. Now, these two algebras are generated by B⊗AMB \otimes_A M and clearly ψ∘ψ′\psi \circ \psi' and ψ′∘ψ\psi' \circ \psi coincide with the identity mapping on B⊗AMB \otimes_A M, whence the conclusion.

4. Direct limit of tensor algebras

Let (Aα,ϕβα)(A_\alpha, \phi_{\beta \alpha}) be a directed direct system of commutative rings and (Mα,fβα)(M_\alpha, f_{\beta \alpha}) a direct system of AαA_\alpha-modules (II, § 6, no. 2); let A=lim⁡→AαA = \lim \rightarrow A_\alpha and M=lim⁡→MαM = \lim \rightarrow M_\alpha, which is an AA-module. For α≤β\alpha \leq \beta an AαA_\alpha-algebra homomorphism (no. 3, formula (8)) fβα′=T(fβα):TAα(Mα)→TAβ(Mβ)f'_{\beta \alpha} = T(f_{\beta \alpha}): T_{A\alpha}(M_\alpha) \to T_{A\beta}(M_\beta) is derived canonically from the AαA_\alpha-homomorphism fβα:Mα→Mβf_{\beta \alpha}: M_\alpha \to M_\beta and it follows from (9) (no. 3) that (TAα(Mα),fβα′)(T_{A\alpha}(M_\alpha), f'_{\beta \alpha}) is a direct system of AαA_\alpha-algebras. On the other hand let fα′:Mα→Mf'_\alpha: M_\alpha \to M be the canonical AαA_\alpha-homomorphism; an AαA_\alpha-algebra homomorphism fα′:TAα(Mα)→TA(M)f'_\alpha: T_{A\alpha}(M_\alpha) \to T_A(M) is derived (no. 3, formula (8)) and it also follows from (9) (no. 3) that the fα′f'_\alpha constitute a direct system of AαA_\alpha-homomorphisms.

Proposition 6

The A-homomorphism* f′=lim⁡→fα′:lim⁡→TAα(Mα)→TA(M)f' = \lim \rightarrow f'_\alpha : \lim \rightarrow T_{A\alpha}(M_\alpha) \to T_A(M) is a graded algebra isomorphism.

To simplify we write E=TA(M)E = T_A(M) and E′=lim⁡→TAα(Mα)E' = \lim \rightarrow T_{A\alpha}(M_\alpha) and let

gα:TAα(Mα)→E′ g_\alpha: T_{A\alpha}(M_\alpha) \to E'

be the canonical AαA_\alpha-homomorphism. Clearly the composite AαA_\alpha-linear mappings Mα→ϕMαTAα(Mα)→gαE′M_\alpha \xrightarrow{\phi_{M_\alpha}} T_{A\alpha}(M_\alpha) \xrightarrow{g_\alpha} E' form a direct system and there is therefore one and only one AA-linear mapping u=lim⁡→(gα∘ϕMα):M→E′u = \lim \rightarrow (g_\alpha \circ \phi_{M_\alpha}): M \to E' such that

u∘fα=gα∘ϕMα u \circ f_\alpha = g_\alpha \circ \phi_{M_\alpha}

for all α\alpha. This mapping itself factorizes uniquely (no. 1, Proposition 1) into M→ϕME→E′M \xrightarrow{\phi_M} E \to E', where hh is an AA-algebra homomorphism. It will suffice to prove that h∘f′=1E′h \circ f' = 1_{E'} and f′∘h=1Ef' \circ h = 1_E.

To this end note that, for every index α\alpha, (no. 3, formula (8))

h∘fα′∘ϕMα=h∘ϕM∘fα=u∘fα=gα∘ϕMα h \circ f'_\alpha \circ \phi_{M_\alpha} = h \circ \phi_M \circ f_\alpha = u \circ f_\alpha = g_\alpha \circ \phi_{M_\alpha}

whence, by the uniqueness assertion of no. 1, Proposition 1,

h∘fα′=gα h \circ f'_\alpha = g_\alpha

for all α\alpha; it follows that (h∘f′)∘gα=gα(h \circ f') \circ g_\alpha = g_\alpha for all α\alpha and hence h∘f′=1E′h \circ f' = 1_{E'} by definition of a direct limit.

On the other hand, by virtue of no. 3, formula (8),

f′∘u∘fα=f′∘gα∘ϕMα=fα′∘ϕMα=ϕM∘fα, f' \circ u \circ f_\alpha = f' \circ g_\alpha \circ \phi_{M_\alpha} = f'_\alpha \circ \phi_{M_\alpha} = \phi_M \circ f_\alpha,

whence again f′∘u=ϕMf' \circ u = \phi_M by definition of a direct limit; we conclude that f′∘h∘ϕM=ϕMf' \circ h \circ \phi_M = \phi_M and the uniqueness property of no. 1, Proposition 1 gives f′∘h=1Ef' \circ h = 1_E.

Proposition 6 can also be shown by observing that, for every integer n≥1n \geq 1, there is a canonical AA-module isomorphism lim⁡→TAαn(Mα)→TAn(M)\lim \rightarrow T^n_{A_\alpha}(M_\alpha) \to T^n_A(M), as follows by induction on nn from II, § 6, no. 3, Proposition 7. It is immediately verified that these isomorphisms are the restrictions of f′f'.

5. Tensor algebra of a direct sum. Tensor algebra of a free module. Tensor algebra of a graded module

Let AA be a commutative ring and M=⨁λ∈LMλM = \bigoplus_{\lambda \in L} M_\lambda the direct sum of a family of AA-modules. It follows from II, § 3, no. 7, Proposition 7, by induction on nn, that Tn(M)T^n(M) is the direct sum of the submodules the images of the canonical injections

Mλ1⊗Mλ2⊗⋯⊗Mλn→Tn(M)=M⊗n M_{\lambda_1} \otimes M_{\lambda_2} \otimes \cdots \otimes M_{\lambda_n} \to T^n(M) = M^{\otimes n}

relative to all the sequences (λi)∈Ln(\lambda_i) \in L^n. Identifying Mλ1⊗Mλ2⊗⋯⊗MλnM_{\lambda_1} \otimes M_{\lambda_2} \otimes \cdots \otimes M_{\lambda_n} with this image, it is seen that T(M)T(M) is the direct sum of all the modules

Mλ1⊗Mλ2⊗⋯⊗Mλn M_{\lambda_1} \otimes M_{\lambda_2} \otimes \cdots \otimes M_{\lambda_n}

where nn runs through N\mathbf{N} and, for each nn, (λi)(\lambda_i) runs through LnL^n.

We first deduce the following consequence:

Theorem 1

Let AA be a commutative ring, MM a free AA-module and (eλ)λ∈L(e_\lambda)_{\lambda \in L} a basis of MM. Then the elements es=eλ1⊗eλ2⊗⋯⊗eλne_s = e_{\lambda_1} \otimes e_{\lambda_2} \otimes \cdots \otimes e_{\lambda_n}, where s=(λ1,…,λn)s = (\lambda_1, \ldots, \lambda_n) runs through the set of all finite sequences of elements of LL and e∅e_\varnothing is used to denote the unit element of T(M)T(M), form a basis of the AA-module T(M)T(M).

The elements of this basis are obviously homogeneous and the multiplication table is given by

eset=est e_s e_t = e_{st}

where stst denotes the sequence of elements of LL obtained by juxtaposing the sequences ss and tt (I, § 7, no. 2).

It is seen that the basis (es)(e_s) of T(M)T(M), with the multiplicative law (10), is canonically isomorphic to the free monoid of the set LL (I, § 7, no. 2), the isomorphism being obtained by mapping each word ss of this monoid to the element ese_s. It follows (§ 2, no. 7) that the tensor algebra T(M)T(M) of a free module MM, with a basis of indexing set LL, is canonically isomorphic to the free associative algebra of LL over AA. In particular (§ 2, no. 7, Proposition 7), for every mapping f:L→Ef : L \to E from LL to an AA-algebra EE, there exists one and only one AA-algebra homomorphism fˉ:T(M)→E\bar{f} : T(M) \to E such that fˉ(eλ)=f(λ)\bar{f}(e_\lambda) = f(\lambda).

Remark

The above results can equally be obtained as a consequence of the universal properties of the free associative algebra of the tensor algebra, using II, § 3, no. 7, Corollary 2 to Proposition 7.

Corollary

If MM is a projective AA-module, T(M)T(M) is a projective AA-module.

MM is a direct factor of a free AA-module NN (II, § 2, no. 2, Proposition 4) and hence T(M)T(M) is a direct factor of T(N)T(N) (no. 2); as T(N)T(N) is free (Theorem 1), this shows that T(M)T(M) is projective (II, § 2, no. 2).

Proposition 7

Let Δ\Delta be a commutative monoid, MM a graded AA-module of type Δ\Delta and (Mα)α∈Δ(M_\alpha)_{\alpha \in \Delta} its graduation. For every ordered pair (α,n)∈Δ×N(\alpha, n) \in \Delta \times \mathbf{N}, let Tα,n(M)T^{\alpha, n}(M) be the (direct) sum of the submodules Mα1⊗Mα2⊗⋯⊗MαnM_{\alpha_1} \otimes M_{\alpha_2} \otimes \cdots \otimes M_{\alpha_n} of Tn(M)T^n(M) such that ∑i=1nαi=α\sum_{i=1}^n \alpha_i = \alpha; then (Tα,n(M))(α,n)∈Δ×N(T^{\alpha, n}(M))_{(\alpha, n) \in \Delta \times \mathbf{N}} is the only graduation of type Δ×N\Delta \times \mathbf{N} compatible with the algebra structure on T(M)T(M) and inducing on M=T1(M)M = T^1(M) the given gradation.

It has been seen at the beginning of this no. that T(M)T(M) is the direct sum of the Tα,n(M)T^{\alpha, n}(M) and the fact that this is a graduation compatible with the algebra structure follows immediately from the definitions. If (T′α,n)({T'}^{\alpha, n}) is another graduation of type Δ×N\Delta \times \mathbf{N} on T(M)T(M) compatible with the algebra structure and such that T′α,1(M)=T′α,1{T'}^{\alpha, 1}(M) = {T'}^{\alpha, 1} for α∈Δ\alpha \in \Delta, it follows immediately from the definitions that, for all n≥1n \geq 1 and all α∈Δ\alpha \in \Delta, Tα,n(M)⊂T′α,nT^{\alpha, n}(M) \subset {T'}^{\alpha, n}; but since T(M)T(M) is also the direct sum of the Tα,n(M)T^{\alpha, n}(M), this implies that T′α,n=Tα,n(M){T'}^{\alpha, n} = T^{\alpha, n}(M) (II, § 1, no. 8, Remark).

6. Tensors and tensor notation

Let AA be a commutative ring, MM an AA-module, M∗M^* the dual of MM (II, § 2, no. 3) and II and JJ two disjoint finite sets; the AA-module ⨂i∈I∪JEi\bigotimes_{i \in I \cup J} E_i, where Ei=ME_i = M if i∈Ii \in I, Ei=M∗E_i = M^* if i∈Ji \in J, is denoted by TJI(M)T^I_J(M); the elements of TJI(M)T^I_J(M) are called tensors of type (I,J)(I, J) over MM. They are called contravariant if J=∅J = \varnothing, covariant if I=∅I = \varnothing and mixed otherwise.

Let I′,I′′I', I'' be two subsets of II and J′,J′′J', J'' two subsets of JJ such that I′∪I′′=II' \cup I'' = I,

I′∩I′′=∅, J′∪J′′=J, J′∩J′′=∅I' \cap I'' = \varnothing,\ J' \cup J'' = J,\ J' \cap J'' = \varnothing; then there is a canonical associativity isomorphism (II, § 3, no. 9)

TJI(M)→TJ′I′(M)⊗ATJ′′I′′(M). T_J^I(M) \to T_{J'}^{I'}(M) \otimes_A T_{J''}^{I''}(M).

Considering the tensor algebra T(M⊕M∗)T(M \oplus M^*), it follows from no. 5 that Tn(M⊕M∗)T^n(M \oplus M^*) is canonically identified with the direct sum of the TJI(M)T_J^I(M) where I runs through the set of subsets of the interval {1,n}\{1, n\} of N\mathbf{N} and JJ is the complement of I in {1,n}\{1, n\}.

When I={1,p}I = \{1, p\} and J={p+1,p+q}J = \{p + 1, p + q\} with integers p≥0, q≥0p \geq 0,\ q \geq 0 (where we replace I (resp. J) by ∅\varnothing when p=0p = 0 (resp. q=0q = 0)), the A-module TJI(M)T_J^I(M) is also denoted by Tqp(M)T_q^p(M); the A-modules T0n(M)T_0^n(M) and Tn0(M)T_n^0(M) are therefore by definition the A-modules Tn(M)T^n(M) and Tn(M∗)T^n(M^*) respectively. When I and J are arbitrary finite sets of cardinals p=Card(I)p = \mathrm{Card}(I) and q=Card(J)q = \mathrm{Card}(J), we give each of them a total ordering; then there exists an increasing bijection of I (resp. J) onto {1,p}\{1, p\} (resp. {p+1,p+q}\{p + 1, p + q\}) and these bijections therefore define an isomorphism

TJI(M)→Tqp(M). T_J^I(M) \to T_q^p(M).

When M is a finitely generated projective A-module, it follows from II, § 2, no. 7, Corollary 4 to Proposition 13 and II, § 4, no. 4, Corollary 1 to Proposition 4 that there is a canonical isomorphism

(TJI(M))∗→TIJ(M). (T_J^I(M))^* \to T_I^J(M).

Suppose now that M is a finitely generated free A-module and let (eλ)λ∈L(e_\lambda)_{\lambda \in L} be a basis of M (L therefore being a finite set). The basis of M∗M^* dual to (eλ)(e_\lambda) (II, § 2, no. 6) is denoted by (eλ)λ∈L(e^\lambda)_{\lambda \in L}. The bases (eλ)(e_\lambda) and (eλ)(e^\lambda) of M and M∗M^* respectively define (no. 5) a basis of TJI(M)T_J^I(M), which we give explicitly as follows: given two mappings f:I→Lf : I \to L and g:J→Lg : J \to L, let efge_f^g be the element ⨂i∈I∪Jxi\bigotimes_{i \in I \cup J} x_i of TJI(M)T_J^I(M) defined by

xi=ef(i)if i∈I,xi=eg(i)if i∈J. x_i = e_{f(i)} \quad \text{if } i \in I, \qquad x_i = e_{g(i)} \quad \text{if } i \in J.

When (f,g)(f, g) runs through the set of ordered pairs of mappings f:I→Lf : I \to L and g:J→Lg : J \to L, the efge_f^g form a basis of the A-module TJI(M)T_J^I(M), said to be associated with the given basis (eλ)(e_\lambda) of M. For z∈TJI(M)z \in T_J^I(M), we can therefore write

z=∑(f,g)αgf(z)⋅efg z = \sum_{(f, g)} \alpha_g^f(z) \cdot e_f^g

where the αgf\alpha_g^f are the coordinate forms relative to the basis (efg)(e_f^g); by an abuse of language, the αgf(z)\alpha_g^f(z) are called the coordinates of the tensor zz with respect to the basis (eλ)(e_\lambda) of the module M. The αgf\alpha_g^f constitute the dual basis of (efg)(e_f^g), in other words they are identified with the elements of the basis of TIJ(M)T_I^J(M), associated with (eλ)(e_\lambda). When I and J are complementary subsets of an interval {1,n}\{1, n\} of N\mathbf{N}, αgf\alpha_g^f (or αgf(z)\alpha_g^f(z)) is denoted as follows: each element f(i)f(i) for i∈Ii \in I is written as an upper index in the i-th place with a dot in the i-th place for i∈Ji \in J; similarly, g(i)g(i) for i∈Ji \in J is written as a lower index in the i-th place with a dot in the i-th place for i∈Ii \in I. For example, for I={1,4},J={2,3}I = \{1, 4\}, J = \{2, 3\}, we write αvρλ⋅⋅μ\alpha_{v\rho}^{\lambda\cdot\cdot\mu}: if f(1)=λ,f(4)=μ,g(2)=ν,g(3)=ρf(1) = \lambda, f(4) = \mu, g(2) = \nu, g(3) = \rho.

Let (e~λ)λ∈L(\tilde{e}_{\lambda})_{\lambda \in L} be another basis of M and P the matrix of passing from (eλ)(e_{\lambda}) to (e~λ)(\tilde{e}_{\lambda}) (II, § 10, no. 8). Then the matrix of passing from (eλ)(e^{\lambda}) to (e~λ)(\tilde{e}^{\lambda}) (dual basis of (e~λ)(\tilde{e}_{\lambda})) is the contragredient tP−1^tP^{-1} of P (II, § 10, no. 8, Proposition 5). It follows (II, § 10, no. 10) that the matrix of passing from the basis (efg)(e^{g}_{f}) of TJI(M)T^{I}_{J}(M) to the basis (e~fg)(\tilde{e}^{g}_{f}) (where ff (resp. gg) runs through the set of mappings of I into L (resp. of J into L)) is the matrix

⨂i∈I∪JQi,where Qi=P if i∈I,Qi=tP−1 if i∈J. \bigotimes_{i \in I \cup J} Q_{i}, \quad \text{where } Q_{i} = P \text{ if } i \in I, Q_{i} = {}^{t}P^{-1} \text{ if } i \in J.

The transpose of this matrix is therefore identified with

⨂i∈I∪JRi,where Ri=tP−1 if i∈I,Ri=P if i∈J. \bigotimes_{i \in I \cup J} R_{i}, \quad \text{where } R_{i} = {}^{t}P^{-1} \text{ if } i \in I, R_{i} = P \text{ if } i \in J.

Suppose now that M is an arbitrary module. Let i∈I,j∈Ji \in I, j \in J and write I′=I−{i},J′=J−{j}I' = I - \{i\}, J' = J - \{j\}; we shall define a canonical A-linear mapping

cji:TJI(M)→TJ′I′(M), c^{i}_{j}: T^{I}_{J}(M) \to T^{I'}_{J'}(M),

called contraction of the index i and the index j. For this, note that the mapping of MI×(M∗)JM^{I} \times (M^{*})^{J}, which associates with every family (xi)i∈I∪J(x_{i})_{i \in I \cup J}, where xi∈Mx_{i} \in M if i∈Ii \in I and xi∈M∗x_{i} \in M^{*} if i∈Ji \in J, the element

⟨xi,xj⟩k∈(I∪J)−{i,j}xk \langle x_{i}, x_{j} \rangle_{k \in (I \cup J) - \{i, j\}} x_{k}

of TJ′I′(M)T^{I'}_{J'}(M), is A-multilinear; cjic^{i}_{j} is the corresponding A-linear mapping.

Suppose now that M is free and finitely generated and let (eλ)λ∈L(e_{\lambda})_{\lambda \in L} be a basis of M. Given two mappings f:I→L,g:J→Lf: I \to L, g: J \to L, let fif_{i} denote the restriction of ff to I′=I−{i}I' = I - \{i\} and gjg_{j} the restriction of gg to J′=J−{j}J' = J - \{j\}; then by virtue of (12)

cji(efg)={0if f(i)≠g(j)efigjif f(i)=g(j) c^{i}_{j}(e^{g}_{f}) = \begin{cases} 0 & \text{if } f(i) \neq g(j) \\ e^{g_{j}}_{f_{i}} & \text{if } f(i) = g(j) \end{cases}

The expression for the coordinates of cji(z)c^{i}_{j}(z) as a function of those of z is obtained; for every mapping f′f' (resp. g′g') of I' into L (resp. of J' into L) and every λ∈L\lambda \in L, let (f′,λ)(f', \lambda) (resp. (g′,λ)(g', \lambda)) denote the mapping of I into L (resp. of J into L) whose restriction to I' (resp. J') is f′f' (resp. g′g') and which takes the value λ\lambda at the element i (resp. j). Then, if the coordinate forms relative to the basis (ef′g′)(e^{g'}_{f'}) of TJ′I′(M)T^{I'}_{J'}(M) are denoted by βg′f′\beta^{f'}_{g'},

βg′f′(cji(z))=∑λ∈Lα(g′;λ)(f′;λ)(z). \beta^{f'}_{g'}(c^{i}_{j}(z)) = \sum_{\lambda \in L} \alpha^{(f'; \lambda)}_{(g'; \lambda)}(z).

Examples of tensors. (1) Let M be a finitely generated projective A-module. We know (II, § 4, no. 2, Corollary to Proposition 2), that there is a canonical A-module isomorphism

θM:M∗⊗AM→End⁡A(M) \theta_M : M^* \otimes_A M \to \operatorname{End}_A(M)

such that θM(x∗⊗x)\theta_M(x^* \otimes x) (for x∈M,x∗∈M∗x \in M, x^* \in M^*) is the endomorphism

y↦⟨y,x∗⟩x. y \mapsto \langle y, x^*\rangle x.

Hence, by means of θM\theta_M, T{1}(2)(M)T^{(2)}_{\{1\}}(M) (isomorphic to T11(M)T^1_1(M)) can be identified with the A-module End⁡A(M)\operatorname{End}_A(M). Suppose that M is a free module and let (eλ)λ∈L(e_\lambda)_{\lambda \in L} be a basis of M; then the coordinates of a tensor z∈M∗⊗Mz \in M^* \otimes M relative to the basis (eμ⊗eλ)(e^\mu \otimes e_\lambda) of this module are denoted by ζμ⋅λ\zeta_{\mu \cdot \lambda}. As θM(eμ⊗eλ)\theta_M(e^\mu \otimes e_\lambda) is the endomorphism y↦⟨y,eμ⟩eλy \mapsto \langle y, e^\mu\rangle e_\lambda, the endomorphism u=θM(z)=θM(∑λ,μζμ⋅λeμ⊗eλ)u = \theta_M(z) = \theta_M\left( \sum_{\lambda, \mu} \zeta_{\mu \cdot \lambda} e^\mu \otimes e_\lambda \right) maps yy to ∑λ,μζμ⋅λ⟨y,eμ⟩eλ\sum_{\lambda, \mu} \zeta_{\mu \cdot \lambda} \langle y, e^\mu\rangle e_\lambda; writing y=eλy = e_\lambda, we obtain the relation

u(eλ)=∑μ∈Lζλ⋅μeμ u(e_\lambda) = \sum_{\mu \in L} \zeta_{\lambda \cdot \mu} e_\mu

in other words, the matrix of the linear mapping u=θM(z)u = \theta_M(z) is that whose element in the row of index μ\mu and column of index λ\lambda is ζλ⋅μ\zeta_{\lambda \cdot \mu}.

The definition of the trace of uu (II, § 4, no. 3) shows immediately that

Tr⁡(θM(z))=c12(z). \operatorname{Tr}(\theta_M(z)) = c_1^2(z).

Therefore the element z0=∑λ∈Leλ⊗eλz_0 = \sum_{\lambda \in L} e^\lambda \otimes e_\lambda (whose coordinates ζμ⋅λ\zeta_{\mu \cdot \lambda} are zero for λ≠μ\lambda \neq \mu and equal to 1 for λ=μ\lambda = \mu), which is such that θM(z0)=1M\theta_M(z_0) = 1_M, is the image of the element 1∈A=T00(M)1 \in A = T^0_0(M) under the mapping the transpose of the contraction

c12:T{1}(2)(M)→A. c_1^2 : T^{(2)}_{\{1\}}(M) \to A.

(2) Suppose always that M is a finitely generated projective A-module; there is a canonical A-module isomorphism

μ:M∗⊗AM∗→(M⊗AM)∗ \mu : M^* \otimes_A M^* \to (M \otimes_A M)^*

(II, § 4, no. 4, Corollary 1 to Proposition 2) and a canonical isomorphism

θ:(M⊗AM)∗⊗AM→Hom⁡A(M⊗AM,M) \theta : (M \otimes_A M)^* \otimes_A M \to \operatorname{Hom}_A(M \otimes_A M, M)

(II, § 4, no. 2, Corollary to Proposition 2); also Hom⁡A(M⊗AM,M)\operatorname{Hom}_A(M \otimes_A M, M) is canonically isomorphic to the A-module L2(M,M;M)\mathscr{L}_2(M, M; M) of A-*bilinear* mappings of M×MM \times M into MM (II, § 3, no. 9). Composing these isomorphisms, a canonical isomorphism is obtained

χM:T{1,2}(3)(M)=M∗⊗M∗⊗M→L2(M,M;M) \chi_M : T^{(3)}_{\{1, 2\}}(M) = M^* \otimes M^* \otimes M \to \mathscr{L}_2(M, M; M)

such that, for x∗,y∗x^*, y^* in M∗M^*, z∈Mz \in M, χM(x∗⊗y∗⊗z)\chi_M(x^* \otimes y^* \otimes z) is the bilinear mapping

(u,v)↦⟨u,x∗⟩⟨v,y∗⟩z. (u, v) \mapsto \langle u, x^*\rangle \langle v, y^*\rangle z.
Hence, by means of χM\chi_M, T{1,2}(3)(M)T_{\{1, 2\}}^{(3)}(M) (isomorphic to T21(M)T_2^1(M)) can be identified with the A-module L2(M,M;M)\mathscr{L}_2(M, M; M). Suppose that M is a free A-module and let (eλ)λ∈L(e_\lambda)_{\lambda \in L} be a basis of M; then the coordinates of a tensor z∈M∗⊗M∗⊗Mz \in M^* \otimes M^* \otimes M relative to the basis (eλ⊗eμ⊗eν)(e^\lambda \otimes e^\mu \otimes e_\nu) of this module are denoted by ζλμν\zeta_{\lambda \mu \nu}. The bilinear mapping χM(z)\chi_M(z) maps the ordered pair (eλ,eμ)(e_\lambda, e_\mu) to
∑ν∈Lζλμνeν \sum_{\nu \in L} \zeta_{\lambda \mu \nu} e_\nu
and therefore the ζλμν\zeta_{\lambda \mu \nu} are just the constants of structure of the (in general non-associative) algebra defined on M by the bilinear mapping χM(z)\chi_M(z), with respect to the basis (eλ)(e_\lambda) (§ 1, no. 7).

Remark 2

Let (eλ)λ∈L(e_\lambda)_{\lambda \in L}, (eˉλ)λ∈L(\bar{e}_\lambda)_{\lambda \in L} be two bases of M and P the matrix of passage from (eλ)(e_\lambda) to (eˉλ)(\bar{e}_\lambda). On account of what was seen in Example 1, the element of P appearing in the row of index λ\lambda and the column of index μ\mu is denoted by αμλ\alpha_\mu^\lambda and the element of the contragredient tP−1{}^tP^{-1} appearing in the row of index λ\lambda and the column of index μ\mu is denoted by βλμ\beta_\lambda^\mu. Then (in the notation introduced above)

{eˉμ=∑λαμλeλeˉμ=∑λβλμeλ \begin{cases} \bar{e}_\mu = \sum_\lambda \alpha_\mu^\lambda e_\lambda \\ \bar{e}^\mu = \sum_\lambda \beta_\lambda^\mu e_\lambda \end{cases}
eˉf′g′=∑(f′,g)(∏i∈Iαf′(i)f(i))(∏j∈Jβg(j)g′(j))efg \bar{e}_{f'}^{g'} = \sum_{(f', g)} \left( \prod_{i \in I} \alpha_{f'(i)}^{f(i)} \right) \left( \prod_{j \in J} \beta_{g(j)}^{g'(j)} \right) e_f^g
for all mappings f′:I→Lf': I \to L and g′:J→Lg': J \to L. The coordinates ζgf\zeta_g^f of a tensor z∈TJI(M)z \in T_J^I(M) with respect to the basis (eλ)(e_\lambda) can therefore be expressed in terms of the coordinates ζˉg′f′\bar{\zeta}_{g'}^{f'} of z with respect to the basis (eˉλ)(\bar{e}_\lambda) using the formulae
ζgf=∑(f′,g′)(∏i∈Iαg′(i)g(i))(∏j∈Jβf′(j)f(j))ζˉg′f′ \zeta_g^f = \sum_{(f', g')} \left( \prod_{i \in I} \alpha_{g'(i)}^{g(i)} \right) \left( \prod_{j \in J} \beta_{f'(j)}^{f(j)} \right) \bar{\zeta}_{g'}^{f'}
The matrix P−1P^{-1} of passage from the basis (eˉλ)(\bar{e}_\lambda) to the basis (eλ)(e_\lambda) is the transpose of tP−1{}^tP^{-1}, so that βλμ\beta_\lambda^\mu is the element which appears in the column of index λ\lambda and the row of index μ\mu of P−1P^{-1}. The calculation of efge_f^g in terms of the eˉf′g′\bar{e}_{f'}^{g'} and that of the ζˉg′f′\bar{\zeta}_{g'}^{f'} in terms of the ζgf\zeta_g^f are therefore made by replacing αμλ\alpha_\mu^\lambda by βλμ\beta_\lambda^\mu and βλμ\beta_\lambda^\mu by αμλ\alpha_\mu^\lambda in the above calculations and exchanging the roles of ff and f′f' and those of gg and g′g'. Then
efg=∑(f′,g′)(∏j∈Jαg′(j)g(j))(∏i∈Iβf′(i)f(i))eˉf′g′ e_f^g = \sum_{(f', g')} \left( \prod_{j \in J} \alpha_{g'(j)}^{g(j)} \right) \left( \prod_{i \in I} \beta_{f'(i)}^{f(i)} \right) \bar{e}_{f'}^{g'}

(23)

ζˉg′f′=∑(f,g)(∏j∈Jαg′(j)g(j))(∏i∈Iβf(i)f′(i))ζgf. \bar{\zeta}_{g'}^{f'} = \sum_{(f,g)} \left( \prod_{j \in J} \alpha_{g'(j)}^{g(j)} \right) \left( \prod_{i \in I} \beta_{f(i)}^{f'(i)} \right) \zeta_g^f.

The above formulae are such that the summation is over indices which appear once as a lower index and once as an upper index. Certain authors allow themselves because of this circumstance to suppress the summation signs.

Exercises

See the exercises for § 5.

Transcribed from the printed edition, 1998, Springer. Not translated.