Bourbaki

§ 4. Finitely generated modules over a principal ideal domain

1. Direct sums of cyclic modules

Let A be a commutative ring. Recall (II, p. 220, Prop. 22) that a cyclic A-module is isomorphic to a quotient module A/aA/a, where a is an ideal of A. We will see later (Sect. 4) that every finitely generated module over a principal ideal domain is a direct sum of finitely many cyclic modules.

Proposition 1

— *Let E be a module over a commutative ring A; suppose E is a direct sum of n cyclic modules Λ/ak\Lambda/\mathfrak{a}_k (1≤k≤n1 \leq k \leq n), where the ak\mathfrak{a}_k are ideals of A;

then, for each integer p>0p > 0, the AA-module ⋀pE\bigwedge^p E is isomorphic to the direct sum of the modules A/aHA / a_H, where for each pp-element subset H={k1,…,kp}H = \{ k_1, \ldots, k_p \} of [1,n][1, n], the ideal aHa_H is ∑j=1pakj\sum_{j=1}^p a_{k_j}.

Let xkx_k be the canonical generator of A/akA / a_k, that is the image of the unit element of AA, so that EE is the direct sum of the AxiAx_i (1≤i≤n1 \leq i \leq n). Then we know (III, p. 515, Prop. 10) that the exterior algebra ⋀nE\bigwedge^n E is isomorphic as an AA-module to the tensor product ⨂i=1n(⋀(Axi))\bigotimes_{i=1}^n (\bigwedge (Ax_i)). Now ⋀(Axi)\bigwedge (Ax_i) is just the direct sum A⊕AxiA \oplus Ax_i, since the exterior product of any two elements of AxiAx_i is zero, and so ⋀pE\bigwedge^p E is the direct sum of the modules MH=(Axk1)⊗…⊗(Axkp)M_H = (Ax_{k_1}) \otimes \ldots \otimes (Ax_{k_p}) as H={k1,…,kp}H = \{ k_1, \ldots, k_p \} runs over the set of pp-element subsets of [1,n][1, n] (with k1<…<kpk_1 < \ldots < k_p) ; now MHM_H is known to be isomorphic to A/a11A / a_{11} (II, p. 257, Cor. 4), which completes the proof.

We now see that, in the notation of Prop. 1, if the ideals aka_k form an increasing sequence, they are completely determined by the module EE. More precisely :

Proposition 2

— *Let AA be a commutative ring, and let EE be a direct sum of nn cyclic modules A/akA / a_k, where the aka_k satisfy a1⊂a2⊂…⊂ana_1 \subset a_2 \subset \ldots \subset a_n. Then, for 1≤p≤n1 \leq p \leq n, the ideal apa_p is the annihilator of ⋀pE\bigwedge^p E; if an≠Aa_n \neq A then ⋀pE≠0\bigwedge^p E \neq 0 for 1≤p≤n1 \leq p \leq n and ⋀mE=0\bigwedge^m E = 0 for m>nm > n.

Indeed, in the notation of Prop. 1, we have ap=as(H)a_p = a_{s(H)}, where s(H)s(H) denotes the greatest element of the subset HH. Since s(H)≥ps(H) \geq p for every pp-element subset HH, and since s(H)=ps(H) = p for H={1,…,p}H = \{ 1, \ldots, p \}, it follows that apa_p is the intersection of ara_r, as HH varies over the set of pp-element subsets of (1,n)(1, n); the ideal apa_p is thus indeed the annihilator of ⋀pE\bigwedge^p E, by Prop. 1.

Corollary

— In the notation of Prop. 2, if an≠Aa_n \neq A, and if EE is also isomorphic to the direct sum of mm cyclic modules A/aj′A / a'_j with a1′⊂a2′⊂…⊂am′≠Aa'_1 \subset a'_2 \subset \ldots \subset a'_m \neq A, then m=nm = n and ak=ak′a_k = a'_k for 1≤k≤n1 \leq k \leq n (uniqueness of the ak∗a_k *).

2. Content of an element of a free module

Let AA be a principal ideal domain, let LL be a free AA-module, and let xx be an element of LL. As ff runs through the set L∗L^* of linear forms on LL, the elements f(x)f(x) form an ideal cL(x)c_L(x) of AA, called the content of xx in LL. An element cc of AA is called a content of xx in LL if it generates the ideal cL(x)c_{L}(x); this amounts to saying that there exists a linear form ff on LL such that f(x)=cf(x) = c and that cc divides g(x)g(x) for every linear form gg on LL. Let (ei)i∈I(e_i)_{i \in I} be a basis of LL; put x=∑aiei,ai∈Ax = \sum a_i e_i, a_i \in A; then the ideal cL(x)c_L(x) consists of sums ∑aibi\sum a_i b_i, as (bi)(b_i) runs through the set A′A'; it follows immediately that an element cc of AA is a content of xx in LL if and only if it is a gcd of the family (ai)(a_i) of coordinates of xx.

We say that xx is indivisible if cL(x)=Ac_L(x) = A, that is if the coordinates of xx with respect to a basis of LL are setwise coprime.

Lemma 1

Let LL be a free module over a principal ideal domain AA and let xx be an element of LL. Then the following conditions are equivalent*:

(i) xx is indivisible ; (ii) *there exists a linear form ff on LL such that f(x)=1f(x) = 1 ; (iii) xx is nonzero and the submodule AxAx is a direct factor of LL ; (iv) xx is an element of some basis of LL.

(i) ⇒\Rightarrow (ii) : this follows from the definition. (ii) ⇒\Rightarrow (iii) : let ff be a linear form on LL such that f(x)=1f(x) = 1 ; then x≠0x \neq 0 and the map y↦f(y)xy \mapsto f(y)x is a projector of LL, with image AxAx. (iii) ⇒\Rightarrow (iv) : let L′L' be a complement of AxAx in LL, and let B′B' be a basis of L′L' (VII, p. 15, Cor. 2) ; then B′∪{x}B' \cup \{x\} is a basis of LL. (iv) ⇒\Rightarrow (i) : trivial.

Remark

— 1) If xx is a nonzero element of LL and cc a content of xx, then there exists a unique element yy of LL such that x=cyx = cy ; denote this element x/cx/c ; then x/cx/c is an indivisible element of LL. 2) The content cL(x)c_L(x) is the annihilator of the torsion module of L/AxL/Ax.

Let LL be a free module over a principal ideal domain AA and let MM be a submodule of LL ; by VII, p. 4, Lemma 1, the family cL(x),x∈Mc_L(x), x \in M, has a maximal element ; if M≠{0}M \neq \{0\} then such a maximal element is nonzero.

Proposition 3

Let LL be a free module over a principal ideal domain AA and let MM be a nonzero submodule of LL. Let xx be an element of MM such that cL(x)c_L(x) is maximal among the contents of elements of MM, let cc be a content of xx in LL, and let ff be a linear form on LL such that f(x)=cf(x) = c.

a) LL is the direct sum of A(x/c)A(x/c) and the kernel KK of ff. b) MM is the direct sum of AxAx and K∩MK \cap M. c) g(M)⊂Acg(M) \subset Ac for every linear form gg on LL.

Put y=x/cy = x/c ; clearly Ay∩K={0}Ay \cap K = \{0\}, since f(y)=1f(y) = 1. Furthermore, for each u∈Lu \in L we have

u=f(u) y+(u−f(u) y), u = f(u)\,y + (u - f(u)\,y),
with f(u)y∈Ayf(u)y \in Ay and u−f(u)y∈Ku - f(u)y \in K ; this proves a). Now note that for u∈Mu \in M we have f(u)∈Acf(u) \in Ac : indeed, let u∈Mu \in M, and let dd be a gcd of f(u)f(u) and cc ; then there exist λ,μ∈A\lambda, \mu \in A such that d=Xf(u)+μc=f(Xu+μx)d = Xf(u) + \mu c = f(Xu + \mu x) ; hence the content of the element λu+μx\lambda u + \mu x of MM divides dd ; by the maximality of cc, this implies that dd is an associate of cc, so f(u)∈Acf(u) \in Ac. Hence for all uu in MM, we can write
u=(f(u)/c) x+(u−(f(u)/c) x)∈Ax+(K∩M), u = (f(u)/c)\,x + (u - (f(u)/c)\,x) \in Ax + (K \cap M),
which shows b). Finally, let gg be a linear form on LL ; by a) there exists a scalar α∈A\alpha \in A and a linear form hh on KK such that g(u)=αf(u)+h(u−f(u) y)g(u) = \alpha f(u) + h(u - f(u)\,y) ; thus by

b) we have g(M)⊂Ac+h(K∩M)g(M) \subset Ac + h(K \cap M). To prove c), it is therefore sufficient to prove that h(K∩M)⊂Ach(K \cap M) \subset Ac for every linear form hh on KK, or equivalently for every linear form hh on LL such that h(x)=0h(x) = 0; now, if u∈K∩Mu \in K \cap M and dd is a gcd of h(u)h(u) and cc, then there exist A,μ∈AA, \mu \in A with d=λh(u)+μcd = \lambda h(u) + \mu c; then (f+h)(Au+μx)=d(f + h)(Au + \mu x) = d, which implies as above that h(u)∈Ach(u) \in Ac, and c) follows.

3. Invariant factors of a submodule

Theorem 1

— Let LL be a free module over a principal ideal domain AA, and let MM be a submodule of LL of finite rank nn. Then there exist a basis BB of LL, nn elements eie_i of BB, and nn nonzero elements αi\alpha_i of AA (1≤i≤n1 \leq i \leq n) such that: a) the αiei\alpha_i e_i form a basis of MM; b) aia_i divides αi−1\alpha_{i-1}, for 1≤i≤n−11 \leq i \leq n-1.

Moreover the module M′M' generated by the (ei)(e_i) and the principal ideals AαiA\alpha_i are uniquely determined by the above conditions; the module M′/MM'/M is the torsion submodule of L/ML/M, and is isomorphic to the direct sum of the AA-modules A/AαiA/A\alpha_i; finally L/ML/M is the direct sum of M′/MM'/M and a free module isomorphic to L/M′L/M'.

1) Existence of the eie_i and the αi\alpha_i.

If M={0}M = \{0\} then the theorem is trivial. If M≠{0}M \neq \{0\} then it follows from Prop. 3 that there exist an element e1e_1 of LL, a nonzero element aa, of AA, and a submodule L1L_1 of LL, such that LL is the direct sum of Ae1Ae_1 and L1L_1, such that MM is the direct sum of Aα1e1A\alpha_1 e_1 and the submodule M1=M∩L1M_1 = M \cap L_1 of LL, and such that g(M)⊂Aα1g(M) \subset A\alpha_1 for every linear form gg on LL.

Now we can proceed by induction on the rank nn of MM. Since L1L_1 is a free module (VII, p. 15, Cor. 2) and M1M_1 has rank n−1n-1, there exist a basis B1B_1 of L1L_1, (n−1)(n-1) elements e2,…,ene_2, \ldots, e_n of B1B_1, and nonzero elements α2,…,a\alpha_2, \ldots, a, of AA such that (α2e2,…,αnen)(\alpha_2 e_2, \ldots, \alpha_n e_n) is a basis of M1M_1, and αi\alpha_i divides αi+1\alpha_{i+1} for 2≤i≤n−12 \leq i \leq n-1. If L′L' is the submodule of LL, generated by the elements of B1B_1 distinct from e1,…,ene_1, \ldots, e_n, then LL is the direct sum of L′L' and the module M′M' generated by e1,…,ene_1, \ldots, e_n; now (e1,…,en)(e_1, \ldots, e_n) is a basis for M′M', and (α1e1,…,αnen)(\alpha_1 e_1, \ldots, \alpha_n e_n) is a basis for MM. It now remains only to show that aia_i divides aia_i; but Aα2A\alpha_2 has the form g(M1)g(M_1), where gg is the linear form on LL defined by g(e2)=1,g(ei)=0g(e_2) = 1, g(e_i) = 0 for i≠2i \neq 2, and g(L′)=(0)g(L') = (0); and we have seen above that g(M1)⊂Aα1g(M_1) \subset A\alpha_1.

2) Uniqueness properties.

Since the aia_i are nonzero, the module M′M' is the set of x∈Lx \in L such that βx∈M\beta x \in M for some β≠0\beta \neq 0 in AA; in other words M′/MM'/M is the torsion submodule of L/ML/M. This uniquely determines M′M'.

It is clear that M′/MM'/M is isomorphic to the direct sum of the nn cyclic modules A/AαiA/A\alpha_i (II, p. 204, formula (26)). Let rr be the number of the ideals AαA\alpha, which are distinct from AA: so that the first n−rn-r ideals AαiA\alpha_i are equal to AA, and the last rare distinct from AA. Then M′/MM'/M is also isomorphic to the direct sum of the modules A/Aαn,…,A/Aαn−r+1A/A\alpha_n, \ldots, A/A\alpha_{n-r+1}, where Aαn⊂Aαn−1⊂…⊂Aαn−r+1≠AA\alpha_n \subset A\alpha_{n-1} \subset \ldots \subset A\alpha_{n-r+1} \neq A.

Thus the conditions of the Cor. to Prop. 2 (VII, p. 16) are satisfied: the ideals AaiAa_i (1≤i≤n1 \leq i \leq n) are thus uniquely determined by the module M′/MM'/M.

Since LL is the direct sum of M′M' and L′L', it follows that L/ML/M is the sum of M′/MM'/M and (L′+M)/M(L' + M)/M, and this sum is direct since (L′+M)∩M′=M(L' + M) \cap M' = M; on the other hand (L′+M)/M(L' + M)/M is isomorphic to L′/(M∩L′)L'/(M \cap L') (I, p. 41, Th. 4, c)), that is to L′L', which shows that (L′+M)/M(L' + M)/M is a free module isomorphic to L/M′L/M'.

Corollary

— A submodule MM of finite rank in a free module LL over a principal ideal domain AA has a complement in LL if and only if L/ML/M is torsion free.

In the notation of Th. 1, if L/ML/M is torsion free, then M=M′M = M', and M′M' has a complement L′L' in LL. Conversely if MM has a complement L′L' in LL, then L/ML/M is isomorphic to L′L', which is free (VII, p. 15, Cor. 2), and a fortiori torsion free.

Remark

It can happen that a submodule MM of infinite rank in a free module LL can be such that L/ML/M is torsion free, but MM has no complement in LL (VII, p. 60, Ex. 6, b)).

Definition 1

— In the notation and hypotheses of Th. 1, the ideals AαiA\alpha_i of AA are called the invariant factors of the submodule MM with respect to the module LL.

In the case where AA is either the ring Z\mathbf{Z} of rational integers or the polynomial ring K[X]K[X] in one indeterminate over a field KK, there is a canonical way of choosing a generator for each ideal of AA: a positive integer in the case of Z\mathbf{Z}, or a monic polynomial in the case of K[X]K[X] (VII, p. 5). In each of these cases, the canonical generator of the invariant factor AaiAa_i is also called an invariant factor of MM with respect to LL, by abuse of language.

4. Structure of finitely generated modules

Theorem 2

— Every finitely generated module EE over a principal ideal domain AA is isomorphic to a direct sum of a finite number mm of cyclic modules A/akA/a_k, where the aka_k are ideals of AA (some of which may be zero) such that a1⊂a2⊂…⊂am≠Aa_1 \subset a_2 \subset \ldots \subset a_m \neq A, and which are uniquely determined by these conditions.

If EE can be generated by qq elements then it is isomorphic to a quotient module L/ML/M, where L=AqL = A^q (II. p. 218). Since MM has finite rank n≤qn \leq q (VII, p. 15, Prop. 1), the conditions of Th. 1 (VII, p. 18) are satisfied. Then in the notation of Th. 1 L/ML/M is isomorphic to a direct sum of a complement L′L' of M′M' in LL and the torsion module M′/MM'/M. The module L′L' is free of finite rank p=q−np = q - n, so isomorphic to ApA^p. If rr is the smallest index such that Aαr≠AA\alpha_r \neq A, then M′/MM'/M is isomorphic to the direct sum of modules A/AαiA/A\alpha_i for r≤i≤nr \leq i \leq n. The stated conditions will then be satisfied if we take m=p+(n−r+1)m = p + (n - r + 1), ak=(0)a_k = (0) for 1≤k≤p1 \leq k \leq p, and aj=Aαn,−,a_j = A\alpha_n , - , for 1≤j≤n−r+11 \leq j \leq n - r + 1. Uniqueness follows from the Cor. of VII, p. 16.

Corollary 1

— *Every finitely generated module E over a principal ideal domain is the direct sum of the torsion submodule of E and a free module.

The torsion submodule of E has in general several distinct complements. For example, if E=Z×(Z/(2))E = \mathbf{Z} \times (\mathbf{Z}/(2)) then the torsion submodule of E is {0}×(Z/(2))\{0\} \times (\mathbf{Z}/(2)); it has as a complement the submodule Z×(0)\mathbf{Z} \times (0), and also the submodule consisting of all elements (n,n‾)(n, \overline{n}), where nn runs through Z\mathbf{Z} and n‾\overline{n} is the residue class of nn mod 2.

Corollary 2

— Every torsion free finitely generated module over a principal ideal domain is free of finite rank. This follows immediately from Cor. 1.

The condition that the module be finitely generated is essential. For example the additive group of the field of fractions K of A is torsion free as an A-module; however it is not a free module if A≠KA \neq K, for on the one hand any two elements of K have a common divisor, and on the other hand K is not a cyclic A-module, for otherwise K=ab−1AK = ab^{-1}A (a,b∈Aa, b \in A), whence b−2=acb−1b^{-2} = acb^{-1} (c∈Ac \in A), so b−1=ac∈Ab^{-1} = ac \in A, and K=AK = A.

Definition 2

— In the notation and hypotheses of Th. 2, the ideals aka_k are called the invariant factors of the module E. As in Def. 1 (VII, p. 19), when A=ZA = \mathbf{Z} or A=K[X]A = K[X], the canonical generator of the ideal aka_k (a positive integer or a monic polynomial) is also called, by abuse of language, an invariant factor of the finitely generated module E.

We must be careful not to confuse the invariant factors of a module E with those of a submodule M of a free module L with respect to the module L (Def. I).

5. Calculation of invariant factors

Proposition 4

— *Let A be a principal ideal domain, let L be a free A-module with finite basis (uj)(u_j) (1≤j≤k1 \leq j \leq k), let M be a submodule of L, let (xi)(x_i) be a system of generators of M and let AαiA \alpha_i (1≤i≤n1 \leq i \leq n) be the invariant factors of M with respect to L. Then for 1≤m≤n1 \leq m \leq n the product δm=α1…αm\delta_m = \alpha_1 \ldots \alpha_m is a gcd of the m-th order minors of the matrix whose columns are the coordinate vectors of the xix_i with respect to the basis (uj)(u_j).

By Th. 1 it is clear that M⊂α1LM \subset \alpha_1 L; hence the coordinates of any element of M are multiples of a1a_1. On the other hand, there exists an element xx of M for which a1a_1 is a content in L. Expressing xx as a linear combination of the xix_i, it follows that a1a_1 is an element of the ideal generated by the coordinates of the xix_i. As these are all multiples of a1a_1, it follows that a1a_1 is indeed their gcd, and our assertion is proved in the case m=1m = 1.

For general mm, consider the m-th exterior power ⋀mM\bigwedge^m M of M (III, p. 507). In the notation of Th. 1 there exists a basis (ai)(a_i) for M where ai=αieia_i = \alpha_i e_i (1≤i≤n1 \leq i \leq n); thus ⋀mM\bigwedge^m M has a basis consisting of the elements ai1∧…∧aima_{i_1} \wedge \ldots \wedge a_{i_m}, as (i1,…,im)(i_1, \ldots, i_m) runs through the set of strictly increasing sequences of mm elements of (1,n)(1, n). Now the elements ei1∧…∧eime_{i_1} \wedge \ldots \wedge e_{i_m} belong to a basis BmB_m of ⋀mL\bigwedge^m L. Hence the canonical map from ⋀mM\bigwedge^m M into ⋀mL\bigwedge^m L is an isomorphism from ⋀mM\bigwedge^m M onto the submodule of ⋀mL\bigwedge^m L having as a basis the elements (αi1…αim) ei1∧…∧eim(\alpha_{i_1} \ldots \alpha_{i_m})\, e_{i_1} \wedge \ldots \wedge e_{i_m}, and we identify this submodule with ⋀mM\bigwedge^m M. Since αj\alpha_j is a multiple of αk\alpha_k for j≥kj \geq k, the elements αi1…αim\alpha_{i_1} \ldots \alpha_{i_m} are all multiples of δm=α1…αm\delta_m = \alpha_1 \ldots \alpha_m, and one of them is equal to δm\delta_m; thus δm\delta_m is a gcd of the coordinates with respect to BmB_m of the elements of a system of generators of ⋀mM\bigwedge^m M. The first part of the argument shows that then δm\delta_m is a gcd of the set of coordinates of any system of generators of ⋀mM\bigwedge^m M, with respect to any basis of ⋀mL\bigwedge^m L. Taking the basis for ⋀mL\bigwedge^m L induced from the basis (uj)(u_j) of LL, and the system of generators for ⋀mM\bigwedge^m M consisting of exterior products of the (xi)(x_i), the expression for the coordinates of these products in terms of determinants (III, p. 528, Prop. 9) gives the stated result.

6. Linear mappings of free modules, and matrices over a principal ideal domain

Let AA be a principal ideal domain. Consider a linear mapping ff from a free AA-module LL of rank mm into a free AA-module L′L' of rank nn. The preceding results allow us, by choosing suitable bases for LL and L′L', to put the matrix for ff in a particularly simple form, called the canonical form of the matrix.

Proposition 5

Let AA be a principal ideal domain, and ff a linear map of rank rr from a free AA-module LL of rank mm into a free AA-module L′L' of rank nn. Then there exist bases (ei)(e_i) (1≤i≤m1 \leq i \leq m) of LL and (ej′)(e'_j) (1≤j≤n1 \leq j \leq n) of L′L' such that f(ei)=αiei′f(e_i) = \alpha_i e'_i for 1≤i≤r1 \leq i \leq r and f(ei)=0f(e_i) = 0 for i>ri > r, where the αi\alpha_i are nonzero elements of AA, each of which divides the next; the ideals AαiA \alpha_i are the invariant factors of f(L)f(L) in L′L', and are thus uniquely determined.

Let L0=f−1(0)L_0 = f^{-1}(0) be the kernel of ff; the quotient L/L0L/L_0 is isomorphic to the module f(L)f(L), which is a submodule of L′L' and so free (VII, p. 15, Cor. 2); thus L0L_0 has a complement L1L_1 in LL (II, p. 218, Prop. 21), and the restriction of ff to L1L_1 is an isomorphism from L1L_1 onto f(L)=M′f(L) = M'. If the ideals AαiA \alpha_i (1≤i≤r1 \leq i \leq r) are the invariant factors of M′M' in L′L', then Th. 1 of VII, p. 18 shows that there exists a basis (ej′)(e'_j) (1≤j≤n1 \leq j \leq n) of L′L' such that (αiei′)(\alpha_i e'_i) (1≤i≤r1 \leq i \leq r) is a basis of M′M'. Since the restriction of ff to L1L_1 is an isomorphism from L1L_1 onto M′M', there exists a basis (ei)(e_i) (1≤i≤r1 \leq i \leq r) of L1L_1 such that f(ei)=αiei′f(e_i) = \alpha_i e'_i. This basis extends to a basis (es)(e_s) (1≤k≤m1 \leq k \leq m) of LL by taking (es)(e_s) (r+1≤s≤mr+1 \leq s \leq m) to be a basis of the kernel L0L_0.

Corollary 1

— Let X be a matrix of rank r, with n rows and m columns, over a principal ideal domain A; then there exists a matrix X0X_0 equivalent to X (II, p. 354) of the form

(α10…00…00α2…00…000…αr0…000…00…000…00…0) \begin{pmatrix} \alpha_1 & 0 & \ldots & 0 & 0 & \ldots & 0 \\ 0 & \alpha_2 & \ldots & 0 & 0 & \ldots & 0 \\ 0 & 0 & \ldots & \alpha_r & 0 & \ldots & 0 \\ 0 & 0 & \ldots & 0 & 0 & \ldots & 0 \\ 0 & 0 & \ldots & 0 & 0 & \ldots & 0 \end{pmatrix}

where the αi\alpha_i are nonzero elements of A, each of which divides the next. Under these conditions the αi\alpha_i are uniquely determined up to multiplication by invertible elements.

Given that two matrices X and X′X' are equivalent if there exist invertible square matrices P and Q, of orders n and m, over A, such that X′=PXQX' = PXQ, Cor. 1 is just Prop. 5 expressed in terms of matrices.

In the notation of Prop. 5 and Cor. 1, the (nonzero) ideals AuAu, are called the invariant factors of the linear map f, or of the matrix X. It then follows immediately from Cor. I that:

Corollary 2

— Two matrices X and X′X' with n rows and m columns over a principal ideal domain A are equivalent if and only if they have the same invariant factors.

Note that when A is a field we can take the αi\alpha_i to be equal to 1, and then we recover Prop. 13 of II, p. 360.

If X is the matrix of the linear mapping f with respect to an arbitrary basis of L and an arbitrary basis of L', then the columns of X are the coordinate vectors, with respect to the basis of L', of elements of L' which form a system of generators for f(L)f(L). The following result is thus an immediate consequence of Prop. 4.

Proposition 6

— Let X be a matrix of rank r over a principal ideal domain A, and let AαiA \alpha_i (1≤i≤r1 \leq i \leq r) be its sequence of invariant factors. Then α1\alpha_1 is a gcd of the elements of X; and the product α1…αq\alpha_1 \ldots \alpha_q is a gcd of the q-th order minors of X for q≤rq \leq r.

7. Finitely generated abelian groups

In the case A=ZA = \mathbf{Z}, the results of section 4 can be expressed

Theorem 3

— Every finitely generated abelian group G is a direct sum of its torsion subgroup F (subgroup of elements of finite order in G) und a free abelian group of finite rank p (isomorphic to Zp\mathbf{Z}^p). The group F is a direct sum of a finite number of cyclic groups of orders n1,n2,…,nqn_1, n_2, \ldots, n_q, where the nin_i are integers >1> 1, each of which divides the previous one; moreover, the integers p,qp, q and nin_i (1≤i≤q1 \leq i \leq q) are uniquely determined by GG.

Remark

While the orders n1,…,nqn_1, \ldots, n_q of the cyclic groups of which F is the direct sum are well defined by the divisibility condition of Th. 3, it is not the same for the groups themselves: for example, in the product G of Z/(p)\mathbf{Z}/(p) with itself (for p prime), the subgroups are precisely the Fp\mathbf{F}_p-vector subspaces, and G is the direct sum of two 1-dimensional subspaces in p(p+1)p(p+1) different ways.

Corollary 1

— *In a finite abelian group G, there exists an element whose order is the lcm of all the orders of elements of G; this order n is the first invariant factor of G.

Corollary 2

— Any finite abelian group G whose order is not divisible by the square of any integer >1> 1 is cyclic. Let us keep to the notation of Th. 3. Then p=0p = 0 since G is finite, and q≤1q \leq 1, for otherwise the order of G would be divisible by nq2n_q^2. Hence G is cyclic.

Corollary 3

— *Let L and M be two free Z\mathbf{Z}-modules of rank n, let (ei)(e_i) be a basis for L and (fi)(f_i) a basis for M (1≤i≤n1 \leq i \leq n), let u be a homomorphism from L into M, and let U be its matrix with respect to the bases (ei)(e_i) and (fi)(f_i). Then Coker u=M/u(L)\mathrm{Coker}\,u = M/u(L) is finite if and only if det⁡(U)≠0\det(U) \neq 0, and then Card(Coker(u))=∣det⁡(U)∣\mathrm{Card}(\mathrm{Coker}(u)) = |\det(U)|. By changing bases in L and M if necessary, we may assume that U has the form described in VII, p. 22, Cor. 1 to Prop. 5 (where the aia_i are in this case integers); the corollary then becomes obvious, since the order of a direct sum of Z\mathbf{Z}-modules Z/αiZ\mathbf{Z}/\alpha_i \mathbf{Z} (1≤i≤n1 \leq i \leq n) is infinite if one of the aia_i is zero, and is equal to ∣α1α2…an∣|\alpha_1 \alpha_2 \ldots a_n| otherwise.

8. Indecomposable modules. Elementary divisors

Definition 3

— A left module M over a ring A is said to be decomposable if it is the direct sum of a family of proper nonzero submodules. Otherwise it is said to be indecomposable. The zero module is thus decomposable, being the direct sum of the empty family of submodules. Let a\mathfrak{a} be a left ideal of the ring A; the submodules of A/aA/\mathfrak{a} are just the quotients b/ab/\mathfrak{a}, where bb is an ideal of A containing a\mathfrak{a} (I, p. 41, Th. 4); if bb and cc are two ideals of A containing a\mathfrak{a}, then the module A/aA/\mathfrak{a} is the direct sum of the submodules b/ab/\mathfrak{a} and c/ac/\mathfrak{a} if and only if A=b+cA = b + c and b∩c=ab \cap c = \mathfrak{a}. As a result:

Lemma 2

— *The module A/aA/\mathfrak{a} is indecomposable if and only if a≠A\mathfrak{a} \neq A and there is no pair (b,c)(b, c) of ideals of A, distinct from A and a\mathfrak{a}, such that A=b+cA = b + c and b∩c=ab \cap c = \mathfrak{a}.

Proposition 7

— Let AA be a commutative ring, let p\mathfrak{p} be a prime ideal of AA (I, p. 117, Def. 3), and let qq be an ideal of AA contained in p\mathfrak{p}. Suppose that for every x∈px \in \mathfrak{p} there exists an integer n>0n > 0 such that xn∈qx^n \in q. Then the AA-module A/qA/q is indecomposable.

Let bb and cc be two ideals of AA, such that A=b+cA = b + c and b∩c=qb \cap c = q. Then bc⊂b∩c=q⊂pbc \subset b \cap c = q \subset \mathfrak{p}; if x∉px \notin \mathfrak{p} and x∈cx \in c, then xb⊂pxb \subset p, so b⊂pb \subset \mathfrak{p} (I, p. 316, Prop. 4); hence either b⊂pb \subset \mathfrak{p} or c⊂pc \subset p. Suppose for example that c⊂pc \subset \mathfrak{p}, so b+p=Ab + p = A; then there exist x∈bx \in b and y∈py \in \mathfrak{p} such that 1=x+y1 = x + y; let n∈Nn \in \mathbf{N} be such that yn∈qy^n \in q; then 1=(x+y)n1 = (x + y)^n. so 1∈xA+ynA⊂b+q⊂b1 \in xA + y^nA \subset b + q \subset b, so b=Ab = A. Lemma 2 now shows that A/qA/q is indecomposable.

Now suppose that AA is a principal ideal domain; by VII, p. 2, Prop. 2, the prime ideals of AA are the ideals (p)(p), where pp is an irreducible element of AA, and the ideal 00; by the previous proposition, the modules AA and A/(pn)A/(p^n), for pp irreducible and n>0n > 0, are indecomposable. Since every cyclic module is a direct sum of modules of this type (VII, p. 3, Prop. 4) and since every finitely generated AA-module is a direct sum of cyclic modules (VII, p. 19, Th. 2), we deduce:

Proposition 8

— Let AA be a principal ideal domain and let MM be a finitely generated AA-module.

a) MM is indecomposable if and only if it is isomorphic to AA or to a module of the form A/(pn)A/(p^n), where pp is an irreducible element of AA and n>0n > 0 is an integer.

b) MM is a direct sum of a finite family of indecomposable submodules.

Part b) of the above proposition can be made more precise as follows:

Proposition 9

— Let AA be a principal ideal domain, let PP be a system of representatives of irreducible elements of AA and let MM be a finitely generated AA-module. Then there exist positive integers m(0)m(0) and m(pn)m(p^n) (p∈P,n>0p \in P, n > 0), uniquely determined by MM and zero except for finitely many of them, such that MM is isomorphic to the direct sum of Am(0)A^{m(0)} and the (A/(pn))m(pn)(A/(p^n))^{m(p^n)} (p∈P,n>0p \in P, n > 0).

The existence of the integers m(0)m(0) and m(pn)m(p^n) (p∈P,n>0p \in P, n > 0) follows from Prop. 8. The integer m(0)m(0) is uniquely determined: it is the rank of the free module which is the quotient of MM by its torsion submodule. Finally, the pp-primary component of MM is isomorphic to the direct sum of the (A/(pn))m(pn)(A/(p^n))^{m(p^n)}; since the family of ideals (pn)(p^n) (n≥1n \geq 1) is totally ordered by inclusion, the uniqueness of the m(pn)m(p^n) follows from the Cor. to Prop. 2 of VII, p. 16.

Definition 4

— In the notation of Prop. 9, those ideals (pn)(p^n) (p∈P,n≥1p \in P, n \geq 1 an integer) such that m(pn)>0m(p^n) > 0 are called the elementary divisors of the module MM, and the integers m(pn)m(p^n) are called their multiplicities; if the integer m(0)m(0) is >0> 0, it is called the multiplicity of the elementary divisor 0.

As for the invariant factors (VII, p. 19, Def. 1), when A=ZA = \mathbf{Z} or A=K[X]A = K[X] (K a commutative field), then the canonical generator of the ideal (pn)(p^n) (a positive integer or a monic polynomial) is also called, by abuse of language, an elementary divisor of the finitely generated module MM.

Remark 1

If MM is a finite abelian group, then its structure can be described by writing down its elementary divisors, each repeated as often as its multiplicity. We will say, for example, that MM is « of type (2,2,4,27,27,25)(2, 2, 4, 27, 27, 25) » (or that it is « a group (2,2,4,27,27,25)(2, 2, 4, 27, 27, 25) ») if it is isomorphic to the product of two groups Z/(2)\mathbf{Z}/(2), one group Z/(22)\mathbf{Z}/(2^2), two groups Z/(33)\mathbf{Z}/(3^3) and one group Z/(52)\mathbf{Z}/(5^2).

Remark 2

If a finitely generated torsion module MM over a principal ideal domain AA is given as a direct sum of cyclic modules isomorphic to A/(ai)A/(a_i) (in particular if the invariant factors of MM are known), then the elementary divisors of MM, and their multiplicities, can be determined by noticing that A/(a)A/(a) is isomorphic to the product of the A/(pn(p))A/(p^{n(p)}), where a=e∏p∈Ppn(p)a = e \prod_{p \in P} p^{n(p)} is the decomposition of aa into irreducible factors (VIIVII, p. 3). Let us study for example the multiplicative group G(464 600)G(464\ 600), where G(n)G(n) denotes the multiplicative group (Z/nZ)∗(\mathbf{Z}/n\mathbf{Z})^* (VIIVII, p. 12). Since 464 600=23⋅5⋅23⋅101464\ 600 = 2^3 \cdot 5 \cdot 23 \cdot 101, this group is isomorphic to the product of the groups G(23)G(2^3), G(52)G(5^2), G(23)G(23) and G(101)G(101) (VIIVII, p. 13, Th. 3); now the last three groups are cyclic of orders 20, 22 and 100, and G(23)G(2^3) is the product of two cyclic groups of order 2 (\emph{loc. cit.}); since 20=22⋅520 = 2^2 \cdot 5, 22=2⋅1122 = 2 \cdot 11 and 100=2′⋅52100 = 2' \cdot 5^2, the group G(464 600)G(464\ 600) is of type (2,2,2,22,22,22,5,52,11)(2, 2, 2, 2^2, 2^2, 2^2, 5, 5^2, 11).

Remark 3

To calculate the invariant factors of a torsion module whose elementary divisors are known, we again lean on the fact that, if the aia_i are pairwise coprime elements of AA, then the product ∏A/(ai)\prod A/(a_i) is a cyclic module isomorphic to A/(a1a2...an)A/(a_1 a_2 ... a_n) (VIIVII, p. 3, Prop. 4). Let us illustrate the method by looking at the example of the group G(464 600)=MG(464\ 600) = M: write the elementary divisors pnp^n of MM which are powers of the same irreducible pp on the same line, beginning with those of greatest exponent; extend these lines to lines of equal length by putting in 1's where necessary:

2222222525111111111. \begin{array}{cccccc} 2^2 & 2^2 & 2 & 2 & 2 \\ 5^2 & 5 & 1 & 1 & 1 \\ 11 & 1 & 1 & 1 & 1 . \end{array}

Then the invariant factors are the products of elements in the same column: 1100,20,2,2,21100, 20, 2, 2, 2. Indeed MM is isomorphic to a product of cyclic groups of orders 1100, 20, 2, 2, 2 by Prop. 4 of VIIVII, p. 3; since each of these orders is a multiple of the next, these are the invariant factors of MM (VIIVII, p. 22, Th. 3).

Remark 4

An AA-module is called simple (II, p. 37) if it is nonzero and has no submodules other than itself and 00; it is then necessarily cyclic, so finitely generated, and indecomposable; since the modules A/(pn)A/(p^n) are not simple for n≠1n \neq 1, while the modules A/(p)A/(p) are, and since AA is simple only if the ring AA is a field, we deduce that the simple AA-modules are: a) free modules of rank 1, when AA is a field; b) modules isomorphic to quotients A/(p)A/(p), where pp is an irreducible element of AA, when AA is not a field.

9. Duality in modules of finite length over a principal ideal domain

In this section AA denotes a principal ideal domain which is not a field (and hence has at least one irreducible element), and KK the field of fractions of AA. For every AA-module MM, put

D(M)=Hom⁡A(M,K/A); D(M) = \operatorname{Hom}_A(M, K/A);

we know that D(M)D(M) is equipped with the structure of an A-module in a natural way, namely for every homomorphism u:M→K/Au : M \to K/A and every α∈A\alpha \in A, the homomorphism αu\alpha u maps xx to αu(x)=u(αx)\alpha u(x) = u(\alpha x). To every homomorphism f:M→Nf : M \to N of A-modules is associated the homomorphism D(f):D(N)→D(M)D(f) : D(N) \to D(M), where D(f)(v)=v∘fD(f)(v) = v \circ f (II, p. 196). For x∈Mx \in M and x′∈D(M)x' \in D(M), we put ⟨x,x′⟩=x′(x)∈K/A\langle x, x' \rangle = x'(x) \in K/A ; then (x,x′)↦⟨x,x′⟩(x, x') \mapsto \langle x, x' \rangle is an A-bilinear map from M×D(M)M \times D(M) into K/AK/A, called canonical.

If M and NN are two A-modules, then to every A-bilinear map φ:M×N→K/A\varphi : M \times N \to K/A are associated the A-linear maps dφ:N→D(M)d_\varphi : N \to D(M) and sφ:M→D(N)s_\varphi : M \to D(N), where dφ(y)(x)=φ(x,y)=sφ(x)(y)d_\varphi(y)(x) = \varphi(x, y) = s_\varphi(x)(y) (II, p. 268, Cor. to Prop. 1). In particular the canonical A-bilinear map M×D(M)→K/AM \times D(M) \to K/A defines an A-linear map (also called canonical)

cM:M→D(D(M)) c_M : M \to D(D(M))

such that ⟨x′,cM(x)⟩=⟨x,x′⟩\langle x', c_M(x) \rangle = \langle x, x' \rangle for x∈Mx \in M and x′∈D(M)x' \in D(M).

Proposition 10

— If M is an A-module of finite length, then D(M)D(M) is (in general non-naturally) isomorphic to M, and the canonical map cM:M→D(D(M))c_M : M \to D(D(M)) is an isomorphism. Using VII, p. 19, Th. 2 and II, p. 203, Cor. 1, we reduce to the case where MM is cyclic. Thus we may suppose that M=A/tAM = A/tA with t≠0t \neq 0. Note that any homomorphism u:A/tA→K/Au : A/tA \to K/A is completely determined by the image ξ∈K/A\xi \in K/A under uu of the class ε\varepsilon of 1 mod tAtA, and this element must satisfy the relation tξ=0t\xi = 0 ; conversely, for any ξ∈K/A\xi \in K/A such that tξ=0t\xi = 0, there exists a homomorphism u:A/tA→K/Au : A/tA \to K/A such that u(ε)=ξu(\varepsilon) = \xi. It follows that D(M)D(M) is isomorphic to t−1A/At^{-1}A/A, and since the homothety by tt is a bijection on K, we also have D(M)D(M) isomorphic to A/tAA/tA, which proves the first assertion. This proves that MM and D(D(M))D(D(M)) are isomorphic, so have the same length ; on the other hand cMc_M is injective, for if y∈Ay \in A is such that the relation tz∈Atz \in A (for z∈Kz \in K) implies yz∈Ayz \in A, then taking z=t−1z = t^{-1} we have y∈tAy \in tA. It follows that the image cM(M)c_M(M) is necessarily equal to D(D(M))D(D(M)).

Corollary

— Let M, N be two A-modules of finite length, and let φ\varphi be an A-bilinear map from M×NM \times N into K/AK/A, such that : 1) the relation φ(x,y)=0\varphi(x, y) = 0 for all y∈Ny \in N implies x=0x = 0 ; and 2) the relation φ(x,y)=0\varphi(x, y) = 0 for all x∈Mx \in M implies y=0y = 0. Then the A-linear maps sφ:M→D(N)s_\varphi : M \to D(N) and dφ:N→D(M)d_\varphi : N \to D(M) associated to φ\varphi are isomorphisms. Indeed the hypotheses on φ\varphi mean that sφs_\varphi and dφd_\varphi are injective and since long(D(N))=long(N)\mathrm{long}(D(N)) = \mathrm{long}(N) and long(D(M))=long(M)\mathrm{long}(D(M)) = \mathrm{long}(M) by Prop. 10, this implies that long(M)=long(N)\mathrm{long}(M) = \mathrm{long}(N), and consequently sφs_\varphi and dφd_\varphi are bijective.

Proposition 11

— If M′→uM→vM′′M' \xrightarrow{u} M \xrightarrow{v} M'' is an exact sequence of A-modules of finite length, then the sequence D(M′′)→D(v)D(M)→D(u)D(M′)D(M'') \xrightarrow{D(v)} D(M) \xrightarrow{D(u)} D(M') is exact [^1].

Let us show first of all that given an exact sequence

(1)

0→M′→M→M′′→0 0 \to M' \to M \to M'' \to 0

the corresponding sequence

0→D(M′′)→D(M)→D(M′)→0 0 \to D(M'') \to D(M) \to D(M') \to 0

is exact; indeed we know that the sequence

0→D(M′′)→D(M)→D(M′) 0 \to D(M'') \to D(M) \to D(M')

is exact (II, p. 227, Th. 1); on the other hand, it follows from (1) that

long⁡(M)=long⁡(M′)+long⁡(M′′) \operatorname{long}(M) = \operatorname{long}(M') + \operatorname{long}(M'')

(II, p. 212, Prop. 16); by Prop. 10 we thus have

long⁡(D(M))=long⁡(D(M′))+long⁡(D(M′′)) \operatorname{long}(D(M)) = \operatorname{long}(D(M')) + \operatorname{long}(D(M''))

in other words long⁡(D(M′))=long⁡(D(M)/D(M′′))\operatorname{long}(D(M')) = \operatorname{long}(D(M)/D(M'')). Since D(M)/D(M′′)D(M)/D(M'') is naturally identified with a submodule of D(M′)D(M'), it must be equal to D(M′)D(M'), which proves our assertion.

This implies immediately that if u:M′→Mu : M' \to M is injective, then D(u):D(M)→D(M′)D(u) : D(M) \to D(M') is surjective; the conclusion then follows from II, p. 199, remark 4.

For every A-module M, let S(M)\mathfrak{S}(M) denote the set of submodules of M. For every submodule N of M (resp. every submodule N' of D(M)), let N0N^0 (resp. N′0{N'}^0) denote the submodule of D(M)D(M) (resp. M) consisting of those x′∈D(M)x' \in D(M) (resp. x∈Mx \in M) such that (y,x′)=0(y, x') = 0 for all y∈Ny \in N (resp. (x,y′)=0(x, y') = 0 for all y′∈N′y' \in N').

Proposition 12

Let M be an A-module of finite length. Then the map which sends every submodule N of M to N0N^0 is a bijection from S(M)\mathfrak{S}(M) onto S(D(M))\mathfrak{S}(D(M)), and the inverse bijection sends every submodule N' of D(M) to the submodule N′0{N'}^0 of M; the module D(N)D(N) is naturally identified with D(M)/N0D(M)/N^0 and D(M/N)D(M/N) with N0N^0. Moreover, we have

(2)

(N1+N2)0=N10∩N20,(N1∩N2)0=N10+N20 (N_1 + N_2)^0 = N_1^0 \cap N_2^0, \quad (N_1 \cap N_2)^0 = N_1^0 + N_2^0

for all submodules N1,N2N_1, N_2 of M.

For each submodule N of M, there is an exact sequence

0→N→M→M/N→0 0 \to N \to M \to M/N \to 0

and hence (Prop. 11) an exact sequence

0→D(M/N)→D(M)→D(N)→0 0 \to D(M/N) \to D(M) \to D(N) \to 0

and since the image of D(M/N)D(M/N) in D(M)D(M) is obviously N0N^0, we see (Prop. 10) that long⁡(N0)=long⁡(M)−long⁡(N)\operatorname{long}(N^0) = \operatorname{long}(M) - \operatorname{long}(N); since MM is identified with D(D(M))D(D(M)) by Prop. 10, we have similarly

long⁡(N00)=long⁡(M)−long⁡(N0)=long⁡(N); \operatorname{long}(N^{00}) = \operatorname{long}(M) - \operatorname{long}(N^0) = \operatorname{long}(N);

in addition it is clear that N⊂N00N \subset N^{00}, so N00=NN^{00} = N. Furthermore the first relation in (2) is obvious, and by applying it to the submodules N10N_1^0 and N20N_2^0 of D(M)D(M), we have (N10+N20)0=N1∩N2(N_1^0 + N_2^0)^0 = N_1 \cap N_2, so N10+N20=(N10+N20)00=(N1∩N2)0N_1^0 + N_2^0 = (N_1^0 + N_2^0)^{00} = (N_1 \cap N_2)^0. This completes the proof of the proposition.

Example 1

For A=ZA = \mathbf{Z}, the Z\mathbf{Z}-modules of finite length are precisely the finite abelian groups; then K=QK = \mathbf{Q}, so K/A=Q/ZK/A = \mathbf{Q}/\mathbf{Z}. Then to define D(M)D(M), we sometimes take, instead of Q/Z\mathbf{Q}/\mathbf{Z}, a Z\mathbf{Z}-module isomorphic to it, such as (V, p. 79, Prop. 2) the group RR of roots of unity (under multiplication) in an algebraically closed field of characteristic 0 ; we then put D(M)=Hom⁡Z(M,R)D(M) = \operatorname{Hom}_{\mathbf{Z}}(M, R). We leave the reader to rewrite the preceeding results for this special case in the corresponding notation.

Example 2

Let aa be a nonzero element of AA. The map x↦x/ax \mapsto x/a from AA into KK induces an isomorphism on quotient modules from A/(a)A/(a) onto the submodule (K/A)(a)(K/A)(a) of K/AK/A consisting of elements annihilated by aa. If MM is an AA-module annihilated by aa, or equivalently an A/(a)A/(a)-module, then the AA-module D(M)D(M) is identified with Hom⁡A/(a)(M,A/(a))\operatorname{Hom}_{A/(a)}(M, A/(a)). We leave the reader to rewrite the preceding results for this special case in the corresponding notation (cf. V, p. 86).

Exercises

See the exercises for § 4.

[^1]: We will see later (A, X, p. 18) that the A-module K/A is injective. It follows that Prop. 11 remains valid for arbitrary A-modules M, M' and M".

Transcribed from the printed edition, 2003, Springer. Not translated.