MODULES OVER A PID

KEITH CONRAD

Every vector space over a field K that has a finite spanning set has a finite basis: it is
isomorphic to K n for some n ≥ 0. When we replace the scalar field K with a commutative
ring A, it is no longer true that every A-module with a finite generating set has a basis: not
all modules have bases. But when A is a PID, we get something nearly as good as that:
(1) Every submodule of An has a basis of size at most n.
(2) Every finitely generated torsion-free A-module M has a finite basis: M ∼ = An for a
unique n ≥ 0.
(3) Every finitely generated A-module M is isomorphic to Ad ⊕ T , where d ≥ 0 and T
is a finitely generated torsion module.
We will prove this based on how a submodule of a finite free module over a PID sits inside
the free module. Then we’ll learn how to count with ideals in place of positive integers.

1. Preliminary results
We start with two lemmas that have nothing to do with PIDs.
Lemma 1.1. If A is a nonzero commutative ring and Am ∼
= An as A-modules then m = n.
Proof. The simplest proof uses a maximal ideal m in A. Setting M = Am and N = An , if
M∼ = N as A-modules then it restricts to an isomorphism mM ∼ = mN and we get an induced
∼ m ∼ n
isomorphism M/mM = N/mN . This says (A/m) = (A/m) as A-modules, hence also as
A/m-vector spaces, so m = n from the well-definedness of dimension for vector spaces. 
Lemma 1.1 says all bases in a finite free module over a nonzero commutative ring have the
same size, and we call the size of that basis the rank of the free module. So the term rank
means dimension when the ring is a field.1 The proof of Lemma 1.1 is valid for free modules
with an infinite basis, so the rank of a free module with an infinite basis is well-defined as
a cardinal number. For our purposes, we only care about free modules of finite rank.
Commutativity in Lemma 1.1 matters: for some noncommutative A, A2 ∼ = A as left
A-modules.
Lemma 1.2. A finitely generated torsion-free module M over an integral domain A embeds
into a finite free A-module. More precisely, if M 6= 0, there is an embedding M ,→ Ad for
some d ≥ 1 such that the image of M intersects each standard coordinate axis of Ad .
Proof. Let K be the fraction field of A and x1 , . . . , xn be a generating set for M as an
A-module. We will show n is an upper bound on the size of each A-linearly independent
subset of M . Let f : An → M be the linear map where f (ei ) = xi for all i. (By e1 , . . . , en
we mean the standard basis of An .) LetPy1 , . . . , yk be linearly independent in M , so their
A-span is isomorphic to A . Write yj = ni=1 aij xi with aij ∈ A. We pull the yj ’s back to
k

1The word “rank” means something completely different in linear algebra – the dimension of the image
of a linear map.
1
2 KEITH CONRAD

An by setting vj = (a1j , . . . , anj ), so f (vj ) = yj . A linear dependence relation on the vj ’s

is transformed by f into a linear dependence relation on the yj ’s, which is a trivial relation
by their linear independence. Therefore v1 , . . . , vk is A-linearly independent in An , hence
K-linearly independent in K n . By linear algebra over fields, k ≤ n.
From the bound k ≤ n, there is a linearly independent subset of M with maximal size,
say t1 , . . . , td . Then dj=1 Atj ∼
= Ad . We will find a scalar multiple of M inside of this. For
P

each x ∈ M , the set {x, t1 , . . . , td } is linearly dependent by maximality of d, so there is a

nontrivial linear relation ax + di=1 ai ti = 0, necessarily with a 6= 0. Thus ax ∈ dj=1 Atj .
P P

Letting x run through the spanning set x1 , . . . , xn there is an a ∈ A − {0} such that
axi ∈ dj=1 Atj for all i, so aM ⊂ dj=1 Atj . Multiplying by a is an isomorphism of M
P P
with aM , so we have the sequence of A-linear maps
d
X
M → aM ,→ Atj → Ad ,
j=1

where the last map is an isomorphism. 

2. Submodules of a finite free module

First we will show submodules of a finite free module over a PID are finitely generated,
with a natural upper bound on the number of generators.
Theorem 2.1. When A is a PID, each submodule of a free A-module of rank n is finitely
generated with at most n generators.
Proof. A free A-module of rank n is isomorphic to An , so we may assume the free A-module
is An . We will induct on n. The case n = 0 is trivial and the case n = 1 holds since A is a
PID: an A-submodule of A is an ideal, hence Aa since all ideals in A are principal.
Suppose n > 1 and the theorem is proved for all submodules of An . Let M ⊂ An+1 be
a submodule. We want to show M has at most n + 1 generators. Write An+1 = An ⊕ A
and let π : An ⊕ A  An be projection to the first component of the direct sum. We look
at the image and kernel of π|M , the restriction of π to M . Its image π(M ) is a submodule
of An , so π(M ) has at most n generators by the inductive hypothesis: π(M ) = ki=1 Ayi
P
where y1 , . . . , yk ∈ An and k ≤ n. We can write yi = π(xi ) for some xi ∈ M , so π(M ) =
Pk ∼
i=1 Aπ(xi ). And ker(π|M ) = M ∩ (0 ⊕ A), with 0 ⊕ A = A as A-modules. Since A is a
PID, A-submodules of A have a single generator, so ker(π|M ) = Ax0 for some x0 ∈ M .
We will show M = ki=0 Axi , so M has at most k + 1 generators and k + 1 6 n + 1. The
P

containment ki=0 Axi ⊂ M is clear. For the reverse containment, pick an arbitrary x ∈ M
P
and the previous paragraph tells us π(x) = a1 π(x1 ) + · · · + ak π(xk ) = π(a1 x1 + · · · + ak xk )
for some a1 , . . . , ak in A. Therefore x − ki=1 ai xi ∈ ker(π|M ), so x − ki=1 ai xi = a0 x0 for
P P

some a0 ∈ A. Thus x = a0 x0 + a1 x1 + · · · + ak xk ∈ ki=0 Axi , so M ⊂ ki=0 Axi .

P P

Let’s refine the previous theorem by showing a submodule of a finite free A-module is
free (has a basis). The proof will be like the one above, but will not logically depend on it.
Theorem 2.2. When A is a PID, each submodule of a free A-module of rank n is free of
rank 6 n.
Proof. As before, since a free A-module of rank n is isomorphic to An , we can assume the
free A-module we use is An . We’ll induct on n.
 MODULES OVER A PID 3

The case n = 0 is trivial and the case n = 1 follows from all submodules of A being (0)
or principal with a nonzero generator, and Aa ∼ = A as A-modules when a is nonzero in A
since A is an integral domain.
Suppose n > 1 and the theorem is proved for all submodules of An . For a submodule M of
An+1 , to show M is free of rank at most n + 1 write An+1 = An ⊕ A and let π : An ⊕ A  An
be projection to the first component. Since π(M ) is a submodule of An , π(M ) is free of
rank 6 n by the inductive hypothesis.
Case 1: π(M ) = 0. Here M ⊂ 0 ⊕ A, so M is free of rank at most 1 since A is a PID.
Case 2: π(M ) 6= 0. Here π(M ) is a nonzero submodule of An , so π(M ) is free of positive
rank d ≤ n. Write a basis of π(M ) as π(e1 ), . . . , π(ed ) where ei ∈ M : π(M ) = di=1 Aπ(ei ).
L
The
P elements e1 , . . . , ed in M are linearly
P independent since their images under π are: if
ai ei = 0 in M then apply π to get ai π(ei ) = 0 in π(M ), so all ai are 0.
For m ∈ M , π(m) = di=1 ai π(ei ) for unique a1 , . . . , ad ∈ A. Then π(m − di=1 ai ei ) = 0,
P P

so m − di=1 ai ei ∈ ker(π|M ). We get inverse maps

P
(
m 7−→ (a1 , . . . , ad , m − di=1 ai ei )
P
d
M ←→ A ⊕ ker(π|M ) by Pd
i=1 ai ei + k ←−[ (a1 , . . . , ad , k)

and both maps are linear (check!). So M ∼ = Ad ⊕ ker(π|M ) as A-modules. Thus M is free
of rank d or d + 1 (depending on ker(π|M ) being {0} or not), and d + 1 ≤ n + 1. 
Theorem 2.2 is always false if A is not a PID, even for the A-module A itself.
Non-example 2.3. If A is not a PID then either it is not an integral domain or it has a
nonprincipal ideal. If A is not an integral domain then we have xy = 0 for some nonzero
x and y in A, and the principal ideal Ax is not a free A-module. If A has a nonprincipal
ideal then that ideal is not a free A-module.
Remark 2.4. Theorem 2.2 holds for infinite-rank free modules too: submodules of all free
modules over a PID are free. The proof for infinite bases uses Zorn’s lemma [5, pp. 650–651].
Remark 2.5. Unlike vector spaces, a spanning set of a finite free A-module might not
contain a basis: Z is spanned by {2, 3} as a Z-module, but {2} and {3} are not Z-bases.2
An analogue of that in Z2 is in Example 5.25.
Corollary 2.6. When A is a PID, every finitely generated torsion-free A-module is a finite
free A-module.
Proof. By Lemma 1.2, such a module embeds into a finite free A-module, so it is finite free
too by Theorem 2.2. 
The term “free” in “torsion-free module” and “free module” means different things: a
torsion-free module has no nonzero torsion elements (all elements have annihilator ideal (0)
aside from the element 0), while a free module has a basis. So Corollary 2.6 is saying a
finitely generated module over a PID that has no torsion elements admits a basis. Corollary
2.6 is false without the finite generatedness hypothesis. For example, Q is a torsion-free
abelian group but it has no basis over Z: every (nonzero) free Z-module has proper Z-
submodules (that is, proper subgroups) of finite index while Q does not.
2 If you know about DVRs (a special type of PID): every minimal spanning set of a finite free module
over a DVR is a basis. The proof uses Nakayama’s lemma.
4 KEITH CONRAD

Corollary 2.7. Let A be a PID. If we have a tower of A-modules M ⊃ M 0 ⊃ M 00 with

M∼ = An and M 00 ∼= An then M 0 ∼
= An .
Proof. Since M is free of rank n and M 0 is a submodule, Theorem 2.2 tells us M 0 ∼
= Am
with m ≤ n. Using Theorem 2.2 on M 00 as a submodule of M 0 , M 00 ∼
= Ak with k ≤ m. By
00 ∼ n
hypothesis M = A , so k = n by Lemma 1.1. Thus m = n. 
Corollaries 2.6 and 2.7 are both generally false when A is not a PID.
√
Non-example 2.8. Let A = Z[ −5] and consider the tower of ideals
√ √ √
3Z[ −5] ⊂ (3, 1 + −5) ⊂ Z[ −5].
The bottom √ and top are principal ideals, so they are free A-modules of rank 1. The middle
ideal (3, 1 + −5) is finitely generated and torsion-free, but is not principal and therefore
is not a free A-module. (A nonzero ideal is a free module only when it is principal, since
all pairs of elements in an ideal are linearly related.)
Corollary 2.9. When A is a PID, M is a finite free A-module, and N is a submodule of
M such that M/N is torsion-free, each A-basis of N can be extended to an A-basis of M .
Proof. The result is obvious if N = 0 or N = M , so we can assume N 6= 0 and N 6= M .
Step 1: M/N is a finite free A-module.
Since M is a finitely generated A-module, so is M/N . Since M/N is a torsion-free
A-module by hypothesis, M/N is a finite free A-module by Corollary 2.6.
Step 2: each basis of N extends to a basis of M .
Let e1 , . . . , er be a basis of N (the basis is nonempty since N 6= 0). Since M/N is finite
free by Step 1, it has an A-basis, say er+1 , . . . , ed (this basis is nonempty since M/N 6= 0).
We will show e1 , . . . , ed is a basis of M .
First we show e1 , . . . , ed spans M as an A-module. For m ∈ M , since m ∈ M/N we have
m = ar+1 er+1 + · · · + ad ed
for some ar+1 , . . . , ad ∈ A. Then m − ar+1 er+1 − · · · − ad ed ∈ N , so we can write
m − ar+1 er+1 − · · · − ad ed = a1 e1 + · · · + ar er
for some a1 , . . . , ar ∈ A. Thus
m = a1 e1 + · · · + ar er + ar+1 er+1 + · · · + ad ed .
Next we show e1 , . . . , ed is A-linearly independent in M . If
(2.1) a1 e1 + · · · + ad ed = 0
then reduce both sides modulo N to get
ar+1 er+1 + · · · + ad ed = 0,
so by the basis property in M/N we get ai = 0 for i = r + 1, . . . , d. Feeding this into (2.1),
a1 e1 + · · · + ar er = 0
in N , so ai = 0 for i = 1, . . . , d. 
√
4
√
Example 2.10. Let M = Z[ 2] and N = Z[ 2], with A = Z. Then
√
4
√ √
4
√ √
4
M = Z + Z 2 + Z 2 + Z 2 2 = N ⊕ N 2,
√ √
M/N ∼
so √ =N 2∼
4
= Z2 is torsion-free. Thus each Z-basis of Z[ 2] extends to a Z-basis of
4
Z[ 2].
 MODULES OVER A PID 5
√
4
√ √
Non-example√ 2.11. Let M = Z[ 2] and N = Z[3 2], with A = Z. Then 2 6= 0 in
M/N√but 3 2 = 0 in M/N , so M/N is not torsion-free. It can be shown that the Z-basis
{1, 3 2} of N is not part of a Z-basis of M (see Exercise 3).

There is a convenient way of picturing a submodule of a finite free module over a PID:
bases can be chosen for the module and submodule that are aligned nicely, as follows.

Definition 2.12. If A is a PID, M is a finite free A-module, and M 0 is a submodule of M ,

then a basis {v1 , . . . , vn } of M and a basis {a1 v1 , . . . , am vm } of M 0 with ai ∈ A − {0} and
m ≤ n is called a pair of aligned bases.

Pictures will show how aligned bases look before seeing the main theorem about them.

Example 2.13. Let A = Z, M = Z[i] and take M 0 = (1 + 2i)Z[i]. So M = Z + Zi and

M 0 = (1 + 2i)Z[i] = (1 + 2i)Z + (1 + 2i)Zi = Z(1 + 2i) + Z(−2 + i).

The obvious Z-bases for M and M 0 are {1, i} and {1 + 2i, −2 + i}. In Figure 1, we shade
a box having each basis as a pair of edges and translate each box across the plane. The
modules M and M 0 are the intersection points of the networks of lines formed by the small
and large boxes, respectively. The lines are only a visual aid, showing how a choice of basis
gives a specific way to generate the module by the basis.

Z[i] = Z + Zi, (1 + 2i) = Z(1 + 2i) + Z(−2 + i)

Figure 1. Nonaligned bases for a modules and submodule

6 KEITH CONRAD

To see a completely different picture of the same two modules M and M 0 , we use new
bases: {1 + 2i, i} for M and {1 + 2i, 5i} for M 0 . These are bases because of the relations
         
1 + 2i 1 2 1 1 + 2i 1 0 1 + 2i
= , = ,
i 0 1 i 5i 2 1 −2 + i
where the two matrices are integral with determinant 1, so the elements of Z[i] in the
vector components on both sides have the same Z-span. These new bases lead to Figure 2,
where the parallelograms with each basis as a pair of edges is shaded and looks quite unlike
the shaded boxes of Figure 1. Translating the parallelograms across the plane produces
two new networks of lines (both sharing all the vertical lines) The intersection points are
the same as before; make sure you can see the vertices of the large box from Figure 1 as
intersection points of lines in Figure 2. In Figure 2 five of the parallelograms for M fill up
of the parallelograms for M 0 . These bases are aligned.

Z[i] = Z(1 + 2i) + Zi, (1 + 2i) = Z(1 + 2i) + Z · 5i

Figure 2. Aligned bases for a module and submodule

Theorem 2.14. Each finite free A-module M of rank n > 1 and nonzero submodule M 0
of rank m 6 n admit a pair of aligned bases: there is a basis v1 , . . . , vn of M and nonzero
a1 , . . . , am ∈ A such that
n
M m
M
0
M= Avi and M = Aaj vj .
i=1 j=1

We can also arrange that a1 | a2 | · · · | am .

 MODULES OVER A PID 7

The condition a1 | a2 | · · · | am at the end of the theorem plays an important role in

the proof of the theorem, but it is not used in any applications presented later except for
Theorem 3.6, so you can largely ignore it.
Proof. Our argument is based on [6, Sect. 1.5]. An A-basis of M gives us coordinate func-
tions for that basis, which are A-linear maps M → A: for an A-basis v1 , . . . , vn of M , its dual
basis is the A-linear maps vj∨ : M → A where vj∨ (x1 v1 + · · · + xd vd ) = xj for x1 , . . . , xd ∈ A.
Bases for M and M 0 as in the theorem are “aligned” coordinate systems on M and M 0 .
Motivation. To explain the main idea in the proof, we’ll first suppose the conclusion of
the theorem is true and derive an important consequence about the images of all A-linear
maps ϕ : M → A (the dual module of M ) when they are restricted to M 0 : there is a nonzero
A-linear map ψ : M → A such that
ϕ(M 0 ) ⊂ ψ(M 0 ) for all ϕ.
To show this, by the formula for M0
in the conclusion of the theorem we get for all ϕ that
Xm  X m
ϕ(M 0 ) = ϕ Aaj vj = Aaj ϕ(vj ) ∈ Aa1 + · · · + Aam .
j=1 j=1

Since A is a PID, Aa1 + · · · + Aam = (a) for some nonzero a ∈ A, so ϕ(M 0 ) ⊂ (a) for all ϕ.3
Write a = c1 a1 + · · · + cm am with cj ∈ A (the
P choice∨ of cj ’s may not be unique, but
fix such a representation for a) and define ψ = m j=1 cj ej as a linear map M → A. Then
Pm Pm Pm
ψ( j=1 aj ej ) = j=1 cj aj = a and j=1 aj ej ∈ M 0 , so
(a) ⊂ ψ(M 0 ) ⊂ (a).
Thus ψ(M 0 ) = (a) and this ideal is the unique maximal member with respect to inclusion
among all ideals ϕ(M 0 ) as ϕ varies over all linear maps M → A. (This is not saying (a) is
a maximal ideal!). This ends the motivation.
Step 1. The set of ideals S := {ϕ(M 0 ) : ϕ : M → A is linear} is not {(0)} and has a
maximal member (a).
Let {v1 , . . . , vn } be a basis of M . Since M 0 6= {0}, at least one coordinate function vj∨
for this basis of M is not identically 0 on M 0 , which makes vj∨ (M 0 ) a nonzero ideal in S.
Each nonzero ideal of A is contained in only finitely many ideals of A since A is a PID: if
(x) is a nonzero ideal then we have (x) ⊂ (y) if and only if y | x. Up to unit multiples, there
are only finitely many possible y since x has only finitely many factors up to unit multiples.
Applying this to (x) = ϕ(M 0 ) for some ϕ where ϕ(M 0 ) 6= (0) and looking at the finitely
many ideals containing (x), S contains a maximal member with respect to inclusion, say
(a) = ψ(M 0 ).
Then a 6= 0, and maximality means that if (a) ⊂ ϕ(M 0 ) for some ϕ, then ϕ(M 0 ) = (a).4
Since a ∈ ψ(M 0 ), there’s some v 0 ∈ M 0 such that a = ψ(v 0 ) .
Step 2: For the ideal (a) in Step 1 and each A-linear map ϕ : M → A, a | ϕ(v 0 ) in A.
Write the ideal (a, ϕ(v 0 )) as (b), where b ∈ A − {0}. For some x and y in A,
b = ax + ϕ(v 0 )y = xψ(v 0 ) + yϕ(v 0 ) = (xψ + yϕ)(v 0 ).
3If a | a | · · · | a , which we don’t want to assume, then we can let a = a since (a ) ⊂ · · · ⊂ (a ) ⊂ (a ).
1 2 m 1 m 2 1
4We anticipate (a) will be the unique maximal member of S by the argument in the motivational section
at the start of this proof, but at the moment (a) is just some maximal member of S.
8 KEITH CONRAD

Since xψ + yϕ is a linear map M → A,

(a) ⊂ (a, ϕ(v 0 )) = (b) ⊂ (xψ + yϕ)(M 0 ),
so the maximality of (a) in Step 1 implies (a) = (b) = (a, ϕ(v 0 )). Therefore a | ϕ(v 0 ).
Step 3: There is an e1 ∈ M such that ψ(e1 ) = 1.
Write v 0 in the basis {v1 , . . . , vn } as v 0 = ni=1 ci vi . Then ci = vi∨ (v 0 ) ∈ (a) by Step 2, so
P
ci = abi where bi ∈ A. Thus
v 0 = a(b1 v1 + · · · + bn vn ).
Set e1 := b1 v1 + · · · + bn vn in M , so a = ψ(v 0 ) = aψ(e1 ) in A. Since a 6= 0, ψ(e1 ) = 1.
Step 4: We have direct sum decompositions
M = Ae1 ⊕ ker ψ, M 0 = Aae1 ⊕ (M 0 ∩ ker ψ).
First we show M = Ae1 + ker ψ. For each v ∈ M , ψ(v − ψ(v)e1 ) = ψ(v) − ψ(v)ψ(e1 ) =
ψ(v) − ψ(v) = 0, so v − ψ(v)e1 ∈ ker ψ. Therefore M ⊂ Ae1 + ker ψ and the reverse
containment is obvious.
The sum in Ae1 + ker ψ is direct since if v ∈ Ae1 ∩ ker ψ then v = αe1 and ψ(v) =
αψ(e1 ) = α, so if ψ(v) = 0 then α = 0.
Next we show M 0 = Aae1 + (M 0 ∩ ker ψ). For each v ∈ M 0 , ψ(v − ψ(v)e1 ) = 0 as before,
and ψ(v) ∈ ψ(M 0 ) = (a), so M ⊂ Aae1 + ker ψ and the reverse containment is obvious.
This sum is direct for the same reason as above. This completes Step 4.
Step 5: Extend the direct sum decompositions in Step 4 to bases of M and M 0 .
If M 0 ∩ ker ψ = {0} then M 0 = Aae1 , so m = 1 and the theorem is proved in this case by
using any basis of ker ψ as e2 , . . . , en .
Now suppose M 0 ∩ ker ψ 6= {0}. Then ker ψ is free of rank n − 1 and M 0 ∩ ker ψ is free
of rank m − 1 ≥ 1 by the direct sum decompositions of M and M 0 in Step 4.
Rename a as a1 . By induction on n (and m), there is a basis e2 , . . . , en of ker ψ and
nonzero a2 , . . . , am in A such that a2 e2 , . . . , am em is a basis of M 0 ∩ ker ψ. Then the direct
sums in Step 4 take the form
M = Ae1 ⊕ (Ae2 ⊕ · · · ⊕ Aen ), M 0 = Aa1 e1 ⊕ (Aa2 e2 ⊕ · · · ⊕ Aam em ).
If we are not interested in a relation like a1 | a2 | · · · | am then we are done with the
proof. If instead we want to prove a1 | a2 | · · · | am then we can bring in a2 | · · · | am by the
inductive hypothesis. Why does a1 | a2 ? Do we need to go back and prove stronger forms of
earlier steps? That isn’t necessary. Consider how ϕ = e∨ ∨ 0
1 + e2 behaves on v = ae1 = a1 e1 :
a1 = a = ϕ(v ) ∈ ϕ(M ), so (a) ⊂ ϕ(M ). By the maximality of (a) in Step 1, (a) = ϕ(M 0 ).
0 0 0

Then a2 = ϕ(a2 e2 ) ∈ ϕ(M 0 ) = (a) = (a1 ), so a1 | a2 . 

In this theorem, an aligned basis for M and M 0 depends on both modules: there need
not be an aligned basis that uses an arbitrary basis for M or an arbitrary basis for M 0 .
Example 2.15. Let M = Z2 and M 0 = { xy ∈ Z2 : x ≡ y mod p} = Z 11 + Z p0 for

prime p. An example of aligned Z-bases here is { 11 , 10 } for M and { 11 , p 10 } for M 0 ,

so [M :
M 0 ] = p. (Another way to see [M : M 0 ] = p is that the mapping M → Z/(p)

where xy 7→ x − y mod p is additive and surjective with kernel M 0 , so M/M 0 ∼
= Z/(p).) A
0
general pair of aligned bases for M and M has the form {v1 , v2 } for M and {a1 v1 , a2 v2 }
for M 0 , so M/M 0 ∼
= Z/(a1 ) × Z/(a2 ) as abelian groups. Then [M : M 0 ] = p = |a1 a2 |, so
{a1 , a2 } = {±1, ±p}. That means v1 or v2 (depending on whether a1 or a2 is ±1) has to
 MODULES OVER A PID 9

be in M 0 , so the basis { 10 , 01 } of M isn’t part of an aligned basis with M 0 since neither

element of that basis is in M 0 . Conversely,
a vector 2

in Z 0 multiplied by p has both of its
1 p+1
coordinates divisible by p, so the basis { 1 , 1 } of M is not part of an aligned basis
with M .
Corollary 2.16. Let A be a PID, M be finite free A-module, and M 0 be a submodule of
M . Then M and M 0 have the same rank if and only if M/M 0 is a torsion module.
Proof. Let M have rank n. The module M 0 is free of some rank m ≤ n. Using aligned
bases for M and M 0 , we can write
n
M m
M
0
M= Avi and M = Aaj vj
i=1 j=1

with nonzero aj ’s. Then M/M 0 ∼

Lm
= j=1 A/(aj ) ⊕ An−m . This is a torsion module if and
only if m = n. 

3. Applications to matrix groups

In this section we give applications of Corollary 2.7 and Theorem 2.14 to matrix groups.
A matrix in GLn (Q) that has finite order does not have to have entries in Z. For example,
−1 x
( 0 1 ) has order 2 for all x ∈ Q. But this matrix is conjugate to a matrix with entries in Z:
( 10 −x/2
1
)( −10 x1 )( 10 −x/2
1
)−1 = ( −10 01 ). More generally, it turns out that every finite subgroup
of GLn (Q) is conjugate to a subgroup with entries in Z. This is a special case of a result
we prove below about subgroups of GLn (K) where K is the fraction field of a PID.
Definition 3.1. Let A be a PID and K be its fraction field. A subset S of K has a common
denominator in A when there is a nonzero a ∈ A such that aS ⊂ A.
Example 3.2. Every finite subset S of K has a common denominator in A: use the product
of the denominators of the elements of S when they are written as ratios of numbers in A.
Pn
Example 3.3. Let S = j=1 Axi be a finitely generated A-submodule of K. It has a
common denominator in A, such as a common denominator of the numbers x1 , . . . , xn .
Non-example 3.4. In Q, the set of reciprocal powers {1/2, 1/4, 1/8, . . . , 1/2n , . . .} has no
common denominator in Z.
Theorem 3.5. Let A be a PID and K be its fraction field. Every subgroup G of GLn (K)
whose matrix entries have a common denominator in A is conjugate to a subgroup with
matrix entries in A: there is a matrix L ∈ GLn (K) such that all matrices in LGL−1 have
entries in A. In particular, every finite subgroup of GLn (K) is conjugate to a subgroup with
matrix entries in A.
The common denominator hypothesis of this theorem means for some nonzero a ∈ A, ag
has matrix entries in A for all g ∈ G.

Proof. For each g ∈ G, g(An ) is a finite free A-submodule of K n . Let M be the A-module
n n n
P
g∈G g(A ), which is the set of finite sums of vectors in K that belong to g(A ) for some
g ∈ G. This is an A-module in K n that contains An (use g = In ). Note M is G-stable
(carried back to itself when acting on it by elements of G).
10 KEITH CONRAD

The common denominator hypothesis says there is a nonzeroPa ∈ A such that each
matrix ag for g ∈ G has all of its entries in A. Therefore aM = g∈G (ag)(An ) ⊂ An , so
M ⊂ (1/a)An in K n . Thus An ⊂ M ⊂ (1/a)An . By Corollary 2.7, M ∼ = An as A-modules.
Let ϕ : An → M be an A-module isomorphism. Then M = ϕ(An ) is the A-linear span
of ϕ(e1 ), . . . , ϕ(en ). These vectors are A-linearly independent since they span a free A-
module of rank n, so they are also K-linearly independent: a nontrivial K-linear relation
would become a nontrivial A-linear relation. Therefore the matrix Φ = [ϕ(e1 ) · · · ϕ(en )] is
invertible, so Φ ∈ GLn (K) and Φ = ϕ on An .
For each g ∈ G, gM ⊂ M since M is G-stable, so gϕ(An ) ∈ ϕ(An ). Therefore g(Φ(An )) ⊂
Φ(An ), so Φ−1 gΦ(An ) ⊂ An . This means the matrix Φ−1 gΦ has entries in A (its columns
are (Φ−1 gΦ)(ei ) for i = 1, . . . , n). Letting g run over G, the group Φ−1 GΦ is conjugate to
G and its elements have matrix entries in A. 

Our second application to matrix groups will be the computation of normalizers. For a
group G and subgroup H of G, the normalizer of H is NG (H) = {g ∈ G : gHg −1 = H} =
{g ∈ G : g −1 Hg = H}: this is the largest subgroup of G in which H is a normal subgroup.
Theorem 3.6. Let A be a PID and K be its fraction field. In the group GLn (K), both
GLn (A) and SLn (A) have normalizer K × GLn (A).
We view K × inside GLn (K) as the nonzero scalar diagonal matrices {cIn : c ∈ K × }.
This is the center of GLn (K).

Proof. The result is trivial when n = 1, since GL1 (K) = K × is commutative, so from now
on we can assume n ≥ 2. Set G = GLn (K). We will show NG (GLn (A)) = K × GLn (A) and
then indicate how that argument can be modified to show NG (SLn (A)) = K × GLn (A).
That K × GLn (A) ⊂ NG (GLn (A)) is obvious. To prove NG (GLn (A)) ⊂ K × GLn (A), pick
g ∈ NG (GLn (A)). We will find c ∈ K × such that cg ∈ GLn (A). The proof is based on an
answer to the Mathoverflow question https://mathoverflow.net/questions/80667. Set
M = g(An ).
Step 1: M is a free A-module of rank n and is GLn (A)-stable: L(M ) ⊂ M for all
L ∈ GLn (A).
The A-module M is spanned by g(e1 ), . . . , g(en ) where e1 , . . . , en is the standard basis
of An , and g(e1 ), . . . , g(en ) is A-linearly independent since g is an invertible matrix over A.
Thus M ∼ = An as A-modules.
To show M is GLn (A)-stable, we use the property g −1 GLn (A)g = GLn (A), which is the
condition of g normalizing GLn (A). For L ∈ GLn (A), L0 := g −1 Lg is in GLn (A) too, so
Lg = gL0 . Therefore
L(M ) = Lg(An ) = gL0 (An ) = g(An ) = M.
Step 2: There is d ∈ A − {0} such that dM ⊂ An .
Let d ∈ A − {0} be a common denominator for the coordinates of g(e1 ), . . . , g(en ) in K n .
Then dg(ei ) ∈ An for i = 1, . . . , n, so dM ⊂ An .
Step 3: There is c ∈ K × such that cg ∈ GLn (A).
By Steps 1 and 2, dM is a submodule of An with rank n, so by Theorem 2.14 we can find
a basis {f1 , . . . , fn } of An and nonzero a1 , . . . , an in A such that {a1 f1 , . . . , an fn } is a basis
of dM and a1 | a2 | · · · | an . We won’t use the full strength of that divisibility condition,
but just a1 | ai for i = 1, . . . , n.
 MODULES OVER A PID 11

In the direct sum decomposition

n
M
dM = Aai fi ,
i=1
divide through by a1 :
n n
d M ai M ai
M= A fi = Af1 ⊕ A fi .
a1 a1 a1
i=1 i=2
Set M0 = (d/a1 )M , so M0 is a submodule of An (each ai /a1 is in A) and M 0 contains the
n
vector f1 that’s part of a basis {f1 , . . . , fn } of A . Since M is GLn (A)-stable by Step 1, so
is M 0 .
The matrix L with columns f1 , . . . , fn (in that order) is in GLn (A) since the columns are
a basis of An , and L(e1 ) = f1 . Then e1 = L−1 (f1 ) ∈ M 0 because L−1 ∈ GLn (A) and M 0
is GLn (A)-stable. By a similar argument, if we permute the columns of L to place f1 in
the jth column for j = 2, . . . , n then we get ej ∈ M 0 . Thus M 0 is a submodule of An that
contains e1 , . . . , en , so M 0 = An . Rewriting that as (d/a1 )g(An ) = An tells us the matrix
(d/a1 )g is in GLn (A), so cg ∈ GLn (A) where c = d/a1 ∈ K × . This completes the proof
that NG (GLn (A)) = K × GLn (A).
To show NG (SLn (A)) = K × GLn (A), we have K × GLn (A) ⊂ NG (SLn (A)) since SLn (A)C
GLn (A). For the reverse containment, if g ∈ NG (SLn (A)) then we want c ∈ K × such that
cg(An ) = An , so cg ∈ GLn (A). The procedure used above for NG (GLn (A)) can be applied to
NG (SLn (A)), with the following changes: M = g(An ) is SLn (A)-stable rather than GLn (A)-
stable in Step 1, and in Step 3 modify the matrix L with columns f1 , . . . , fn so that its last
column is u−1 fn where u = det(f1 · · · fn ) ∈ A× , as that makes det(L) = uu−1 = 1, so
L ∈ SLn (A). 

4. Finitely generated modules

As an application of aligned bases for a submodule of a finite free module, we describe
the structure of every finitely generated module over a PID.
Theorem 4.1. Let A be a PID. Every finitely generated A-module has the form F ⊕ T
whereLF is a finite free A-module and T is a finitely generated torsion A-module. Moreover,
T ∼ m
= j=1 A/(aj ) for some m with nonzero ai .
Proof. Let M be a finitely generated A-module, with generators x1 , . . . , xn . Define f : An 
M by f (ei ) = xi . Then there is a surjective linear map An  M , so M is isomorphic
 to
a quotient A /N . As in the proof of Corollary 2.16, A /N ∼
n n m n−m
L
= j=1 A/(aj ) ⊕ A for
some m ≤ n and nonzero aj ’s. The direct sum of the A/(aj )’s is a torsion module and
An−m is a finite free A-module. 
Example 4.2. Taking A = Z, every finitely generated abelian group is isomorphic to
Zn ⊕ G where n ≥ 0 and G is a finite abelian group (a finitely generated torsion abelian
group must be finite).
Corollary 4.3. Every finitely generated torsion module over a PID A is a direct sum of
cyclic torsion modules: it is isomorphic to A/(a1 )⊕· · ·⊕A/(ak ), where the ai ’s are nonzero.
Some ai ’s might be units, making A/(ai ) = 0.
12 KEITH CONRAD

Proof. A finitely generated module in Theorem 4.1 is a torsion module if and only if F = 0.
The description of T in Theorem 4.1 gives the desired cyclic decomposition for finitely
generated torsion modules. 
Corollary 2.6 could be regarded as a consequence of Theorem 4.1: a finitely generated
module is F ⊕ T where F is finite free and T is torsion, and being torsion-free forces T = 0,
so the module is free.
When a finitely generated A-module is written as F ⊕T , where F is a finite free submodule
and T is a torsion submodule, the choice of F is not unique but T is unique: T is the set
of all elements in F ⊕ T with nonzero annihilator ideal, which is a description that makes
no reference to the direct sum decomposition. The best way to see F is not unique is by
examples.
Example 4.4. Using A = Z, Z×Z/(2) has generating set {(1, 0), (0, 1)} and {(1, 1), (0, 1)}.
Therefore it can be written as F1 ⊕ T1 , where F1 = h(1, 0)i ∼ = Z and T1 = 0 ⊕ Z/(2), and
∼
also as F2 ⊕ T2 where F2 = h(1, 1)i = Z and T2 = 0 ⊕ Z/(2) = T1 .
√ √
Example 4.5. It can be shown that every√unit in Z[ 2] has the form ±(1 + 2)k some
choice of sign ±1 and some ×
√ × integer k, so Z[ 2] √is a finitely generated√ abelian group. The
torsion subgroup of Z[ 2] is {±1}, while 1 + 2 and −(1 + 2) each generate different
free subgroups that complement {±1}:
√ √ √
Z[ 2]× = {±1} × (1 + 2)Z = {±1} × (−(1 + 2))Z
√ ×
This leads to two isomorphisms√ of Z[ 2] with Z×Z/2Z, which √ identify different subgroups
with Z (the powers of 1 + 2 and the powers of −(1 + 2)) but both identify the same
subgroup {±1} with Z/(2).
√ √ × √ Z √ Z √ √ Z
Example 4.6. It can √ be shown
√ that
√ Z[ √2, 3] = ±(1 + 2) (2 +√ 3) ( 2 +
√ 3) ,
where
√ the
√ units 1 + 2, 2 + 3, and 2 + 3 (with respective
√ inverses 2
√ b √ − 1, 2 +√ 3, and
3 − 2) have no multiplicative relations over Z: if (1 + 2)a (2 √ + √ 3) ( 2 + 3)c = 1
for integer exponents a, b, and c, then a = b√ = c√= 0. Therefore√Z[ 2, √3]×√ ∼
= Z3 √ × Z/2Z.
Two √ examples of complements to ±1 in Z[ 2, 3]× are h1 + 2, 2 + 3, 2 + 3i and
√ √ √
h1 + 2, −(2 + 3), −( 2 + 3)i.
While the free part of a direct sum decomposition is not unique, the rank of the free
part is unique: writing M = F ⊕ T with F finite free and T necessarily being the torsion
submodule Mtor of M , we have F ∼ = M/T = M/Mtor so the rank of F is the rank of the
finite free module M/Mtor . The rank of a finitely generated module over a PID A is defined
to be the rank of its free part, which is well-defined even though the free part itself is not.
In down-to-earth terms, the rank is the largest number of linearly independent torsion-free
elements (over Z, it is the largest number of independent elements of infinite order).
Our focus in this section was on describing the abstract structure of a finitely generated
module over a PID, not proving a module over a PID arising somewhere in mathematics is
finitely generated. The fact that certain abelian groups (Z-modules) are finitely generated
can be a major theorem (look up the Mordell–Weil theorem) and a formula for the rank
can be a major theorem or conjecture (look up Dirichlet’s unit theorem or the Birch and
Swinnerton-Dyer conjecture).
Theorem 4.1 says for a finitely generated module M over a PID A, M = Mtor ⊕ F where
F is a free A-module. In particular, Mtor is a direct summand of M . That conclusion can
 MODULES OVER A PID 13

become false if (i) A is a PID and M is not finitely generated or (ii) M is finitely generated
and A is not a PID. We’ll give examples of this using A = Z for (i) and A = Z[x] for (ii).
Q
Example 4.7. Set M = p Z/(p), which is a Z-module in a natural way. We will show
the torsion submodule T of M is not a direct summand: M 6= T ⊕ N for a submodule N .
The argument is Lbased on [4, Theorem 10.2].
Step 1: T = p Z/(p) and T 6= M .
To prove the claim, one containment is immediate: since elements in the direct sum have
only finitely many nonzero coordinates and each Z/(p) is killed off by a single prime p, each
element of the direct sum is killed off by a product of finitely many primes and thus is in
T . Conversely, if m := (ap mod p)p is an element of T and km = 0 where k ∈ Z − {0}, then
kap ≡ 0 mod p for each prime p. When p - k, ap ≡ 0 mod p, so m has only finitely many
possible
L nonzero coordinates (its p-coordinates where p is a prime factor of k). Therefore
m ∈ p Z/(p), which finishes Step 1. Because there are infinitely many primes, T 6= M .
Step 2: We can’t write M = T ⊕ N for a Z-submodule N of M .
Assume there is such a direct sum decomposition. Then N ∼
= M/T , so M has a submodule
isomorphic to M/T . We will get a contradiction by showing M and its submodules all share
a property that is not true
T for M/T .
The property is this: p pM = {0}, where the intersection runs over all primes. Indeed,
for each prime
T p the p-coordinate of an element of pM has to be 0, T so all coordinates of an
element of p pM equal 0. Therefore if N is a submodule of M , p pN = {0}.
T
In contrast to that, we will show p p(M/T ) 6= {0}. Specifically, let v := (1, 1, 1, . . .) be
the element of M with p-coordinate 1 mod p for each prime p. We’ll show
(i) v 6∈ T
T, so v 6= 0 in M/T ,
(ii) v ∈ p p(M/T ).
Proof of (i): If v ∈ T , then kv = 0 for some nonzero integer k since v 6= 0. Looking at
the coordinates of kv, we get k ≡ 0 mod p for each prime p, so k has infinitely many prime
factors and that forces k = 0, a contradiction.
Proof of (ii): Fix a prime p. We will find a w ∈ M and t ∈ T (depending on p). such
that v = pw + t, so v ≡ pw mod T , so v ∈ p(M/T ).
All the coordinates of v are 1. For each prime q other than p, p is invertible mod q so
we can define wq ∈ Z/(q) by the congruence pwq ≡ 1 mod q. Define wp = 0 mod p. Then
w = (wq )q is in M and pw has q-coordinate 1 for each q 6= p and pw has p-coordinate
0. Therefore v = pw + t where t has p-coordinate 1 and every other coordinate 0, which
means t ∈ T . That completes Step 2.

Remark 4.8. This construction of a Z-module whose torsion submodule is not a direct
summand of it carries over without change to modules over a PIDQA that has infinitely
many non-associate irreducible elements. Let M be the A-module (π) A/(π), where the
direct product is taken over all distinct
L maximal ideals (π). It is left to the reader to prove
the torsion submodule T of M is (π) A/(π) and M 6= T ⊕ N for an A-submodule N of M
by an argument similar to the Q case A = Z above.
The infinite direct product p Z/(p) in Example 4.7 may seem artificial. What about
using the abelian group S 1 , whose torsion submodule is the group µ∞ of all roots of unity?
It seems plausible that we can’t write S 1 = µ∞ × H for a subgroup H of S 1 , but that is
possible by the Axiom of Choice since µ∞ is a divisible abelian group. See Corollary 2.5 in
14 KEITH CONRAD

https://kconrad.math.uconn.edu/blurbs/zorn1.pdf.
L Note that unlike µ∞ , the torsion
submodule p Z/(p) in Example 4.7 is not divisible.

Example 4.9. Let A = Z[x] and r be an integer
besides 0 and ±1, so the ideal a := (r, x)
in A is not principal. We will show M := A2 /a xr , which is a finitely generated A-module,
does not have Mtor as a direct summand. This example and the argument behind it are
from https://math.stackexchange.com/questions/3593455/.
Since A is an integral domain and a is a proper ideal in A, xr 6∈ a xr , so xr 6= 0 in M .

Claim: Mtor = A xr .

To prove the claim, we have xr ∈ Mtor since x xr = 0 and x 6= 0. (Also r xr = 0.) Thus

A xr ⊂ Mtor . For the reverse containment, let
ab ∈ Mtor for a and b in A, so c ab = 0 for

some nonzero c ∈ A. That means ca r

cb is in a x , so ca = dr and cb = dx for some d ∈ a.
We want to show ab is in A xr .

If d = 0 then ca and cb are 0, so a and b are 0 since c is nonzero and Z[x] is an integral
domain. Thus ab = 0 ∈ A xr .

Suppose instead that d 6= 0. From ca = dr and cb = dx we have b(dr) − a(dx) =

bca − acb = 0, so d(br − ax) = 0 and thus br = ax. Since x | br and x - r (note r is
nonzero in Z), x | b. Write b = xs where s ∈ A, so ax = br = xsr. Thus a = sr, so
a sr r a r






b = sx = s x , which implies b ∈ A x in M . That proves the claim.
To show Mtor is not a direct summand of M , suppose M = Mtor ⊕ N for some submodule
N . The projection M → Mtor associated to that hypothetical direct sum decomposition of
M is A-linear and fixes each element of Mtor . We will show every A-linear map f : M → Mtor
must be zero on Mtor , which is incompatible with f fixing Mtor since Mtor is nonzero. Since
Mtor is the A-multiples of xr , it suffices to show f xr = 0. Write f 10 = a xr and

f 01 = b xr for some a and b in A. Then

     !        
r 1 0 1 0 r r
f =f r +x = rf + xf = ra + xb =0
x 0 1 0 1 x x

r


since x is killed by both r and x.
Remark 4.10. Example 4.9 works in the same way if we replace Z[x] by R[x] where R is
an integral domain that is not a field and r is a nonzero nonunit in R. What we used about
r and x in the example is that (r, x) is a proper ideal of Z[x], x is prime in Z[x], and x - r.
It wasn’t important that r is an integer. (For instance, we could have used r = x + 2).
Therefore we could replace Z[x] by an integral domain A containing a proper ideal of the
form (r, p) where p is prime
√ in A and p - r. Such a domain A consisting of numbers rather
than √
polynomials√is Z[ n] where the integer n is not a square and √ n ≡ 1 mod 4: √ the ideal
(1 + n, 2) in Z[ n] is proper (it has index 2), 2 is prime in Z[ n], and 2 - (1 + n).
There are some integral domains A that are not a PID but every finitely generated A-
module M has Mtor as a direct summand. This is true if A is a Dedekind domain, which is a
type of generalization of a PID. When we write M = Mtor ⊕ N , the complementary module
N is torsion-free but need not be free (that is, it need not have an A-basis). For example,
√ √ 1+√−3 √ √
let A = Z[ −6] and M = Z[ 2, 2 ], so A ⊂ M since M contains 2 and −3. Then
Mtor = {0} and M is finitely generated as an A-module (it is already finitely generated as
 MODULES OVER A PID 15

a Z-module, and Z ⊂ A), but it can be shown that M is not a free A-module. See Example
2.3 in https://kconrad.math.uconn.edu/blurbs/gradnumthy/notfree.pdf.

5. Cardinality and Index over a PID

A finite abelian group is the same thing as a finitely generated torsion module over Z, so
finitely generated torsion modules over a PID are generalizations of finite abelian groups.
Corollary 4.3 provides a method of defining a “size” for finitely generated torsion modules
over a PID as an ideal that generalizes the size of a finite abelian group.
Definition 5.1. Let T be a finitely generated torsion module over the PID A. Writing
T ∼
= A/(a1 ) ⊕ · · · ⊕ A/(am ), define the A-cardinality of T to be the ideal
cardA (T ) = (a1 a2 · · · am ).
The term A-cardinality is adapted from [2, pp. 35], where they use “A-cardinal”.5 When
A = Z, so T is a finite abelian group, cardZ (T ) is the size (cardinality) of T .
A finitely generated torsion module can be written as a direct sum of cyclic modules in
more than one way (including different numbers of cyclic components), e.g.,
(5.1) Z/(3) × Z/(4) × Z/(5) × Z/(5) ∼= Z/(5) × Z/(60)× ∼ = Z/(15) × Z/(20)
and
(5.2) R[X]/(X) × R[X]/(X) × R[X]/(X − 1) ∼
= R[X]/(X 2 − X) × R[X]/(X),
so we need to check A-cardinality is well-defined. First let’s look at a few examples of
A-cardinality, assuming that it is well-defined.
Example 5.2. If G is a finite abelian group and G ∼ = Z/(a1 )×· · ·×Z/(am ), then cardZ (G) =
(a1 · · · am )Z, whose positive generator is |a1 · · · am | = |G|.
Example 5.3. If T = {0} then each (ai ) is (1), so cardA (T ) = (1) = A. The converse holds
too: if cardA (T ) = (1) then a1 · · · am is a unit, so each (ai ) is (1) and thus T is trivial.
Example 5.4. When A is a field, so A-modules are vector spaces, the only (finitely gener-
ated) torsion module over A is T = {0}, and each ideal (ai ) has to be (1), so cardA (T ) = (1).
Example 5.5. Let V = R2 , viewed as an R[X]-module with X acting on R2 as the
identity matrix. Then V as an R[X]-module is the direct sum of submodules Re1 ⊕ Re2 ∼
=
R[X]/(X − 1) ⊕ R[X]/(X − 1), so cardR[X] (V ) = (X − 1)2 as an ideal in R[X].
Example 5.6. Let V = R2 ,
viewed as an R[X]-module with X acting on R2 as ( 10 11 ). Then
X(e1 ) = e1 and X(e2 ) = 11 = e1 + e2 , so {e2 , X(e2 )} is a basis of R2 . Since X 2 (e2 ) =
2e1 + e2 = 2(e1 + e2 ) − e2 = 2X(e2 ) − e2 , so (X 2 − 2X + 1)(e2 ) = 0, as an R[X]-module
V ∼= R[X]/(X 2 −2X +1) = R[X]/(X −1)2 . Therefore cardR[X] (V ) = (X −1)2 , which is the
same ideal as in the previous example. The difference between R[X]/(X −1)⊕R[X]/(X −1)
and R[X]/(X − 1)2 as R[X]-modules is similar to that between Z/(2) × Z/(2) and Z/(4)
as abelian groups, which are nonisomorphic but have the same size.
The ideal cardA (T ) need not be the annihilator ideal AnnA (T ) = {a ∈ A : aT = 0},
which in terms of the cyclic decomposition of T is (lcm(a1 , . . . , am )) (see Exercise 10). For
example, the Z-modules in (5.1) have Z-cardinality (300) and annihilator ideal (60), while
the R[X]-modules in (5.2) have R[X]-cardinality (X 3 −X 2 ) and annihilator ideal (X 2 −X).
5Aluffi [1, Remark 5.8] uses the term “characteristic ideal,” which can be motivated by Exercise 7(a).
16 KEITH CONRAD

When A = Z and G is a finite abelian group, AnnZ (G) is generated by the least positive
integer whose power (or multiple, in additive notation) kills everything in the group and is
traditionally called the exponent of G. The size and exponent of G are equal exactly when
G is cyclic, and likewise cardA (T ) = AnnA (T ) exactly when T is a cyclic A-module.
Now we show cardA (T ) is well-defined. You may want to skip the proof on a first reading.
Theorem 5.7. If A/(a1 ) ⊕ · · · ⊕ A/(am ) ∼ = A/(b1 ) ⊕ · · · ⊕ A/(bn ) as A-modules, where the
ai ’s and bj ’s are nonzero, then (a1 a2 · · · am ) = (b1 b2 · · · bn ) as ideals.
Proof. If A is a field then the theorem is obvious (both ideals are (1)), so assume A is
not a field: it has irreducible elements. For each irreducible π in A, we will show the
highest powers of π in a1 a2 . . . am and b1 b2 · · · bn are equal. Then by unique factorization
(a PID is a UFD) a1 a2 . . . am and b1 b2 · · · bn are equal up to multiplication by a unit, so
(a1 a2 . . . am ) = (b1 b2 · · · bn ).
Set T = A/(a1 ) ⊕ · · · ⊕ A/(am ). For each irreducible π in A we look at the descending
chain of modules
T ⊃ πT ⊃ π 2 T ⊃ · · · ⊃ π i T ⊃ · · ·
The quotient of successive modules π i−1 T /π i T is an A-module on which multiplication by
π is 0, so this is an A/(π)-vector space. Since T is finitely generated, so is π i−1 T (multiply
the generators of T by π i−1 ) and thus so is its quotient π i−1 T /π i T , so π i−1 T /π i T is a finite-
dimensional A/(π)-vector space. The dimensions dimA/(π) (π i−1 T /π i T ) will be the key. We
will show the highest power of π in a1 a2 . . . am is i≥1 dimA/(π) (π i−1 T /π i T ).
P

Step 1: T = A/(a) is a cyclic torsion module.

For i ≥ 1 we will show when T = A/(a) that
(
1, if π i | a,
dimA/(π) (π i−1 T /π i T ) =
0, otherwise.
Since T is generated as an A-module by 1, π i−1 T /π i T is spanned as an A/(π)-vector space
by π i−1 , so π i−1 T /π i T has dimension ≤ 1 over A/(π). It is 0-dimensional if and only if
π i−1 T = π i T , which is equivalent to π i−1 ∈ π i A + (a) = (π i , a). If π i | a then (π i , a) ⊂ (π i ),
and obviously π i−1 6∈ (π i ), so π i−1 T /π i T is 1-dimensional over A/(π). If π i does not divide
a then the greatest common divisor of π i and a is π j for some j ≤ i − 1, and therefore
π i−1 ∈ (π j ) = (π i , a). Thus
X
dimA/(π) (π i−1 T /π i T ) = |{i ≥ 1 : π i | a}|,
i≥1

which is the highest power of π dividing a. In particular, if π - a then the sum is 0.

Step 2: T is a finitely generated torsion module.
Writing T = A/(a1 ) ⊕ · · · ⊕ A/(am ), we have an A/(π)-vector space isomorphism
m
π i−1 T /π i T ∼
M
= π i−1 (A/(ak ))/π i (A/(ak )),
k=1
so
m
X
dimA/(π) (π i−1 T /π i T ) = dimA/(π) (π i−1 (A/(ak ))/π i (A/(ak )))
k=1
so summing over all i ≥ 1 gives us on the right the sum of the highest powers of π in all
ak ’s, which is the highest power of π in a1 a2 · · · am . So if π - a1 · · · am then the sum is 0.
 MODULES OVER A PID 17

Step 3: Comparing isomorphic cyclic decompositions.

The number dimA/(π) (π i−1 T /π i T ), which does not depend on a cyclic decomposition, is
unchanged if T is replaced by an isomorphic A-module because an A-module isomorphism
T → T 0 induces an A/(π)-vector space isomorphism π i−1 T /π i T → π i−1 T 0 /π i T 0 in a natural
way. So a1 a2 . . . am and b1 b2 · · · bn are divisible by the same highest power of π, for all π. 
Example 5.8. We have cardA (T ) = (1) if and only if each ai in a cyclic decomposition is
a unit, which means T = 0.
Example 5.9. For a ∈ A−{0}, cardA (A/(a)) = (a). For nonzero a and b, cardA (A/(ab)) =
(ab) = (a)(b) = cardA (A/(a)) cardA (A/(b)).
Example 5.10. When A = Z and G is any of the isomorphic groups in (5.1), which have
A-cardinality (300) = (22 · 3 · 52 ), let’s recover the multiplicity of 5 in this by looking at
dimZ/5Z (5i−1 G/5i G) for i ≥ 1. We have
G∼ = Z/(3) × Z/(4) × Z/(5) × Z/(5) ∼ = Z/(5) × Z/(60)× ∼ = Z/(15) × Z/(20)
and
5G ∼
= Z/(3) × Z/(4) × (0) × (0) ∼
= (0) × 5Z/(60)× ∼
= 5Z/(15) × 5Z/(20)
and 5i G = 5G for i ≥ 2. Then dimZ/5Z (G/5G) = 2 and dimZ/5Z (5i−1 G/5i G) = 0 for i ≥ 2,
so the sum of these dimensions is 2, which is the multiplicity of 5 in 300.
Example 5.11. When A = R[X] and T is an R[X]-module in (5.2), for which cardA (T ) =
(X 3 −X 2 ) = (X 2 (X −1)), the multiplicity of X in T is the sum of dimR[X]/(X) (X i−1 T /X i T )
for i ≥ 1. We have
T ∼= R[X]/(X) × R[X]/(X) × R[X]/(X − 1) ∼ = R[X]/(X 2 − X) × R[X]/(X)
and
XT ∼= (0) × (0) × R[X]/(X − 1) ∼ = XR[X]/(X 2 − X) × (0)
and X i T = XT for i ≥ 2. Then dimR[X]/(X) (T /XT ) = 2 and dimR[X]/(X) (X i−1 T /X i T ) = 0
for i ≥ 2, so the sum of the dimensions is 2, which is the multiplicity of X in X 3 − X 2 .
Corollary 5.12. If T is a finitely generated torsion A-module then cardA (T ) ⊂ AnnA (T ).
Proof. It is obvious from a cyclic decomposition T ∼ = A/(a1 ) ⊕ · · · ⊕ A/(am ) that multiplica-
tion by a1 a2 · · · am kills T , so a1 · · · am ∈ AnnA (T ), and therefore cardA (T ) ⊂ AnnA (T ). 
Theorem 5.13. For two finitely generated torsion A-modules T1 and T2 , cardA (T1 ⊕ T2 ) =
cardA (T1 ) cardA (T2 ).
Proof. Combine cyclic decompositions of T1 and T2 to get one for T1 ⊕ T2 . 
Some properties of cardA (T ) that generalize results about orders of finite abelian groups
(e.g., Cauchy’s theorem) are in the exercises.
The notion of index, not just cardinality, also generalizes from abelian groups to finitely
generated modules over a PID.
Definition 5.14. If M is a finitely generated module over the PID A with submodule M 0
such that M/M 0 is a torsion module, we set the A-index of M 0 in M to be the A-cardinality
of their quotient:
[M : M 0 ]A = cardA (M/M 0 ).
18 KEITH CONRAD

Example 5.15. If M 0 ⊂ M then M 0 = M if and only if [M : M 0 ]A = (1), since M/M 0 = 0

if and only if cardA (M/M 0 ) = (1).
Example 5.16. For nonzero a and b in A, [A2 : aA ⊕ bA]A = (ab) since A2 /(aA ⊕ bA) ∼
=
A/(a) ⊕ A/(b).
Example 5.17. Let A = Z[i],
   
2 1 0
M = Z[i] = Z[i] + Z[i] ,
0 1
and        
0 3 0 1 0
M = Z[i] + Z[i] = Z[i]3 + Z[i](1 + 2i) .
0 1 + 2i 0 1
Since M/M 0 ∼= Z[i]/(3) ⊕ Z[i]/(1 + 2i), [M : M 0 ]Z[i] = (3(1 + 2i)).
Using different bases for these two modules,
   
3 1 − 2i
M = Z[i] + Z[i]
1 + 2i 2
and    
0 3 1 − 2i
M = Z[i] + Z[i]3(1 + 2i) ,
1 + 2i 2
so M/M 0 ∼
= Z[i]/(3(1 + 2i)). Thus again we compute [M : M 0 ]Z[i] = (3(1 + 2i)).
If M is a finite free A-module and M 0 is a submodule, by Corollary 2.16 the A-index
[M : M 0 ]A is defined if and only if M and M 0 have equal rank. This is also true in the
general case (Exercise 8). When A = Z, [M : M 0 ]Z is the subgroup of Z generated by the
positive integer |M/M 0 |, which is the usual index [M : M 0 ], so the A-index generalizes the
index in group theory.
Theorem 5.18. Let M be a finitely generated A-module with submodules M 0 ⊃ M 00 and
assume M/M 00 is a torsion module. Then
[M : M 00 ]A = [M : M 0 ]A [M 0 : M 00 ]A .
Proof. In terms of A-cardinalities, this says
cardA (M/M 00 ) = cardA (M/M 0 ) cardA (M 0 /M 00 ).
The ideal cardA (M/M 00 ) is defined by hypothesis. Since M 0 /M 00 is a submodule of M/M 00
and M/M 0 is a quotient module of M/M 00 , both are finitely generated torsion modules so
cardA (M/M 0 ) and cardA (M 0 /M 00 ) are both defined.
Setting T = M/M 00 and T 0 = M 0 /M 00 , the identity we want to prove becomes
cardA (T ) = cardA (T 0 ) cardA (T /T 0 ),
which is Exercise 9. 
A finite free Z-module M looks like Zn , and a submodule M 0 also of rank n has finite
index in M . We will prove a determinant formula for the index of M 0 in M and then
generalize to the case of an arbitrary PID in place of Z.
Theorem 5.19. Let M be a finite free Z-module with rank n and M 0 be a submodule of
M with 0
Pnrank n. Let x1 , . . . , xn be a basis of M0 and y1 , . . . , yn be a basis of M . Writing
yj = i=1 cij xi with cij ∈ Z, the index [M : M ] equals | det(cij )|.
 MODULES OVER A PID 19

Proof. Our proof will be an application of aligned bases in a finite free abelian group and
finite-index subgroup: Theorem 2.14 with A = Z.
The n-tuples x1 , . . . , xn and y1 , . . . , yn do not have the same Z-span (unless M 0 = M ),
but morally they should have the same Q-span. To make this idea precise, we transfer
the data of M and 0 n
n
PnM into the vector space Q . Let e1 , . . . , en be the standard basis of
Q and set fj = i=1 cij ei , using the same coefficients that express y1 , . . . , yn in terms of
x1 , . . . , xn . Identify M with Zn by xi ↔ ei (extended by Z-linearity). This isomorphism
identifies yj with fj for all j, so M 0 inside M is identified with the Z-span of f1 , . . . , fn
inside Zn .
A Z-linear map Zn → Zn is determined by its effect on the standard basis e1 , . . . , en of
n n n
Z . Let ϕ P:nZ → Z be the Z-linear map determined by ϕ(e1 ) = f1 , . . . , ϕ(en ) = fn . Then
ϕ(ej ) = i=1 cij ei for j = 1, . . . , n, so (cij ) is the matrix representation of ϕ with respect
to the standard basis e1 , . . . , en of Zn and
ϕ(Zn ) = Zϕ(e1 ) ⊕ · · · ⊕ Zϕ(en ) = Zf1 ⊕ · · · ⊕ Zfn ,
so [M : M 0 ] = [Zn : ϕ(Zn )]. We will show [Zn : ϕ(Zn )] = | det(ϕ)| = | det(cij )|.
The bases e1 , . . . , en and f1 , . . . , fn of Zn are usually not aligned with each other. By
Theorem 2.14, there is a set of aligned bases for Zn and its submodule ϕ(Zn ):
Zn = Zv1 ⊕ · · · ⊕ Zvn , ϕ(Zn ) = Za1 v1 ⊕ · · · ⊕ Zan vn ,
where the ai ’s are nonzero integers. Then
n
Zn /ϕ(Zn ) ∼
M
= Z/ai Z,
i=1

which tells us
(5.3) [Zn : ϕ(Zn )] = |a1 a2 · · · an | .
We want to prove | det(ϕ)| = |a1 a2 · · · an | too.
A Z-linearly independent set of size n in Qn is a Q-basis of Qn , so all four sets {ei },
{ϕ(ei )}, {vi } and {ai vi } are Q-bases for Qn . For two Q-bases of Qn there is a unique
Q-linear map Qn → Qn taking one basis to the other. The Q-linear map Qn → Qn taking
ei to ϕ(ei ) is the natural extension of ϕ : Zn → Zn from a Z-linear map to a Q-linear map,
so we will also call it ϕ (its matrix representation with respect to the standard basis of Qn
is (cij ), just like ϕ as a Z-linear map on Zn ). Consider the diagram of Q-linear maps

e_i 
ϕ ϕ
Qn / Qn / fi
O O
α γ where α γ
   _
Qn / Qn vi / ai vi
β β

This diagram commutes: ϕ = γ ◦ β ◦ α. Taking determinants of these Q-linear maps

Qn → Qn ,6
(5.4) det(ϕ) = det(γ) det(β) det(α).

6We are using Q-linear maps throughout because it is nonsense to talk about the determinant of a Z-
linear map Zn → ϕ(Zn ) when ϕ(Zn ) 6= Zn : the different bases don’t all have the same Z-span but they do
all have the same Q-span.
20 KEITH CONRAD

Using the Q-basis {vi } of Qn , the matrix representation [β] is diagonal with ai ’s along its
main diagonal, so
det(β : Qn → Qn ) = a1 a2 · · · an .
What is det(α)? The map α identifies two Z-bases of the same Z-module Zn :
α(c1 e1 + · · · + cn en ) = c1 v1 + · · · + cn vn , ci ∈ Z.
Therefore α is invertible as a Z-linear map of Zn to itself. Using {ei } as the basis in which
a matrix for α is computed, the matrices of α : Qn → Qn over Q and α : Zn → Zn over Z
are the same. Since an invertible Z-linear map Zn → Zn has determinant ±1,
det(α : Qn → Qn ) = det(α : Zn → Zn ) = ±1.
A similar argument shows
det(γ : Qn → Qn ) = det(γ : ϕ(Zn ) → ϕ(Zn )) = ±1
since γ identifies two Z-bases of ϕ(Zn ). Feeding these determinant formulas into the right
side of (5.4),
(5.5) det(ϕ) = ±a1 a2 · · · an .
Comparing (5.3) and (5.5), |det(ϕ)| = |a1 a2 · · · an | = [Zn : ϕ(Zn )] = [M : M 0 ]. 
Example 5.20. Let M = Z[i] = Z + Zi and M 0 = (1 + 2i)Z[i] = Z(1 + 2i) + Z(−2 + i). In
Example 2.13, we saw [M : M 0 ] = 5 by finding aligned Z-bases for M and M 0 . We will now
compute this index in a new way by using a determinant for a matrix expressing a Z-basis
of M 0 in terms of a Z-basis of M that are not aligned bases.
A Z-basis for Z[i] is {1, i} and a Z-basis for (1 + 2i) is {1 + 2i,

−2 + i}. The isomorphism
of abelian groups (Z-modules) Z[i] → Z2 where x + yi 7→ xy identifies the ideal (1 + 2i)
with Z 12 + Z −2 1 −2 2 5.19 [M : M 0 ] = det( 12 −2



1 = ( 2 1 )(Z ), so by Theorem
√ 1 ) = |5| = 5.
Similarly,
√ for a pure quadratic√ ring Z[ m] where m is an integer that √ is not a perfect
square, Z[ m] has Z-basis {1, √m} and a √ nonzero principal ideal√(a + b m) √ (for a, b ∈ Z
not both 0) has Z-basis {a + b m, mb + a m}, so the index [Z[ m] : (a + b m)] equals
a b ) = |a2 − mb2 |.
det( mb a
√ √ √
Example 5.21. In √ the ring Z[ 10] let a be the ideal (2+5 10, 4+7 10). We √ will compute
the index of a in Z[√ 10] as abelian
√ groups using (unaligned) Z-bases
√ for Z[ √ 10] and a.
A Z-basis for Z[ 10] is {1, 10}. A Z-basis for a is {2 + 5 10, 4 + 7 10}, but this
requires verification because it is a stronger condition to generate a as a Z-module than to
generate it as an ideal7: the initial definition of a tells us that
√ √ √ √
a = Z[ 10](2 + 5 10) + Z[ 10](4 + 7 10)
√ √ √ √
= (Z + Z 10)(2 + 5 10) + (Z + Z 10)(4 + 7 10)
√ √ √ √
= Z(2 + 5 10) + Z(50 + 2 10) + Z(4 + 7 10) + Z(70 + 4 10).
√ √
For a to be spanned over Z by 2 + 5 10 and 4 + 7 10, we need to write the other two
Z-module generators in terms of them. After some linear algebra, we can do this:
√ √ √
50 + 2 10 = −57(2 + 5 10) + 41(4 + 7 10),
√ √ √
70 + 4 10 = −79(2 + 5 10) + 57(4 + 7 10).
7For example, in Z[i], (17, 3 + 5i) 6= Z17 + Z(3 + 5i) since the ideal contains 4 + i = 17i − (3 + 5i)(2 + 2i)
but 4 + i 6= 17a + (3 + 5i)b for integers a and b: look at the imaginary parts.
 MODULES OVER A PID 21
√ √ √ √
Therefore a = Z(2 + 5 10) + Z(4 + 7 10) and the numbers 2 + 5 10 and 4 + √ 10 are2
7
obviously Z-linearly
√ independent, so they are a Z-basis of a. The isomorphism Z[ 10] → Z
where x + y 10 7→ xy identifies the ideal a with Z 25 + Z 47 = ( 25 47 )(Z2 ), so Theorem




√
5.19 tells us that the index of a in Z[ 10] is |det( 25 47 )| = | − 6| = 6. The absolute value is
important: the index is not −6.
Theorem 5.22. Let M be a finite free module over the PID A with rank n and M 0 be a
submodule with
Pnrank n. Let x1 , . . . , xn be a basis of M and y1 , . . . , yn be a basis of M 0 .
0
Writing yj = i=1 cij xi with cij ∈ A, (det(cij )) = [M : M ]A . In particular, det(cij ) 6= 0.
Proof. Let K be the fraction field of A. If we run through the proof of Theorem 5.19 with
A in place of Z, K in place of Q, and use alignedLbases v1 , . . . , vn and a1 v1 , . . . , an vn for
M and M 0 , with all ai nonzero in A, so M/M 0 ∼ n
= i=1 A/(ai ), then the proof of Theorem
5.19 shows det(cij ) = det ϕ = ua1 . . . an , where u = det α det γ is a unit in A. The K-linear
operators α and γ on K n have unit determinant since they are also A-linear operators on
An and ϕ(An ) sending a basis to a basis.
By the definition of A-index, [M : M 0 ]A = cardA (M/M 0 ) = (a1 a2 · · · an ). 
Corollary 5.23. Let M bePa finite free A-module with rank n and basis x1 , . . . , xn . For
y1 , . . . , yn in M , write yj = ni=1 cij xi with cij ∈ A. Then y1 , . . . , yn is linearly independent
if and only if det(cij ) 6= 0.
Proof. If y1 , . . . , yn is a linearly independent set, then its A-span in M is a free A-submodule
of rank n, so det(cij ) 6= 0 by Theorem 5.22.
Conversely, if det(cij ) 6= 0, then we want to show an A-linear relation nj=1 cj yj = 0 with
P
cj ∈ A must have all cj equal to 0. Writing yj in terms of the xi ’s,
 
n n n n n
!
X X X X X
0= cj yj = cj cij xi =  cij ci  xi ,
j=1 j=1 i=1 i=1 j=1
Pn
so looking at the coefficients of x1 , . . . , xn tells us j=1 cij ci = 0 for all i. As a matrix
equation this says     
c11 c21 . . . cn1 c1 0
 c12 c22 . . . cn2   c2   0 
..   ..  =  ..  .
    
 .. .. . .
 . . . .  .   . 
c1n c2n . . . cnn cn 0
n n
This is an equation in A ⊂ K , where K is the fraction field of A. The matrix is the
transpose of (cij ). Since det((cji )> ) = det(cij ) 6= 0 the vector vanishes, so all ci are 0. 
Corollary 5.24. If M is a finite free module over the PID A with basis x1 , . . . , xn , a set
of n elements y1 , . . . , yn in M is a basis of M if and only if the matrix (cij ) expressing the
y’s in terms of the x’s has unit determinant.
Proof. If y1 , . . . , yn is a basis of M then (det(cij )) = [M : M ]A = (1) by Theorem 5.22,
so det(cij ) ∈ A× . Conversely, if det(c ×
P ij ) ∈ A then y1 , . . . , yn is linearly
P independent by
Corollary 5.23 and the A-index of Ayj in M is (det(cij )) = (1), so Ayj = M . 
Example 5.25. The Z-span of 22 , 36 , and 05 is Z2 since 10 = −4 22 + 3 36 − 2 05

and 01 = 12 22 − 8 36 + 5 05 , but no 2-element subset is a Z-basis since det ( 22 36 ) = 6,

det ( 22 05 ) = 10, and det ( 36 05 ) = 15. This is a rank-2 analogue of Remark 2.5.
22 KEITH CONRAD

Example 5.26. A set of n vectors v1 , . . . , vn in Zn is a basis of Zn if and only if the matrix

(v1 v2 · · · vn ) with the v’s as the columns has determinant ±1, since this matrix expresses
v1 , . . . , vn in terms of the standard basis of Zn .
The definition of the ideal cardA (T ) for a torsion module T was based on a decomposition
of T as a direct sum of cyclic A-modules. To prove cardA (T ) is well-defined (i.e., it is
independent of the choice of cyclic decomposition of T ), the proof in Theorem 5.7 gave a
formula for the multiplicity of each prime π in cardA (T ) that makes no reference to a cyclic
decomposition of T : that multiplicity is
X
dimA/(π) (π i−1 T /π i T ),
i≥1

where each quotient module π i−1 T /π i T is a vector space over the field A/(π), and the sum
is finite since the ith term is 0 for large enough i (depending on T ). This suggests the
following weaker notion of size for finitely generated torsion modules.
Definition 5.27. For a finitely generated torsion module T over the PID A, set the primary
cardinality of T to be
X
ωA (T ) = dimA/(π) (T /πT ) ∈ {0, 1, 2, 3, . . .}
(π)

where the sum runs over all nonzero prime ideals (π) of A.
Example 5.28. Let A = Z and T = Z/100Z ∼ = Z/4Z ⊕ Z/25Z. If p is 2 or 5 then
T /pT = (Z/100Z)/(pZ/100Z) ∼ = Z/pZ, so dimZ/(p) (T /pT ) = 1. If p is a prime other than
2 or 5 then pT = T (p is invertible modulo 100), so T /pT = 0 and dimZ/(p) (T /pT ) = 0.
Thus
X
ωZ (Z/100Z) = dimZ/(p) (T /pT ) = dimZ/(2) (T /2T ) + dimZ/(5) (T /5T ) = 1 + 1 = 2.
prime p

For general n, ωZ (Z/nZ) is the number of prime factors of n, which explains the name
“primary cardinality” for ωA (T ). (In number theory, ω(n) is the standard notation for the
number of prime factors of n.) The multiplicity of the prime factors of n does not play
a role in ωZ (Z/nZ). For example, if we run through the same calculation for Z/10Z and
10Z/100Z (both cyclic of order 10) then ωZ (Z/10Z) = 2 and ωZ (10Z/100Z) = 2. So unlike
the ordinary notion of size, a finite abelian group and a proper subgroup (like Z/100Z and
its subgroup 10Z/100Z) could have the same primary cardinality.
In the sum defining ωA (T ), each term dimA/(π) (T /πT ) is finite since T being finitely
generated as an A-module makes T /πT finitely generated as an A/(π)-vector space. The
following lemma explains why the sum defining ωA (T ) has only finitely many nonzero terms.
Theorem 5.29. Let T be a finitely generated torsion A-module. The ideal AnnA (T ) =
{a ∈ A : aT = 0} is not (0) and dimA/(π) (T /πT ) > 0 if and only if π | AnnA (T ), which
occurs for only finitely many prime ideals (π).
In the paragraph following Example 5.6, we saw AnnA (T ) 6= (0) by a formula for this
ideal in terms of a cyclic decomposition of T . The point of reproving AnnA (T ) 6= (0) again
is to see it can be done without relying on a cyclic decomposition.
 MODULES OVER A PID 23

Proof. To show AnnA (T ) 6= (0) we will use a finite spanning set of T as an A-module, say
{t1 , . . . , tk }. Since each element of T is an A-linear combination of t1 , . . . , tk , for a ∈ A we
have aT = {0} if and only if at1 = 0, . . . , atk = 0. Since T is a torsion module, each of
t1 , . . . , tk is killed by some nonzero element of A, and the product of such elements is a single
nonzero element of A killing all of t1 , . . . , tk and thus killing all of T . Thus AnnA (T ) 6= (0).
The ideal AnnA (T ) is principal, so it has a generator that is not 0, say AnnA (T ) = (α).
We will show dimA/(π) (T /πT ) > 0 if and only if π | α, or equivalently dimA/(π) (T /πT ) = 0
if and only if π - α.
If π - α then (α, π) = (1) since π is prime, so αx + πy = 1 for some x and y in A. Thus
for each t ∈ T , t = (αx + πy)t = x(αt) + π(yt) = π(yt) since αt = 0. This shows T ⊂ πT ,
and the reverse containment is obvious, so πT = T and thus T /πT = 0.
If π | α, write α = πα0 . Then α0 6∈ (α) (since α - α0 ), so α0 T 6= {0}: for some t ∈ T we
have α0 t 6= 0. We can’t have t ∈ πT since then α0 t ∈ α0 πT = αT = {0}, which isn’t true.
So T /πT 6= {0}.
We have dimA/(π) (T /πT ) > 0 if and only if T /πT 6= 0, and we showed that is the same
as π | α. A nonzero element of A has only finitely many prime factors up to unit multiple,
so π | α for only finitely many prime ideals (π). 
Corollary 5.30. We have ωA (T ) = 0 if and only if T = 0.
Proof. Obviously if T = 0 then ωA (T ) = 0. Conversely, if ωA (T ) = 0 then for all prime
elements π we have dimA/(π) (T /πT ) = 0, so T = πT . By Theorem 5.29, the ideal AnnA (T )
has no prime factors, so AnnA (T ) = (1). That 1 ∈ AnnA (T ) implies 1 · T = 0, so T = 0. 
Theorem 5.31. For finitely generated torsion A-modules T1 and T2 , ωA (T1 ⊕T2 ) = ωA (T1 )+
ωA (T2 ).
Proof. For each prime π, π(T1 ⊕ T2 ) = πT1 ⊕ πT2 , so (T1 ⊕ T2 )/π(T1 ⊕ T2 ) ∼
= (T1 /πT1 ) ⊕
(T2 /πT2 ). Isomorphic vector spaces have the same dimension, so
dimA/(π) (T1 ⊕ T2 )/π(T1 ⊕ T2 ) = dimA/(π) ((T1 /πT1 ) ⊕ (T2 /πT2 ))
= dimA/(π) (T1 /πT1 ) + dimA/(π) (T2 /πT2 ),
and summing over all (π) gives us ωA (T1 ⊕ T2 ) = ωA (T1 ) + ωA (T2 ). 
Example 5.32. For a ∈ A − {0}, A/(a) is isomorphic to a direct sum of cyclic modules of
the form A/(π k ) for prime π by the Chinese remainder theorem:
a = uπ k1 · · · π kr =⇒ A/(a) ∼
1 r = A/(π k1 ) ⊕ · · · ⊕ A/(π kr ),
1 r
where u ∈ A×and π1 , . . . , πr are nonassociate primes in A. So each nonzero finitely gener-
ated torsion A-module T is isomorphic to a direct sum of cyclic torsion A-modules:
T ∼
= A/(π e1 ) ⊕ · · · ⊕ A/(π en ),
1 n
where π1 , . . . , πn are prime and e1 , . . . , en are positive integers. (The primes could be the
same when T is not cyclic, e.g.,P(Z/15Z)× = h2 mod 15i ⊕ h−1 mod 15i ∼ = Z/4Z ⊕ Z/2Z.)
n ej ej
By Theorem 5.31, ωA (T ) = j=1 ωA (A/(πj )). We have ωA (A/(πj )) = 1, so ωA (T ) =
n. Thus we have a concrete interpretation for ωA (T ): it is the number of terms in a
decomposition of T as a direct sum of cyclic modules having a prime-power annihilator
ideal (AnnA (A/(π e )) = (π e ) when π is prime). For instance, if a finite abelian group G is
expressed as a direct sum of cyclic abelian groups of prime-power order then ωZ (G) is the
number of direct summands, e.g., ωZ ((Z/15Z)× )) = ωZ ((Z/4Z) ⊕ Z/2Z) = 2.
24 KEITH CONRAD

Exercises.

1. Let A be a nonzero commutative ring. Show that if every submodule of every finite
free A-module is a free A-module, then A is a PID.
2. Suppose A is a PID and π is irreducible in A. Inside A2 , set
      
1 0 x 2
M =A +A 2 = : y ≡ 0 mod π
0 π y
and
      
π 1 x 2
N =A +A = : y ≡ 0 mod π, πx ≡ y mod π .
0 π y
a) Find a basis {e1 , e2 } of A2 and a1 and a2 in A such that {a1 e1 , a2 e2 } is a basis
of N . (Such an aligned pair of bases obviously exists for A2 and M .)
b) Show there is no basis {e1 , e2 } of A2 and a1 , a2 , b1 , b2 in A such that {a1 e1 , a2 e2 }
is a basis of M and {b1 e1 , b2 e2 } is a basis of N . That is, the submodules M and N
of A2 do not admit bases simultaneously aligned with a single basis of A2 .
3. Let A be a PID, M be a finite free A-module, and N be a submodule. Consider the
following conditions on M and N :
(i) the quotient module M/N is torsion-free,
(ii) every A-basis of N extends to an A-basis of M ,
(iii) some A-basis of N extends to an A-basis of M .
Corollary 2.9 shows (i) implies (ii), and trivially (ii) implies (iii). Prove (iii) implies
(i), so the three conditions are equivalent.
4. If M is a finitely generated module over a PID A and M 0 is a submodule, is it
always possible to align their decompositions into free parts and torsion parts: can
we write M = F ⊕ T and M 0 = F 0 ⊕ T 0 such that F and F 0 are free, T and T 0
are torsion, and F 0 ⊂ F and T 0 ⊂ T ? If A is not a field, show the answer is no by
picking an irreducible π in A and using M = A ⊕ A/(π) and M 0 = A(π, 1). (Hint:
first show M 0 is free.)
5. Let A be a PID.
a) Prove an analogue of Cauchy’s theorem from group theory: for a finitely
generated torsion A-module T and irreducible π in A such that π divides cardA (T ),
meaning π divides a generator of the ideal cardA (T ), show there is some t ∈ T with
“order” π: the annihilator ideal AnnA (t) = {a ∈ A : at = 0} is πA.
b) Let T and T 0 be finitely generated torsion A-modules such that cardA (T ) =
cardA (T 0 ). If f : T → T 0 is an A-linear map, show f is one-to-one if and only if it is
onto. (When A = Z this is the familiar statement that a homomorphism between
finite abelian groups of equal size is one-to-one if and only if it is onto.)
c) Show a finitely generated A-module M is a torsion module if and only if there
is some a 6= 0 in A such that aM = 0. (This is false without a hypothesis of finite
generatedness, e.g., Q/Z is a torsion abelian group and for all nonzero integers a
we have a(Q/Z) = Q/Z.)
d) For a finitely generated A-module M and submodule N , show M/N is a torsion
module if and only if there is some a 6= 0 in A such that aM ⊂ N .
6. Let A be a PID and R be a commutative ring containing A such that R is a finite free
A-module. Prove there is an A-basis of R that contains 1. (Hint: Let e1 , . . . , en be
an A-basis of R and write 1 = ni=1 ci ei and d = gcd(c1 , . . . , cn ). Prove ϕ(R) ⊂ dA
P
 MODULES OVER A PID 25

for all A-linear maps ϕ : R → A. Deduce that d = 1 and then that there’s an
A-linear map ϕ : R → A such that ϕ(1) = 1. Conclude from the method of proof of
Theorem 2.14 that R = A ⊕ N for an A-submodule N of R. )
7. Let V be a finite-dimensional vector space over a field K and f : V → V be an
K-linear operator. Write Vf for V as a K[X]-module where X acts on V as f :
Xv = f (v) for v ∈ V .
a) Show cardK[X] (Vf ) = (χf (X)), where χf is the characteristic polynomial of f .
b) If χf (X) decomposes into linear factors in K[X], so all the eigenvalues of f
are in K and we can find a matrix for f in Jordan canonical form, show ωK[X] (Vf )
is the number of Jordan blocks in that matrix.
8. For a pair of finitely generated A-modules M ⊃ M 0 , show M/M 0 is a torsion module
if and only if M and M 0 have the same rank (that means the free parts of M and
M 0 have equal rank). This generalizes Corollary 2.16 in the case of free modules.
9. Let T be a finitely generated torsion module over the PID A and T 0 be a submodule.
Show cardA (T ) = cardA (T 0 ) cardA (T /T 0 ). In particular, if cardA (T 0 ) = cardA (T )
then cardA (T /T 0 ) = (1), so T /T 0 = {0} (Example 5.3) and thus T 0 = T , so if T 0 is
strictly contained in T then cardA (T ) ( cardA (T 0 ).
10. Let T be a finitely generated torsion module over the PID A. Write a cyclic decom-
position of T as A/(a1 ) ⊕ · · · A/(am ) for nonzero ai in A.
a) Show AnnA (T ) = (lcm(a1 , . . . , am )).
b) For t and t0 in T , let AnnA (t) = (a) and AnnA (t0 ) = (a0 ). Show there is an
A-linear combination of t and t0 whose annihilator ideal is (lcm(a, a0 )). (Hint: when
A = Z, this becomes the fact from group theory that in a finite abelian group, if g
has order m and h has order n, then some element of hg, hi has order lcm(m, n).)
c) Using (a) and (b), show there is a t0 ∈ T such that AnnA (t0 ) = AnnA (T ).
(Hint: when A = Z, this is the statement that in a finite abelian group, the least
common multiple of the orders of its elements is the order of some element of the
group.)
11. Prove Corollaries 5.23 and 5.24 if A is an integral domain, not necessarily a PID.

References
[1] P. Aluffi, Algebra: Chapter 0, Amer. Math. Soc., Providence, 2009.
[2] H. Cohen and H. W. Lenstra, Jr., “Heuristics on class groups of number fields,” pp. 33–62 in Number
theory, Noordwijkerhout 1983, Springer Lecture Notes in Mathematics 1068, Springer-Verlag, Berlin,
1984.
[3] S. Roman, Advanced Linear Algebra, 2nd ed., Springer-Verlag, New York, 2005.
[4] J. Rotman, An Introduction to the Theory of Groups,4th ed., Springer-Verlag, NY, 1995.
[5] J. Rotman, Advanced Modern Algebra, Prentice-Hall, Upper Saddle River, NJ, 2002.
[6] P. Samuel, Algebraic Theory of Numbers, Dover, 2008.