Motivation
A finite abelian group is assembled from finite cyclic groups, but finite generation does not imply finiteness: itself has one generator and infinitely many elements. The general classification has two parts. A free part records independent directions of infinite order; a finite torsion part records elements killed by nonzero integers. Neither part should be inferred from the number of generators in a particular presentation alone.
We derive the abelian-group classification from the previously proved PID module structure theorem. We then explain which pieces are intrinsic, calculate from relation matrices, and give tests that can be performed without knowing the original generators.
Definitions and the integer-module bridge
Definition 13.1.1
(Finitely generated abelian groups as -modules)
Write the group operation additively. An abelian group is finitely generated if finitely many elements have every group element as an integer linear combination. For and , let mean repeated addition, using the additive inverse when . Commutativity makes the distributive laws hold, so is a -module. Conversely every -module is an abelian group under addition. Finite generation has the same meaning in both languages.
Definition 13.1.2
(Torsion elements and the torsion subgroup)
An element is torsion if for some positive integer . The torsion elements form a subgroup : if and , then and . A group is torsion-free when its only torsion element is zero. The set is intrinsic: every homomorphism sends torsion elements to torsion elements, and an isomorphism identifies the two torsion subgroups.
Theorem 13.1.3
(Classification of finitely generated abelian groups)
Every finitely generated abelian group is isomorphic to
The nonnegative integer and the list are uniquely determined. Empty sums are allowed. The torsion subgroup is the displayed finite direct sum, and .
Proof and the role of each hypothesis
Proof
(From the PID structure theorem)
The integers form a principal ideal domain. Apply the previously proved PID structure theorem to as a finitely generated -module. Its nonzero nonunit invariant factors can be chosen positive, giving the displayed divisibility chain. A unit factor is the zero group and is omitted. A free coordinate is not a factor of order zero: it contains elements of infinite order.
Any element in the finite cyclic part is killed by the product . A nonzero element of cannot be killed by any positive integer, because at least one of its integer coordinates is nonzero. It follows that the finite part is exactly , and the quotient is . The PID theorem gives uniqueness of and of the normalized invariant factors. The finite-group chapter separately explains recovery from primary subgroups; we do not reprove Smith reduction here.
The free rank is also visible without selecting a decomposition: the rational vector space has dimension . Tensoring with turns each summand into and each finite cyclic summand into zero, since every positive integer becomes invertible. Thus free rank is invariant under isomorphism, even though a chosen free subgroup inside need not be canonical.
Proof X-ray
A canonical subgroup and a noncanonical complement
The set of elements killed by some positive integer is determined by the group, so is canonical. The theorem supplies an isomorphism , but it does not select a unique copy of inside . Both statements are compatible.
Splitting need not choose a unique free subgroup
Take . Its torsion subgroup is . Both subgroups generated by and by are infinite cyclic, intersect trivially, and together with give all of . They are different subgroups, since is not an integer multiple of . Thus the quotient has a well-defined free rank although its embedding back into the original group is not uniquely specified.
The quotient map sends both generators to the same generator of . Picking either lift defines a splitting. This distinction becomes important whenever a later construction is said to be natural: existence of some decomposition alone does not make the chosen complement preserved by every automorphism.
From relations to a group
Suppose has commuting generators and a specified finite list of integer-linear relations. Put the coefficient column of each relation into an integer matrix with rows. The presented abelian group is
Unimodular row operations change the free coordinates; unimodular column operations replace the relation generators. Both give isomorphic quotients. The earlier Smith theorem gives a diagonal matrix with positive nonzero entries and remaining zero coordinates. Delete entries equal to one; each remaining entry contributes , and each unfilled row contributes a free .
Theorem 13.1.4
(Reading Smith data)
If an integer relation matrix has nonzero Smith entries, then the free rank of its cokernel is . The nonunit Smith entries, in divisibility order, are its invariant factors. In particular the group is finite exactly when .
The finite matrix need not be square, and the number of written relations does not by itself determine : relations can be dependent. If is square and has nonzero determinant, its cokernel is finite of order ; in the diagonal form this is the product of its entries. If the determinant is zero, it is wrong to call the group a finite group of order zero. At least one free coordinate remains.
Worked example 13.1.5
(A relation matrix with a free direction)
Let
Subtract three times row one from row two, then subtract twice column one from column two. After changing the second row's sign the matrix has diagonal entries and a zero third row. Thus
Its torsion subgroup has eight elements; the whole group is infinite because of the free coordinate. All operations used are invertible over .
Worked example 13.1.6
(Dependent relations leave a free coordinate)
Let have commuting generators and relations and . The second is twice the first. The relation matrix is . Subtract three times row one from row two, then subtract twice column one from column two, giving diagonal entries . Hence . Two written relations did not kill both generators.
Why the integer operations preserve the quotient
A column operation replaces the listed relations by another generating set for the same subgroup of . For example, replacing the second column by the second minus twice the first loses no relation: the old second column is recovered by adding twice the first to the new second. Row operations act on the ambient free group by an invertible integer matrix. They can change the coordinates of all the relations simultaneously, but the inverse row operation transports the new relation subgroup back to the old one. These are different reasons for invariance, and both require an integer inverse.
A nonunit scaling is different. In the one-generator presentation , replacing the single relation by changes a two-element group to the zero group. Replacing it by changes it to a four-element group. Even if rational row reduction would make all three relation matrices have the same rank, integer quotients retain their distinct torsion.
For a square full-rank matrix, multiplying its Smith entries gives the absolute determinant because unimodular matrices have determinant . This determinant is an efficient consistency check for the order of the finite quotient, but it does not determine the invariant factors individually. For example, diagonal matrices with entries and both have determinant four; their cokernels are and , which differ in whether an element of order four exists.
Intrinsic tests and generator counts
For , write and . The first is a subgroup, and the quotient is intrinsic. Both can be read from the decomposition.
Theorem 13.1.7
(Torsion counts, quotients and generators)
For the normalized decomposition above and ,
The minimum number of generators of is ; for the trivial group it is zero.
Proof
(Counting summand by summand, then reducing modulo a prime)
On , multiplication by positive has zero kernel and quotient of order . On , let . The congruence has exactly solutions: divide by and use coprimality to see that is a multiple of modulo . The kernel therefore has size , and the multiplication map has image of size , so its quotient also has size . Products over the direct summands give both formulas.
The decomposition gives generators, proving an upper bound. If , choose a prime dividing . Because , the same prime divides every . Then is a vector space of dimension over . Any generators of map to a spanning set of that vector space, so at least are needed. If , the same argument with any prime gives lower bound . For the zero group the empty set generates it.
Proof X-ray
Choosing the prime that sees every summand
The upper bound costs nothing: the decomposition itself lists generators. The lower bound needs a quotient in which counting generators becomes counting dimensions. Reduction modulo a prime makes a vector space over , and the image of any generating set spans it. The real move is the choice of . In general is plus the number of divisible by ; choosing and using the divisibility chain makes every finite summand contribute. Without the chain the bound fails: has two summands but is cyclic, and no single prime sees both.
Worked example 13.1.8
(A quotient and its killed elements)
Let . With ,
The free part contributes no nonzero elements killed by four, but contributes residue classes to the quotient. The invariant factors are , so four generators are necessary and sufficient.
Worked example 13.1.9
(Two descriptions of the same torsion part)
The Chinese remainder theorem gives . Hence . The invariant-factor description has rank one and one factor ; its minimum number of generators is two, not three. Coprime cyclic summands can be combined.
Worked example 13.1.10
(A torsion-free group outside the theorem)
The additive group has no nonzero torsion, but it is not finitely generated. Given finitely many rational generators, choose a common denominator ; every integer combination has denominator dividing in that expression, while cannot be represented as for any integer . Thus torsion-free alone does not imply free of finite rank, and the finite-generation hypothesis cannot be dropped.
Counterexample mode
The same rank and number of torsion elements need not determine the group
and both have rank one and torsion subgroup of order four. They are not isomorphic: the first has an element of order four, while every torsion element of the second is killed by two. The actual invariant factors, not just their product, matter.
Common mistakes and scope
Common mistake
(Commutativity, finite generation and integer inverses)
The decomposition requires an abelian group and finite generation. A presentation of a nonabelian group cannot be replaced by an integer matrix without first imposing commutativity. Do not divide an integer relation by a nonunit: replacing by destroys a nonzero torsion element.
Common mistake
(Rank is not a count of infinite-order elements)
The free rank is not the number of elements of infinite order, which can be infinite as soon as . The torsion subgroup is canonical, but a complementary free subgroup may not be. An invariant-factor list must satisfy divisibility and exclude unit factors. We use the earlier PID theorem, including its uniqueness proof; this chapter does not reprove that theorem or solve arbitrary nonabelian presentation problems.
Summary
Finite generation allows the PID structure theorem to separate into a free part and finite cyclic torsion. The torsion subgroup and rank are intrinsic; normalized invariant factors classify the remaining finite part. Integer relation matrices, finite quotients and prime reduction turn the theorem into explicit calculations.
Exercises
Quick check
(Determine the rank and torsion subgroup of .)
Solution · Answer
The rank is two, and the torsion subgroup is . The factors form the divisibility chain .
Quick check
(Compute the cokernel of over .)
Solution · Answer
Subtract three times row one from row two and then twice column one from column two. The result is , so the cokernel is .
Quick check
(Find the minimum number of generators of .)
Solution · Answer
The invariant factors are , so and . Four generators suffice by the displayed decomposition and are necessary because reduction modulo two has vector-space dimension four.
Quick check
(Compute the sizes of and for .)
Solution · Answer
The killed subgroup has size . The quotient has size , because the free coordinate contributes three residue classes.
Quick check
(Explain why does not contradict the theorem that a finitely generated torsion-free abelian group is free.)
Solution · Answer
is torsion-free but not finitely generated: any finite collection of rational generators has a common denominator, so its integer span cannot contain all rational numbers. The finite-generation hypothesis fails.