The Quillen–Suslin theorem is easy to state: if is a principal ideal domain, then every finite projective module over is free. Equivalently, every vector bundle over is trivial. Here and throughout, “finite” means finitely generated.
The proof is broken into two parts:
- Every finite projective module over is stably free.
- Every stably free module over is free.
As will become clear, the method of proving each part will be quite different: the first consists mostly of manipulating exact sequences and resolutions, while the second deals with matrices of elements of rings. This post gives Serre’s reduction, the first part of the proof; we will complete the argument in Quillen–Suslin 2: Vaserstein’s proof.
Stably free modules and resolutions
In this section, all modules without qualification will be over some fixed commutative ring .
Definition 1.¶ A module is stably free if there is a finite free module such that is finite free. Two modules are stably isomorphic if there are finite free modules such that
In this case we write .
When is stably free it is finite projective, as a summand of a finite free module. It is also clear that if is stably isomorphic to a stably free module, then is stably free. Indeed, if and is finite free, then is finite free as well.
Throughout this post, we will generally reserve the letter for finite free modules and for stably free modules. Much of the theory of stably free modules is about resolutions.
Definition 2.¶ A finite free resolution of a module is a finite exact sequence
with each finite free. A finite stably free resolution is defined in the same way, with each term stably free.
The finite free dimension , respectively the finite stably free dimension , is the minimal length of such a resolution, and is if no such resolution exists. Thus dimension zero means that itself is finite free, respectively stably free.
We start with a comparison of presentations.
Lemma 3 (Schanuel’s lemma).¶ Suppose that for the sequences
are exact, with projective. Then
Proof. Consider the pullback . Projection onto either factor gives exact sequences
Both split by projectivity, so the pullback is isomorphic to both and . □
Proposition 4.¶ Let be modules such that , and suppose that
are exact with stably free. Then .
Proof. Choose finite free and an isomorphism . Adding the identity of to the th presentation gives two projective presentations of the same module. Schanuel’s lemma gives
Choose finite free such that is finite free. Adding to both sides makes the summands accompanying and finite free. This is the required stable isomorphism. □
Lemma 5.¶ Let and suppose . If
is exact with each stably free, then
is stably free, where when .
Proof. It suffices to show that is stably isomorphic to a stably free module. By assumption there is a resolution
of length , extending by zeros if necessary. Write , and set
In particular, and .
For , Proposition 4 gives . For we have exact sequences
and similarly for the primed modules. The same proposition now gives inductively. In particular , as desired. □
Corollary 6.¶ Assume is Noetherian, and suppose
is exact with stably free. If , then
Proof. If , then is stably free, hence projective, so . Adding finite free complements for and shows that is stably free.
For , the module and the successive kernels of finite free presentations are finite by Noetherianity. Choose enough such presentations to extend to a partial stably free resolution with terms in degrees . Its last kernel is stably free by Lemma 5. Removing gives a stably free resolution of of length at most . □
Resolvable modules
Definition 7.¶ A module is resolvable if it admits a finite free resolution, i.e. if . A ring is pointwise resolvable if every prime ideal is resolvable as an -module.
The distinction between prime ideals and their quotient modules is harmless here: the sequence
and the next proposition will show that one is resolvable if and only if the other is.
First, a small observation lets us use stably free terms whenever convenient. A finite stably free resolution can always be changed into a finite free resolution. Choose a finite free such that is finite free, and add the exact complex in degrees and . Now the term in degree is free and the term in degree is still stably free. Repeat in degrees and , then and , and so on. After the last step, all terms are finite free, with at most one extra term at the left.
Proposition 8 (The 2-of-3 property).¶ Assume is Noetherian. If
is exact and any two of the modules are resolvable, then so is the third.
Proof. We give the three cases explicitly.
Suppose first that and have finite free resolutions and . Lift through , using freeness. Together with , this gives a surjection
Its kernel fits into a short exact sequence whose outer terms are the kernels of and . Indeed, projecting a pair in the middle kernel to gives a map onto the second kernel; a preimage can be corrected using , and its kernel is the first kernel. Repeating this construction gives a resolution of with terms . It is finite because the original resolutions are finite. This is the horseshoe construction.
For the other cases, recall the following explicit construction. Given a homomorphism and finite free resolutions and , projectivity lets us lift to maps commuting with the differentials. At degree this is the lifting property for . At each later degree, the map already constructed lands in the appropriate kernel, which is the image of the next differential, so the same lifting property applies.
The mapping cone has terms
and differential
with the second component omitted when . Its homology is
For completeness, these identities follow from the degreewise split exact sequence of complexes
where denotes the complex with in degree and differential . The connecting map in homology is precisely , and the resolutions have no positive homology. Equivalently, the same identities follow by substituting the displayed formula for and chasing kernels and images.
If and are resolvable, apply this construction to the injection . The cone has no positive homology and has , so it is a finite free resolution of .
Finally, suppose and are resolvable, and apply the construction to . This time , so is surjective. Since is finite free, it splits. Thus
so is stably free. The homology calculation gives an exact sequence
This is a finite stably free resolution of , and the preceding observation converts it to a finite free resolution. □
In fact, the proof of the 2-of-3 property did not need Noetherianity. It will be crucial in what follows, however, because it gives finite filtrations. This is to be expected: our results concern finite resolutions, and Serre’s problem asks about sheaves which are locally free of finite rank.
Prime filtrations
To prove the following lemma, we will need associated primes. For an -module , write
An associated prime of is a prime ideal for some nonzero . Equivalently, has a submodule isomorphic to .
If is Noetherian and , then always has an associated prime. Indeed, choose a maximal element of
under inclusion, say . This ideal is proper. If and , then and . Maximality gives equality, so . Thus is prime.
Lemma 9 (Prime filtration).¶ Let be Noetherian and a finite -module. Then there are prime ideals and a finite filtration
such that for each .
Proof. By Noetherianity, satisfies the ascending chain condition on submodules. Start with . If , choose an associated prime of . The corresponding submodule of has an inverse image satisfying
In particular . This process must stop by the ascending chain condition, and it can only stop when . □
Adjoining one variable
Our next goal is the following theorem.
Theorem 10.¶ Suppose that is Noetherian and pointwise resolvable. Then is pointwise resolvable.
From this Serre’s reduction will follow relatively easily, but its proof requires more than the exact sequence manipulation we have used so far. We first prove an intermediate step. For an ideal , write .
Lemma 11.¶ Let be Noetherian, let be prime, and put . Suppose . Then there is a finite -module with the following properties:
- There is an exact sequence
- The module admits a finite filtration
such that , where each is prime and
Proof. Let , a Noetherian domain, and let be its fraction field. Since , the quotient
is a nonzero ideal of . The ring is Noetherian, so is generated by finitely many polynomials .
Choose a nonzero of minimal degree. Division in gives
for each nonzero remainder. Let be a product of denominators such that for all . Then
If this polynomial were nonzero, it would have degree less than , contradicting minimality. Hence for every , and
Set , considered as an -module through the quotient map . There is an exact sequence
Multiplication by identifies with : it is onto by definition and injective because is a domain and . This gives the first required sequence.
Choose a lift of , so . The inclusion shows that , and by the definition of the -module structure. Thus
The module is finite, so Lemma 9 gives a filtration with quotients . Every element annihilating also annihilates these quotients, hence
Therefore , and the inclusion is strict since but . □
We can now prove Theorem 10.
Proof. First, each extended prime is resolvable as an -module. Indeed, pointwise resolvability of gives a finite free resolution of . The extension is flat, and
Tensoring the resolution with therefore gives the required finite free resolution of . The 2-of-3 property also shows that is resolvable.
Suppose now that some prime of is not resolvable. By Noetherianity of , the set
has a maximal element. Choose such a and write . The preceding paragraph implies , so we may apply Lemma 11.
In its filtration of , each prime has contraction strictly larger than . By maximality, is resolvable. The exact sequence
then shows that is resolvable. Starting from and applying the 2-of-3 property to
shows inductively that is resolvable.
We already know that is resolvable, so the first sequence of Lemma 11 shows that is resolvable. Finally,
shows that is resolvable, a contradiction. □
Corollary 12 (Serre).¶ Suppose that is Noetherian and pointwise resolvable. Then every finite -module is resolvable.
Proof. By Theorem 10, all prime ideals of are resolvable. If is a finite -module, take a prime filtration
The sequence shows that each quotient is resolvable. The sequences
then show by induction that every , and in particular , is resolvable. □
Serre’s reduction
It remains to connect resolutions back to stably free modules.
Lemma 13.¶ Let be a projective module. Then is stably free if and only if it is resolvable.
Proof. If with finite free, then
is a finite free resolution.
Conversely, suppose has a finite free resolution, and induct on its length. In length zero there is nothing to prove. Otherwise write its last part as
Since is projective, this sequence splits, so and is projective. The remaining terms form a shorter finite free resolution of , so is stably free by induction. Choose finite free with . Then
is finite free. Thus is stably free. □
Theorem 14 (Serre’s reduction).¶ Let be a principal ideal domain. Then every finite projective module over is stably free.
Proof. A PID is Noetherian and every ideal is finite free: the zero ideal is free of rank zero, and a nonzero principal ideal is isomorphic to because is a domain. In particular is pointwise resolvable.
Repeatedly applying Theorem 10 shows that is Noetherian and pointwise resolvable. Hence, for , Corollary 12 shows that every finite -module is resolvable. Its finite projective modules are therefore stably free by Lemma 13.
For , every finite projective -module is already free. To recall the elementary reason, any submodule is finite free: project onto the last factor, whose image is a principal ideal, and use induction on on the kernel. The resulting short exact sequence splits because the image is free. A finite projective module is a summand, hence a submodule, of some . □
Notice that Serre’s reduction works more generally over any Noetherian pointwise resolvable ring: its finite projective modules, and those over its polynomial rings, are stably free. The argument applies in particular when every prime ideal is stably free, since stably free modules are resolvable.
A natural question is where in this argument the ring may be replaced by a scheme and the modules by quasicoherent sheaves. The biggest issue is that finite freeness, which we would like to replace with triviality, is not local. In addition, vector bundles need not be projective objects of . For these reasons, the best setting for this sequence of results is the affine case.
We have now reduced the Quillen–Suslin theorem to a cancellation problem: if adding a finite free summand makes a module free, must it already be free? We answer this for polynomial rings over a PID in Quillen–Suslin 2: Vaserstein’s proof. For the original exposition and its geometric context, see Serre Schemes and the Quillen–Suslin Theorem.