← All writingMathematics blog

Quillen–Suslin 2: Vaserstein’s proof

Unimodular vectors, Horrocks’ theorem, and Vaserstein’s local-to-global argument, with the monic-polynomial step over a PID.

In this post we complete the proof of the Quillen–Suslin theorem using Vaserstein’s simplified proof. In Quillen–Suslin 1: Serre’s reduction, we showed that every finite projective module over a polynomial ring over a PID is stably free. Our remaining task is to show that every stably free module over the same ring is free.

The focus is the following definition.

Unimodular vectors

Definition 1. Let AA be a ring.

  1. A unimodular vector (f1,…,fm)(f_1,\ldots,f_m) is an ordered list of elements of AA which generate the unit ideal.
  2. A unimodular vector has the unimodular extension property (UEP) if there is a matrix in GL⁡m(A)\GL_m(A) whose first column is (f1,…,fm)t(f_1,\ldots,f_m)^t.
  3. The ring AA has the unimodular extension property if every unimodular vector over AA has the UEP.

We regard vectors as columns from now on. This is a fundamentally geometric notion: unimodularity says precisely that the principal open sets D(fi)D(f_i) cover Spec⁡(A)\Spec(A). Equivalently, the section ff of the trivial rank-mm bundle generates a trivial line subbundle, and the UEP asks whether it extends to a frame of the whole bundle. The following proposition explains why the definition is useful, in light of the previous post.

Proposition 2. A ring AA has the UEP if and only if every finite stably free AA-module is free.

Proof. Suppose first that AA has the UEP. By induction on the number of free summands, it suffices to show that if E⊕AE\oplus A is finite free, then EE is free.

Choose an identification E⊕A≅AmE\oplus A\cong A^m, and let u1u_1 be the image of (0,1)(0,1). Projection to the second summand gives a linear map p ⁣:Am→Ap\colon A^m\to A such that p(u1)=1p(u_1)=1, so the coordinates of u1u_1 generate AA. We may therefore complete u1u_1 to a basis u1,u2,…,umu_1,u_2,\ldots,u_m of AmA^m.

For j≥2j\geq2, replace uju_j by

vj=uj−p(uj)u1.v_j=u_j-p(u_j)u_1.

These replacements preserve the basis and put vjv_j in ker⁡p=E\ker p=E. If v∈Ev\in E is written in the basis u1,v2,…,vmu_1,v_2,\ldots,v_m, applying pp shows that its u1u_1-coefficient is zero. Hence v2,…,vmv_2,\ldots,v_m is a basis of EE.

Conversely, let f∈Amf\in A^m be unimodular. Choose ai∈Aa_i\in A such that ∑iaifi=1\sum_i a_if_i=1 and define

p ⁣:Am⟶A,p(v1,…,vm)=∑iaivi.p\colon A^m\longrightarrow A, \qquad p(v_1,\ldots,v_m)=\sum_i a_iv_i.

Since p(f)=1p(f)=1, we have Am=Af⊕ker⁡pA^m=Af\oplus\ker p. The second summand is stably free, hence free by assumption. A basis of it, together with ff, completes the required matrix. □

The first example of a ring with the UEP is a PID. There is a brute-force induction proof, but the following argument is a nice trick.

Proposition 3. Every PID has the UEP.

Proof. Let f∈Amf\in A^m be unimodular, choose aia_i with ∑iaifi=1\sum_i a_if_i=1, and use the map pp from the preceding proof. Then

Am=Af⊕ker⁡p.A^m=Af\oplus\ker p.

Since AA is a PID, the submodule ker⁡p⊆Am\ker p\subseteq A^m is free. It has rank m−1m-1, as one sees after tensoring with Frac⁡(A)\Frac(A). Thus ff extends to a basis. □

Definition 4. For f,g∈Amf,g\in A^m, we say that ff and gg are AA-equivalent, and write f∼gf\sim g, if there is M∈GL⁡m(A)M\in\GL_m(A) such that f=Mgf=Mg.

This is an equivalence relation. A unimodular vector has the UEP exactly when it is equivalent to the first standard basis vector e1e_1. Permuting components, multiplying one component by a unit, and adding a multiple of one component to another all preserve equivalence.

Two small cases will be useful. A unimodular vector of length one is a unit. For a unimodular vector (f1,f2)t(f_1,f_2)^t, choose a1f1+a2f2=1a_1f_1+a_2f_2=1; then

(f1−a2f2a1)\begin{pmatrix} f_1&-a_2\\ f_2&a_1 \end{pmatrix}

has determinant 11 and completes the vector. Thus vectors of length at most two have the UEP over every commutative ring.

Horrocks’ theorem

We next need Horrocks’ theorem. This is the machinery behind the inductive proof of the Quillen–Suslin theorem. We give the polynomial argument, including the degree reduction which makes the induction work.

Lemma 5. Let RR be a ring, let f∈R[x]f\in R[x] be monic of degree d>0d>0, and let

g=b0+b1x+⋯+bsxs,s<d.g=b_0+b_1x+\cdots+b_sx^s, \qquad s<d.

If the coefficients bib_i generate RR, then the ideal (f,g)(f,g) contains a monic polynomial of degree d−1d-1.

Proof. Division by the monic polynomial ff gives a unique remainder of degree less than dd. For 0≤i≤s0\leq i\leq s, let hih_i be the remainder of xd−1−igx^{d-1-i}g modulo ff, and let cic_i be its coefficient of xd−1x^{d-1}. Each hih_i belongs to (f,g)(f,g).

The term bixib_ix^i contributes bib_i to cic_i, terms with index less than ii have degree less than d−1d-1, and terms with index greater than ii contribute RR-multiples of their coefficients. Consequently,

ci=bi+∑j>irijbj(rij∈R).c_i=b_i+\sum_{j>i}r_{ij}b_j \qquad(r_{ij}\in R).

Descending induction on ii shows that (c0,…,cs)=(b0,…,bs)=R(c_0,\ldots,c_s)=(b_0,\ldots,b_s)=R. Choose aia_i with ∑iaici=1\sum_i a_ic_i=1. Then ∑iaihi\sum_i a_ih_i has degree d−1d-1 and leading coefficient 11, as required. □

Theorem 6 (Horrocks). Let (R,m)(R,\m) be a local ring. If f∈R[x]mf\in R[x]^m is a unimodular vector with a monic component, then ff has the UEP.

Proof. The cases m=1,2m=1,2 were handled above, so assume m≥3m\geq3. After permuting components, let f1f_1 be monic of degree dd. We induct on dd.

If d=0d=0, then f1=1f_1=1 and subtracting fif_i times the first component from the iith component gives e1e_1. Suppose d>0d>0. Divide each fif_i, for i≥2i\geq2, by f1f_1 and subtract the quotient times f1f_1. These operations let us assume that

deg⁡fi<d(i≥2).\deg f_i<d\qquad(i\geq2).

At least one coefficient of one of these remaining components is a unit in RR. Otherwise all the fif_i for i≥2i\geq2 would vanish in (R/m)[x](R/\m)[x], and unimodularity would say that the monic degree-dd polynomial f1‾\overline{f_1} generates the unit ideal of (R/m)[x](R/\m)[x], which is impossible. Relabel so that f2f_2 has a unit coefficient.

By Lemma 5, there are u,v∈R[x]u,v\in R[x] such that

h=uf1+vf2h=uf_1+vf_2

is monic of degree d−1d-1. Let cc be the coefficient of xd−1x^{d-1} in f3f_3. Replace f3f_3 by

f3+(1−c)uf1+(1−c)vf2=f3+(1−c)h.f_3+(1-c)uf_1+(1-c)vf_2=f_3+(1-c)h.

This is an elementary change, and its coefficient of xd−1x^{d-1} is c+(1−c)=1c+(1-c)=1.

We have obtained an equivalent unimodular vector with a monic component of degree d−1d-1. The induction hypothesis completes the proof. □

Corollary 7. If RR is a local ring and f∈R[x]mf\in R[x]^m is unimodular with a monic component, then f∼f(0)f\sim f(0) over R[x]R[x].

Proof. Evaluating a relation ∑igifi=1\sum_i g_if_i=1 at zero shows that f(0)f(0) is unimodular over RR. At least one fi(0)f_i(0) is a unit, since otherwise all of them would lie in m\m. Hence f(0)∼e1f(0)\sim e_1 over RR, and Horrocks’ theorem gives f∼e1f\sim e_1 over R[x]R[x]. □

From local rings to a domain

The local theorem tells us when a polynomial vector is equivalent to its value at zero. The next step is to patch this equivalence over a domain. The useful intermediate statement concerns translations of the variable.

Lemma 8. Let AA be a domain and S⊆A∖{0}S\subseteq A\setminus\{0\} a multiplicative subset. If f(x)∼f(0)f(x)\sim f(0) over S−1A[x]S^{-1}A[x], then there is c∈Sc\in S such that

f(x+cy)∼f(x)f(x+cy)\sim f(x)

over A[x,y]A[x,y].

Proof. Choose M(x)∈GL⁡m(S−1A[x])M(x)\in\GL_m(S^{-1}A[x]) with M(x)f(0)=f(x)M(x)f(0)=f(x), and set

G(x,y)=M(x)M(x+y)−1.G(x,y)=M(x)M(x+y)^{-1}.

Then

G(x,y)f(x+y)=f(x),G(x,0)=Im.G(x,y)f(x+y)=f(x), \qquad G(x,0)=I_m.

Thus G(x,y)=Im+yH(x,y)G(x,y)=I_m+yH(x,y) for a matrix HH over S−1A[x,y]S^{-1}A[x,y]. Choose a common denominator c∈Sc\in S for its finitely many coefficients. We then have

G(x,cy)=Im+cyH(x,cy)∈Mat⁡m(A[x,y]).G(x,cy)=I_m+cyH(x,cy)\in\operatorname{Mat}_m(A[x,y]).

It remains to check invertibility over A[x,y]A[x,y], rather than just over its localization. A unit in a polynomial ring over a domain is constant, so det⁡M(x)\det M(x) is a unit of S−1AS^{-1}A. Therefore

det⁡G(x,cy)=det⁡M(x)det⁡M(x+cy)=1.\det G(x,cy) =\frac{\det M(x)}{\det M(x+cy)}=1.

The adjugate supplies an inverse over A[x,y]A[x,y]. The displayed identity now proves the desired equivalence. □

Proposition 9. Let AA be a domain and f∈A[x]mf\in A[x]^m a unimodular vector with a monic component. Then f(x)∼f(0)f(x)\sim f(0) over A[x]A[x].

Proof. Let

J={c∈A:f(x+cy)∼f(x) over A[x,y]}.J=\{c\in A:f(x+cy)\sim f(x)\text{ over }A[x,y]\}.

We first check that JJ is an ideal. Clearly 0∈J0\in J. If c∈Jc\in J and a∈Aa\in A, substituting ayay for yy in an equivalence gives ac∈Jac\in J. For addition, suppose

Mc(x,y)f(x)=f(x+cy),Md(x,y)f(x)=f(x+dy).M_c(x,y)f(x)=f(x+cy), \qquad M_d(x,y)f(x)=f(x+dy).

Substitute x+dyx+dy for xx in the first identity and compose with the second. This gives

Mc(x+dy,y)Md(x,y)f(x)=f(x+(c+d)y).M_c(x+dy,y)M_d(x,y)f(x)=f(x+(c+d)y).

Both factors are invertible over A[x,y]A[x,y], so c+d∈Jc+d\in J.

Now let p⊂A\p\subset A be a prime ideal. The ring ApA_\p is local, so Corollary 7 gives f(x)∼f(0)f(x)\sim f(0) over Ap[x]A_\p[x]. Applying Lemma 8 with S=A∖pS=A\setminus\p produces c∈J∖pc\in J\setminus\p. Thus JJ is contained in no prime ideal, and J=AJ=A.

In particular, there is M(x,y)∈GL⁡m(A[x,y])M(x,y)\in\GL_m(A[x,y]) with

M(x,y)f(x)=f(x+y).M(x,y)f(x)=f(x+y).

Evaluating xx at zero gives M(0,y)f(0)=f(y)M(0,y)f(0)=f(y). The specialized matrix is invertible because its inverse specializes as well. Renaming yy as xx proves the proposition. □

Making a component monic over a PID

We can now discard a variable whenever a component is monic in it. Over a field, a suitable change of variables makes any nonzero polynomial have constant nonzero leading coefficient, which we may then scale to 11. Over a PID, that coefficient need not be a unit. We therefore need a stronger version of the monic trick.

We will use unique factorization to prove exactly the version needed here. Recall that polynomial rings over a PID are noetherian UFDs. Saying that an ideal has no common irreducible divisor means that it is not contained in (q)(q) for any irreducible element qq.

Lemma 10. Let BB be a noetherian UFD and I⊆B[t]I\subseteq B[t] a nonzero ideal with no common irreducible divisor. The set

L(I)={0}∪{leading coefficients in t of nonzero elements of I}L(I)=\{0\}\cup\{\text{leading coefficients in }t\text{ of nonzero elements of }I\}

is an ideal of BB with no common irreducible divisor.

Proof. To add two leading coefficients, multiply their polynomials by suitable powers of tt to give them the same degree and then add. If the coefficients cancel, their sum is zero, which is included in L(I)L(I). Multiplication by elements of BB is equally immediate. Thus L(I)L(I) is an ideal.

Suppose, for a contradiction, that L(I)⊆(q)L(I)\subseteq(q) for an irreducible q∈Bq\in B. Choose the finitely many minimal primes P1,…,PrP_1,\ldots,P_r over II. Noetherianity gives an integer NN such that

(P1⋯Pr)N⊆I.(P_1\cdots P_r)^N\subseteq I.

Indeed, the product is contained in I\sqrt I, and a finitely generated ideal contained in I\sqrt I has some power contained in II. Leading coefficients multiply in the domain B[t]B[t], so

L(P1)N⋯L(Pr)N⊆L(I)⊆(q).L(P_1)^N\cdots L(P_r)^N\subseteq L(I)\subseteq(q).

Since (q)(q) is prime, L(P)⊆(q)L(P)\subseteq(q) for at least one of the PiP_i, which we now call PP. Put p=P∩B\p=P\cap B. Then p⊆(q)\p\subseteq(q). In a UFD the only primes contained in (q)(q) are (0)(0) and (q)(q): a nonzero prime contained in (q)(q) contains an irreducible factor of any of its nonzero elements, and that factor must be associated to qq.

If p=(q)\p=(q), every coefficient of every polynomial in PP lies in (q)(q). To see this, its leading coefficient lies in (q)(q) by assumption; subtract that leading term, which belongs to qB[t]⊆PqB[t]\subseteq P, and repeat. Hence P=qB[t]P=qB[t], contradicting the hypothesis on I⊆PI\subseteq P.

If p=(0)\p=(0), localize at B∖{0}B\setminus\{0\}. The nonzero prime PP becomes a nonzero proper prime of the PID Frac⁡(B)[t]\Frac(B)[t], and is therefore generated there by an irreducible polynomial. Clear denominators and divide out its content to obtain a primitive irreducible polynomial h∈B[t]h\in B[t]. Gauss’ lemma says

(hFrac⁡(B)[t])∩B[t]=hB[t].\big(h\Frac(B)[t]\big)\cap B[t]=hB[t].

A prime disjoint from the multiplicative set equals the contraction of its localization, so P=hB[t]P=hB[t]. Again I⊆PI\subseteq P has a common irreducible divisor, a contradiction. □

Lemma 11 (The monic trick over a PID). Let AA be a PID and I⊆A[x1,…,xn]I\subseteq A[x_1,\ldots,x_n] a nonzero ideal with no common irreducible divisor. If n≥1n\geq1, there are polynomial coordinates s1,…,sns_1,\ldots,s_n over AA in which II contains a polynomial monic in s1s_1.

Proof. For n=1n=1, the ideal L(I)⊆AL(I)\subseteq A from Lemma 10 has no common irreducible divisor. Since AA is a PID and L(I)≠0L(I)\neq0, it follows that L(I)=AL(I)=A. Thus II contains a monic polynomial.

Suppose n>1n>1 and put B=A[x1,…,xn−1]B=A[x_1,\ldots,x_{n-1}]. Regard II as an ideal in B[xn]B[x_n]. By the preceding lemma, L(I)L(I) has no common irreducible divisor in BB. By induction we may choose coordinates s1,…,sn−1s_1,\ldots,s_{n-1} for BB such that L(I)L(I) contains a polynomial

g=s1d+gd−1s1d−1+⋯+g0,gi∈A[s2,…,sn−1].g=s_1^d+g_{d-1}s_1^{d-1}+\cdots+g_0, \qquad g_i\in A[s_2,\ldots,s_{n-1}].

By the definition of L(I)L(I), choose

F=gxne+∑j<ehjxnj∈I,hj∈B.F=gx_n^e+\sum_{j<e}h_jx_n^j\in I, \qquad h_j\in B.

Choose a positive integer NN so large that

Ne+d>Nj+deg⁡s1hjNe+d>Nj+\deg_{s_1}h_j

for every nonzero hjh_j with j<ej<e. Set sn=xn−s1Ns_n=x_n-s_1^N, so xn=sn+s1Nx_n=s_n+s_1^N. These are still polynomial coordinates over AA. Written in these coordinates, the leading s1s_1-term of gxnegx_n^e is s1Ne+ds_1^{Ne+d} with coefficient 11, and every term coming from hjxnjh_jx_n^j has smaller s1s_1-degree. Therefore FF is monic in s1s_1. □

This is the UFD case of Suslin’s monic-polynomial argument; the more general formulation uses the height of an ideal. See Suslin, §6. The point of the preceding proof is that it applies to a PID without dividing by a nonunit coefficient.

Lemma 12. Let AA be a PID and m≥3m\geq3. Given a unimodular vector f∈A[x1,…,xn]mf\in A[x_1,\ldots,x_n]^m, with n≥1n\geq1, elementary operations and a change of polynomial coordinates produce a vector with a monic component.

Proof. Permute the components so that f1≠0f_1\neq0. If f1f_1 is a unit, adding a suitable multiple of f1f_1 to f2f_2 makes f2=1f_2=1, and we are done. Otherwise let p1,…,prp_1,\ldots,p_r be its distinct irreducible factors in the UFD A[x1,…,xn]A[x_1,\ldots,x_n].

For each jj such that pjp_j divides f2f_2, unimodularity supplies an index ij≥3i_j\geq3 for which pjp_j does not divide fijf_{i_j}. Replace f2f_2 by

f2+∑{j:pj∣f2}(∏k≠jpk)fij,f_2+\sum_{\{j:p_j\mid f_2\}} \left(\prod_{k\neq j}p_k\right)f_{i_j},

where the condition in the sum refers to the original f2f_2. This is a sequence of elementary operations. Modulo any pjp_j which originally divided f2f_2, exactly its designated summand is nonzero. Modulo any other pjp_j, all the added summands vanish and f2f_2 was already nonzero. Thus the new f2f_2 is divisible by none of the pjp_j, and f1,f2f_1,f_2 have no common irreducible divisor.

Apply Lemma 11 to the ideal (f1,f2)(f_1,f_2). After changing coordinates, there is a polynomial

h=af1+bf2h=af_1+bf_2

which is monic in s1s_1. Choose k≥0k\geq0 with k+deg⁡s1h>deg⁡s1f3k+\deg_{s_1}h>\deg_{s_1}f_3. Replacing f3f_3 by

f3+s1kaf1+s1kbf2f_3+s_1^kaf_1+s_1^kbf_2

makes the third component monic in s1s_1. Again, these are elementary operations. □

Vaserstein’s theorem

We can finally complete the induction. The proof is an induction on the number of variables, not on the length of the vector.

Theorem 13 (Vaserstein). Let AA be a PID and let f∈A[x1,…,xn]mf\in A[x_1,\ldots,x_n]^m be unimodular. Then ff has the UEP.

Proof. For n=0n=0, this is Proposition 3. Suppose the result holds for n−1n-1 variables. Vectors of length one or two can be completed over any ring, so assume m≥3m\geq3.

By Lemma 12, after elementary operations and a change of coordinates we may assume that a component of ff is monic in s1s_1, where

A[x1,…,xn]=A[s1,…,sn].A[x_1,\ldots,x_n]=A[s_1,\ldots,s_n].

Apply Proposition 9 with coefficient ring A[s2,…,sn]A[s_2,\ldots,s_n]. It gives

f(s1,s2,…,sn)∼f(0,s2,…,sn)f(s_1,s_2,\ldots,s_n)\sim f(0,s_2,\ldots,s_n)

over A[s1,…,sn]A[s_1,\ldots,s_n]. The vector on the right is unimodular over A[s2,…,sn]A[s_2,\ldots,s_n], so the induction hypothesis completes it to an invertible matrix over that ring. The same matrix remains invertible after adjoining s1s_1.

Hence our transformed vector is equivalent to e1e_1. Undo the elementary operations to obtain a completion of the original vector. A change of coordinates does not affect the conclusion: an isomorphism of polynomial rings sends an invertible matrix, together with its inverse and its specified first column, to an invertible matrix with the corresponding first column. □

Corollary 14 (Quillen–Suslin). Let AA be a PID. Then every finite projective module over A[x1,…,xn]A[x_1,\ldots,x_n] is free. Equivalently, every finite-rank vector bundle on AAn\mathbb A_A^n is trivial.

Proof. Vaserstein’s theorem says that A[x1,…,xn]A[x_1,\ldots,x_n] has the UEP, so Proposition 2 shows that every finite stably free module over it is free. By Theorem 14, every finite projective module over the same ring is stably free. Combining the two statements proves the theorem. □

The two parts of the argument do rather different jobs. Serre’s reduction produces a stable basis; the unimodular-vector argument removes the additional free summands. Horrocks’ theorem and the translation ideal are what let us carry this cancellation from a local ring back to the polynomial ring itself.

For the paper version, including the geometric background and further applications, see Serre Schemes and the Quillen–Suslin Theorem.

Back to the blog →