betaFeedback

WorksMathematics

Vasconcelos' conjecture on the module of differentials

Abstract

Let kk be a field and let SS be a local kk-algebra essentially of finite type whose module of Kähler differentials ΩS/k\Omega_{S/k} has finite projective dimension. A conjecture of Vasconcelos, in the form studied by Herzog and Briggs, asserts that SS is then a complete intersection when char⁡k=0\operatorname{char}k=0. We give a complete proof, in particular one independent of the Eisenbud–Mazur conjecture, that SS is a reduced complete intersection when char⁡k=0\operatorname{char}k=0, and, over an arbitrary field, when some minimal prime of SS is a generic point at which SS is a field separably generated over kk; in the latter case pd⁡SΩS/k≤1\operatorname{pd}_S\Omega_{S/k}\le1 and SS is essentially smooth at all of its generic points. In positive characteristic pp, Vasconcelos' conjecture for reduced algebras holds for every such algebra exactly when [k:kp]≤p[k:k^p]\le p.

Introduction

Let kk be a field and let SS be a local ring essentially of finite type over kk, that is, a localization of a finitely generated kk-algebra at a prime ideal. Vasconcelos conjectured that the projective dimension pd⁡SΩS/k\operatorname{pd}_S \Omega_{S/k} of the module of Kähler differentials can only take the values 00, 11 and ∞\infty, for affine algebras over an arbitrary field and leaving questions of separability aside [12 §4, (C2)]. In the form studied by Herzog [7 §1, (C2)] and Briggs [4], for char⁡k=0\operatorname{char} k = 0, the conjecture asserts that

if pd⁡SΩS/k<∞\operatorname{pd}_S \Omega_{S/k} < \infty, then SS is a complete intersection. (C2)

See the introduction of [4] for the history of this conjecture and of its companion [12 §0, (C1)]: for an ideal II of finite projective dimension in a local ring RR,

if pd⁡R/I(I/I2)<∞\operatorname{pd}_{R/I}(I/I^2) < \infty, then II is generated by a regular sequence. (C1)

In characteristic zero, the graded cases of both conjectures were settled by Avramov and Herzog, (C1) in [AH94, Thm. 3] and (C2) in [AH94, Thm. 1]. Briggs proved (C1) in full [4 Thm. A]. For the module of differentials, Briggs proved the conjecture for algebras without non-trivial evolutions [4 Thm. 3.4], so that in characteristic zero it would follow from the Eisenbud–Mazur conjecture on evolutions and symbolic squares. Since the Eisenbud–Mazur conjecture has been shown to fail [9], that reduction does not prove Herzog’s (C2), and a different argument is needed.

In this paper we prove the conjecture for ordinary Kähler differentials in characteristic zero, and we determine exactly what survives in positive characteristic.

Throughout, (R,n,κ)(R,\mathfrak{n},\kappa) denotes a local ring RR with maximal ideal n\mathfrak{n} and residue field κ\kappa, and κ(p)\kappa(\mathfrak{p}) is the residue field of a ring at a prime ideal p\mathfrak{p}. For a finite module MM over RR we write pd⁡RM\operatorname{pd}_{R} M for its projective dimension, μR(M)\mu_{R}(M) for its minimal number of generators and Ass⁡RM\operatorname{Ass}_{R} M for its set of associated primes; ht⁡I\operatorname{ht} I is the height of an ideal II, and grade⁡I\operatorname{grade} I is the common length of the maximal RR-regular sequences in a proper ideal II [8 Thm. 16.7]. If RR is Cohen–Macaulay, for instance regular, then [8 Thms. 17.3 and 17.4]

grade⁡I=ht⁡Ianddepth⁡Rp=ht⁡p\operatorname{grade} I=\operatorname{ht} I \qquad\text{and}\qquad\operatorname{depth} R_{\mathfrak{p}}=\operatorname{ht}\mathfrak{p}

for every prime ideal p\mathfrak{p}.

A ring is non-trivial if it is not the zero ring. A generic point of a ring is a minimal prime ideal of it.

In positive characteristic pp we write Fp\mathbb{F}_{p} for the prime field and kpk^{p} for the subfield of pp-th powers of a field kk.

A local kk-algebra is essentially smooth if it is a localization of a smooth kk-algebra, and a local kk-algebra essentially of finite type is called a complete intersection over kk if it is of the form R/(g1,…,gc)R/(g_{1},\ldots,g_{c}) with RR a localization of a polynomial ring over kk at a prime ideal and g1,…,gcg_{1},\ldots,g_{c} an RR-regular sequence. Whether a given presentation S=R/IS=R/I with RR regular has II generated by a regular sequence does not depend on the presentation [8 Thm. 21.2(ii)].

Theorem 1.1. Let kk be a field of characteristic zero and let SS be a non-trivial local kk-algebra essentially of finite type. If pd⁡SΩS/k<∞\operatorname{pd}_{S}\Omega_{S/k}<\infty, then

  • SS is a reduced complete intersection over kk, and

  • pd⁡SΩS/k≤1\operatorname{pd}_{S}\Omega_{S/k}\leq1.

Reducedness is a conclusion here, not a hypothesis. In positive characteristic the conclusion can fail even for domains, and the correct statement involves the generic points of SS.

Theorem 1.2. Let kk be any field and let SS be a non-trivial local kk-algebra essentially of finite type with pd⁡SΩS/k<∞\operatorname{pd}_{S}\Omega_{S/k}<\infty. Suppose that for some minimal prime η\eta of SS the local ring SηS_{\eta} is a field which is separably generated over kk. Then

  • SS is a reduced complete intersection over kk,

  • the ring Sη′S_{\eta'} is essentially smooth over kk for every minimal prime η′\eta' of SS, and

  • pd⁡SΩS/k≤1\operatorname{pd}_{S}\Omega_{S/k}\leq1.

Corollary 1.3. Let kk be a perfect field and let SS be a non-trivial local kk-algebra essentially of finite type with pd⁡SΩS/k<∞\operatorname{pd}_{S}\Omega_{S/k}<\infty. If the local ring SηS_{\eta} is reduced for some minimal prime η\eta of SS, then

  • SS is a reduced complete intersection over kk,

  • the ring Sη′S_{\eta'} is essentially smooth over kk for every minimal prime η′\eta' of SS, and

  • pd⁡SΩS/k≤1\operatorname{pd}_S \Omega_{S/k} \leq1.

In particular every reduced such SS is a complete intersection over kk.

For reduced algebras the dividing line is the imperfection degree [k:kp][k:k^p] of the base field kk.

Theorem 1.4. Let kk be a field of positive characteristic pp. The following are equivalent.

(1) [k:kp]≤p[k:k^p] \leq p.

(2) Every non-trivial reduced local kk-algebra SS essentially of finite type with pd⁡SΩS/k<∞\operatorname{pd}_S \Omega_{S/k} < \infty is a complete intersection over kk.

Outline of the proof

Outline of the proof. The proof descends in three steps. Differentials occur only in the first step.

Step 1: from differentials to a quotient of the conormal module. Present SS as R/IR/I with RR a localization of a polynomial ring in vv variables over kk. The conormal sequence

I/I2→δS⊗RΩR/k≅Sv⟶ΩS/k⟶0I/I^2 \xrightarrow{\delta} S \otimes_R \Omega_{R/k} \cong S^v \longrightarrow\Omega_{S/k} \longrightarrow0

shows that the image δ(I/I2)\delta(I/I^2) of δ\delta, the module generated by the rows of the Jacobian matrix, is a quotient of I/I2I/I^2 with pd⁡Sδ(I/I2)<∞\operatorname{pd}_S \delta(I/I^2) < \infty exactly when pd⁡SΩS/k<∞\operatorname{pd}_S \Omega_{S/k} < \infty. At a generic point of SS whose local ring is a separably generated field, the Jacobian criterion says that δ(I/I2)\delta(I/I^2) has rank at least ht⁡I\operatorname{ht} I; hence δ(I/I2)\delta(I/I^2) needs at least ht⁡I\operatorname{ht} I generators. The conjecture is thereby reduced to a statement in which differentials do not appear: a quotient of I/I2I/I^2 of finite projective dimension cannot have more than ht⁡I\operatorname{ht} I generators, and if it has that many, then II is generated by a regular sequence.

To state it, let (R,n,κ)(R,\mathfrak n,\kappa) be a local ring, I⊊RI \subsetneq R an ideal and S=R/IS=R/I. Let

σ:I/I2⟶N\sigma:I/I^2 \longrightarrow N

be an SS-linear map into a finitely generated SS-module, and for g∈Ig\in I write [g][g] for the image of σ(gˉ)\sigma(\bar g) in N/nNN/\mathfrak nN, where gˉ\bar g is the class of gg in I/I2I/I^2. The statement does not require RR to be regular: it holds whenever SS has finite projective dimension over RR, with the grade of II in place of its height.

Theorem 1.5. Let (R,n,κ)(R,\mathfrak n,\kappa) be a Noetherian local ring and I⊊RI \subsetneq R an ideal such that S=R/IS=R/I has finite projective dimension over RR. Let σ:I/I2↠N\sigma:I/I^2 \twoheadrightarrow N be a surjective SS-linear map onto a finitely generated SS-module with pd⁡SN<∞\operatorname{pd}_S N<\infty.

(1) μS(N)≤grade⁡I\mu_S(N) \leq\operatorname{grade} I.

(2) If μS(N)≥grade⁡I\mu_S(N) \geq\operatorname{grade} I, then equality holds and II is generated by an RR-regular sequence whose images form a basis of N/nNN/\mathfrak nN.

(3) If μS(N)=grade⁡I−1\mu_S(N)=\operatorname{grade} I-1 and depth⁡Rp=grade⁡I\operatorname{depth} R_{\mathfrak p}=\operatorname{grade} I for every associated prime p\mathfrak p of the RR-module SS, then II is generated by an RR-regular sequence.

If RR is regular, then pd⁡RS<∞\operatorname{pd}_R S<\infty by Serre’s theorem [10 Tag 00O7], grade⁡I=ht⁡RI\operatorname{grade} I=\operatorname{ht}_R I and depth⁡Rp=ht⁡Rp\operatorname{depth} R_{\mathfrak p}=\operatorname{ht}_R\mathfrak p; in parts (2) and (3) the ring SS is then a complete intersection. Step 1 applies the theorem in this case to the image N=δ(I/I2)N=\delta(I/I^2), with σ\sigma induced by δ\delta. For N=I/I2N=I/I^2 and σ\sigma the identity one has

μS(N)=μR(I)≥ht⁡RI≥grade⁡I,\mu_S(N)=\mu_R(I)\geq\operatorname{ht}_R I\geq\operatorname{grade} I,

so part (2) recovers the theorem of Briggs [4 Thm. A], that is, (C1): if pd⁡RS<∞\operatorname{pd}_R S<\infty and pd⁡S(I/I2)<∞\operatorname{pd}_S(I/I^2)<\infty, then II is generated by a regular sequence. For regular RR and a free module NN of rank ht⁡RI\operatorname{ht}_R I, part (2) is the classical criterion of Vasconcelos on free conormal summands [Vas68, §2, Prop.]. The Jacobian image is a priori neither free nor all of I/I2I/I^2, and this is why the general statement is needed.

Granting Theorem 1.5, part (2) shows that SS is a complete intersection in the situation of Theorem 1.2. Only afterwards do we obtain injectivity of the conormal map, the bound pd⁡SΩS/k≤1\operatorname{pd}_S\Omega_{S/k}\leq1, and reducedness together with smoothness at the generic points, all from one computation at the associated primes of SS. There the module of differentials, having finite projective dimension over a ring of depth zero, is free of the same rank as at the given generic point. Hence the Jacobian matrix of the regular generators of II has an invertible maximal minor, SS is smooth over kk there, and a regular local ring of depth zero is a field. In characteristic zero the hypothesis on the generic point holds automatically: the module of differentials is free at every generic point, and in characteristic zero this already makes the local ring regular. Thus Theorem 1.1 is a corollary of Theorem 1.2. Over an imperfect field with [k:kp]≤p[k:k^p]\leq p the rank of NN at a generic point may drop by one, and part (3) takes over. Over every field with [k:kp]>p[k:k^p]>p a one-dimensional local domain with free differentials which is not a complete intersection is obtained from the monomial curve (t3,t4,t5)(t^3,t^4,t^5) by a purely inseparable change of coordinates (Proposition 7.2).

Step 2: from a quotient of the conormal module to a change of rings. To prove Theorem 1.5 one chooses an RR-regular sequence g1,…,gℓg_1,\ldots,g_\ell in II such that [g1],…,[gℓ][g_1],\ldots,[g_\ell] are linearly independent and which cannot be prolonged by an element of II with these two properties, and passes to

Q=R/(g1,…,gℓ),J=I/(g1,…,gℓ),Q/J=S.Q=R/(g_1,\ldots,g_\ell),\qquad J=I/(g_1,\ldots,g_\ell),\qquad Q/J=S.

The residual ideal JJ has grade

grade⁡J=grade⁡I−ℓ.\operatorname{grade} J=\operatorname{grade} I-\ell.

If SS has finite projective dimension over QQ, two elementary facts finish the proof. First, an ideal of grade zero with a cyclic quotient of finite projective dimension is zero; so JJ contains a nonzerodivisor unless J=0J=0, and prime avoidance shows that [g1],…,[gℓ][g_1],\ldots,[g_\ell] span N/nNN/\mathfrak nN, for otherwise the sequence could be prolonged. Second, an ideal of grade one with such a quotient is principal if QQ has depth one at every associated prime of the quotient. The second fact rests on the factorization of ideals with finite free resolutions due to Buchsbaum and Eisenbud [2 Cor. 5.2]; this is the only result beyond the textbooks which is quoted in this paper, and it is needed only for part (3)(3), that is, only for Theorem 1.4. Everything therefore depends on transferring finite projective dimension from RR to QQ.

Theorem 1.6 (change of rings along a regular sequence). Let (R,n,κ)(R,\mathfrak{n},\kappa) be a Noetherian local ring and I⊊RI \subsetneq R an ideal such that S=R/IS=R/I has finite projective dimension over RR; this holds for every II when RR is regular. Let σ:I/I2→N\sigma:I/I^2 \to N be an SS-linear map into a finitely generated SS-module with pd⁡SN<∞\operatorname{pd}_S N<\infty. If g1,…,gt∈Ig_1,\ldots,g_t\in I is an RR-regular sequence such that [g1],…,[gt][g_1],\ldots,[g_t] are linearly independent over κ\kappa, where [g][g] denotes the image of σ(g‾)\sigma(\overline{g}) in N/nNN/\mathfrak{n}N, then

pd⁡R/(g1,…,gt)S≤pd⁡RS−t<∞.\operatorname{pd}_{R/(g_1,\ldots,g_t)} S \le\operatorname{pd}_R S-t<\infty.

The regular sequence in this theorem is chosen in advance and is not extracted from homological data, and neither freeness of NN nor injectivity of σ\sigma is assumed. This is the main technical result of the paper.

Step 3: the change of rings. Theorem 1.6 is proved by removing one regular element at a time, in the residue algebra B=κ⊗RRB=\kappa\otimes_R\mathcal{R} of a free DG algebra resolution R\mathcal{R} of SS over RR which has a variable eje_j of degree one for each generator gjg_j of II. A finite free resolution of NN, of length rr, yields for every SS-linear functional λ:N→κ\lambda:N\to\kappa a derivation DλD_\lambda of BB of degree −1-1 with Dλ(ej)=λσ(gj‾)D_\lambda(e_j)=\lambda\sigma(\overline{g_j}) whose commutator with the differential takes values in the ideal generated by the variables of degree less than rr. This alone makes the commutator nilpotent in each degree, within a range depending only on rr (Proposition 3.1); the construction originates in the proofs of [4 Lem. 2.6 and Thm. 2.7]. A derivation DD of this kind with D(e1)=1D(e_1)=1 forces the top group Tor⁡pd⁡RSR/(g1)(S,κ)\operatorname{Tor}^{R/(g_1)}_{\operatorname{pd}_R S}(S,\kappa) to vanish, so the projective dimension drops by one (Lemma 4.1). The derivations dual to the selected elements descend to the successive quotients, and the step can be repeated tt times without changing NN. The argument is self-contained; in particular, the Eisenbud–Mazur conjecture is not used.

Organization

The sections follow the three steps in reverse order, so that every result is proved before it is used.

Section 22 prepares Step 33. It fixes the DG algebra resolution R\mathcal{R} of SS and its residue algebra BB, and shows that a free DG module with triangular differential over R\mathcal{R} resolves its reduction (Lemma 2.1).

Section 33 carries out the first half of Step 33. It constructs the derivation DλD_\lambda of BB from a finite free resolution of NN and proves that its commutator with the differential is nilpotent within a range depending only on rr (Proposition 3.1).

Section 44 completes Step 33. One regular element is removed in Lemma 4.1, and repeating this along the prescribed sequence proves Theorem 1.6.

Section 5 carries out Step 2. It treats the residual ideal in grades zero and one (Lemmas 5.1 and 5.2) and proves Theorem 1.5 with a maximal regular sequence whose images in N/nNN/\mathfrak{n}N are linearly independent.

Section 6 carries out Step 1. It computes the rank of the Jacobian image at a generic point (Lemma 6.1) and proves Theorem 1.2, Theorem 1.1 and Corollary 1.3.

Section 7 carries out Step 1 in positive characteristic. Theorem 1.5(3) gives Corollary 7.1 and the implication (1)⇒(2)(1)\Rightarrow(2) of Theorem 1.4; Proposition 7.2 gives the converse.

Declaration of generative AI: As the outcome of about three weeks of intensive discussion between the second author and a ChatGPT model, a proof was proposed which rests on two deep results, Briggs’ theorem on radical elements of the homotopy Lie algebra [4 Thm. 2.7] and a theorem of Briggs, Grifo and Pollitz on cohomological support varieties [3]. A careful study of their proofs showed which parts of them the argument actually needs, and this led to the comparatively elementary proof given here, which uses neither result. The authors reviewed and edited the content as needed and take full responsibility for the content of the article.

Conventions and resolutions

This section fixes conventions and constructs the resolution of SS in which the computations of § 3 and § 4 take place. All rings are commutative and Noetherian. In a minimal free resolution of a finite module MM over a local ring (T,m,κ)(T,\mathfrak{m},\kappa) all differentials vanish after tensoring with κ\kappa; therefore, if Tor⁡qT(M,κ)=0\operatorname{Tor}^{T}_{q}(M,\kappa)=0 for some q≥1q\geq1, then pd⁡TM≤q−1\operatorname{pd}_{T}M\leq q-1.

DG algebras and modules

Let S=R/IS=R/I be a quotient of a local ring (R,n,κ)(R,\mathfrak{n},\kappa) and let g∈Ig\in I. Finite projective dimension of SS over the quotient R/(g)R/(g) will be detected on Tor⁡R/(g)(S,κ)\operatorname{Tor}^{R/(g)}(S,\kappa). To compare it with Tor⁡R(S,κ)\operatorname{Tor}^{R}(S,\kappa) we resolve SS over RR by an algebra in which gg is the differential of a variable ee; dividing out gg and ee then yields a resolution over R/(g)R/(g). Resolutions of this kind were introduced by Tate [11 §2]. The multiplication is what makes derivations available, and a derivation of a free algebra is determined by its values on the variables.

We use homological grading, with differentials of degree −1-1. Every differential is denoted by ∂\partial, the degree of a homogeneous element aa by ∣a∣|a|, and the component of degree kk of a graded object MM by MkM_k; H(M)H(M) is the homology of MM in all degrees together. A DG algebra is strictly graded-commutative and concentrated in nonnegative degrees. A DG module PP over a DG algebra T\mathcal{T} satisfies

∂(ap)=∂(a)p+(−1)∣a∣a∂(p),\partial(ap)=\partial(a)p+(-1)^{|a|}a\partial(p),

and we put

pa=(−1)∣a∣∣p∣ap.pa=(-1)^{|a||p|}ap.

A derivation of degree −1-1 from T\mathcal{T} to PP is a map β\beta, linear over the ring T0\mathcal{T}_0, which lowers degrees by one and satisfies

β(ab)=β(a)b+(−1)∣a∣aβ(b).\beta(ab)=\beta(a)b+(-1)^{|a|}a\beta(b).

For a ring TT and a set YY of variables of positive degrees, T[Y]T[Y] denotes the free strictly graded-commutative TT-algebra on YY: polynomial in the variables of even degree and exterior in those of odd degree. A monomial is a product of variables; the monomials form a TT-basis of T[Y]T[Y]. A derivation of degree −1-1 from T[Y]T[Y] to a graded module is determined by its values on YY, and these may be prescribed arbitrarily in the right degrees.

A free resolution by a DG algebra

Let (R,n,κ)(R,\mathfrak n,\kappa) be a local ring, I⊊RI\subsetneq R an ideal with generators g1,…,gng_1,\ldots,g_n, not necessarily minimal, and S=R/IS=R/I. Let R=R[X]\mathcal{R}=R[X] be a free DG RR-algebra on variables X=X1⊔X2⊔⋯X=X_1\sqcup X_2\sqcup\cdots, with XkX_k in degree kk, such that

X1={e1,…,en},∂ej=gj(1≤j≤n),X_1=\{e_1,\ldots,e_n\},\qquad\partial e_j=g_j\quad(1\leq j\leq n),
H0(R)=S,Hk(R)=0(k≥1).(1)H_0(\mathcal{R})=S,\qquad H_k(\mathcal{R})=0\quad(k\geq1). \tag*{(1)}

Such an algebra exists [11 Thm. 1]: start from the Koszul complex R[X1]R[X_1] and, for k=1,2,…k=1,2,\ldots in turn, adjoin variables of degree k+1k+1 whose differentials are cycles representing generators of the RR-module HkH_k of the algebra built so far. The elements of degree at most kk are unchanged by this step, so the homology in degrees below kk is unchanged and HkH_k becomes zero. Each Rk\mathcal{R}_k is a free RR-module and R0=R\mathcal{R}_0=R, so R\mathcal{R} is a free resolution of SS over RR. We put

B=κ⊗RR=κ[X],Hj(B)=Tor⁡jR(S,κ).B=\kappa\otimes_R\mathcal{R}=\kappa[X],\qquad H_j(B)=\operatorname{Tor}^{R}_j(S,\kappa).

Then B0=κB_0=\kappa, and every eje_j is a cycle of BB because gj∈ng_j\in\mathfrak n. For g∈Ig\in I we write gˉ\bar{g} for its class in I/I2I/I^2. When an SS-linear map σ:I/I2→N\sigma:I/I^2\to N into a finitely generated SS-module is given, [g][g] denotes the image of σ(gˉ)\sigma(\bar{g}) in N/nNN/\mathfrak nN, as in the Introduction.

A DG module as in the next lemma is built by adjoining free generators one degree at a time, and such a module cannot distinguish quasi-isomorphic algebras: this is the DG form of the fact that tensoring with a bounded below complex of free modules preserves quasi-isomorphisms. We use the lemma twice: to lift a finite resolution over SS to a DG module over RR, and to compute Tor⁡\operatorname{Tor} over a quotient of RR by a regular element.

Lemma 2.1. Let T→U\mathcal{T}\to U be a surjective quasi-isomorphism of DG algebras. Let MM be a DG T\mathcal{T}-module which is graded free on homogeneous basis elements of nonnegative degrees, such that the differential of every basis element is a T\mathcal{T}-linear combination of basis elements of smaller degree. Then M→U⊗TMM\to U\otimes_{\mathcal{T}}M is a surjective quasi-isomorphism.

Proof. Let k=ker⁡(T→U)\mathfrak{k}=\ker(\mathcal{T}\to U), an acyclic DG ideal. The kernel of M→U⊗TMM\to U\otimes_{\mathcal{T}}M is the direct sum of the kv\mathfrak{k}v over the basis elements vv, and it suffices to show that it is acyclic. Let zz be a homogeneous cycle in it, let qq be the largest degree of a basis element occurring in zz, and write the corresponding part of zz as ∑∣v∣=qavv\sum_{|v|=q}a_vv with av∈ka_v\in\mathfrak{k}. Since ∂v\partial v involves only basis elements of degree less than qq, the coefficient of vv in ∂z=0\partial z=0 is ∂av\partial a_v; hence ∂av=0\partial a_v=0. Choose bv∈kb_v\in\mathfrak{k} with ∂bv=av\partial b_v=a_v. Then z−∂(∑∣v∣=qbvv)z-\partial(\sum_{|v|=q}b_vv) is a cycle in the kernel which involves only basis elements of degree less than qq. As zz involves finitely many basis elements, of nonnegative degrees, repeating this step exhibits zz as a boundary. □

Derivations from a finite free resolution

This section constructs, from a finite free resolution of NN, the derivations with which a regular element will be divided out in § 4.

Motivation

3.1. Motivation. Keep the notation of § 2.2, let σ:I/I2→N\sigma:I/I^2\to N be an SS-linear map into a finitely generated SS-module with pd⁡SN<∞\operatorname{pd}_S N<\infty, and suppose that g=g1g=g_1 is a nonzerodivisor of RR with [g]≠0[g]\ne0. Put Q=R/(g)Q=R/(g). We want to show that pd⁡QS<∞\operatorname{pd}_Q S<\infty.

The prototype is the argument of Vasconcelos for a free module NN [Vas68, §2, Prop.]. The homomorphism S→I/gIS\to I/gI, 1↦g1\mapsto g, is well defined. Since gg is a nonzerodivisor on RR and on II, a finite free resolution of II over RR reduces modulo gg to a free resolution of I/gII/gI over QQ, so

pd⁡Q(I/gI)≤pd⁡RI=pd⁡RS−1.\operatorname{pd}_Q(I/gI)\le\operatorname{pd}_R I=\operatorname{pd}_R S-1.

If NN is free, there is an SS-linear map N→SN\to S taking σ(gˉ)\sigma(\bar{g}) to 11, and composing it with I/gI→I/I2→σNI/gI\to I/I^2\xrightarrow{\sigma}N splits the map S→I/gIS\to I/gI. Thus SS is a direct summand of I/gII/gI over QQ, and pd⁡QS≤pd⁡RS−1\operatorname{pd}_Q S\le\operatorname{pd}_R S-1. When NN is not free there is no such splitting: the only functionals at our disposal are the maps λ:N→κ\lambda:N\to\kappa with λσ(gˉ)=1\lambda\sigma(\bar{g})=1. The construction of this section is what remains of the splitting over the residue field.

In terms of resolutions, we want to compare the homology of BB, which is Tor⁡R(S,κ)\operatorname{Tor}^R(S,\kappa) and is bounded when pd⁡RS<∞\operatorname{pd}_R S<\infty, with the homology of C=B/(e1)C=B/(e_1), which will turn out to be Tor⁡Q(S,κ)\operatorname{Tor}^Q(S,\kappa) (Lemma 4.1). As a graded algebra BB is the tensor product over κ\kappa of CC and the exterior algebra on e1e_1, and the simplest way to pass from BB to CC is to contract against e1e_1: one looks for a derivation DD of BB of degree −1-1 with D(e1)=1D(e_1)=1. If DD commuted with the differential in the graded sense, that is, if

θ=∂D+D∂\theta=\partial D+D\partial

vanished, then BB would split as a complex into CC and a shifted copy of CC, and boundedness would pass from H(B)H(B) to H(C)H(C) at once. This is the chain-level form of the splitting above, and it is what happens when NN is free. In general θ≠0\theta\ne0. The map which θ\theta induces on H(C)H(C) lowers degrees by two; it plays the role, on Tor, of the operators of degree two attached to the hypersurface R/(g)R/(g) by Gulliksen [Gul74, Thm. 3.1] and Eisenbud [5 §1], for which finite projective dimension over R/(g)R/(g) is a nilpotence property. We use this comparison only as a guide and do not need to make it precise. The task is to choose DD in such a way that θ\theta is visibly nilpotent in each degree. The values D(ej)D(e_j) are forced on us only through a functional on the conormal module. Indeed, a variable xx of degree two has ∂x=∑j=1najej\partial x = \sum_{j=1}^{n} a_j e_j with ∑j=1najgj=0\sum_{j=1}^{n} a_j g_j = 0, and for θ(x)\theta(x) to have no component in degree zero one needs

∑j=1naˉjD(ej)=0,\sum_{j=1}^{n} \bar{a}_j D(e_j) = 0,

where aˉj\bar{a}_j is the image of aja_j in κ\kappa; so gj↦D(ej)∈κg_j \mapsto D(e_j) \in\kappa should come from an SS-linear map I/I2→κI/I^2 \to\kappa. Extending DD to the higher variables of R\mathcal{R} means solving one equation per variable, and over κ\kappa these equations have no reason to be solvable. The idea, which goes back to the proofs of [4 Lem. 2.6 and Thm. 2.7], is to solve them one level higher. If the functional factors as λσ\lambda\sigma through the module NN, with a finite free resolution GG of length rr, one first lifts σ\sigma to a derivation β\beta from R\mathcal{R} into a DG module PP lifting GG; there the equations are solvable because PP is a resolution of NN. One then composes with the projection of PP onto κ\kappa given by λ\lambda. The only step which does not commute with differentials is this last projection, and it fails to commute only through basis elements of GG of degree at least two, with defects of positive degree at most r−1r-1. Thus θ\theta takes values in the ideal generated by the variables of degree less than rr, and on each of these variables a bounded number of applications of θ\theta reaches degree zero. This is all that is needed for nilpotence.

The construction

Proposition 3.1. Let (R,n,κ)(R,\mathfrak{n},\kappa), I=(g1,…,gn)I=(g_1,\ldots,g_n), S=R/IS=R/I, R\mathcal{R} and BB be as in § 2.2. Let σ:I/I2→N\sigma:I/I^2\to N be an SS-linear map into an SS-module NN which has a resolution G→NG\to N by finitely generated free SS-modules G0,…,GrG_0,\ldots,G_r. For every SS-linear map λ:N→κ\lambda:N\to\kappa there is a derivation Dλ:B→BD_\lambda:B\to B of degree −1-1 such that

Dλ(ej)=λσ(gˉj)(1≤j≤n)(2)D_\lambda(e_j)=\lambda\sigma(\bar{g}_j) \qquad(1\leq j\leq n) \tag*{(2)}

and such that the commutator

θλ=[∂,Dλ]=∂Dλ+Dλ∂\theta_\lambda=[\partial,D_\lambda]=\partial D_\lambda+D_\lambda\partial

satisfies

θλu(Bk+2u)=0(3)\theta_\lambda^u(B_{k+2u})=0 \tag*{(3)}

for all integers k≥0k\geq0 and u≥1u\geq1 with

u>k⌊r/2⌋.u>k\left\lfloor r/2\right\rfloor.

Moreover θλ\theta_\lambda is a chain map and a derivation of BB of degree −2-2, its image is contained in the ideal of BB generated by the variables of degrees 1,…,r−11,\ldots,r-1, and θλ=0\theta_\lambda=0 if r≤1r\leq1.

Proof. Write ρ:G0→N\rho: G_{0} \to N for the augmentation and choose bases of the GkG_{k}.

Step 1: lifting the resolution. We construct a DG R\mathcal{R}-module PP, graded free on a finite basis Π=Π0⊔⋯⊔Πr\Pi= \Pi_{0} \sqcup\cdots\sqcup\Pi_{r} with Πk\Pi_{k} in degree kk and in bijection with the chosen basis of GkG_{k}, such that the differential of each element of Πk\Pi_{k} involves only Π0,…,Πk−1\Pi_{0}, \ldots, \Pi_{k-1} and such that S⊗RP=GS \otimes_{\mathcal{R}} P = G. Put ∂p=0\partial p = 0 for p∈Π0p \in\Pi_{0}. Suppose that the differential has been defined on Π0,…,Πk−1\Pi_{0}, \ldots, \Pi_{k-1}, so that the free module P<kP^{<k} on these elements is a DG module with S⊗RP<k=G<kS \otimes_{\mathcal{R}} P^{<k} = G_{<k}, the part of GG in degrees less than kk. By Lemma 2.1, applied to the augmentation R→S\mathcal{R} \to S, the projection P<k→G<kP^{<k} \to G_{<k} is a surjective quasi-isomorphism. For p∈Πkp \in\Pi_{k} with image pˉ∈Gk\bar{p} \in G_{k}, the element ∂pˉ∈Gk−1\partial\bar{p} \in G_{k-1} is a cycle. Lift it to an element of P<kP^{<k} of degree k−1k-1; the differential of this lift lies in the acyclic kernel of the projection and is a cycle there, so after subtracting a primitive in the kernel we obtain a cycle zp∈P<kz_{p} \in P^{<k} lifting ∂pˉ\partial\bar{p}. Put ∂p=zp\partial p = z_{p}. This defines PP. By Lemma 2.1 again, φ:P→G\varphi: P \to G is a surjective quasi-isomorphism, hence so is

ε=ρφ:P⟶N.\varepsilon= \rho\varphi: P \longrightarrow N.

Only the basis of PP is finite; PP itself is not bounded.

Step 2: lifting the conormal map. We construct a derivation β:R→P\beta: \mathcal{R} \to P of degree −1-1 with

∂β=−β∂,εβ(ej)=σ(gˉj)(1≤j≤n).(4)\partial\beta= -\beta\partial,\qquad\varepsilon\beta(e_{j}) = \sigma(\bar{g}_{j})\quad(1 \leq j \leq n). \tag*{(4)}

Choose β(ej)∈P0\beta(e_{j}) \in P_{0} with εβ(ej)=σ(gˉj)\varepsilon\beta(e_{j}) = \sigma(\bar{g}_{j}). The first identity holds on eje_{j} because ∂P0=0\partial P_{0} = 0 and β\beta vanishes on R0=R\mathcal{R}_{0} = R. Let x∈Xkx \in X_{k} with k≥2k \geq2, and suppose that β\beta has been defined on the variables of degree less than kk and satisfies ∂β=−β∂\partial\beta= -\beta\partial on the subalgebra they generate, which contains ∂x\partial x. Then ∂β(∂x)=−β(∂2x)=0\partial\beta(\partial x) = -\beta(\partial^{2}x) = 0, so β(∂x)\beta(\partial x) is a cycle of PP of degree k−2k-2. If k≥3k \geq3 it is a boundary because Hk−2(P)=0H_{k-2}(P) = 0. If k=2k = 2, write ∂x=∑j=1najej\partial x = \sum_{j=1}^{n} a_{j}e_{j} with aj∈Ra_{j} \in R; then ∑j=1najgj=0\sum_{j=1}^{n} a_{j}g_{j} = 0 and

εβ(∂x)=∑j=1najσ(gˉj)=σ(∑j=1najgj)=0,\varepsilon\beta(\partial x) = \sum_{j=1}^{n} a_{j}\sigma(\bar{g}_{j}) = \sigma\left(\sum_{j=1}^{n} a_{j}g_{j}\right) = 0,

so β(∂x)\beta(\partial x) is again a boundary, since ε\varepsilon induces H0(P)≅NH_{0}(P) \cong N. Choose β(x)\beta(x) with ∂β(x)=−β(∂x)\partial\beta(x) = -\beta(\partial x). Since ∂β+β∂\partial\beta+ \beta\partial is additive, RR-linear and satisfies the product rule ab↦(∂β+β∂)(a)b+a(∂β+β∂)(b)ab \mapsto(\partial\beta+ \beta\partial)(a)b + a(\partial\beta+ \beta\partial)(b), it vanishes on R\mathcal{R} once it vanishes on the variables. The map β\beta does not depend on λ\lambda.

Step 3: projecting to κ\kappa. Put Pˉ=κ⊗RP\bar{P} = \kappa\otimes_{R} P, a graded free BB-module on Π\Pi, and βˉ=κ⊗Rβ\bar{\beta} = \kappa\otimes_{R} \beta. Let γλ:Pˉ→B\gamma_{\lambda}: \bar{P} \to B be the BB-linear map of degree zero with

γλ(p)={λρ(pˉ)if p∈Π0,0if p∈Πk, k≥1,\gamma_{\lambda}(p) = \begin{cases} \lambda\rho(\bar{p}) & \text{if } p \in\Pi_{0},\\ 0 & \text{if } p \in\Pi_{k},\ k \geq1, \end{cases}

and put Dλ=γλβˉD_{\lambda} = \gamma_{\lambda}\bar{\beta}. Then DλD_{\lambda} is a derivation of degree −1-1, and (3.1) follows from (3.3). By (3.3),

θλ=∂γλβˉ+γλβˉ∂=αλβˉ,αˉλ=∂γλ−γλ∂.(5)\theta_{\lambda} = \partial\gamma_{\lambda}\bar{\beta} + \gamma_{\lambda}\bar{\beta}\partial= \alpha_{\lambda}\bar{\beta},\qquad\bar{\alpha}_{\lambda} = \partial\gamma_{\lambda} - \gamma_{\lambda}\partial. \tag*{(5)}

The defect αλ\alpha_\lambda satisfies αλ(ap)=(−1)∣a∣aαλ(p)\alpha_\lambda(ap)=(-1)^{|a|}a\alpha_\lambda(p) for a∈Ba\in B. It vanishes on Π0\Pi_0. On p∈Π1p\in\Pi_1 its value is

αλ(p)=−γλ(∂p)=−λρ(∂p‾)=0,\alpha_\lambda(p)=-\gamma_\lambda(\partial p)=-\lambda\rho(\overline{\partial p})=0,

because ρ∂=0\rho\partial=0 on G1G_1. On p∈Πjp\in\Pi_j with j≥2j\ge2 its value is an element αλ(p)∈Bj−1\alpha_\lambda(p)\in B_{j-1}, of positive degree at most r−1r-1. Let J⊆B\mathfrak{J}\subseteq B be the ideal generated by the variables of degrees 1,…,r−11,\ldots,r-1. A monomial of positive degree jj contains a variable of degree at most jj, so every homogeneous element of BB of positive degree at most r−1r-1 lies in J\mathfrak{J}. Hence

θλ(B)⊆J,(6)\theta_\lambda(B)\subseteq\mathfrak{J}, \tag*{(6)}

and θλ=0\theta_\lambda=0 if r≤1r\le1, in which case J=0\mathfrak{J}=0. As the commutator of two derivations of odd degree, θλ\theta_\lambda is a derivation of degree −2-2, and

∂θλ=∂Dλ∂=θλ∂.\partial\theta_\lambda=\partial D_\lambda\partial=\theta_\lambda\partial.

Step 4: powers of θλ\theta_\lambda. If r≤1r\le1, then θλ=0\theta_\lambda=0 and (3) holds. Let r≥2r\ge2 and put h=⌊r/2⌋≥1h=\lfloor r/2\rfloor\ge1. For a variable yy of degree at most r−1r-1, the element θλ(y)\theta_\lambda(y) lies in J\mathfrak{J} by (6) and has degree

∣y∣−2h≤r−1−2h≤0.|y|-2h\le r-1-2h\le0.

As J\mathfrak{J} has no nonzero elements of degree zero, θλh(y)=0\theta_\lambda^h(y)=0. We claim that

θλhn+1(B)⊆Jn+1(n≥0).(7)\theta_\lambda^{hn+1}(B)\subseteq\mathfrak{J}^{n+1}\qquad(n\ge0). \tag*{(7)}

For n=0n=0 this is (6). Let n≥1n\ge1 and b∈Bb\in B, and write θλ(b)=∑yayy\theta_\lambda(b)=\sum_y a_y y, a finite sum over variables yy of degrees 1,…,r−11,\ldots,r-1 with ay∈Ba_y\in B. Since θλ\theta_\lambda is a derivation of even degree,

θλhn(ayy)=∑i=0hn(hni)θλhn−i(ay)θλi(y).\theta_\lambda^{hn}(a_y y)=\sum_{i=0}^{hn}\binom{hn}{i}\theta_\lambda^{hn-i}(a_y)\theta_\lambda^i(y).

The terms with i≥hi\ge h vanish. For i<hi<h the factor θλi(y)\theta_\lambda^i(y) lies in J\mathfrak{J}, by (6) if i≥1i\ge1, and

θλhn−i(ay)=θλh(n−1)+1(θλh−1−i(ay))∈Jn\theta_\lambda^{hn-i}(a_y)=\theta_\lambda^{h(n-1)+1}\left(\theta_\lambda^{h-1-i}(a_y)\right)\in\mathfrak{J}^n

by induction. Hence θλhn+1(b)∈Jn+1\theta_\lambda^{hn+1}(b)\in\mathfrak{J}^{n+1}, which proves (7). Finally, J\mathfrak{J} is generated in positive degrees, so Jk+1\mathfrak{J}^{k+1} has no nonzero elements of degree kk. If u>hku>hk, then u≥hk+1u\ge hk+1, and (7) gives

θλu(Bk+2u)⊆(Jk+1)k=0,\theta_\lambda^u(B_{k+2u})\subseteq(\mathfrak{J}^{k+1})_k=0,

which proves (3). □

Change of rings

One regular element

The next lemma isolates what a derivation of degree −1-1 with nilpotent commutator gives. Its proof is the chain-level form of the argument by the change-of-rings exact sequence below. Let SS be a quotient of a local ring (T,m,κ)(T,\mathfrak{m},\kappa) with pd⁡TS<∞\operatorname{pd}_{T} S < \infty, and let ff be a nonzerodivisor of TT which maps to zero in SS. For U=T/(f)U = T/(f) there is an exact sequence

⋯⟶Tor⁡kT(S,κ)⟶Tor⁡kU(S,κ)→χTor⁡k−2U(S,κ)⟶Tor⁡k−1T(S,κ)⟶⋯\cdots\longrightarrow\operatorname{Tor}^{T}_{k}(S,\kappa) \longrightarrow\operatorname{Tor}^{U}_{k}(S,\kappa) \xrightarrow{\chi} \operatorname{Tor}^{U}_{k-2}(S,\kappa) \longrightarrow\operatorname{Tor}^{T}_{k-1}(S,\kappa) \longrightarrow\cdots

so χ\chi is surjective onto Tor⁡kU(S,κ)\operatorname{Tor}^{U}_{k}(S,\kappa) for k≥pd⁡TSk \geq\operatorname{pd}_{T} S, and the group Tor⁡pd⁡TSU(S,κ)\operatorname{Tor}^{U}_{\operatorname{pd}_{T} S}(S,\kappa) vanishes as soon as some power of χ\chi kills Tor⁡pd⁡TS+2uU(S,κ)\operatorname{Tor}^{U}_{\operatorname{pd}_{T} S+2u}(S,\kappa). This general sequence is not used below. In the proof its role is played by the short exact sequence of complexes given by multiplication by ee, and the connecting map of that sequence is the map induced by [∂,D][\partial,D] for any derivation DD with D(e)=1D(e)=1.

Lemma 4.1. Let (T,m,κ)(T,\mathfrak{m},\kappa) be a local ring, let SS be a non-trivial quotient of TT with pd⁡TS<∞\operatorname{pd}_{T} S < \infty, and let f∈ker⁡(T→S)f \in\ker(T \to S) be a nonzerodivisor of TT. Let T=T[Y]\mathcal{T} = T[Y] be a free DG TT-algebra with H0(T)=SH_{0}(\mathcal{T}) = S and Hk(T)=0H_{k}(\mathcal{T}) = 0 for k≥1k \geq1, having a variable ee of degree one with ∂e=f\partial e = f, and put B=κ⊗TTB = \kappa\otimes_{T} \mathcal{T}.

(1) The DG algebra T/(f,e)\mathcal{T}/(f,e) is a free DG algebra over U=T/(f)U = T/(f) on the variables Y∖{e}Y \setminus\{e\}, and it is a free resolution of SS over UU. In particular H(B/(e))=Tor⁡U(S,κ)H(B/(e)) = \operatorname{Tor}^{U}(S,\kappa).

(2) Suppose that DD is a derivation of BB of degree −1-1 with D(e)=1D(e)=1 and that θ=[∂,D]\theta= [\partial,D] satisfies θu(Bpd⁡TS+2u)=0\theta^{u}(B_{\operatorname{pd}_{T} S+2u}) = 0 for some integer u≥1u \geq1. Then

pd⁡T/(f)S≤pd⁡TS−1.\operatorname{pd}_{T/(f)} S \leq\operatorname{pd}_{T} S - 1.

Proof. (1) Let K=T[e]⊆T\mathcal{K} = T[e] \subseteq\mathcal{T}, the Koszul complex on ff. Since ff is a nonzerodivisor, K→U\mathcal{K} \to U is a surjective quasi-isomorphism. As a graded K\mathcal{K}-module, T\mathcal{T} is free on the monomials in the variables other than ee, and the differential of such a monomial of degree qq is a K\mathcal{K}-linear combination of monomials of degree less than qq. By Lemma 2.1,

T⟶U⊗KT=T/(f,e)\mathcal{T} \longrightarrow U \otimes_{\mathcal{K}} \mathcal{T} = \mathcal{T}/(f,e)

is a quasi-isomorphism. The target is U[Y∖{e}]U[Y \setminus\{e\}] as a graded algebra, so it is a complex of free UU-modules resolving SS, and tensoring it with κ\kappa gives C:=B/(e)C := B/(e).

(2) We have pd⁡TS≥1\operatorname{pd}_{T} S \geq1: the nonzerodivisor ff annihilates the non-trivial ring SS, so SS is not free over TT. By (1) and the remark at the beginning of § 2 it suffices to prove Hpd⁡TS(C)=0H_{\operatorname{pd}_{T} S}(C) = 0. Recall that Hk(B)=Tor⁡kT(S,κ)=0H_{k}(B) = \operatorname{Tor}^{T}_{k}(S,\kappa) = 0 for k>pd⁡TSk > \operatorname{pd}_{T} S.

As a graded algebra BB is free over the exterior algebra on the cycle ee, the projection B→CB \to C has kernel eBeB, and ∂(eb)=−e∂b\partial(eb) = -e\partial b. Hence left multiplication by ee gives a short exact sequence of complexes

0⟶C[1]→eB⟶C⟶0,0 \longrightarrow C[1] \xrightarrow{e} B \longrightarrow C \longrightarrow0,

where C[1]k=Ck−1C[1]_k = C_{k-1} with differential −∂-\partial. Its long exact homology sequence [10 Tag 0111] contains the exact sequences

Hk+2(C)→τHk(C)⟶Hk+1(B),H_{k+2}(C) \xrightarrow{\tau} H_k(C) \longrightarrow H_{k+1}(B),

so τ\tau is surjective for k≥pd⁡TSk \ge\operatorname{pd}_T S. Now

θ(e)=∂(1)+D(∂e)=0,\theta(e) = \partial(1) + D(\partial e) = 0,

so the derivation θ\theta preserves the ideal eBeB and induces a chain map θˉ\bar{\theta} on CC. To compute τ\tau, let cˉ∈Ck+2\bar{c} \in C_{k+2} be a cycle, lift it to c∈Bk+2c \in B_{k+2} and write ∂c=eb\partial c = eb. Then τ[cˉ]=[bˉ]\tau[\bar{c}] = [\bar{b}], while D(e)=1D(e) = 1 gives

θc=∂(Dc)+D(eb)=∂(Dc)+b−eD(b),\theta c = \partial(Dc) + D(eb) = \partial(Dc) + b - eD(b),

that is,

bˉ=θˉ cˉ−∂Dc‾in C.\bar{b} = \bar{\theta}\,\bar{c} - \overline{\partial Dc} \qquad\text{in } C.

Thus τ\tau is the map H(θˉ)H(\bar{\theta}) induced by θˉ\bar{\theta}. Consequently

τu=H(θˉ u):Hpd⁡TS+2u(C)⟶Hpd⁡TS(C)\tau^u = H(\bar{\theta}^{\,u}) : H_{\operatorname{pd}_T S+2u}(C) \longrightarrow H_{\operatorname{pd}_T S}(C)

is surjective, as a composite of surjections. It is also zero, because θˉ u(Cpd⁡TS+2u)\bar{\theta}^{\,u}(C_{\operatorname{pd}_T S+2u}) is the image of θu(Bpd⁡TS+2u)=0\theta^u(B_{\operatorname{pd}_T S+2u}) = 0. Hence Hpd⁡TS(C)=0H_{\operatorname{pd}_T S}(C) = 0. □\square

Note that DD is applied only to elements of BB; it need not induce a map on C=B/(e)C = B/(e). The lemma holds in every characteristic and involves no completeness hypothesis.

Proof of Theorem 1.6

If t=0t = 0 there is nothing to prove. Let t≥1t \ge1 and extend the given sequence to a finite generating list g1,…,gng_1,\ldots,g_n of II. Form R\mathcal{R} and BB as in § 2.2. Since [g1],…,[gt][g_1],\ldots,[g_t] are linearly independent in N/nNN/\mathfrak{n}N, there are SS-linear maps λi:N→N/nN→κ\lambda_i:N \to N/\mathfrak{n}N \to\kappa with

λiσ(gˉj)=δij(1≤i,j≤t).\lambda_i\sigma(\bar{g}_j) = \delta_{ij} \qquad(1 \le i,j \le t).

Let rr be the length of a finite free resolution of NN over SS. Proposition 3.1 gives derivations Di=DλiD_i = D_{\lambda_i} of BB with Di(ej)=δijD_i(e_j) = \delta_{ij} for i,j≤ti,j \le t, whose commutators θi=[∂,Di]\theta_i = [\partial,D_i] satisfy θiu(Bk+2u)=0\theta_i^u(B_{k+2u}) = 0 whenever u>k⌊r/2⌋u > k\lfloor r/2\rfloor.

For 0≤j≤t0 \le j \le t put

Qj=R/(g1,…,gj),R(j)=R/(g1,…,gj,e1,…,ej).Q_j = R/(g_1,\ldots,g_j), \qquad\mathcal{R}^{(j)} = \mathcal{R}/(g_1,\ldots,g_j,e_1,\ldots,e_j).

We prove by induction on jj that R(j)\mathcal{R}^{(j)} is a free DG QjQ_j-algebra on the variables X∖{e1,…,ej}X \setminus\{e_1,\ldots,e_j\} which resolves SS, and that

pd⁡QjS≤pd⁡RS−j.(8)\operatorname{pd}_{Q_j} S \le\operatorname{pd}_{R} S - j. \tag*{(8)}

For j=0j = 0 this is the choice of R\mathcal{R}. Let 1≤j≤t1 \le j \le t and assume the statement for j−1j - 1. The image of gjg_j in Qj−1Q_{j-1} is a nonzerodivisor which maps to zero in SS, and it is the differential of the variable eje_j of R(j−1)\mathcal{R}^{(j-1)}. Lemma 4.1(1), applied to T=Qj−1T = Q_{j-1}, T=R(j−1)\mathcal{T} = \mathcal{R}^{(j-1)} and this element, gives the assertion about R(j)\mathcal{R}^{(j)}. The residue algebra of R(j−1)\mathcal{R}^{(j-1)} is B/(e1,…,ej−1)B/(e_1,\ldots,e_{j-1}). The derivation DjD_j maps eibe_i b to −eiDj(b)-e_iD_j(b) for i<ji<j, since Dj(ei)=0D_j(e_i)=0; hence it induces a derivation D‾j\overline{D}_j of B/(e1,…,ej−1)B/(e_1,\ldots,e_{j-1}), still with D‾j(ej)=1\overline{D}_j(e_j)=1. Its commutator with the differential is the map induced by θj\theta_j, so its uu-th power vanishes in degree k+2uk+2u whenever u>k⌊r/2⌋u>k\lfloor r/2\rfloor. Take k=pd⁡Qj−1Sk=\operatorname{pd}_{Q_{j-1}}S, which is finite by the induction hypothesis, and any uu with u>k⌊r/2⌋u>k\lfloor r/2\rfloor. Lemma 4.1(2) gives

pd⁡QjS≤pd⁡Qj−1S−1≤pd⁡RS−j.\operatorname{pd}_{Q_j}S\leq\operatorname{pd}_{Q_{j-1}}S-1\leq\operatorname{pd}_{R}S-j.

This completes the induction, and (8) for j=tj=t is the theorem. □

The module NN and its resolution are never replaced by quotients in this induction; what changes is only the algebra on which the fixed derivations act. This is essential. An induction on tt which replaces RR, II and NN by R/(g1)R/(g_1), I/(g1)I/(g_1) and N/Sσ(g‾1)N/S\sigma(\overline{g}_1) does not work, because the last module may have infinite projective dimension. For instance, let

R=κ[x,y,z](x,y,z),I=(x2,y2),N=S/(z)⊕S/(z+x),R=\kappa[x,y,z]_{(x,y,z)},\qquad I=(x^2,y^2),\qquad N=S/(z)\oplus S/(z+x),

with σ:I/I2≅S2↠N\sigma:I/I^2\cong S^2\twoheadrightarrow N the coordinatewise quotient map. Then pd⁡SN=1\operatorname{pd}_S N=1, since zz and z+xz+x are nonzerodivisors on SS, and for g1=x2+y2g_1=x^2+y^2 one finds

N/Sσ(g‾1)≅S/(z,x).N/S\sigma(\overline{g}_1)\cong S/(z,x).

This module has infinite projective dimension over SS: because the annihilator of xx in SS is xSxS, the minimal free resolution of S/(x)S/(x) is infinite, and zz is a nonzerodivisor on S/(x)S/(x).

Quotients of finite projective dimension

The residual ideal

Theorem 1.6 is used as follows. In the situation of Theorem 1.5, choose a regular sequence g1,…,gℓg_1,\ldots,g_\ell in II with independent images in N/nNN/\mathfrak nN, as long as possible, and pass to

Q=R/(g1,…,gℓ),J=I/(g1,…,gℓ),Q/J=S.Q=R/(g_1,\ldots,g_\ell),\qquad J=I/(g_1,\ldots,g_\ell),\qquad Q/J=S.

Then pd⁡Q(Q/J)<∞\operatorname{pd}_Q(Q/J)<\infty, and the grade of the residual ideal JJ is grade⁡I−ℓ\operatorname{grade} I-\ell. The question what II looks like has become the question which ideals of small grade in a local ring have a cyclic quotient of finite projective dimension. In grade zero, that is, for ideals consisting of zerodivisors, the answer is that only the zero ideal does: an ideal with a finite free resolution has a well-defined rank, namely the alternating sum of the ranks of the free modules, and this rank is seen at every associated prime. In grade one the answer is that the ideal is principal, provided the ring has depth one at every associated prime of the quotient; this follows from the factorization theorem of Buchsbaum and Eisenbud [2 Cor. 5.2]. We state the grade-zero fact for modules, because the module form is used again in § 6, at the associated primes of a complete intersection.

Lemma 5.1. Let TT be a local ring and MM a finitely generated TT-module with pd⁡TM<∞\operatorname{pd}_T M<\infty.

  1. The modules MpM_{\mathfrak p} which are free over TpT_{\mathfrak p}, for prime ideals p\mathfrak p of TT, all have the same rank, and MpM_{\mathfrak p} is free for every $\mathfrak p\in\operatorname{Ass}T.

(2) If Mp0M_{\mathfrak{p}_0} is free of rank ss for some prime ideal p0\mathfrak{p}_0 of TT, then every surjection π:Ts↠M\pi:T^s \twoheadrightarrow M is an isomorphism.

(3) If M=T/JM=T/J for an ideal J⊊TJ\subsetneq T which is contained in an associated prime of TT, then J=0J=0.

Proof. (1) Let EE be the alternating sum of the ranks of the modules in a finite free resolution of MM. For a prime p\mathfrak{p} such that MpM_{\mathfrak{p}} is free, the localized resolution is a split exact sequence of free modules, so MpM_{\mathfrak{p}} has rank EE. For p∈Ass⁡T\mathfrak{p}\in\operatorname{Ass}T the ring TpT_{\mathfrak{p}} has depth zero, so MpM_{\mathfrak{p}} is free by the Auslander–Buchsbaum formula [10 Tag 090V].

(2) The case s=0s=0 is trivial. By (1), MpM_{\mathfrak{p}} is free of rank ss for every p∈Ass⁡T\mathfrak{p}\in\operatorname{Ass}T, so πp\pi_{\mathfrak{p}} is a surjection between free modules of the same finite rank over a local ring, and therefore an isomorphism. The kernel of π\pi is a submodule of TsT^s which vanishes at every associated prime of TsT^s. Its associated primes are among those of TsT^s [10 Tag 02M3], so it has none, and it is zero [10 Tag 0587].

(3) If J⊆p0∈Ass⁡TJ\subseteq\mathfrak{p}_0\in\operatorname{Ass}T, then (T/J)p0(T/J)_{\mathfrak{p}_0} is nonzero and cyclic, and it is free by (1), hence free of rank one. Part (2), applied to T↠T/JT\twoheadrightarrow T/J, gives J=0J=0.

Equivalently, a nonzero proper ideal with a cyclic quotient of finite projective dimension contains a nonzerodivisor; this ideal form is classical.

Lemma 5.2. Let QQ be a local ring and JJ an ideal with pd⁡Q(Q/J)<∞\operatorname{pd}_Q(Q/J)<\infty. If depth⁡Qp=1\operatorname{depth}Q_{\mathfrak{p}}=1 for every associated prime p\mathfrak{p} of Q/JQ/J, then J=aQJ=aQ for a nonzerodivisor aa of QQ.

Proof. The ideal JJ is nonzero, since otherwise the associated primes of QQ, at which QQ has depth zero, would be associated to Q/JQ/J. It has a finite free resolution, so by [2 Cor. 5.2] we have J=aJ′J=aJ' for a nonzerodivisor aa and an ideal J′J' which is either QQ or contains a QQ-regular sequence of length two. Multiplication by aa is an isomorphism of Q/J′Q/J' onto aQ/aJ′=aQ/J⊆Q/JaQ/aJ'=aQ/J\subseteq Q/J, hence

Ass⁡(Q/J′)⊆Ass⁡(Q/J).\operatorname{Ass}(Q/J')\subseteq\operatorname{Ass}(Q/J).

If J′≠QJ'\ne Q, choose p∈Ass⁡(Q/J′)\mathfrak{p}\in\operatorname{Ass}(Q/J'). Then p⊇J′\mathfrak{p}\supseteq J' contains a QQ-regular sequence of length two, which is also QpQ_{\mathfrak{p}}-regular, so depth⁡Qp≥2\operatorname{depth}Q_{\mathfrak{p}}\geq2, while depth⁡Qp=1\operatorname{depth}Q_{\mathfrak{p}}=1 because p∈Ass⁡(Q/J)\mathfrak{p}\in\operatorname{Ass}(Q/J). This contradiction shows J′=QJ'=Q and J=aQJ=aQ.

If QQ is Cohen–Macaulay, the hypothesis of Lemma 5.2 says that every associated prime of Q/JQ/J has height one.

Proof of Theorem 1.5

Choose an RR-regular sequence g1,…,gℓg_1,\ldots,g_\ell in II such that [g1],…,[gℓ][g_1],\ldots,[g_\ell] are linearly independent in N/nNN/\mathfrak{n}N and which is maximal with these two properties. It exists, since the empty sequence has them and linear independence bounds the length by μS(N)\mu_S(N). Let

Q=R/(g1,…,gℓ),J=I/(g1,…,gℓ).Q=R/(g_1,\ldots,g_\ell),\qquad J=I/(g_1,\ldots,g_\ell).

Then Q/J=SQ/J=S is non-trivial, and Theorem Theorem 1.6 gives pd⁡QS<∞\operatorname{pd}_Q S<\infty. If J≠0J\ne0, then JJ is contained in no associated prime of QQ by Lemma Lemma 5.1(3); by prime avoidance [10 Tag 00DS] it contains a nonzerodivisor of QQ, whose lift to II prolongs g1,…,gℓg_1,\ldots,g_\ell to an RR-regular sequence in II. Hence

ℓ<grade⁡Iif J≠0.(9)\ell< \operatorname{grade} I \qquad\text{if } J \ne0. \tag*{(9)}

The images span. We show that [g1],…,[gℓ][g_1],\ldots,[g_\ell] span N/nNN/\mathfrak{n}N. If J=0J=0, then g1,…,gℓg_1,\ldots,g_\ell generate II, and their images span because σ\sigma is surjective. Let J≠0J\ne0, and suppose that the span WW of [g1],…,[gℓ][g_1],\ldots,[g_\ell] is a proper subspace. The set

I′′={g∈I:[g]∈W}I''=\{g\in I : [g]\in W\}

is an ideal of RR, and I′′≠II''\ne I because I→N/nNI\to N/\mathfrak{n}N is surjective. Let p1,…,ps\mathfrak{p}_1,\ldots,\mathfrak{p}_s be the inverse images in RR of the associated primes of QQ; none of them contains II, since JJ is contained in no associated prime of QQ. By prime avoidance [10 Tag 00DS] there is gℓ+1∈Ig_{\ell+1}\in I outside I′′∪p1∪⋯∪psI''\cup\mathfrak{p}_1\cup\cdots\cup\mathfrak{p}_s. Then g1,…,gℓ+1g_1,\ldots,g_{\ell+1} is RR-regular and [g1],…,[gℓ+1][g_1],\ldots,[g_{\ell+1}] are linearly independent, contradicting maximality. Thus

ℓ=μS(N).\ell=\mu_S(N).

Parts (1) and (2). A regular sequence in II has length at most grade⁡I\operatorname{grade} I [8 Thm. 16.7], so μS(N)=ℓ≤grade⁡I\mu_S(N)=\ell\le\operatorname{grade} I, which is (1). If μS(N)≥grade⁡I\mu_S(N)\ge\operatorname{grade} I, then ℓ=grade⁡I\ell=\operatorname{grade} I, and (5.1) gives J=0J=0. Hence I=(g1,…,gℓ)I=(g_1,\ldots,g_\ell), and the ℓ\ell independent elements [gi][g_i] form a basis of N/nNN/\mathfrak{n}N. This proves (2).

Part (3). Suppose μS(N)=grade⁡I−1\mu_S(N)=\operatorname{grade} I-1 and depth⁡Rp=grade⁡I\operatorname{depth} R_{\mathfrak{p}}=\operatorname{grade} I for every p∈Ass⁡RS\mathfrak{p}\in\operatorname{Ass}_R S. Then ℓ=μS(N)=grade⁡I−1\ell=\mu_S(N)=\operatorname{grade} I-1. The associated primes of the QQ-module Q/J=SQ/J=S are the ideals p‾=p/(g1,…,gℓ)\overline{\mathfrak{p}}=\mathfrak{p}/(g_1,\ldots,g_\ell) with p∈Ass⁡RS\mathfrak{p}\in\operatorname{Ass}_R S. For such p\mathfrak{p} the sequence g1,…,gℓg_1,\ldots,g_\ell is RpR_{\mathfrak{p}}-regular and Qp‾=Rp/(g1,…,gℓ)Q_{\overline{\mathfrak{p}}}=R_{\mathfrak{p}}/(g_1,\ldots,g_\ell), so [10 Tag 090R]

depth⁡Qp‾=depth⁡Rp−ℓ=1.\operatorname{depth} Q_{\overline{\mathfrak{p}}}=\operatorname{depth} R_{\mathfrak{p}}-\ell=1.

Lemma 5.2 gives J=aQJ=aQ with aa a nonzerodivisor of QQ. For a lift gℓ+1∈Ig_{\ell+1}\in I of aa, the sequence g1,…,gℓ,gℓ+1g_1,\ldots,g_\ell,g_{\ell+1} is RR-regular and generates II. □

Only one regular sequence is chosen in this proof. The three parts correspond to three possibilities for the residual ideal: its grade cannot be negative, it vanishes in grade zero, and it is principal in grade one under the depth condition of part (3). When JJ has grade two or more there is no such conclusion: in Proposition 7.2 the ring RR is regular, the residual ideal has height two and N=0N=0.

Differentials

The Jacobian image

Let kk be a field and SS a non-trivial local kk-algebra essentially of finite type. Fix a polynomial ring k[x1,…,xv]k[x_1,\ldots,x_v], an ideal aa of it contained in a prime ideal qq, and put Λ=k[x1,…,xv]/a\Lambda=k[x_1,\ldots,x_v]/a, such that S=Λq/aS=\Lambda_{q/a}. Let

R=k[x1,…,xv]q,I=aR,S=R/I.R=k[x_1,\ldots,x_v]_q,\qquad I=aR,\qquad S=R/I.

Then (R,n,κ)(R,n,\kappa) is a regular local ring, so that Theorem 1.5 applies with grade I=ht⁡RII=\operatorname{ht}_{R} I, and ΩR/κ\Omega_{R/\kappa} is free on dx1,…,dxvdx_{1},\ldots,dx_{v}. The conormal sequence of κ→R→S\kappa\to R\to S reads

I/I2→δS⊗RΩR/κ≅Sv⟶ΩS/κ⟶0,δ(fˉ)=∑i=1v∂f∂xi dxi.I/I^{2}\xrightarrow{\delta} S\otimes_{R}\Omega_{R/\kappa}\cong S^{v}\longrightarrow\Omega_{S/\kappa}\longrightarrow0,\qquad\delta(\bar{f})=\sum_{i=1}^{v}\frac{\partial f}{\partial x_{i}}\,dx_{i}.

We call the image δ(I/I2)\delta(I/I^{2}) of δ\delta the Jacobian image, and we write σ\sigma for δ\delta regarded as a map onto it; it plays the role of NN in Theorem 1.5. Thus

σ:I/I2↠δ(I/I2),0⟶δ(I/I2)⟶Sv⟶ΩS/κ⟶0.(10)\sigma:I/I^{2}\twoheadrightarrow\delta(I/I^{2}),\qquad0\longrightarrow\delta(I/I^{2})\longrightarrow S^{v}\longrightarrow\Omega_{S/\kappa}\longrightarrow0. \tag*{(10)}

Since SvS^{v} is free, pd⁡Sδ(I/I2)<∞\operatorname{pd}_{S}\delta(I/I^{2})<\infty if and only if pd⁡SΩS/κ<∞\operatorname{pd}_{S}\Omega_{S/\kappa}<\infty. In that case δ(I/I2)\delta(I/I^{2}) is a quotient of the conormal module with finite projective dimension, and Theorem 1.5 applies. The number μS(δ(I/I2))\mu_{S}(\delta(I/I^{2})) is the minimal number of generators of the module generated by the rows of the Jacobian matrix of a generating set of a\mathfrak{a}; it is at least the rank of that matrix at the closed point. Theorem 1.5(1) says that

μS(δ(I/I2))≤ht⁡I,\mu_{S}(\delta(I/I^{2}))\leq\operatorname{ht} I,

and what has to be shown is that μS(δ(I/I2))\mu_{S}(\delta(I/I^{2})) is not smaller. Since the number of generators of a module can only drop under localization, it suffices to bound the rank of δ(I/I2)\delta(I/I^{2}) from below at a generic point, where it is computed by the classical Jacobian criterion.

For a homomorphism of rings A→BA\to B we write H1(LB/A)H_{1}(L_{B/A}) for the first homology of the naive cotangent complex of BB over AA [10 Tag 00S0]. It vanishes when BB is a localization of a polynomial ring over AA [10 Tag 08Q1, Tag 00S7], it is b/b2\mathfrak{b}/\mathfrak{b}^{2} when A→BA\to B is surjective with kernel b\mathfrak{b} [10 Tag 07BP], and for a field extension K/κK/\kappa it vanishes if and only if KK is separable over κ\kappa [10 Tag 0322]. For homomorphisms A→B→CA\to B\to C such that ΩB/A\Omega_{B/A} is a free BB-module there is the exact Jacobi–Zariski sequence [10 Tag 00S2]

H1(LB/A)⊗BC⟶H1(LC/A)⟶H1(LC/B)⟶ΩB/A⊗BC⟶ΩC/A⟶ΩC/B⟶0.(11)\begin{aligned} H_{1}(L_{B/A})\otimes_{B}C&\longrightarrow H_{1}(L_{C/A})\longrightarrow H_{1}(L_{C/B})\\ &\longrightarrow\Omega_{B/A}\otimes_{B}C\longrightarrow\Omega_{C/A}\longrightarrow\Omega_{C/B}\longrightarrow0. \tag*{(11)} \end{aligned}

Lemma 6.1. Let η\eta be a minimal prime of SS such that K=SηK=S_{\eta} is a field, let p⊆R\mathfrak{p}\subseteq R be its inverse image. Then

dim⁡Kδ(I/I2)η=ht⁡Rp−dim⁡KH1(LK/κ),\dim_{K}\delta(I/I^{2})_{\eta}=\operatorname{ht}_{R}\mathfrak{p}-\dim_{K}H_{1}(L_{K/\kappa}),

and H1(LK/κ)=0H_{1}(L_{K/\kappa})=0 if KK is separably generated over κ\kappa. Consequently

μS(δ(I/I2))≥ht⁡Rp−dim⁡KH1(LK/κ)≥ht⁡RI−dim⁡KH1(LK/κ).\mu_{S}(\delta(I/I^{2}))\geq\operatorname{ht}_{R}\mathfrak{p}-\dim_{K}H_{1}(L_{K/\kappa})\geq\operatorname{ht}_{R}I-\dim_{K}H_{1}(L_{K/\kappa}).

Proof. Since Rp/IRp=SηR_{\mathfrak{p}}/IR_{\mathfrak{p}}=S_{\eta} is a field, IRp=pRpIR_{\mathfrak{p}}=\mathfrak{p}R_{\mathfrak{p}}, and the localization of δ\delta at η\eta is the map

pRp/(pRp)2⟶ΩRp/κ⊗RpK,f⟼df,(12)\mathfrak{p}R_{\mathfrak{p}}/(\mathfrak{p}R_{\mathfrak{p}})^{2}\longrightarrow\Omega_{R_{\mathfrak{p}}/\kappa}\otimes_{R_{\mathfrak{p}}}K,\qquad f\longmapsto df, \tag*{(12)}

with image δ(I/I2)η\delta(I/I^{2})_{\eta}. In the Jacobi–Zariski sequence (11) of κ→Rp→K\kappa\to R_{\mathfrak{p}}\to K the first term vanishes, because RpR_{\mathfrak{p}} is a localization of a polynomial ring 、】【 over kk, and H1(LK/Rp)=pRp/(pRp)2H_{1}(L_{K/R_{\mathfrak p}})=\mathfrak pR_{\mathfrak p}/(\mathfrak pR_{\mathfrak p})^{2} maps to ΩRp/k⊗RpK\Omega_{R_{\mathfrak p}/k}\otimes_{R_{\mathfrak p}}K by (12). Hence H1(LK/k)H_{1}(L_{K/k}) is the kernel of (12). The ring RpR_{\mathfrak p} is a regular local ring [10 Tag 0AFS] of dimension ht⁡Rp\operatorname{ht}_{R}\mathfrak p, so the source of (12) has dimension ht⁡Rp\operatorname{ht}_{R}\mathfrak p, and dim⁡Kδ(I/I2)η=ht⁡Rp−dim⁡KH1(LK/k)\dim_{K}\delta(I/I^{2})_{\eta}=\operatorname{ht}_{R}\mathfrak p-\dim_{K}H_{1}(L_{K/k}). If KK is separably generated over kk, then it is separable over kk [10 Tag 030X], so H1(LK/k)=0H_{1}(L_{K/k})=0. Finally μS(δ(I/I2))≥dim⁡Kδ(I/I2)η\mu_{S}(\delta(I/I^{2}))\geq\dim_{K}\delta(I/I^{2})_{\eta}, and ht⁡Rp≥ht⁡RI\operatorname{ht}_{R}\mathfrak p\geq\operatorname{ht}_{R}I because I⊆pI\subseteq\mathfrak p. □

The vector space H1(LK/k)H_{1}(L_{K/k}) depends only on the field extension K/kK/k, and it vanishes exactly when KK is separable over kk [10 Tag 0322]. Its dimension is what the Jacobian matrix fails to see in positive characteristic. Localizing (10) at η\eta gives the first of the equalities

dim⁡Kδ(I/I2)η=v−dim⁡KΩK/k,ht⁡Rp=v−trdeg⁡kK\dim_{K}\delta(I/I^{2})_{\eta}=v-\dim_{K}\Omega_{K/k},\qquad\operatorname{ht}_{R}\mathfrak p=v-\operatorname{trdeg}_{k}K

and the second is the dimension formula for k→k[x1,…,xv]k\to k[x_{1},\ldots,x_{v}], which is an equality because a field is universally catenary [10 Tag 02IJ]. Comparing with Lemma 6.1,

dim⁡KH1(LK/k)=dim⁡KΩK/k−trdeg⁡kK\dim_{K}H_{1}(L_{K/k})=\dim_{K}\Omega_{K/k}-\operatorname{trdeg}_{k}K

which is Cartier’s equality [8 Thm. 26.10]; there it is stated over a perfect base field, here the prime field k0k_{0} of kk, with the kernel of K⊗kΩk/k0→ΩK/k0K\otimes_{k}\Omega_{k/k_{0}}\to\Omega_{K/k_{0}} in place of H1(LK/k)H_{1}(L_{K/k}); the two agree by the Jacobi–Zariski sequence of k0→k→Kk_{0}\to k\to K, since H1(LK/k0)=0H_{1}(L_{K/k_{0}})=0. It is not used in this paper.

Proof of Theorem 1.2

Theorem 1.5 needs no injectivity of the conormal map, so it is applied first. Everything else then follows from one computation at the associated primes of SS. Let η\eta be the minimal prime of the theorem, let K=SηK=S_{\eta} and p\mathfrak p be as in Lemma 6.1.

Complete intersection. By Lemma 6.1, where H1(LK/k)=0H_{1}(L_{K/k})=0 because KK is separably generated,

μS(δ(I/I2))≥ht⁡Rp≥ht⁡RI.\mu_{S}(\delta(I/I^{2}))\geq\operatorname{ht}_{R}\mathfrak p\geq\operatorname{ht}_{R}I.

Theorem 1.5(2) gives μS(δ(I/I2))=ht⁡RI\mu_{S}(\delta(I/I^{2}))=\operatorname{ht}_{R}I and shows that II is generated by an RR-regular sequence of length ht⁡RI\operatorname{ht}_{R}I; then ht⁡Rp=ht⁡RI\operatorname{ht}_{R}\mathfrak p=\operatorname{ht}_{R}I. Thus SS is a complete intersection over kk. Clearing denominators, which are units of RR, we may take the generators f1,…,fht⁡RIf_{1},\ldots,f_{\operatorname{ht}_{R}I} in a\mathfrak a. Let

Φ:Sht⁡RI⟶Sv\Phi:S^{\operatorname{ht}_{R}I}\longrightarrow S^{v}

be the SS-linear map sending the ii-th basis vector to δ(fˉi)\delta(\bar f_{i}). It is the composite of the surjection Sht⁡RI↠I/I2S^{\operatorname{ht}_{R}I}\twoheadrightarrow I/I^{2} given by the fˉi\bar f_{i} with δ\delta, so its image is δ(I/I2)\delta(I/I^{2}) and its cokernel is ΩS/k\Omega_{S/k} by (10).

At the associated primes. By Lemma 6.1, dim⁡Kδ(I/I2)η=ht⁡Rp=ht⁡RI\dim_{K}\delta(I/I^{2})_{\eta}=\operatorname{ht}_{R}\mathfrak p=\operatorname{ht}_{R}I, so (ΩS/k)η(\Omega_{S/k})_{\eta} is free of rank v−ht⁡RIv-\operatorname{ht}_{R}I over KK. Let r∈Ass⁡S\mathfrak r\in\operatorname{Ass}S. By Lemma 5.1(1), (ΩS/k)r(\Omega_{S/k})_{\mathfrak r} is free of rank v−ht⁡RIv-\operatorname{ht}_{R}I. Hence the projection Srv→(ΩS/k)rS_{\mathfrak r}^{v}\to(\Omega_{S/k})_{\mathfrak r} splits, and its kernel δ(I/I2)r\delta(I/I^{2})_{\mathfrak r} is free of rank ht⁡RI\operatorname{ht}_{R}I. The map Φr\Phi_{\mathfrak r} maps Srht⁡RIS_{\mathfrak r}^{\operatorname{ht}_{R}I} onto this kernel, so it is an isomorphism onto it, and Φr\Phi_{\mathfrak r} is a split injection. Its reduction modulo rSr\mathfrak{r}S_{\mathfrak{r}} is therefore injective, so some maximal minor Δ\Delta of the Jacobian matrix (∂fi/∂xj)(\partial f_i/\partial x_j) is a unit in SrS_{\mathfrak{r}} (take Δ=1\Delta=1 if I=0I=0). Let r~⊆q\tilde{\mathfrak{r}}\subseteq\mathfrak{q} be the inverse image of r\mathfrak{r} in k[x1,…,xv]k[x_1,\ldots,x_v]. After renumbering the variables, the kk-algebra k[x1,…,xv]Δ/(f1,…,fht⁡RI)k[x_1,\ldots,x_v]_\Delta/(f_1,\ldots,f_{\operatorname{ht}_R I}) is standard smooth [10 Tag 00T8], hence smooth [10 Tag 00T7], and its local ring at r~\tilde{\mathfrak{r}} is regular [10 Tag 00TT]. This local ring is SrS_{\mathfrak{r}}, because f1,…,fht⁡RIf_1,\ldots,f_{\operatorname{ht}_R I} generate a\mathfrak{a} after localization at q\mathfrak{q}, hence also after localization at r~\tilde{\mathfrak{r}}. A regular local ring is a domain [10 Tag 00NP], and SrS_{\mathfrak{r}} has depth zero, so SrS_{\mathfrak{r}} is a field which is essentially smooth over kk.

Conclusion. The kernel of Φ\Phi is a submodule of Sht⁡RIS^{\operatorname{ht}_R I}, and the nilradical of SS is an ideal of SS. Both vanish at every associated prime of SS, the first because Φr\Phi_{\mathfrak{r}} is injective and the second because SrS_{\mathfrak{r}} is a field. As in the proof of Lemma 5.1(2), both are zero [10 Tag 02M3, Tag 0587]. Thus SS is reduced, every associated prime of SS is minimal, which gives the assertion on the minimal primes, and Φ\Phi is injective. Since Φ\Phi factors through the surjection Sht⁡RI↠I/I2S^{\operatorname{ht}_R I}\twoheadrightarrow I/I^2, both maps are isomorphisms, and (6.1) becomes the exact sequence

0⟶I/I2≅Sht⁡RI→δSv⟶ΩS/k⟶0.(13)0\longrightarrow I/I^2\cong S^{\operatorname{ht}_R I}\xrightarrow{\delta}S^v\longrightarrow\Omega_{S/k}\longrightarrow0. \tag*{(13)}

Hence pd⁡SΩS/k≤1\operatorname{pd}_S\Omega_{S/k}\leq1. No separability is used after the first step. □

Characteristic zero and perfect fields

In characteristic zero there is no hypothesis on generic points in Theorem 1.1, so one must show that the local rings of SS at its generic points are fields. They have depth zero, so their modules of differentials are free, and in characteristic zero freeness of the module of differentials characterizes regularity.

Proof of Theorem 1.1. Let η\eta be a minimal prime of SS and η~\tilde{\eta} its inverse image in Λ\Lambda, so that Sη=Λη~S_\eta=\Lambda_{\tilde{\eta}}. The ring SηS_\eta has depth zero, so (ΩΛ/k)η~=(ΩS/k)η(\Omega_{\Lambda/k})_{\tilde{\eta}}=(\Omega_{S/k})_\eta is free by the Auslander–Buchsbaum formula. Since char⁡k=0\operatorname{char} k=0, the local ring Λη~\Lambda_{\tilde{\eta}} is therefore regular [10 Tag 00TX], hence a domain [10 Tag 00NP]. Being of dimension zero, it is a field. This field is finitely generated over kk, hence separably generated since char⁡k=0\operatorname{char} k=0 [10 Tag 0322, Tag 030O]. Apply Theorem 1.2. □

Proof of Corollary 1.3. Let η\eta be a minimal prime with SηS_\eta reduced. A reduced local ring of dimension zero is a field, and SηS_\eta is finitely generated over kk as a field. Every extension of a perfect field is separable [10 Tag 030Y, Tag 030Z], and a finitely generated separable extension is separably generated [10 Tag 030O]. Apply Theorem 1.2. □

Remark 6.2. The hypotheses of Theorem 1.2 and Corollary 1.3 cannot be dropped. For every field kk of positive characteristic pp the ring

S=(k[x,y]/(x2p,xpyp,y2p))(x,y)S=\left(k[x,y]/(x^{2p},x^py^p,y^{2p})\right)_{(x,y)}

has ΩS/k≅S2\Omega_{S/k}\cong S^2, because the differentials of the three equations vanish, and it is not a complete intersection, because its defining ideal has height two and three minimal generators. For k=Fpk = \mathbf{F}_p the residue field of its unique minimal prime η\eta is separable over kk; so the hypothesis of Theorem 1.2 concerns the local ring SηS_\eta and not only its residue field, and reducedness cannot be dropped in Corollary 1.3.

The imperfection-degree threshold

Theorems 1.1 and 1.2 settle the conjecture when some generic point is separably generated. It remains to determine what happens for reduced algebras over an imperfect field, where the Jacobian image can be too small.

One element short

When no generic point is separably generated, Lemma 6.1 still bounds μS(δ(I/I2))\mu_S(\delta(I/I^2)) from below by ht⁡RI−dim⁡KH1(LK/k)\operatorname{ht}_R I-\dim_K H_1(L_{K/k}). If this dimension is at most one, the regular sequence of Theorem 1.5 is at most one element short, and Theorem 1.5(3) applies.

Corollary 7.1. Let kk be a field and let SS be a non-trivial reduced local kk-algebra essentially of finite type with pd⁡SΩS/k<∞\operatorname{pd}_S\Omega_{S/k}<\infty. If dim⁡SηH1(LSη/k)≤1\dim_{S_\eta}H_1(L_{S_\eta/k})\leq1 for every minimal prime η\eta of SS, then SS is a complete intersection over kk.

Proof. Use the notation of §6.1. Since SS is reduced, every SηS_\eta is a field and Ass⁡RS\operatorname{Ass}_R S consists of the inverse images pη\mathfrak{p}_\eta of the minimal primes η\eta of SS. Then ht⁡Rpη≥ht⁡RI\operatorname{ht}_R\mathfrak{p}_\eta\geq\operatorname{ht}_R I. Lemma 6.1 gives

μS(δ(I/I2))≥ht⁡Rpη−dim⁡SηH1(LSη/k)≥ht⁡RI−1\mu_S(\delta(I/I^2))\geq\operatorname{ht}_R\mathfrak{p}_\eta-\dim_{S_\eta}H_1(L_{S_\eta/k})\geq\operatorname{ht}_R I-1

for every η\eta. If μS(δ(I/I2))≥ht⁡RI\mu_S(\delta(I/I^2))\geq\operatorname{ht}_R I, Theorem 1.5(2) shows that SS is a complete intersection. Otherwise μS(δ(I/I2))=ht⁡RI−1\mu_S(\delta(I/I^2))=\operatorname{ht}_R I-1 and

ht⁡Rpη≤μS(δ(I/I2))+1≤ht⁡RI,\operatorname{ht}_R\mathfrak{p}_\eta\leq\mu_S(\delta(I/I^2))+1\leq\operatorname{ht}_R I,

so ht⁡Rpη=ht⁡RI\operatorname{ht}_R\mathfrak{p}_\eta=\operatorname{ht}_R I for every η\eta. As RR is regular, this says that depth⁡Rp=grade⁡I\operatorname{depth}R_{\mathfrak{p}}=\operatorname{grade} I for every associated prime p\mathfrak{p} of the RR-module SS, and Theorem 1.5(3) applies. □

In positive characteristic pp the dimension of H1(LK/k)H_1(L_{K/k}) is bounded by the size of a pp-basis of kk: the Jacobi–Zariski sequence of Fp→k→K\mathbf{F}_p\to k\to K embeds H1(LK/k)H_1(L_{K/k}) into K⊗kΩk/FpK\otimes_k\Omega_{k/\mathbf{F}_p}, and Ωk/Fp\Omega_{k/\mathbf{F}_p} has a basis indexed by a pp-basis of kk. This explains the form of Theorem 1.4: H1(LK/k)H_1(L_{K/k}) has dimension at most one for all extensions K/kK/k exactly when [k:kp]≤p[k:k^p]\leq p.

Proof of Theorem 1.4, (1)⇒(2). Write [k:kp]=pj[k:k^p]=p^j with j≤1j\leq1, so that dim⁡kΩk/Fp=j\dim_k\Omega_{k/\mathbf{F}_p}=j by the description of differentials through pp-bases [10 Tag 07P2]. Let K/kK/k be a field extension. Every extension of the perfect field Fp\mathbf{F}_p is separable [10 Tag 030Y, Tag 030Z], so H1(LK/Fp)=0H_1(L_{K/\mathbf{F}_p})=0, and the Jacobi–Zariski sequence (11) of Fp→k→K\mathbf{F}_p\to k\to K gives an injection

H1(LK/k)↪K⊗kΩk/Fp.H_1(L_{K/k})\hookrightarrow K\otimes_k\Omega_{k/\mathbf{F}_p}.

Hence

dim⁡KH1(LK/k)≤j≤1.\dim_K H_1(L_{K/k})\leq j\leq1.

Thus every reduced SS as in (2) satisfies the hypothesis of Corollary 7.1, which proves (2). □

A domain with free differentials

For the converse we need, over every field with [k:kp]>p[k:k^p]>p, a reduced algebra with differentials of finite projective dimension which is not a complete intersection. The construction rests on one observation: in positive characteristic pp a polynomial in pp-th powers has zero differential, so equations in pp-th powers do not affect the module of differentials. Over a perfect field such equations produce nilpotents, as in Remark 6.2. Over an imperfect field one can shift pp-th powers by constants which are not pp-th powers and obtain domains. Two pp-independent constants suffice to transport the monomial curve (t3,t4,t5)(t^3,t^4,t^5), which is not a complete intersection; with only one, H1(LK/k)H_1(L_{K/k}) has dimension at most one and Corollary 7.1 applies.

Proposition 7.2. Let kk be a field of positive characteristic pp containing two pp-independent elements a,ba,b, for instance k=Fp(a,b)k=\mathbb{F}_p(a,b). In the polynomial ring k[x,y,z]k[x,y,z] put

U=xp−a,V=yp−b,W=zp−a,U=x^p-a,\qquad V=y^p-b,\qquad W=z^p-a,
I=(V2−UW, VW−U3, W2−U2V),(14)I=(V^2-UW,\ VW-U^3,\ W^2-U^2V), \tag*{(14)}

let m=(U,V,z−x)\mathfrak{m}=(U,V,z-x) and R=k[x,y,z]mR=k[x,y,z]_{\mathfrak{m}}, S=R/IRS=R/IR. Then SS is a one-dimensional local domain over kk, it is not a complete intersection, and

ΩS/k≅S3,pd⁡SΩS/k=0.(15)\Omega_{S/k}\cong S^3,\qquad\operatorname{pd}_S\Omega_{S/k}=0. \tag*{(15)}

Proof. The monomial curve. For independent variables U,V,WU,V,W the homomorphism

k[U,V,W]/(V2−UW, VW−U3, W2−U2V)⟶k[t],k[U,V,W]/(V^2-UW,\ VW-U^3,\ W^2-U^2V)\longrightarrow k[t],
(U,V,W)⟼(t3,t4,t5),(16)(U,V,W)\longmapsto(t^3,t^4,t^5), \tag*{(16)}

is an isomorphism onto Γ:=k[t3,t4,t5]\Gamma:=k[t^3,t^4,t^5]: the three relations reduce every monomial to a k[U]k[U]-linear combination of 1,V,W1,V,W, and the images of these normal forms have exponents in distinct residue classes modulo three, so they are linearly independent. This holds for every pp, including p=2,3p=2,3.

Primality and dimension. Since k[U,V,W]=k[xp,yp,zp]k[U,V,W]=k[x^p,y^p,z^p], the ring k[x,y,z]k[x,y,z] is free of rank p3p^3 over k[U,V,W]k[U,V,W], and by (16)

k[x,y,z]/I≅Γ[x,y,z]/(xp−a−t3, yp−b−t4, zp−a−t5),k[x,y,z]/I\cong\Gamma[x,y,z]/(x^p-a-t^3,\ y^p-b-t^4,\ z^p-a-t^5),

which is free of rank p3p^3 over Γ\Gamma because the three equations are monic in separate variables. In particular dim⁡k[x,y,z]/I=1\dim k[x,y,z]/I=1. Let F=k(t)=Frac⁡ΓF=k(t)=\operatorname{Frac}\Gamma. The generic fibre F[x,y,z]/(xp−(a+t3), yp−(b+t4), zp−(a+t5))F[x,y,z]/(x^p-(a+t^3),\ y^p-(b+t^4),\ z^p-(a+t^5)) is a field if and only if a+t3a+t^3, b+t4b+t^4, and a+t5a+t^5 are pp-independent over Fp\mathbb{F}_p, that is, if and only if their differentials are linearly independent in ΩF/Fp\Omega_{F/\mathbb{F}_p} [10 Tag 07P2]. Since k→Fk\to F is a purely transcendental, hence separable, extension, the map F⊗kΩk/Fp→ΩF/FpF\otimes_k\Omega_{k/\mathbb{F}_p}\to\Omega_{F/\mathbb{F}_p} is injective and ΩF/Fp=(F⊗kΩk/Fp)⊕F dt\Omega_{F/\mathbb{F}_p}=(F\otimes_k\Omega_{k/\mathbb{F}_p})\oplus F\,dt, so dada, dbdb, dtdt are linearly independent. With respect to these three elements the differentials of a+t3a+t^3, b+t4b+t^4, a+t5a+t^5 have coefficient matrix

(103t2014t3105t4).\begin{pmatrix} 1 & 0 & 3t^2 \\ 0 & 1 & 4t^3 \\ 1 & 0 & 5t^4 \end{pmatrix}.

Its determinant 5t4−3t2=t2(5t2−3)5t^4-3t^2=t^2(5t^2-3) is a nonzero polynomial in the transcendental element tt in every characteristic (t4+t2t^4+t^2 for p=2p=2, 2t42t^4 for p=3p=3). Hence the generic fibre is a field of degree p3p^3 over FF, and the free Γ\Gamma-algebra k[x,y,z]/Ik[x,y,z]/I embeds into it, so II is a prime ideal of height two.

The maximal ideal. The residue ring k[x,y,z]/(U,V,z−x)k[x,y,z]/(U,V,z-x) is the field k(a1/p,b1/p)k(a^{1/p},b^{1/p}), so m\mathfrak m is maximal with ht⁡m=3\operatorname{ht}\mathfrak m=3, and U,V,τ:=z−xU,V,\tau:=z-x is a regular system of parameters of the regular local ring RR. One has W=(x+τ)p−a=U+τpW=(x+\tau)^p-a=U+\tau^p, so I⊆m2I\subseteq\mathfrak m^2 and S=R/IRS=R/IR is a one-dimensional local domain with ht⁡RIR=2\operatorname{ht}_R IR=2.

Not a complete intersection. Let κ=R/mR\kappa=R/\mathfrak mR and write u∗u^* for the initial form of u∈Ru\in R in the associated graded ring gr⁡mR=κ[U∗,V∗,τ∗]\operatorname{gr}_{\mathfrak m}R=\kappa[U^*,V^*,\tau^*]. The initial forms of the three generators are V∗2−U∗2V^{*2}-U^{*2}, U∗V∗U^*V^* and U∗2U^{*2} (note W∗=U∗W^*=U^* since τp∈m2\tau^p\in\mathfrak m^2), which are linearly independent in m2/m3\mathfrak m^2/\mathfrak m^3. Since mIR⊆m3\mathfrak mIR\subseteq\mathfrak m^3, the three generators are linearly independent modulo mIR\mathfrak mIR, so μR(IR)=3>2=ht⁡RIR\mu_R(IR)=3>2=\operatorname{ht}_R IR. An ideal of RR generated by a regular sequence of length cc has height cc and is minimally generated by cc elements, so IRIR is not generated by an RR-regular sequence, and SS is not a complete intersection [8 Thm. 21.2(ii)].

Free differentials. Each generator of II is a polynomial in U,V,WU,V,W, and dU=dV=dW=0dU=dV=dW=0 in Ωk[x,y,z]/k\Omega_{k[x,y,z]/k}. Hence the conormal map I/I2→S⊗RΩR/k=S3I/I^2\to S\otimes_R\Omega_{R/k}=S^3 is zero and ΩS/k≅S3\Omega_{S/k}\cong S^3, proving (15).

Proof of Theorem 1.4, (2)⇒\Rightarrow(1). If [k:kp]>p[k:k^p]>p, then kk contains two pp-independent elements a,ba,b, and the ring SS of Proposition 7.2 is a reduced local kk-algebra essentially of finite type with free module of differentials which is not a complete intersection. Hence (2) fails. This includes fields of infinite imperfection degree.

Shanghai Institute for Mathematics and Interdisciplinary Sciences (SIMIS), Shanghai 200433, China

Email address: [email protected]

Fakultät 8 – Mathematik und Physik, Universität Stuttgart, Stuttgart, Germany

Email address: [email protected]

References

  1. [2]D. A. Buchsbaum and D. Eisenbud, Some structure theorems for finite free resolutions, Adv. Math. 12 (1974), 84–139. doi:10.1016/S0001-8708(74)80019-8.DOI
  2. [3]B. Briggs, E. Grifo and J. Pollitz, Bounds on cohomological support varieties, Trans. Amer. Math. Soc. Ser. B 11 (2024), 703–726. doi:10.1090/btran/182.DOI
  3. [4]B. Briggs, Vasconcelos’ conjecture on the conormal module, Invent. Math. 227 (2022), no. 1, 415–428. doi:10.1007/s00222-021-01070-0.DOI
  4. [5]D. Eisenbud, Homological algebra on a complete intersection, with an application to group representations, Trans. Amer. Math. Soc. 260 (1980), no. 1, 35–64. doi:10.1090/S0002-9947-1980-0570778-7.DOI
  5. [7]J. Herzog, Homological properties of the module of differentials, in: Atas da 6ª Escola de Álgebra (Recife, 1980), Coleção Atas, Sociedade Brasileira de Matemática, 1981. Reprinted with an appendix by B. Briggs and S. B. Iyengar as arXiv:2502.14159.
  6. [8]H. Matsumura, Commutative ring theory, Cambridge Studies in Advanced Mathematics 8, Cambridge University Press, Cambridge, 1986.DOI
  7. [9]C. P. Mok and Y. Xu, A counterexample to the Eisenbud–Mazur conjecture, preprint, Hexagon 2610.00384v1, 2026.hub.valency.io/works/W_yyc6af9c
  8. [10]The Stacks Project Authors, Stacks Project, https://stacks.math.columbia.edu, 2018.
  9. [11]J. Tate, Homology of Noetherian rings and local rings, Illinois J. Math. 1 (1957), no. 1, 14–27. doi:10.1215/ijm/1255378502.DOI
  10. [12]W. V. Vasconcelos, On the homology of I/I², Comm. Algebra 6 (1978), no. 17, 1801–1809. doi:10.1080/00927877808822322.DOI

Paper details

vlnc:W_jz2pnafm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

OpenAI ChatGPT GPT-6 Astra, OpenAI ChatGPT GPT-6.2 Astra, Anthropic Claude Opus 5.5, Anthropic Claude Fable 5.1

View statement
Looking for related papers…
Contents