Regular rings and finite projective resolutions

Recall that a local ring (R, m, K) is regular if its embedding dimension, dimK (m/m2 ),
which may also be described as the least number of generators of the maximal ideal m, is
equal to its Krull dimension. This means that a minimal set of generators of m is also a
system of parameters. Such a system of parameters is called regular. Another equivalent
condition is the the associated graded ring of R with respect to m be a polynomial ring,
in which case the number of variables is the same as dim(R). A regular local ring is a
domain.
The following fact about regular local rings comes up frequently.
Proposition. Let (R, m, K) be a regular local ring. Let J ⊆ m be a proper ideal of R.
Then R/J is regular if and only if J is generated by part of a minimal set of generators
for m, i.e., part of a regular system of parameters. (This is true if J = 0, since we may
take the set to be empty.)
Proof. If J is generated by x1 , . . . , xk , part of a minimal set of generators for m, then x1 is
not in an minimal prime, since R is a domain, and both the dimension and the embedding
dimension of R/x1 R are one less than the corresponding number for R. It follows that
R/x1 R is again regular, and the full result follows by a straightforward induction on k.
To prove the other direction, we also use induction on dim(R). The case where dim(R) =
0 is obvious. Suppose dim(R) > 0 and 0 6= J ⊆ m2 . Then R/J is not regular, for
its dimension is strictly less than that of R, but its embedding dimension is the same.
Thus, we may assume instead that there exists an element x1 ∈ J with x1 ∈ / m2 , so
that x1 is part of a minimal set of generators for m. Then R/x1 R is again regular, and
(R/x1 R)/(J/x1 R) ∼ = R/J is regular. It follows that J/x1 R is generated by part of a
minimal system of generators x2 , . . . , xk for m/x1 R, where xj is the image in R/x1 R of
xj ∈ m, 2 ≤ j ≤ k. But then x1 , . . . , xk is part of a minimal set of generators for m. 

Projective resolutions
We want to characterize regular local rings in terms of the existence of finite free reso-
lutions. Note that over a local ring, a finitely generated module is flat iff it is projective
iff it s free.
Over any ring R, every module has a projective resolution. That is, given M , there is
a (usually infinite) exact sequence · · · → Gn → · · · → G1 → G0 → M → 0, such that all
of the Gi are projective. In fact, we may take them all to be free.
One can construct a free resolution as follows.
L First choose a set of generators {uλ }λ∈Λ
for M , and then map the free module G0 = λ∈Λ Rbλ on a correspondingly indexed set of
generators {bλ }λ∈Λ onto M : there is a unique R-linear map G0  M that sends bλ 7→ uλ
1
2

for all λ ∈ Λ. Whenever we have such a surjection, the kernel M1 of P  M is referred

to as a first module of syzygies of M . If 0 → M1 → P0 → M → 0 is exact with P0 free,
we may repeat the process and form an exact sequence 0 → M2 → P1 → M1 → 0. Then
the sequence 0 → M2 → P1 → P0 → M → 0 is also exact, where the map P1 → P0 is the
composition of the maps P1  M1 and M1 ,→ P0 .
Recursively, we may form short exact sequences 0 → Mn → Pn−1 → Mn−1 → 0 for all
n ≥ 1 (where M0 = M ), and then one has that every n ≥ 1, the sequence

(∗) 0 → Mn → Pn−1 → Pn−2 → · · · → P3 → P2 → P1 → P0 → M → 0

is exact. A module Mn that occurs in such an exacct sequence (∗) in which all the Pi are
projective modules is called an n th module of syzygies of M . Equvialently, an n th module
of syzygies may be defined recursively as a first module of syzygies of any n − 1 st module
of syzygies. Note that the (usually infinite) sequence

(∗∗) · · · → Pn → Pn−1 → · · · → P3 → P2 → P1 → P0 → M → 0

is exact as well, and so is a projective resolution of M .

A projective resolution is called finite if Pn = 0 for all n  0. If M has a finite projective
resolution, it is said to have finite projective dimension. The projective dimension of M is
defined to be −1 if M = 0 and to be 0 if M is nonzero and projective. In general, if M has
finite projective dimension and is not projective, the projective dimension n of M , which we
denote pdR M (or simply pd M if R is understood from context), is the smallest integer n
for which one can find a finite projective resolution 0 → Pn → · · · → P1 → P0 → M → 0. If
M does not have finite projective dimension, it is said to have infinite projective dimension,
and one may write pd M = ∞.
If R is Noetherian and M is finitely generated, one may also construct a module of syz-
gies by mapping a finitely generated free module onto M . The first module of syzygies will
then be a submodule of this finitely generated free module, and, hence, finitely generated
again. Therefore, M has a free resolution by finitely generated free modules.
In the sequel, we prove the following result:
Theorem. Let (R, m, K) be a local ring. Then the following conditions are equivalent.
(1) R is regular.
(2) The residue field K = R/m has a finite free resolution.
(3) Every finitely generated R-module has a finite free resolution.
Before giving a proof, which will be based on elementary properties of Tor, we note an
important consequence of this characterization of regularity.
Corollary. If R is a regular local ring Q is a prime ideal of R, then RQ is regular.
Proof. Since R is regular, R/Q has a finite R-free resolution by R-modules. We may then
localize at Q to obtain a finite RQ -free resolution of the residue class field RQ /QRQ ∼
=
(R/Q)Q . 
 3

Thus, a Noetherian ring has the property that its localization at every prime ideal is
regular if and only if it has the property that its localization at every maximal ideal is
regular. A Noetherian ring with these equivalent properties is called regular.

Minimal free resolutions over local rings

Let (R, m, K) be local. We keep the notation of the preceding section. In constructing
a free resolution for a finitely generated R-module M , we may begin by choosing a minimal
set of generators for M . Then, at every stage, we may choose a minimal set of generators
of Mn−1 and use that minimal set to map a free R-module onto Mn−1 . A resolution
constructed in this way is called a minimal free resolution of M . Thus, a free resolution
· · · → Pn → · · · → P1 → P0 → M → 0 iof M is minimal precisely if every Pn that occurs
has a free basis that maps to a minimal set set of generators of the image Mn of Pn .
Our discussion shows that minimal free resolutions exist. We also note the following
fact: a free resolution → Pn → Pn−1 → · · · P1 → P0 → M → 0 (where the Pj are finitely
generated free R-modules) is minimal if and only if for all n ≥ 1, the image of Pj in Pj1 is
contained in mPj−1 . The reason for this is that image of Pj consists of elements of Pj−1
that give generators for the relations on the generators of Mj−1 . These generators will be
minimal generators if and only if they have no relation with a coefficient that is a unit,
i.e., all of the generating relations are in mPj−1 . An equivalent way to phrase this is that
entries of matrices for the maps Pj → Pj−1 have all of their entries in m.
Recall that Torn (M, N ) may be defined as the homology module at the n th spot of
the complex P• ⊗R N , where P• is the complex · · · → Pn → · · · → P0 → 0 obtained by
replacing M by 0 in a projective resolutons · · · → Pn → · · · → P0 → M → 0. It is is
independent of the specific projective resolution chosen up to canonical isomorphism. We
assume familiarity with a few basic properties of Tor over a ring R as described in the
supplement entitled Exact sequences with a flat cokernel and a sketch of properties of Tor.
The specific facts that we need about TorR (we frequently omit the superscript) are these:
(1) If · · · → Pn → · · · → P0 → M is a projective resolution of M , then Torn (M, N ) is
the homology of the complex · · · → Pn ⊗ N → · · · → P0 ⊗ N → 0 at the Pn ⊗ N spot.
Hence, Torn (M, N ) = 0 if n > pd M .
(2) Torn (M, N ) = 0 for n < 0, and Tor0 (M, N ) ∼ = M ⊗ N.
(3) Torn (M, N ) ∼ = Tor n (N, M ).
(4) T orn (M, ) (respectively, Torn ( , M )) is a covariant functor from R-modules to
R-modules
(5) If 0 → M 0 → M → M 00 → 0 is exact there is a long exact sequence:
· · · → Torn (M 0 , N ) → Torn (M, N ) → Torn (M 00 , N ) → Torn−1 (M 0 , N ) → · · ·
(6) The map induced on Torn (M, N ) by multiplication by r ∈ R on M (or N ) is multi-
plication by r.
(7) The module Torn (M, N ) is killed by AnnR M + AnnR N .
(8) If M or N is flat (e.g., if either is free or projective), then Torn (M, N ) = 0 for n ≥ 1
(9) If R is Noetherian and M , N , are finitely generated, so is Torn (M, N ) for all n.
From our discussion of minimal resolutions we obtain:
4

Theorem. Let M be a finitely generated module over a local ring (R, m, K). The modules

Tori (M, K) are finite-dimensional vector spaces over K, and dimK Tori (M, K) is the
same as the rank of the i th free module in a minimal free resolution of M .
Moreover the following conditions on M are equivalent:
(1) In a minimal free resolution P• of M , Pn+1 = 0.
(2) The projective dimension of M is at most n.
(3) Torn+1 (M, K) = 0.
(4) Tori (M, K) = 0 for all i ≥ n + 1.
It follows that a minimal free resolution of M is also a shortest possible projective
resolution of M . In particular, M has finite projective dimension (respectively, infinite
projective dimension) if and only its minimal free resolution is finite (respectively, infinite.)
Proof. If we take a minimal free resolution P• of M , because the image of every Gj is in
mGj−1 , when we apply ⊗R K the maps become 0, while Gi ⊗ K is a vector space Vi
over K whose dimension is the same as the rank of Gi . Hence, the homology of P• ⊗R K
at the i th spot is Vi , and the first statement follows. It is clear that (1) ⇒ (2) ⇒ (3) ⇒
(1) (note that once one of the Pj is 0, all the Pk for k ≥ j are 0). The last implication
follows from the first assertion of the Theorem. It is also clear that (1) ⇒ (4) ⇒ (3) ⇒ (1).
There cannot be a projective resolution shorter than the minimal resolution, for if Pj 6= 0
then Torj (M, K) 6= 0, and if there were a shorter resolution it could be used to compute
Torj (M, K), which would have to vanish. The final statement is then clear. 
We shall next use some elementary facts about Tor to prove that over a regular local
ring, every finitely generated module has finite projective dimension. We first note:
Proposition. Let R be a ring and let x ∈ R an element.
(a) Given an exact sequence Q• of modules

· · · → Qn+1 → Qn → Qn−1 → · · ·

(it may be doubly infinite) such that x is a nonzerodivisor on all of the modules Qn ,
the complex Q• obtained by applying ⊗ R/xR, which we may alternatively describe
as
· · · → Qn+1 /xQn+1 → Qn /xQn → Qn−1 /xQn−1 → · · · ,
is also exact.
(b) If x is a nonzerodivisor in R and is also a nonzerodivisor on the module M , while
xN = 0, then for all i, TorR ∼ R/xR
i (M, N ) = Tori (M/xM, N ).
x·
Proof. (a) We get a short exact sequence of complexes 0 −
→ Q• −→ Q• − → Q• −→ 0 which,
x·
at the n th spots, is 0 −
→ Qn −→ Qn − → Qn /xQn − → 0 (exactness follows because x is a
nonzerodivisor on every Qn ). The snake lemma yields that

· · · → Hn (Q• ) → Hn (Q• ) → Hn (Q) → Hn−1 (Q• ) → · · ·

is exact, and since Hn (Q• ) and Hn−1 (Q• ) both vanish, so does Hn (Q• ).
 5

(b) Consider a free resolution (∗) · · · → Pn → · · · P0 → M → 0 for M . By part (a),

this remains exact when we apply ⊗R R/xR, which yields a free resolution of M/xM
over R/xR. Let P• be the complex (∗) with M replaced by 0. Then TorR n (M, N ) is the
homology at the n th spot of P• ⊗R N . Since x kills N , (R/xR) ⊗ ∼

R/xR N = N . Thus,
Torn (M, N ) is the homology at the n th spot of P• ⊗R (R/xR) ⊗R/xR N , and since
R/xR
P• ⊗R R/xR is a free resolution of M/xM over R/xR, this is also Torn (M, N ). 
We can now prove:
Theorem. If (R, m, K) is a regular local ring of Krull dimension d, then for every finitely
generated R-module M , the projective dimension of M is at most d.
Proof. We use induction on dim(R). If dim(R) = 0, then the maximal ideal of R is
generated by 0 elements, and is a field, so that every R-module is free and has projective
dimension at most 0.
Now suppose dim(R) ≥ 1. Let M be a finitely generated R-module. It suffices to prove
that Torn (M, K) = 0 for n > d. We can form a short exact sequence 0 → M1 → P →
M → 0 where P is free. Since M1 ⊆ P , if we choose a regular paramter x ∈ M , x is
not a zerodivisor on M1 . Hence, TorR ∼ R/xR
n (M1 , K) = Torn (M1 /xM1 , K) by the preceding
Proposition. The long exact sequence for Tor coming from the short exact sequence 0 →
M1 → P → M → 0 shows that TorR ∼ R ∼ R/xR
n+1 (M, K) = Torn (M1 , K) = Torn (M1 /xM1 , K)
for n ≥ d, and the last term vanishes by the induction hypothesis, since R/xR is again
regular. 
The converse is much more difficult. We need several preliminary results. We write
pdR M or, if R is understood from context, pd M for the projective dimension of M over
R. We first note:
Theorem. Let (R, m, K) be a local ring.
Given a finite exact sequence of finitely generated R-modules such that every term but
one has finite projective dimension, then every term has finite projective dimension.
In particular, given a short exact sequence

0 → M2 → M1 → M0 → 0

of finitely generated R-modules, if any two have finite projective dimension over R, so does
the third. Moreover:
(a) pd M1 ≤ max {pd M0 , pd M2 }.
(b) If pd M1 < pd M0 are finite, then pd M2 = pd M0 − 1. If pd M1 ≥ pd M0 , then pd M2 ≤
pd M1 .
(c) pd M0 ≤ max{pd M1 , pd M2 + 1}.
Proof. Consider the long exact sequence for Tor:

· · · → TorR R R
n+1 (M1 , K) → Torn+1 (M0 , K) → Torn (M2 , K)

→ TorR R
n (M1 , K) → Torn (M0 , K) → · · ·
6

If two of the Mi have finite projective dimension, then two of any three consecutive terms
are eventually 0, and this forces the third term to be 0 as well.
The statements in (a), (b), and (c) bounding some pd Mj above for a certain j ∈ {0, 1, 2}
all follow by looking at trios of consecutive terms of the long exact sequence such that the
middle term is TorR n (Mj , K). For n larger than the specified upper bound for pdR Mj , the
Tor on either side vanishes. The equality in (b) for the case where pd M1 < pd M0 follows
because with n = pd M0 − 1, TorR R
n+1 (M0 , K) injects into Torn (M2 , K).

The statement about finite exact sequences of arbitrary length now follows by induction
on the length. If the length is smaller than three we can still think of it as 3 by using terms
that are 0. The case of length three has already been handled. For sequences of length 4
or more, say
0 → Mk → Mk−1 → · · · → M1 → M0 → 0,
either Mk and Mk−1 have finite projective dimension, or M1 and M0 do. In the former
case we break the sequence up into two sequences

0 → Mk → Mk−1 → B → 0

and
(∗) 0 → B → Mk−2 → · · · → M1 → M0 → 0.
The short exact sequence shows that pd B is finite, and then we may apply the induction
hypothesis to (∗). If M1 and M0 have finite projective dimension we use exact sequences

0 → Z → M1 → M0 → 0

and
0 → Mk → Mk−1 → · · · → M2 → Z → 0
instead. 
Lemma. If M has finite projective dimension over (R, m, K) local, and m ∈ Ass (R),
then M is free.
Proof. If not, choose a minimal free resolution of M of length n ≥ 1 and suppose that the
left hand end is
A
0 → Rb −→ Ra − → ···
where A is an a × b matrix with entries in m. The key point is that the matrix A cannot
give an injective map, because if u ∈ m − {0} is such that AnnR u = m, then A kills a
column vector whose only nonzero entry is u. 
Lemma. If M has finite projective dimension over R, and x is not a zerodivisor on R
and not a zerodivisor on M , then M/xM has finite projective dimension over both R and
over R/xR.
Proof. Let P• be a finite projective resolution of M over R. Then P• ⊗R R/xR is a finite
complex of projective R/xR-modules whose homology is TorR n (M, R/xR), which is 0 for
 7

n ≥ 1 when x is not a zerodivisor on R or M . This gives an (R/xR)-projective resolution

of M over R/xR. The short exact sequence
x
0→P −
→ P → P/xP → 0
shows that each P/xP has projective dimension at most 1 over R, and then M/xM has
finite projective dimension over R by the Proposition above. 
Lemma. Let (R, m, K) be local, let In denote the n × n identity matrix over R, let x be an
element of m − m2 , and let A, B be n × n matrices over R such that xIn = AB. Suppose
that every entry of A is in m. Then B is invertible.
Proof. We use induction on n. If n = 1, we have that (x) = (a)(b) = (ab), where a ∈ m.
Since x ∈/ m2 , we must have that b is a unit. Now suppose that n > 1. If every entry of B
is in m, the fact that xIn = AB implies that x ∈ m2 again. Thus, some entry of B is a
unit. We permute rows and columns of B to place this unit in the upper left hand corner.
We multiply the first row of B by its inverse to get a 1 in the upper left hand corner. We
next subtract multiples of the first column from the other columns, so that the first row
becomes a 1 followed by a string of zeros. We then subtract multiples of the first row from
the other rows, so that the first column becomes 1 with a column of zeros below it. Each
of these operations has the effect of multiplying on the left or on the right by an invertible
n × n matrix. Thus, we can choose invertible n × n matrices U and V over R such that
B 0 = U BV has the block form  
0 1 0
B = ,
0 B0
where the submatrices 1, 0 in in the first row are 1 × 1 and 1 × (n − 1), respectively, while
the submatrices 0, B0 in the second row are (n − 1) × 1 and (n − 1) × (n − 1), respectively.
Now, with
A0 = V −1 AU −1 ,
we have
A0 B 0 = V −1 AU −1 U BV = V −1 (AB)V = V −1 (xIn )V = x(V −1 In V ) = xIn ,
so that our hypothesis is preserved: A0 still has all entries in m, and the invertibility of B
has not been changed. Suppose that
 
0 a ρ
A =
γ A0
where a ∈ R (technically a is a 1 × 1 matrix over R), ρ is 1 × (n − 1), γ is (n − 1) × 1, and
A0 is (n − 1) × (n − 1). Then
   
0 0 a(1) + ρ(0) a(0) + ρB0 a ρB0
xIn = A B = =
γ(1) + A0 (0) γ(0) + A0 B0 γ A0 B0
from which we can conclude that xIn−1 = A0 B0 . By the induction hypothesis, B0 is
invertible, and so B 0 is invertible, and the invertibility of B follows as well. 
The following is critical in proving that if K has finite projective dimension over
(R, m, K) then R is regular.
8

Theorem. If M is finitely generated and has finite projective dimension over the local
ring (R, m, K), and x ∈ m − m2 kills M and is not a zerodivisor in R, then M has finite
projective dimension over R/xR.
Proof. We may assume M is not 0. M cannot be free over R, since xM = 0. Thus, we may
assume pdR M ≥ 1. We want to reduce to the case where pdR M = 1. If pdR M > 1, we
can think of M as a module over R/xR and map (R/xR)⊕h  M for some h. The kernel
M1 is a first module of syzygies of M over R/xR. By part (b) of the second Theorem on
p. 5, pdR M1 = pdR M − 1. Clearly, if M1 has finite projective dimension over R/xR, so
does M . By induction on pdR M we have therefore reduced to the case where pdR M = 1.
To finish the proof, we shall show that if x ∈ m − m2 is not a zerodivisor in R, xM = 0,
and pdR M = 1, then M is free over R/xR.
Consider a minimal free resolution of M over R, which will have the form

A
0 → Rn −→ Rk → M → 0

where A is an k × n matrix with entries in m. If we localize at x, we have Mx = 0, and so

0 → Rxn → Rxk → 0

is exact. Thus, k = n, and A is n × n. Let ej denote the j th column of the identity matrix
In . Since xM = 0, every xej is in the image of A, and so we can write xej = Abj for
a certain n × 1 column matrix bj over R. Let B denote the n × n matrix over R whose
columns are b1 , . . . , bn . Then xIn = AB. By the preceding Lemma, B is invertible, and
so A and AB = xIn have the same cokernel, up to isomorphism. But the cokernel of xIn
is (R/xR)⊕n ∼= M = Coker (A), as required. 
We can now prove the result that we are aiming for, which completes the proof of the
Theorem stated at the end of the previous lecture.
Theorem. Let (R, m, K) be a local ring such that pdR K is finite. Then R is regular.
Proof. If m ∈ Ass (R), then we find that K is free. But K ∼
= Rn implies that n = 1 and R
is a field, as required. We use induction on dim (R). The case where dim (R) = 0 follows,
since in that case m ∈ Ass (R).
Now suppose that dim (R) ≥ 1 and m ∈ / Ass (R). Then m is not contained in m2 nor
any of the primes in Ass (R), and so we can choose x ∈ m not in m2 nor in any associated
prime. This means that x is not a zerodivisor in R. By the preceding Theorem, the fact
that K has finite projective dimension over R implies that it has finite projective dimension
over R/xR. By the induction hypothesis, R/xR is regular. Since x ∈ / m2 and x is not
a zerodivisor, both the least number of generators of the maximal ideal and the Krull
dimension drop by one when we pass from R to R/xR. Since R/xR is regular, so is R. 
We have finially proved the result we were aiming for, and we have now completed the
argument given much earlier that a localization of a regular local ring is regular. We also
note:
 9

Theorem. Let R → S be a faithfully flat homomorphism of Noetherian rings. If S is

regular, then R is regular.
Proof. Let P be a maximal ideal of R. Then P S 6= S, and there is a maximal ideal Q of
S lying over P . It suffices to show that every RP is regular, and we have that RP → SQ
is flat and local. Thus, we have reduced to the case where R and S is local and the map
is local. Take a minimal free resolution of R/P over R. If R is not regular, this resolution
is infinite. Apply S ⊗R . Since S is R-flat, we get a free resolution of S/P S over S.
Since P maps into Q, this resolution is still minimal. Thus, S/P S has infinite projective
dimension over S, contradicting the fact that S is regular.