MATH 249B NOTES: ALTERATIONS

LECTURES BY BRIAN CONRAD,

NOTES BY TONY FENG AND AARON LANDESMAN

C ONTENTS
1. Resolution of singularities 2
2. Preparations for the proof of de Jong’s Theorem 9
3. Preliminary reduction steps 17
4. Construction of good curve fibrations 21
5. Projection. . . What’s the point? 24
6. Three-point lemma 37
7. Passage to a union of sections 42
8. Stable curves 43
9. Moduli of stable curves 51
10. Application of moduli stacks 63
11. Semistable curves over a discrete valuation ring 78
12. Resolving semistable curves over a regular base 95
References 116

1
2 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

1. R ESOLUTION OF SINGULARITIES
1.1. First definition. We first introduce the notion of resolution of singular-
ities, which we will later refine and modify.
Definition 1.1.1. Let X be a reduced (locally) noetherian scheme. A resolu-
tion of singularities of X is a map f : X 0 → X such that
(1) X 0 is regular,
(2) f is proper,
(3) f is birational: there exists a dense open U 0 ⊂ X 0 and U ⊂ X such
that f restricts to an isomorphism
F |U 0 : U 0 ' U.
Exercise 1.1. Show that f is birational if and only if there exists a dense open
U ⊂ X such that f −1 (U ) ⊂ X 0 is dense open and f −1 (U ) ' U. [Hint: shrink
U in the definition by removing f ( X 0 − U 0 ).]
There are stronger notions of resolution of singularities. To motivate
them, let us first mention that for “nice” X (where nice means “excellent”),
the regular set Reg( X ) ⊂ X, defined as { x ∈ X : Ox regular}, is Zariski open
in X. If X is finite type over an algebraically closed field, this can be seen
concretely in terms of Jacobians.
For resolutions of singularities, one could further ask for the open sub-
scheme U in (1.1) to be the full regular locus, and ( X 0 − f −1 (U ))red ⊂ X 0 to
have nice geometric structure, namely to be a strict normal crossings divisor
(to be defined soon).
Remark 1.2. At the end of the day, we’ll be most interested in varieties over
an algebraically closed fields. However, in the middle of arguments one
sometimes passes out of this realm (such as to work over completions of lo-
cal rings), so it’s useful to have a more general framework. Also, de Jong’s
method in [deJ] adapts to apply to a relative situation over discrete valua-
tion rings (so, for schemes which are not over a field); this is important for
arithmetic applications and is addressed in the final part of his original pa-
per (and for reasons of time we will say nothing about that in this course,
but we will treat some intermediate steps in wide enough generality to be
applicable in the adaptation to working over discrete valuation rings).
1.2. Strict normal crossings divisors.
Definition 1.2.1. Let S be a regular scheme. A strict normal crossings divisor
(sncd) in S is a closed subscheme D defined by an invertible ideal ID ⊂ OS
(=: effective Cartier divisor) such that
(1) D is reduced,
(2) the (reduced) irreducible components { Di }i∈ I of D are regular,
 MATH 249B NOTES: ALTERATIONS 3

(3) (most important) for any finite subset J ⊂ I,

\
Dj = Dj
j∈ J

is regular of codimension #J at all points; i.e. for all ξ ∈ D J

dim OD J ,ξ = dim OS,ξ − #J.
Exercise 1.3. Show that the definition is equivalent to saying that for all
ξ ∈ D and the set { Di }i∈ Iξ of Di ’s through ξ, we have
IDi ,ξ = ti OS,ξ
for such i with {ti }i∈ Iξ part of a regular system of parameters of OS,ξ .

One unfortunate feature of the definition sncd is that it is not “very local”;
i.e., not local for the étale or “analytic” topologies. This is due to the basic
but fundamental fact that irreducibility is not étale-local (or more concretely,
a completion of a noetherian local domain need not be a domain):
Example 1.4. Let D = {y2 = x2 ( x + 1)} ⊂ A2k , for char k 6= 2. This integral
divisor fails to be a sncd because it is non-regular at the origin. However,
“very locally” at the origin (namely, upon passing to the finite étale cover
near (0, 0) defined by u2 = 1 + x) it looks like the intersection of two trans-
versely intersecting regular divisors y = ±ux, which is what an sncd looks
like at its non-regular points. Another way to make this precise is to pass to
completions: Spec ObA2 ,0 ⊃ Spec ObD,0 is a sncd (since the completion con-
√
tains 1 + x, by Hensel’s Lemma).
Definition 1.2.2. A reduced Cartier divisor D in a regular scheme S is a nor-
mal crossings divisor (ncd) if for all ξ ∈ D, there exists an étale neighborhood
S0 → S of ξ such that such that (S0 , D 0 := S0 ×S D ) is an sncd.
Remark 1.5. For S excellent, D ⊂ S is an ncd if and only if for all ξ ∈ D the
closed subscheme Spec ObD,ξ ⊂ Spec ObS,ξ is an ncd. This is not obvious; it
uses Artin-Popescu approximation (which we will discuss later); the hard
part is to show that being an sncd after completion at a point ensures the ex-
istence of an étale neighborhood (much less drastic than completion!) over
which D becomes an sncd. This is useful in practice because completions
are often a more convenient framework to perform concrete calculations
(e.g., the nature of étale maps simplifies a lot after passing to completions).
To get back to a sncd from a ncd at the end of the day, we will need to
know the relationship between the two notions.
4 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Exercise 1.6. Show that a normal crossings divisor is a strict normal crossing
divisor precisely when its irreducible components Di (given the reduced
structure) are regular.
1.3. Hironaka’s Theorem. Hironaka proved that in characteristic 0, vari-
eties have the best imaginable resolutions of singularities. Here is a sample
version of his result (other variants are also available, in analytic settings
and for “embedded resolution”, etc.):
Theorem 1.3.1 (Hironaka). Let X be reduced, separated of finite type over
a field k of characteristic 0. There exists a resolution of singularities X 0 → X
such that
(1) f is an isomorphism over U := Reg( X ) = X sm ⊂ X.
(2) ( X 0 − f −1 (U ))red ⊂ X 0 is ncd.
Remark 1.7. What Hironaka proved was much stronger. For example, he
showed that f can be built as a composition of blow-ups along smooth
closed subschemes (we will discuss blow-ups in detail later).
It is a wide open question whether this holds in characteristic p > 0. de
Jong’s alterations provide a substitute notion that works in most applica-
tions, in any characteristic (and even in some relative situations).
1.4. Alterations. For most applications, one can weaken the birationality
aspect of resolution of singularities as follows.
Definition 1.4.1. Suppose X is an integral noetherian scheme. An alteration
of X is a map f : X 0 → X such that
(1) X 0 is integral and regular,
(2) f is proper and dominant,
(3) [k ( X 0 ) : k ( X )] < ∞.
Exercise 1.8. Show that the third condition is equivalent to the existence of a
dense open U ⊂ X such that f −1 (U ) → U is finite (and flat, after shrinking
again); note that f −1 (U ) is automatically dense in X 0 since it is a non-empty
open subset (by dominance of f ) and X 0 is assumed to be irreducible (even
integral).
deJong’s main theorem is a weakened form of Hironaka’s Theorem, with
alterations in place of resolution of singularities, in which one loses all con-
trol over the open subscheme U over which the morphism is actually finite
(in particular, we cannot guarantee that it contains any particular closed
point in the regular locus):
Theorem 1.4.2. Let X be a variety over a field k; i.e., an integral separated
scheme of finite type over k. For Z ⊂ X any proper closed subset, there
 MATH 249B NOTES: ALTERATIONS 5

exists an alteration f : X 0 → X and an open immersion

0
J : X 0 ,→ X
such that
0
(1) X is regular and projective,
0 0
(2) f −1 ( Z ) ∪ ( X − j( X 0 )))red ⊂ X is sncd.
| {z }
complement of f −1 ( X − Z )

Remark 1.9. In the statement of the theorem, why not simply take Z to
be empty? The answer is that the theorem will be proved by an inductive
means, and we really need the flexibility of a closed subscheme Z for the
inductive step (even if in practice we may often only care about the case
of empty Z). Informally, the role of Z is inspired by Hironaka’s result on
embedded resolution of singularities (which we did not define).
Remark 1.10. If k is perfect, the proof will give that we can arrange for the
extension k ( X 0 )/k ( X ) to be separable (so f : X 0 → X is finite étale over
some dense open in X). The proof doesn’t control the degree of this field
extension, but Gabber later made an improvement arranging the degree not
to be divisible by any desired finite set of primes away from char(k ). In
later work, deJong proved an equivariant version of his result for the action
by finite groups. But we stress again that the method cannot control U: it
doesn’t even guarantee finiteness of f over any particular closed point of X.
1.5. Applications. We now give some applications of alterations (where
they can be used as a substitute for resolution of singularities), and also
an example where de Jong’s Theorem appears to be insufficient to replace
Hironaka’s Theorem.
1.5.1. Grauert-Remmert Theorem. Let X be a normal C-scheme, locally of fi-
nite type. Given a surjective finite étale map E → X, the analytification is a
finite-degree covering space Ean → X an . This defines a functor
Fét( X ) → Fét( X an )
between finite étale X-schemes and analytic spaces finite étale over X an
(which is categorically “the same” as finite-degree proper local homeomor-
phisms to X (C)); these finite étale maps may not be surjective.
The theorem of Grauert and Remmert is that this functor is an equiva-
lence of categories. This is amazing because it holds even when X is not
proper! The idea of the proof is to use Hironaka to reduce to the case where
X is the complement of a normal crossings divisor in a smooth projective
variety. In the projective case, we can use GAGA. The details are carried
over [SGA1, XII].
6 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Exercise 1.11. Adapt Grothendieck’s argument in [SGA1] to use just alter-

ations, rather than resolution of singularities.
1.5.2. Artin comparison theorem. Let X be a separated, finite type scheme
over C, and F a constructible abelian sheaf on Xét , or a constructible `-adic
sheaf.
Theorem 1.5.1. The natural pullback map
∼
(1.1) H∗ ( Xét , F ) −
→ H∗ ( X an , F an )
is an isomorphism.
Remark 1.12. This is also true, and much easier to prove, for cohomology
with compact supports. Why is that case easier to prove? The reason is that
cohomology with compact supports is much better behaved with respect to
stratification by locally closed sets due to the excision sequence involving
H∗c ’s throughout; in contrast, the excision sequence for ordinary cohomol-
ogy involves an anomalous term H∗Z ( X, ·) for cohomology with supports
along a closed set, and that messes up the induction. The upshot of the
more convenient excision sequence for H∗c is that it permits a reasonably
rapid reduction to the case of curves.
The proof of (1.1) uses resolution of singularities in dimension ≤ dim X.
For an analogue with non-Archimedean étale cohomology allowing posi-
tive characteristic, an entirely different proof was developed by Berkovich
that adapts almost verbatim to the complex-analytic setting above for con-
structible abelian F . So for this much, resolution of singularities is not
needed. However, it is needed to bootstrap to the `-adic case (that, oddly
enough, does not seem to be documented in the literature); this is handled
by a beautiful argument of Deligne that can also be carried out using alter-
ations.

1.5.3. Deligne’s theory of mixed Hodge structures. There is a notion of “proper

hypercovering”, which is an abstraction of Cech coverings and allows other
types of maps than “open embeddings” to be used in the study of suffi-
ciently topological cohomological investigations. In particular, due to the
proper base change theorem in étale cohomology, it is fruitful to work with
hypercoverings whose constituents are proper smooth varieties with suitable
“sncd boundary”, and such hypercovers can be built by systematic use of
de Jong’s theorem instead of Hironaka’s Theorem.

1.5.4. A non-application. Here is an example in which de Jong’s alterations

do not appear to suffice in place of Hironaka’s Theorem. Let X be a smooth
 MATH 249B NOTES: ALTERATIONS 7

scheme over C, separated of finite type. We can form the algebraic de Rham
complex
d d d
Ω•XC = (OX −
→ Ω1X/C −
→ Ω2X/C −
→ . . .)
Grothendieck defined the algebraic de Rham cohomology to be the hyper-
cohomology of this complex:
i
HdR ( X/C) = Hi (Ω•X/C ).
There’s a natural “analytification” Ω•Xan /C (being careful about the fact that
the d-maps are not OX -linear!), so one has a map
i
θX i
: HdR ( X/C) = Hi (Ω•X/C ) → Hi (Ω•Xan /C ).
By the ∂-Poincaré Lemma, the latter is isomorphic to Hi ( X (C), C). In the
proper case, GAGA combined with a spectral sequence for hypercohomol-
i is an isomorphism.
ogy implies that θ X
Grothendieck proved in IHES 29 that θ X i is an isomorphism even without

properness. A striking consequence is that if X is affine, then Hi ( X (C), C) '

Hi (Ω•X/C ) = Hi (Ω•A/C ) (the first isomorphism by Grothendieck’s result,
and the second by acyclicity of coherent sheaves on affines). That is, topo-
logical C-valued cohomology of X (C) for smooth affine X can be computed
entirely in terms of algebraic differential forms!
How did Grothendieck achieve this? The idea is to bootstrap from the
proper case by first forming a compactification X ,→ X and then using a
strong enough form of Hironaka’s theorem (applied to X) to find a reso-
0
lution of singularities X → X such that if X 0 is the pre-image of X then
0
not only is X 0 → X an isomorphism but also X − X 0 is a normal crossings
divisor. In other words, we have a Cartesian diagram

X0  / X0

  
 /
X X
whose left side is an isomorphism that thereby identifies X as the comple-
0
ment of a normal crossings divisor in the smooth proper X . Grothendieck
related the de Rham cohomology of X = X 0 to the hypercohomology of the
an
deRham complex on X 0 modified to permit controlled poles along that
normal crossings divisor.
The key point is that one knows that X is untouched by the resolution of
singularities, being in the smooth locus of X. This is what breaks down for
alterations. In de Jong’s version, one cannot control where the map restricts
8 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

to a finite morphism, let alone an étale map. This is bad because formation
of differential forms interacts badly with respect to a map that is not étale,
so one can’t relate Ω•X/C to Ω• 0 in any useful way.
X /C

Remark 1.13. Bhatt proved a result that gets around this issue in the lci case,
establishing a version of Grothendieck’s isomorphism beyond the proper
case with lci singularities by working with the so-called infinitesimal site (in
lieu of the deRham complex, to which it is closely related only in the smooth
case) and using more advanced homological tools than were available in the
1960’s.
1.6. Regular vs smooth. We want to emphasize the technical distinction
between regular and smooth. Let S be a locally finite type scheme over a
field k.
Definition 1.6.1. The scheme S is regular if all the local rings OS,s are reg-
ular. (It is sufficient to verify this at closed points s ∈ S, in view of Serre’s
theorem that the localization at a prime ideal for any regular local ring is
again regular.)
The scheme S is smooth over k if Sk is regular. There are several equivalent
formulations of this “geometric” condition:
• Sk0 is regular for all finite extensions k0 /k,
• Sk0 is regular for all extensions k0 /k,
• Sk0 is regular for one perfect extension k0 /k,
• S → Spec k satisfies the infinitesimal criterion for smoothness.
Smoothness over a field k always implies regularity, but the converse is
false (if k is not perfect). For instance, smoothness is preserved by ground
field extension, but regularity is not:
Exercise 1.14. Let k be an imperfect field of characteristic p > 0, and a ∈
k − k p . Pick m > 1 such that p - m (e.g., m = 2 when p is odd). Check
that {ym = x p − a} ⊂ A2k is Dedekind, but obviously Ck = {ym = z p } (for
z := x − a1/p ) is not regular at (0, 0).
Example 1.15. Let K/k( x ) be a finite extension. Consider the diagram

K ⊃ A
finite integral closure

K ( x) ⊃ k[ x]

Then Spec A is dense open in a unique projective k-curve C with function

field K. It can happen that C is geometrically irreducible over k yet Ck is
 MATH 249B NOTES: ALTERATIONS 9

nowhere reduced, even if k ⊂ K is relatively algebraically closed. (We de-

mand this final condition on how k sits inside K to avoid the lame example
in which k admits a non-trivial purely inseparable extension inside K.)

Exercise 1.16 (MacLane). Take k = F p (s, t), Spec A = {sx p + ty p = 1} ⊂

A2 . Show that A is Dedekind, and k ⊂ Frac( A) is relatively algebraically
closed. Evidently, after adjoining the pth roots of s and t the curve Spec A
becomes everywhere non-reduced.

In de Jong’s theorem with imperfect k one cannot expect to arrange that

the k-variety X has an alteration X 0 → X with X 0 generically smooth over k.
To see what is an obstruction to this, suppose a smooth alteration exist, so
we have between dense opens a finite flat surjection

U0 ⊃ X
finite flat

U ⊃ X0

Regularity descends across faithfully flat maps [Mat2, 23.7(i)]. So if U 0 were

k-smooth (as we could arrange if X 0 were generically k-smooth) then Uk0
were regular, so Uk would be regular, so X would be generically smooth;
equivalently, k( X )/k would be separable in the sense of field theory (i.e.,
k ( X ) ⊗k k is reduced for all finite extensions k0 /k). But this latter property
can fail even if k ⊂ k( X ) is relatively algebraically closed, as illustrated with
MacLane’s example above. Thus, regularity without generic smoothness is
the best we can hope for when k is imperfect.

Remark 1.17. The regularity will be achieved in de Jong’s proof by building

an alteration X 0 that is smooth over a finite extension (typically inseparable)
of k. That is, even if k is algebraically closed in k( X ), we do not arrange that
k is also algebraically closed in k ( X 0 ).

2. P REPARATIONS FOR THE PROOF OF DE J ONG ’ S T HEOREM

2.1. Basic outline. Our object of interest is a variety X over k. By a trivial
application of Chow’s Lemma, we may assume that X is quasi-projective.
Also, we will see that it suffices to replace k with k; eventually we will chase
coefficients in k to make the desired construction over k from one over k.
The method involves induction on d = dim X. Using normalization, the
cases d ≤ 1 are easy. In general, with d ≥ 2, we’ll blow up X appropriately
10 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

so that it is “fibered in curves”, meaning that there is a surjective map

X
f

Pdk −1

With all geometric fibers connected of dimension 1 (but possibly with nasty
singularities or non-reduced), and such that the generic fiber Xη is smooth.
Next we will use properties of the moduli stack M g,n and the Semistable
Reduction Theorem for curves (which will be discussed shortly), to pro-
duce an alteration Y → Pdk −1 and also an alteration X  X fitting into a
commutative diagram

alteration /
X X

f0 f

 
Y /
Pdk −1
alteration

where f 0 is flat with semistable geometrically connected fibers of dimension

1 (and smooth generic fiber). By induction, we can find a regular alteration
of Y 0 → Y, so X 0 := X ×Y Y 0 (which remains integral) is an alteration of X
that is semistable over a regular Y 0 (with smooth generic fiber). Then, in this
nice situation, we resolve X 0 by hand (using carefully-chosen blow-ups).

Remark 2.1. Non-trivial extensions of function fields arise in the method

(i.e., alterations rather than birational maps) due to the role of field exten-
sions in the Semistable Reduction Theorem as discussed below.

2.2. Semistable curves. We first discuss the notion of semistable curves

over a field k; the version over more general base schemes will be crucial
later on, at which time we will revisit the topic from a wider point of view.
Theorem 2.2.1. Let X be a finite type scheme over k of pure dimension 1.
For closed points x ∈ X, the following are equivalent:

(1) As k-algebras, ObXk ,x is isomorphic to k [[t]] or k[[u, v]]/(uv) for some

k-point x over x,
 MATH 249B NOTES: ALTERATIONS 11

(2) x lies in the open k-smooth locus X sm or there is an étale neighbor-

hood ( X 0 , x 0 ) → ( X, x ) that is also étale of the crossing of the coordi-
nate axes in the plane over k:
(X 0 , x0 )
ét ét
z )
( X, x ) ({uv = 0} ⊂ A2k , (0, 0))
Definition 2.2.2. We say that X is semistable at x if it satisfies the above
equivalent conditions, and is semistable if it is so at all closed points.
Remark 2.2. Condition (2) in the preceding theorem implies that k ( x )/k is
separable, due to ( X 0 , x 0 ) being an étale neighborhood over k for both ( X, x )
and of the origin in {uv = 0} (ensuring that as extensions of k, k ( x 0 ) is
finite separable over both k( x ) and k). This property of the residue field
is not automatic for isolated non-smooth points on geometrically integral
curves over imperfect fields in general. For instance, consider the curve
C = {v2 = u p − a} with a ∈ k − k p for imperfect k of characteristic p > 0.
This becomes v2 = z p over k”, and at the unique non-smooth point x ∈ C
the residue field is k ( a1/p ) over k.
Remark 2.3. The proof that (1) implies (2) in Theorem 2.2.1 is serious: it uses
Artin approximation. A reference is [FK, Ch. III, §2], where they investigate
the structure of “relative ordinary double point singularities”. This eluci-
dates the special feature of this geometric singularity in contrast with the
example at the end of Remark 2.2: it is formally cut out by a non-degenerate
quadratic form.
At the end of [SGA7I , Exp. I] there is given Deligne’s elegant proof of:
Theorem 2.2.3 (Semistable Reduction for abelian varieties). Let A be an
abelian variety over a field K = Frac( R), where R is a discrete valuation
ring. There exists a finite separable extension K 0 /K such that for the R-finite
integral closure R0 of R in K 0 (a semi-local Dedekind domain), the Néron
model of AK0 has special fibers over R0 with identity component an exten-
sion of an abelian variety by a torus (so there is no occurrence of G a as a
subgroup of those geometric special fibers).
Remark 2.4. Although it is not stated explicitly, the proof shows that one
can take any K 0 such that AK0 [`] is K 0 -split for ` ∈ R× that is an odd prime
or 4.
Applying this to the Jacobians of curves and doing more hard work (us-
ing results of Raynaud relating Néron models of Jacobians to relative Picard
schemes), in [DM] Deligne and Mumford proved:
12 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Theorem 2.2.4 (Semistable Reduction for curves). For K, R as above and X

a smooth proper geometrically connected curve over K, there exist K 0 , R0 as
above such that XK0 = XK0 0 for X 0 a proper flat R0 -scheme with semistable
geometrically connected fibers.
Where does this X 0 come from (when Pic0X 0 /K0 has semi-abelian reduc-
K
tion)?
(1) The first step is to analyzes the “minimal regular proper model”
for XK0 . But why does any regular proper flat model exist for this
generic fiber? In the 1960’s Shafarevich and Lichtenbaum indepen-
dently showed (via inspiration from the theory of minimal surfaces
over C) that if there exists a regular proper flat model, then there is
a minimal one that is moreover unique in a precise sense when X
has positive genus. This is already a non-trivial achievement, but it
rests on the existence of some regular proper flat model, and that lies
deeper.
To get started with making a regular proper flat model, now work-
ing over K and R for ease of notation, pick a closed immersion X ,→
PnK over K. Viewing this inside PnR , let X be the schematic closure;
this is a proper flat model over R with XK = X. Pass to the normal-
ization X f; this is R-finite over X (due to some auxiliary consider-
ations with excellence over R). b It could have horrible singularities
in its special fiber, but then we can use a form of resolution of sin-
gularities for excellent surfaces: a hard theorem of Lipman over R b
(which is excellent) and a trick of Hironaka says that the process of
“normalize, then blow up the finitely many non-regular points, and
repeat” terminates.
For a reference on the proof of this resolution result, see Artin’s
article [A] on Lipman’s proof and Chinburg’s article [Ch] for the
Lichtenbaum-Shafarevich result on minimal models. We need to in-
fer information about the special fiber X0 from information about
the Néron model, at least when the Néron model has semistable re-
duction; this is what Deligne and Mumford do in [DM], building on
work of Raynaud that relates such Néron models to relative Picard
schemes.

Example 2.5. Suppose g = 1 and X (K ) 6= ∅, so X is an elliptic curve. It

turns out (but is far from obvious) that the relative smooth locus X sm coin-
cides with Néron( X ). There are two cases with non-smooth special fiber:
• Additive reduction: Xk has a cusp,
• Multiplicative reduction: Xk has a node.
 MATH 249B NOTES: ALTERATIONS 13

The second of these two cases is the “semistable reduction” case (when the
special fiber isn’t entirely smooth), and if R is complete one can always ar-
range that [K 0 : K ] | 24 in the semistable reduction theorem for such X (with
X (K ) non-empty!).
2.3. Excellence. We need to understand a technical condition called “ex-
cellence”. Its main purpose for us will be to transfer information both ways
between a noetherian local ring and its completion. The good news, as we
shall see, its that basically everything we care about is excellent (but that
miracle is not to be taken lightly, since its proof is not trivial).
A reference for what follows is [Mat1, Ch. 13] and [EGA, IV2 §5-§7] (es-
pecially §7.8ff. in loc. cit.). First, we make a definition for a class of rings
whose dimension theory exhibits the familiar features for schemes of finite
type over fields:
Definition 2.3.1. A noetherian ring A is catenary if for all primes p ⊂ p0 , all
maximal chains of primes
p = p0 ⊂ p1 ⊂ . . . ⊂ pn = p0
have the same length n. We say that A is universally catenary if all finite type
A-algebra A0 are catenary.
Example 2.6. By [Mat2, 17.4], quotients of Cohen-Macaulay rings are cate-
nary. In fact, any such ring is automatically universally catenary as well.
This is because if B  A is a surjection with B Cohen-Macaulay, then
A[ X1 , . . . , Xn ] is a quotient of B[ X1 , . . . , Xn ], which is Cohen-Macaulay.
This implies that basically everything we care about is universally cate-
nary, since basically every ring we care about is a quotient of a regular ring
(which is Cohen-Macaulay). Note in particular that by the Cohen Structure
Theorem, every complete local noetherian ring is a quotient of a regular ring,
and hence is universally catenary. (There do exist non-catenary noetherian
rings, but they are rather hard to construct. It is a non-trivial theorem that
Dedekind domains are universally catenary, so every finitely generated ring
over Z or a discrete valuation ring is catenary.)
2.3.1. G-rings. We now introduce the most important condition, since the
motivation for excellence is to codify a notion that ensures properties trans-
fer well between a ring and its completion.
Definition 2.3.2. A noetherian ring A is a G-ring if for all primes p ⊂ A, the
(faithfully flat) map
Spec Abp → Spec Ap
has “geometrically regular” fibers. This means that for all q ∈ Spec Ap ,
the fiber algebra κ (q) ⊗ Ap A
bp is geometrically regular over κ (q); i.e., this
14 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

(visibly noetherian!) κ (q)-algebra (which is almost never finite type over

κ (q)) is regular and remains so after any finite extension on κ (q). (If this
fiber algebra were finite type over κ (q), then geometric regularity would
just be one of the equivalent definitions of κ (q)-smoothness.)
For many interesting properties of local noetherian rings, they can be
transferred up and down through a faithfully flat local map if all fiber al-
gebras satisfy the given property (and regularity subsumes essentially all
properties of interest such as: normality, CM, reducedness, etc.).
2.3.2. Excellent rings.
Definition 2.3.3. A Noetherian ring A is excellent if it satisfies the following
three conditions:
(1) A is universally catenary.
(2) A is a G-ring.
(3) For all finitely generated A-algebras A0 , the locus Reg( A0 ) ⊂ Spec A0
of regular points is open.
The most important condition for our purposes is (2).
Remark 2.7. There are some more robust equivalent versions of (3) given in
[EGA, IV2 6.12.4] and [Mat1, §32B, Thm. 73].
Remark 2.8. The combination of (2) and (3) is called quasi-excellence. Since
in practice every ring is universally catenary, there is no real difference in
practice.
Example 2.9. From the definitions, if A = R is a discrete valuation ring with
fraction field K then A is a G-ring if and only if K 0 ⊗K Kb is reduced for all
0
finite extensions K /K (which is one of the equivalent characterizations of
the typically huge non-algebraic extension K/K
b being separable in the sense
of field theory). This is of course automatic in characteristic 0.
Exercise 2.10. Check the the preceding criterion for discrete valuation rings
does hold for the local ring R = OC,c at a closed point c on a regular curve
C over a field. This is not a trivial matter, in view of the next example.
Example 2.11. There are discrete valuation rings that are not G-rings (in
characteristic p > 0, obviously). See [BLR, §3.6, Ex. 11]. for such examples.
Example 2.12. All fields and all Dedekind domains with generic character-
istic 0 are G-rings.
The following result summarizes some highlights from [EGA, IV2 , §7.9].
Theorem 2.3.4. Grothendieck-Nagata]
(1) Every complete local noetherian ring and Dedekind domain with frac-
tion field of characteristic 0 (e.g. Z) is excellent.
 MATH 249B NOTES: ALTERATIONS 15

(2) Excellence is inherited by any localization, and any finitely generated

algebra.
(3) If A is excellent and reduced then the normalization A → A e is finite.
(4) Assume A is excellent. If P is any of a long list of “nice” homological
conditions on local noetherian rings, and P( A) denotes the set of points
p ∈ Spec A such that Ap satisfies P then
• P( A) is open,
• for any ideal I ⊂ A, and A b the I-adic completion, the natural map
f : Spec Ab → Spec A satisfies f −1 (P( A)) = P( A
b).
In particular, the last assertion implies that if A is excellent local then
upon taking I to be m A we have that A satisfies P if and only if Ab does!
Definition 2.3.5. We say that a local noetherian scheme X is excellent if it has
an affine open cover X = {Spec Aα } with each Aα excellent. (This easily
implies that every affine open Spec A ⊂ X has A excellent.)
The theorem implies that every algebra finitely generated over a field or
Dedekind domain, or localizations thereof, is excellent. This covers most of
the rings we ultimately care about (but the excellence of certain completions
will be technically important later on as well).
Remark 2.13. A very interesting further class of cases are local rings on
complex analytic spaces. Rather generally, for regular Q-algebras, there is
a “Jacobian criterion” for excellence [Mat1, Thm. 101] which holds for the
local ring OCann ,0 (whose noetherianity is a serious fact from the theory of
several complex variables, and then regularity is seen by computing the
completion). Thus, passing to quotients of such rings yields that local rings
on all complex analytic spaces are excellent! This implies that if X is a locally
finite type C-scheme then for x ∈ X (C), the local ring OX,x has P if and
only if OXan ,x has P, because the two local rings are excellent and have the
same completion. This is used all the time in the study of analytification of
algebraic C-schemes.
2.3.3. Relationship between resolution of singularities and the G-ring property.
We won’t use the following dicussion, but it is good for awareness as to
why excellence (or really the G-ring property) is intimately tied up with
any general result towards resolution of singularities for a wide class of
schemes.
Theorem 2.3.6. Suppose X is a locally noetherian scheme. Assume the fol-
lowing:
for all Zariski-open U ⊂ X and finite Y → U with Y integral,
Y admits a resolution of singularities.
16 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Then X is quasi-excellent. In particular, all the local rings OX,x are G-

rings.

Remark 2.14. The statement of this result in [EGA, IV2 , 7.9.5] omits the open
subset U, but that makes the hypotheses appear to be not local enough for
the proof to work (especially since we are considering locally noetherian Y
and not just noetherian Y, so extending a finite scheme over an open subset
to a finite scheme over the entirety of X is not so clear).

Let’s discuss the essential and most striking part of the proof: that the lo-
cal rings OX,x must be G-rings. The problem is Zariski-local due to the local
nature of the hypotheses, so we may assume that X = Spec A is affine. Pick
p ∈ Spec A and q ⊂ p in A (so q corresponds to a point in Spec Ap , which
we denote by the same name). We want to show that κ (q) ⊗ Ap A bp is geo-
metrically regular over κ (q). Replacing A with ( A/q)p , we may assume that
A is a local domain (by inspecting that we may spread out finite algebras
over this localization to finite algebras over some basic affine open around
p to access the hypothesis concerning resolutuon for integral schemes finite
over an open subset of X). Let K = Frac( A). We want to show that for all
finite extensions K 0 /K, the noetherian ring K 0 ⊗ A A
b = K 0 ⊗K (K ⊗ A A b) is
regular.
First we reduce to just showing the statement for K 0 = K. We can find an
A-finite A0 ⊂ K 0 such that K 0 = Frac( A0 ), so

K0 ⊗ A A
b = K 0 ⊗ A0 ( A 0 ⊗ A A
b).

Now, A0 ⊗ A A 0 has fraction field K 0 , so we can rename

A0 m , where Am
b ' ∏ m0 c
Am0 as A and thus reduce to showing that K ⊗ A b is regular. (The A-scheme
0 A
0
Spec A plays the role of Y in the hypothesis, at least after some “spreading
out” from our current situation over a local ring.)
Now we handle the case K 0 = K. Let Z → Spec A = X be a resolution of
singularities. Consider the base change

h0 /
Z0 Z
f0 f
 
Spec A
b / Spec A
h

We want the generic fiber of h to be regular. The map f is proper bira-

tional, hence an isomorphism over a non-empty open set in Spec A, so the
same holds for f 0 over Spec A.This
b implies that the local rings of K ⊗ A A
b
 MATH 249B NOTES: ALTERATIONS 17

occur already as local rings on Z 0 , so it is enough to show that Z 0 is a reg-

ular scheme. It may seem surprising that Z 0 is regular since we have not
assumed anything particularly nice about the fiber algebras for h.
Now for the brilliant step: since A b is excellent (by Theorem 2.3.4, all com-
plete local noetherian rings are excellent!), the regular locus Reg( Z 0 ) ⊂ Z 0
is open. But Z 0 is A-proper,
b so to show that Reg( Z 0 ) = Z 0 it suffices to show
that the regular locus contains the special fiber of f 0 . Denote the special
fibers by Z0 and Z00 , so it is enough to show that the completed local rings
of Z 0 at points of Z00 are regular. But these completed local rings are limits of
the rings of functions on the infinitesimal special fibers. For z0 ∈ Z00 we have
ObZ,h0 (z0 ) ' ObZ0 ,z0 because f , f 0 have the same infinitesimal special fibers, as
Z 0 = Z ⊗ A A. b Thus, the regularity of Z (by design) saves the day.

2.4. Strict transforms. For later purposes, we introduce one final general
definition before we get started With the proof of deJong’s theorem. Let
f : X → S be a finite type separated morphism, with X and S noetherian
and integral. Let ψ : S0 → S be a proper birational morphism (so S0 is also
integral). Consider the fiber square

X ×S S0 / X
f
 
S0 / S
ψ

Let η ∈ S be the generic point of S, and η 0 ∈ S0 the generic point of S0 .

Definition 2.4.1. The strict transform X 0 of X → S with respect to ψ is the
schematic closure of Xη ×η η 0 = Xη in X ×S S0 :

X0 / X ×S S0 / X
f
# 0 
S / S
ψ

Remark 2.15. In many situations later on there will be an evident closed S0 -

flat subscheme X 0 ⊂ X ×S S0 with full generic fiber; in such cases we claim
X 0 = X 0 . The reason is that because S0 is integral, S0 -flatness implies that
X 0 is the schematic closure of its own generic fiber Xη00 .

3. P RELIMINARY REDUCTION STEPS

We are now going to start the proof.
18 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

3.1. Statement of the main theorem.

Theorem 3.1.1 (de Jong, 4.1). Let X be a variety over a field k (i.e. an integral,
separated, finite type k-scheme). Let Z ( X be a proper closed subset. Then
there exists
 j
X0  / X0
ϕ

X
With j a dense open immersion and ϕ an alteration such that
0
• X is a regular projective k-scheme,
0 0 0
• ( ϕ−1 ( Z ) ∪ ( X − j( X 0 )))red ⊂ X is sncd in X .
| {z }
complement of ϕ−1 ( X − Z )
If in addition k is perfect then we can arrange ϕ to be generically étale.
0
Remark 3.1. We view ( ϕ−1 ( Z ) ∪ ( X − j( X 0 )))red as “∂ X1 ( X1 − ϕ−1 ( Z ))”,
and will use the latter notation in the future.
Remark 3.2. We do not control the algebraic closure of k in k ( X 0 ), even if X
is geometrically integral over k. We’ll build X 0 as smooth over some finite
extension k0 /k, over which we have little control. In fact, we’ll mostly work
over algebraically closed k, and then spread things out to get the desired
result over k.
We’ll induct on d = dim X ≥ 0 (with k fixed).

If d = 0, then there is nothing to do. [In this case X = Spec k0 for a finite
extension k0 /k, and Z = ∅.]

3.2. Reduction to algebraically closed base field. Let’s grant the full result
in dimension d over k, and deduce it over k. Consider ( Xk )red . This has pure
dimension d, because X does.
Pick an irreducible component X 0 ⊂ ( Xk )red , with the reduced scheme
structure. It is easy to check that π : X 0  X is surjective, so Z 0 = π −1 ( Z ) (
X 0 . By the assumed Theorem 3.1.1 in dimension d over k, we have a generi-
cally étale alteration
 j10
X10  / 0
X1
ϕ10

X0
 MATH 249B NOTES: ALTERATIONS 19

0 0
where X 1 is k-smooth and projective, and ∂ X 0 ( X10 − Z10 ) is an sncd in X 1 with
1
Z10 := ( ϕ10 )−1 ( Z 0 ) ⊂ X10 .

Remark 3.3. Recall that a sncd is a union of k-smooth irreducible hypersur-

faces such that all finite intersections are k-smooth with the “expected” pure
dimension (or equivalently, “expected” pure codimension).

Now we want to bring this down to some finite extension of k. By “stan-

dard” direct limit formalism [EGA, IV3 §8, §9, §11, . . . ] since k = limα k α for
−→
[k α : k] < ∞, we can descend the entire situation over k down to some finite
extension K/k inside k: there exists a finite type K-scheme X 00 such that
• X 00 ⊗K k = X 0 inside ( Xk )red with X 00 ⊂ ( XK )red , and ( XK )red ⊗K k =
( Xk )red .

Remark 3.4. There is a technical but important point here. Exten-

sion of scalars does not preserve reducedness in general. However,
since the nilradical for a noetherian ring is finitely generated, a nil-
radical formed over k has generators on an affine open all defined
over some finite extension K/k inside k. Thus, after killing the nil-
radical over such a K we see that further extension of scalars from K
to k preserves reducedness. This is important, because after guaran-
teeing that condition we may need to further extend K to obtain the
additonal properties below.

• There exists
 j100 /
X100 
00
X1
ϕ10

X 00

with j100 an open embedding, ϕ100 a generically étale alteration, and X100
00
projective and smooth over K, and ∂( X100 − Z100 ) is a sncd in X 1 .
Now X 00 ⊂ ( XK )red is an irreducible component, and XK  X is
surjective, so X 00 ,→ ( XK )red → X is finite and generically flat:

X 00  / ( XK )red


$ 
X
20 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Therefore, the composite map

X100
ϕ10

X 00


X
is an alteration that one sees settles the problem for X. Note that by
design, X100 is smooth over the finite extension K/k over which we
had very little control.
If k is perfect then K/k is separable, so XK is reduced with XK → X
finite étale, so X 00 → X is generically étale because XK → X is (since
some dense open in X 00 is also open in XK ). The upshot is that if k is
perfect, then X100 → X is generically étale.
3.3. Reduction to projective case. By Chow’s Lemma (see [EGA, II, §5] for
a version that applies to separated schemes of finite type over a noetherian
base, not requiring any properness hypotheses), there exists a modification
π : X0 → X
with X 0 integral and quasi-projective, so after replacing ( X, Z ) with ( X 0 , Z 0 =
π −1 ( Z )) we pass to the case where X is quasi-projective. Now we can form
a compactification X ,→ X with X projective. Let Z = X − ( X − Z ) =
(closure of Z ) ∪ ( X − X ); this is also what we’ve been calling ∂ X ( X − Z ). So
Z ( X is proper closed, with Z ∩ X = Z.
Suppose we have the solution for the projective X; i.e., an alteration

X1
ϕ1

X
with X 1 smooth, and ϕ1−1 ( Z ) ⊂ X 1 a sncd. Look at X1 = ϕ1−1 ( X ) ,→ X 1 : we
claim that

X1  /X
1
ϕ 1 | X1

X
is the desired alteration. The alteration property and regularity of X 1 are
guaranteed by construction. We have to check what happens over Z. For
 MATH 249B NOTES: ALTERATIONS 21

this, note that we rigged Z (which was not just the closure of Z, but also
included the boundary of X in X) so that

∂ X1 ( X1 − ϕ1−1 ( Z )) = ∂( X 1 − ϕ1−1 ( Z ))

so it is an sncd by construction.

3.4. Reduction to blowup. If Z 6= ∅, then we can pass to BlZ ( X ) so that

Z is the underlying space of a Cartier divisor. Notice that this condition is
preserved by pre-image with respect to any alteration on X. So we may as-
sume that Z is the underlying space of a Cartier divisor (i.e., Zariski-locally
is the zero locus of a single nonzero element of a domain).

3.5. Reduction to the normalization. We can pass to the (finite birational!)

normalization X e → X, so X is normal. This completes the case d = 1,
because a normal curve is regular and any finite set of closed points in such
a curve is a sncd (!).

3.6. Increasing Z. We now show that, under our running assumptions, we

can increase Z without loss of generality. What this means is:

Lemma 3.5. If ϕ : X 0 → X is an alteration and Z ⊂ Z

e ( X such that ϕ−1 ( Z
e) ⊂
0 − 1 0
X is an sncd, then ϕ ( Z ) ⊂ X is also sncd.

Proof. Since Z is Cartier, ϕ−1 ( Z ) is also Cartier, and lies inside the sncd
ϕ −1 ( Z
e ). So for dimension reasons, ϕ−1 ( Z ) is a union of irreducible compo-
nents of the sncd ϕ−1 ( Z e ). But it is easy to see that the union of irreducible
components of an sncd is still an sncd. 
Remark 3.6. Any alteration ϕ : X 0 → X must be finite over some open set
U ⊂ X with codimX ( X − U ) ≥ 2. This is an exercise using the valuative
criterion of properness; see the handout for details. (The point is to prove
finiteness at the codimension-1 points.)

4. C ONSTRUCTION OF GOOD CURVE FIBRATIONS

Now we want to realize X as a “curve fibration” over Pd−1 at the cost of
some mild blowup.

4.1. Warmup. As a warm-up case, we consider X = Pd . For p ∈ Pd , we

have
Bl p (Pd ) = {(q, `) ∈ Pd × Pd−1 | q ∈ `}
22 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Viewing Pd−1 = { lines ` 3 p}. This has two “fibrations.”

Bl p (Pd )
π ϕ
z #
P d −1 Pd
Let’s think about what these maps look like.
• The map ϕ is an isomorphism over Pd − p, and ϕ−1 ( p) = Pd−1 .
• The map π is a P1 -bundle for the Zariski topology, e.g. π −1 (`) is the
line ` viewed as a subscheme of Pd .
Consider the closed set Z ⊂ Pd missing p, so ϕ−1 ( Z ) ' Z. We want to
think about how this sits over the Pd−1 .
• If dim Z < d − 1 (which is the case we’ll most often be interested in),
does ϕ−1 ( Z ) map birationally onto its image under π, for “generic”
p?
• If dim Z = d − 1, is ϕ−1 ( Z ) → Pd−1 generically étale (again, for
generic p)?
Since π −1 ({`}) = ` ⊂ Pd as schemes, the {`}-fiber of

π | ϕ −1 ( Z ) : ϕ −1 ( Z ) → P d −1

is ϕ−1 ( Z ) ∩ `, which is isomorphic to Z ∩ ` as schemes.

Note that Z ∩ ` ( ` is finite. By a suitable Bertini Theorem (cf. Jouanolou),
if Z is pure dimension d − 1 and reduced, so that Zsing ⊂ Pd has codimen-
sion at least 2, then there exists a dense open locus of lines ` ⊂ Pd such that
Z ∩ ` = Zsm ∩ ` is smooth of dimension 0, hence étale. To construct a gener-
ically étale projection π, first pick such an `0 , and then p ∈ `0 − (`0 ∩ Z ).
This ensures that
π : ϕ− 1
p (Z) → P
d −1

Has étale fiber at {`0 }. But then, what about flatness? This may come for
free at points, if one knows smoothness of the source and target...

4.2. The technical lemma. That was just a warm-up example to illustrate
the relevant issues. In general, we want a similar procedure using X in place
of Pd - we want to blow up X at a suitable finite subset avoiding Z ( X.
Lemma 4.1 (de Jong, 4.11). Consider pairs ( X, Z ) under our running assump-
tions, except that X is not necessarily normal. There exists a finite subset S ⊂
 MATH 249B NOTES: ALTERATIONS 23

X sm (k ) − Z (k) such that for

X 0 = BlS ( X )
f ϕ

x %
P d −1 X
The fibers of f are pure dimension 1 and generically smooth, and
F : ϕ −1 ( Z ) → P d −1
is generically étale and finite.
Furthermore, if X is normal then we can arrange that there exists a dense open
U ⊂ Pd−1 such that f −1 (U ) → U is smooth with geometrically connected fibers.
Remark 4.2. Why haven’t we assumed that X is normal? We will want to
apply this Lemma to X which are not the same as in Theorem 3.1.1.
The idea is to perform a projective version of Noether normalization to
h
find a finite map X − → Pd which is “nice” with respect to Z, and take S =
−
h ( p) where p is “nice” with respect to h( Z ) ⊂ Pd (via the warm-up §4.1).
1

“Nice” should mean, for instance, that

• h : Z → h( Z ) is birational onto its image.
• h−1 ( p) is smooth (so that the blowup of X along h−1 ( p) is still nor-
mal).
This will requiring composing “projections from points” for X ,→ PkN .
4.3. Review on projections from points. For p ∈ P N , let
π p : P N → P N −1
denote the projection map. For a fixed hyperplane H = P N −1 ⊂ P N (with
p∈/ H), we can visualize π p (q) as the (unique!) intersection point of the line
qp with H. See the visualization in Figure 1.
Desired property. For “generic” p ∈ / X, we want that
• If d = N − 1, then π p : X → P N −1 is generically étale, finite, and Z
maps birationally onto its image.
• If d < N − 1, then π p : X → P N −1 is birational onto its image, and
likewise for Z.
Observe that (π p | X )−1 ([`]) = ` ∩ X as a scheme inside X, since π − 1
p ([`]) =
` − { p} inside P N − { p}. This is how one can control the étaleness of a fiber.
The subtlety is in ensuring that the choice of genericity is retained for
compositions of such processes, which is not obvious because the situation
for each step depends on the one before.
24 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

F IGURE 1. A picture of the projection of a curve X from a

point p onto a hyperplane H.

5. P ROJECTION . . . W HAT ’ S THE POINT ?

Projection from a point can be described in coordinate-free terms, as we
now explain. As always, we follow Grothendieck’s convention that P(V ) =
Proj(Sym(V )), so this represents the functor of (isomorphism classes of)
line-bundle quotients of V; it is contravariant in V (which will be convenient
later). Suppose dim V ≥ 2 and let p ∈ P(V ) be a k-point, so p = P(V/W ) ∈
P(V ) and the open subscheme P(V ) − { p} classifies line bundle quotients
λ : VS  L over k-schemes S such that λs |Ws 6= 0 for all s ∈ S, which is to
 MATH 249B NOTES: ALTERATIONS 25

say λ carries WS onto L too. Thus, functorially we have

π p : P(V ) − P (V/W ) → P(W )
λ 7 → λ |W .
The important point to observe is how the target depends on W, and that it
is not naturally “inside” P(V ) (keep in mind that P(V ) is contravariant in
V).
Example 5.1. For V = k N +1 , p = [1, 0, . . . , 0], and W = { x0 = 0} we have
naturally P(V ) = PkN , P(W ) = PkN −1 , and π p : PkN − { p} → PkN −1 is given
by the habitual formula
[ x0 , . . . , x N ] 7 → [ x1 , . . . , x N ].
Returning to our setup of interest, we have
X ⊂ PkN − { p} → PkN −1 = {` 3 p} ,
And

−1
π p |X ({`}) = ` ∩ X
As schemes (not just as sets); this will be very important in some subsequent
calculations.
Proposition 5.2 ([deJ, Prop. 2.11]). Let k be a field and X ⊂ PkN a closed sub-
scheme that is generically smooth of pure dimension d ≤ N − 1. There exists a
dense open U ⊂ PkN − X so that for all finite separable extensions k0 /k and points
p ∈ U (k0 ), the projection
π p : Xk0 → PkN0 −1
satisfies
(α) π p is birational onto its image if d < N − 1,
( β) π p is generically étale onto PkN0 −1 if d = N − 1.
Remark 5.3. We do not assume X is irreducible so that we may apply this re-
sult to Z later on. Hence, for (α) we cannot treat the irreducible components
separately because we also need that distinct components have distinct im-
ages. For ( β) we can reduce to the irreducible case, but not for (α).
Proof. To start, let’s reduce to the case k is separably closed. Assume we
know the statement over k s , for some dense open U 0 ⊂ PkNs . Let’s now
deduce the general case. For any p ∈ PkN − X (k0 ), whether or not π p

satisfies (α) (respectively ( β)) is the same for π p ⊗k0 k s (using a choice of
embedding k0 → k s over k). Thus, it suffices to find a non-empty open
U ⊂ PkN − X such that UkS ⊂ U 0 . Certainly U 0 = Vks for some dense open
26 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

V ⊂ PKN − XK , for some finite Galois K/k inside k s , by standard limit argu-
ments (as k s is the direct limit of finite Galois extensions of k). Hence, the
intersection
 
γ∗ (V ) ⊂ PKN − XK = PkN − X
\
K
γ∈Gal(K/k)

is a Gal(K/k )-stable non-empty open subscheme. This descends to an open

subscheme U ⊂ PkN − X that does the job. Hence, without loss of generality,
now k = k s .
We first treat ( β), so d = N − 1. Note that the non-smooth locus X sing :=
X − X sm ⊂ P N has codimension at least 2 in P N . We want to find lines that
“cut X nicely.” Since X sing ⊂ PkN has codimension at least 2, it misses most
lines ` ⊂ PkN . Let V = X sm = ä Vj where Vj are the open subsets where
the smooth loci of irreducible components of X avoid the other irreducible
components of X. For most lines ` ⊂ PkN we have
` ∩ X = ` ∩ V = ä ` ∩ Vj


j

We’ll now need the following setup for a Bertini theorem, following [Jou].
Let F be a field and Z ⊂ P FN a locally closed subscheme of pure dimension
d ≤ N − 1. (It will be useful to permit quasi-projective Z, not just projec-
tive Z.) For a choice of r ≤ d, let G := Gr(r, N ) be the Grassmannian of
codimension-r subspaces of P N . Consider the universal subspace

V P N × G ⊃ ZG := Z × G
(5.1)

G
Example 5.4. For example, if F 0 /F is a field and L ∈ G( F 0 ), the map on
L-fibers is given by

VL = L P FN0
(5.2)

{ L} = Spec F 0 .
Thus, passing to fibers over the F 0 -point { L} for the diagram
V ∩ ZG ZG
(5.3)

G
 MATH 249B NOTES: ALTERATIONS 27

recovers the closed subscheme L ∩ ZF0 ,→ ZF0 .

The following “Bertini theorem” combines several results proved in the
beautiful book [Jou].
Theorem 5.5 (Bertini theorem). Consider all F 0 /F and L ∈ G( F 0 ) as above.
(i) If Z is smooth then there is a dense open subscheme Ω1 ⊂ G such that
{ L} ∈ Ω1 ( F 0 ) if and only if L ∩ ZF0 is smooth of pure dimension d − r.
(ii) If Z is geometrically reduced then there is a dense open subscheme Ω2 ⊂ G
such that { L} ∈ Ω2 ( F 0 ) if and only if L ∩ ZF0 is geometrically reduced of
pure dimension d − r.
(iii) If Z is geometrically irreducible and r ≤ d − 1, then there is a dense open
subscheme Ω3 ⊂ G such that { L} ∈ Ω3 ( F 0 ) if and only if L ∩ ZF0 is
geometrically irreducible of pure dimension d − r.
Remark 5.6. Observe that the first two conditions are local on Z, but the
third is not.
Remark 5.7. If r > d then there exists a dense open Ω ⊂ G so that for all
L ∈ Ω( F 0 ) we have ZF0 ∩ L = ∅.
We’ll now use Theorem 5.5(ii) in the case r = d = N − 1. That is, we will
slice by lines ` ⊂ P N . Again, G is the Grassmannian of all lines ` ⊂ P N and
k = k s , so any dense open subset of a non-empty smooth k-scheme has lots
of k-points. By Theorem 5.5 we get a line ` ⊂ PkN (with k = k s ) so that (with
V = X sm )
` ∩ X = ` ∩ V = ä ` ∩ Vj ,


j

with all ` ∩ Vj non-empty and étale over k. In particular, the intersection is

finite.
Fix such a line `0 . Choose any point p ∈ (`0 − `0 ∩ X ) (k) with p ∈
/ `0 ∩ X,
so for the map
π p : X → PkN −1 = {` 3 p}

we see that π − 1
p ({`0 }) is étale and meets X inside X
sm touching every ir-

reducible component of X. By openness on the source for the quasi-finite

locus of a map of finite type between noetherian schemes (i.e., openness of
the locus of points isolated in their fibers), we conclude that π p is generically
quasi-finite and hence dominant (and even surjective, due to properness).
Since π p is a dominant map, π p is flat on some dense open W. We claim
that the open locus where W meet the open complement of the support of
the coherent sheaf Ω1X/P N −1 to touch each irreducible component of X; at
28 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

such points π p is étale (flat and unramified). Thus, it suffices to show that
the complement of the support of Ω1X/P N −1 is dense in X.
Consider x0 ∈ π − 1
p ({`0 }); there is such a point x0 on each irreducible
component of X. Since the formation of Ω1 commutes with base change,
the fiber Ω1X/P N −1 ( x0 ) of the coherent sheaf Ω1X/P N −1 at the point x0 is the
same as the fiber at x0 for Ω1 on the π p -fiber through x0 . But we have used
Bertini’s theorem to ensure that this fiber scheme is étale, so we conclude
that Ω1X/P N −1 ( x0 ) = 0. By Nakayama’s Lemma, it follows that Ω1X/P N −1
vanishes near x0 .
To complete ( β), it only remains to show the locus of points p we have
been using sweeps out at least a dense open locus in P N . We shall do this
by analyzing an incidence correspondence:

( x, `) ∈ P N × G : x ∈ `

pr2
(5.4)
pr1
PN G = ` ⊂ PN .


Observe that pr1 is a Zariski P N −1 -bundle over P N , and pr2 is a Zariski P1 -

bundle over G. Letting Ω ⊂ G be the non-empty open locus from the above
application of part (ii) in Bertini’s theorem, the set of p we’re looking for is
   
pr1 pr2−1 (Ω) ∩ P N − X .
These two sets are both non-empty opens because pr1 is an open map. This
completes the treatment of ( β).
Now we move on to (α), so d ≤ N − 2.
Goal 5.8. We seek a dense open U ⊂ P N − X so that for all p ∈ U (k) the
map
π p : X → P N −1

is birational onto its image.

Without loss of generality we may assume X is reduced, since it is gener-
ically reduced, and the higher codimension non-reduced locus has no im-
pact on the birationality assertion. We first seek a good line for slicing each
irreducible component of X, where the lines will be found inside a generic
codimension-d linear subspaces W ⊂ P N (so dim W = N − d ≥ 2) that
meets X nicely.
 MATH 249B NOTES: ALTERATIONS 29

To be precise, let X j be the irreducible components of X with reduced
closed subscheme structure. We have dim X j = d (recall the assumption
that X was pure dimensional). Let
 
Vj := X sm
j − ∪ X
i6= j j
sm
∩ X i ,

a smooth dense open in X j . We seek lines ` j ⊂ PkN so that

` j ∩ X j = ` j ∩ Vj = Spec k.
Let G = Gr(d, N ) be the Grassmannian classifying linear subspaces of
PN with codimension d. By Bertini’s theorem (i.e., Theorem 5.5, where Z is
only required to be quasi-projective rather than projective) applied to each
X j and Vj , we can find W ∈ G(k ) so that for every j we have
(1) W ∩ X j is finite and non-empty,
(2) W ∩ X j − Vj = ∅ (so W ∩ X j = W ∩ Vj ),

(3) W ∩ Vj is étale (so a disjoint union of finitely many k-points since

k = k s ).
Since each W ∩ X j = W ∩ Vj is a finite non-empty set of k-points, we may
pick one such k-point q j for each j.
Recall W has dimension N − d ≥ 2. Pick a line ` j ⊂ W through q j missing
the rest of the finitely many points where W meets X = ∪ X j . Pick a k-point


p j ∈ ` j − ` j ∩ X j and consider the map
π p j : X j → P N −1 = ` 3 p j .


This map has fiber over {`} equal to the scheme ` ∩ X j that is finite since
it is closed in the line ` but misses the point p j ∈ ` − (` ∩ X j ). We have as
schemes that
π− 1


pj `j = Xj ∩ `j


= Xj ∩ W ∩ `j
= q j ' Spec k.
The upshot is that
π pj : Xj → π pj (Xj )


is a finite surjection between varieties such that its fiber scheme over ` j
is a single k-point. The final part of the next lemma completes the proof of
Proposition 5.2. 
30 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Lemma 5.9. With notation as above,

(1) π p j is birational onto its image,
(2) For each j, there exists a non-empty open Uj ⊂ PkN − X j so that all points
in Uj (k) arise as such a p j .
(3) For a suitable dense open U ⊂ ∩ j Uj ⊂ P N − X and all p ∈ U (k ), the
images π p ( X j ) are pairwise disjoint and hence
π p : X → π p (X)
is birational.
Proof. Let’s prove each in order.
(1) It suffices to show more generally that for any finite map f : X → Y
to a noetherian scheme Y and any pair of points y, η ∈ Y with y in
the closure of η (e.g., η could be the generic point if Y is irreducible),
the fiber scheme Xy and Xη that are finite over the respective residue
fields k(y) and k (η ) satisfy
degk(y) Xy ≥ degk(η ) Xη
where degk Z := dimk k [ Z ] for any finite scheme Z over a field k. In-
deed, once this is proved, it follows that the non-empty finite fiber of
π p j : X j → π p j ( X j ) has degree 1 (since we found a k-point of the tar-
get whose fiber-degree is 1), and that expresses exactly birationality
of this map between varieties.
We may assume y 6= η, so (by arguments with the Krull-Akizuki
theorem, or with blow-ups) there exists a discrete valuation ring R
and a map
Spec R → Y

carrying the generic point to η and the special point to y (but with
possibly gigantic residue field extension over each, especially at y).
The formation of the degree of a finite scheme over a field is unaf-
fected by extension of the field, so we may apply base change along
such a map to arrange that Y = Spec R.
Now X = Spec A with R-finite A. If K denotes the fraction field of
R and k denotes its residue field then our assertion is dimK ( AK ) ≤
dimk ( Ak ), which holds with A replaced by any finitely generated
R-module.
(2) Recall dim W ≥ 2, so we may pick p ∈ W not on any line joining
two points of the finite étale k-scheme W ∩ X = ä W ∩ Vj . For




q j ∈ W ∩ Vj (k ), the line λ j := pq j meets X only in q j , since we
 MATH 249B NOTES: ALTERATIONS 31

ensured p is not on the line joining any two distinct points of W ∩ X.

The finite map

π p : X → P N −1 = {` 3 p}
therefore satisfies
π− 1


p λj = qj ∈
/ Xi
for all i 6= j. Hence, this p not only works as a common choice of
p j for all j but moreover π p ( X j ) changes as we vary j, so π p : X →
π p ( X ) is birational.
One can use an incidence correspondence argument similar to the
one employed for lines in the treatment of ( β), but now working
with Gr(d, N ), to obtain that the possibilities for such p exhaust at
least a dense open locus in PkN − X.
(3) This follows from the discussion in the previous part.

Now we return to our original setting of interest over k = k with Z (
X ⊂ PkN where X is a closed subvariety of dimension d < N with d ≥ 2
and Z is the support of a Cartier divisor (and so has pure dimension d − 1)
equipped with the reduced structure. We now apply Proposition 5.2 N − d
times to both X and Z to arrive at the following:

X Pd
(5.5)
π
Z π (Z)

where the top map is finite and generically étale and the bottom map is
birational.
Let Ω ⊂ Pd − π ( Z ) be a dense open such that π −1 (Ω) → Ω is étale.
Choose ξ ∈ Ω(k) generic with respect to π ( Z ) in the sense of our warm-up
example; that is, the finite projection

prξ : π ( Z ) → Pd−1 = {` 3 ξ }

is generically étale.
We aim to use X 0 := BlS ( X ) for S := π −1 (ξ ). Because blow-up commutes
with flat base change, we will be able to study this situation using our model
calculation for Pd . The idea is that X 0 is fibered over the space Pd−1 of lines
` through ξ, with `-fiber equal to π −1 (`).
32 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

5.1. Completing the proof of Lemma 4.1. We have composed N − d pro-

jections from sufficiently generic choices of points, and each choice of point
depended on the previous choices. The total collection of choices made cor-
responds to a collection of N − d independent points in P N , or more specif-
ically an ( N − d − 1)-plane in P N disjoint from X away from which we are
projecting. We’d like to understand how generic that linear subspace is (in
the corresponding Grassmannian). We’ll come back to this, but first let’s
recall the genericity conditions imposed on the choice of
ξ ∈ Pd − π ( Z )
above:
(1) ξ ∈ Ω, with Ω a dense open in Pd for which
π −1 ( Ω ) −
→Ω
π

is étale,
(2) ξ is “generic” with respect to π ( Z ), meaning that the natural map
Blξ (Pd ) → Pd−1 = {` 3 ξ }
is generically étale when restricted to
π ( Z ) ⊂ Pd − ξ ⊂ Blξ (Pd ).
Let S := π −1 (ξ ). By the above assumptions, S ⊂ X sm since π : X → Pd
has smooth target and is étale over ξ. Therefore, BlS ( X ) is normal if X is
because BlS ( X ) is a gluing of X − S and BlS ( X sm ) along X sm − S. Here we
are using that the blow-up Bl p (W ) of a smooth k-scheme W at a k-point p is
smooth as well. There are two easy ways to see this:
(1) The completed local rings on Bl p (W ) at points of the exceptional di-
visor are identified with ones on Bl0 (An ) and hence are all regular.
(2) By shrinking W around p (as we may do), we can arrange that there
is an étale map
f : W → An
p 7 → {0}
for which p is the entire fiber over 0. Then by compatibility of blow-
up with flat base change we have a fiber square
Bl p (W ) Bl0 (An )
(5.6)

W An
and Bl0 (An ) is smooth by inspection.
 MATH 249B NOTES: ALTERATIONS 33

In view of our explicit description of blow-up of projective space at a point

via an incidence relation, in the fiber square

X0 Blξ Pd
(5.7)

Pd .
π
X
we have
n o
X 0 = ( x, `) ∈ X × Pd−1 : x ∈ π −1 (`)

where Pd−1 is the space of lines through ξ. Also, π is flat and even étale over
Ω 3 ξ, so by compatibility of blow up with flat base change we have X 0 =
BlS ( X ). Thus, BlS ( X ) now has an incidence-relation description similar to
that of Blξ (Pd ); this will be very useful below.
Let’s now study the second projection f := pr2 , with pr2 as in

X0
pr2
(5.8)
pr1
X P d −1 .
Note that on
Z ⊂ X − S ⊂ BlS ( X ) = X 0
we have that f | Z is generically étale and finite onto Pd−1 because of our
choice of ξ (generic with respect to π ( Z )!).
We make the following observations, in sequence:
(1) For {`} ∈ Pd−1 we have
f −1 ({`}) = π −1 (`)
as schemes. Remember that π −1 (`) is 1-dimensional since π is a fi-
nite surjection. In fact, f −1 ({`}) has dimension ≥ 1 at all points (i.e.,
no isolated points) due to the fact that ` ⊂ Pd is locally defined by
d − 1 equations. Thus, π −1 (`) ⊂ X is pure dimension 1 everywhere.
(2) Each π −1 (`) is generically smooth. Indeed, π −1 (`) → ` carries each
irreducible component of π −1 (`) surjectively onto ` for dimension
reasons because it is proper and π is finite. Yet, this is a (scheme
theoretic) base change of the map π : X → Pd that is étale over an
open neighborhood Ω of ξ, so π −1 (`) → ` is étale over ` ∩ Ω. Thus,
π −1 (`) is smooth on the dense open π −1 (` ∩ Ω).
34 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

In fact, not only are the fibers of f generically smooth, but we claim that
f is a smooth map at k-points that are smooth in their fibers. This is a proto-
type for a property we will need in some other contexts later (bootstrapping
from a fibral property to a relative property without flatness assumptions),
and will follow from the next lemma that is useful but weaker variant of
[deJ, Lemma 2.8].
Lemma 5.10 (Weak [deJ, Lemma 2.8]). Let h : A → B be a local map between
complete local noetherian rings such that
(1) A is a domain of dimension δ
(2) dim B = δ + r
(3) B/m A B ' k A [[t1 , . . . , tr ]] as k A -algebras.
Then B ' A[[ T1 , . . . , Tr ]] as A-algebras.
The crucial feature is that we do not assume h to be flat (but we require A
to be a domain, as occurs when we form completed local rings on a smooth
scheme over a field such as Pd−1 ).

Proof. Pick T1 , . . . , Tr ∈ mB lifting t1 , . . . , tr ∈ B/m A B. We have a map of

A-algebras
φ
A[[ X1 , . . . , Xr ]] −
→B
Xi 7→ Ti .
Modulo m A this is just the map
φ
k A [[ X1 , . . . , Xr ]] −
→ k A [[t1 , . . . , tn ]],
Xi 7 → t i
that is an isomorphism, so by successive approximation and m A -adic com-
pleteness (and separatedness) we see that φ is surjective.
Now, we claim that in fact φ is an isomorphism. This holds because the
source of φ is a domain with dimension δ + r and B (which we don’t know yet
to be a domain or not!) has dimension δ + r, and the quotient of a noetherian
local domain modulo any nonzero ideal has strictly smaller dimension. 
Remark 5.11. More generally Lemma 5.10 shows that if f : X → Y is a map
of finite type between noetherian schemes and x is a rational point in the
smooth locus of its fiber X f ( x) such that the “dimension formula holds at x”
(l.e., dim OX,x = dim OY, f ( x) + dim OX f (x) ,x ) then x is in the smooth locus for
f provided that the completed local ring on Y at f ( x ) is a domain (such as
occurs when Y is regular at f ( x ), or normal and excellent).
 MATH 249B NOTES: ALTERATIONS 35

By applying the preceding lemma to the map of completions

OP∧d−1 , f ( x) → OX∧0 ,x0 .

for x 0 any k-point of X 0 smooth in its fiber X 0f ( x0 ) (as is applicable because

completed local rings of the smooth Pd−1 at k-points are formal power series
rings over k in d − 1 parameters), we conclude that f is smooth at any k-
point x 0 that is a smooth point in its fiber X 0f ( x0 ) .
To complete the proof of Lemma 4.1, it remains to prove the next result.
Lemma 5.12. If X is normal then for a suitable choice of ξ:
(1) there exists a dense open U ⊂ Pd−1 so that f −1 (U ) → U is smooth and
(2) f : X 0 → Pd−1 is its own Stein factorization (so it has geometrically con-
nected fibers, by [EGA, III, 4.3.3]).
Let us first prove (1). The relative smooth locus sm( X 0 /Pd−1 ) ⊂ X 0 is
open, so by properness of f : X 0 → Pd−1 it is enough for this open set to
contain one fiber Xy0 . By Lemma 5.10, this is equivalent to Xy0 being smooth,
so we just need to find some ` ⊂ Pd through ξ so that π −1 (`) ⊂ X is smooth.
Recall that π is constructed by projecting from N − d + 1 independent
points in P N , and ξ corresponds to another point. Overall, this corresponds
to an ( N − d + 1)-plane Λ ⊂ PkN .
Since X is normal, we know X − X sm ⊂ X has dimension at most d − 2.
But N − d + 1 = N − (d − 1). Therefore, general such linear spaces Λ ⊂ P N
of codimension d − 1 meet X entirely in X sm . Thus, as long as the possible
planes Λ ⊂ P N encoding all of our choices from iterated Bertini theorems
really sweep out at least a dense open set in the appropriate Grassmannian
we are done. That such genericity does hold for Λ is addressed in a handout
on “iterated genericity”.

5.2. A Stein factorization. It remains (for the proof of Lemma 4.1) to estab-
lish part (2) in Lemma 5.12. Recall that after passing to BlS ( X ) (renamed as
X and still normal) we were in the following situation:
(1) all Xy are pure of dimension 1
(2) there exists a dense open U ⊂ Y so that f −1 (U ) → U is smooth
(3) sm( X/Y ) ∩ Xy ⊂ Xy is dense for all geometric points y ∈ Y.
Recall that by design of f , such Xy are controlled by Bertini methods, at least
over a dense open in Y. These Xy are X ∩ W for W ⊂ PkN sufficiently general
linear spaces of codimension d − 1. So, W ∩ X = W ∩ X sm is irreducible and
smooth of dimension 1 (since Bertini applies to the smooth quasi-projective
X sm of dimension d). The upshot is that there exist many y0 ∈ U (k ) so that
36 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Xy0 is geometrically irreducible and smooth. We will use the existence of

this geometrically connected fiber y0 to settle our Stein factorization prob-
lem:
Lemma 5.13. The map f : X → Y is its own Stein factorization (i.e., OY '
f ∗ OX ), and in particular all geometric fibers of f are connected.
Proof. We have OY → f ∗ OX is an injection as f is a surjection of varieties.
Note that Y is normal, so this inclusion of coherent sheaves of domains is
an equality if and only if it is an equality over the generic point of Y. In fact,
Letting T be the localization of Y at y0 , we’ll show
OY,y0 → ( f ∗ OX )y0 = H 0 ( XT , OXT )

is an isomorphism. This is a special case of the next result. 

Proposition 5.14. Let ( R, m, k ) be a local noetherian ring and Z a non-empty
proper flat R-scheme so that the special fiber Z0 is geometrically reduced and ge-
ometrically connected over the residue field k. Then, the natural map φ : R →
H 0 ( Z, OZ ) is an isomorphism.
Proof. Surjectivity of R → H 0 ( Z, OZ ) will use cohomology and base change,
and injectivity will use flatness. Note that f is both open and closed and Z
is connected (and non-empty), so f is surjective.
First let’s show φ is injective. Pick z0 ∈ Z0 . By flatness of f , we have a
map R → OZ,z0 which is a flat local map. Hence it is faithfully flat, and thus
injective. But the map factors as
φ
→ H 0 (OZ ) → OZ,z0
R−
so ker φ = 0.
We’ll now verify surjectivity. If Y is proper over a field F and geometri-
cally reduced and geometrically connected over F, then F ' Γ(Y, OY ) (as we
see via scalar extension to the case F is algebraically closed). This implies
k ' H 0 ( Z0 , OZ0 ), so via the factorization

H 0 ( Z, OZ )/m H 0 ( Z0 , OZ0 )
α

(5.9) '

k
involving the base change morphism α we see that α is surjective. The the-
orem on cohomology and base change applies to OZ since Z is R-flat and
proper, so α is an isomorphism. This implies that the module-finite ring
 MATH 249B NOTES: ALTERATIONS 37

map
φ : R → H 0 ( Z, OZ )
is surjective modulo m and so is surjective (by Nakayama’s Lemma). 

6. T HREE - POINT LEMMA

Remark 6.1 (Summary of the situation thus far). Before continuing, we re-
view the current situation. We have established all of the following proper-
ties except (vi )(e) and (iv)( f ) below:
(i)-(iv) X is a projective variety over k = k of dimension d ≥ 2 with Z ⊂ X
the support of a non-empty Cartier divisor.
(v) X is normal
(vi) there exists a surjective map X → Y where Y is a projective variety
of dimension d − 1 satisfying the following properties:
(a) All fibers of f are geometrically connected of pure dimension 1,
(b) the smooth locus sm( X/Y ) for f is fiberwise dense over Y,
(c) f is smooth over a dense open U ⊂ Y,
(d) Z → Y is finite (hence surjective) and generically étale,
(e) sm( X/Y ) ∩ Z meets each geometric fiber Xy for y ∈ Y a geomet-
ric point in at least 3 points per irreducible component of Xy .
(f) We have
Z = ∪ri=1 σi (Y )
for sections σi : Y → X so that {σi (ηY )} are pairwise distinct
(which is preserved under further alterations of Y).
The motivation for the condition with 3 points in (vi)(e), which we will
arrange below by suitable enlargement of Z, is that irreducible components
isomorphic to P1 in stable curves need 3 distinguished points.
Lemma 6.2 (The three-point lemma, [deJ, Lemma 4.13]). Suppose we have
f : X → Y is a map of projective varieties over k = k with d = dim X ≥ 2 and
assume that (vi )( a) and (vi )(b) from Remark 6.1 hold. Then there exists a Cartier
divisor D ⊂ X so that
(1) D → Y is finite and generically étale and
(2) for all geometric points y ∈ Y, sm( X/Y ) ∩ D meets each irreducible com-
ponent of Xy in at least 3 points.
Remark 6.3. Once we complete the proof of Lemma 6.2, we can replace Z
with ( Z ∪ D )red , which is still the support of a Cartier divisor. We also gain
condition (vi )(e) of Remark 6.1.
38 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Proof. We’ll build D as X ∩ H for a generic hyperplane H in a suitable em-

bedding ι : X → P M for M large.
Pick a very ample L on X. Consider

ιL : X ,→ P = P (Γ( X, L )) ,
X 7→ [evx : Γ( X, L ) → L ( x ) := Lx /mx Lx ] .

This is a map to the space of hyperplanes in Γ( X, L ). Note that hyperplanes

H ⊂ P correspond to lines in Γ( X, L ), which correspond to hyperplanes in
Γ( X, L )∨ , the space of which is P(Γ( X, L )∨ ) =: P∨ .
We seek H such that H ∩ Xy is 0-dimensional for all geometric points
y ∈ Y (and we will also seek additional properties). That is, we wish to
avoid the situation where H ∩ Xy is 1-dimensional for some y. To do so, we
consider the incidence correspondence:

T := ( H, y) ∈ P∨ × Y : dim H ∩ Xy = 1 ⊂ P∨ × Y.


This is closed because it is the jumping locus for fiber-dimension of a mor-

phism: we have a map

{( H, x ) ∈ P∨ × X : x ∈ H }
(6.1)

{( H, y) ∈ P∨ × Y }


for which the fiber over a point ( H0 , y0 ) is equal to { H0 } × H0 ∩ Xy0 . Thus,
all fibers of this map have dimension 1 or 0, and so by definition T is the
locus where the fiber dimension has at least 1. This proves that T is closed,
by upper-semicontinuity of fiber dimension.
Consider the projections

T
pr2
(6.2)
pr1
P∨ Y

To find H making H ∩ Xy be 0-dimensional for all y ∈ T, we only need to

check pr1 ( T ) ⊂ P∨ is not all of P∨ , since we know the image is closed by
properness. We will study the fibers of h := pr2 to show dim T < dim P∨ ,
provided we replace L by a large tensor power.
 MATH 249B NOTES: ALTERATIONS 39

For a geometric point y0 ∈ Y, label the irreducible components of Xy0 as

C1 , . . . , Cr , so
h−1 (y0 ) = H ∈ P∨ : H contains some Cj


= ∪1≤ j≤r H ⊃ Span ιL (Cj )
Define
ΛL



Cj : = Span ιL C j .

Now, observe that h−1 (y0 ) is a union of linear subspaces of P∨ of codimen-

sion 1 + ΛL
Cj , so

dim T ≤ dim Y + sup dim h−1 (y)

y ∈Y ( k )
  
= dim Y + sup dim P∨ − 1 + dim ΛL
C
y∈Y (k),C ⊂ Xy irreducible components
 
= dim Y + dim P∨ − inf 1 + dim ΛL
C .
y∈Y (k),C ⊂ Xy

We just need that all dim ΛL

C are uniformly bigger than dim Y.
L
The linear span ΛC ⊂ P (Λ ( X, L )) is
 
P Λ ( X, L ) /WCL

for
WCL := ker (Λ ( X, L ) → Γ (C, L |C )) .
Hence,
1 + dim ΛL L
C = dim Γ ( X, L ) /WC
= dim (im (Γ ( X, L ) → Γ (C, L |C ))) .
We want the dimension of these images to be uniformly large. Even better,
the dimensions of these images are uniformly big for all irreducible closed
curves C ⊂ X:
Lemma 6.4. With the notation as above, for n ≥ 1, we have
im Γ( X, L ⊗n ) → Γ(C, L ⊗n |C )

has dimension at least n + 1.

Proof. Since L is very ample, we can find two hyperplane slices of C in P
that are disjoint and non-empty.
40 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

This corresponds to two global sections

s, s0 ∈ Γ( X, L ) − {0}
such that
s|C , s0 |C ∈ Γ (C, L |C )
have disjoint non-empty zero loci. Therefore, they are linearly independent
over k, so
s0 |C
f := ∈ k(C ) − k.
s|C
Observe that f is transcendental over k, as k is algebraically closed. There-
fore, the monomials s0 j |C sn− j |C ∈ Γ(C, L ⊗n |C ) are linearly independent
over k, since a nontrivial dependence relation would give an algebraic de-
pendence on f over k. So, s0 j sn− j |C form the desired n + 1 independent
sections in the image. 
Now, rename L ⊗n as L for such big enough n, taken to be at least 3.
Thus, for generic H ∈ P∨ , D := H ∩ X meets each Xy in a finite set.
Remark 6.5. For any irreducible component C of any fiber Xy , we know
D meets C in at least 3 points with multiplicity by Bezout’s theorem, since
deg L |C⊗n ≥ n in the earlier situation. Thus, if D ∩ Xy0 ⊂ sm( X/Y ) and is
in the étale locus for D over Y then then we are in good shape.
Note that D → Y is finite for D := H ∩ X with H as above, as D → Y is
proper and quasi-finite. Further, D has pure dimension d − 1 due to being
Cartier in X. We want to arrange that D is generically étale over Y for most
H ∈ P∨ :
Lemma 6.6. For generic H ∈ P∨ , the map D := H ∩ X → Y is generically étale.
Proof. Choose y1 ∈ Y (k ) and consider Xy1 ⊂ P. Let H1 be a hyperplane so
that H ∩ X → Y is quasi-finite (and hence finite) and Dy1 := H1 ∩ Xy1 ⊂
H1 ∩ sm( X/Y ) is étale and meets each irreducible component of Xy1 in at
least three points. The condition that H ∩ X be Y-finite holds for generic H
due to our earlier passage to a large tensor power of an initial choice of very
ample L on X, and the other conditions (involving just Xy1 and sm( X/Y )y1 )
happen for a generic choice of H1 due to Bertini’s theorem (since by design
sm( X/Y ) meets the fiber Xy1 in a dense open subset). Set D1 := H1 ∩ X.
It remains only to show (as we will do using a variant of earlier calcula-
tions upgrading smoothness of a point in a fiber to smoothness of a mor-
phism at that point, under some hypotheses on the base) that D → Y is
étale at points that are étale in its fibers. Once we show this, by Y-finiteness
 MATH 249B NOTES: ALTERATIONS 41

of D we would obtain an open U1 ⊂ Y containing y1 so that ( D1 )U1 → U1

is étale. In particular, for all geometric points y ∈ U we would have that
( D1 )y is étale and thus (for degree reasons) has at least 3 points per irre-
ducible component of Xy .
We will next exploit that
Dy1 ⊂ sm( X/Y ).
Let h ∈ mx ⊂ OX,x be the local equation for H ∩ X near x, so the structure
map D1 → Y induces
OY,y1 → OX,x /(h) = OD1 ,x .
Proving étaleness at x amounts to the natural map of completions
O
[Y,y1 → O D1 ,x
[
being an isomorphism. By design x ∈ sm( X/Y ), so
O
[ D,x = OX,x / ( h ) = OY,y1 [[ t ]] / ( h )
d [
and hence it suffices to show h ∈ k× · t + m
b y1 [[t]] (as then we would get
the desired isomorphism between completions for D → Y at x and y1 via
successive approximation calculations). This task for h only involves its
image in
Od my1 [[t]] = Od
X,x /d X,x /my1 · OX,x
d

=O
\ Xy1 ,x
= k[[t]].
Thus, we just need that the image of h in here is tu for a unit u. But
k [[t]]/(h) = O\
( D1 )y ,x 1

and
O( D1 )y ,x =k
1

since ( D1 )y1 is k-étale. Therefore, k [[t]]/(h) = k, so h has the desired form

modulo m d y1 .

By Lemma 6.6 for our arbitrary initial choice of point y1 ∈ Y (k ) we have
found H1 and an open neighborhood U1 of y1 so that D1 := H1 ∩ X is Y-
finite and étale over U1 . But away from U1 we have not controlled the fiber
of D1 → Y at y ∈ (Y − U1 )(k ), and so for such y the fiber ( D1 )y might not
have at least 3 points on each irreducible component C of Xy (even though
the scheme ( D1 )y ∩ C has large degree; perhaps it is a very fat point).
42 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

If the open set U1 from Lemma 6.6 is equal to Y then we are done. If not,
choose y2 ∈ (Y − U1 ) (k ), and run the same procedure to find D2 := H2 ∩ X
that is Y-finite and étale over some open neighborhood U2 of y2 in Y. We
can also arrange (by genericity of the choice of H2 ) that H2 does not contain
any irreducible component of D1 , so the effective Cartier divisor D1 + D2
(i.e., the closed subscheme of X corresponding to ID1 · ID2 ) agrees with
( D1 ∪ D2 )red on a dense open. (Note that each Di is generically reduced due
to generic étaleness over the reduced scheme Y.) Thus D1 + D2 → Y is still
generically étale. If U1 ∪ U2 6= Y, we continue in the same way. we conclude
using noetherian induction. 
By replacing Z with ( Z ∪ D )red , as we may certainly do, we have arranged
that property (iv)(e) From Remark 6.1 holds.

7. PASSAGE TO A UNION OF SECTIONS

We next arrange that condition Remark 6.1(iv)(f) holds, by finally apply-
ing a (mild) generically étale alteration to Y (and adjusting X and Z in a
suitable manner to be described).
Remark 7.1. If ψ : Y 0 → Y is a generically étale alteration, then the data
X 0 := X ×Y Y 0 red

Z 0 := Z ×Y Y 0 red

Y0
satisfies all of our running conditions except possibly that X 0 may not be
normal. The only aspect of this that requires some thought is to justify that
X 0 is irreducible, but this is done in a handout (the point being that flatness
of the fiberwise-dense sm( X/Y ) allows us to detect irreducible components
of X ×Y Y 0 by looking at the generic fiber over Y 0 , which in turn is a base
change of the generic fiber of X → Y that is smooth and geometrically con-
nected (so geometrically irreducible!).
e→
As a simple illustration, we can take ψ to be the finite normalization Y
Y to arrange that Y is normal (at the possible cost of losing normality of X).
Lemma 7.2. By passing to a further alteration of Y, we may arrange that (iv)(f)
of Remark 6.1 holds.
Proof. We want to arrange for a generically étale alteration ψ : Y 0 → Y that
Z 0 = ∪1≤ j≤r σj (Y 0 )
for distinct sections σj : Y 0 → X 0 .
 MATH 249B NOTES: ALTERATIONS 43

Let η be the generic point of Y, so Zη is a nonempty finite
étale η-scheme.

We can choose η 0 → η a big finite Galois point such that Zη ×η η 0 is a finite
disjoint union of copies of η 0 .
Let ψ : Y 0 → Y be the normalization of Y in the finite Galois extension
K (η 0 ) of K (η ) = K (Y ). Since Yη0 = η 0 as schemes, we can use Remark 7.1 to
arrange that
N
Zη = ä η.
i =1

Let Z1 , . . . , Zr be the irreducible components of the pure-dimensional Z

with reduced structure. Thus, the finite maps Zi → Y are surjective for
dimension reasons, so the generic points of the Zi ’s are precisely the points
of Zη . Therefore each finite map Zi → Y is birational. But Y is normal, so
each Zi → Y is an isomorphism. Thus, the inverse maps Y ' Zi → X define
sections σi satisfying
[
Z= σi (Y )
i

as subsets in X. 
Remark 7.3. Above, we crucially used (generically étale) alterations involv-
ing function field extensions in order to split Z into a union of sections.
Later we will require such alterations for less explicit (and deeper) geomet-
ric reasons.
To go further, we need to digress and review some basic facts concerning
relative stable marked curves and their associated moduli spaces/stacks
(which admits a “smooth scheme chart” built via Hilbert schemes, so we
also need to review some facts about Hilbert schemes).

8. S TABLE CURVES
8.1. Definitions and examples.
Definition 8.1. Fix n, g ≥ 0 so that 2g − 2 + n > 0. (That is, either g ≥ 2 or
g = 1 with n ≥ 1, or g = 0 with n ≥ 3.) An n-pointed stable genus-g curve
over a scheme S is a proper flat finitely presented map f : C → S equipped
with sections σ1 , . . . , σn ∈ C (S) so that:
(1) all geometric fibers Cs are connected semistable and the (arithmetic)
genus h1 (Cs , O ) is equal to g,
(2) the sections σi are pairwise disjoint and supported in the Zariski-
open S-smooth locus C sm .
44 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

(3) for any geometric point s of S, any irreducible component C of Cs

isomorphic to P1 contains at least three special points: some σi (s) or
a point where C meets another irreducible component of Cs .
Exercise 8.2. Show that C obtained from P1 via n “semistable self-gluing”
are semistable irreducible curves of genus n over an algebraically closed
field. We can take “semistable self-gluing” to mean semistable with nor-
malization P1 , and the content is to prove that C can be canonically recon-
1
structed o from its normalization C ' P equipped with the data of the pairs
e
n
xei , xe0 over each singularity of C (and that any such finite collection of
i
pairs gives rise to a “self-gluing” P1 → C that is initial for maps out of P1
having the same restriction to each k-point in such a pair).
The idea is to consider the exact sequence

(8.1) 0 OC ν∗ OCe Q 0

for ν the normalization map. Here, the coherent sheaf Q is a skyscraper

supported at Csing and the semistability implies
Q ' ⊕ x∈Csing κ ( x )
(which we leave as an exercise, using the description of ObC,x and that nor-
malization commutes with completion in these situations, as can be proved
by elementary means without recourse to hard theorems involving excel-
lence). Passing to the cohomology sequence, we get

(8.2) 0 k k h 1 ( OC ) h1 (OCe ) 0,
so
h1 (OC ) = #Csing .
Remark 8.3. In general, if C is an irreducible semistable proper curve over
an algebraically closed field then the same calculation with the normaliza-
tion sequence yields
h1 (OC ) = h1 (OCe ) + #Csing .

Example 8.4. Consider the union P1 ∪ C where P1 has two sections and
meets a nodal C1 at a point. One can use the normalization sequence to
show this has (arithmetic) genus
g = 1 + g(C
f1 ).

Exercise 8.5. Let k = k.

 MATH 249B NOTES: ALTERATIONS 45

(1) Show that a connected proper semistable curve over k with genus 0
is precisely a connected finite tree of copies of P1 . The point is that
the homology you get from the exact sequences for the normalization
comes from the homology of the dual graph (so arithmetic genus 0
amounts to the dual graph having no loops; i.e. it is a tree).
(2) The only stable curve over k with g = 0, n = 3, is P1 with the ordered
triple of markings (0, 1, ∞) up to unique isomorphism.
(3) A stable curve of genus 1 with n = 1 is a single elliptic curve with
a marked point or a nodal cubic (i.e., P1 with one semistable self-
gluing) having a marked point.
If one allows n > 1, there are many more possibilities. We will just
describe the underlying semistable curve, since the stable curves are
obtained by adding marked points. Any such curve can be written as
C ∪ D1 ∪ · · · ∪ Dn , with C, Di defined as follows: Di are all semistable
genus 0 curves as described above (a connected finite tree of copies of
P1 ) so that each Di meets C at points p1 , . . . , pn , and Di does not meet
D j , and C is a smooth genus-1 curve or a so-called Néron polygon: a
curve whose normalization is a collection of copies of P1 so that the
dual graph is a polygon (a nodal P1 , or a banana – two copies of P1
glued at two points, or a triangle – three copies of P1 each glued to
the other two at a single point – and so on).
(4) For a stable curve with irreducible component C satisfying h1 (OC ) =
1, there must be at least one special point on C.
Remark 8.6. The two key references on the topic of stable curves are the
following: the case n = 0 with any g ≥ 2 is [DM, §1], and the case of
arbitrary n is addressed in [Knu, pp. 161-199]. Both involve much use of
coherent duality over rings (not just over fields).
8.2. Moduli for stable curves. A key input for the construction of a “uni-
versal stable marked curve” (for a given g and n) Will require applying the
following openness result to a universal object over a variant of a Hilbert
scheme:
Lemma 8.7. Let f : X → S be a proper flat finitely presented map of schemes.
Then
U := {s ∈ S : Xs is a connected semistable curve of arithmetic genus g.}
is open in S.
Proof. The openness results here are special cases of general openness theo-
rems on the base for properties of geometric fibers of flat finitely presented
proper scheme maps (though in the present case one may be able to give
46 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

more direct arguments; we just wish to convey that there are general prin-
ciples at work here, applicable way beyond the setting of relative curves).
Openness of the set of pure 1-dimensional fibers is [EGA, IV3 , 12.2.2(ii)].
Openness for the condition for geometrically reduced and geometrically
connected fibers is [EGA, IV3 , 12.2.4(vi)]. The idea of these proofs is usually
by first attacking problems locally on the source (proving openness results
in X without any properness hypotheses for flat finitely presented maps, by
bootstrapping from constructibility results via generization considerations),
and then in the proper setting the bad locus in the base is the image of the
complementary closed bad locus in X.
Once we have passed to the open locus cut out by the preceding condi-
tions, we have the equality h1 (O ) = 1 − χ(O ) on geometric fibers and can
invoke the constancy Zariski-locally on the base for the fibral Euler char-
acteristic s 7→ χ( Xs , Fs ) for any proper finitely presented f : X → S and
S-flat finitely presented quasi-coherent OX -module F (see [Mum, Ch. II, §5,
Cor. 1(b)] for noetherian S; one reduces to this case by some nontrivial re-
sults in EGA concerning spreading-out for flatness and properness).
Finally, what about openness for the semistability condition? We just
need to prove openness of the semistable locus on X (implying openness
on S via properness considerations, due to the closed non-semistable in X
having closed image in S), and this is shown using Artin approximation in
[FK, Ch. III, Prop. 2.7] as a special case of a general étale-local description
for ordinary double-points in geometric fibers (in any odd relative dimen-
sion). 
Example 8.8 (Why stable curves can’t be classified by a universal object
over a moduli scheme). The obstruction to existence of a universal stable
marked curve (for a given g and n) is, roughly speaking, that there exist
non-constant families with pairwise isomorphic geometric fibers. (This is
a more precise version of the informal obstruction that objects can admit
non-trivial automorphisms.)
Consider g = n = 1 and a field k with char k 6= 2. Let E be an elliptic
curve over k and consider the map
f : ( E × Gm )/h−1i → Gm
( x, t) 7→ t2 .
where the action of −1 is given by
(−1) ( x, t) = (− x, −t) .
(Let’s briefly discuss the existence of this quotient by the action of Z/(2).
That action is free on geometric points since negation on Gm over k is fixed-
point free, and for any quasi-projective scheme V over a field equipped with
 MATH 249B NOTES: ALTERATIONS 47

an action by a finite group Γ that is free on geometric points over any alge-
braically closed field, one can cover V by Γ-stable affine opens and thereby
form a quasi-projective V/Γ for which V → V/Γ is a finite étale Γ-torsor;
see [Mum, ?].)
We take the marking σ to be the image of the zero section t 7→ (0, t). All
k-fibers of f are isomorphic to E. If there were actually a universal 1-point
stable genus-1 curve over k then we could recover f as the left side of a fiber
square

X (E univ , σuniv )
(8.3)
q
S M1,1

and hence the classifying map q : S = Gm → M1,1 would send the entirety
of S(k ) through a single k-point (corresponding to ( E, 0)). But then q would
have to factor through that k-point, forcing X → S to be a constant stable
marked curve. The following is an instructive exercise:
Exercise 8.9. Complete the above example by verifying that X is not S-
isomorphic to E × S.
8.3. Equivalent definitions of stable curves. We’d like to explain why we
require three points to be special on each P1 , and also the reason that we
require 2g − 2 + n to be positive. Keep in mind that we are just surveying
some key highlights in the basic theory of stable curves and their moduli so
that we can use these ideas to progress further into deJong’s proof; hence,
we will not have time to develop this material in full and so will provide
literature references for various details.
Lemma 8.10. Let X be a connected semistable curve over an algebraically closed
field k, let g := h1 (O ), and let σ1 , . . . , σn ∈ X sm (k ) be disjoint sections. The
following conditions are equivalent.
(1) ( X; σ1 , . . . , σn )k[ε] (viewed as a curve over Spec k [ε] with ε2 = 0) has no
nontrivial k [ε]-automorphism reducing to the identity modulo ε.
(2) The locally finite type k-scheme (actually always finite type, due to working
with curves)
Aut(X;σ1 ,...,σn )/k
is étale.
(2’) The group Aut ( X; σ1 , . . . , σn ) is finite.
(3) The following two conditions hold:
48 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

(i) every irreducible component isomorphic to P1 has at least 3 special

points,
(ii) every irreducible component with arithmetic genus at least 1 has at
least 1 special point.
Remark 8.11. Note that when 2g − 2 + n > 0, condition (ii) in (3) automati-
cally holds (so one never sees (ii) mentioned in literature on stable pointed
curves); this is not the reason that 2g − 2 + n is assumed to be positive. Also,
the equivalence of (1) and (2) is due to each expressing the vanishing of the
tangent space at the identity on the Aut-scheme (i.e., a group scheme lo-
cally of finite type over a field is étale if and only if its tangent space at the
identity point vanishes).
The fact that the Aut-scheme in (2) is really finite type and not just locally
finite type can be proved in several ways, but perhaps the most concep-
tually satisfying way is to exploit the fact that fixing a Hilbert polynomial
carves out a finite-type clopen subscheme of a Hilbert scheme in the projec-
tive setting. We will come back to this point later in the proof of Lemma 9.6
(which does not required the present considerations, so there is no circular-
ity involved).
Proof. An automorphism as in (1) is a map
φ : OX [ε] → OX [ε]
f + gε 7→ f + ( g + D ( f )) ε
for a k-linear derivation
D : OX → OX .
So we can view D instead as an OX -linear map D : Ω1X/k → OX . Hence, on
X sm we can view D as a vector field ~v, since Ω1X/k | Xsm = Ω1Xsm /k is a vector
bundle with local frame
 
∂
∂t
for étale t : U → A1k for Zariski-open U ⊂ X sm .
b denotes the k[ε]-automorphism of Xk[ε] corresponding to φ then the
If φ
condition that φ b((σj )k[ε] ) = (σj )k[ε] as k [ε]-points is exactly the condition that
~v(σj ) = 0. That is, the vector field on X sm associated to D has zeros at these
k-points of X sm . Under this constraint, we want to show that necessarily
D = 0 when (3) holds (so (3) implies (1)).
The case n = 0, g ≥ 2 is treated [DM, §14] (where (3)(ii) is never men-
tioned since it holds automatically, and there is no appearance of marked
points). The method amounts to some calculations in coherent duality on
 MATH 249B NOTES: ALTERATIONS 49

semistable curves, generalizing the classical fact that a smooth connected

proper curve over k with genus at least 2 has no nonzero global vector fields.
The same method generalizes without much difficulty to show (3) implies
(1) in general (with the marked points, and incorporating (3)(ii)).
The other equivalences are easier and are omitted. 
Remark 8.12. Suppose we are given a projective scheme Y over an alge-
braically closed field and form its Aut-scheme, which is made out of a Hom-
scheme that in turn is built from the Hilbert scheme of Y × Y (by viewing
endomorphisms through their graphs). The Hilbert scheme is built as a
countable union of finite type k-schemes, so the same holds for this Aut-
scheme. Hence, the identity component of AutY/k is of finite type (though
one can also show abstractly that any locally finite type k-group scheme
which is connected is in fact finite type and irreducible) and more impor-
tantly the component group ΓY of AutY/k is countable. But might ΓY have
some reasonable finiteness properties, such as being finitely generated or
finitely presented?
Note that if E is a non-CM elliptic curve then the Aut-scheme of E × E
(without respecting (0, 0)!) has identity component E × E via translations
and component group GL2 (Z) that is finitely presented (as are all arithmetic
groups). In general what can be said about the structure of ΓY , say if Y is a
variety (i.e., reduced and irreducible)? Even the case of smooth projective
surfaces Y over C is very mysterious!
There is one natural idea that comes quickly to mind for attacking this but
0
almost as quickly dies: the Néron–Severi group NS(Y ) = PicY/k /PicY/k is
a finitely generated abelian group (Theorem of the Base, due to Néron and
Lang), and ΓY naturally acts on this, so it provides a natural representation

ΓY → GL(NS(Y )/torsion) = GL N (Z).

If the image were an arithmetic subgroup, hence finitely presented, that

might give us a handle on the problem. Alas, Borcherds gave examples of
smooth projective surfaces over C for which this image is not arithmetic, so
those hopes are dashed.
I once came upon this puzzle as a side issue during some work on finite-
ness questions over global function fields, and lacking any idea on how to
proceed with it I decided to ask around. I asked a lot of people, including
Mumford, Oort, de Jong, Gabber, and so on. Nobody had any idea. What
a stumper! I posed this baffler on Math Overflow to raise awareness about
such a natural question. The happy ending is that John Lesieutre saw that
MO question and recently came up with smooth geometrically connected
counterexamples with trivial geometric étale fundamental group over any
50 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

field of characteristic 0 (a bit beyond dimension 2 for now, but he has some
candidates in dimension 2 as well).
8.4. Dualizing sheaves. We’d next like to give a criterion for ampleness of
a line bundle attached to a marked semistable curve. To do so, we’ll first
need a brief digression into coherent duality.
Consider
( X; σ1 , . . . , σn )
As in the setup for Lemma 8.10. the k-scheme X is Gorenstein, as the Goren-
stein property of local Noetherian rings can be checked on the completion
and we know the completions at closed points of semistable curves over
k = k are either k[[t]] or k[[u, v]]/(uv). Consequently, the relative dualizing
complex for X/k is an invertible sheaf.
Remark 8.13. The Cohen-Macaulay property of a finite-type scheme over
a field (or more generally, but we omit that here) is equivalent to the con-
dition that its dualizing complex in the sense of Grothendieck’s theory of
coherent duality (which is characterized by features not requiring proper-
ness!) is simply a sheaf (thereby called the dualizing sheaf), denoted ωX/k .
The Gorenstein property (stronger than the CM condition) is equivalent to
the dualizing sheaf being invertible.
The formation of the dualizing sheaf ωX/k commutes with étale pullback
on X (this is a general property of the dualizing complex), so one can de-
scribe ωX/k using étale models. There is a useful concrete description in the
case of semistable curves via 1-forms with controlled poles and residues on
the normalization, but since we are not getting into gritty details of calcula-
tions with dualizing sheaves here (though they are lurking in the references
to [DM] below!) we will not discuss these explicit descriptions here.
Let
e X/k := ωX/k ∑ σi = ωX/k ⊗ Iσ−1 1 ⊗ · · · ⊗ Iσ−n 1 ,


(8.4) ω
where Iσ is the ideal sheaf of σ for σ ∈ X sm (k ).
For a line bundle L on X we have a notion of degree defined by the
condition
χ(L ⊗m ) = deg L · m + χ(OX ).
See [BLR, §9.1] for an elegant discussion of this notion, its relation to Cartier
and Weil divisors and normalization, and its properties are rather arbitrary
proper curves over fields (including additivity in L ). In particular, the de-
gree of the inverse ideal sheaf of a rational point in the smooth locus is equal
to 1, as in the familiar setting of smooth connected projective k-curves. In
 MATH 249B NOTES: ALTERATIONS 51

particular, if L is ample then for very large m there are many global sec-
tions of L ⊗m , so deg L > 0 in such cases. (The converse fails when X is
reducible, as L might have degree-0 restriction to some irreducible com-
ponents. Since ampleness can be checked on the irreducible components
equipped with their reduced structure, ampleness on pure curves is equiv-
alent to positivity of the degree of the restriction to each irreducible compo-
nent equipped with its reduced structure.)
e X/k = 2g − 2 + n, with ω
Lemma 8.14. We have deg ω e X/k as defined in (8.4).
Proof. Coherent duality implies
deg ωX/k = 2g − 2
For g := h1 (OX ) = 1 − χ(OX ). Thus, deg ω
e X/k = 2g − 2 + n. 
Lemma 8.15. Condition (3) in Lemma 8.10 is equivalent to ω
e X/k being ample,
(which in turn implies 2g − 2 + n > 0).
Proof. For n = 0, g ≥ 2 the implication “ ⇒” is [DM, Theorem 1.2], proved
via computations with coherent duality; the converse in such cases is easier.
This method adapts to any ( g, n) with g ≥ 2 or when g = 0, 1 with with
n ≥ 4. For g ≤ 1, n ≤ 3, see [Knu, Corollary 1.10]. 
Remark 8.16. For f : X → S proper of finite presentation, an invertible
sheaf L on X is ample over all affine opens in S (or equivalently on each
member of an open affine cover of S) if and only if Ls on Xs is ample over
every s ∈ S. The amazing implication here (in view of the absence of any
flatness hypotheses, thereby preventing any use of base change theorems
for coherent cohomology) is “⇐”, which is [EGA, IV3 , 9.6.4]. This condition
on L is called S-ampleness.
9. M ODULI OF STABLE CURVES
To work in the relative setting, we need a generalization of ωX/k from
the earlier considerations with semistable curves over fields. Here is “the
bitter pill.” For finitely presented maps with Cohen-Macaulay fibers there
is a finitely presented quasi-coherent dualizing sheaf ωX/S compatible with
any base change on S and with étale pullback on X and having some nice
properties with respect to coherent duality (that we do not have time to flesh
out here). Furthermore, ωX/S is invertible if all fibers Xs are Gorenstein.
Consider a proper finitely presented map
X
(9.1) f

S
52 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

whose geometric fibers are connected semistable curves, and suppose we

are also given pairwise disjoint sections σ1 , . . . , σn ∈ sm( X/S).
Proposition 9.1. In the above situation,
Sst := {s ∈ S : ( Xs ; {σi (s)}) is stable }
is Zariski-open in S.
Proof. Consider the invertible sheaf L = ωX/S (∑ σi ) whose formation is
compatible with base change. (Here we are twisting by the inverse ideal
sheaves of the sections σi , leaving it as an exercise to check that the ideal
sheaf of a section through the relative smooth locus of a relative flat finitely
presented curve is always invertible and that the formation of this ideal sheaf
commutes with any base change.) Our preceding discussion in the context
of semistable curves over algebraically closed fields gives the crucial equal-
ity
Sst = {s ∈ S : Ls is ample on Xs } .
But for any proper finitely presented map Y → S and any invertible L
whatsoever on Y, it is a general fact that
{s ∈ S : Ls is ample on Ys }
is Zariski-open: see [EGA, IV3 9.9.6]. 
Generalizing classical arguments with cohomology and base change by
which a smooth proper relative curve with geometrically connected fibers
of genus g ≥ 2 admits a closed immersion (“tri-canonical embedding”) into
a projective-space bundle of dimension 5g − 6, the preceding results on the
relative setting lead to:
Theorem 9.2. Let ( X → S, {σi }) be an n-pointed stable genus-g curve over a
⊗m
scheme S. For any integer m ≥ 4, the sheaf f ∗ (LX/S ) is a vector bundle whose for-
⊗m
mation commutes with any base change on S, and the natural map f ∗ f ∗ (LX/S )→
LX/S is surjective. Moreover, the resulting natural map
⊗m
X → P( f ∗ (LX/S ))
⊗m
into a projective-space bundle is a closed immersion, the rank of f ∗ (LX/S ) is equal
to
N (n, g, m) = m(2g − 2 + n) + 1 − g.
There is a universal such structure equipped with the data of an isomorphism
⊗4
P( f ∗ (LX/S )) ' P N (n,g,4)−1 .
 MATH 249B NOTES: ALTERATIONS 53

Proof. See the handout “Universal Stable Curve” on the course website.
(The only reason we use the 4th tensor power rather than the 3rd is to avoid
some headaches that arise in certain cases with irreducible geometric fibers
when g ≤ 1. by artful use of the stability condition for such cases, as ex-
plained in [Knu], one can push through the use of m = 3 in general. For our
purposes, even using m = 538 would be sufficient.) 
The proof of the preceding result provides a universal 4-canonically em-
bedded n-pointed stable genus g curve,

(X ; {τi })
(9.2)

S
with S a Z-scheme that is a quasi-projective PGL N ( g,n,4)−1 -torsor over an
open inside the n-fold fiber product Z ×HΦ · · · ×HΦ Z , where HΦ is Hilbert
scheme for a specific projective space and a specific degree-1 polynomial
Φ = Φ g,n (t) ∈ Q[t].
Any n-pointed stable genus-g curve f : ( X; {σi }) → S over any scheme is
a pullback of
(X ; {τi }) → S
Zariski-locally on S upon trivializing f ∗ (ωX/S (∑ σi )⊗4 ) Zariski-locally on
S; there is a PGL( N ( g, n, 4))’s worth of choices. In particular, using the
natural action of PGL( N (n, g, 4)) on S arising from its universal property,
the vague dream is that the quotient S / PGL( N (n, g, 4)) (whatever it might
mean) is a moduli space for such data of n-pointed stable genus-g curves
( X → S, {σi }) over variable base schemes.
Remark 9.3. The preceding method is often used to build moduli spaces:
we have something too structureless to make a geometric construction, so
we put extra structure on it to connect it to something we can work with
(such as a trivialization of a vector bundle or a projective-space bundle, or
a level-structure on an abelian variety, etc.). Then, we quotient the output
about by a suitable group action to obtain the desired moduli space for the
original more structureless data.
There are two potential problems with the preceding dream involving
S / PGL( N (n, g, 4)):
(1) The PGL( N (n, g, 4))-action on S is very far from free (say on geo-
metric points with values in a fixed but arbitrary algebraically closed
54 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

field), since stable curves (and even smooth curves) can have non-
trivial automorphisms. Namely, by using a “4-canonical embed-
ding”, if the data ( X0 , {si }) over an algebraically closed field k ad-
mits non-trivial automorphisms then such automorphism transfer
to automorphisms of the ambient projective space preserving the
marked curve inside that projective space. This gives a nontrivial
element of PGL( N (n, g, 4) − 1)(k ) that fixes the point in S (k ) corre-
sponding to ( X0 , {si }) equipped with the trivialization of its ambient
projective space.
(2) Even on a dense open of S where the PGL-action is free (if such
were to exist, though in this case does not) it is not at all clear how to
make sense of such a quotient as a scheme. This gets involved with
very difficult matters involving “GIT over Z” (not to be confused
with GIT-hub), but anyway is totally inadequate here since we can-
not ignore the stable marked curves with nontrivial automorphisms.
Instead, we require a “geometry of functors” (or of fibered categories),
which is Artin’s approach to moduli problem: we abandon trying to make
actual schemes and instead identify a class of functors (or fibered categories)
near enough to schemes that we can develop for them much of the famil-
iar machinery of algebraic geometry. Of course, making definitions is not
enough: one also needs an “EGA” for these things (largely done in the case
of algebraic spaces by Donald Knutson in his PhD thesis under Artin).
Definition 9.4. For a scheme S, we let M g,n (S) denote the category of n-
pointed stable genus g curves ( X, σ) over S, using only isomorphisms over
S as the morphisms (so this category is a groupoid).
For maps S0 → S we have a base change functor
bS0 /S : M g,n (S) → M g,n (S0 )
and for S00 → S0 → S we have an isomorphism of functors
bS00 /S0 ◦ bS0 /S ' bS00 /S ,
meaning an isomorphism of the two maps
M g,n (S) ⇒ M g,n (S00 ),
and an evident associativity condition relative to S000 → S00 → S0 → S.
Remark 9.5. The assignment S M g,n (S) is a sheaf of categories (aka
stack) for the fpqc topology in the following sense:
Suppose we are given an fpqc map S0 → S and a pair
X0, θ


 MATH 249B NOTES: ALTERATIONS 55

for X 0 ∈ M g,n (S0 ) and θ a descent datum relative to S0 → S. Let us now

explain what we mean by a descent datum. By definition, θ is an isomor-
phism

θ : p1∗ ( X 0 ) ' p2∗ ( X 0 )

in the category M g,n (S0 ×S S0 ), where p1 , p2 are the base change functors
along the two projections S0 ×S S0 ⇒ S0 , satisfying the cocycle condition on
the induced isomorphisms in

M g,n S0 ×S S0 ×S S0

with respect to the three pullbacks

S0 ×S S0 ×S S0 → S0 ×S S0 .

This is a commutativity condition on maps, which we won’t write out here.

For any such ( X 0 , θ ), the “stack” condition is that there exists a unique
object X ∈ M g,n (S) up to unique isomorphism equipped with an isomorphism
XS0 ' X 0 in M g,n (S0 ) carrying the canonical S0 /S-descent datum on XS0
(arising from X) over to θ.
In our context with objects ( X, σ ) we have the associated line bundle
ωe X/S = ωX/S (∑ σi ) that is S-ample on X and canonically attached to the given
data over S in the sense of being functorial with respect to base change on
S and canonically compatible with all isomorphisms in such ( X, σ). In par-
ticular, ω e X/S is naturally compatible with any descent datum on ( X, σ ), so
by Grothendieck’s effectivity criterion for fpqc descent in the presence of
ample line bundles compatible with the descent datum it follows that for
any fpqc S0 → S the above unique effective descent property holds for any
( X 0 , σ0 ) ∈ M g,n (S0 ) equipped with a descent datum relative to S0 → S.
We therefore say M g,n is a stack (in groupoids) for the fpqc topology on
the category of schemes.

We have the following key lemma.

Lemma 9.6. The diagonal map

∆ : M g,n → M g,n × M g,n

x 7→ ( x, x )

is relatively representable in quasi-compact and even finite type scheme maps.

56 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Proof. For any S we have the fiber square

IsomS ( x, y) S
(9.3) ( x,y)

∆
M g,n M g,n × M g,n

for x, y ∈ M g,n (S), by definition of fiber products of fibered groupoids,

where the upper-left is the Isom-functor (carrying any S-scheme T to the set
IsomT ( x T , y T )).
Grothendieck proved Isom-functors are represented by locally finitely
presented S-schemes, constructing such Isom-schemes from Hom-schemes
that in turn are built inside Hilbert schemes. The point is that if X and Y are
the curves underlying x and y then one builds IsomS ( x, y) inside HilbX ×Y/S
via the graphs of isomorphisms Γ f → X × Y. Since Hilbert schemes are built
(in the presence of a relatively ample line bundle) as a countable disjoint
union of finitely presented schemes over the base (even quasi-projective, at
least Zariski-locally on the base), the same holds for the Isom-schemes by
design.
The crucial point is to show that in our setting with curves, the relevant
Isom-schemes are even finite type over the base, not merely locally of finite
type. This amounts to controlling Hilbert polynomials of graphs of isomor-
phisms. More specifically, for a field k and ( X, σ) , ( X 0 , σ0 ) ∈ M g,n (k ) we
claim that the graph Γ f ⊂ X × X 0 of any isomorphism

f : ( X, σ ) ' X 0 , σ0

has precisely one possibility for its Hilbert polynomial with respect to the
ample line bundle
L := p1∗ (ω
e X/k ) ⊗ p2∗ (ω
e X 0 /k )
on X × X 0 : it must be Φ(2t) (where Φ = Φ g,n is as in our earlier discussion).
This comes down to an elementary calculation: via the identification of
Γ f with X using the inclusion (1, f ) : X → X × X 0 , we have
e X/k ⊗ f ∗ ω
L |Γ f ' ω e X 0 /k
⊗2
'ω
e X/k .

Proposition 9.7. The map S → M g,n is a representable in smooth scheme covers;
more precisely, for S → M g,n corresponding to an n-pointed stable genus-g curve
 MATH 249B NOTES: ALTERATIONS 57

f : ( X, σ ) → S, the fiber product

F S
(9.4)
S M g,n

is a scheme that is a PGL( N ( g, n, 4))-torsor over S.

In particular, M g,n is a quasi-separated Artin stack of finite type over Z.
Proof. By the universal property of the structure over S that underlies the
definition of the bottom horizontal arrow (via “forgetting” the trivialization
of a projective-space bundle), the fiber product is the Isom-functor over S
⊗4
for P N (n,g,4)−1 and P( f ∗ (ω
e X/S )). This is a PGL( N ( g, n, 4))-torsor over S.
By definition, a quasi-separated Artin stack of finite type over Z is pre-
cisely a stack in groupoids M for the fppf (let alone fpqc) topology on the
category of schemes such that (i) ∆ M : M → M × M is represented in quasi-
compact scheme maps, and (ii) there exists a “smooth cover by a scheme of
finite type over Z”: a map Y → M from a Z-scheme Y of finite type such that
for any scheme S and map S → M corresponding to an object in M(S) the
fiber product Y × M S is a scheme that is smooth surjective onto S. Hence,
M g,n is an Artin stack of the asserted type. 
In [DM] the notion of Artin stack was not discussed and instead the
smooth scheme cover of M g,n was (without too much explanation) “sliced”
down to an étale cover by a scheme. This slicing ultimately relied on the
fact that Aut-schemes at geometric points are étale, a property we noted in
Lemma 8.10. The same technique works more generally:
Theorem 9.8. If X is an Artin stack of finite presentation over a scheme T and its
finite-type Aut-schemes at geometric points are étale then X admits an étale cover
by a scheme; i.e., it is a Deligne–Mumford stack.
We note that “finite presentation” includes quasi-separatedness (so M g,n
is finitely presented over Z, which is stronger than “finite type” precisely
because of the quasi-compactness of the diagonal that we have seen in-
volved an actual argument, not just trivialities). The Aut-schemes are lo-
cally finite type due to the Artin stack being locally of finite presentation
over the base, and they are finite type (or equivalently, quasi-compact) due
to the quasi-separatedness of the Artin stack. Indeed, if k is a field and
x ∈ X (k ) is an object then its Aut-functor is the base change of ∆Xk /k along
( x, x ) : Spec (k) → Xk × Xk :
x ×X x = Spec (k ) ×Xk ×k Xk Xk .
58 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Proof. To prove Theorem 9.8, we first make an important observation about

the fiber product
X ×∆X ,X ×T X S
for any T-scheme S and T-map ( x, y) : S → X × X : this is not only a
finitely presented algebraic space over S (as follows from the very definition
of X being a finitely presented Artin stack over T), but has vanishing rela-
tive Ω1 , or in other words is unramified over S in the sense of Grothendieck.
This is a property that is sufficient to check on geometric fibers over S, so
it comes down to the property that for any algebraically closed field k and
objects x, y ∈ X (k ) the finite type Isom-scheme Isom( x, y) is étale. But if
non-empty then (since k = k) this is a copy of the Aut-scheme Aut( x ) that
is étale by hypothesis. With the unramifiedness of ∆X /T established, the
Deligne–Mumford property of X is exactly [LMB, Thm. 8.1]. 
The preceding work yields the following result of Deligne and Mumford
that we shall use extensively:
Corollary 9.9. The finitely presented Artin stack M g,n → Spec (Z) is actually a
Deligne-Mumford stack.
Here is an important refinement:
Theorem 9.10. The finitely presented Deligne-Mumford stack M g,n over Z is
smooth and proper.
The proof involves an induction on n (the base cases being n = 0, g ≥ 2
and (n, g) = (1, 1), (3, 0), with the base cases for g ≥ 2 involving serious
work in deformation theory in [DM, §1]). We’ll say a bit about the proof
soon, especially a beautiful geometric idea of Knudsen.
Question 9.11. Where are we at?
Recall that we have X → Y with sections τ1 , . . . , τn satisfying conditions
coming out of the 3-point lemma (and Z = ∪τi (Y )). For a dense open U ⊂ Y
over which X is smooth and the τi ’s are pairwise disjoint, this corresponds
to a map h : U → Mg,n ⊂ M g,n into the open substack classifying smooth
marked curves. Inspired by the standard device of extending the domain
of a rational map via blow-up of the source, we can dream of altering Y so
that h extends to a map h : Y → M g,n . This is a geometric problem.
Once that is done, the plan is to exploit the “3-point lemma” property of
∗
the τi ’s to show that the resulting pullback by h of the universal marked sta-
ble family, which agrees over U with ( X, τ )|U , can dominate ( X, τ ) over the
entirety of Y via a birational map (possibly after a further alteration of Y).
Then we would be in very good shape: studying a semistable fibration with
 MATH 249B NOTES: ALTERATIONS 59

smooth geometrically connected generic fiber over Y of dimension d − 1.

Then we apply induction to alter Y to be smooth, and so are brought to such
a fibration over a smooth base; that seems very tractable for even running a
concrete resolution of singularities via blow-ups! (But don’t forget: we need
to keep track of Z = ∪τi (Y ) throughout this process.).
Before returning to our regularly scheduled program of de Jong’s proof,
we say a few things concerning the proof of:
Theorem 9.12 (Deligne-Mumford, Knudsen). The finitely presented Deligne-
Mumford stack M g,n over Z is smooth and proper over Z.
Remark 9.13. Smoothness is detected by deformation rings, which are com-
pletions of strictly henselian local rings at points of M g,n . (These deforma-
tion rings might not exist at points for Artin stacks in general, though there
are versal deformations.) Properness is detected by a valuative criterion.
Proof. We first discuss the case n = 0, g ≥ 2. We then induct on n ≥ 0 with
fixed g ≥ 2. At the end we treat the cases with g ≤ 1.
The case n = 0, g ≥ 2 is treated in [DM, §1]. The issue for smoothness is
to understand the deformation theory of reducible semistable curves. This
is used to compute the deformation rings to check Z-smoothness. In fact,
they even describe the deformation rings for Mg,n ⊂ M g,n , and the descrip-
tion of deformation rings yields M g,n − Mg,n is a Z-flat Cartier divisor, so
Mg,n ⊂ M g,n is fiberwise dense over Spec Z (which will be crucial for the
valuative criterion).
Using the uniqueness of stable models, Deligne and Mumford showed
the quasi-finite Isom-schemes for stable genus-g curves are proper (hence
finite). This expresses properness for the quasi-compact diagonal of M g,n ,
so this stack is separated. (The diagonal for an Artin stack is represented in
algebraic spaces, and the definition of separatedness for stacks is that the di-
agonal is proper. Since the valuative criteria for separatedness and proper-
ness continue to hold for quasi-separated algebraic spaces, it follows that
the usual formulation of the valuative criterion for separatedness holds for
Artin stacks.)
There is a Chow’s Lemma for separated DM stacks of finite type over
a noetherian ring A, with the birational condition relaxed to a generically
étale condition. This is recorded in the handout on avoiding Gabber’s The-
orem, and it gives that for M a DM stack separated and finite type over
A, there exists a surjective proper generically étale map Y → M with Y
quasi-projective over A.
This yields a valuative criterion for properness of M in which we allow
finite extensions on the discrete valuation ring (which can be taken to be
60 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

complete with an algebraically closed residue field). The idea is that by

Chow’s lemma, M fails to be proper if and only if the quasi-projective Y is
not proper, and the latter provides a point outside Y that is contained in the
closure of Y in some projective space over A. We pick a map to Y from the
spectrum of a discrete valuation ring in a such a way that the generic point
maps into the dense open locus étale over M and the closed point lands
outside Y.
Exercise 9.14. In the setting of a map of schemes, if we weaken the valuative
criterion to only require that maps extend after a finite extension on the
discrete valuation ring then in fact the map extended before extension of the
discrete valuation ring. Hence, when the valuative criterion of properness
for stacks is specialized to schemes it recovers exactly the criterion usually
used for schemes.
To verify the valuative criterion of properness for M g,n , it is enough to
consider discrete valuation rings whose generic point maps into Mg,n The
semistable reduction theorem for smooth geometrically connected curves
of genus g ≥ 2 over the fraction field of a discrete valuation ring R involves
exactly such an extension on R. In [DM, §2] it is shown in cases with R/m
algebraically closed (as is sufficient for us) how to find a stable R-model
after a suitable finite extension on Frac( R); this construction rests on ma-
nipulation of the minimal regular proper model, guided by the theory of
abelian varieties (especially work of Raynaud relating the Néron model of
the Jacobian to the minimal regular proper model).
Now, for a fixed g, we will induct on n. Knudsen’s strategy is to show
that the fibered category M g,n+1 is isomorphic to the universal curve Zg,n
over M g,n .
Warning 9.15. We have to describe the map M g,n+1 → M g,n ; this entails a
bit of a subtlety because simply forgetting a point may lose stability.
As motivation, consider Y → S proper and finitely presented (even pro-
jective) with universal family
Z Y × HilbY/S
(9.5)

HilbY/S
What is the functor of points of Z ?
Lemma 9.16. The functor of points of Z is
Z ( T ) = {( Z, σ) : Z ∈ HilbY/S ( T ), σ ∈ Z ( T )} .
 MATH 249B NOTES: ALTERATIONS 61

That is, Z ⊂ YT is finitely presented and T-flat and σ is a section over T of the map
Z → T.

Proof. To give

T Z
(9.6)
S

amounts to giving a map

T Z
(9.7) h
φ
S

and a lift of h through φ The map h is precisely the T-flat Z ⊂ YT , with

Z = T ×h,HilbY/S ,φ Z ⊂ YT . But then to give a section σ : T → Z over T is
precisely to give a map

T→Z

lifting h through φ. 

Similarly to Lemma 9.16, Zg,n is a fibered category of data

(( X; σ1 , . . . , σn ) , σn+1 )

Beware that σn+1 may not be in sm( X/S) and may meet some σi for i ≤ n.
Thus, this data is not an (n + 1)-pointed stable curve in general. To get
around this issue, Knudsen defined contraction and stabilizations functors.
That is, he defined maps of fibered categories
(1) Zg,n → M g,n+1 (which needs “stabilization”) and
(2) M g,n+1 → Zg,n forgetting σn+1 (which needs “contraction”).
The idea of the second map is that if forgetting a marked point on a P1 -
component in a geometric fiber ruins stability then we should contract that
fibral P1 -component. To do this in the relative setting for f : X → S we have
!
n +1
f ∗ (ωX/S ( ∑ σi )⊗m )
M
X = Proj
m ≥0 i =1
62 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

by S-ampleness of ωX/S (∑in=+01 σi ), so we are led to try defining a “relative

contraction” by
!
n
f ∗ (ωX/S ( ∑ σi )⊗m )
M
c( X ) := Proj
m ≥0 i =1

with sections
σ i : S → X → c ( X ).
One has to check this is is a semistable S-curve whose formation commutes
with base change and has the desired effect over algebraically closed fields,
and that forgetting σn+1 yields a stable n-pointed genus-g curve.
Defining stabilization as a functor in the other direction involves blow-
ups and more work (to ensure good behavior under base change, stability,
and so on). These two operations turn out to be naturally inverse to each
other, so we obtain an isomorphism
'
M g,n+1 Zg,n
(9.8) π

M g,n

with
π ( X; σ1 , . . . , σn+1 ) = (c( X ); σ1 , . . . , σn ) .
This gives that π is proper, so M g,n is Z-proper for all n ≥ 0 (with the fixed
value of g ≥ 2) by induction.
What about Z-smoothness of M g,n when g ≥ 2? Proceeding by induction
on n ≥ 0 with fixed g ≥ 2, by the preceding isomorphism it is the same to
show that Zg,n is Z-smooth when M g,n is Z-smooth. The issue is local at the
geometric points ξ of Zg,n is singular in the fibers over M g,n . Knudsen’s de-
formation theory analysis and Artin approximation (to be discussed later)
show that étale locally f : Zg,n → M g,n near ξ looks like

Spec ( R[t, x, y]/( xy − t))

(9.9)

Spec R[t]

for suitable rings R, with ξ corresponding to ( x, y, t) = (0, 0, 0) over some

s ∈ Spec R. The ring R[t] is Z-smooth near t = 0 over s by induction on
n, since it shares an étale neighborhood in common with (M g,n , f (ξ )), so
 MATH 249B NOTES: ALTERATIONS 63

R is Z smooth near s because Spec ( R[t]) → Spec ( R) is a smooth cover.

Note that although Spec ( R[t, x, y]/( xy − t)) is not smooth over Spec R[t],
it is smooth over Spec R! Hence, R[t, x, y]/( xy − t) is smooth over Z near
(0, 0, 0) over s. The latter shares an étale neighborhood in common with
(ξ, Zg,n ), so Z-smoothness of Zg,n is proved. The cases with g ≥ 2 (and any
n ≥ 0) are now settled.
The cases with g = 1 (so n ≥ 1) proceed by induction in exactly the same
way once the base case n = 1 is settled. The smoothness and properness of
M 1,1 over Z was proved by Deligne and Rapoport in their work on gener-
alized elliptic curves.
Finally, suppose g = 0 (so n ≥ 3). In this case one can likewise carry out
the same inductive arguments after the case n = 3 is settled. In that base
case a little miracle happens: M 0,3 = Spec Z! This says that the only object
over any ring (or equivalently over any scheme) is the automorphism-free
object
(P1 , {0, 1, ∞}).
That this is the only object over an algebraically closed field was noted some
time ago, and it then follows that the same holds over any field. Thus,
any such object ( X; σ1 , σ2 , σ3 ) over any base S has X → S smooth with its
geometrically connected fibers of genus 0, so X is a Zariski P1 -bundle by
deformation-theoretic arguments with coherent base change. The 3 mark-
ings rigidify this uniquely up to unique isomorphism over any local ring,
and then Zariski-locally on the base by spreading-out, and then globally
due to the local rigidity. 

10. A PPLICATION OF MODULI STACKS

For our needs, it is the properness of M g,n over Spec Z rather than its
smoothness that matters. We discussed its smoothness both for general
awareness and because knowledge of smoothness aspects of this stack was
used in the proof of properness (to justify the sufficiency of checking the
valuative criterion when the generic point lands in Mg,n ).
We shall use this proper stack to make a generically étale alteration on Y
so that the restriction of ( X, τ ) → Y over some dense open U ⊂ Y extends
to a stable n-pointed genus-g family (C ; σ) over Y. We’ll then have two
families over Y, the stable family and the “3-point lemma” family, and an
isomorphism between them over U. This isomorphism will be extended
to a proper birational map C → X from the stable family onto the “3-point
lemma” family after a suitable further modification of Y; it is in constructing
this Y-map C → X that the true importance of the 3-point lemma will be
seen.
64 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Recall that Z = ∪in=1 σi (Y ) ⊂ X for f : X → Y with sections σi collec-

tively satisfying the “3-point” property; in particular, n ≥ 3. By design, for
some dense open U ⊂ Y the map XU → U is smooth with geometrically
connected fibers of some genus g ≥ 0.
We can shrink U if necessary so that {σi |U } are pairwise disjoint, since
{σi (ηY )} are pairwise distinct. Since 2g − 2 + n ≥ n − 2 > 0, it makes sense
to consider the universal stable marked curve
(Zg,n , ζ ) → M g,n
with M g,n now denoting the proper stack over our field k = k obtained
by base change from Z. (This base change has the same moduli-theoretic
meaning on the category of k-schemes.) Note that
( XU → U, σ|U )
arises via pullback
 
( XU , σ|U ) Zg,n , ζ
(10.1)
φ
U M g,n
and φ factors through Mg,n since the fibers over U are smooth.
Here is our first goal:
Goal 10.1. Our dream (which will not quite be fulfilled) is to find a U-
modification fiber square
U0 Y0
(10.2) '

U Y
so that there exists a commuting triangle
U0 Y0
(10.3)
M g,n
or in other words
( XU 0 , σ|U 0 )
extends to a stable family (C , τ ) over Y 0 , and then further normalize Y 0 .
(The reason to express this task in terms of extending a map, rather than in
 MATH 249B NOTES: ALTERATIONS 65

terms of its raw moduli-theoretic meaning, is because extending a map is a

more familiar type of geometric problem and is the viewpoint that will lead
us to a solution.)
If M g,n were a scheme we could just take the graph Y 0 = Γφ ⊂ Y × M g,n
(which is a U-modification because its restriction over U is the graph Γφ →
U × M g,n that is a closed immersion since M g,n is separated); note that Y 0
is Y-proper because M g,n is proper.
Warning 10.2. There are at least two problems with forming this graph clo-
sure, due to M g,n being a stack and not a scheme:
(1) the diagonal ∆M g,n is really not even monic (let alone not a closed
immersion), yet the graph of ψ is a base change of this diagonal, so
this graph is generally not a substack of U × M g,n .
(2) Even if we ignore the first problem by simply forming the “schematic
image” of the graph (which is unlikely to be at all useful), this is
merely a stack and not a scheme.
These two issues are overcome in the 1-page handout that makes Y 0 →
Y as a generically étale alteration (so not generally as a modification); the
version of Chow’s Lemma for DM stacks involves such an alteration rather
than a birational map. The appearance of alterations here is no surprise,
much as the valuative criterion for properness of M g,n requires making a
finite separable extension on the fraction field of a discrete valuation ring in
order to extend the map as in the valuative criterion.
The upshot is that after a suitable generically étale alteration on Y (and
then further normalizing it) we can assume there exists a stable family
(C , τ ) → Y
and an isomorphism
ψ : (C , τ )U ' ( X, σ)U
over U.
Here is a further goal.
Goal 10.3. At the cost of a U-modification (or maybe a generically étale
alteration) Y 0 → Y and then replacing X with ( X ×Y Y 0 )red as usual (which
preserves all of our running hypotheses upon arranging for Y 0 to also be
normal), we want to extend ψ to a Y-map
e:C →X
ψ
that is necessarily birational (as it is an isomorphism over U) and satisfies
τi 7→ σi because it does so over U.
66 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Remark 10.4. Note that C is Y-proper with a Y-ample line bundle, so since
Y is projective over k it follows that C is also projective over k. Thus, C is
a projective variety (it is integral since it is semistable over Y with smooth
geometrically connected generic fiber).
e after a suitable change of Y, we will use the stability of
To make such a ψ
(C , τ )
and the “3-point” property of
( X, σ) .
Often, when one wants to extend a map to a proper target after a modifi-
cation of the source, a natural strategy is to pass to the closure of the graph
in an appropriate fiber product. We will do the same here.
Goal 10.5. Consider the schematic closure
T := Γψ ⊂ C ×Y X.
Observe that TU = Γψ ⊂ CU ×U XU so that TU → U is smooth with
geometrically connected fibers of dimension 1. The ideal situation is that
the Y-morphism pr1 : T → C is an isomorphism. We know pr1 is at least
proper birational, since it is an isomorphism over U.
We’ll see later that if Y is normal (as we may arrange) then C is normal
as well (using Serre’s “R1 + S2 ” criterion). More generally, it will be shown
that any semistable curve over a normal base with smooth generic fiber has
normal total space. Granting this, it would suffice for pr1 : T → C to be
quasi-finite (and hence finite, by properness) since it would then be a finite
birational map from a variety to a normal variety and thus an isomorphism.
Such quasi-finiteness is a geometric problem for Ty ⊂ Cy × Xy for all geo-
metric points y ∈ Y.
To make progress on that geometric problem for Ty , there are at least two
things we must address beforehand:
(1) Is each Ty at least a curve (rather than of dimension 0, or filling up
Cy × Xy )?
(2) If so, is each Ty connected?
The key issues this will come down to are to arrange that:
(1’) T is Y-flat,
(2’) T → Y is its own Stein factorization.
Remark 10.6. We do have a bit of geometric information - the sections
(τi , σi ) : Y → C ×Y X
 MATH 249B NOTES: ALTERATIONS 67

factor through T = Γψ since we can check that over U. Therefore, the fiber
Ty ⊂ Cy × Xy
contains the points
(τi (y), σi (y))
for all i. This will provide some geometric control on the possibilities due
to the stability and 3-point conditions on these markings, provided that Ty
really is a connected curve.
Lemma 10.7. In the above setup, the map T → Y is its own Stein factorization.
(Recall that we normalized Y before forming T.)
Proof. This is a map of projective varieties over k and over U it is Γψ → U or
equivalently XU → U. Therefore, the map OY → h∗ OT is an equality over
U. But, h∗ OT is a coherent sheaf of domains that coincides with OU over U,
so Hence, h∗ OT = OY since Y is normal. 
The preceding lemma accomplishes (20 ) in Goal 10.5, and hence (2) in
Goal 10.5. To show (1) and (10 ) in Goal 10.5, we first note that TU → U is
just the smooth map XU → U whose geometrically connected fibers have
dimension 1. The following lemma thereby shows that (10 ) implies (1).
Lemma 10.8. Let Z → Y be flat and finite type between irreducible noetherian
schemes. For d := dim ZηY ≥ 0, all geometric fibers Zy have pure dimension d.
Proof. This is standard - see a handout on the course website. 
Remark 10.9. One can drop irreducibility (and even the noetherian hy-
potheses) if instead Z → Y is proper flat and finitely presented, in which
case y 7→ dim Zy is locally constant; see [EGA, IV3 , 12.2.1(ii)].
Example 10.10. Without properness, one cannot drop irreducibility from
the lemma (even with separatedness). For example, consider Y = Spec R
for R a discrete valuation ring with fraction field K and residue field κ. Glue
A1R along inversion on (Gm )K ⊂ A1R to a K-scheme that is a union of A1K
and a plane meeting at the origin (so the gluing has generic fiber that is a
connected union of P1K and a plane). This gluing process takes place entirely
over the generic fiber, so the glued scheme Z is R-flat, and it is separated.
See Figure 2.
This gluing Z is not R-proper because we “glued in” a new irreducible
component on the generic fiber so that this new component is closed in the
entire space (as we may check Zariski-locally!) but is supported entirely
over the generic point of Y. By design, the generic fiber of the gluing is
reducible of dimension 2 but the (irreducible) special fiber is 1-dimensional.
68 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

F IGURE 2. A picture of a flat (but non-proper) morphism to a

dvr whose geometric fibers are not pure dimensional.

We now want to show (10 ) in Goal 10.5. Consider TU = Γψ ⊂ VU for

V := C ×Y X ⊂ PYN . We seek a U-modification Y 0 → Y so that the schematic
closure
TU 0 ⊂ VY 0
is Y 0 -flat. Once this is achieved and Y 0 is renamed as Y then the formation
of such closure commutes with any further modification on Y since a Y-flat
separated scheme is always the closure of its restriction over a dense open
in Y. By this same principle, it suffices to find Y 0 → Y so that there exists
a Y 0 -flat closed T ⊂ VY 0 that restricts to TU 0 over U 0 ' U (as V is then
automatically the schematic closure of TU 0 , since a flat closed subscheme of
a scheme over a reduced base is the schematic closure of its restriction over
any dense open in the base).
There is a deep general result [RG, 5.2.2] by Raynaud and Gruson on
“flattening by blow-up” which provides the desired flat closed T ⊂ VY 0 for
a suitable Y 0 , but in the projective setting we can use the following simpler
trick with Hilbert schemes. Observe that the U-flat closed subscheme TU ⊂
VU ⊂ PU N has (by U-flatness) all fibers with a common Hilbert polynomial
 MATH 249B NOTES: ALTERATIONS 69

Φ ∈ Q[t]. Therefore, TU → U is a pullback

Φ
TU Z ⊂ V × HilbV/Y
(10.4)
Φ
U HilbV/Y

of a universal flat family over a projective Hilbert scheme over Y. Then,

Φ
take Y 0 to be the graph closure Γ f ⊂ Y × HilbV/Y and re-normalize this Y 0
(and rename that as Y) before forming T, which will now be Y-flat! In this
way, both goals have been achieved, so Ty is a connected curve for all y.
To recap where we are now, via the isomorphism

ψ
CU XU
(10.5)

we defined T := Γψ ⊂ C ×Y X and arranged that T → Y is flat. It was

deduced from such properties and the normality of Y that T → Y has geo-
metrically connected fibers of dimension 1, and that we have a factorization

(τi , σi ) : Y → T ⊂ C ×Y X.

Our aim is to show p1 : T → C is quasi-finite because this has been shown

to imply T → C is an isomorphism (so as to extend ψ to a proper birational
Y-morphism) provided that also C is normal. Thus, let’s now address why
C is normal, and then we will take up the geometric task of proving p1 is
quasi-finite. More generally, we have the following normality result:

Lemma 10.11. If X → S is a semi-stable curve over a normal noetherian S (no

properness assumptions) with smooth generic fibers then X is normal.

Example 10.12. Here is a model case that can be done by bare hands, to
appreciate that some actual work is involved. Consider a normal noetherian
domain A with fraction field K, and for a choice of a ∈ A − {0} let R =
A[u, v]/(uv − a). By the fibral flatness criterion this is A-flat (as uv − c is
not a zero-divisor in k[u, v] for any field k and c ∈ k), so it is a semistable
curve; its generic fiber is a hyperbola.
We leave it as an instructive exercise to the reader to directly prove that
R is a domain, and even integrally closed. Here is a hint: R is A-free with
70 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

basis 1, ui , v j for i, j ≥ 1 (since uv = a ∈ A), so we have an injection

R ,→ R ⊗ A K
= K [u, v]/(uv − a)
= K [u, u−1 ].
Thus, R is certainly a domain, and since K [u, u−1 is already an integrally
closed domain we only need to study elements of R ⊗ A K integral over R.
The preceding example will arise in an important way when we study
resolution of singularities for relative semistable curves.
Proof of Lemma 10.11. In the general case we shall use Serre’s necessary and
sufficient normality criterion “R1 + S2.” Without loss of generality S is con-
nected, hence irreducible with generic point denoted η. For x ∈ X, f is flat
at x and so by the dimension formula for flat maps we have
dim OX,x = dim OS, f ( x) + dim OX f ( x),x .
Note that
(
1 if x is closed in X f ( x) ,
dim OX f ( x),x =
0 if x is a generic point ofX f ( x) .
Recall that R1 means OX,x is regular when its dimension is at most 1, and
given R1 the condition S2 is equivalent to saying that when OX,x has dimen-
sion at least 2 then it has a regular sequence of length 2.
If dim OS, f ( x) ≥ 2, then OS, f ( x) has a regular sequence of length 2 by nor-
mality of S (using Serre’s “R1 + S2 ” criterion). By flatness of f , the pullback
of these elements to OX,x is a regular sequence. Thus, these cases are set-
tled and so we may and do now assume that either OS, f ( x) is a field (i.e.,
f ( x ) = η) or is 1-dimensional. Recalling that Xη is assuming to be a smooth
curve, all local rings are regular for x in the generic fiber.
To complete the proof, it suffices to deal with the case when OS, f ( x) is
1-dimensional, and hence a dvr (as S is normal). So now we may assume
S = Spec R for R a dvr, and consider x ∈ X0 a point in the special fiber.
Since dim OX,x = 1 + dim OX0 ,x , there are two cases:
(1) Suppose x is generic in X0 . In this case, OX,x is 1-dimensional and
so we want to show that it is a dvr. Let π ∈ R be a uniformizer.
Note that π ∈ OX,x is a regular element of mx by R-flatness. Also,
OX,x /(π ) = OX0 ,x is the local ring at a generic point of X0 , which is a
field. Therefore, π generates mx . Since OX,x is a 1-dimensional local
noetherian ring in which the maximal ideal is generated by an ele-
ment that is not a zero-divisor, OX,x is a dvr by [Se, Ch. I, §2, Prop. 2].
 MATH 249B NOTES: ALTERATIONS 71

(2) Say x is closed in X0 , so OX,x has dimension 2. We seek a 2-term

regular sequence. Since π ∈ mx is a not a zero divisor by R-flatness, it
suffices to find a regular element in the maximal ideal of the quotient
OX,x /(π ) = OX0 ,x ,
which is a 1-dimensional local noetherian ring. That is, we are asking
whether OX0 ,x satisfies S1 . By Serre’s reducedness criterion, a local
noetherian ring is reduced if and only if it satisfies “R0 + S1 ”. Thus,
we are done because X0 is reduced (by definition of semistability for
curves).

We next aim to study the fibers of T → C . Recall T ⊂ C ×Y X. We have
T
p2
p1
(10.6) C h X

Y
with h flat and Y is normal, and all geometric fibers
Ty ⊂ Cy × Xy
are connected curves. Our aim is to show that each map
( p1 )y : Ty → Cy
has finite fibers (as we saw this would imply that p1 : T → C is an isomor-
phism, as desired).
Hence, we want to show that ( p1 )y doesn’t crush any irreducible compo-
nent of Ty to a point.
10.1. Setup for Lemma 10.13. Consider the irreducible components
Ty = T1 ∪ · · · ∪ Tt ⊂ Cy × Xy
( p2 ) y
( p1 ) y
C1 ∪ · · · ∪ Cs = Cy X1 ∪ · · · ∪ X r = X y .

Note that p1 , p2 are proper birational, hence surjective, so ( p1 )y , ( p2 )y are

surjective. Thus, for all Xi there exists some j for which Tj surjects onto Xi
and for all Cα where exists a Tγ that surjects onto Cα . We need to know
(among other things) that such j, γ are unique:
72 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Lemma 10.13. In the preceding setup, the following results hold.

(i) For each 1 ≤ i ≤ r there exists a unique 1 ≤ j(i ) ≤ t so that Tj(i)  Xi .
Also, there exists an open V ⊂ X meeting Xy densely such that p2−1 (V ) '
V.
(*) Moreover,
(10.7) Tj(i) → Cy
is not constant.
(ii) For each 1 ≤ α ≤ s there exists a unique 1 ≤ γ(α) ≤ t so that Tγ(α) 
Cα . Also, there exists an open W ⊂ C meeting Cα densely such that
p1−1 (W ) → W is an isomorphism.
Remark 10.14. The proof of the condition (10.7) will crucially use the prop-
erties from the 3-point Lemma.
Before we prove the preceding lemma, we record its main consequence
for our needs:
Corollary 10.15. The map ( p1 )y is quasi-finite.

Proof. If ( p1 )y is not quasi-finite, it crushes some Tj to a point in Cy . There-

fore, by (10.7), Tj cannot map onto any Xi , so it must be crushed to a point
by ( p2 )y . But Tj ⊂ Cy × Cy , and since the map to each component is a point,
Tj must itself be a point. However, Tj is a curve! Contradiction. 
Using that p1 is quasi-finite, we can even say it is an isomorphism:
Corollary 10.16. The map p1 : T → C is an isomorphism.
Proof. Since ( p1 )y is quasi-finite for all geometric points y (Corollary 10.15),
p1 is quasi-finite. Note also that C is normal by Lemma 10.11. Therefore,
p1 is a proper birational quasi-finite map to a normal variety C , hence an
isomorphism. 
10.2. Proof of Lemma 10.13(i) and (ii). We now prove (i) without (*), and
the same exact method will give (ii) since the argument will not use the
sections σi (y) or τj (y) that break the symmetry between X and C (in other
words, the argument will treat Xy with methods that apply equally well to
Cy ).
Suppose for some j0 6= j that Tj , Tj0 both map onto Xi under ( p2 )y . Since
Tj ∩ Tj0 is finite (and possibly empty) this would imply the fibers of

p2−1 ( Xi ) → Xi
 MATH 249B NOTES: ALTERATIONS 73

would have size at least 2 away from a finite subset of Xi . Therefore, to get
a contradiction it is enough to find an open V ⊂ X meeting Xy densely (in
particular, meeting Xi ) such that

p2−1 (V ) → V

is an isomorphism
Recall that sm( X/Y ) meets all Xy in a dense open. Also, ( p2 )y : Ty → Xy
has only finitely many positive dimensional fibers. Hence, the subset
n o
Ω= x∈X: p2−1 ( x ) is 0-dimensional

of X meets Xy away from a finite set of closed points. By upper semiconti-

nuity of fiber dimension, Ω is open in X.
We conclude that

V := Ω ∩ sm( X/Y ) ⊂ X

is an open subset of X meeting Xy densely and

p2−1 (V ) → V

is a finite birational map (as it is proper quasi-finite, and p2 : T → X is

birational between varieties because T = Γψ with ψ : CU ' XU ).
Thus, it remains to show that this finite birational map between varieties
is an isomorphism. For that purpose it is enough to show V is normal. But
V → Y is smooth (due to how V was built) and Y is normal, so V is also
normal! (Here we essentially used that the smooth locus of X → Y meets
every fiber in a dense open subset.)
It remains to prove Lemma 10.13(*).

10.3. Setup and proof of Lemma 10.13(*). Recall our setup: We have

(10.8) C X
h τ
σ f
Y
 74 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

with T flat over Y, both Y and C normal, and T defined to be the closure in
C ×Y X of the graph of a U-isomorphism
ψ
CU XU
(10.9)

U
that carries τi |U to σi |U for all i. We have seen all Ty are connected curves
inside Cy × Xy and the section
(τi , σi ) : Y → C ×Y X
factors through T (as may be checked on U ⊂ Y). In particular, Ty contains
(τi (y), σi (y)) for all i.
10.3.1. Proof of Lemma 10.13(*). Using the notations of Lemma 10.13, we
want to show that for each i, the unique Tj(i) mapping onto Xi is not crushed
to a point in Cy . Assume to the contrary for some i that
( p1 )y ( Tj(i) ) = {c}
for some c ∈ C (y). We seek a contradiction.
The surjective map Tj(i)  Xi hits all markings σj (y) ∈ Xi , and there are
at least 3 of these; let’s label them as
xα := σα (y), x β := σβ (y), xγ := σγ (y)
in Xi ∩ sm( X/Y )y . In particular, xα , x β , xγ do not lie in any components of
X other than Xi .
We will use crucially that
tα := (τα (y), σα (y)) ∈ Ty
and likewise for β, γ. We’ll argue separately depending on whether or not
c ∈ τα (y) , τβ (y) , τγ (y) ⊂ (C )sm

y .

That is, we shall treat two separate cases:


Case 1. c ∈/ τα (y) , τβ (y) , τγ (y) . In this case, we’ll find three irreducible
components of Cy through c, which is a contradiction since Cy is
semistable.
/ τα (y) , τβ (y) , τγ (y) . In this case c ∈ (C )sm

Case 2. c ∈ y and we’ll find two
irreducible components of Cy through c, again a contradiction.
Note that tα 7→ σα (y) = xα , and likewise for β, γ. We label Cases 1 and 2 as
Proposition 10.17 and Proposition 10.19 below.
 MATH 249B NOTES: ALTERATIONS 75

Proposition 10.17. If

c∈
/ τα (y) , τβ (y) , τγ (y) .
then there are three irreducible components of Cy through c.
Proof. For this, we need the following lemma.
Lemma 10.18. The non-empty fibers
p2−1 ( xα ), p2−1 ( x β ), p2−1 ( xγ )
are each connected chains of irreducible components of Ty . Moreover, their p1 -
images
p1 ( p2−1 ( xα )), p1 ( p2−1 ( x β )), p1 ( p2−1 ( xγ ))
are also connected chains of irreducible components of Cy .

Proof. Since p2−1 ( x ) ⊂ Cy × { x } for x ∈ X (y), clearly p2−1 ( x ) → p1 ( p2−1 ( x ))

is an isomorphism. Hence, we can focus on the claim concerning p2 -fibers
(though in fact it will proceed via the p1 -images of such fibers).
First, we show that the visibly pairwise disjoint fibers
p2−1 ( xα ), p2−1 ( x β ), p2−1 ( xγ )
are each either a point or a connected chain of irreducible components of Ty .
Since Tj(i) → Xi is surjective, certainly each of these p2 -fibers meets Tj(i) . We
claim that each is connected, for which it suffices to show that p2 : T → X
is its own Stein factorization over the open subset sm( X/Y ) ⊂ X whose
y-fiber contains xα , x β , xγ . Since p2 is a proper birational map and sm( X/Y )
inherits normality from Y, the claim concerning Stein factorization follows.
We conclude that each of
p2−1 ( xα ), p2−1 ( x β ), p2−1 ( xγ )
is either a point or a connected chain of irreducible components of the con-
nected curve Ty .
To complete the proof, we want to show none of
p2−1 ( xα ), p2−1 ( x β ), p2−1 ( xγ )
are points. Let’s just proceed to show p2−1 ( xα ) is not a point, as the situation
is symmetric with respect to α, β, γ.
It suffices to show p1 ( p2−1 ( xα )) is not a point. By hypothesis ( p1 )y ( Tj(i) ) =
{c} and
Tj(i)  Xi 3 xα
76 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

so c ∈ p1 ( p2−1 ( xα )). Since xα = σα (y), we have

τα (y) = p1 (τα (y), σα (y)) ∈ p1 ( p2−1 ( xα )).
This point is distinct from c by hypothesis, so p2−1 ( xα ) is indeed not a point.

Consider p1 ( p2−1 ( xα )). By Lemma 10.18, we know this is a connected
chain of irreducible components of Cy . In particular this has an irreducible
component through c in Cy . To get three distinct irreducible components of
Cy through c, (using x β , xγ also), we just need to ensure
p1 ( p2−1 ( xα )), p1 ( p2−1 ( x β )), p1 ( p2−1 ( xγ ))
have no irreducible components in common. Since
p2−1 ( xα ), p2−1 ( x β ), p2−1 ( xγ )
are pairwise disjoint, and each p2 -fiber has p1 -image that is a union of irre-
ducible components of Cy obtained as p1 -images of irreducible components
of Ty that it contains, we only need to show that for j 6= j0 if p1 ( Tj ) and
p1 ( Tj0 ) happen to be irreducible components of Cy , then these components
are distinct. But this is exactly Lemma 10.13(ii). 
Proposition 10.19. If

c ∈ τα (y) , τβ (y) , τγ (y) .
then there are two irreducible components of Cy through c.
Proof. The proof of this is completely analogous to that of Proposition 10.17,
using that if c = τα (y), then

c∈
/ τβ (y), τγ (y)
(and similarly with α replaced by β or γ). 
10.4. Reducing to the case when X → Y is a semistable curve. By Corol-
lary 10.16 the map p1 : T → C is an isomorphism, so composing its inverse
with p2 : T → X yields a Y-map
β
C X
(10.10) h f

τ σ
Y
such that βU is an isomorphism (for a dense open U ⊂ Y), β ◦ τi = σi for all
i, and
Z = ∪i σi (Y )
 MATH 249B NOTES: ALTERATIONS 77

is the support of a Cartier divisor in X.

Since β is proper birational (a “modification”), we can replace ( X, Z ) with
 
C , β −1 ( Z )

except for the issue is that β−1 ( Z ) may not be Y-finite. But the fibration over
Y is just an auxiliary device in the service of our real goal, which is to find
a generically étale alteration of X that is smooth and in which the preimage
of Z is the support of a strict normal crossings divisor.
The properties that Z → Y is finite and generically étale will never be
used anymore (the only role for those conditions were to create the sections
σi satisfying the properties in the 3-point lemma that has finally served its
role above to create β via p1 : T → X being an isomorphism). Thus, we now
drop those running hypotheses on Z ⊂ Y. Since moreover
β−1 ( Z ) ⊂ Z := ∪ j τj (Y ) ∪ h−1 ( D )

for D := (Y − U )red , for the purpose of our real goal of finding a suitable
generically étale alteration for ( X, Z ) it is enough to work with
(C , Z ) .
By induction on dimension for our main goal in this course, there exists a
generically étale alteration
F : Y0 → Y
so that
(1) Y 0 is smooth,
(2) F −1 ( D )red is a strict normal crossings divisor in Y 0 .
Since CY 0 → C is a generically étale alteration, instead of working with
(C , Z ) it suffices to work with
   
CY 0 , Z 0 := ∪ j τj0 (Y 0 ) ∪ h0−1 ( D 0 )
where
h0 : CY 0 → Y 0
is the structure map, τi0 = (τi )Y 0 , and CY 0 is a semistable Y 0 -curve and D 0 ⊂
Y 0 is a sncd. The stability of CY 0 → Y 0 is not required in what follows, only
that it is semistable with smooth generic fiber and that the sections τi0 satisfy
τj0 (Y 0 ) ⊂ sm(CY 0 /Y 0 ).
We are now in a geometrically very favorable situation! The next order of
business is to understand a global intrinsic process for resolution of singu-
larities for semistable curves with smooth generic fiber over a regular base
78 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

scheme, beginning with the key case of a discrete valuation ring as the base.
Eventually we need to incorporate “Z” into this process, but first we should
understand such resolution without the distraction of “Z”.
Remark 10.20. If we were only in the case of projective varieties, we would
not need to keep track of Z throughout the argument. But, if we started
with a non-proper X at the start of the main goal of this course then it is
very important is useful to have Z in order to arrange that the final smooth
generically étale alteration of X is the complement of a sncd in a smooth
projective variety. Note that the way we applied the inductive hypothesis
above was for the pair (Y, D ) where we know nothing at all about D. For
this reason, it is essential for the induction that the proper closed set Z is
allows to be rather arbitrary.
Up to the issue of making Z an sncd, we have essentially reduced our
task to resolving singularities on a semistable curve with smooth generic
fiber over a regular base. As a warm-up, we’ll start with the case that the
base is a discrete valuation ring.

11. S EMISTABLE CURVES OVER A DISCRETE VALUATION RING

Let’s start with a toy (or local) version.
Example 11.1. Take R to be a discrete valuation ring, π a uniformizer, and
n ≥ 1. Consider
Cn := Spec R[ x, y]/ ( xy − π n ) .
We’ll see later using Artin approximation that any point on a semistable
curve with smooth generic fiber over R has an étale neighborhood that is
étale over Cn for some n.
The generic fiber of Cn is clearly smooth and its special fiber is smooth
away from one point, call it ξ. Then,
n
O
[ Cn ,ξ = RJx, yK/ ( xy − π )
b

with xy − π n ∈ m2 if and only if n ≥ 2. Therefore, this is not regular if and

only if n ≥ 2.
We will later work out some blow-up calculations which will show that
Blξ (Cn ) is regular if n = 2 and that this blow-up is covered by copies of
Cn−2 if n ≥ 3 (hence also regular if n = 3, but not so if n ≥ 4).
It follows that we have a coordinate-free intrinsic resolution process in
this case: blow up all finitely many non-regular points, and repeat. This
eventually ends.
 MATH 249B NOTES: ALTERATIONS 79

We want a version of this in general, but we will have to link a given

abstract semistable curve X → Spec R to such Cn ’s in a way that permits
transferring over information about blow-up calculations. There will be
some issues to address due to non-rational non-regular points.
We’ll also later have to generalize to a higher-dimensional base than R as
above. For X → Y a semistable curve fibration over a connected regular Y
with dimension more than 1, we’ll have a blow-up process to push X sing :=
X − reg( X ) into codimension at least 3. Some calculations are required to
affirm that the blow-ups we make of semistable curves are still semistable.

11.1. Describing the étale local structure of semistable curves. Let Y be

any scheme and f : X → Y a semi-stable curve, with no properness assump-
tions; i.e., f is flat and locally finitely presented and for y ∈ Y a geometric
point, Xy is a reduced curve such that if x ∈ Xy is not smooth then

O
[ Xy ,x ' κ ( y )Ju, vK/ ( uv ).

Our goal is to describe the étale-local structure of f . The key to linking this
completion to our toy model via a common étale neighborhood will be Artin
approximation, which is discussed in [deJ, §2.21-2.23,§3.1-3.5] and (with full
proof) in [BLR, §3.6].
Remark 11.2. We will only need Artin’s version below, which concerns for
A essentially finite type over a field or an excellent Dedekind domain. But
we will state the ultimate general version below (due to further deep work
of Popescu). Note that excellent Dedekind domains include Z and any com-
plete dvr. The final part of [deJ] requires Artin approximation over a com-
plete dvr; for us the version for A essentially of finite type over a field will
be sufficient (though is no easier to prove than Artin’s result in his own
generality).
Definition 11.3. Let A be a local ring. Then the henselization of A, denoted
Ah is by definition
Ah := lim A0
−→ 0
A→ A

where the limit is taken over all local-étale maps with trivial residue field.
(The collection of such local-étale maps is directed and rigid: there is at most
one local A-algebra map between any two such A0 , and any two are “dom-
inated” by a common third one. Thus, such a direct limit makes sense.)
Fact 11.4. If A is noetherian then Ah is noetherian with the same completion
of A; see [EGA, 0III , 10.3.1.3]. Further, it is excellent when A is.
80 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Theorem 11.5 (Artin-Popescu approximation). Let ( A, m) be an excellent local

noetherian ring. Let B = A[ x1 , . . . , xn ]/( f 1 , . . . , f m ) be a finite type A-algebra.
The map
b) = lim Hom A ( B, A/mn+1 )
Hom A ( B, Ah ) → Hom A ( B, A
←− n
A

has dense image (in the m A -adic topology on the target).

Proof. See [BLR, §3.6]. 
Exercise 11.6. Using the ramification theory of dvr’s, one can show without
Artin approximation that if A is a dvr then Ah ⊂ A b is the algebraic closure
of A inside A.b Artin approximation yields the same with A any normal
excellent local domain.
Remark 11.7. Let’s deduce a very useful enhancement of the preceding re-
sult. For C essentially finite type over A, meaning (for our purposes) a local
ring at a prime on a finite-type A-algebra, we have
b) = HomC (C ⊗ A B, C
Hom A ( B, C b)

Now, applying Theorem 11.5, we obtain a map

Hom A ( B, C h ) = HomC (C ⊗ A B, C h ) → HomC (C ⊗ A B, C
b) = Hom A ( B, C
b)

which has dense image and coincides with the evident natural map.
Corollary 11.8. For such ( A, m) as in Theorem 11.5 and n ≥ 2, and C1 , C2 two
local essentially finite type A-algebras. Given an isomorphism
c2 ' C
f :C c1

over A
b (equivalently over A), there exists a residually trivial local-étale extension
C1 → C10 and a local A-algebra map
φ : C2 → C10
so that
(1) the induced map
b:C
φ c0 = C
c2 → C c1
1

is an isomorphism (so φ is local-étale and residually trivial),

b ≡ f mod mn .
(2) φ cC1

Before proving the corollary, we consider an example.

 MATH 249B NOTES: ALTERATIONS 81

Example 11.9. Suppose A = k is a field and X1 , X2 are two finite type

schemes over k and xi ∈ Xi (k ) for i = 1, 2. Given a k-algebra isomorphism
O
\ X1 ,x1 ' OX2 ,x2
\
there exists a common residually trivial étale neighborhood

( X10 , x10 )
(11.1)

( X1 , x 1 ) ( X2 , x 2 )

by setting Ci = OXi ,xi and spreading out C10 to make ( X10 , x10 ) .
Proof of Corollary 11.8. We have C2 = Bp for B of finite type over A. Then,
letting Hom A,loc denote the set of local A-algebra maps, we have
 
f ∈ Hom A,loc C2 , C1
c c

= Hom A,loc (C2 , C

c1 )
n o
= h ∈ Hom A ( B, C c1 ) : h−1 (m c ) = p
C1
n o
0 −1
≈ lim 0 h ∈ Hom A ( B, C1 ) : h (mCc ) = p
C1 →C1 1

= lim 0 Hom A,loc (C2 , C10 )

C1 →C1

where the “≈” step uses approximations modulo mnc and so in particular
C1
maintains the condition on the h-preimage.
Given a local A-algebra map φ : C2 → C10 so that
b ≡ f mod mn ,
φ
we would like to show φ
b is an isomorphism. The composite map
f φ
c0 = C
b
c1 ←
C −Cc2 → C c1
1

obtained by inverting the isomorphism f is the identity modulo m2 , so the

composite map is surjective due to successive approximation. But if R is a
noetherian ring then any surjective map R → R as rings is an isomorphism,
so the composite map is an isomorphism and hence so is φ b. 
Example 11.10. Let k be a field and X a curve over k. Let x ∈ X (k ) with
Od
X,x ' kJu, vK/ ( uv ).
82 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Then, for C = {uv = 0} ⊂ A2k , then ( X, x ) and (C, (0, 0)) have a common
residually trivial étale neighborhood.
For example, X could even be irreducible, say y2 = x3 + x2 , the nodal
cubic.
Example 11.11. We could ask for a variant of the previous example. Let k
be a field and X a curve over k. Let x ∈ X is a closed point with
O
[ Xk ,x ' kJu, vK/ ( uv ).

Does the same conclusion of Example 11.10 hold, but without the residually
trivial assumption? That is, can we find a common étale neighborhood?
The issue is that we need to worry about inseparable extensions. If so, then
k ( x )/k is separable (assuming that n is even if the characteristic is 2).
The reason we care about non-algebraically closed fields because we’ll
be applying this to the residue field at a generic point of a singular locus,
and in varieties over any fields of positive characteristic (even algebraically
closed) the residue fields at generic points of positive-dimensional closed
subschemes are never perfect. So, we’ll absolutely need to handle the above
issue for general fields k.
11.2. Quadratic forms in a geometric setting.
Definition 11.12. Say X → S is a flat finitely presented map, s ∈ S is a point.
A closed point x ∈ Xs is called an ordinary double point if for compatible
“algebraic” geometric points s over s and x over x and s we have an k(s)-
algebra isomorphism
O
[ Xs,x ' k ( s )Jt1 , . . . , tn K/q
for a nonzero quadratic form q that is non-degenerate in the sense that the
projective quadric (q = 0) ⊂ Pn−1 is smooth.
It is easy to check that in the preceding definition the choice of x doesn’t
impact the existence or not of such an isomorphism.
Remark 11.13. Given a quadratic form q on a nonzero finite-dimensional
vector space V over a field F, consider the associated symmetric bilinear
form
Bq (v, w) := q (v + w) − q(v) − q(w)
on V. Note that Bq is non-degenerate implies the projective quadric (q =
0) ⊂ P(V ∗ ) is smooth, and the converse holds except precisely when n is
odd and char s = 2; see HW2 Exercise 4 in the course “Algebraic Groups
I”. For example, xy + z2 is a counterexample the converse in characteristic 2
(for this case, the z-axis is a line that is Bq -perpendicular to everything).
 MATH 249B NOTES: ALTERATIONS 83

Example 11.14. Let A be a local ring and Q = ∑ ay ti t j residually non-

degenerate over A. Let a ∈ m A . Consider
X := Spec ( A [t1 , . . . , tn ] /( Q − a))
(11.2)

S := Spec A.

Let s be the closed point of S and x = ~0 ∈ Xs . If the quadratic form Q is

residually non-degenerate then some coefficient of Q is a unit, so
Q − a ∈ A [ t1 , . . . , t n ]
is nowhere a zero-divisor on the fibers of the S-flat AnS → S and hence X is
S flat by the local flatness criterion after reducing to noetherian A.
Lemma 11.15. In the above setup, for all b ∈ S, Qb is non-degenerate over k (b).
Proof. The structure map h : V ( Q) ⊂ PnA−1 → S is finitely presented, proper,
and (by the reasoning in the preceding example, with a = 0) flat. Hence,
verifying smoothness of h at a point ξ ∈ V ( Q) is equivalent to showing that
the fiber V ( Q)h(ξ ) is smooth at ξ. We know that the locus Ω ⊂ V ( Q) of
points at which h is smooth is Zariski-open and it contains the entire special
fiber by our assumption that Qs is non-degenerate (with s ∈ S the closed
point). Therefore Ω = V ( Q) by S-properness of V ( Q). 
Lemma 11.16. Let F be a field, q a non-degenerate quadratic form on F n , and
c ∈ F (e.g., F = k (b) for b ∈ S, c = a(b), q = Qb ). Then
Spec ( F [t1 , . . . , tn ] /(q − c))
is smooth except when
(
c = 0 at ~0,
c 6= 0 at a unique generic point for n odd when char F = 2.
Proof. One can assume F is algebraically closed. The case n = 1 is easy, so
we assume n ≥ 2. We can choose coordinates so that
(
t t2 + · · · + tn−1 tn if n is even
q = qn := 12
t 0 + q n −1 if n ≥ 3 is odd.
Now one can apply the Jacobian criterion. 
Corollary 11.17. When n is even and X → S as in Example 11.14 then X → S is
smooth, except precisely along the 0 section over SpecA/a ⊂ S.
Proof. This follows immediately from Lemma 11.16. 
84 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

11.3. The main structure theorem for ordinary double points.

Theorem 11.18. Say X → S is a flat finitely presented map. Let x ∈ Xs be
an ordinary double point with Xs of dimension at least 1 near x. Define n =
1 + dimx ( Xs ), and assume n is even if char s = 2. Then,
(1) Choose any residually non-degenerate quadratic form
Q ∈ OS,s
sh
[ t1 , . . . , t n ],
where Ash denotes the strict henselization of a local ring A. There exists
a ∈ mO sh so that there is an OS,s
sh -algebra isomorphism
S,s
 sh
OX,x
sh
' OS,s
sh
{t1 , . . . , tn } /( Q − a) = OS,s
sh
[ t1 , . . . , t n ] / ( Q − a ) .
(ms ,t1 ,...,tn )
By spreading out, this says there is a commutative diagram of pointed
maps
(11.3)
(X 0 , x0 )
f1
f2
( X, x ) a spread out version of Example 11.14 over OS0 ,s0

(S, s) (S0 , s0 )
f3

where the labeled maps f 1 , f 2 , f 3 are étale.

sh is intrinsic to the above setup. In particular, if some such a is
(2) Further aOS,s
not a zero divisor (so it generates an invertible ideal) then all such possible
a are a unit multiple of that one.
The proof is a hard but spectacular application of Artin approximation.
For (1) we may assume S is noetherian, and even finite type over Z, via
limit considerations. The argument in this case via Artin approximation is
given in [FK, §2, Ch III]. (Note that in [FK], they define “non-degenerate” to
require n even in characteristic 2.) That proof characterizes

aO
d sh
S,s

as an annihilator in O
d sh
S,s of a specific completed stalk of a module of relative
differentials for X over S; such a uniqueness for that ideal in the completion
implies uniqueness in general (by faithful flatness of completion for local
noetherian rings).
 MATH 249B NOTES: ALTERATIONS 85

Remark 11.19. In fact, the proof gives a little more: if k ( x ) = k (s) then one
can replace strict henselizations with henselizations, except one cannot con-
trol the (non-degenerate) reduction of Q over k (s). This issue is already seen
in the case of elliptic curves with non-split multiplicative reduction. The
actual choice of Q lifting such an uncontrolled reduction does not matter
because over OS,sh any two quadratic forms in n variables with isomorphic

non-degenerate reductions are in fact isomorphic, as is proved in a handout

on quadratic forms.
Having talked about ordinary double points in more generality, we’ll
now come back to discussing the relative dimension 1 case, which for us
is the main example of ordinary double points that we care about.
Let X → Spec A be a flat semistable curve over a reduced local noether-
ian ring A. We do not make any properness or geometric connectivity hy-
potheses. Let’s also assume the generic fibers are smooth. Let x ∈ X0 be a
non-smooth point in the special fiber X0 .
Proposition 11.20. Under the above assumptions, there exists a ∈ m A unique up
to A× -multiple so that ( X, x ) and
(Spec A[u, v]/(uv − a), 0 := (u, v, m A ))
have a common étale neighborhood. Further, a is not a zero divisor in A.
Remark 11.21. The key thing is that we can find a inside A, and not just in
a local-étale extension (as is the case in Theorem 11.18!).
Proof. First, we verify uniqueness of a, and then address its exists. By the
n = 2 case of Theorem 11.18, there exists a0 in the maximal ideal of a local-
étale extension A0 of A that does the job (using A0 [u, v]/(uv − a0 )). Note that
A0 is reduced, being an étale A-algebra. We want to show that this a0 is not a
zero-divisor in A0 , which will imply a0 is unique up to A0× -multiple by part
(2) of Theorem 11.18. It would then follow by faithful flatness of A0 over
A that any possible a ∈ m A is unique up to A× -multiple and is not a zero
×
divisor in A since necessarily a ∈ a0 A0 by the uniqueness over A0 . (We will
×
then have to find an A0 -multiple of a0 that comes from A, necessarily from
m A since m A0 = A0 ⊗ A m A due to A0 being local-étale over A).
To show a0 is not a zero-divisor (and hence the uniqueness of a0 up to unit
multiple in A0 ), it amounts to showing that a0 is nonzero each irreducible
component of the reduced noetherian scheme Spec A0 . Hence, passing to
such components with their reduced (and even integral) structure allows
us to assume that A = A0 is a domain.
Now, since A is a domain, we just seek a contradiction if a = 0. Let K be
the fraction field of A. We will use the vanishing of a to create a non-smooth
86 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

point on XK , contradicting our hypotheses on the generic fibers of X. With

a = 0, we have pointed étale maps

(Y, y)
h
(11.4)


( X, x ) Spec A[u, v]/(uv), 0 .

Thus, to show XK has a non-smooth point, it is enough to show that the

non-smooth point (0, 0) in the K-fiber of Spec A[u, v]/(uv) is in the image
of h. We know the image h(Y ) ⊂ Spec A[u, v]/(uv) is open by étaleness of
h, and it meets the 0-section (which is isomorphic to Spec A) in at least the
closed point 0 = h(y). But Spec A is local, so the open h(Y ) must contain
the entirety of that 0-section (as a local scheme has no proper open subset
containing the closed point). Thus, h(Y ) contains the generic point of the
0-section, which is the point (0, 0) in the K-fiber.
To summarize, we have shown that if such an a as desired exists, it must
not be a zero divisor and is unique up to multiplication by A× . To construct
a, consider the invertible ideal I 0 := a0 A0 ⊂ m A0 (invertible since a0 isn’t a
zero-divisor). Any generator of I 0 must be a unit multiple of I 0 , so we seek
a generator coming from A (necessarily then coming from m A because I 0
is a proper ideal of A0 ). It is enough to show I 0 = I ⊗ A A0 for an ideal
I ⊂ A. Indeed, if we could show this then I would be an invertible A-
module (since invertibility descends through faithfully flat ring extensions,
such as A → A0 ), and thus I = aA for some a ∈ A because A is local. This
would show aA0 = I 0 , so (as explained above) we’d be done.
To make the construction, inspired by the proof of Theorem 11.18, as
given in [FK, §2, Ch. III] we shall try
 
I := ann A Ω2X A /A,x

where ΩnB/A := ∧nB (Ω1B/A ) for any ring map A → B. By design of (Y, y) we
have a pointed étale map (Y, y) → ( X A0 , x 0 ) for some x 0 over x, so

I ⊗ A A0 = ann A0 ((Ω2X/A,x ) A0 ) by flatness of A → A0

= ann A0 (Ω2X 0 /A0 ,x0 ) for locality reasons
 A 
= ann A0 ΩY/A0 ,y
2
Y → X A0 étale

A pointed étale map h : (Y, y) → (Spec A0 [u, v]/(uv − a0 ), 0) exists by

design of Y, so likewise ann A0 (ΩY/A
2
0 ,y ) = ann A0 ( Ω
2
( A0 [u,v]/(uv− a0 ))/A0 ,0
). But
 MATH 249B NOTES: ALTERATIONS 87

this final annihilator is something we can compute! Indeed, direct calcula-

tion of Ω1( A0 [u,v]/(uv−a0 ))/A0 and passing to its second exterior power gives

Ω2( A0 [u,v]/(uv−a0 ))/A0 = ( A0 [u, v]/(uv − a0 , u, v))du ∧ dv,

and as a module over A0 [u, v]/(uv − a0 ) this collapses to A0 /( a0 ) that is al-
ready local (so localizing at 0 has become redundant) and visibly has A0 -
annihilator equal to ( a0 ) = I 0 ! 
As an application, let X → Spec R be an “open” (meaning no properness
or geometric connectivity assumptions) semistable curve over a dvr R with
uniformizer π, fraction field K, and residue field k. We assume XK is smooth
and x0 ∈ Xk is a non-smooth point. Recall from the general structure of
ordinary double point singularities that automatically k ( x0 )/k is separable.
(This is very important for later considerations!)
Since all nonzero elements of mR are R× -multiples of a unit π n with n > 0,
there exists a unique n x0 ≥ 1 such that ( X, x ) has an étale neighborhood in
common with
Spec R[u, v]/(uv − π nx0 ), 0 .

Passing to completions of local rings, we see that x0 ∈ reg( X ) if and only

if n x0 = 1 (so one may call n x0 a “measure of irregularity”). Our primary
interest is therefore in the case n x0 ≥ 2.
Proposition 11.22. Assume in the above setup that n x0 ≥ 2. Let
X 0 := BlX0 ( X ).
The R-scheme X 0 is a semistable curve over R and:
(1) if n X0 is either 2 or 3 then X 0 is regular over x0
(2) if n X0 ≥ 4 then X 0 has a unique non-regular point x00 over x0 and
n x 0 = n x0 − 2
0

and k ( x00 ) = k ( x0 ).
The proof will consist of three steps:
(1) A review of blow-up charts
(2) Transfer our problem to the model case (keeping careful track of
residue fields and points over x0 ).
(3) Doing blow-up calculations in the model case (which is also [Liu,
§8.3, Ex. 3.53]).
We shall carry out (1), refer to a short handout for the (technically impor-
tant but neither difficult nor interesting) details for (2), and then work out
(3) based on (1).
88 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Let A be a ring and I = ( f 1 , . . . , f n ) be a finitely generated ideal. Define

Z := Spec A/I → Spec A := X. The blow-up of X along Z, denoted
e := Bl I ( A) = BlZ ( X ),
X
is final among X-schemes Y → X with the property that the quasi-coherent
ideal sheaf
IOY ⊂ OY
is invertible.
Remark 11.23. We are really using the ideal sheaf IOY and not I ⊗ A OY .
That is, IOY is the image of I ⊗ A OY → OY (which may not be injective if Y
is not flat over X).
One standard definition/construction is
e = Proj (⊕m≥0 I m ) ⊂ Proj A[ T1 , . . . , Tn ] = Pn−1
X A
in which the graded algebra ⊕m≥0 I m is generated over A by f 1 , . . . , f n in
degree 1. This looks uncomputable away from very special situations, so
we want another description which yields more useful descriptions of
Xe ∩ D+ ( Ti )
for 1 ≤ i ≤ n.
The idea is to “cover” the moduli problem for the blow-up with sev-
eral refinements (corresponding to “open subfunctors”), using that if L
is invertible on a scheme W and generated by global sections s1 , . . . , sn ∈
Γ(W, L ), then
Wi := {w ∈ W : si is a basis of L near w }
is an open cover of W.
For any Y as above, on the open locus Yi where f i is a local basis it is
necessarily a global basis and so we can write f j = hij f i for a unique hij ∈
O (Yi ) if j 6= i. Then f i generates L and and in fact freely generates it over
Yi . Thus, if we consider the A-algebra
Ai := A Tij | j6=i /( f j − Tji ) /( f i∞ -torsion ) ⊂ A[1/ f i ]
 

generated by the f j / f i modulo “ f i∞ -torsion” (i.e., the directed union of the

f in -torsion for all n) then the relation immediately yield that Yi = Spec Ai is
final among A-schemes Y on which f i is a basis for IOY .
For i0 6= i, we have an identification of open subschemes
Yi ⊃ { Ti0 i 6= 0} ' { Tii0 6= 0} ⊂ Yi0
via the relation Tii0 = 1/Tii0 (and related transition formulas for Tji and Tji0
for j 6= i, i0 ) over the open sets of each on which f i , f i0 are both a basis for
 MATH 249B NOTES: ALTERATIONS 89

IOY . One can easily check this construction satisfying the triple overlap
condition to glue them to make Y = Bl I ( A).
The gluing itself respects each closed immersion Spec Ai → AnA−1 so glob-
alize to a closed subscheme
Y → PnA−1
making the commutative diagrams

Spec Ai AnA−1
(11.5)

Y PnA−1

be Cartesian (with AnA−1 as D+ ( Ti )).

Remark 11.24. If { f 1 , . . . , f n } is a regular sequence in A then the f i∞ -torsion
above vanishes [FL, Ch. IV, Thm. 2.2, Cor. 2.5]. That handles things such as
blow-up of a regular scheme along a regular closed subscheme, but we will
need go to beyond that setting.
Remark 11.25. The universal property of blow-up is not on all A-schemes!
It only applies to A-schemes Y with IOY invertible. An analogue to keep in
mind is normalization: normalization is final only with respect to dominant
maps from normal schemes. Consequently, the behavior under base change
can be delicate (since applying a base change may take us outside the rather
limited category of test objects for the universal property).
Consider A → A0 and define I 0 = I A and X 0 = Spec A0 . We have a
commutative diagram

f0
X X
e
(11.6)

X0 X

since IOX 0 = Ie0 pulls back to an invertible ideal sheaf on X

f0 , but the natural
map
A0 ⊗ A IOXe  IOXe
A0

may not be injective and in particular the target ideal sheaf might not be
invertible! Hence, the diagram above may not be Cartesian. This problem
does not arise if A → A0 is flat: the construction (especially the formation
of f i∞ -torsion) is compatible with such base change in an evident manner,
90 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

making the natural map X f0 → Xe A0 an isomorphism. This is compatible with

the Proj description of blow-ups, using that naturally

m
A0 ⊗ A I m = I A0

when A0 is A-flat.

This globalizes, allowing us to make the blow-up of any scheme X along

any finitely presented closed subscheme Z. By the universal property, the
formation of blow-up is Zariski-local on X (over affine opens in X, it is the
construction provided above). The natural map q : X e → X thereby restricts
to an isomorphism over U := X − Z because blow-up of the ideal sheaf (1)
on U is tautologically an isomorphism; in other words, Bl∅ (U ) = U.
On the other hand,

q−1 ( Z ) = Proj (⊕m≥0 I m ) ⊗ A A/I

= Proj ((⊕m≥0 I m ) ⊗ A A/I )
 
= Proj ⊕m≥0 I m /I m+1 .

This is called the exceptional divisor in X

e (it is Cartier by design of blow-
ups).

Example 11.26. Suppose X is finite type over a field k and x ∈ X is a closed

point, so the residue field k0 := κ ( x ) is k-finite. Let m ⊂ OX,x be the corre-
sponding maximal ideal. Let’s blow up X along { x }. Its exceptional divisor
fits into a fiber square

Proj ⊕m≥0 mm /mm+1

Blx ( X )
(11.7)

Spec k0 X.

If x 0 is a regular point (not necessarily a k-smooth point!) we have as k0 -

algebras

k0 [t1 , . . . , td ] ' ⊕m≥0 mm /mm+1

{ti } 7→ a k0 -basis of m/m2 ,

so Blx ( X ) has fiber Pκd(−x1) when x ∈ reg( X ).

 MATH 249B NOTES: ALTERATIONS 91

11.4. A detailed blow-up computation. This completes our review of gen-

eralities on blow-ups, so we now turn to the task of proving Proposition
11.22. A handout carries out the reduction to proving it in the following
“model case”: for a discrete valuation ring R with residue field k and uni-
formizer π, we consider
A := R [u, v] / (uv − π n )
with n ≥ 2. Thus, X := Spec A → Spec R is a flat semistable curve such that
X is normal and also is R-smooth away from the point ξ = 0 ∈ Xk at which
X is not regular because n ≥ 2. We want to compute Blξ ( X ) = Bl I ( A) for
I = (u, v, π ). and check that it satisfies the assertions in Proposition 11.22.
Note that Bl I ( A) is covered by
D+ ( u ) , D+ ( v ) , D+ ( π ) ,
the three affine opens where I is free on the elements u, v, and π respectively.
We want to compute these three affine schemes, their open overlaps, and
study their k-fibers over Xk (especially over ξ, to determine regularity at
such points on the total space of the blow-up).
11.4.1. D+ (u) computation. Let’s start by computing D+ (u). Introduce new
variables v0 and π 0 that are the universal multipliers against u:
v := v0 u
π 0 := π 0 u
The coordinate ring of D+ (u) is
( A v0 , π 0 /(v − v0 u, π − π 0 u))/(u∞ -torsion).
 

We have
A[v0 , π 0 ]/(v − v0 u, π − π 0 u) = R[u, v, v0 , π 0 ]/(uv − π n , v − v0 u, π − π 0 u)
and since π n = uv we also have
π 0 n u n = u2 v 0
and hence
 
u 2 π 0 n u n −2 − v 0 = 0
is a relation in this quotient. Thus, by killing u∞ -torsion we acquire the
relation v0 = π 0n un−2 , so the variable v0 can be eliminated everywhere via
that substitution.
Since v = v0 u = π 0n un−1 , we have
D+ (u) = Spec (( R[u, π 0 ]/(π − π 0 u))/(u∞ -torsion))
92 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

F IGURE 3. A depiction over Xk of the special fibers of the

charts D+ (u) and D+ (v) in the blow-up.

but Spec R[u, π 0 ]/(π − π 0 u) is already a (regular) domain, so there is no

more u∞ -torsion to kill. Hence,
D+ (u) = Spec R[u, π 0 ]/(π − π 0 u),
so D+ (u)k = Spec (k [u, π 0 ]/(uπ 0 )) → Xk = Spec (k [u, v]/(uv)) identifies
the u-axes on each but crushes the π 0 -axis into the origin ξ ∈ Xk .

11.4.2. D+ (v) computation. The affine open D+ (v) is essentially the same,
up to some change of letters:
e0 ]/(π − vπ
D+ (v) = Spec R[v, π e0 )
with a similar description of its k-fiber as Spec k[v, π e0 ]/(vπe 0 ) whose v-axis
maps isomorphically onto that of Xk and whose π e 0 -axis is crushed into the
origin ξ ∈ Xk . See Figure 3.
 MATH 249B NOTES: ALTERATIONS 93

Note that both D+ (u) and D+ (v) are regular, as can be seen by inspection,
but are not R-smooth: their special fibers are crossing axes.
11.4.3. Overlap of D+ (u) and D+ (v). Let’s compute the overlap of D+ (u)
and D+ (v) in the special fibers. Recall
D+ (u) = Spec R[u, π 0 ]/(uπ 0 − π ), D+ (v) = Spec R[v, π 0 ]/(vπe0 − π )
with v n= v0 u and v0 = π 0n un−2 in the first chart and with u = u0 v and
u0 = πe0 vn−2 . We have D+ (u)k ∩ D+ (v)k is the locus where v0 = π 0n un−2 is
n
invertible on D+ (u)k and u0 = πe0 vn−2 is invertible mod D+ (v)k . When n ≥
3 (so n − 2 > 0) it follows that these special fibers do not meet at all, whereas
if n = 2 they overlap exactly away from the complement of the origin on
the lines u = 0 and v = 0, in each via the relation π 0 πe0 = π 2 /(uv) = 1 on
the total space. Hence, the glued k-fiber contains a copy of P1k covered by
the π 0 -axis and π
e 0 -axis of each.
11.4.4. Computing D+ (π ). On this affine open chart we have u = u00
π, v =
v00 π, so π n = uv = u00 v00 π 2 . Since n ≥ 2, we have π 2 π n−2 − u00 v00 = 0.
Since π ∞ -torsion vanishes on D+ (π ), in the coordinate ring of D+ (π ) we
obtain the relation
u00 v00 − π n−2 .
After killing this relation, we conversely obtain the old relation π n = uv
from the relation u00 v00 = π n−2 and hence
D+ (π ) = Spec ( R[u00 , v00 ]/(u00 v00 − π n−2 )).
(This is easily seen to be R-flat, so it really is the correct affine open part of
the blow-up.) When n = 2 this gives u00 v00 = 1 on D+ (π ), so
D+ ( π ) k ⊂ D+ ( u ) k ∪ D+ ( v ) k .
Hence, when n = 2, the third chart is redundant on the special fiber and
the exceptional divisor is P1k with all points regular in the total space (as
D+ (u) and D+ (v) are regular). Consider the more interesting case n ≥ 3,
so D+ (π )k on the special fiber misses the two points ∞ on the lines u00k = 0
and v00k = 0. The total special fiber contains affine lines with coordinates
uk , u00k , v00k , vk , πk0 , πe0 k . Among these, the uk -axis meets the u00k -axis with k-
rational crossing point inside the regular D+ (u), the vk -axis meets the v00k -
axis with k-rational crossing point contained in the regular D+ (v), and the
u00k -axis meets the v00k -axis with k-rational crossing point contained in D+ (π )
whose unique non-smooth point in the special fiber is a k-point with mea-
sure of irregularity n − 2. See Figure 4 for a picture of the case n = 2 and
Figure 5 for a picture of the case n ≥ 3.
94 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

F IGURE 4. A depiction of the total special fiber when n = 2.

F IGURE 5. A depiction of D+ (π )k when n ≥ 3, followed by

a picture of the total special fiber (gluing of D+ (u)k , D+ (v)k ,
and D+ (π )k ) when n ≥ 3.

For n ≥ 3, we conclude that the exceptional divisor is the union of two

copies of P1k crossing at a single k-point whose measure of irregularity is
n − 2. We conclude that the blow-up is regular when n ≥ 3 and that if n ≥ 4
it has a unique non-regular point that is moreover k-rational with measure
of irregularity n − 2.
 MATH 249B NOTES: ALTERATIONS 95

12. R ESOLVING SEMISTABLE CURVES OVER A REGULAR BASE

For semistable curves over a general base, we shall now put a canonical
scheme structure on the non-smooth locus, given by the annihilator of Ω2 .
Over a dvr this will recover the reduced structure on isolated non-smooth
points in the special fiber.
Definition 12.1. An open semistable curve over S is a flat finitely presented
surjection f : X → S whose fibers Xs are semistable curves.
Define sm( X/S) to be the Zariski-open S-smooth locus on X (which is
fiberwise dense), and sing( X/S) to be X − sm( X/S). Note that even for
regular S, sing( X/S) may meet reg X.
Remark 12.2. When someone says “consider a singular point” on a scheme
that isn’t (locally) of finite type over a perfect field, you should always with-
out exception ask: do you mean “not regular” or “not smooth”?
Lemma 12.3. The annihilator ideal sheaf
annOX (Ω2X/S )

(quasi-coherent since Ω2X/S is finitely presented) is finitely generated with under-

lying zero-scheme sing( X/S). Moreover:
(1) The formation of OX / annOX (Ω2X/S ) commutes with any base change on
S.
(2) Via this scheme structure, the finitely presented map sing( X/S) → S
is quasi-finite with étale fibers (equivalently, “unramified” in the sense of
Grothendieck).
Remark 12.4. The unramifiedness condition is equivalent to the vanishing
of Ω1sing(X/S)/S . The reason for this is that the vanishing of this finitely
presented quasi-coherent sheaf is suffices to check on fibers over S (due
to Nakayama’s Lemma), and for X of finite type over a field k we have
Ω1sing(X/k)/k = 0 if and only if X is k-étale (as may be checked over k).

Proof. To show
annOX (Ω2X/S )
is finitely generated and satisfies (1), the problem is étale-local on X and S.
Thus, it is enough to consider the model case.
That is, it suffices to consider
X = Spec B := Spec ( A[u, v]/ (uv − a))
96 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

with S = Spec A and a ∈ A. Then,

 
Ω2B/A = ∧2B Ω1B/A
B du ⊕ B dv
 
2
= ∧B
(u dv + v du)
= ( B/(u, v)) du ∧ dv,
so annB (Ω2B/A ) = (u, v) ⊂ B (visibly finitely generated). Thus,

B/ annOX (Ω2X/S ) = A/( a)

as A-algebras. It is clear that the formation of this quotient of B is compati-
ble with any scalar extension on A, as desired.
To verify the support of OX / annOX (Ω2X/S ) inside X is as claimed, and
to verify (2), it now suffices to check on fibers. That is, we can assume
S = Spec k for a field k. Further, we can assume k = k, via flat base change.
On X sm the sheaf Ω1 is invertible, so Ω2 = 0. Therefore
(annOX Ω2X/S )| Xsm = OXsm ,
so it only remains to verify what is happening at the non-smooth points.
That is, we need to show (annOX Ω2X/S ) x = mx for x ∈ sing( X/k ). Since k is
algebraically closed, so k( x ) = k, we can work étale-locally to reduce to the
model case
( X, x ) = ((uv ⊂ A2k ), (0, 0))
that is easily handled by a direct calculation. 
We return to the original situation. Recall we have a proper semistable
curve
X
(12.1) f

D Y
where D is an sncd in a smooth projective variety Y of dimension d − 1 ≥ 1
and X is a proper semi-stable curve over Y that is smooth over U := Y − D
with geometrically connected fibers.
Modulo the desired sncd condition for Z = f −1 ( D ) ∪ (∪σi (Y )) for some
sections σi ∈ X (Y ), we need to resolve singularities of X as a k-scheme. It is
this latter aspect that will be our primary focus; handling the task of making
Z an sncd will come at the end, so we set Z aside for now.
 MATH 249B NOTES: ALTERATIONS 97

If d = 2 then Y is a smooth curve and so sing( X/Y ) maps quasi-finitely

to some finite set of points y1 , . . . , yn ⊂ Y (k ). For Ri := OY,yi a dvr, when
we localize at yi we obtain the situation

X Ri
(12.2)

Spec Ri

that is exactly our earlier focus (semistable curves with smooth generic fiber
over discrete valuation rings).
Thus, by repeatedly blowing up points in X that are not k-smooth (hence
not Y-smooth, so there are only finitely many of these), the process pre-
serves semistability over Y and improves the measure of irregularity with-
out ever increasing the number of non-regular points. Therefore, this pro-
cess eventually terminates at a curve semistable over Y with measure of
irregularity 1 at each of the finitely many points not smooth over Y, so it
resolves singularities on the k-surface X.
Now suppose instead d ≥ 3. In this case the proper map
sing( X/Y ) → D
is quasi-finite and hence finite to the sncd D in the k-smooth Y, and D has
pure dimension d − 2. Thus, each irreducible component T of sing( X/Y )
has codimension at least 2 in X. Further, by dimension reasons, the codi-
mension of such a T in X equals 2 if and only if T surjects onto some irre-
ducible component Di ⊂ D (which we give the reduced structure, so Di is
k-smooth since D is an sncd in the k-smooth Y).
Our main focus is when such T of codimension 2 is in the non-regular
locus of X (i.e., sing( X/k ) has codimension 2 rather than ≥ 3). Given such
a T, we want to “improve” its non-regularity via a blow-up, and repeat
the process until we manage to reach a situation in which sing( X/k ) has
codimension ≥ 3 in X.
Consider an irreducible component T of sing( X/k ) with codimension 2
in X (recall that X is normal, being a semistable curve with smooth generic
fiber over a connected normal base, so sing( X/k ) has codimension ≥ 2.
Thus, T is also an irreducible component of sing( X/Y ) with codimension
2 in X. Give T the reduced structure. We have a finite (since proper quasi-
finite) surjection T  Di for some i, so the resulting map between generic
points ηT → ηDi is étale because R = OY,ηD is a dvr for which ηT is a non-
i
smooth point in the special fiber of XR . The aims in this situation are:
(1) to show T → Di is étale,
98 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

(2) to define a measure of irregularity along the entirety of T inside X,

enabling us to carry out local calculation along T (not only near ηT ).
If ηT ∈/ Reg( X ) then we want to blow up X along T, but we need to
verify that BlT ( X ) it remains semistable over Y (and that in a suitable sense
this blow-up improves the non-regularity). Once we push sing( X/k) out
of codimension 2 into codimension 3, the possibilities for the structure for
sing( X/k ) (such as the dimensions of its irreducible components and their
overlaps) will turn out to be rather nice.

As in [deJ, 3.1], we now make the following extended setup for our cur-
rent situation. We shall first proceed to resolve singularities in codimension
2 (in Theorem 12.5) after possibly passing to a completed local ring on the
base, so we’ll describe a setup that encompasses this.
Let S be a connected regular excellent scheme and f : X → S a proper
semistable curve with D ⊂ S sncd such that f is smooth over S − D with
geometrically connected fibers. (Recall that by Stein factorization, the geo-
metric connectivity of all fibers follows from smoothness and geometric
connectivity of the generic fiber since the base S is normal.) We know that
X is normal so that X sing := X − Reg( X ) has codimension at least 2. Fur-
ther, write D = ∪i Di with Di the reduced irreducible components of D (all
regular, since D is an sncd in S). Each Di ⊂ S is Cartier and for every subset
J ⊂ I of the index set I, the intersection
\
D J := Dj ⊂ S
j∈ J

is regular with pure codimension #J. Thus, for all points d ∈ D j , generators
of the invertible ideals IDj ,d ⊂ OS,s constitute part of a regular sequence of
parameters for OS,d .
Note that since S is regular,

sing( X ) ⊂ sing( X/S)

:= X − sm( X/S)
⊂ f −1 ( D ).

For irreducible components T of sing( X ), we want to get to the situation

where codimX ( T ) ≥ 3 for all T. Note that sing( X/S) ⊂ X has codimension
at least 2 everywhere since XηS is smooth and all Xs are generically smooth.
For irreducible components T ⊂ sing( X ) of codimension 2 in X, we want
to show BlT ( X ) “improves things.”
 MATH 249B NOTES: ALTERATIONS 99

Theorem 12.5. There exists a modification φ : X1 → X (to be constructed as a

composite of several blow-up maps) that is an isomorphism over

reg X ⊃ sm( X/S) ⊃ f −1 (S − D ),

such that
(1) X1 → S is a semistable curve
(2) codimX (sing( X1 )) ≥ 3.

Here is the method of proof: Without loss of generality, there exists an ir-
reducible component T ⊂ X sing with codimension 2 (or else there is nothing
to do), and for X 0 := BlT ( X ) → X we’ll show
(i) X 0 → S is a semistable curve
(ii) in an appropriate sense, sing( X 0 ) are “less severe” than sing( X ) by
using a “measure of irregularity along an entire irreducible compo-
nent of sing( X ) with codimension 2.
Once we make (ii ) precise and we prove it, we can iterate the process to
reach the desired X1 in finitely many steps. A source of some further com-
plications beyond our calculations over dvr’s is that we are blowing-up
along the closed subscheme T that is no longer a point.
To prove actually prove Theorem 12.5, we will need to establish several
ingredients. First, we need to describe ( X, T ) étale-locally without losing
contact with irreducibility of T. This rests on the following subsidiary result:

Proposition 12.6. The codimension-2 component T maps finitely onto some DiT ,
and T → DiT is étale. (Recall that ηT 7→ ηDi is étale, with the generic point ηT of
T
T being in codimension 2.)

The integral T maps onto an irreducible component DiT of D. Indeed, the

inclusion T ⊂ sing( X ) implies that for all t ∈ T, t ∈ X f (t) is not smooth due
to the regularity of S. Since the fibers Xs are generically smooth, T → D
must be quasi-finite (so finite by properness) and then onto some DiT (via a
finite surjection) for dimension reasons.
In particular, the generic point ηT of T maps to the generic point ηiT of
DiT . Since ηiT has codimension-1 in S, the point ηT is a non-smooth point in
the special fiber of XR → Spec R for the dvr R = OS,ηi . Thus, ηT 7→ ηiT is
T
étale, so T → DiT is étale when restricted to dense opens on the source and
100 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

target. We want to show T → DiT is everywhere étale. We have

T sing( X/S)
(12.3)

Di T S

with sing( X/S) → S having étale fiber schemes. Thus, the map T → DiT
to a regular (hence normal) scheme is a finite surjection between integral
noetherian schemes with étale fibers. Now apply the following lemma:
Lemma 12.7. Say B → C is a finite injection of noetherian domains with B normal
and Ω1C/B = 0(equivalent to SpecC → SpecB having étale fibers, by Nakayama’s
Lemma).
Then, C is B-étale (so B-flat).
Proof. The key point is that a map of finite type between noetherian schemes
with vanishing Ω1 is Zariski-locally a closed immersion into a scheme étale
over the base. Thus, Zariski-locally Spec C is a closed subscheme of an étale
B-scheme E. By normality, the connected components of E are irreducible.
By dimension reasons, the part of the dominant B-finite Spec C meeting
any component of E must exhaust that component, so Spec B is C-étale. For
details, see [FK, Ch. I, Lemma 1.5]. 
This completes the proof of Proposition 12.6. Next, we describe some
completions. Choose x ∈ sing( X/S) over s = f ( x ) ∈ D (e.g., any point in
T). Let Di1 , . . . , Dim be the irreducible components of D through s. We want
to describe O cx as an Obs -algebra using the generators t j of ID ⊂ OS,s .
ij

Warning 12.8. Keep in mind that κ ( x ) is finite separable over κ (s) (and usu-
ally this extension is non-trivial). We’ll need to make a maneuver to pass to
the case κ ( x ) = κ (s).
We now reduce the task of describing such a completion to the case κ ( x ) =
κ (s). Let R be the unique local finite étale Obs -algebra with residue field
κ ( x )/κ (s). Explicitly,we pick a primitive element
κ ( x ) = κ (s)[U ]/( f )
with f monic, so for a monic lift F ∈ Obs [U ] the quotient ring
R = Obs [U ]/( F )
is finite flat over Obs with R/m0 R = κ ( x ) a field, so R is local and visibly
finite étale over Obx .
 MATH 249B NOTES: ALTERATIONS 101

By the henselian property for Obx , there is a unique map of Obs -algebras

Obs R
(12.4)
t
Obx
lifting the equality of residue fields. We have a fiber diagram
Y0 Y X

(12.5) Spec R Spec Obs S

Ds0 D
bs D
Note that
Y 0 3 Spec κ R ×S x
 
= Spec κ ( x ) ⊗κ (s) κ ( x )

We can write Spec (κ ( x ) ⊗κ (s) κ ( x )) as a union of ∆( x ) with some other

scheme, and we define y0 := ∆( x ).
Exercise 12.9. Verify that as R-algebras,
ObY 0 ,y0 ' ObX,x .
The upshot is that we can pass to (Y 0 , y0 ) over Spec R to focus on the case
κ ( x ) = κ (s). In order to continue our proof of Proposition 12.6, we require
the following lemma:
Lemma 12.10. Suppose S = SpecA for a regular complete local noetherian ring
A and s ∈ S the closed point. Let x ∈ Xs be non-smooth with κ ( x ) = κ (s). Then
there is an isomorphism of A-algebras
O
cx ' AJu, vK/( Q − a)
for a residually non-degenerate quadratic form Q and a ∈ m A − {0}, and a =
nj
∏ t j for local generators t j of the ideal sheaves
IDj ,s ⊂ A
/ Reg( X ) if and only if ∑ j n j ≥ 2.
of the D j passing through s and n j ≥ 0. Also, x ∈
102 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

Remark 12.11. We can always change a by an A× -multiple by absorbing a

unit into the quadratic form.
To prove Lemma 12.10, note that the refined structure theorem for ordi-
nary double points (Theorem 11.18 and Remark 11.19) gives
Obx ' AJu, vK/( Q − a)
for some Q and a ∈ m A − {0} unique up to A× -multiple.
nj
We shall show Spec A/a ⊂ D, which would imply a ∈ A× · ∏ t j for
some n j ≥ 0 since A is regular local, hence a UFD. The main input in the
proof will be Artin approximation and the smoothness over S − D:
Lemma 12.12. We have |Spec( A/a)| ⊂ D.
Proof. It is enough to show that f is not smooth over any points in S0 :=
Spec A/a ⊂ S. That is, we claim Xs0 is not smooth for all s0 ∈ S0 .
Using Artin approximation, there exists a common residually trivial étale
neighborhood
(12.6)
(X 0 , x0 )
q2
q1
( X, x ) (Y = Spec A[u, v]/( Q − a), y = (m A , u, v))
f

(S, s) S0 = Spec A/a

with q1 , q2 both étale. It is enough to show that X 0 has nonsmooth fiber over
all s0 ∈ S0 , since then the q1 -images of such points do the job for f : X → S.
Therefore, it is enough to show that the open image q2 ( X 0 ) ⊂ Y contains the
entire 0-section σ : S0 → Y since the origin u = v = 0 in Y over points in
the base where a = 0 are non-smooth points in their fibers over S0 (as q2 is
étale).
The overlap q2 ( X 0 ) ∩ σ (S0 ) is open in the local scheme σ(S0 ) and contains
the closed point y = q2 ( x 0 ). The only open subscheme of a local scheme
that contains the closed point is the entire space, so q2 ( X 0 ) must contain all
of σ(S0 ) as desired. 
This completes the proof of Lemma 12.10.
Corollary 12.13. Prior to completing the base S, if x ∈ sing( X/S) and s = f ( x )
then we obtain intrinsic integers ni ( x ) ≥ 0 attached to all Di 3 s.
 MATH 249B NOTES: ALTERATIONS 103

Proof. For any point, we have a description of the completed local ring. We
n
get the element a = ∏im=1 ti i , unique up to unit, via Lemma 12.10. This
provides intrinsic multiplicities ni ( x ) ≥ 0 for each Di 3 s. 
Now it is time to analyze niT ( x ) in Lemma 12.10 for varying x ∈ T when
T is an irreducible component of sing( X/S) of codimension 2 in X. (For
example, we could take T to be an irreducible component of X sing of codi-
mension 2 in X.)
The étaleness of T over the regular scheme DiT implies that T is regular,
so for x ∈ T the local ring O
d T,x is regular, hence a domain.
The next lemma Lemma 12.14 defines a measure of irregularity n T ≥ 1.
Lemma 12.14. For all x ∈ T, the exponent niT ( x ) is independent of x.
Proof. Let s = f ( x ) and define Obs0 to be the a local finite étale Obs -algebra
with residue field κ ( x ). So,
Obx ' Obs0 Ju, vK/( Q x (u, v) − ∏ ti i )
n (x)

as Obs0 -algebras, where ti are local generators of IDi ⊂ OS near s for those
Di 3 s and Q x is a residually non-degenerate quadratic form (so disc( Q x ) ∈
(Obs0 )× , a property retained under any Zariski-localization on the base). By
using Artin approximation, we may use any Q x with a specific reduction
over κ ( x ). Hence, for a suitable étale neighborhood
(Spec A, ξ ) → (S, s)
we have
(X 0 , x0 )
q2
q1
(12.7) n
( X, x ) Spec A[u, v]/( Q − ∏i ti i , (ξ, (0, 0)))

(S, s)
with q1 and q2 étale, x 0 7→ x ∈ T ⊂ X, and T  DiT . Let’s localize S at ηiT .
The open q1 ( X 0 ) meets T at x, so it contains all generizations of x in T
(such as ηT ). Therefore, there is some y0 ∈ X 0 with q1 (y0 ) = ηT . Look at the
image q2 (y0 ). After localizing at ηiT , all t j for j 6= i T become units. Therefore,
the localization of the model case at q2 (y0 ) is a Zariski-localization of
ni T
 
R[u, v]/ Q − tiT
104 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

with R a dvr with uniformizer tiT (explicitly, R is the localization of A at

q2 (y0 )) and disc( Q) ∈ R× . Hence, the intrinsic measure of irregularity at ηT
that has nothing to do with the specific choice of x ∈ T is equal to niT ( x ). 
12.1. Proving Theorem 12.5. Our goal is now to complete the proof of The-
orem 12.5. Again, recall the setup: we have
T X
(12.8) f

Di T Y

with DiT an irreducible component of the sncd D ⊂ Y for smooth projective

Y over k = k, f semistable proper with geometrically connected fibers and
smooth over Y − D, and
T ⊂ X sing ⊂ sing( X/Y )
an irreducible component of X sing with codimension 2 in X (so T is also an
irreducible component of sing( X/Y ), the situation considered above).
Since ηT ∈
/ Reg( X ) we have n T ≥ 2 (recall n T = niT ( x ) for all x ∈ T, by
Lemma 12.14). We have
nj
X,x ' (O f ( x ) ) Ju, vK/ ( Q − ∏ t j ( x ))
0
Od b
j

where (Obf ( x) )0 is the unique local finite étale O [ f ( x ) algebra with residue field
κ ( x ) and
(1) all D j ’s are irreducible components of D through f ( x ), with
t j = IDj , f ( x) ⊂ OY, f ( x)

(2) n j ( x ) ≥ 0 (so niT ( x ) = n T ≥ 2).

Our aim is to show:
Goal 12.15. The blow-up X e := BlT ( X ) is a proper semistable Y-curve with
geometrically connected fibers, smooth over Y − D, with X e sing “simpler”
than X sing with respect to the irreducible components of codimension 2:
either the number of codimension-2 components has gone down, or there is
a unique one over T, call it T 0 , and n T 0 = n T − 2.
Remark 12.16. Certainly Xe → Y is proper and over Y − D is smooth with
geometrically connected fibers of dimension 1 since T ⊂ f −1 ( D ), so
e |Y − D ' X |Y − D .
X
 MATH 249B NOTES: ALTERATIONS 105

e → Y it is
Therefore, to verify the desired semistability properties for X
enough to check that X e → Y is open semistable, which is an étale-local
problem on X and Y (as we have seen in earlier discussions via Stein factor-
e
ization).
Provided the blow-up is Y-semistable, which we have noted is an étale-
local problem, the following argument establishes Theorem 12.5. Grant-
ing this Y-semistability, it follows from our work over with blowing up
semistable curves with smooth generic fiber over discrete valuation rings
that the codimension-2 irreducible components T e sing correspond bi-
e of X
jectively via the projection to codimension-2 irreducible components T1 of
X sing with the exception that we omit T if n T is either 2 or 3.
The point is that we can detect any such T e looking over the generic points
of D. Further, only over ηDi has anything changed, and there we under-
T
stand how it has changed by localizing the base Y at that generic point, to
arrive at the situation over a dvr. Then, over that point ηDi there is only
T
one non-regular point in the blow-up over the given non-regular point on
X, and its measure of irregularity drops by 2 by Proposition 11.22. If n T > 3
then there is a unique T e1 over T and it satisfies n e = n T − 2, whereas if
T1
n T = 2, 3 then there is no such codimension-2 component mapping onto
T. Hence, there are at most as many codimension-2 irreducible components
of sing( X
e ), and the total irregularity (the sum over all ni for T irreducible
T
components) drops in the blow-up.
12.1.1. Justifying semistability of the blowup. It remains to show the blow-up
Xe is open-semistable over Y. This will be done in Proposition 12.17, in
combination with a handout that works out calculations in a “model case”.
Our task is étale local on X around each x ∈ T. Recall the formation of
Xe = BlT ( X ) commutes with flat base change on X, such as passage to an
étale neighborhood of ( X, x ).
As a first step, it is harmless to pass to an étale neighborhood
Y 0 , y0 → (Y, f ( x )) .

Indeed,
T ×Y Y 0 → T
is étale, so the connected components of T ×Y Y 0 are its irreducible com-
ponents because regularity of T (see Proposition 12.6) implies T ×Y Y 0 is
regular. Checking local properties of the blow-up along a disjoint union of
closed subschemes amounts to checking the blow-up along each component
separately. (Beware that we do not claim that blow-up along a reducible
106 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

closed subscheme whose irreducible components meet can be achieved by

successive blow-up of such components individually.)
We can thereby increase the residue field at f ( x ) to make Q residually
split and so may arrange that Q = uv. Thus, now by Artin approximation
we have a common étale neighborhood

(W, w)
q2
q1

nj
 
(12.9) T ( X, x ) Spec B = R[u, v]/(uv − ∏ j t j ), ξ

Di T (Y, y) (Spec R, y)
e := q−1 ( T ) (which
with q1 , q2 étale, y = f ( x ), and ξ = (y, 0, 0). We define T 1
might be disconnected).
In the model case of Spec B, observe that
Y := V (t1 , u, v) = Spec R/(t1 ) ⊂ Spec B
is a codimension-2 irreducible component of Spec Bsing passing through ξ.
(The irregularity is clear since we can check it at the generic point, using
that n1 = n T ≥ 2 to ensure ηT ∈/ reg X.)
Proposition 12.17. In order to show BlT ( X ) over x is semistable, it suffices to
show that BlY (Spec( B)) over ξ.
Proof. This proof proceeds along standard lines.
(1) First, we reduce to the case T e is irreducible As we have initially con-
structed T,e it is possible that it may be reducible. But because T e→T
is étale with w 7→ x, and T is regular, it follows that T
e is regular with
a unique connected component through w and that in turn must be
irreducible by regularity. Hence, as far as studying what is happen-
ing in the blow up around x, we may as well shrink around w so that
only the component passing through w appears. In other words, we
may shrink W around w for the Zariski topology to make T e irre-
ducible.
(2) It is enough to show that BlTe (W ) is Y-semistable at points over w.
Indeed, by flat base change, BlTe (W ) = BlT ( X ) × X W with w 7→ x,
and W → X is étale.
(3) We can also shrink W further if necessary so that q2−1 (ξ ) = w. Then
we have:
 MATH 249B NOTES: ALTERATIONS 107

Lemma 12.18. Y = V (t1 , u, v) ⊂ Spec( B) is the unique codimension-2

irreducible component of sing(Spec( B)) through ξ.
Proof. Since q2 : (W, w) → (Spec B, ξ ) is étale,
q2−1 (sing(Spec B)) = W sing .
Hence, the uniqueness of Te implies the uniqueness of Y through ξ
because we arranged that q2−1 (ξ ) = {w}. 
− 1 − 1
e = q (Y ) since q (Y ) is the union
The preceding result implies T 2 2
of the components of the non-regular locus through q2−1 (ξ ) = {w}
with codimension 2 in W, and we arranged that this is T.
e
(4) We conclude that
BlTe (W ) = BlY (Spec B) ×Spec B W,
so to show BlTe (W ) is semistable near points over w it suffices to
prove that
nj
Bl(u,v,t1 ) ( R[u, v]/(uv − ∏ t j )
j

is R-semistable for a regular local ring R (e.g., R = OY,y in our case)


and t j ⊂ mR part of a regular system of parameters with n1 ≥ 2.
This is a hands-on calculation generalizing what we did over dvr’s,
and is done in a handout.

Now, by Theorem 12.5,
 
codim X sing , X ≥ 3.

Lemma 12.19. For all x ∈ X sing , we have ni ( x ) ≤ 1 for all i.

Proof. Suppose some ni0 ( x ) ≥ 2, so O
cx has a local finite étale extension that
is the completion of an étale neighborhood of
(W := Spec ( R[u, v]/(uv − ∏ ti i
n (x)
)), w = (u, v, mR ))
i

with R local finite étale over O

[ f ( x ) . Note that W is normal since it is an open-
semistable curve over R with smooth generic fiber, so the non-regular locus
of W (which is closed by excellence) is supported entirely in codimension
≥ 2. The integral closed subset

E := u = v = ti0 = 0 ' Spec ( R/(ti0 ))
108 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

has codimension 2 in W, and W is not regular at the generic point of E since

the completion of E at that point is easily seen to be the visibly non-regular
ni ( x )
local noetherian ring R0 [[u, v]]/(uv − ti0 0 ) where R0 is the completion of
the regular local ring ( Rp )∧ for p = (u, v, ti0 ) ⊂ R (note that t j ∈ Rp× for
sing
j 6= i0 ). Thus, E is an irreducible component of OW,w ; i.e., the non-regular
locus of W near w has an irreducible component of codimension 2.
The faithfully flat map from any excellent local noetherian ring to its max-
adic completion is a regular map (i.e., flat with each fiber algebra regular and
remaining so after any finite extension of the ground field of the fiber alge-
bra). Hence, the non-regular locus of ObX,x is the full preimage of the non-
regular locus of that of OX,x , and likewise for any local finite étale extension
of ObX,x . The same considerations apply to OW,w and its completion, so we
conclude that the non-regular locus of OX,x has codimension 2 (even if some
of its irreducible components become reducible after completion) since the
dimension of a local noetherian ring is the same as that of its completion.
We have contradicted that
 
codim X sing , X ≥ 3.

12.2. Setup for handling when codim( X sing , X ) ≥ 3. We now have the
following setup

x ∈ X sing ⊂ sing( X/Y ) X

(12.10) f

y := f ( x ) ∈ ∪in=1 Di = D Y

where D ⊂ Y is an sncd, Y is connected excellent regular, and f : X → Y is

an open semistable curve smooth over Y − D. We want to describe the local
structure of X sing near x, so we may shrink Y to ensure y ∈ Di for all i.
Let A = O d 0
Y,y and A the a local finite étale A-algebra with residue field
κ ( x ). Further,
IDi ,y = ti OY,y ,

with {ti } part of a regular system of parameters of OY,y . As we have used

multiple times already, Od 0
X,x is uniquely an A -algebra over its A-algebra
structure compatibly with its residue field identification with that of A0 , and
 MATH 249B NOTES: ALTERATIONS 109

as A0 -algebras
µ
X,x ' A Ju, vK/ ( Q − ∏ ti )
0
Od
i =1

for some residually non-degenerate quadratic form Q ∈ A0 [u, v] and 1 ≤

µ ≤ n. In fact, 2 ≤ µ ≤ n since Od 0
X,x is not regular (whereas A [[ u, v ]] is
regular) due to the hypothesis x ∈ X sing .
Proposition 12.20. (1) The closed set X sing is covered by the closed sets (given
the reduced structure)
Eij := X sing ∩ f −1 ( Di ∩ D j )
for i < j, and each Eij is regular (hence its irreducible components are its
connected components) with codimension 3 in X. Moreover, the finite map
Eij → Di ∩ D j
is étale. In particular, all irreducible components E of X sing with reduced
structure are regular.
(2) If E, E0 are distinct irreducible components of X sing (taken with the reduced
structure and E ∩ E0 6= ∅ then E ∩ E0 is regular with codimension 4 or 5
in X for each of its irreducible (or equivalently connected) components.
Proof. The idea is to use excellence to pass to the completion, then pass to
an étale model where we can perform an explicit computation.
Note that passing to completion commutes the formation of nilradicals
since these schemes are excellent, so the completion at a point for a reduced
structure on a closed subscheme is the reduced structure on the completion
of the closed subscheme at that point. We will implicitly use this repeatedly
in the following.
For the first part, we pick x ∈ X sing and note that since Spec O cx →
Spec Ox is regular, the non-regular locus of Spec O cx is the full preimage of
the non-regular locus of Spec Ox . Further, we can pass to a local finite étale
extension of the local finite étale algebra A0 over A := Od X,y with residue
field κ ( x )) so that are usual description of completions of étale neighbor-
hoods of ( X, x ) has Q become uv residually, so Q is isometric to uv (as we
are working over a complete base ring). This reduces us to studying the
model case Spec B with
µ
B := R[u, v]/(uv − ∏ ti )
i =1
110 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

with R a complete regular local ring, {ti } a subset of a regular system of

parameters in R (with ti R the ideal for Di ⊂ Spec R), and x = (u, v, mR ).
For the second part, we shall analyze Eij ∩ Ei0 j0 in Spec B
cx for (i, j) 6=
0 0
(i , j ). We will show the intersection is codimension 4 if i or j belong to
{i0 , j0 } and is codimension 5 if i, j ∈
/ { i 0 , j 0 }.
By excellence, the subset
µ
Spec Bsing ⊂ sing( B/R) = {u = v = 0, ∏ ti = 0}
i =1

is closed in Spec B, and the given description of the non-smooth locus for
µ
Spec B over R is correct because it is exactly the zero-section over ∪i=1 Di .
For points w of Spec Bsing , we claim that at least two ti ’s vanish at s =
µ
f (w) ∈ Spec R. At least one of them must vanish (as ∏i=1 ti vanishes at
w), and suppose some ti0 (s) vanishes but t j (s) 6= 0 for all j 6= i0 . We seek a
contradiction.
By design

u = v = t i0 = 0
is non-regular at the point w away from f −1 ( D j ) for all j 6= i0 . The quotient
B/(u, v, ti0 ) = R/(ti0 ) is regular and so remains regular after inverting all t j
for j 6= i0 . After inverting the t j ’s for j 6= i0 , killing ti0 is the same as killing
µ
∏i=1 ti . This forces regularity at w, a contradiction.
The converse also holds. That is, we claim that
Spec R/(ti , t j ) = u = v = ti = t j = 0 ⊂ Spec Bsing


for any i < j. This quotient of R is visibly local regular, so its spectrum is
irreducible, and the singular locus in Spec B is closed (by excellence), so it
is enough to check the completion of the total space of Spec B at the generic
point (u, v, ti , t j ) of the closed subset {u = v = ti = t j = 0} of Spec B is
non-regular. This completion is

(u,v,ti ,t j ) = R(ti ,t j ) Ju, vK/ ( uv − ti t j ),

B\ \

which is not regular since uv − ti t j is in the square of the maximal ideal.

We have described the non-regular locus of Spec B as the union of the reg-
ular irreducible closed subsets Zij := {u = v = ti = t j = 0} for i < j, and
so it is clear that Spec Bsing ∩ f −1 ( Di ∩ D j ) = Zij for any i < j. From this ex-
plicit description, we obtain the desired codimension assertions. Moreover,
( Zij )∧x = Spec ( R/(ti , t j )) → ( Di ∩ D j )∧x
 MATH 249B NOTES: ALTERATIONS 111

is an isomorphism by inspection, so in the global setting (before passing to

completions and increasing residue fields by finite separable extensions via
suitable local-étale extensions) we see that the finite map Eij → Di ∩ D j is
étale. 

12.3. Returning to the examination of Z. In the original setup over k = k,

sing


to achieve codim X , X ≥ 3 we applied a composite of blowups
φ
X1 X
(12.11) f1
f
Y
over the complement of the k-smooth locus X sm of X (even over the com-
plement of the Y-smooth locus of X) at each step. How does this interact
with Z? Recall that
Z = f −1 ( D ) ∪ (∪i τi (Y ))
for pairwise disjoint τi : Y → sm( X/Y ), so each τi lands inside an open
subscheme of X away from the blow-up locus. Therefore, the τi give rise to
sections τi0 : Y → sm( X1 /Y ) such that
φ−1 ( Z ) = f 1−1 ( D ) ∪ ∪i τi0 (Y ) =: Z1 .

Thus, even accounting for Z, we can assume codim X sing , X ≥ 3. In par-

ticular, for d = 2 we have reached the case of k-smooth X but we still need
to do more work even there (to arrange that Z is an sncd in X, possibly after
some further modification).
To wrap up the case d = 2 (which involves content via the sncd property
we wish to achieve for Z), and more generally to get rid of the consideration
of the pair (Y, D ) that has largely done everything we need from it, we shall
prove the following:
Lemma 12.21. The proper reduced closed subset
Z ∩ X sm ⊂ X sm
is an ncd.
Proof. The property of being ncd is étale-local on X sm near each z ∈ Z (k ).
We shall do some mild case-analysis. First consider the situation away from
f −1 ( D ); i.e., over Y − D. Over this Y-smooth open subset of X, Z is a disjoint
union of sections to a smooth curve over Y, and such a section is always
sncd. That is, near any z ∈ Z ∩ f −1 (Y − D ) in the section τi (Y ) for a unique
112 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

i there is a Zariski-open neighborhood U in f −1 (Y − D ) admitting an étale

map h as in the diagram

h
U A1V
(12.12) f

for open V := f (U ) ⊂ Y such that τi restricts to a section τi |V : V → U

whose composition with h is the zero-section 0V of the affine line over V.
Since h is étale, it follows that h−1 (0V ) is étale over the regular 0V = V
and hence τi (V ) is a connected component of h−1 (0V ). Thus, the global
coordinate on A1V pulls back to a local function on U near z that cuts out Z
near z.
Next, we consider the case z ∈ f −1 ( D ), which is to say f (z) ∈ D. This
breaks into two sub-cases, depending on whether or not z ∈ sm( X/Y ).
Suppose z does belong to the Y-smooth locus in X, so if z belongs to some
τi (Y ) (with i then unique) the data ( X, Z, z) is étale over

(AY1 , 0Y ∪ A1D = f −1 ( D ), 0 f (z) )

with D ⊂ Y an sncd, so the ncd property is clear. If such z doesn’t belong

to any τi (Y ), so Z near z coincides with f −1 ( D ) then the proof of the ncd
property is essentially the same but easier (as we don’t need to pay attention
to the zero-section of the affine line over Y).
There remains the subcase z ∈ X sm with z not smooth in the fiber X f (z) .
Hence, z ∈ / τi (Y ) for all i (as the τi are supported in sm( X/Y ))), so Z near z
coincides with

f −1 ( D ) ∩ X sm

The task is thereby reduced to showing that the closed subset

f −1 ( D ) ∩ X sm ⊂ X sm

(given the reduced structure) is ncd at each of its k-points z that is non-
smooth in X f (z) .
As we saw in the étale-local model case, X near its regular k-point z ∈
−
f 1 ( D ) has an étale neighborhood in common with

(W = Spec A[u, v]/(uv − t1 ), w)

 MATH 249B NOTES: ALTERATIONS 113

for A a regular local ring, D = Spec ( A/(∏rj=1 t j )) for {t1 , . . . , tr } part of a

regular system of parameters of A, and w = (u, v, m A ).) Clearly the preim-
age
r
f −1
( D ) = V (∏ t j ) ⊂ W
j =1

coincides with
r
Spec ( A/(∏ t j ))[u, v]/(uv − t1 ) ⊂ W.
j =1

In the local ring at w, this closed subscheme is also cut out by uv ∏rj=2 t j (the
product over j means 1 if r = 1), and so to verify the ncd (even sncd) prop-
erty for f −1 ( D ) near w it suffices to show the collection of r + 1 elements
{u, v, t2 , . . . , tr } in the maximal ideal of the regular local ring
( A[u, v]/(uv − t1 ))w
is part of a regular system of parameters. The quotient by (u, v, t2 , . . . , tr ) is
A/(t1 , t2 , . . . , tr ) that is regular (!) with dimension r + 1 less than that of the
regular local ring ( A[u, v]/(uv − t1 ))w , so we are done. 

12.4. Completing the proof of the main result. To complete the proof for
d = 2, we have already reduced to the case of smooth X, and it remains to
turn the ncd Z ⊂ X into an sncd. The final handout shows that for smooth
X of any dimension and Z ⊂ X an ncd, blow up along a suitable closed
subscheme of Z preserves smoothness of the ambient scheme and makes
the preimage of Z in the blow-up“closer” to being sncd; more specifically,
the handout provides an algorithm that reaches a k-smooth blow-up of X
in finitely many steps for which the preimage of Z is an sncd. Beware that
these blow-ups of X typically destroy the semi-stability property over Y,
but that doesn’t matter anymore when d = 2 (and similarly won’t matter in
what follows when d > 2). This completes the case d = 2.
To complete the proof when d ≥ 3, we have the following new setup, sup-
pressing any mention Y and D. We want to make a statement only involving
X and Z because the subsequent blow-ups will usually ruin semistability
over Y.
We are in the following new axiomatic situation. Let X be a projective
variety over an algebraically closed field k, with d = dim X ≥ 3. Let Z ⊂ X
be a proper reduced closed subset so that
(i) The closed subset Z ∩ X sm ⊂ X sm is ncd.
114 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

(ii) The reduced closed subset X sing ⊂ X has pure codimension 3 (if it is
nonempty), with all irreducible components Eα smooth and the non-
empty scheme-theoretic intersections Eα ∩ Eβ for α 6= β also smooth
(but possibly not transverse) with each irreducible component of
such Eα ∩ Eβ of codimension 4 or 5 in X (see Proposition 12.20).
(iii) For all z ∈ ( Z ∩ X sing )(k ), we have
s
O X,z = k [ u, v, t1 , . . . , td−1 ] / ( uv − ∏ t j )
d
j =1

for some 2 ≤ s ≤ d − 1 (we have s 6= 1 since z is not a regular point

on X) and
r
O Z,z = OX,x / ( ∏ t j )
d d
j =1

for some s ≤ r ≤ d − 1. (The reason we have r ≥ s is that Z near

such z in our situation of interest coincides with f −1 ( D ) and some
ni (z) might equal 0 rather than 1.)
Our remaining goal is to construct blow-ups
φ
e−
X →X
such that for
e := φ−1 ( Z )red
Z
the pair ( X,
e Ze ) satisfies (i)-(iii) with X
e “more nearly” k-smooth in a sense we
will make precise below.
In finitely many steps, we’d then reach the case of (i)-(iii) with X smooth
everywhere, so we could use the final handout (as we did at the end of the
treatment of the case d = 2) to make the preimage of Z an sncd in some
smooth modification of X. In particular, if X is already smooth everywhere
then (as in the case d = 2) we are done by the final handout. Thus, we may
and do now assume X sing 6= ∅.
The preceding axioms have two immediate consequences:
(1) For x ∈ X sing (k ), by our usual arguments with Artin approximation
we know that ( X, x ) has a common étale neighborhood with
(Spec k[u, v, t1 , . . . , td−1 ]/(uv − t1 , . . . , ts ),~0)
(with 2 ≤ s ≤ d − 1).
(2) The reduced closed subset Z ⊂ X is Cartier (i.e., IZ is invertible);
this can be checked in the completed local ring of X at all points of
Z (k) (treating separately those that lie in X sm and in X sing ).
 MATH 249B NOTES: ALTERATIONS 115

Now a miracle happens (as explained in the final handout); for each irre-
ducible component E of X sing , the pair
(BlE ( X ), Z-preimage)
satisfies (i)-(iii), and with gain that the irreducible components of BlE ( X )sing
are precisely the strict transforms of the irreducible components E0 of X sing
distinct from E. Therefore, the total number of irreducible components of
the singular locus drops under this process, so after blowing up finitely
many times we arrive at the situation of (i)-(iii) with X now k-smooth every-
where, a situation from which we have already discussed how to conclude
(via some computations for which we punted to a handout).
116 BRIAN CONRAD, TONY FENG, AND AARON LANDESMAN

R EFERENCES
[A] M. Artin, “Lipman’s proof of resolution of singularities for surfaces”, Chapter XI
in Arithmetic Geometry (Cornell/Silverman ed.), Springer-Verlag, New York, 1986.
[BLR] S. Bosch, W. Lütkebohmert, M. Raynaud, Néron Models, Springer-Verlag, New
York, 1990.
[Ch] T. Chinburg, “Minimal models for curves over Dedekind rings”, Chapter XIII in
Arithmetic Geometry (Cornell/Silverman ed.), Springer-Verlag, New York, 1986.
[DM] P. Deligne, D. Mumford, The irreducibility of the space of curves of given genus,
Publ. Math. IHES 36, 75-109.
[FK] E. Freitag, R. Kiehl, Étale cohomology and the Weil conjecture, Springer-Verlag, New
York 1988.
[FL] W. Fulton, S. Lang, Riemann-Roch Algebra, Springer-Verlag, New York, 1985.
[EGA] A. Grothendieck, Eléments de Géométrie Algébrique, Publ. Math. IHES 4, 8, 11, 17, 20,
24, 28, 32, 1960–7.
[SGA1] A. Grothendieck, Séminaire de géométrie algébrique 1, Paris, 1961.
[SGA7I ] A. Grothendieck (with M. Raynaud, D.S. Rim), Seminaire de géométrie algébrique 7,
Lecture Notes in Math 288, Springer-Verlag, New York (1972).
[deJ] A.J. de Jong, Smoothness, semistability, and alterations, Publ. Math. IHES 83 (1996),
51-93
[Jou] J-P. Jouanolou, Théorèmes de Bertini et applications, Progress in Mathematics 42,
Birkhauser, Boston, 1983.
[Knu] F. Knudsen, The projectivity of the moduli space of curves II, III, Math. Scand. 52 (1983),
161-212.
[LMB] G. Laumon, L. Moret-Bailly, Champs Algébriques, Springer-Verlag, New York, 1970.
[Liu] Q. Liu, Algebraic Geometry and Arithmetic Curves, Oxford University Press, 2002.
[Mat1] H. Matsumura, Commutative Algebra, W.A. Benjamin Co., New York, 1970.
[Mat2] H. Matsumura, Commutative Ring Theory, Cambridge University Press, Cambridge,
1990.
[Mum] D. Mumford, Abelian Varieties, Oxford University Press, 1970.
[RG] M. Raynaud, L. Gruson, Critères de platitude et de projectivité: Techniques de platifica-
tion d’un module, Inventiones Math. 13 (1971), 1-89.
[Se] J-P. Serre, Local Fields, GTM 67, Springer-Verlag, New York, 1979.