Castelnuovo-Mumford Regularity and Computing

the de Rham Cohomology of Smooth Projective

Varieties
Peter Scheiblechner∗
Hausdorff Center for Mathematics,
Endenicher Allee 62, 53115 Bonn, Germany
peter.scheiblechner@hcm.uni-bonn.de

Abstract
We describe a parallel polynomial time algorithm for computing the
topological Betti numbers of a smooth complex projective variety X. It is
the first single exponential time algorithm for computing the Betti num-
bers of a significant class of complex varieties of arbitrary dimension. Our
main theoretical result is that the Castelnuovo-Mumford regularity of the
sheaf of differential p-forms on X is bounded by p(em + 1)D, where e, m,
and D are the maximal codimension, dimension, and degree, respectively,
of all irreducible components of X. It follows that, for a union V of generic
hyperplane sections in X, the algebraic de Rham cohomology of X \ V
is described by differential forms with poles along V of single exponential
order. By covering X with sets of this type and using a Čech process, this
yields a similar description of the de Rham cohomology of X, which allows
its efficient computation. Furthermore, we give a parallel polynomial time
algorithm for testing whether a projective variety is smooth.
Key words Castelnuovo-Mumford regularity, de Rham cohomology, algorithm,
complexity, parallel polynomial time, smooth projective variety, Betti numbers,
Cech cohomology, hypercohomology
Mathematics Subject Classification (2010) 14Q20, 68Q25, 14F40, 68W30
Communicated by Elizabeth Mansfield

1 Introduction
A long standing open problem in algorithmic real algebraic geometry is to con-
struct a single exponential time algorithm for computing the Betti numbers of
semialgebraic sets (for an overview see [4]). The best result in this direction
∗ Partially supported by DFG grant SCHE 1639/1-1.

1
is in [3] saying that for fixed ℓ one can compute the first ℓ Betti numbers of a
semialgebraic set in single exponential time.
In the complex setting one approach for computing Betti numbers is to com-
pute the algebraic de Rham cohomology. A result of Grothendieck [30] states
that the de Rham cohomology of a smooth complex variety is canonically iso-
morphic to the singular cohomology. An algorithm based on D-modules for
computing the de Rham cohomology of the complement of a complex affine
variety is given in [46, 61]. This algorithm is used in [62] to compute the coho-
mology of a projective variety. However, these algorithms are not analyzed with
regard to their complexity. Furthermore, they use Gröbner basis computations
in a non-commutative setting, so that a good worst-case complexity is not to be
expected. Indeed, already in the commutative case, computing Gröbner bases
is exponential space complete [40, 39].
Via the well known Hodge decomposition, the de Rham cohomology of a
smooth projective variety X is related to the sheaf cohomologies of the sheaves
of differential forms on X. Algorithms for computing sheaf cohomology are
described in [58, 54, 19] and implemented in Macaulay2 [28], see also [21].
A parallel polynomial time algorithm for counting the connected components
(i.e., computing the zeroth Betti number) of a complex affine variety is given
in [12]. Although this problem can also be solved by applying the corresponding
real algorithms, the algorithm of [12] is the first one using the field structure
of C only. It also extends to counting the irreducible components. In [52]
it is described how one can compute equations for the components in parallel
polynomial time.
Concerning lower bounds it is shown in [51] that it is PSPACE-hard to com-
pute some fixed Betti number of a complex affine or projective variety given over
the integers. Note that the varieties constructed in this reduction are highly sin-
gular. We do not know of any lower bound result for the problem of testing
whether a variety is smooth.

1.1 Main Result

In this paper we describe an algorithm for computing the algebraic de Rham
cohomology of a smooth projective variety running in parallel polynomial time.
It is based on the same techniques as the algorithm in [12] using squarefree
regular chains. Namely, by applying the algorithm of Szántó [56, 57] we can
construct a linear system of equations describing the ideal (up to a given degree)
of an affine variety. This allows to compute a linear system describing the
vanishing of differential forms of some degree. Given a smooth projective variety
X ⊆ Pn of dimension m and generic hyperplane sections H0 , . . . , Hm ⊆ X with
H0 ∩ · · · ∩ Hm = ∅, their complements Ui = X \ Hi form an open affine cover
of X. Under the additional assumption that the hypersurface H0 ∪ · · · ∪ Hm has
normal crossings, we are able to compute the cohomologies of the affine patches
Ui0 ···iq = Ui0 ∩ · · · ∩ Uiq and by a Čech process also the cohomology of X.
More precisely, let D be the maximal degree and e the maximal codimension
of all irreducible components of X, and choose s ≥ m(em + 1)D. Now consider

2
the double complex
M
K p,q := Γ(X, ΩpX ((s + p)(Hi0 ∪ · · · ∪ Hiq )))
i0 <···<iq

together with the Čech differential δ : K p,q → K p,q+1 and the exterior differen-
tial d : K p,q → K p+1,q . We show that the de Rham cohomology of X can be
computed as the cohomology of the total complex of K •,• , i.e.,
•
HdR (X) ≃ H • (tot• (K •,• )).

To describe the output of the algorithm explicitly, let the Hi be defined by

k
the linear forms ℓi . The cohomology HdR (X) is then represented by a basis
consisting of vectors of rational differential forms ω = (ωi0 ···iq ) of the form

1 X j ···j
ωi0 ···iq = ωi01···iqp dXj1 ∧ · · · ∧ dXjp , (1)
(ℓi0 · · · ℓiq )t
0≤j1 <···<jp ≤n

j ···j
where p + q = k, t ≤ s + p, and the polynomials ωi01···iqp ∈ C[X0 , . . . , Xn ] are
homogeneous of degree t(q + 1) − p.
Furthermore, we show how to test whether a variety X is smooth and how
to choose sufficiently generic generic hyperplanes Hi with normal crossings in
parallel polynomial time. In summary, we prove the following theorem.

Theorem 1.1. Given homogeneous polynomials f1 , . . . , fr ∈ C[X0 , . . . , Xn ] of

degree at most d, one can test whether X := Z(f1 , . . . , fr ) ⊆ Pn is smooth
•
and if so, compute the algebraic de Rham cohomology HdR (X) in parallel time
O(1) O(n4 )
(n log d) and sequential time d .

As for the necessity of squarefree regular chains in our algorithm we remark

that for smooth varieties there are single exponential bounds for the degrees in
a Gröbner basis (cf. §1.2). So perhaps one could replace the squarefree regular
chains in our approach by Gröbner bases. But we do not know whether such
an algorithm would be well-parallelizable.
The new aspect of our result compared to the methods of [46, 61, 62] is the
complexity analysis of the algorithm, which is the main purpose of this paper.
We do not expect that our algorithm as stated would yield a practically efficient
implementation. However, it seems conceivable that a variation of the approach
(perhaps using Gröbner bases) could produce a reasonable implementiation.

1.2 Castelnuovo-Mumford Regularity

The main theoretical result of this paper is a bound on the Castelnuovo-Mum-
ford regularity of the sheaf of regular differential p-forms on a smooth projective
variety X. This result allows to bound the degrees of the differential forms
one has to deal with in computing the cohomology of X. More precisely, the

3
regularity yields a bound on the order t in (1), which in turn determines the
degree of the coefficients.
The Castelnuovo-Mumford regularity was defined in [43] for sheaves on Pn .
The definition has been modified in [20] to apply to a homogeneous ideal I.
This notion was related to computational complexity in [6] by showing that the
regularity of I equals the maximal degree in a reduced Gröbner basis of I with
respect to the degree reverse lexicographic order in generic coordinates. In this
respect upper bounds on the regularity of a homogeneous ideal are of particular
interest in computational algebraic geometry and commutative algebra. For a
general ideal, double exponential upper bounds were shown in [26] and [24]. The
famous example of [40] shows that this is essentially best possible. However,
there are several results giving better bounds for the regularity in special cases,
such as [31, 48, 37, 49, 47, 55, 35, 25]. A nice overview over these kinds of results
is given in [5]. This paper also contains a bound on the regularity of the ideal of
a smooth veriety X, which is asymptotic to the product of the degree and the
dimension of X. A more precise bound in terms of the degrees of generators of
the ideal of X is proved in [8]. More generally, the authors prove the vanishing
of the higher cohomology of powers of the sufficiently twisted ideal sheaf of X
(cf. Proposition 3.6). Our bound on the regularity of the sheaf of differential
forms on X is deduced from this result. We are not aware of any other bounds
on the regularity of a sheaf other than a power of an ideal sheaf.

Acknowledgements
The author is very grateful to Saugata Basu for being his host, many important
and interesting discussions, and recommending the book [38]. Without him this
work wouldn’t have been possible. The author also thanks Manoj Kummini
for fruitful discussions about the Castelnuovo-Mumford regularity, Christian
Schnell for a discussion about the cohomology of hypersurfaces, Nicolas Perrin
and Martı́ Lahoz for useful discussions about exterior powers of sheaves, as well
as the anonymous referees for valuable comments and suggestions.

2 Preliminaries
2.1 Basic Notations
Denote by Pn := Pn (C) the projective space over C. A (closed) projective variety
X ⊆ Pn is defined as the zero set

X = Z(f1 . . . , fr ) := {x ∈ Pn | f1 (x) = · · · = fr (x) = 0}

of homogeneous polynomials f1 . . . , fr ∈ C[X0 , . . . , Xn ]. Note that X may be

reducible. Occasionally we write ZX (f ) := X ∩ Z(f ) for f ∈ C[X0 , . . . , Xn ]. A
quasi-projective variety is a difference X \ Y , where X and Y are closed pro-
jective varieties. The term variety will always mean quasi-projective variety.
The homogeneous (vanishing) ideal I(X) of the variety X is defined as the ideal

4
generated by the homogeneous polynomials vanishing on X. The homogeneous
coordinate ring of X is C[X] := C[X0 , . . . , Xn ]/I(X). By the (weak) homo-
geneous Nullstellensatz we have Z(f1 , . . . , fr ) = ∅ iff there exists N ∈ N with
(X0 , . . . , Xn )N ⊆ (f1 , . . . , fr ). Its effective version states that one can choose
N = (n + 1)d − n, when deg fi ≤ d [36, Théorème 3.3]. According to the affine
effective Nullstellensatz, for f1 , . . . , fr ∈ C[X1 , . . . , Xn ] of degree ≤ d, we have
Z(f1 , . . . ,P fr ) = ∅ iff there exist polynomials g1 , . . . , gr with deg(gi fi ) ≤ dn such
that 1 = i gi fi [10, 34, 22, 33].
The dimension dim X is the Krull dimension of X in the Zariski topology.
A variety all of whose irreducible components have the same dimension m is
called (m-)equidimensional. The local dimension dimx X at x ∈ X is defined
as the maximal dimension of all components through x. A hypersurface of a
variety X is a closed subvariety V with dimx V = dimx X − 1 for all x ∈ V .
We often identify Pn \ Z(Xi ) ≃ Cn , 0 ≤ i ≤ n, via (x0 : · · · : xn ) 7→
( xi , . . . , c
x0 xi xn
xi , . . . , xi ), where as usual b denotes omission. Under this identifica-
tion, a homogeneous polynomial f ∈ C[X0 , . . . , Xn ] corresponds to its dehomog-
enization f i by setting Xi := 1. We thus get a surjection i : C[X0 , . . . , Xn ] →
C[X0 , . . . , X ci , . . . , Xn ], and the image of I(X) under this map is the vanish-
ing ideal of the affine variety X \ Z(Xi ). Now let f1 , . . . , fr be homogeneous
polynomials defining the hypersurfaces H1 , . . . , Hr of Pn . Then we say that the
closed variety X is scheme-theoretically cut out by the hypersurfaces H1 , . . . , Hr
iff for each i, the dehomogenizations f1i , . . . , fri generate the image of I(X) in
C[X0 , . . . , X ci , . . . , Xn ].
For a polynomial f ∈ C[X1 , . . . , Xn ] its differential
P ∂f at x ∈ Cn is the linear
function dx f : C → C defined by dx f (v) := i ∂Xi (x)vi . The tangent space
n

of the variety X at x ∈ X \ Z(Xi ) is defined as the vector subspace

Tx X := {v ∈ Cn | ∀f ∈ I(X) dx f i (v) = 0} ⊆ Cn .

If X is scheme-theoretically cut out by the hypersurfaces defined by the ho-

mogeneous polynomials f1 , . . . , fr , then Tx X = Z(dx f1i , . . . , dx fri ). We have
dim Tx X ≥ dimx X for all x ∈ X. We say that x ∈ X is a smooth point in X iff
dim Tx X = dimx X. The variety X is smooth iff all of its points are smooth.
The degree deg X of an irreducible closed variety X of dimension m is defined
as the maximal cardinality of X∩L over all linear subspaces L ⊆ Pn of dimension
n − m [45, §5A]. We define the (cumulative) degree deg X of a reducible variety
X to be the sum of the degrees of all irreducible components of X. It follows
essentially from Bézout’s Theorem that if X is defined by polynomials of degree
≤ d, then deg X ≤ dn [14].

2.2 Coherent Sheaves

Let X be a closed variety in Pn . Then every graded C[X]-module M gives
f of OX -modules on X such that, on a principal open set
rise to a sheaf M
f are given by Γ(U, M
U = X \ Z(f ), the sections of M f) = M(f ) , the degree 0

5
part of the localization of M at f . A sheaf F on X is called coherent iff F = M f
with a finitely generated graded C[X]-module M .
An important example is of course the structure sheaf OX = C[X]. ] We
^ for k ∈ Z, where C[X](k)d :=
also define the twisting sheaf OX (k) := C[X](k)
C[X]k+d . The twisting sheaf behaves well with respect to direct and inverse im-
ages under maps. In particular, for a closed embedding i : X ֒→ Y of projective
varieties we have i∗ (OY (k)) ≃ OX (k) and i∗ (OX (k)) ≃ (i∗ OX )(k). Further-
more, we have OX (k) ⊗ OX (ℓ) ≃ OX (k + ℓ) [32, II, Proposition 5.12]. The sheaf
OX (1) is called the very ample line bundle on X determined by the embedding
X ֒→ Pn . For any sheaf of OX -modules F on X we define the twisted sheaf
F(k) := F ⊗ OX (k) for k ∈ Z. The ideal sheaf IX of X is defined as the kernel
of the restriction map OPn → i∗ OX . We have IX = I(X), ] hence the ideal sheaf
is coherent.
Fundamental for us is the sheaf of regular differential forms, which is defined
as follows. Let U := {Ui | 0 ≤ i ≤ s} be an open affine cover of X (we can take
Ui := {Xi 6= 0}, but any other open cover works too). Denote by C[Ui ] the affine
coordinate ring of Ui , and let ΩUi be the C[Ui ]-module of Kähler differentials
ΩC[Ui ]/C [18, Chapter 16]. Then, ΩX is defined as the sheaf on X obtained by
glueing the sheaves on Ui corresponding to ΩUi . This means that one determines
a section s on an open set U ⊆ X by giving for each i a section si ∈ ΩUi ∩U with
the property that for all i, j we have si |Ui ∩Uj ∩U = sj |Ui ∩Uj ∩U . The universal
derivations d : C[Ui ] → ΩUi glue together to give a map d : OX → ΩX of sheaves,
Vp
which is a derivation on the stalks. The p-fold exterior power ΩpX := ΩX
is called the sheaf of regular (p-)forms. The derivation d : OX → ΩX uniquely
extends to the exterior derivative d : ΩpX → Ωp+1 X satisfying Leibnitz’ rule and
d ◦ d = 0, so that we get the (algebraic) de Rham complex
d d d
Ω•X : 0 −→ OX −→ Ω1X −→ · · · −→ Ωm
X −→ 0,

where m = dim X.
For X = Pn there is a more explicit description of this complex, which
we give now (cf. §6.1 of [17]). Denote by Λ the module of Kähler differen-
tials ΩC[X0 ,...,Xn ]/C , which is the free module generated by dX0 , . . . , dXn [18,
PropositionP16.1]. The universal derivation Vp d : C[X0 , . . . , Xn ] → Λ is given
∂f p
by df = i ∂Xi dXi . Then, Λ := Λ is the free module generated by
dXi1 ∧ · · · ∧ dXip , 0 ≤ i1 < · · · < ip ≤ n. Furthermore, the exterior derivative
d : Λp → Λp+1 yields the de Rham complex
d d d d
Λ0 = C[X0 , . . . , Xn ] −→ Λ1 −→ · · · −→ Λn −→ Λn+1

The modules Λp are graded by setting

deg(gdXi1 ∧ · · · ∧ dXip ) := deg(g) + p, g ∈ C[X0 , . . . , Xn ] homogeneous.

Then d is a map of degree 0. There is another derivation ∆ : Λp → Λp−1

of degree 0, which can be defined as the contraction with the Euler vector field

6
P ∂
Xi ∂X
i i
. It is uniquely determined by Leibnitz’ rule and the formula ∆(df ) =
deg f ·f for a homogeneous polynomial f , and satisfies ∆(dα)+d(∆α) = deg α·α
for any homogeneous α ∈ Λp .
Now put M p := ker(∆ : Λp → Λp−1 ). One can define the sheaf of differential
p-forms on Pn by setting ΩpPn := M gp . Hence, for a homogeneous polynomial f
of degree k, each differential p-form on Pn \ Z(f ) is of the form
α
ω= with deg α = tk and ∆(α) = 0, (2)
ft
where α ∈ Λp is homogeneous and t ∈ N. By the usual quotient rule one can
p
extend d : Λp → Λp+1 to localizations. Then one easily checks that d(M(f )) ⊆
p+1
M(f ) for homogeneous f ∈ C[X0 , . . . , Xn ]. This defines the exterior derivative
d : ΩpPn → Ωp+1
Pn on the sheaf level.
A sheaf F on X is said to be locally free iff it is locally isomorphic to a
direct sum of copies of OX . The local rank of F is the number of copies of the
structure sheaf needed, which is a locally constant function.

If X is smooth and
m-equidimensional, then ΩpX is locally free of rank m p .
The isomorphism classes of locally free sheaves on X of rank k are in one-
to-one correspondence with those of vector bundles over X of rank k [32, II,
Exercise 5.18]. A locally free sheaf of rank 1 is called an invertible sheaf or
line bundle. The tensor product L ⊗ M of two line bundles L, M is also a line
bundle. For any line bundle L, its dual sheaf Ľ := Hom(L, OX ) is another
line bundle satisfying L ⊗ Ľ ≃ OX [32, II, Proposition 6.12]. In particular, the
sheaves OX (k) defined above are line bundles, and we have OX (k)ˇ = OX (−k).

2.3 Divisors and Line Bundles

Let X be a smooth variety. A divisor on X is an element of the free abelian
group Div X generated by the irreducible hypersurfacesPof X. This means that
each D ∈ Div X is a formal linear combination D = i mi Vi , where mi ∈ Z
and Vi ⊆ S X are irreducible hypersurfaces. The support of D is defined as
supp D := mi 6=0 Vi . Each D ∈ Div X defines a line bundle OX (D), which is
a subsheaf of the sheaf K of total quotient rings of OX [43, p. 61]. Q The stalk
OX (D)x is OX,x , if x 6∈ supp D. For x ∈ supp D, the stalk is i fi−mi OX,x ,
where fi is a local equation of Vi at x. The rule D 7→ OX (D) maps Div X
bijectively onto the invertible subsheaves of K, and it maps sums to tensor
products [32, II, Proposition 6.13].
Let X be closed and H a hyperplane in Pn meeting X properly, i.e., H does
not contain any irreducible component of X, so that X ∩ H is a hypersurfaceP
in X. Then the hyperplane section V := X∩H defines a divisor H·X = i mi Vi ,
where the Vi are the irreducible components of V , and mi is the intersection
multiplicity i(X, H; Vi ) between X and H along Vi , which can be defined as
follows. Choose x ∈ Vi and a (reduced) local equation f ∈ OPn ,x of H. Then
i(X, H; Vi ) is the order of vanishing ordVi (f ) of f along Vi , i.e., the maximal
k ∈ N such that f = ghk with some g in OX,x , where h is a local equation of Vi

7
at x. Note that OX,x is factorial, since X is smooth. The line bundle OX (H ·X)
is isomorphic to the very ample line bundle OX (1).
A hypersurface V ⊆ X is said to have normal crossings iff for each x ∈ V
contained in k irreducible components V1 , . . . , Vk of V , there exist local equations
fi ∈ OX,x of Vi around x, such that dx f1 , . . . , dx fk are linearly independent
in the dual (Tx X)∗ . Note that the case k = 1 implies that all irreducible
components of V are smooth.

2.4 Sheaf Cohomology

Let F be a coherent sheaf and U := {Ui | 0 ≤ i ≤ s} an open cover of the
variety X. For 0 ≤ q ≤ s and 0 ≤ i0 < · · · < iq ≤ s set Ui0 ···iq := Ui0 ∩ · · · ∩ Uiq .
L
The Čech complex is defined by C q := C q (U , F) := i0 <···<iq F(Ui0 ···iq ), with
the Čech differential δ : C q → C q+1 given by
q+1
X
(δ(ω))i0 ···iq+1 := (−1)ν ωi0 ···biν ···iq+1 |Ui0 ···iq+1 for ω = (ωi0 ···iq ) ∈ C q . (3)
ν=0

Then one easily checks that δ ◦ δ = 0, hence (C • , δ) is indeed a complex. Its

cohomology H i (U , F) := H i (C • , δ) is called the i-th Čech cohomology of F with
respect to U . The i-th Čech cohomology (or sheaf cohomology) of F is defined as
the direct limit over all open covers U of X, directed by refinements. A sheaf F
on X is called acyclic iff H i (X, F) = 0 for all i > 0.
A cover U of X is called a Leray cover for F iff F is acyclic on Ui0 ···iq for
all i0 < · · · < iq . Leray’s Theorem states that in this case we have H i (X, F) =
H i (U , F) [32, III, Exercise 4.11]. Since F is a coherent sheaf, this is true for
any affine cover [32, III, Theorem 3.5]. It easily follows that for a morphism
f : X → Y there is a natural isomorphism H i (X, F) ≃ H i (Y, f∗ F), where f∗ F
denotes the direct image of the sheaf F under f [32, III, Exercise 4.1].

2.5 Hypercohomology and de Rham Cohomology

The material in this section is explained e.g. in [29]. Let X be a smooth variety
and consider a complex of coherent sheaves (F • , d) on X with F p = 0 for p < 0.
Then, for an open cover U , the Čech complexes C • (U , F p ) as defined in §2.4 fit
together to the Čech double complex C •,• := C •,• (U , F • ) by setting
M
C p,q (U , F • ) = F p (Ui0 ···iq ) for all p, q ≥ 0.
i0 <···<iq

The two differentials are the one induced by the differential d of F and the
Čech differential δ defined by (3). Denote by H• (U , F • ) := H • (tot• (C •,• ))
the cohomology of the total complex of C •,• (U , F • ). Then the hypercohomol-
ogy Hi (X, F • ) of the complex of sheaves F • is defined as the direct limit of
H• (U , F • ) over all open covers U of X, directed by refinement. As for any

8
double complex [41, §2.4], there are two spectral sequences
′
E2p,q = Hdp (H q (X, F • )) ⇒ Hp+q (X, F • ) and
′′ p,q q p • p+q •
E2 = H (X, H (F )) ⇒ H (X, F ),

where the cohomology sheaf is defined by

Hp (F • ) := ker(d : F p → F p+1 )/im (d : F p−1 → F p ).

The first spectral sequence implies that if all the sheaves F p are acyclic, then
H• (X, F • ) = H • (Γ(X, F • )) is the cohomology of the complex of global sections.
Similarly as for sheaf cohomology we have Hi (X, F • ) ≃ H• (U , F • ), if U is a
Leray cover for all F p .
A map of complexes of sheaves f : F • → G • is called a quasi-isomorphism
iff it induces an isomorphism H• (F • ) → H• (G • ). By comparing the second
spectral sequences of the hypercohomologies of F • and G • it follows that f
∼
induces an isomorphism H• (X, F • ) → H• (X, G • ).
The algebraic de Rham cohomology of X is defined as the hypercohomology
•
HdR (X) := H• (X, Ω•X )
•
of the de Rham complex. If X is affine, then HdR (X) can be computed by
taking the cohomology of the complex Γ(X, ΩX ) of global sections, since all ΩpX
•

are acyclic.

2.6 Computational Model

Our model of computation
F is that of algebraic circuits over C, cf. [60, 11].
We set C∞ := n∈N Cn . The size of an algebraic circuit C is the number of
nodes of C, and its depth is the maximal length of a path from an input to an
output node. We say that a function f : C∞ → C∞ can be computed in parallel
time d(n) and sequential time s(n) iff there exists a polynomial-time uniform
family of algebraic circuits (Cn )n∈N over C of size s(n) and depth d(n) such
that Cn computes f |Cn .

2.7 Efficient Parallel Linear Algebra

We use differential forms to reduce our problem to linear algebra, for which
efficient parallel algorithms exist. In particular, we need to be able to solve the
following problems:

1. Given A ∈ Cn×m and b ∈ Cn , decide whether the linear system of equa-

tions Ax = b has a solution and if so, compute one.

2. Compute a basis of the kernel of a matrix A ∈ Cn×m .

3. Compute a basis of the image of a matrix A ∈ Cn×m .

9
 4. Given a linear subspace V ⊆ Cn in terms of a basis, and given linearly
independent v1 , . . . , vi ∈ V , extend them to a basis of V .

These problems are easily reduced to inverting a regular square-matrix (thus to

computing the characteristic polynomial) and computing the rank of a matrix.
For instance, the last problem boils down to rank computations as follows. Let
b1 , . . . , bm ∈ V be the given basis. Set B := (v1 , . . . , vi ). For all j = 1, 2, . . . , m
do: if rk (B, bj ) > rk B then append bj to B.
Mulmuley [42] has reduced the problem of computing the rank to the com-
putation of the characteristic polynomial of a matrix. Since we need his con-
′
struction, we describe it here. Let A ∈ Cm×m be a matrix. Then
 
0 A
rk = 2 rk A,
AT 0

so we can assume m = m′ . Introducing the additional variable T , define the di-

agonal matrix X := diag(1, T, . . . , T m−1 ), and consider the characteristic poly-
nomial pA (Z) of XA over the field C(T ), pA (Z) := det(XA − ZI). Then the
rank of A equals m − s, where s is the maximal integer with Z s |pA (Z). We will
call pA (Z) the Mulmuley polynomial of A.
The characteristic polynomial of an m×m matrix can be computed in parallel
(sequential) time O(log2 m) (mO(1) ) with the algorithm of [7]. If the matrix has
polynomial entries of degree d in n variables, then the Berkowitz algorithm can
be implemented in parallel (sequential) time O(n log m log(md)) ((md)O(n) ) [50].

3 Castelnuovo-Mumford Regularity
A nice exposition about various versions of Castelnuovo-Mumford regularity and
vanishing results is contained in the book [38]. Let X ⊆ Pn be a smooth closed
subvariety. Recall from §2.2 that OX (1) denotes the very ample line bundle
on X determined by the embedding X ֒→ Pn , and that for a coherent sheaf F
on X we put F(k) := F ⊗OX (k). The following definition is due to [43] building
on ideas of Castelnuovo.

Definition 3.1. The coherent sheaf F on X is called k-regular iff

H i (X, F(k − i)) = 0 for all i > 0. (4)

The Castelnuovo-Mumford regularity reg (F) of F is defined as the infimum over

all k ∈ Z such that F is k-regular.

Remark 3.2. (i) A fundamental result of [43] is that if F is k-regular, then F

is ℓ-regular for all ℓ ≥ k.

(ii) Let
0→F →G→H→0

10
 be a short exact sequence of coherent sheaves on X. The long exact
cohomology sequence shows that
reg (H) ≤ max{reg (F) − 1, reg (G)} and
reg (G) ≤ max{reg (F), reg (H)}.

(iii) Let i : X ֒→ Pn be the closed embedding, and F a coherent sheaf on X.

The projection formula shows i∗ (F ⊗ i∗ OPn (t)) = i∗ F ⊗OPn OPn (t) for all
t ∈ Z [32, II, Exercise 5.1], hence
reg (F) = reg (i∗ F).
This is usually used as an argument that one can restrict to the case
X = Pn . However, we are dealing with exterior powers and thus need the
more general situation, since a direct image of a locally free sheaf is not
locally free, and reasonable formulas for exterior powers hold for locally
free sheaves only (in particular, Corollary 3.13).
(iv) Let X ⊆ Pn be a subscheme of dimension m with ideal sheaf I = IX , and
let k > 0. Then I is k-regular if and only if H i (Pn , I(k − i)) = 0 for all
0 < i ≤ m + 1 [38, Example 1.8.29]. This follows from the short exact
sequence 0 → I → OPn → i∗ OX → 0 using that H i (X, F) = 0 for all
i > m and any coherent sheaf F [32, III, Theorem 2.7].
Example 3.3. (i) Theorem 5.1 in Chapter III of [32] shows reg (OPn ) = 0.
(ii) The structure sheaf OX of a hypersurface X ⊆ Pn of degree D has reg-
ularity D − 1. This follows from the isomorphism I ≃ OPn (−D) for the
ideal sheaf I of X and the exact sequence 0 → I → OPn → i∗ OX → 0.
(iii) Example (i) together with the exact sequence of [32, II, Theorem 8.13]
implies reg (ΩPn ) = 2.
The aim of this section is to prove the following theorem.
Theorem 3.4. Let X ⊂ Pn be a smooth closed subvariety of dimension m.
Let D be the maximal degree and e the maximal codimension of all irreducible
components of X. Then
reg (ΩpX ) ≤ p(em + 1)D for p > 0,
reg (OX ) ≤ e(D − 1).
Remark 3.5. For X = Pn the first claim is false as Example 3.3 (iii) shows.
We will reduce this theorem to the following vanishing result of [8].
Proposition 3.6. Let I be the ideal sheaf of a smooth irreducible closed variety
in Pn of codimension e, which is scheme-theoretically cut out by hypersurfaces
of degrees at most D. Then
H i (Pn , I a (k)) = 0 for a ≥ 0, i > 0, k ≥ (a + e − 1)D − n,
where I a denotes the a-th power of the ideal sheaf I.

11
Remark 3.7. In [8] there is proved a more precise bound in terms of the indi-
vidual degrees of the hypersurfaces which cut out X, but we do not need this
here.
We also use the following result of [44].

Proposition 3.8. Each smooth irreducible closed projective variety of degree D

is scheme-theoretically cut out by hypersurfaces of degree D.

Corollary 3.9. Let I be the ideal sheaf of a smooth irreducible projective vari-
ety X in Pn of degree D and codimension e. Then

reg (I a ) ≤ (a + e − 1)D − e + 1.

Proof. This follows immediately from Propositions 3.6 and 3.8 together with
part (iv) of Remark 3.2 (note that I a is the ideal sheaf of some subscheme of
the same dimension as X).
Before we prove Theorem 3.4, let us first gather some basic properties of
regularity. In the following one can always assume X to be irreducible. The next
two lemmas are a version of a well known technique to characterize regularity
by free resolutions.
Lemma 3.10. Let

FN → FN −1 → · · · → F0 → F → 0

be an exact sequence of coherent sheaves on X, where N + 1 ≥ dim X =: m.

Then

reg (F) ≤ max{reg (F0 ), reg (F1 ) − 1, . . . , reg (Fm−1 ) − m + 1}.

Proof. This follows easily by chasing through the complex [38, Proposition
B.1.2], taking into account that H i (X, F) = 0 for all i > m and any coher-
ent sheaf F. Another proof is given in [1, Lemma 3.9].

For a finite dimensional vector space V and a coherent sheaf F on X we

denote by V ⊗ F the sheaf U 7→ V ⊗C F(U ). If v1 , . . . , vN is a basis of V , then
LN
V ⊗ F = i=1 vi ⊗ F. The following lemma is proved as Corollary 3.2 in [1].
Set Reg (X) := max{1, reg (OX )}.

Lemma 3.11. Let F be a k-regular coherent sheaf on X. Then there exist finite
dimensional vector spaces Vi and an exact sequence

· · · → Vi ⊗ OX (−k − iR) → · · · → V1 ⊗ OX (−k − R) → V0 ⊗ OX (−k) → F → 0,

(5)
where R := Reg (X).

Using this we prove a bound on the regularity of tensor products.

12
Proposition 3.12. Let F, G be coherent sheaves on X, where G is locally free,
and denote m := dim X and R := Reg (X) as above. Then

reg (F ⊗ G) ≤ reg (F) + reg (G) + (m − 1)(R − 1).

Proof. The proof parallels the one of the special case X = Pn [38, Proposition
1.8.9]. Let k := reg (F), and consider the resolution (5) of F, which exists
according to Lemma 3.11. Tensoring with G yields

· · · → Vi ⊗ G(−k − iR) → · · · → V1 ⊗ G(−k − R) → V0 ⊗ G(−k) → F ⊗ G → 0.

Since tensoring with a locally free sheaf is an exact functor, this sequence is
exact. Furthermore, reg (Vi ⊗G(−k−iR)) ≤ k+iR+reg (G), hence reg (F ⊗G) ≤
k + reg (G) + (m − 1)(R − 1) by Lemma 3.10.

Corollary 3.13. Let F be a locally free sheaf on X. Then for p > 0

reg (Λp F) ≤ p · reg (F) + (p − 1)(m − 1)(R − 1).

Proof. The same bound for the p-th tensor power of F clearly follows from
Proposition 3.12. Since the exterior power is a direct summand of the tensor
power [9, III, §7.4], this implies the claim.

Proposition 3.14. Let X ⊆ Pn be a smooth irreducible subvariety of codimen-

sion e and degree at most D ≥ 2. Then

reg (ΩX ) ≤ (e + 1)D − e.

Proof. Denote with I the ideal sheaf of X and let i : X ֒→ Pn be the inclusion.
Part (ii) of Remark 3.2 applied to the exact sequence

0 → I → OPn → i∗ OX → 0

implies reg (OX ) = reg (i∗ OX ) ≤ max{reg (I) − 1, reg (OPn )}. Using Corol-
lary 3.9 and Example 3.3 (i) we conclude

reg (OX ) ≤ eD − e. (6)

Furthermore, from the exact sequence 0 → I 2 → I → I/I 2 → 0 it follows

reg (I/I 2 ) ≤ max{reg (I), reg (I 2 ) − 1} ≤ (e + 1)D − e. (7)

The last exact sequence we consider is the conormal sequence

0 → I/I 2 → OX ⊗ ΩPn → ΩX → 0 (8)

on X [32, II, Theorem 8.17]. More precisely, the sheaf in the middle is the
inverse image
i∗ ΩPn = OX ⊗i−1 OPn i−1 ΩPn

13
of ΩPn under i as an OX -module. The sheaf on the left is the restriction i−1 I/I 2 ,
which is automatically an OX -module.
We want to push the sequence (8) forward to Pn . In general, the direct image
of sheaves does not commute with stalks. However, for closed immersions it
does: if F is a sheaf on X, we have (i∗ F)x = Fx for x ∈ X, and 0 otherwise [32,
II, Exercise 1.19(a)]. It follows that the functor i∗ is exact.
Furthermore, since ΩPn is locally free of finite rank, the projection formula
shows i∗ i∗ ΩPn = i∗ OX ⊗OPn ΩPn . Since the stalk of I/I 2 at x ∈ / X vanishes,
we have i∗ i−1 I/I 2 = I/I 2 . Thus, applying i∗ to (8) yields the exact sequence

0 → I/I 2 → i∗ OX ⊗ ΩPn → i∗ ΩX → 0.

Hence, reg (ΩX ) = reg (i∗ ΩX ) ≤ max{reg (i∗ OX ⊗ ΩPn ), reg (I/I 2 )}. Now
Proposition 3.12, Example 3.3 (iii), and (6) imply reg (i∗ OX ⊗ΩPn ) ≤ reg (OX )+
reg (ΩPn ) ≤ eD − e + 2, so that reg (ΩX ) ≤ (e + 1)D − e by (7).
Remark 3.15. (i) In the case D = 1 we have reg (ΩX ) ≤ 2.
(ii) One can check that for a hypersurface X ⊆ Pn of degree D ≥ 3 one has
reg (ΩX ) = 2D − 2, so that our bound is essentially sharp.
Proof of Theorem 3.4. The claim can easily be checked for D = 1. Also, we
can assume X to be irreducible. The claim for p = 0 is (6). For the case p ≥ 1,
Proposition 3.14 and Corollary 3.13 imply

reg (ΩpX ) ≤ p((e + 1)D − e) + (p − 1)(m − 1)(e(D − 1) − 1)

< p ((e + 1)D − e + (m − 1)e(D − 1))
< p(em + 1)D.

4 Cohomology of Hypersurface Complements

4.1 Theory
Let X ⊆ Pn be a smooth closed subvariety. Using our result on the Castelnuovo-
Mumford regularity of the sheaf of differential forms, one can compute the de
Rham cohomology of certain hypersurface complements in X as the cohomology
of finite dimensional complexes.
To describe these complexes, let H0 , . . . , Hq ⊆ X be hyperplane
S sections and
denote by U the complement of the hypersurface V := ν Hν in X. Assume
P V hasPnormal crossings (see §2.3). We also consider V as a divisor V =
that
ν Hν = i mi Vi , where the Vi are the irreducible components of V (cf. §2.3).
Then, since OX (Hν ) ≃ OX (1), it follows that
O
OX (V ) ≃ O(Hν ) ≃ OX (1)⊗(q+1) ≃ OX (q + 1). (9)
ν
P
Now let A = i ai Vi be any divisor with support in V , and let j : U ֒→ X
be the inclusion. Define the subsheaf ΩpX (A) := ΩpX ⊗ OX (A) of j∗ ΩpU , which

14
consists of those rational differential p-forms on X, which are regular on U and
have poles (zeros if ai < 0) of order |ai | along Vi . Define the sheaves
p
KX (A) := ΩpX (A + pV ).
p p+1 •
Note that d(KX (A)) ⊆ KX (A), so that KX (A) is in fact a subcomplex of j∗ Ω•U .
For A = V it is the zeroth term of the polar filtration [16, 17].
The next lemma is the crucial fact that allows us to compute the algebraic
de Rham cohomology of U by a finite dimensional complex. Its proof requires
to consider holomorphic differential forms. So let Ω•U an denote the complex
of holomorphic differential forms on U an regarded as a complex manifold, and
• •
let KX an (A) be the holomorphic version of KX (A). The following lemma is
proved analogously to the corresponding statement for the logarithmic complex
(cf. [15, 29, 59]). The calculation can be found in [2, Lemma 4.1].
P
Lemma 4.1. Let A = i ai Vi be a divisor with ai > 0 for all i, and assume
• •
that V has normal crossings. Then the inclusion KX an (A) ֒→ j∗ ΩU an is a quasi-

isomorphism.
The following is the main result of this section and the key for our algorithm.
Theorem 4.2. Let X ⊆ Pn be a smooth closed subvariety with dimension at
most m ≥ 1. Let D ≥ 2 and e be upper bounds on the degree and the codimension
of all irreducible components of X. Let H0 , . . . , Hq be hyperplane sections of X
such that V = H0 ∪ . . . ∪ Hq has normal crossings, and denote U := X \ V . For
s ∈ N set Ks• := Γ(X, KX •
(sV )). Then we have
•
HdR (U ) ≃ H • (Ks• ) for s ≥ m(em + 1)D.

Proof. It follows from Lemma 4.1 that

H• (X an , KX
• •
an (sV )) ≃ H (X
an
, j∗ Ω•U an ). (10)
•
Since KX (sV ) and j∗ Ω•U are coherent sheaves on X, by GAGA [53] the hyper-
cohomologies in (10) can be replaced by their algebraic versions. But we have
H i (X, j∗ ΩpU ) = H i (U, ΩpU ) = 0 for i > 0 since U is affine (see §2.2). Hence, the
•
right hypercohomology in (10) is HdR (U ).
On the other hand, Theorem 3.4 implies s ≥ reg (ΩpX ) for all 0 ≤ p ≤ m,
thus

H i (X, ΩpX ((s + p)V )) = H i (X, ΩpX ((s + p)(q + 1))) = 0 for all i>0

(use (9)). It follows that the left side of (10) is H • (Ks• ) as claimed.

4.2 Computation
We adopt the notations and assumptions of the last section. We choose s ac-
cording to Theorem 4.2 and set K • := Ks• . In this section we describe this finite
dimensional complex more explicitly and show how to compute its cohomology.

15
 Let Hν = ZX (ℓν ) with linear forms ℓν , 0 ≤ ν ≤ q. We can assume w.l.o.g.
that ℓ0 = X0 . With f := X0 ℓ1 · · · ℓq we have V = ZX (f ) and U = X \ V . On
the ambient space we define Ve := Z(f ) and U e := Pn \ Ve . Recall from §2.2 that
e
each p-form on U is given by
α
ω= with deg α = t(q + 1) and ∆(α) = 0, (11)
ft
where α is a homogeneous p-form on Cn+1 and ∆ denotes contraction with the
Euler vector field. Note that for fixed t such forms are precisely the global
sections of the sheaf ΩpPn (tVe ). We have the following
Lemma 4.3. With s ≥ m(em + 1)D the restriction map

Γ(Pn , ΩpPn (tVe )) −→ Γ(X, ΩpX (tV ))

is surjective for all t ≥ s and p ≥ 0.
Proof. Let i : X ֒→ Pn be the inclusion and consider the exact sequence
α
0 → I ⊗ ΩpPn → ΩpPn → i∗ OX ⊗ ΩpPn → 0 (12)
on Pn , as well as
β
0 → ker β → i∗ ΩpPn → ΩpX → 0 (13)
on X. The restriction map from the lemma coincides with the composition of
α∗
H 0 (Pn , ΩpPn (t(q+1))) −→ H 0 (Pn , i∗ OX ⊗ΩpPn (t(q+1))) ≃ H 0 (X, i∗ ΩpPn (t(q+1)))
with
β∗
H 0 (X, i∗ ΩpPn (t(q + 1))) −→ H 0 (X, ΩpX (t(q + 1))).
We show that α∗ and β ∗ are surjective for t ≥ s. For α∗ this follows immediately
by twisting the short exact sequence (12) and considering the induced long exact
sequence of cohomology, taking into account that H 1 (Pn , I ⊗ΩpPn (t(q +1))) = 0,
since
reg (I ⊗ ΩpPn ) ≤ reg (I) + p · reg (ΩPn ) ≤ e(D − 1) + 1 + 2p ≤ s + 1.
We can prove that β ∗ is surjective by the same method, once we bound the
regularity of ker β. By [32, II, Exercise 5.16] the exact sequence
0 → I/I 2 → i∗ ΩPn → ΩX → 0
induces a filtration of locally free sheaves on X
p
^
(i∗ ΩPn ) = F 0 ⊇ F 1 ⊇ · · · ⊇ F p ⊇ F p+1 = 0

and for all 0 ≤ j ≤ p an exact sequence

j
^
0→F j+1 j
→F → I/I 2 ⊗ Ωp−j
X →0 (14)

16
 Vp ∗
In particular, for j = 0 this sequence coincides with (13), since (i ΩPn ) =
i∗ ΩpPn [32, II, Exercise 5.16], so that ker β ≃ F 1 .
The same calculation as in the proof of Theorem 3.4 using Proposition 3.14
and Corollary 3.13 shows
j
^

reg I/I 2 ⊗ Ωp−j
X ≤ p(em + 1)D ≤ s.

Furthermore, by Remark 3.2 (ii) the sequence (14) yields

reg (F j ) ≤ max{reg (F j+1 ), s},

which inductively implies reg (F j ) ≤ s for all j, in particular reg (ker β) ≤ s.

It follows from the Lemma that each element of K p = Γ(X, ΩpX ((s + p)V ))
is the restriction of a form in Ωp := Γ(Pn , ΩpPn (tVe )).
We identify Cn ≃ {X0 6= 0} ⊆ Pn and set X 0 := X \ Z(X0 ). As with
polynomials one can dehomogenize a homogeneous differential form α on Cn+1
by setting X0 = 1 and dX0 = 0 to get a form α0 on Cn . Hence for ω ∈ Ωp one
gets a regular form ω 0 on Cn \ Z(f 0 ). Its restriction defines a regular form on
the dense open subset U = X 0 \ Z(f 0 ) of X 0 .
We useTthe algorithm of Szántó [56, 57] to compute a decomposition I :=
I(X 0 ) = j Ij , where each Ij is the saturated ideal of a squarefree regular
chain Gj . Note that Z(Ij ) is equidimensional. We will construct for all j a
linear system of equations describing the identity ω = 0 on Z(Ij ) for ω ∈ K p .
For simplicity we assume that I is represented by a single G = {g1 , . . . , ge }.
In the general case one only has to combine all the linear systems to one large
system.
Let k ∈ N. In [12] we have constructed a linear system of equations

prem k (f, G) = 0 (15)

in the coefficients of f ∈ C[X1 , . . . , Xn ], whose solution space is I≤k , the set of

polynomials of degree ≤ k vanishing on X 0 .
Szántó’s algorithm also yields a polynomial h which is a non-zerodivisor
mod I, and such that the module of differentials on X 0 \ Z(h) is the free module
generated by m of the dXj , where m = dim X 0 . More precisely, let X1 , . . . , Xm
denote the free variables, and Y1 , . . . , Ye the dependent variables, where m + e =
n. For a polynomial F ∈ C[X1 , . . . , Xm , Y1 , . . . , Ye ] we denoteLF := F mod I ∈
m
C[X 0 ]. Then by [12, Propositon 3.13] we have ΩC[X 0 ]h /C = i=1 C[X 0 ]h dX i .
Furthermore, for all F ∈ C[X1 , . . . , Xm , Y1 , . . . , Ye ]
m  −1 !
X ∂F ∂F ∂g ∂g
dF = − dX i , (16)
i=1
∂Xi ∂Y ∂Y ∂Xi

∂g
where g := (g1 , . . . , ge )T . Note that h is a multiple of det( ∂Y ), so that the
∂g −1
entries of ( ∂Y ) lie in h C[X1 , . . . , Xn ]. Using (16) for the coordinates Yj ,
−1

17
one can write the restriction of a form ω ∈ Ωp to U \ Z(h) in terms of the free
generators of ΩpC[U ]h /C , which are dX i1 ∧· · ·∧dX ip , where 1 ≤ i1 < · · · < ip ≤ m.
ωh
It follows that ω = h(f 0 )t , where

X
ωh = (ωh )i1 ···ip dX i1 ∧ · · · ∧ dX ip ∈ ΩpC[X 0 ]/C .
1≤i1 <···<ip ≤m

Then, since U \ Z(h) is dense in U and in X 0 , we have

ω = 0 on U ⇐⇒ ωh = 0 on U \ Z(h)
⇐⇒ ∀i1 < · · · < ip : (ωh )i1 ···ip = 0 on U \ Z(h)
⇐⇒ ∀i1 < · · · < ip : (ωh )i1 ···ip ∈ I≤k (17)

for sufficiently large k.

Now we compute the cohomology of K • . First note that the contraction
with the Euler vector field ∆ (cf. §2.2) can be easily computed, so that we can
compute a basis for Ωp . Consider the commutative diagram

d
Ωp / Ωp+1

π π
 
d / K p+1 ,
Kp

where π is the restriction of forms to U . Let N p := ker(π : Ωp → K p ). Ac-

cording to (17) and (15), N p is the solution set of a linear system of equations.
Since π is surjective, we have K p ≃ Ωp /N p , thus K p can be identified with any
complementary subspace of N p in Ωp . So we compute a basis of N p and extend
it to a basis of Ωp to get a basis of K p via this identification. The differential
d : K p → K p+1 is just the restriction of the differential d : Ωp → Ωp+1 , which
we can evaluate efficiently. Hence we can compute the matrix of d : K p → K p+1
with respect to the computed bases of K • . By computing kernel and image of
this matrix and taking their quotient we get the cohomology of K • .
Proposition 4.4. Under the notations and assumptions of Theorem 4.2, let
X ⊆ Pn be given by equations of degree ≤ d. Then one can compute the coho-
4
•
mology HdR (U ) in parallel time (d log n)O(1) and sequential time dO(n ) .
Proof. It remains to analyze the algorithm described above. Let δ denote the
maximal degree of the polynomials in the squarefree regular chain G. Then
the system (15) has asymptotic size O((nkδ e )n ) and can be computed in par-
4
allel time (n log(kδ))O(1) and sequential time (kδ)O(n ) [12]. Since the nu-
merator of each ω ∈ Ωp has degree (s + p)(q + 1), the dimension of Ωp is


n+1 (s+p)(q+1)−p+n


p n = O(sn nO(n) ). Furthermore, for ω ∈ Ωp , the degree of
the coefficients of ωh is bounded by (s + p)(q + 1) + p(e + 1)δ, hence we must
choose the k in (17) of that order. Thus, N p is described by a linear system of

18
equations of size O(nn ((s + n)n + n2 δ)n δ en ) ≤ nO(n) sn δ (e+1)n . Now let X be
given by equations of degree d. According to Theorem 4.2, we have to choose s
2
of order n3 deg X ≤ n3 dn . Furthermore, by [57] we have δ = dO(n ) . Hence
4
the size of this system is dO(n ) . The algorithms of §2.7 imply the claimed
bounds.

5 Patching Cohomologies
Let X be a smooth projective variety of dimension at most m ≥ 1. Our aim is
to compute the de Rham cohomology of X by way of an open affine cover.
So let H0 , . . . , Hm ⊆ X be hyperplane sections with H0 ∩ · · · ∩ Hm = ∅ and
set Ui := X \ Hi . Then U := {Ui | 0 ≤ i ≤ m} is an open affine cover of X.
Consider the Čech double complex C •,• := C •,• (U , Ω•X ) asL defined in §2.5. Recall
that with Ui0 ···iq = Ui0 ∩· · ·∩Uiq we have C p,q (U , Ω•X ) = i0 <···<iq ΩpX (Ui0 ···iq ).
Since U is a Leray cover for all the sheaves ΩpX , we have
Lemma 5.1. HdR •
(X) ≃ H• (U , Ω•X ) = H • (tot• (C •,• )), where tot• (C •,• ) de-
notes the total complex associated to C •,• .
To compute this cohomology, we replace the infinite dimensional double
complex C •,• by a finite dimensional one, which is built from the complex of
the last section for each Ui0 ···iq . More precisely, let e, D have the meanings of
Theorem 4.2, and choose s ≥ m(em + 1)D. For a hypersurface V in X we
denote K p (V ) := Γ(X, ΩpX ((s + p)V )). This corresponds to the complex Ks•
from Theorem 4.2. Now we define the double complex
M
K p,q := K p (Hi0 ∪ · · · ∪ Hiq )
i0 <···<iq

together with the differential δ : K p,q → K p,q+1 , which is the restriction of the
Čech differential (3), and the exterior differential d : K p,q → K p+1,q . Then K •,•
is a subcomplex of C •,• .
•
Lemma 5.2. We have HdR (X) ≃ H • (tot• (K •,• )).
Proof. Clearly, the inclusion K •,• ֒→ C •,• induces a morphism of spectral se-
quences ′′ Er (K •,• ) → ′′ Er (C •,• ) between the second spectral sequences of these
double complexes. Theorem 4.2 implies that this is an isomorphism
M
′′ p,q p
E1 (K •,• ) ≃ HdR (Ui0 ···iq ) = ′′ E1p,q (C •,• )
i0 <···<iq

on the first level of these spectral sequences. According to [41, Theorem 3.5],
this induces an isomorphism on their ∞-terms and, since the corresponding
filtrations are bounded, also on the cohomologies of the total complexes, so
H • (tot• (K •,• )) ≃ H • (tot• (C •,• )). Together with Lemma 5.1 this completes
the proof.

19
Proposition 5.3. Assume that one is given homogeneous polynomials of degree
at most d defining the smooth variety X S⊆ Pn , and linear forms defining the
hyperplane sections H0 , . . . , Hm such that i Hi has normal crossings. Then one
4
can compute HdR •
(X) in parallel time (d log n)O(1) and sequential time dO(n ) .
Proof. By Lemma 5.2 Lone has to compute the cohomology of the total complex
T k := totk (K •,• ) = p+q=k K p,q with the differential

dT : T k → T k+1 , (ωp,q )p+q=k 7→ (dωp−1,q + (−1)p δωp,q−1 )p+q=k+1 .

As in §4.2 one can compute bases for K p,q and hence for T • within the claimed
bounds. Since the differential dT is easily computable and the vector spaces
4
have dimension dO(n ) , the cohomology of this complex can be computed using
the algorithms of §2.7.
We conclude this section with an
Example 5.4. Consider the plane curve X = Z(X03 + X13 + X23 ) ⊆ P2 . One easily
checks that X is smooth, and the well known genus formula shows that its Betti
numbers are
b0 (X) = b2 (X) = 1, b1 (X) = 2.
Furthermore, the hyperplanes H e 0 := Z(X0 ) and H e 1 := Z(X1 ) intersect X
transversally, and their hyperplane sections H0 := ZX (X0 ) and H1 := ZX (X1 )
do not intersect. Hence, Ui := X \ Hi , i ∈ {0, 1}, form an open cover of X
satisfying our assumptions.
Our theorems yield the lower bound s ≥ 6, but let us find the smallest pos-
sible s. The bounds of Theorem 3.4 are reg (OX ) ≤ 2 and reg (ΩX ) ≤ 6, but the
more precise Proposition 3.14 yields reg (ΩX ) ≤ 5. However, computations with
Macaulay2 [28] give H 1 (OX (1)) = H 1 (ΩX (1)) = 0 and H 1 (OX ) 6= 0, H 1 (ΩX ) 6=
0, hence reg (OX ) = reg (ΩX ) = 2 (since dim X = 1). A close look at the proofs
shows that the conclusions of Theorem 4.2 and Lemma 5.2 hold for s = 1.
But Macaulay2 also computes dim H 0 (ΩP2 (2)) = 3 and dim H 0 (ΩX (2)) = 6,
so the restriction map in Lemma 4.3 cannot be surjective. In order to get sur-
jectivity, note that reg (I) = 3, thus the restriction map Γ(Pn , OPn (tH e 0 )) →
Γ(X, OX (tH0 )) is surjective for t ≥ 2. As for the one-forms, we compute
reg (I ⊗ ΩP2 ) = reg (I/I 2 ) = 5. It follows that the corresponding restriction
map is surjective for t ≥ 4, so we can choose s = 3.
The double complex we have to consider is
d1 / K 1,1
K 0,1
O O
δ0 δ1

d0 / K 1,0 ,
K 0,0

where
K 0,0 = K 0 (H0 ) ⊕ K 0 (H1 ) = Γ(X, OX (3H0 )) ⊕ Γ(X, OX (3H1 )),

20
 K 1,0 = K 1 (H0 ) ⊕ K 1 (H1 ) = Γ(X, ΩX (4H0 )) ⊕ Γ(X, ΩX (4H1 )),
K 0,1 = K 0 (H0 ∪ H1 ) = Γ(X, OX (3H0 + 3H1 )),
K 1,1 = K 1 (H0 ∪ H1 ) = Γ(X, ΩX (4H0 + 4H1 )),
together with the differentials

δ 0 (f, g) = g − f, d0 (f, g) = (df, dg), δ 1 (ω, η) = η − ω, d1 (f ) = df.

The dimensions of these vector spaces are

dim K 0,0 = dim K 0,1 = 18, dim K 1,0 = dim K 1,1 = 24.

Put Ωij = Xi dXj − Xj dXi (for simplicity we write Xi instead of X i ). Then,

bases for these spaces are given by
1
K 0 (H0 ) : (X 3 , X 2 X1 , X02 X2 , X0 X12 , X0 X1 X2 , X0 X22 , X13 , X12 X2 , X1 X22 )
X03 0 0
1
K 0 (H1 ) : (X 3 , X 2 X1 , X02 X2 , X0 X12 , X0 X1 X2 , X0 X22 , X13 , X12 X2 , X1 X22 )
X13 0 0
1
K 1 (H0 ) : (X 2 Ω01 , X0 X1 Ω01 , X0 X2 Ω01 , X12 Ω01 , X1 X2 Ω01 , X22 Ω01 ,
X04 0
X02 Ω02 , X0 X1 Ω02 , X0 X2 Ω02 , X12 Ω02 , X1 X2 Ω02 , X1 X2 Ω12 )
1
K 1 (H1 ) : (X 2 Ω01 , X0 X1 Ω01 , X0 X2 Ω01 , X12 Ω01 , X1 X2 Ω01 , X22 Ω01 ,
X14 0
X02 Ω02 , X0 X1 Ω02 , X0 X2 Ω02 , X12 Ω02 , X1 X2 Ω02 , X1 X2 Ω12 )
K 0 (H0 ∪ H1 ) :
1
(X 6 , X 5 X1 , X05 X2 , X04 X12 , X04 X1 X2 , X04 X22 , X03 X13 , X03 X12 X2 ,
(X0 X1 )3 0 0
X03 X1 X22 , X02 X14 , X02 X13 X2 , X02 X12 X22 , X0 X15 , X0 X14 X2 ,
X0 X13 X22 , X16 , X15 X2 , X14 X22 )
K 1 (H0 ∪ H1 ) :
Ω01
(X 6 , X 5 X1 , X05 X2 , X04 X12 , X04 X1 X2 , X04 X22 , X03 X13 , X03 X12 X2 ,
(X0 X1 )4 0 0
X03 X1 X22 , X02 X14 , X02 X13 X2 , X02 X12 X22 , X0 X15 , X0 X14 X2 ,
X0 X13 X22 , X16 , X15 X2 , X14 X22 ),
Ω02
(X 6 , X 5 X1 , X05 X2 , X04 X12 , X04 X1 X2 , X03 X12 X2 )
(X0 X1 )4 0 0

The corresponding total complex T • = tot• (K •,• ) is

dtot,0 dtot,1
T 0 = K 0,0 −→ T 1 = K 0,1 ⊕ K 1,0 −→ T 2 = K 1,1 ,

21
where
dtot,0 = (δ 0 , d0 ), dtot,1 = d1 − δ 1 .
The matrix of dtot,0 with respect to the given bases is
 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 
 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 
 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 
 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 .
 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 −3 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 −2 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 −2 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −2 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 −1 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2

22
The matrix of dtot,1 is the transpose of
 −3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 −2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 
 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 −2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 
 0 0 0 0 0 −2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 −2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 
 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 −2 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 −1 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 −2 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 3 0 1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 −1 
 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 .
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 −1 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 −1 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 −1 0 
 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 −1 
 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 1 0 0 0 
 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 
 0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
−1
0
0
−1
0
0
0
0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0
0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1

Since the rank of dtot,1 is 23, HdR

2
(X) = H 2 (T • ) has the right dimension, and
since the seventh column of its matrix is zero, the corresponding basis vector is
2
not in the image, hence HdR (X) is spanned by

Ω01 dX1 dX0

= − ∈ K 1 (H0 ∪ H1 ).
X0 X1 X1 X0

23
The nullity of dtot,1 is 19, and a basis for its kernel is given by the columns of
 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 
 0 0 0 0 −1 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 
 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 
 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 1 0 0 0 0 
 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 
 0 0 −1 0 0 −1 0 0 0 0 0 0 0 −1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 
 0 0 2 0 0 2 0 0 0 0 0 0 0 1 0 0 0 0 0 .
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 
 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 −2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 
 0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

The rank of dtot,0 is 17, and by some rank computations one sees that the
third and fourth column of the above matrix are a basis of the first cohomology
1
HdR (X) = H 1 (T • ), which correspond to
   
X1 X2 X2 Ω01 Ω02 X22 X1 X2 Ω12 X0 X2 Ω02
− , 2 , , − , − ,
X02 X03 X12 X0 X1 X04 X14

∈ K 0 (H0 ∪ H1 ) ⊕ K 1 (H0 ) ⊕ K 1 (H1 ).

Finally, the nullity of dtot,0 is 1, and the third last column cancels with the first,
0
so HdR (X) = H 0 (T • ) is generated by

(1, 1) ∈ K 0 (H0 ) ⊕ K 0 (H1 ).

Now let us see if we can find the Hodge decomposition

1
HdR (X) ≃ H 1 (OX ) ⊕ H 0 (ΩX )

in our double complex K •,• . For that purpose, note that δ 0 is the submatrix
of dtot,1 above the line, and δ 1 is the transpose of the submatrix of dtot,1 below
the line. Since the 12th row of δ 0 is zero, H 1 (OX ) is spanned by

X22
∈ K 0 (H0 ∪ H1 ).
X0 X1

24
Furthermore, one finds that the kernel of δ 1 and hence H 0 (ΩX ) is generated by
   
X2 Ω01 − X1 Ω02 Ω02 −Ω12 Ω02
, = , ∈ K 1 (H0 ) ⊕ K 1 (H1 ).
X33 X12 X02 X12

So indeed, we find generaters for the sheaf cohomologies of O

 X and ΩX in our
p,• −Ω12 Ω02
vertical subcomplexes K . It is interesting to note that 0, X 2 , X 2 can
0 1
1 X2
also be chosen as one vector in a basis of HdR (X), but ( X0 X2
1
, 0, 0) cannot. In
fact, it is easy to see that any element of the form (f, 0, 0) in the kernel of dtot,1
is in the image of dtot,0 .

6 Testing Smoothness
In this section we describe how one can test in parallel polynomial time whether
a closed projective variety X is smooth.
Crucial is Proposition 3.8 implying that if X is smooth, then X is scheme-
theoretically cut out by hypersurfaces of degree ≤ D = deg X. Using the linear
system of equations (15) one can compute a vector space basis f1 , . . . , fN of
I≤D , where I = I(X). Let Ui be the open subset X \ Z(Xi ) for 0 ≤ i ≤ n.
Then (cf. §2.1) the tangent space of X at each x ∈ Ui is

Tx X = Z(dx f1i , . . . , dx fN
i
) ⊆ Cn ≃ Pn \ Z(Xi ).

Hence, if we assume X to be m-equidimensional and denote with Lix the linear

i
subspace Z(dx f1i , . . . , dx fN ), we have
^
X smooth ⇐⇒ ∀ x ∈ Ui dim Lix = m, (18)
i

which is the Jacobian criterion. Indeed, if X is not smooth at x ∈ Ui , then

dim Lix ≥ dim Tx X > m, since in general Tx X ⊆ Lix .
Now our algorithm reads as follows.

Algorithm Smoothness Test

input X given by homogeneous polynomials of degree ≤ d.

1. Compute the equidimensional decomposition X = Z0 ∪· · ·∪Zn , where Zm

is either empty or m-equidimensional.

2. if Zn 6= ∅ then output “Yes”.

3. for 0 ≤ m < m′ < n do if Zm ∩ Zm′ 6= ∅ then output “No”.

4. Set D := dn .

5. Compute a basis f1 , . . . , fN of I≤D , where I = I(X).

25
 6. for 0 ≤ m < n do
7. for 0 ≤ i ≤ n do
 
∂fνi
8. Compute the matrix A := ∂Xµ , where fνi is the dehomogeniza-
µ,ν
tion of fν with respect to Xi .
9. Compute the Mulmuley polynomial p(Z) of A (see §2.7), which lies
in C[X0 , . . . , X ci , . . . , Xn , T, Z]. Write p(Z) = p0 + p1 Z + · · · + pK Z K ,
and let F1 , . . . , FL ∈ C[X0 , . . . , X ci , . . . , Xn ] be the coefficients of all T k
in p0 , . . . , pm .
10. if Zm ∩ Z(F1 , . . . , FL ) ∩ {Xi 6= 0} 6= ∅ then output “No”.
11. output “Yes”.
Proposition 6.1. The algorithm Smoothness Test is correct and can be imple-
4
mented in parallel time (n log d)O(1) and sequential time dO(n ) .
Proof. Correctness: If X is smooth, then it clearly passes the test in step 3.
By (18) we have dim Lix = m for all m, i and all x ∈ Zm . Denote by Ax the
matrix A evaluated at x, and similarly for p(Z). Then Lix = ker Ax , hence

dim Lix > m ⇐⇒ Z m+1 | px (Z) ⇐⇒ F1 (x) = · · · = FL (x) = 0.

If X is not smooth, then it doesn’t pass the test in step 3 or some Zm is not
smooth. In the latter case at some point x ∈ Zm ∩ Ui we will have dim Lix > m.
Analysis: All the algorithms we use are well-parallelizable. We therefore
state only the sequential time bounds. The equidimensional decomposition in
2
step 1 can be done in time dO(n ) with the algorithm of [27]. For each m, this
2
algorithm returns dO(n ) polynomials of degree bounded by deg Zm = O(dn )
whose zero set is Zm . Testing feasibility of a system of r homogeneous equations
of degree d can be done in time r(nd)O(n) using the effective homogeneous
2
Nullstellensatz. Hence step 3 takes time dO(n ) . Szántó’s algorithm in step 5
4 2
runs in time dO(n ) , and clearly N = O(Dn+1 ) = dO(n ) . Furthermore, the
3
computation of the Mulmuley polynomial in step 9 can be done in time dO(n )
2 2
by §2.7, and we have L = O(N 2 m) = dO(n ) and deg Fi ≤ N D = dO(n ) . Thus
4
step 10 takes time dO(n ) by the affine effective Nullstellensatz.

7 Finding Generic Hyperplanes

The algorithm for computing the cohomology of X described in §4.2 and §5
depends on a choice of sufficiently generic hyperplane sections Hν of X. More
precisely, it is required that V := H0 ∪ · · · ∪ Hm is a hypersurface with normal
crossings in X, where m = dim X. Note that as a consequence Hi0 ∪ · · · ∪ Hiq
has normal crossings for each tuple i0 < · · · < iq . Here we describe how to find
sufficiently generic hyperplanes deterministically in parallel polynomial time.

26
 Throughout this section we assume X to be smooth, and let us first assume
that X is m-equidimensional. We will formulate a sufficient condition for normal
crossings in terms of transversality. Recall that a linear subspace L ⊆ Pn is
called transversal to X in x ∈ X ∩ L, written X ⋔x L, iff dim(Tx X ∩ Tx L) =
dim Tx X + dim Tx L − n. Now let the hypersurfaces H0 , . . . , Hm be given by the
linear forms ℓ0 , . . . , ℓm ∈ C[X0 , . . . , Xn ]. Denote Li0 ···iq := Z(ℓi0 , . . . , ℓiq ) for all
0 ≤ q ≤ m and all 0 ≤ i0 < · · · < iq ≤ m.

Lemma 7.1. If ℓ0 , . . . , ℓm are linearly independent and for all 0 ≤ q ≤ m and

all 0 ≤ i0 < · · · < iq ≤ m we have

∀x ∈ X ∩ Li0 ···iq : X ⋔x Li0 ···iq , (19)

then V ⊆ X is a hypersurface with normal crossings.

Proof. Suppose that the condition (19) holds. First note by choosing q = 0
that Li = Z(ℓi ) is transversal to X at all x, thus V is indeed a hypersurface.
Furthermore, Hi = X ∩ Li is smooth in x, so that x lies in only one irreducible
component of Hi , and ℓi ∈ OX,x is a local equation of that component. By
transversality we have dim(Tx X ∩ Tx Li0 ···iq ) = m − q − 1. But Tx X ∩ Tx Li0 ···iq
is the kernel of the linear map ϕ := (dx ℓi0 , . . . , dx ℓiq ) : Tx X → Cq+1 , which thus
must be surjective. Hence dx ℓi0 , . . . , dx ℓiq are linearly independent on Tx X,
which proves the claim.

In order to work with condition (19) algorithmically, we introduce some

notation. Set I := I(X) and D := deg X. Recall from §6 that if f1 , . . . , fN is a
vector space basis of I≤D , then

Tx X = Z(dx f1i , . . . , dx fN
i
) ⊆ Cn

for all x ∈ Ui = X \ Z(Xi ) and 0 ≤ i ≤ n. For each tuple i0 < · · · < iq and
each i we define the matrix
 
dx f1i
 .. 
 . 
 
i 
 dx f N
i 
Ai0 ···iq :=   c
i  ∈ C[X0 , . . . , Xi , . . . , Xn ]
(N +q+1)×n
. (20)
 dx ℓi0 
 .. 
 . 
dx ℓiiq

Then the kernel of Aii0 ···iq is the kernel of ϕ of the proof of Lemma 7.1. Assume
that ℓ0 , . . . , ℓm are linearly independent. Then condition (19) is eqivalent to the
statement that the nullity of Aii0 ···iq is m − q − 1, its minimal possible value,
at each point x ∈ Ui ∩ Li0 ···iq . Note that this condition also implies the linear
independence. Now let p(Z) be the Mulmuley polynomial of Aii0 ···iq , which
ci , . . . , Xn , T, Z]. Let F1 , . . . , FL ∈ C[X0 , . . . , X
lies in C[X0 , . . . , X ci , . . . , Xn ] be

27
the coefficients of all T k in the coefficient of Z m−q in p(Z). Then a sufficient
condition for (19) is
^
Ui ∩ Li0 ···iq ∩ Z(F1 , . . . , FL ) 6= ∅. (21)
i

Using this formula we can prove

Proposition 7.2. Given polynomials of degree ≤ d defining a smooth subva-

riety X ⊆ Pn of dimension m, one can compute in parallel time (n log d)O(1)
4 S
and sequential time dO(n ) linear forms ℓ0 , . . . , ℓm such that V = j Hj is a
hypersurface with normal crossings, where Hj = ZX (ℓj ).
Proof. First we set D := dn and compute a basis f1 , . . . , fN of ID . Then we
compute the equidimensional components Zm of X. We find the linear forms
ℓ0 , . . . , ℓm successively, one at a time. So assume that ℓ0 , . . . , ℓj−1 have been
already found, and take ℓj = α0 X0 + · · · + αn Xn with indeterminate coefficients
α = (α0 , . . . , αn ). Now consider the conjunction of the conditions (21) for all
m ≤ dim X and i0 < · · · < iq = j, which is a first order formula with free
variables α. Note that here one has to take Ui = Zm ∩ {Xi 6= 0}. By quantifier
elimination compute an equivalent quantifier-free formula Φ(α) in disjunctive
normal form. Let G1 , . . . , GM be all polynomials accuring Φ(α). Since Φ(α) is
satisfied for generic α, it is easy to see that Gν (α) 6= 0 for all ν implies Φ(α) [13,
proof of Theorem 3.8]. Let δ be the maximal degree of the Gν . Now take a
set S ⊆ C of cardinality > M nδ and test for all b ∈ S in parallel, whether
Pν (b) := Gν (1, b, . . . , bn ) 6= 0 for all 1 ≤ ν ≤ M . Since Pν is a univariate
polynomial of degree ≤ nδ, there must exist a successful b. Then we can take
ℓj = X0 + bX1 + . . . + bn Xn .
Analysis: The computation of f1 , . . . , fN , of the equidimensional decompo-
sition, and of the Mulmuley polynomial can be done within the claimed time
bounds. Recall that N, L, the degrees of the defining equations for Zm , as well
2
as deg Fi are of order dO(n ) . Condition (21) is a universal first order formula
2
with O(n) free and bounded variables, and dO(n ) atomic formulas involving
2
polynomials of degree dO(n ) . According to [23], one can eliminate the universal
quantifier and hence compute the polynomials Gν in parallel time (n log d)O(1)
4 4
and sequential time dO(n ) . Furthermore, M and δ are also bounded by dO(n ) .
4
Hence the cardinality of the set S is dO(n ) and our claim follows.

Theorem 1.1 follows from the Propositions 6.1, 7.2, and 5.3.

References
[1] D. Arapura. Frobenius amplitude and strong vanishing theorems for vector
bundles. Duke Math. J., 121(2):231–267, 2004. With an appendix by Dennis
S. Keeler.

28
 [2] M.F. Atiyah, R. Bott, and L. Gȧrding. Lacunas for hyperbolic differential
operators with constant coefficients. II. Acta Math., 131:145–206, 1973.

[3] S. Basu. Computing the first few Betti numbers of semi-algebraic sets in
single exponential time. J. Symbolic Comput., 41(10):1125–1154, 2006.

[4] S. Basu. Algorithmic semi-algebraic geometry and topology – recent

progress and open problems. In Surveys on discrete and computational
geometry, volume 453 of Contemp. Math., pages 139–212. Amer. Math.
Soc., Providence, RI, 2008.

[5] D. Bayer and D. Mumford. What can be computed in algebraic geometry?

In Computational algebraic geometry and commutative algebra (Cortona,
1991), Sympos. Math., XXXIV, pages 1–48. Cambridge Univ. Press, Cam-
bridge, 1993.

[6] D. Bayer and M. Stillman. A criterion for detecting m-regularity. Invent.

Math., 87:1–11, 1987.

[7] S.J. Berkowitz. On computing the determinant in small parallel time using
a small number of processors. Inf. Process. Lett., 18(3):147–150, 1984.

[8] A. Bertram, L. Ein, and R. Lazarsfeld. Vanishing theorems, a theorem

of Severi, and the equations defining projective varieties. J. Amer. Math.
Soc., 4(3):587–602, 1991.

[9] N. Bourbaki. Elements of mathematics. Algebra, Part I: Chapters 1-3.

Hermann, Paris, 1974. Translated from the French.

[10] W.D. Brownawell. Bounds for the degrees in the Nullstellensatz. Ann. of
Math. (2), 126(3):577–591, 1987.

[11] P. Bürgisser and F. Cucker. Variations by complexity theorists on three

themes of Euler, Bézout, Betti, and Poincaré. In J. Krajı́ček, editor, Com-
plexity of computations and proofs, volume 13 of Quaderni di Matematica
[Mathematics Series], pages 73–152. Department of Mathematics, Seconda
Università di Napoli, Caserta, 2004.

[12] P. Bürgisser and P. Scheiblechner. On the complexity of counting compo-

nents of algebraic varieties. J. Symb. Comp., 44(9):1114–1136, 2009.

[13] P. Bürgisser and P. Scheiblechner. Counting irreducible components of

complex algebraic varieties. Comp. Compl., 19(1):1–35, 2010.

[14] L. Caniglia, A. Galligo, and J. Heintz. Equations for the projective closure
and effective Nullstellensatz. Discr. Appl. Math., 33(1-3):11–23, 1991.

[15] P. Deligne. Équations différentielles à points singuliers réguliers. Lecture

Notes in Mathematics, Vol. 163. Springer-Verlag, Berlin, 1970.

29
[16] P. Deligne and A. Dimca. Filtrations de Hodge et par l’ordre du pôle
pour les hypersurfaces singulières. Annales scientifiques de l’École Normale
Supérieure, 23(4):645–656, 1990.
[17] A. Dimca. Singularities and Topology of Hypersurfaces. Universitext.
Springer Verlag, 1992.

[18] D. Eisenbud. Commutative Algebra with a View Toward Algebraic Geom-

etry, volume 150 of Graduate Texts in Mathematics. Springer-Verlag, New
York, 1995.

[19] D. Eisenbud, G. Fløystad, and F.-O. Schreyer. Sheaf cohomology and free
resolutions over exterior algebras. Trans. Amer. Math. Soc., 355(11):4397–
4426 (electronic), 2003.

[20] D. Eisenbud and S. Goto. Linear free resolutions and minimal multiplicity.
J. Algebra, 88(1):89–133, 1984.

[21] D. Eisenbud, D. Grayson, M. Stillman, and B. Sturmfels (Eds.). Computa-

tions in algebraic geometry with Macaulay 2. Number 8 in Algorithms and
Computations in Mathematics. Springer Verlag, 2001.

[22] N. Fitchas and A. Galligo. Nullstellensatz effectif et conjecture de Serre

(théorème de Quillen-Suslin) pour le calcul formel. Math. Nachr., 149:231–
253, 1990.

[23] N. Fitchas, A. Galligo, and J. Morgenstern. Precise sequential and parallel

complexity bounds for quantifier elimination over algebraically closed fields.
J. Pure Appl. Alg., 67:1–14, 1990.

[24] A. Galligo. Théorème de division et stabilité en géométrie analytique locale.

Ann. Inst. Fourier (Grenoble), 29(2):vii, 107–184, 1979.

[25] D. Giaimo. On the Castelnuovo-Mumford regularity of connected curves.

Trans. Amer. Math. Soc., 358(1):267–284 (electronic), 2006.

[26] M. Giusti. Some effectivity problems in polynomial ideal theory. In EU-

ROSAM ’84: Proceedings of the International Symposium on Symbolic
and Algebraic Computation, pages 159–171, London, UK, 1984. Springer-
Verlag.

[27] M. Giusti and J. Heintz. Algorithmes -disons rapides- pour la

décomposition d’une variété algébrique en composantes irréductibles et
équidimensionnelles. In T. Mora C. Traverso, editor, Effective Methods
in Algebraic Geometry (Proceedings of MEGA’90), volume 94 of Progress
in Math., pages 169–193, New York, NY, USA, 1991. Birkhäuser.

[28] D.R. Grayson and M.E. Stillman. Macaulay2, a software

system for research in algebraic geometry. Available at
http://www.math.uiuc.edu/Macaulay2/.

30
[29] P. Griffiths and J. Harris. Principles of algebraic geometry. Wiley, New
York, 1978.

[30] A. Grothendieck. On the de Rham cohomology of algebraic varieties. Pub-

lications Mathématiques IHES, 39:93–103, 1966.

[31] L. Gruson, R. Lazarsfeld, and C. Peskine. On a theorem of Castelnuovo,

and the equations defining space curves. Invent. Math., 72(3):491–506,
1983.

[32] R. Hartshorne. Algebraic geometry. Springer-Verlag, New York, 1977.

[33] Z. Jelonek. On the effective Nullstellensatz. Invent. Math., 162(1):1–17,

2005.

[34] J. Kollár. Sharp effective Nullstellensatz. J. Amer. Math. Soc., 1(4):963–

975, 1988.
[35] S. Kwak. Generic projections, the equations defining projective varieties
and Castelnuovo regularity. Math. Z., 234(3):413–434, 2000.

[36] D. Lazard. Resolution des systemes d’equations algebriques. Theoret.

Comp. Sci., 15:77–110, 1981.

[37] R. Lazarsfeld. A sharp Castelnuovo bound for smooth surfaces. Duke Math.
J., 55(2):423–429, 1987.

[38] R. Lazarsfeld. Positivity in algebraic geometry. I, volume 48 of Ergebnisse

der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Sur-
veys in Mathematics [Results in Mathematics and Related Areas. 3rd Series.
A Series of Modern Surveys in Mathematics]. Springer Verlag, Berlin, 2004.
Classical setting: line bundles and linear series.

[39] E.W. Mayr. Some complexity results for polynomial ideals. J. Compl.,
13(3):303–325, 1997.

[40] E.W. Mayr and A.R. Meyer. The complexity of the word problems for
commutative semigroups and polynomial ideals. Advances in Mathematics,
46(3):305–329, 1982.

[41] J. McCleary. A User’s Guide to Spectral Sequences, volume 12 of Mathe-

matics Lecture Series. Publish or Perish, Inc, Wilmington, Delaware, 1985.

[42] K. Mulmuley. A fast parallel algorithm to compute the rank of a matrix

over an arbitrary field. Combinatorica, 7(1):101–104, 1987.

[43] D. Mumford. Lectures on Curves on an Algebraic Surface. With a section

by G. M. Bergman, Annals of Mathematics Studies, No. 59. Princeton
University Press, Princeton, N.J., 1966.

31
[44] D. Mumford. Varieties defined by quadratic equations. In Questions on
Algebraic Varieties (C.I.M.E., III Ciclo, Varenna, 1969), pages 29–100.
Edizioni Cremonese, Rome, 1970.
[45] D. Mumford. Algebraic Geometry I: Complex Projective Varieties, volume
221 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag,
Berlin Heidelberg New York, 1976.

[46] T. Oaku and N. Takayama. An algorithm for de Rham cohomology groups

of the complement of an affine variety via D-module computation. Journal
of Pure and Applied Algebra, 139:201–233, 1999.

[47] I. Peeva and B. Sturmfels. Syzygies of codimension 2 lattice ideals. Math.

Z., 229(1):163–194, 1998.

[48] H.C. Pinkham. A Castelnuovo bound for smooth surfaces. Invent. Math.,
83(2):321–332, 1986.

[49] Z. Ran. Local differential geometry and generic projections of threefolds.

J. Differential Geom., 32(1):131–137, 1990.

[50] P. Scheiblechner. On the complexity of counting irreducible components

and computing Betti numbers of complex algebraic varieties. PhD Thesis,
2007.

[51] P. Scheiblechner. On the complexity of deciding connectedness and comput-

ing Betti numbers of a complex algebraic variety. J. Compl., 23(3):359–379,
2007.

[52] P. Scheiblechner. On a generalization of Stickelberger’s theorem. J. Symb.

Comp., 45(12):1459 – 1470, 2010. MEGA’2009.

[53] J.P. Serre. Géométrie algébrique et géométrie analytique. Ann. Inst.

Fourier, Grenoble, 6:1–42, 1955–1956.

[54] G. Smith. Computing global extension modules,. J. Symb. Comp., 29:729–

746, 2000.

[55] J. Stückrad and W. Vogel. Castelnuovo’s regularity and multiplicity. Math.

Ann., 281(3):355–368, 1998.

[56] Á. Szántó. Complexity of the Wu-Ritt decomposition. In PASCO ’97:

Proceedings of the second international symposium on Parallel symbolic
computation, pages 139–149, New York, NY, USA, 1997. ACM Press.
[57] Á. Szántó. Computation with polynomial systems. PhD Thesis, 1999.

[58] W.V. Vasconcelos. Computational methods in commutative algebra and

algebraic geometry, volume 2 of Algorithms and Computation in Mathe-
matics. Springer-Verlag, Berlin, 1998.

32
[59] C. Voisin. Hodge theory and complex algebraic geometry. I, volume 76 of
Cambridge Studies in Advanced Mathematics. Cambridge University Press,
Cambridge, 2002. Translated from the French original by Leila Schneps.
[60] J. von zur Gathen. Parallel arithmetic computations: a survey. In
MFOCS86, number 233 in LNCS, pages 93–112. SV, 1986.

[61] U. Walther. Algorithmic computation of de Rham cohomology of comple-

ments of complex affine varieties. J. Symb. Comp., 29(4-5):795–839, 2000.

[62] U. Walther. Algorithmic determination of the rational cohomology of com-

plex varieties via differential forms. In Symbolic computation: solving equa-
tions in algebra, geometry, and engineering (South Hadley, MA, 2000), vol-
ume 286 of Contemp. Math., pages 185–206. Amer. Math. Soc., Providence,
RI, 2001.

33