JOURNAL OF PURE AND APPLIED ALGEBRA
Journal of Pure and Applied Algebra 117 & 118 (1997) 277-3 17
ELSEVIER
Lower bounds for diophantine M. Giusti a,*, J. Heintzb, L.M.
approximations’
K. Htigeleb,
Pardob,
J.L.
J.E.
Moraisb,
Montaiia”
il GAGE, Centre de MathCmatiques, .&ColePolytechnique, F-91128 Palaiseau Cedex, France b Dcpartamento de Matemriticas, Estadistica y Computackin, Facultad de Ciencias, Universidad de Cantabria, E-39071 Santander, Spain c Departamento de Matemritica e Informcitica, Campus de Arrosadia, Universidad Ptiblica de Navarra, E-31006 Pamplona, Spain
Abstract We introduce a subexponential algorithm for geometric solving of multivariate polynomial equation systems whose bit complexity depends mainly on intrinsic geometric invariants of the solution set. From this algorithm, we derive a new procedure for the decision of consistency of polynomial equation systems whose bit complexity is subexponential, too. As a byproduct, we analyze the division of a polynomial modulo a reduced complete intersection ideal and from this, we obtain an intrinsic lower bound for the logarithmic height of diophantine approximations to a given solution of a zero-dimensional polynomial equation system. This result represents a multivariate version of Liouville’s classical theorem on approximation of algebraic numbers by rationals. A special feature of our procedures is their polynomial character with respect to the mentioned geometric invariants when instead of bit operations only arithmetic operations are counted at unit cost. Technically our paper relies on the use of straight-line programs as a data structure for the encoding of polynomials, on a new symbolic application of Newton’s algorithm to the Implicit Function Theorem and on a special, basis independent trace formula for affine Gorenstein algebras. @ 1997 Published by Elsevier Science B.V. 1991 Math.
Subj.
Class.:
68Q25,
14Q15,
11568
1. Introduction The present
paper represent
the design of algorithms
a continuation
of intrinsic
of [25, 26, 441. These
type to solve systems
papers concern
of polynomial
equations.
* Corresponding author. E-mail:
[email protected].
’ Research was partially supported by the following French, Spanish and Argentinian 1026 MEDICIS,
PB93-0472~C02-02,
UBACYT-EX-001
0022-4049/97/$17.00 @ 1997 Published PII SOO22-4049(97)00015-7
and PID CONICET
3949192.
by Elsevier Science B.V. All rights reserved
grants:
GDR CNRS
278
Solving
M. Giustiet al. I Journal of Pure and Applied Algebra II 7& 1161(1997) 277-317
is then applied
By “intrinsic semantical
to decide
type” we mean algorithms
and the syntactical
both for the improvement
character
equation
of systems
of polynomial
which are able to distinguish of the input system
of the complexity
With respect to bit complexity polynomial
consistency
equations. between
the
in order to profit from
estimates.
we show how the time necessary
to solve a given
system is related to the affine degree and the (affine) logarithmic
height of the corresponding diophantine variety. In this sense, the results of [25, 261 show already that the (affine) degree of an input system is associated with the complexity when measured in terms of number of arithmetic operations. However, there are still some drawbacks
to this approach.
The first one is that the algorithms procedure and that the algorithms in the case of [25] the algorithms
developed
in [26] require iterative
calls to the
in [25] rely on the use of algebraic numbers. Thus are not “rational” although their inputs and outputs
are. A second drawback concerns the modeling tity of arithmetic operations of an algorithm
used to measure complexity. does not explain sufficiently
pens when we are using it on a “real-life” computer. models of bit complexity (Turing machines, random
Years of experience show that access machines or equivalent
models) represent more realistic patterns for practical computing. of bit complexity of the intrinsic complexity of the algorithms necessary. In the present paper we deal with both disadvantages
The quanwhat hap-
In this sense, a study of [25, 261 becomes
of the algorithms
in [25,26],
giving practicable solutions (cf. also Section 3). First, our new algorithm is completely rational. It does not require any constants other than those in the field of coefficients of the input system. Secondly, to improve the bit complexity estimates, we introduce a suitable notion of height of afline diophantine varieties which is inspired by the corresponding notion introduced shown in Section 2, our notion geometric
elimination
procedures
for projective varieties in [5,9,21,42,43,45-48]. As of height is strongly related to the bit complexity of of any kind.
Our notion of height combined with a new algorithmic interpretation of duality theory for complete intersection ideals yields a new Liouville estimate (cf. Sections 7 and 4). Liouville estimates can be applied to get lower time bounds for the numerical analysis approach to solving systems of polynomial equations. In particular, we show the practical inefficiency of both floating point and binary encoding for rational approximation of zero-dimensional multivariate polynomial equation systems even in the algorithmically well-suited cases. To illustrate these introductory observations, let us first consider the following two problems. Problem 1. Let be given a sequence of n integer polynomials with coefJicients in (0, l}, using three variables each:
of degree
at most 3,
M. Giusti et al. I Journal of Pure and Applied Algebra 117&c 118 (1997) 277-317
219
Decide whether the following system of polynomial equations has a solution: X&-X,,k=O
,..., X,$-X,,,k=O,Yf-Y,=O
,...,
1
Yjn-Y3,,=0,
n
rI i=l
Problem 2. Given in decimal representation an integer k E N and the polynomials:
XT - x, = 0,. . . ,x,’ - x, = 0, k - (Xl + 2x2+. . . +2”-‘XJ
= 0.
Decide whether this system has a solution. Both problems
have a similar
“syntactical
form” (i.e., they look very similar,
as
a consistency question for “syntactically easy” polynomial equation systems). However, they possess completely different “semantical” characters: the first problem is a translation of a well-known NP-complete problem (3_satisfiability, for short 3SAT), while the second one just concerns the binary encoding of k with n bits (if there exists such an encoding). This means that the second system is consistent if and only if k may be encoded using at most n binary digits (cf. [31, 441). Traditional symbolic procedures deal with both problems using in each case the same general treatment. However these two problems demand for different algorithms that may profit from their different semantical features. The construction of algorithms which are able to distinguish between equation systems which are semantically well suited and such which are not is the main goal of our paper. We shall also consider polynomial equation systems in the following terms:
the question
of consistency
of
Problem 3 (Effective Nullstellensatz). Given a sequence of polynomials fi , . . , fS E , . . . ,X,,], decide whether the afine algebraic variety Z[Xl V(f ,,..., fs):={XE@“:f,(X)=
..’
=fs
=O}
is empty or not. As both Problems 1 and 2 above can be written as special cases of this more general Problem 3, we have to find a way to distinguish (within the context of Problem 3) between the different levels of difficulty Problems 1 and 2 represent. rithms which solve this problem taking care of the special features instance of the input system, we need two major geometric invariants: and the affine height of the system (cf. Section 2.2). Assuming these are going to show in this paper the following result.
To design algoof the particular the affine degree two notions we
280
M. Giusti et al. I Journal of Pure and Applied Algebra 117&118
(1997) 277-317
Theorem 4. There exists a bounded error probability Turing machine which solves the following task: given polynomials f,, . . . , fn+l E Z[Xl, . . .,X,1 of degree at most 2 and of (logarithmic) V( f,,
height h, the Turing machine decides whether the variety
, fn+l ) is empty or not. Moreover, iJ’ 6 is the intrinsic (affine) degree of the
system and n is the intrinsic (logarithmic) height of the system, the Turing machine answers in (bit) time
using a total amount of arithmetic operations in Q of (n6)‘(‘). Our algorithm
first solves a suitable polynomial
equation
system and then uses this
system fi,. . . , fn+l. Solving is let us assume that the polynomials fi, . . , fn form a regular sequence, each ideal (fi, . . . , fi), 1 < i 5 n, being radical. Then the algorithm proceeds in n steps, solving at each stage 1 5 i < n the system
information for the consistency test of the original done inductively. In order to explain the procedure
f, =o )...) fi=o. The corresponding intermediate algebraic varieties are obtained by a lifting process from special zero-dimensional varieties which we call the lifting jibers. The lifting process is based on a division-free symbolic version and represents a algorithmic version of the Implicit
of the Newton-Hensel algorithm Function Theorem. In this way,
the procedure constructs the lifting fiber of step i + 1 from the lifting fiber of step i. Each inductive step includes two cleaning phases: one is performed to throw away extraneous projective components (mainly those at infinity) and a second one reduces the size of the representation of the integers which appear during the process. This second cleaning
phase is included
to avoid uncontrolled
of intermediate polynomials. Finally, we show how the following division problem in the Nullstellensatz resolution
of systems of multivariate
growth of integer coefficients
two computational and the numerical
polynomial
problems are related: the analysis approach for the
equations.
Problem 5 (Division problem in the Effective Nullstellensatz). Let be given polynomials fi,. . . ,fs E Z[X~, . . . ,X,,] without any common zeroes (i.e., the polynomials satisfy
the condition
V( f,, .
, fS) = 0). Compute a non-zero integer a E Z, and poly-
nomials 91,. . . , gS E Z[X,, . . . ,X,] such that
holds. Problem 6 (Numerical analysis approach). Given a real number E > 0 and a regular sequence of polynomials fi, . . , fn E Z[X,, . . ,X,,] of degree at most d defining a zerodimensional a$fine variety V := V( f,, . . . , fn), compute for any point c(E V an
M. Giusti et al. I Journal of Pure and Applied Algebra 117& I18 (1997) 277-317
281
approximation a E @” at level E > 0, i.e., find by an eflective procedure a point a E @” such that Ila - x/I < 6 holds. At first glance Problems
5 and 6 seem to be unrelated.
However,
the solution of Problem 5 provides lower bounds for any solution shall state this observation in terms of Liouville estimates. In order to understand
the relation between
the context of Problem 6 any approximation nal (i.e., such an approximation
Problems
we shall see how of Problem
6. We
5 and 6, let us observe that in
a computer may output is necessarily
must be a point a belonging
is assumed to be encoded in binary (i.e., by the binary expansion numerators of the coordinates of a). In the past the division problem in the Nullstellensatz
ratio-
to Q[i]“). Such an output of denominators
and
was studied in both number
theory and computer science simultaneously. The first results mainly established upper bounds for the degree [lo, 13, 12, 331 and height [3,4] of the polynomials 91,. . , gS appearing in Problem 5. The complexity study in [22, 34, 351 yields optimal upper bounds for both invariants (see also [19]) and can be applied estimates in the following sense (cf. also Section 2.5). A Liouville
Liouville
estimate is a lower bound for the binary length (the logarithmic
of the numerators and denominators V, the point cx and the real number Theorem
to derive
7 gives such a Liouville
of the coordinates E in the statement
height)
of a in terms of the variety of Problem 6. The following
estimate.
Let us introduce the following notions and notations: let VcC” be a zero-dimensional diophantine variety not intersecting Q[iln and let u be a point of V. Let E > 0 be a real number.
We call a point a = (al/q, , . . . , a,/q,) E O[i]” a rational approximation
of x at
level E > 0, if
holds. (Here 11.II denotes the norm associated with the usual hermitian product in C”.) A regular sequence f, , . . . , fn E Z[&, . . . ,X,] is said to be smooth if for any 1 < i 5 n the Jacobian J( f 1,. . . , fi) is not a zero-divisor modulo the ideal (fi,. . . , f2). Observe that for a smooth regular sequence the ideals (fi , . . . , fi) are always radical. Theorem 7. There exists a universal constant C with the following property: given a smooth regular sequence of polynomials f,, . . . , fn E Z[&, . . . ,A?,,] of degree at most d and of logarithmic height at most h. Furthermore let be given a point a E Q[i]” and a real number 0 ( E 5 1. Suppose that the variety V := V(f,, . . ,fn) following conditions: - V tl Q[iln = 0. ~ there exists a point SIE V such that ((U- Xl1<
E 5
1
verifies the
282
hf. Giusti et al. I Journal of Pure and Applied Algebra I1 7& 118 (1997) 277-317
holds. Then we have for any denominator q of a the following inequality: - (h + q) 5 log, q. (ndG)C where 6 denotes the degree and ye the (afine)
Proof. According
to Proposition
E-1
logarithmic height of V.
20 below we have
q2.
lgn+l(colw4l+ 1) 5
Thus, we look for upper bounds
for IIc((J and lgn+i(C()/. First, from Proposition
14
below we deduce (21la1] + 1) < v&2”? On the other hand, combining log, lgn+i(CoI < (nd@‘(‘)(h From these two upper bounds We observe
Theorem
21 and Lemma 23 below we conclude
that
+ y + log, 4). we easily deduce the bounds
that for n = 1, Theorem
7 represents
0
of the theorem.
just a restatement
of Liouville’s
classical result with somewhat coarser bounds (cf. e.g. [50, Kapitel I, Satz 11). The height estimates in [3, 4, 19, 34, 351, combined with the methods described in Section 2.5 (Proposition 20) produce Liouville bounds that relate the syntactical description of V as given by the input, the approximation E and the height of the denominator of a. In fact, these estimates yield the inequality: - log, E < d”“‘h log, q.
(1)
On the other hand, degree and height of the solution time bounds
for the numerical
analysis
approach
set represent
to polynomial
reasonable
equation
lower
solving
by
rational point approximation. Exponential degrees and exponential height produce exponential lower time bounds for numerical methods of polynomial equation solving (cf. [44]). A lower bound like inequality (1) above, i.e., a lower bound for log, q, represents a lower bound for the output length and, therefore, a lower time bound for numerical methods of polynomial system solving. The results of [24,8,9,19,34,35,46-48] imply that an approximation level of e := 2- do’“’ is sufficient in order to characterize (and to distinguish adequately) the solutions of the variety V in Problem 5. On the other hand, the examples of [39, Chapitre 4, Proposition 121 and [29,30, Example 31 show that even in case of semantically and syntactically “easy” systems such an approximation level may be necessary. Therefore, it becomes reasonable to fix an approximation level for numerical solving of E := 2-d”, where d” represents the Btzout number of the input system. Under this assumption, lower bounds are required.
inequality
(I ) becomes
meaningless
and more precise
M. Giusti et al. IJournal of Pure and Applied Algebra 117& 118 (1997) 277-317
However
an exponential
tained in the following Corollary
lower time bound
corollary
to Theorem
for numerical
solving
283
is implicitly
con-
7.
8. Given a smooth regular sequence fi,. . . , fn E Z[X,, . . . ,_&I, and points
c(E V, a E Q[iln verifying the conditions of Theorem 7, let E > 0 be a level of approximation such that -log,
E =d”
(where d” represents
the BCzout number of the
system). Then we have d” (ndS)C
(h + YI)I log, 9,
where C is a suitable universal constant as in Theorem 7. In particular, for systems of “small” degree and height, the output length for numerical solving methods is necessarily exponential in the number of variables. For instance, polynomials:
we may consider
the following
smooth regular
sequence
of quadratic
This sequence defines a zero-dimensional variety with just two points (the degree of the variety is 2) of small height (the logarithmic height is 1). Thus, for a level of approximation E > 0 with - log, E = 2”, the lower bound obtained from inequality (1) says just: 2” -< 2O(“) log, q, whereas the lower bound level E of any solution consequence
from Corollary
8 states that every rational
of this system has exponential
of this corollary
is that both binary
bers in Q[i] are not efficient for reaching
in [52-54,56,17,18].
approximation
length.
of num-
level of approximation.
in the following
of
Thus, the main
and floating point encoding
the appropriate
alternative encoding is therefore required. Another way out of this dilemma may consist in [51] and further developed
binary
approach
Instead of approximating
An
initiated rationally
(or similarly
by floating point arithmetic) the solutions of the zero-dimensional input level E = 2Yd”, we just try to system ,fi, . . . , fn up to the appropriate approximation find approximate zeroes (in the sense of [54]) of the system, i.e., we try to find points a E Q[iln from which a suitable version of Newton’s algorithm converges quadratically to a true zero of the system.
2. Notions,
notations
and results
We are going to study the consistency question of Problem 3 (i.e., the decisional problem in the effective Nullstellensatz) only when the input system of polynomial
284
M. Giusti et al. /Journal of’ Pure and Applied Algebra 117& IICI (1997) 277-317
equations _ r
fi, . . . , fr E T&JCL,.. . ,A’,] verifies the conditions:
_ the sequence
J;,
...,fr_lis a smooth regular sequence.
These two additional an effective
version
conditions of Bertini’s
are not really restrictive. Theorem
allows
As shown in [22,24,27,34,35]
to reduce
a general
input
system
fl,..., .fs E w-l ,...,X,] to such a system. We just have to find generic Z-linear combinations
of the input polynomials
coefficients
in Z of logarithmic
such that these linear combinations
and d is an upper bound for the degrees of the input polynomials. tion of this preprocessing
we may suppose,
equations satisfy the following conditions: _ the ideals (f,, ...,fi)are radical, 1
1 the varieties
of dimension
contain
height O(n log, d), where n is the number
I$ = V(fi,.
without
Under the assump-
loss of generality,
1, . ,J;:) are complete
only
of variables
that our input
intersection
affine
n - i.
Notation 1. Given R c B an extension of a commutative ring (which converts B into a R-algebra), and an element b E B, we denote by vb : B + B the R-linear endomorphism induced by the multiplication of the elements of B by b (in the following we shall call such a linear map a homothety). If B is a free R-module of finite rank, we denote by Mb the matrix of the homothety
?Ib and by Xb E R[T] the characteristic
polynomial
of qb. Moreover, if R is a unique factorization domain, we shall denote by mb the primitive minimal polynomial of ?‘Ib. Observe that Xb and mb are manic polynomials of R[T], which respectively.
for short we will call characteristic
In order to decide whether
(fi,
compute the following items: - A linear change of coordinates
and minimal
,,...,
Y,_,+,]+O[Y,
of 6,
...,fr)represents the trivial ideal we just need to (Xl,.
. ,X,) + (Y,, . . . , Y,,) such that the following
Q-algebra homomorphism represents a Noether normalization V(,fi, , fr-, ),with the variables Yi, . . , Y+,.+l being free: R:=Q[Y
polynomial
,...,
Y,J(f
I,...,
of
the
variety
fr-,) =:B.
Observe that this means that B is an integral extension of R. _ the matrix Mf; of the homothety ~5 with respect to an appropriate R-module basis. Suppose that such a Noether normalization is available. Then the Q-algebra homomorphism R + B is injective and B is a free R-module of rank at most deg V,-, (cf. [27,28]). Under this assumption the ideal (fi, . .,fr)is trivial if and only if the matrix yf; is unimodular what means that the determinant of MI; is non-zero and belongs to Q. This comment equation systems in a very specific exactly we mean
shows how the original Problem 3 of testing consistency of polynomial can be reduced to the problem of solving polynomial equation systems geometric sense. In the next subsection we are going to explain what by this, namely geometric solving.
285
M. Giusti et al. I Journal of Pure and Applied Algebra II 7& 118 (1997) 277-317
2.1.
Geometric solving
The previous nomial
equation
considerations
puting a Noether normalization homothety
reduce the search for a consistency
system fi = 0,. . . , fr = 0, namely of the variety
~1; with respect to a suitable
Problem
test for the poly-
3, to the problem
of com-
V(J) . . . , J>_ 1) and the matrix Mb of the
R-module
basis. Assume
that Xi,. . , X, are
in Noether position with respect to the variety I’(fi, . , fr_l ), the variables being free. Then we consider the following integral ring extension: Xl,‘..,Xn-r+l already
R := C&Y, ,...,Xn-r+ll~Q~~l,...,~nll~fl,...,fr-1) Our assumptions
on fi , .
, Jr_,
imply that B is a reduced
R-module.
We are now going to explain
“geometric
solution”.
Definition
=: B.
First we need the following
9. Let R be a ring of polynomials
algebra
and a finite free
what we mean by “geometric
solving”
or
notion of primitive element of B:
over Q and R C B an integral ring exten-
sion such that B is reduced and B is a free R-module of finite rank. An element u E B is called a primitive element of the ring extension R 2 B if the degree of the minimal polynomial
m, of u equals the rank of B as free R-module,
i.e., if
deg m, = rankR B holds. Let K be the quotient field of R and B’ = K @RB be the localization of B by the non-zero elements of R. An element u E B is a primitive element of B if and only if for D:=rank,B=dimKB’ the set {l,u,...,&’ } represents a K-vector space basis of B’. The computation of the matrix Mfr is a consequence of the following “generic point” description of the K-algebra B’: this algebra is characterized by the following items (which our algorithm will compute): _ a K-vector space basis of B’; - for n - r + 2 5 i 5 n the matrices Mx, of the homotheties the given basis (these matrices
describe
the multiplication
qx, : B -+ B with respect to tensor of the K-algebra
B’ and hence also of the R-algebra B). We then obtain the matrix MfV by substituting for the variables X-,.+2,. . ,X, appearing in the polynomial f,(Xi , . . , X, ) the matrices Mx,_, i-2,. . , Mx,.(In the sequel we shall write for short MJ = f,.(M~,i_~+~,. . , Mx, ) for this substitution, interpreting ,fi as an element of the polynomial ring R[Xn_-r+~,. . ,X,,]). In this sense, geometric solving means just computing both a basis of the K-algebra B’ and the matrices M,y,_,.,, , . . , Mx~. This is done making use of a suitable primitive element of B. In the context of this paper, the primitive element u E B will always be chosen as the image in B of a generic Z-linear form of the variables &_,.+z,. ,X,. In particular, we may assume that u is the image of a linear form U = &_,.+zX~_~+Z+. . . +2,X,, with
286
hf. Giustiet al. IJournal of Pure and Applied Algebra 117& 118 (1997) 277-317
i, E L for n - r + 2 2 i < n. Let T be a new indeterminate. m,(T)
of u as an element
K-algebra
of the R-algebra
B’) will be a manic polynomial
We shall choose
this minimal
The minimal
B (or equivalently in O[Xt,. . . ,&-,.+I,
polynomial
as an element
T] = R[T]. In the sequel we shall pay special attention
Gr+l, where we have R = Z, K = Q and where m, is a polynomial
polynomial
as an element
of the
T] = Q @,zR[T]. Z[Xt,. . . ,
of the ring
to the case r = 12+ 1,
of Z[T] of positive degree
(the polynomial m, is then trivially manic over Q[T] since we may divide it by its leading coefficient, which is a non-zero integer). Discarding the content (the maximum common
divisor)
m,, by its primitive
of the coefficients counterpart
of m,, we may replace the minimal
polynomial
which we shall always denote by qu. Similarly,
case Y 5 n, we shall replace the minimal
polynomial
for the
m, E Z[&, . . . ,Xn_,.+l, T] = R[T]
by some “cleaner” (however not necessarily primitive) version qu. Finally, as { 1, u,. . ,uD-’ } is a basis of the K-vector space B’, for n - Y + 2 < i 5 n there are polynomials al” E R[T] and non-zero elements pj”) E R such that pf”‘Xi v)“‘(U) belongs to the ideal (Ji , . . , fr_ 1) in K[Xn_r+~, . ,X,1. In particular, the following identity between ideals of K[X,_,.+2,. . . ,X,1: (fl,...,
fr-1)=(4u(U),P~IU+2Xn--r+2
- $!v+2(m
..,P?‘&
we have
- $w)).
Moreover, if M denotes the companion matrix of the homothety Q, with respect to the basis { 1, U, . . . , @-I}, the matrices A~,Y_,~~,. . . , &IX, characterizing the multiplication tensor of the R-algebra B (or equivalently of the K-algebra B’) are given by the formula Mx
=
’
for n - r+2
(p!“’>-’ .p(M) I
I
5 i 5 n. To simplify notations
we shall often omit the superindex
(u) when
referring to these polynomials. After these explanations we shall assume, without loss of generality, that we have s := IZ+ 1 in the statement of Problems 3 and 5. Consequently, we restrict the meaning polynomials
of “geometric
solving” to the case where we have given as input
Ji , . . . , fn E Z[& , . . ,X,] forming a regular sequence
in Q[Xt,. . .,X,1 (we
shall say in the future for short that ft, . . . , fnEZ[X,,. . ,X,J is a regular With these conventions “geometric solving” means the following: Definition
10. An algorithm
for geometric
sequence).
solving is a procedure which from a smooth
regular sequence f,, . . , fn Ei&Y,,.. . ,X,J as input produces: _ a primitive element u of the ring extension Q -_[x,,...,x,]/(fi,..., fn) represented by a Z-linear form U = I,&, + . . . +i,X,, - the primitive minimal polynomial qu E Z[T], - the parametrizations of the variety V(fr,. . , fn) by the zeroes of qu, namely the (unique) primitive polynomials p(,u’Xt -u:“‘(T), . . . , p$f’X, -v?)(T) with py), . . . , pp’ non-zero integers and vi”‘, . . , up’ E Z[T] which satisfy the conditions max {deg$),...,degv?)}
287
M. Giusti et al. I Journal of Pure and Applied Algebra I17& 118 (1997) 277-317
In the sequel we shall refer to the polynomials (with
image
u in Q[_Xl,...,x,l/(f~,...,f~>>,
(u) PI , . . , p?’ E Z
U = Al&+
consider
we cannot
system f, = 0,.
solving”
..,fn= 0.
has a long history, which
give here in full detail for mere lack of space. One might
[ 15,231 as early references
time in complexity
E Z[X,, . . ,X,]
qu(U, v(,U’(T),...,~~‘(T)EZ[TIand
as geometric solution of the equation
Let us remark here that this notion of “geometric unfortunately
. . . +A,&
where this notion was implicitly
used for the first
theory.
2.2. Intrinsic parameters We said in the introduction that we are interested ing which are able to profit from “good” geometrical of polynomial
equations
might possess,
in algorithms for geometric solvproperties which an input system
This makes it necessary
to precise what such
“good” geometrical properties may be and how to find measures for them. Therefore, we are going to define two geometric invariants in this subsection that will arise as parameters arithmic
of the complexity
height of complete
of our procedures: intersection
the (affine) degree and the (affine) log-
varieties.
The notion of degree has been taken
directly from [28], while our notion of height is strongly inspired by the corresponding notion developed for projective varieties in [42,43,4548]. Let us first recall the notion
of degree of an affine algebraic
defining the degree of a zero-dimensional positive-dimensional complete intersection To fix notations,
variety.
We begin by
variety and then we extend this notion algebraic varieties.
let V C C” be an algebraic
subset (variety)
to
given as the set of com-
mon zeroes in C=”of a smooth regular sequence of polynomials f,, . . . , fi E i&Y,, . . ,X,] of degree at most d. If V is zero dimensional (i.e., if i = n) the degree of V is defined as the number of points of V (points at infinity are not counted in this definition). In the general
case (when dim V = n - i > 1 holds)
affine linear subspaces
of C” of dimension
i (defined
let us consider
the class 9 of all
as the set of solutions
in C” of
a linear equation system Li = 0,. . . ,L,_i = 0 where Lk = aklX,+. . +ah,X, + ako is an affine linear polynomial with coefficients akj E Z for 1 5 k
and assumptions
be as before. The degree of V is defined
of the degrees of the intersections of V with affine linear We denote the degree of V by deg V.
spaces
As observed in [28], this definition of the degree never yields infinity, but gives always a natural number. Our definition is equivalent to the following one: Consider all linear changes of coordinates in C” defined by non-singular matrices with integer entries, i.e., linear changes of the type
(Xl,...,&)
H
(YI>...,Y,),
288
M. Giusii et al. IJournal of‘ Pure and Applied Algebra 117& 118 (1997) 277-317
. +ak,X,
where Yk= ~1x1~.
is a linear form with integer coefficients
Any generic linear change of coordinates
induces an integral ring extension
~o[~l,...,r,-il-o~~I,...,ml/(fl,...,fi)
This rank is the same for any generic
equals the degree of the variety
as follows:
= Q[Vl.
It turns out that the ring O[V] is a free Q[Y,, . . . , Y,_i]-module for instance).
for 1
of finite rank (cf. [27]
linear coordinate
change
and it
V.
In order to define the height of an affine diophantine the zero-dimensional case and then the case of positive
variety, we also consider dimension.
To start with let us first say what we mean by the (logarithmic)
first
height of an integer,
a vector of integers, a matrix over Z and a polynomial with integer coefficients. Let a EL be an integer, then the height of a is defined as ht(a) := max{log, Ial, 1). It is obvious
that the height measures
the bit length of a. On the other hand, we shall
see soon that this simple notion of height of an integer has a natural extension to algebraic varieties where it plays the role of “arithmetic degree” (see [9, 5, 21, 34, 35, 46-481). The main outcome of this paper will be the reinterpretation of the notion of height for algebraic varieties as a measure for the bit complexity of an elimination procedure. In this sense, our contribution justifies a posteriori Northcott’s terminology of “complexity” for the height of an algebraic variety [57]. For a vector of integer numbers ~:=(a,, . . , a, ) E Z”, we define the height ht (a) as the maximum of the heights of its coordinates. For a matrix A E J&(Z) with integer entries, its height ht (A) is defined as the height of A as vector. Similarly, we define the height of a polynomial Definition
12. Given
f E Z[Xl,
. .,X,,] as the height of the vector of its coefficients.
a zero-dimensional
diophantine
algebraic
variety
V G C” and
a linear form U = i&l+. . . +A,& with integer coefficients representing a primitive element u of the ring extension Q + Q[V], we define the height of V with respect to U as the maximum X, - v?‘(T)
of the heights of the polynomials
(see Definition
In our first approach
qu(T),py)&
- u{~)(T),
. . . ,pt’
10) and denote this height by ht ( V; U).
to find a suitable
notion
define the height of the given zero-dimensional
of height for algebraic variety
V as the function
varieties
we
ht, : N + N
which associates to any natural number c E RJ the value htv(c) := max{ht (V; u): ht(U)Lc} if the ring extension Q + O[V] has a primitive element of height at most c and which associates to c the value 1 if no such primitive element exists. This notion of height is related to the hermitian norm and the denominator of the points of the variety V. In order to explain this relation, let us introduce the following notations. Taking into account that the zero-dimensional variety V is contained in C”, we define the norm [[VII as
11 VII :=max{llccI1 : a
E
V}.
289
M. Giusti et al. I Journal of Pure and Applied Algebra 117& 118 (1997) 277-317
Furthermore,
a natural
number
d E N is called a denominator of V if all elements
in
the set
d.V:={d,a:rE have algebraic
V}
integers
the denominator
as coordinates.
The smallest
denominator
of V will be called
of V and denoted by dv.
With these notations dimensional varieties.
we are able to state the following
height estimation
for zero-
Lemma 13. There exists a universal constant K > 0 such that for any zero-dimen-
sional diophantine subvariety V of @” and any c E N the following inequality holds:
htdc) 5 WC + log,(ndv IIVII1). (This means the function htv is of order htv(c) = 6’(‘)(c + log2(ndvI( V(l))). Proof. Let V c @" be a zero-dimensional
diophantine
variety
of degree
6 E N. The
inequality is trivial for htY(c) = 1. Therefore, we may suppose, without loss of generality, that the ring extension Q --f Q[ V] has a primitive element u which is the image of a linear form U=2,& +... + ,I,& E Z[Xl, . . ,X,] of height at most c. We just show the inequality:
and leave the corresponding
inequalities
for py’&
- vl”‘( r), . . . , pfiu?y, - v?‘( 2’) to the
reader. Since dv is a denominator for V and U has integer coordinates, dv is also a denominator for the set of algebraic numbers u(V) contained in C. Therefore, if OL,,.
, CQE C” are the points of V and if 7’ is a new indeterminate,
f(T):=fI(T
the polynomial
- dvu(ai))
i=l
has integer coefficients,
i.e., f(T)
belongs
to H[T]. Moreover,
f (dyT)
the polynomial
vanishes on the set u(V). Since f and q,, have the same degree 6 and since qu is the primitive minimal polynomial of the image u of U in Q[V], the vanishing off (dyT) on u(V) implies the existence of a non-zero integer b E Z such that f(T) = bqu(T) holds. Taking into account Ib(2 1 and that the coefficients of f are elementary symmetric functions in the points of the set dv . u(V) c @ we deduce the following inequalities:
2ht(~~)<2hr(/)<(2dv(lu(V)ll)6<(2d~)6(n2C)6((V(ld<(2C+‘ndvI/VII)6. Again, let V be a zero-dimensional diophantine estimate now the quantities dv and 1)V/j.
subvariety
0
of C” of degree 6. We
290
M. Giusti et al. IJournal of‘ Pure and Applied Algebra 117&118
Proposition
(1997) 277-317
14. Let c E N be a natural number such that the ring extension Q +
Q[V] has a primitive element u := ,?,A!, + . . + ;I,,& E Z[X,, . . .,X,1 of height at most c. Then dr and (/V 1) can be estimated as follows:
(i) where a~ is the leading coefficient of q,(T) and nb, pi(‘I is the discriminant obtained from the polynomials in Definition 10. It follows that V has denominators of height O((n + Wv(c)). (ii) the norm 11 V /( of V satisjes 11V]j < +Y26h’” Cc). Proof. The integer la&l is a denominator of the set u(V) because it is the leading coefficient of the polynomial qu(T) which defines u(V). Since the polynomials py) , . . . , p?‘X, -v?)(T) “parametrize” the variety V in function of the zeroes of qU(T), i.e., in function of the set u(V), we deduce that la:-’ ny=, p$‘)/ is a denominator
X,-$)(T)
of V (notice that the degree of the polynomials From [39, Chapitre IV, Theo&me
U?‘(T), . . . , up’(T)
is bounded by 6- 1).
2 (ii)], one deduces
liU( V)]l 5 2ht’,(c).
V and l
Thus, for u=(cI,,...,R~)E
we have
IQ] = Ip~‘I-‘~c’“‘(u@))l. k Since pr’
is an integer
and ok(T) is a polynomial
of degree at most 6 - 1 and of
standard height at most 2hty(C), we conclude IOLk 15 62h” Cc)\\U( v)((“-’
5S26htc
This implies that
[[VI]:=max{ I/MI]: a E V} < fiLi2”h”(C). Remark 1. The height is strongly
related with complexity
issues in polynomial
equa-
tion solving. In fact, given a regular sequence of polynomials fi , . . . , fn E E[Xl, . . . ,X,] defining a zero-dimensional diophantine variety V, we have an upper bound for the length of the output of any algorithm which solves the system fi = 0,. . . , fn = 0 geometrically, namely: (n + 1) deg V htY(c). Here we assume tacitly that the output is given by a linear form representing a primitive element of the ring extension Q + Q[ V] and by polynomials as in Definition 10 and that these polynomials are given in dense and their coefficients in bit representation.
292
M. Giusti et al. IJournal of’ Pure and Applied Algebra 117&118
In a more down-to-earth 6 = ~‘(a)
language
to be smooth
points
this means
(Observe & implies notations
that everything
of
deg r;; of elements
of I;; to
. , Y&i]-module
Q[ V]. Furthermore,
we ask
a primitive
comes
that we ask all the elements
of V and the number
be equal to the rank of the free Q[Yl,. the linear form U to generate
(1997) 277-317
together
that the zero-dimensional
element
of the ring extension
since the smoothness
Q-algebra
of the elements
O[&] is reduced.)
we model our notion of height of algebraic
Q + Q[ K].
Maintaining
of
these
varieties by means of the follow-
ing function. Definition
17. Given
V a complete
intersection
diophantine
variety as before, a linear
change of coordinates AE .Nv, a linear form U E Z[Yn_i+l,. . . , G] whose image u in Q[ V] is a primitive element of the integral ring extension (2) and given a point a E Z& we define the height of V with respect to the triple (A, U,a) as ht(V;(A,U,a)):=ht(&U). Our first approximation
to the notion of height of a diophantine
complete intersection
variety is given by the function htv : N -+ N which associates to any natural number c E N the value htv(c) := max{ht(V; (A, U,a)) (ht(A, U,a)
programs subsections
we discussed
the mathematical
form and syntactical
en-
coding of the output of the algorithm for geometric solving we are going to exhibit. However, we also need a suitable encoding for input and intermediate results of our algorithm. As mathematical objects our inputs are polynomials with integer coefficients f E Z[X,,
. ,X,1. These polynomials
yields the first two possible sparse encoding.
can be written
encodings
Thus, if d is the degree of f
height, the length of f under dense encoding
Let us remark that the binomial
as lists of monomials
for input and intermediate
coefficients
results:
and this dense and
and if h is an upper bound
for its
is
appearing
in this expression
are poly-
nomial in the number of variables in the case of “small” degree polynomials (e.g., the length of the dense encoding of f is h. (“l*) z h. n* if the degree of f is at most 2). Analogously, the length of the dense encoding of f is polynomial in d for d + n. Alternatively, our input polynomial f may be given by representing just all its nonzero coefficients: this is the sparse encoding of the polynomial f.Then, if h is a bound of the height, d the degree and N the number of monomials with non-zero coefficients of f, the length of the sparse encoding of f becomes hNn log, d.
M. Giusti et al. I Journal of Pure and Applied Algebra 117&118
However, programs. version
in many
practical
applications
This is, for instance,
of the 3SAT problem
in the following Definition
the approach
(Problem
we shall mainly restrict ourselves
our input
in the encoding
to the straight-line
may be given
of the inputs
program encoding
as
of the
In the sequel of polynomials
sense.
18. A generic straight-line of the graph. 9 contains
the variables
polynomials
293
277-317
1) we stated in the introduction.
program
r’ over Z is a pair r’ = (99, Q), where
$9 is a directed acyclic graph and Q is an assignment vertices)
(1997)
Xl,.
of instructions
it + 1 gates of indegree
. , X, and by the constant
to the gates (i.e.,
0 which are Q-labeled
by
1 E Z. They are called the input gates of
r’. We define the depth of a gate v of the graph 9 as the length of the longest path joining
v and some input gate. Let us denote the gates of the directed acyclic graph 9’
by pairs of integer numbers corresponding
(i,j),
value of an arbitrary
where i represents numbering
(this encoding can be seen in [26,34,40,41,44]). (i,j) we have the following operation:
where Ayj,B$’
are indeterminates
puted values corresponding
the depth of the gate and j is the
imposed
on the set of gates of depth i
Associated
called parameters
to the intermediate
gate
of r’ and Q,.s,Qr/sl are precom-
to the gates (Y,s) and (Y’,s’).
We denote by 2 = (Ay), B = (B$“) the list of all parameters in the straight-line program r’. The intermediate results Q;j of r’ are therefore polynomials belonging to Z[k,B,X, , . . . ,X,] and r’ represents a procedure which evaluates them. A (finite) set of polynomials
fi, . . . , J1 E Z[X,, . . . , X,] is said to be evaluated
by a straight-line
pro-
gram of generic type r’ with parameters in a set 9 C Z if specializing the coordinates of the parameters k and B in r’ to values in 9, there exist gates (il,jl), . . . , (&,j,) of r’ such that
holds for 15 k 5s. Specializing in the indicated way the parameters of r’ into values of 8 we obtain a copy r of the directed acyclic graph 9 underlying the generic straight-line program r’ and of its instruction assignment Q. We call this copy r a straight-line program (of generic type P) in Z[X,, . ,X,,] with parameters in F. The gates of r correspond to polynomials belonging to Z[X,,. . . ,X,1. These polynomials are obtained from the intermediate results Qij of r’ by specializing the parameters on which they depend. We shall call these polynomials the ate results of the straight-line program I-. Furthermore, the polynomials fi, called (final) results or outputs of r. Alternatively, we shall say that fi, represented, computed or evaluated by r
adequately intermedi. . . , fs are . , fs are
294
A4. Giusti et al. IJournal of’ Pure and Applied Algebra 117~6 118 (1997) 277-317
The current complexity measures for the generic straight-line - the size of r’= the size of the graph 3, - the non-scalar depth (r’) = the depth of the graph 3. The size and non-scalar gously
defined.
parameters
depth of the specialized
Additionally,
for any straight-line
in F G Z, we call the maximum
height of the parameters nomial f E Z[Xt, .
straight-line program
r
program
r’ are:
program
r are analo-
in Z[Xt,. . . ,X,]
of the heights of the elements
of r or, for short, the height of r. The encoding
,X,] by a straight-line
in 9
with the
of a poly-
program r in Z[Xt, . . . ,X,] with parameters
of height h has (bit) length 2(L2h + L log, L). Note that the notion of straight-line program encoding covers as well the both notions of dense and sparse encoding of polynomials. This can be seen as follows: a polynomial of degree at most d and height at most h, given in dense encoding,
can be evaluated
by
a straight-line program of size O(d (“r)) an d non-scalar depth O(log, d) with parameters of height at most h. Similarly, if a polynomial f E Z[X,, . . . ,X,] has degree at most d, height h and N non-zero program
of size O(dN)
coefficients,
and non-scalar
it can be evaluated
by a straight-line
depth O(log, d) with parameters
of height at
most h. 2.4.
Complexity
of geometric
solving
Now that the notions of geometric
solving, degree and height of polynomial
equation
systems and straight-line programs and its distinct complexity measures have been introduced, we are able to state our main result, namely the following Theorem 19. This theorem represents our principal contribution to the solution of Problem 3 in Section 1.
Theorem 19. There exists a bounded error probabilistic
Turing machine
that from
a smooth regular sequence fi,. . , fn E Z[Xl,. .,X,1 outputs a linear form UER [XI,. . . ,X,,] representing a primitive element u of the ring extension Q + Q[ V], where V := V( fi , . . , fn) is the algebraic variety defined by f,, . . . , fn, and polynomials of the form q,,(T), p\“‘_& - v~“‘(T), . . , p??u, - vr)( T) with pi’), . . . , pr’ non-zero integers (u) “I E Z[T] such that these polynomials represent a geometric solution and qU,v1 , . . , on of the equation system fi = 0,. . . , fn = 0 (see Definition 10). The Turing machine finds this solution in time (counting the number of bit operations executed) (ndhLog)O(‘%> n+‘) using only (ndLG)‘(‘) arithmetic operations in Z at unit cost. We assume that the polynomials have degree at most d and that they are given by a straight-line program
f,, . . , fn of size L
295
M. Giusti et al. 1Journal of Pure and Applied Algebra lI7& 118 (1997) 277-317
and non-scalur depth 8 with parameters of height at most h. The quantity 6 dejined by 6:=max{deg V(f, ,..., fi): 1 liln} is the degree of the system f~ =0 ,..., fn =O. Finally, the quantity n is the height of the system f, = 0,. . , fn = 0 defined as y := max{htv,(cs((log, n + a) log, 6)): 15 i 0 is a suitable universul constant independent of the specijk input J;, . . . , fn (or even its size). Theorem 19 follows from the description of the algorithm given in Section 3. Theorem 4 follows immediately from Theorem 19 by putting d := 2, L := n2, e := 2. Let us remark here that Theorem
19 improves
and extends the main result of [25,26]
to the bit complexity model. It sheds also new light on the main complexity of the papers [54,20, 111.
outcome
2.5. A division step in the Nullstellensatz In this subsection
we show how Nullstellensatz
bounds
imply Liouville
estimates.
This establishes a close connection between Problems 5 and 6 in the Introduction. us recall the assumptions in the statement of Theorem 7.
Let
There is given a smooth regular sequence of polynomials f,, . . . , fn E Z[X,, . . . ,X,J of degree at most d and of height at most h, defining a zero-dimensional affine variety V:=V(fi,...,f,)C@“.Moreover,thereisgivenapointa=(ccl,...,cc,)EVandareal number 0
of Theorem
7. For this purpose consider
integers
such that p/q is an approximation at level E to the algebraic We introduce the following polynomial:
p E Z[i] and q E IkJ
number
~(1.
fn+l :=(q& - p)(q& - j) of p). The assumption V n (Q[i] x C-’ ) = 0 f,, . . . , fn+l has no common zero in C”. Thereinteger b E Z and polynomials gi, . . . , gn+i E Z[Xi, . . . ,&I
(here p stands for the complex conjugate implies that the sequence of polynomials fore, there exists a non-zero such that the following B&out
b = ah Evaluating
identity
holds:
+ . . . + gn+lfn+l
this identity
at the point CIE V, we obtain
1I Ibl = lgn+l(~)l+m This yields the following
PIN
-
iii.
estimate.
20. With assumptions and notations as before, for any rational approximation a := (PI/q,. . , p,,/q) E Q[i]” at level E to the point IXE V we have the
Proposition
296
M. Giusti et al. I Journal of’ Pure and Applied Algebra 117& I18 (1997) 277-317
inequality:
In particular, for [Ia - ~111
E-1
Ic7n+~(~>l(21141 + l>“** (Here pl,. . , p,, E Z[i] are Gaussian integers and q E N is a natural number.) From the second inequality we deduce that it is sufficient to bound the values of llclll and jgn+t (a)/ in order to obtain a Liouville estimate for the rational approximation a of the point CI.Note that the bounds for llcl]l and jgn+r(a)l which can easily be deduced from the known Nullstellensltze (as, e.g., in [3,4, 341) would be insufficient for our purpose (proving
Theorem 7) as they imply only a unspecific
- log, E dc”h < log,
general estimate, namely,
I41
for a suitable constant C > 0. A Liouville estimate of this type does not take into account the specific properties of the variety V expressed through its degree and height. The more specific bounds rem 7 by means result:
for I/M]/ and [g,,+,(a)] as required
of Proposition
20 are an immediate
for the proof of Theo-
consequence
of the following
Theorem 21. Let f,,
, fn, fn+, E Z[X,, . . . ,X,,] polynomials having no common zero in C” and vertfying the following assumptions: _ there exists a straight-line program of size L and non-scalar depth e with parameters of height h that evaluates the polynomials fi, . . . , fn, f,,+l; _ the degrees of the polynomials fi, . . , fn+l are bounded by d and h’ is an upper bound for their heights. Let us furthermore assume that f,,. . ., fn form a smooth regular sequence which
dejines a zero-dimensional afine algebraic variety V = V( fi, . . . , fn ). We also consider the following quantities: ~ o:=deg(V); - v]:= min ht (V) (i.e., n is the minimal value of htv distinct from 1). Then there exists a straight-line program of size L(ndG)‘(‘) and non-scalar depth of order O(log, n + log, d + log, 6 + 8) with parameters of height at most O(max{h, h’, n, log, n, 8)) which evaluates a non-zero integer a E Z and a polynomial gn+l E i&Y,, .,&I such that
a-
gn+l . fn+l
belongs to the ideal (fi,. . ,, fn) generated by fi,. . ., f,, in E[X,,. . .,X,].
M. Giusti et al. I Journal of Pure and Applied Algebra 1178~ 118 (1997) 277-317
The proof of this theorem in Section 4. Applying bound
for the value
will follow from the description
Theorem
21 and Lemma
Jgn+i(C()]. From this bound
we then deduce easily Theorem
of the algorithm
23, we obtain together
297
given
a more precise upper
with Proposition
14 and 20
7.
3. An algorithm for geometric solving The aim of this section is to establish a proof for Theorem 19. We describe an algorithm which implies Theorem 19. This algorithm works inductively on the codimension of the varieties recursion.
6 := V( ,f,, . . . , fi),
1 5 i < n, and our main goal is to describe this
Recall that our input is a smooth regular sequence fi, . . , fn E Z[& , . . . ,X,] of degree at most d. We assume that this input is encoded by a straight-line program r of size L and non-scalar depth C with parameters of height at most h that evaluates the polynomials ,f; . . . , fn. Our algorithm computes a geometric solution of the zero-dimensional algebraic variety V := V( fi, ..,fn ) C C".In order to describe for 1 < i < n the ith recursive
step of our algorithm,
algebraic varieties - 6, := deg( K), ~ 6:=max{6;: _
we shall refer to the intermediate
as I$ := V( f,, . . . , f;)and introduce
complete
the following
intersection
parameters:
1
q, := ht,( C), where C is a suitably chosen natural number of order O((log, n + t’) log, 6) such that Q[K] has a primitive element of height C with respect to a suitable
Noether position
of I$.
- q:=max{y,: 1 . “n Y(‘))
we shall obtain geometric computes
for every
solutions
of the algebraic
1 5 i < n a Z-linear
change of
>
such that the ring extension R; :=
is integral.
O[r,“‘,..., &y’,l-+B; The Rj-module
:=
Q[y,(‘),. . . ) r,c”ll(fi,...,fi>=a[~l
Bi is a free module
of rank D, 5 hi. Let us denote
xi : r/;+ Cnpi the projection on the first n - i coordinates phism TI, of affine varieties is finite.
of (I’,(‘), . . , gi’).
by
The mor-
Definition 22. Let assumptions and notations be as before. A lifting point for W := I/; of the finite morphism ri is a point P = (PI,. . . , pn-, ) E Zn-i with the following properties: _ the zero-dimensional fiber Wp := n;‘(P) has degree (i.e., cardinality) equal to the rank of B, as free &-module (this means deg( Wp) = Di)
298
M. Giusti et al. IJournal of Pure and Applied Algebra 117&118
_ the fiber W, contains
only smooth points. (This is equivalent
to saying that the Jaco-
we obtain from fi ( Yin), . . . , K?), . . . , fi( q@), . , IL ’ )
bian matrix of the polynomials by substituting
(1997) 277-317
for Yi’ ,. . ., l$)i the coordinates
p1 ,...,p
,,_ I 0fP
is ’ regular at evk:,
point of the fiber WC’; P. If P E TFi
is a lifting point of the morphism
of W = K. Observe that the elements
rci we will call its fiber Wp a lifting $ber
of a lifting fiber of rti are smooth points of W.
The lifting fibers Wp have the property that a geometric be reconstructed
from the projection
solution of the variety K can
xi and any geometric
such a fiber (see Section 3.2). Our algorithm lifting point I”r:E Zflhi of the morphism
solution
3.2, treating
how it is possible
“lifting
to reconstruct
of
will choose for each 1 < i 5 n a suitable
ni. In the sequel we shall denote the lifting
fiber of this point by Vc. The following Section 3 is divided into two well-distinguished (i) Section
of the equations
by a symbolic
a geometric
Newton
solution
parts, namely:
method”,
of the equations
where we show fi, . . , fi from
the lifting fiber Ve. (ii) Section 3.3, where we show how to find a linear coordinate change (Xi,. .., X,) (1), . . . , I$“), the lifting point P and a geometric solution of the equations of the -
of V,
by the quantity
Ci+ l)(si + 2)?/i. 3.1. Elementary
operations
and bounds for straight-line
programs
In this subsection we collect some elementary facts about straight-line programs. We start with an estimate for the degree and height of a polynomial given by a straightline program. In order to state our result with sufficient generality, let us observe that the notion of height makes sense mutatis mutandis for polynomials over any domain equipped with an absolute value. Lemma 23 (Krick and Pardo [34]). Let R be a ring equipped with an absolute 1.1: R--t 53. Suppose f E R[Xl, . . ,X,] is a polynomial which can be evaluated
value by a
M. Giusti et al. IJournal of Pure and Applied Algebra 1178~ 118 (1997) 277-317
299
straight-line program T in R[&, . . , , X,,] of size L and non-scalar depth & with parameters of height h. Let H > 0 be a real number and let u = (ccl,. . . , a,,) be a point in KY’such that log, ]uil 5 H holds for 1 5 i 5 n. Then f and f(a)
satisfy the following
estimations: - d&f) ~ ht(f)
5 2”, 5 (2r+1 - l).(h+log,L), - log, IS(a)1 5 (2/” - 1). (max{h,H}
+ log, L).
One of the main ingredients used in our procedure below is the efficient computation of the coefficients of the characteristic polynomial of a matrix. We shall use for this task Berkowitz’ historical
division-free
predecessors
and well parallelizable
of this algorithm).
algorithm
This is the content
[6] (cf. also [ 16,7] for of the next lemma.
Lemma 24 (Berkowitz [6] and Krick and Pardo [34]). Let R be a domain. There exists a straight-line program of size No(‘) and non-scalar depth O(log, N) with parameters in { - 1, 0, 1} that from the entries of any N x N input matrix over R computes all coeficients
of the characteristic polynomial of the given matrix.
We use this algorithm
not only for the computation
of the characteristic
of a given matrix but also for the computation
of the greatest common
given univariate
in a unique
polynomials
with coefficients
factorization
polynomial
divisor of two domain
(this
task can be reduced to solving a suitable linear equation system corresponding to the Bezout identity over the ground domain. See [34] for details). The following result is an immediate consequence of the formal rules of derivation. Lemma 25. Let R be a domain. For a given a finite set of polynomials fi, . . , ,x7 of R[X,, . . ,X,,] which can be evaluated by a straight-line program p in R[Xl, . . . ,X,,] of size L and non-scalar depth t?, there exists a straight-line program in R[X,,.. .,X,,] of size (2n + 1)L and non-scalar depth e + 1 with the same parameters as p which evaluates f,, . . . , fS and all the first partial derivatives:
Combining this lemma with Lemma 24, one concludes: let f,, . . . , fn be a family of polynomials of R[X, , . . . , X,] which can be evaluated by a straight-line program fl of size L and non-scalar depth e. Then there exists a straight-line program in R[Xl, . . . ,X,,] of size n’(‘)L and non-scalar depth O(~@+log, n) with the same parameters as p which evaluates the Jacobian determinant:
J(fl,..‘,
af;
( >
fn):=det ax,
,<,jln’ -2
In some exceptional cases the straight-line programs we are going to consider might contain divisions as operations. Since we are only interested in division-free straight-line
300
M. Giusti et al. I Journal of Pure and Applied Algebra II 7& 118 (1997) 277-317
programs,
the following
[58] becomes Proposition
“Vermeidung
von Divisionen”
technique
due to V. Strassen
crucial. 26 (Krick and Pardo [34] and Strassen
[SS]). Let P be a (division-free)
struight-line program in Z[X,, . ,X,] of size L und depth G with parameters of height h that computes a finite set of polynomials fb, . . , J;n of Z[X,, . . .,X,,]. Assume that fo # 0 holds and that fo divides f; in Z[X, , . . . ,X,J jbr any 1 < i 5 m. Then there exists a (again division-free) straight-line program P’ in Z[X,, . . . ,X,] lowing properties:
with the fol-
(i) r’ computes polynomials PI,. . . , Pm of Z[X, , . . . ,X,,] and a non-zero integer 8 such that for any 1 < i <: m holds
p.&C
fo.
(ii) r’ has size of order 0(22’(L + n + 2’ + m)), depth of order O(L) and its parameters have height of order max{h,O(l)}. Moreover, the height of 6 is of order: 2’t’)(max{h,
e} + log, L).
The proof of this proposition is based on the computation of the homogeneous components of a polynomial given by a straight-line program. This proof also provides an algorithm computing the homogenization of a polynomial given by a straight-line program.
This is the content
of the next lemma.
Lemma 27 (Krick and Pardo [34]). Suppose that we are given a polynomial P:= CPpXF’ . ..XF in Z[Xl , . . . ,X,,] which can be evaluated by a straight-line program r of size L and depth e with parameters in a given set 9 c Z. Let be given a natural number D. Then there exists a straight-line program T’ in Z[X,, . . . ,X,,] which
computes all the homogeneous components of P having the following properties: S uses parameters from 9, has size (D + l)*L and non-scalar depth 2d. In Section 4 we shall work with a specific polynomial which we call the pseudoJacobian determinant of a given regular sequence. We introduce now this polynomial and say how it can be evaluated. Let R be a domain containing Q. Let K be the field of fractions of R and let f,, . . . , fn E R[X, , . . . ,X,J be a regular sequence in K[&, . . . ,X,1. Furthermore, let I’j, . . . , K be new variables. We write Y=(K,...,Y,). Fix l
301
M. Giusti et al. I Journal of’ Pure and Applied Algebra 117& 118 (1997) 277-317
with
lk,j
R[ 8,. . . , Y,,X,,
E
A =(lk,jh
. . , X,,]. Let us consider
the determinant
A of the matrix
namely,
A := det(A). This determinant nomials
is called the pseudo-Jacobian determinant of the regular sequence
f, , . . , fn. If d is a bound
polynomials
are given by a straight-line
for the degrees
program
of
of fi , . . . , fn and these poly-
/I of size L and non-scalar
depth P, then
there is a straight-line program /I’ of size (nd) O(‘)L and non-scalar depth O(log, n + e) which evaluates the pseudo-Jacobian determinant A. The straight-line program p’ uses apart from the same parameters as /I only parameters of Z of height O(log, d). We shall also consider the execution of straight-line programs in matrix rings. The situations
where we apply these considerations
will be of the following
type: Let R be
a domain. Suppose that there is given a polynomial g E R[Xl, . . . ,X,] by a straight-line program ,5 in R[X, , . . . ,X,] of size L and non-scalar depth e. Suppose also that there are given n commuting D x D matrices Ml,. . . , M, over R. In such a situation the entries of the matrix y(Mi , . . . ,M,,) can be computed from the entries of Ml,. . . ,M,, and the parameters of fl by a straight-line program /I’ in R of size D’(‘)L and non-scalar depth O(Y). The new parameters of /I’ are just the values 0,l (see [27] for details). Another important aspect of our main algorithm is its probabilistic (or alternatively its non-uniform) character. This is the content of the next definition and proposition.
Definition 28. Let be given a set of polynomials
YY c Z[Xi, . . . ,X,1. A finite set Q C Z”
is called a correct test sequence (or questor set) for w belonging to w the following implication holds:
if for any polynomial
J
f(x) = 0 for all x E Q implies f = 0. of Z[Xi,. . . ,X,]
Denote
by “Ilr(n,L,e),
the class of all polynomials
evaluated
by straight-line
programs of size at most L and of non-scalar
The following of moderate
result says that for the class W(n,L,&)
which can be depth at most e.
exist many correct test sequences
length.
Proposition 29 (Heintz and Schnorr [32] and Krick and Pardo [34]). Let be natural numbers n, e, L with L 2 n + 1 and consider the following quantities: u := (2/+’ - 2)(2’ + l)*
and
given
t := 6 (eL)2.
correct test seThen the jinite set { 1,. . . , u},’ c Z”’ contains at least u”’(1 - u+) quences of length t for W(n, L, e). In particular, the set of correct test sequences for W(n, L, /) of length t containing only test points from { 1,. . , u}~ is not empty. From Lemma 23 we deduce the following complexity estimate: Let f E Z[XI , . . .,X,] be a polynomial given by a straight-line program of size L and non-scalar depth e with parameters of height h. Let c1E Z” be a point of height h’
302
M. Giusti et al. I Journal qf’ Pure and Applied Algebra 117& I18 (1997) 277-317
given in bit representation. which computes
the bit representation
In the next subsection Newton iteration.
Then there exists a (deterministic) of the value f(a)
ordinary Turing machine
in time (2e~max{h,h’})o(1).
we shall make use of a problem adapted version of the Hensel-
We are now going to describe a suitable division-free
symbolic
form
of this procedure. Let R be a polynomial
ring over Q, let K be its field of fractions and let fi, . . . , fn E
R WI , . . . ,X,] be polynomials a (division-free) straight-line assume that the Jacobian nomials
of degree at most d. Suppose that fi, . . . , fn are given by program /zI of size L and non-scalar depth 8. Let us also
matrix D(f)
:=D(fi,.
. . ,fn) := (dfi/i3Xj),,,,jl,
fi , . . . , fn is regular. We consider now the following
Nf(X,,...,X,):=
(1)
-D(f)-’
The
next
($ ,..., $) lemma
gives
operator:
(3)
functions of K(Xl,. . . ,X,). This is which we denote by Nj. For any
k E N there exist numerators (#‘, . . . , g$‘) E R[X, , . . ,X,J and a non-zero /z(‘) E R[&, . . . ,X,J such that Nj can be written as
N;=
of the poly-
.
(;;;Y;::)
This operator is given as a vector of n rational also true for the kth iteration of this operator,
Newton-Hensel
denominator
EK(X ,,..., X,,)“.
a description
R[&, . . ,X,,] that evaluates numerators
of a division-free and denominators
straight-line
program
in
for Nj.
Lemma 30. Let notations and assumptions be as before. Let k be a natural number. There exists a straight-line program in R[X, , . . . ,X,,] of size O(kd’n’L) and non-scalar
depth O((log, n + Qk) with the same parameters as p which evaluates numerators WI (k) 91 Y...,Sil and a (non-zero) denominator hck) for the k-fold iteration Nf” of the Newton-Hensel operator Nf. Proof. Let A(f) = (av)l
adjoint matrix of D( f ). can be evaluated by a O(log, n + a), as it can the operator Nf as
(4) The entries au of the matrix A(f) are polynomials of the ring R[Xl, . . . ,X,] having degree at most (n- l)(d - 1). Moreover, the Jacobian determinant J(f) is a polynomial
hf. Giusti et al. I Journal of Pure and Applied Algebra 1178~ 118 (1997) 277-317
303
of R[Xl, . . . ,X,] having degree at most n(d - 1). For 1 5 i 5 n we consider
Si
:=J(fWi
-
2
ai,jh.
j=l
All polynomials
appearing
have degree bounded
on the right-hand
side of the definition
of gi as summands
by v := nd + 1. Thus the degree of any gi is bounded
by v. Let
, . . . ,X,) E R[&, . . . ,X,] be the homogenization of gi by a new variable X0 . . . , X,) E R[Xo,. . .,X,1 be the homogenization of the Jacobian and let ‘J( f&Y&, determinant J(f) by X0.
h.4i(X0,XI
We introduce now the following homogeneous polynomials (forms): - Gi(Xo,. ,A’,) :=X~-deg(gz)(hgi), ~ N(&, . . . ,X,) := X;-deg(J(f))(hJ( f)). According to Lemma 27, there exists a division-free straight-line program
in R[X,,
. . .,X,,] of size 0(d2(n7 + n3L)) and non-scalar depth O(log, n + e) which evaluates the forms G1 , . . , G,, H. We now define recursively the following polynomials: - for k=l, l 2, 1 5 i < n let gIk’ := Gi(hck-‘), g(lk-‘I,. . . , &,k-l’), hck) := H(h(k-‘),g~-‘),
. ..) gn(k-1) ). It is easy to see that these polynomials nomial
hCk) is a denominator
gik’, . . . , gy’ are numerators and that the polyof the iterated Newton-Hensel operator Nj. A straight-
line program evaluating them is obtained by iterating k times the straight-line program which computes G,, . . . , G, and H. No new parameters are introduced by this procedure. Putting all this together we obtain the complexity bounds in the statement of Lemma 30. 0 3.2. Lifting jibers by the symbolic Newton-Hensel The idea of using a symbolic was introduced algebraic lifting
in [25]. For technical
parameters
algorithm
adaptation
for the lifting
algorithm
of Newton-Hensel
reasons, process.
for lifting fibers
in this paper it was necessary
We present
in which the use of algebraic
iteration
numbers
trix with integer entries. The whole procedure therefore The new lifting process is described in the statement
here a new version is replaced
to use of this
by a certain ma-
becomes completely rational. of the next theorem and its
proof. Let notations and assumptions be as the same as at the beginning of this section. We fix 1 < i 5 n and assume for the sake of notational simplicity that the variables Xl,. . . , X, are already in Noether position with respect to the variety K, the variables Xl,. .,X,_, being free. We suppose that the lifting point fi, the coordinates of the Z-linear form Ui and a geometric solution for the equations of the lifting fiber Ve are explicitly given. With these conventions we state the main result of this subsection as follows.
M. Giusti et al. I Journal of‘ Pure and Applied Algebra 117& 118 (1997) 277-317
304
Theorem 31. There exists a (division-free) straight-line program & in the polynomial
ring Z[X, , . . .,X,_,, Uj] of size (id&L)‘(‘) and non-scalar depth O((log, i + k) log, Si) using as parameters - the coordinates of S, _ the integers appearing in the geometric solution of the equations of the lofting fiber VI, and - the parameters of the input program T such that the straight-line program C computes _ the minimal polynomial qi E Z[X,, . . . , Xn-i, Ui] of the primitive element ui of the ring extension Q[Xr , . . . , ~~-il~~[~l~~~xl~~~~~~~ll~fl~~~~~~~~ _ polynomials p”:. , . . . ,Xn_i], p := nL=,_,, pf’ and polynomials n ,+,,...J)EZ[Xl n v(l) ,,+,, . . , vf’ E Z[Xl,. . ,Xn-i, U,] with max{degurvj’); r
-V~~I(Vi),...,p~‘X,-V~‘(Uj))p
holds. Without loss of generality, we may assume that c represents the coeficients of the polynomials qi and vf$, , vt’ with respect to Ui. Proof. Under our hypotheses, namely that XI,. . . , X, are in Noether position with respect to the variety K, the variables Xl,. . . ,Xn_i being free, we have the following integral ring extension
of reduced rings:
Ri I= O[X, ,...,X,-i]--tBi:=~[X,,...,X,]/(fi,...,J;). Let fl=( ~1,. . . , pn-i)E Znpi be a lifting point of the morphism rc, and let V, = ni’(i9) be its lifting fiber. We have deg V, = Di = rankRZ Bi 5 deg J$ = 6i. Let Vi = &_i+rXn-i+t + ” ’ + 2,X, E Z[Xn-ifI7 . . .,X,1 generate a primitive element of the ring extension Q +Q[F$] (i.e., Ui separates the points of V,). The minimal polynomial of the image of Ui in Bi has degree at least the cardinality of the set Ui(Vc). Since the linear form Ui separates the points of I$, this cardinality is deg Ve = rankR,Bi. We conclude that
U, generates also a primitive
element of the ring extension Ri+Bi. In the sequel we denote the primitive element generated by the linear form Vi in the ring extensions Q-tQ[V~] and Ri -fBi by the same letter ui. By hypothesis
the following
data is given explicitly
(i.e., by the bit representation
of their coefficients): - the primitive minimal equation q E Z[T] of the primitive element u; of Q[b], _ the parametrization of Vc by the zeroes of q, given by the equations Xr - p1 =O,...,Xn__i
- Pn__i=O,
4(T) = 0,
-v~-i+l(T)=O,...,p,X~-v~(T)=o,
Pn-i+lXn-i+l
. . , v,, are polynomials of Z[T] of degree strictly less than Di = deg q integers. Check again Definition 10 to see that the polynomials pn-i+lXa-i+t - an-i+r( T), . . . , p,X,, - vn(T) are assumed to be primitive. Let a E Z be the leading coefficient of q. where
t+-i+l,
.
and on-i+l,..., pn are non-zero
305
M Giusti et al. I Journal of Pure and Applied Algebra I17& 118 (1997) 277-317
We consider elements
now
fi,. ..,fias polynomials
of the polynomial
be the corresponding
Recall that the Newton
. . ,X,1. Let
ring Ri[X,_i+l,.
Jacobian matrix off operator
in the variables
Xn_i+l,. . .,X,,, i.e., as
f :=(fi , ...,f;)and let
(with respect to the variables XnPr+r,. . . ,X,).
with respect to these variables
is defined as
Using Lemma 30, we deduce the existence of numerators ~~_~+l,. . . , g,, and a non-zero denominator h in the polynomial ring Ri[Xn_i+l,. . . ,X,] such that the following K-fold iterated Newton operator has the form
Let K be a natural number.
cl.-,+1
h
Iv;=
0 :
.
9_ h
Now,
let MC_,+,
Q-algebra
, . . . ,Mx, be the matrices describing the multiplication tensor of the Q[Q] (recall that by assumption a geometric solution of the polynomial
equation system defining are known).
Vj is given and that therefore
the matrices a-‘q(
Let M denote the companion matrix of the polynomial 1
Moreover,
T) E Q[T}.
Let n - i +
Vj(M).
the matrices
pj#l-‘Mx,
have integer entries. Let K := [l + log, Sil and note
that K 2 1 + log, Di holds. Our straight-line a subroutine
I&_,+, , . . . , Mx,
program
fi will execute K Newton steps in
which we are going to explain now: Let us consider the following
column
vector of matrices Nn_i+r,, . . , N,, with entries in Q(Xr,. . . yXn_i):
(?+j
:=&-f(M)=
(‘“;z’), > n
whereM = Wfx,_,+, , . . . ,Mx,) and gn-i+l, nator polynomial
of Lemma 30. Finally,
A := Ui(Nn-i+r,.
I,_
. . . , g,, are the numerator
let us consider
. . , N,) = jbn_i+lNn-i+l
+
and h the denomi-
the matrix
+ A,Nn
This matrix is a matrix whose entries are rational functions of C&Y,, . . .,X,-i). From the fact that P,=(p,,..., pn_i) is a lifting point and from the proof of Lemma 30 one deduces easily that in fact the entries of A’Z belong to the local ring Re :=(Ri)(~,--pl,...,
~,_,-p,_,)=a[x,,...,x,-il(x~-p,
,,.,,x,-,-p._,).
M. Giusti et al. I Journal of Pure and Applied Algebra 1176~ 118 (1997) 277-317
306
Let T be a new variable.
With these notations
and assumptions
we have the following
result:
Lemma 32. Let xEQ(X,,..., Xn_i)[T] be the characteristic polynomial of A%?and m, E Ri[T] the minimal integral equation of the primitive element ui of Bi over Ri. Let x(T) = TDl+CfAi’ akTk and m, = Tol+C~~~’ bkTk with ak, bk E 0(x,, . . ,Xn-i) for 0 di+l, where orde denotes the usual order function (additive valuation) of the local ring R8. Proof. Let V, = {
fiber V, and from Hensel’s
Lemma
(which
represents
a
symbolic version of the Implicit Function Theorem) we deduce that there exist formal power series Rf!,,, , . . . , R~~‘EC[[X~ - pl,...,X,_i - pn-i]] with R~~i+l(P)=~~)i+l,
. . ..R.)(P;)=@
such that for R(‘):= (XI - p,, . . .,X,_,
- pn_-i,RyLi+,, . . . , RLt’) the
identities
f,(R’t’)=O ,.._,. fi(R(t))=O
(5)
hold in @[[Xl - PI , . . . ,X,_, - pnJ]. Let z4(‘):= L’i(R(‘)) = An-i+lRfl,‘li+l+ . . +A,R$“. As shown in 1251, the minimal polynomial m, of the primitive element ui verifies in @[[Xl - ~1,. . ,Xn_i - p,_i]][T] the identity
m, = II
(T - u(t)).
(6)
above which produces the matrix J@ E (Re )D1xD3 starting with the matrix M E QD3xDC (recall that M is the companion matrix of the polynomial a-‘q(T) E Q[T]). The same construction transforms the Di distinct eigenLet us now consider the construction
values of the diagonalizable matrix M, namely the values c, = Ui(tl), 1 < 15 Di, into eigenvalues of JY. (Observe that by construction 4 is a rational function of the matrix M and therefore the same rational function applied to any eigenvalue of M produces an eigenvalue of JY.) As shown in [25], in this way we obtain Di distinct rational functions zZ(‘)E @(Xl,. . . ,Xn-i), which are eigenvalues of ~4’. Moreover, these rational functions are all defined in the point Pi (this means that u”(‘)E @[Xl,. . .,X,,_&_, ,,,,,,x,,_,_~+,) holds). For 1 i: 1
- pn-;)6S+‘.
- p,,-i]]
and they satisfy in this
(7)
(For a proof of these congruence relations see [25].) Let us now consider the characteristic polynomial x of the matrix J%‘. Since the coefficients of JY belong to RF:= O[Xl , . . . ,xn-il(x,-p,,...,x,~,-p,_,), the c~~~cient~of x do so too. Therefore, x can be interpreted as a polynomial in the variable T with coefficients in the power series ring a[[& - ~1,. . ,Xn-i - pm-i]]. From the fact that the
M. Giusti et al. I Journal of Pure and Applied Algebra 1 I7& 118 (1997) 277-317
rational
u”(l), . . . , iPI)
functions
x(T)=
J-J I
eigenvalues
of J&’ we deduce that (8)
the k-th elementary
symmetric
function
in Di arguments.
(8) implies that for 0 < k < Di - 1 we can write the coefficient
of the polynomial
relations
m, satisfies bk = (- 1 )“‘-kek(U(‘),
(7) and the identities
the congruence
. . . , z&@)). From the congruence
(6) and (8) we conclude
now that for 0 5 k 5 Dj - 1
relations
ak - bk E (xl - p1,. . . ,Xn-i - Pn_i)6f+’ hold in O[[Xi - ~1,. . . ,Xn_i - p,_i]]. We continue coefficients
The
ak of x( 7’) as
ii@~)). From the identity (6) we deduce that the kth coefficient
ak=(-l)D’-kok(U”“‘,..., bk
D, distinct
(T-G(“)
holds. Let csk denote identity
represent
307
This proves the lemma.
now the proof of Theorem
of 1, let us consider
(9) 0
3 1. In order to see how to evaluate
6 := det (h(g))
the
E Ri and the matrix
A!, := 9./Z. This matrix -Ai has entries in Ri. Now we have the identities det(T-IdD, Let &T)=TDf
- A)=det(T.Id~~
- 8-1~,)=8-D’det((BT)Id~,
-&I).
+ $D,-~T~~-’
-A$. From the identities of x can be written as
+ .. . + ~$0E Ri[T] be the characteristic polynomial of above we deduce that for 0 5 k 5 Di - 1 the kth coefficient
Executing K = [l + log, Sil steps in the Newton iteration (as given at the beginning of this proof) to produce the entries of the matrix M, 8, JZ and finally Ai (in this order) and applying non-scalar parameters
Lemma 24 we produce
a straight-line
program
q! in Q(Xi,. . . ,Xn-i)
of
size O(log, &d3i6L) and non-scalar depth O((log, i + /) log, Si) using as those given by the statement of Theorem 31 such that 4’ evaluates the
family of polynomials 1,8,. . . , BDs, 40,. . ,40, -1 E Ri. Now, applying Strassen’s Vermeidung von Divisionen technique (Proposition 23) we obtain a division-free straight-line program 4.” in Q[A’l, . . . ,Xn-i] of size (idGiL)‘( non-scalar depth O((log, i + t) log, Si) using as parameters the coordinates pi,. . . ,pn-, of P,, the coefficients of the linear change of coordinates for the Noether normalization for E, the rational numbers appearing as coefficients in the geometric solution of the lifting fiber Vp8and the parameters of r such that 4” evaluates for each 0 5 k < D; - 1 the expansion of ak in Q[[Xi - ~1,. . .,X,_, - p,_,]] up to terms of degree of order 6; + 1. Taking into account the congruence relations (9) we see that the division-free straight-line program ry evaluates polynomials go, . . , 90, _ 1 E fil![&, . . . ,Xn-i] such that bk - gk E (Xl -
pl,.
. . ,Xn_j - Pn-i)“+’
308
M.
Giusti et al. IJournal
holds in O[[Xi - pi,. . ., X,-i that the degrees
of Pure and Applied Algebra 117&118 (1997) 277-317 - Pn-i]]
of the coefficients
exceed 6;. Putting
all this together,
for any 0 5 k < Di - 1. From [25] we deduce
bk of the minimal we conclude
polynomial
m, E Ri[r]
do not
that
bk=gk
holds for any 0 5 k < D, - 1. This means that the division-free
straight-line
program
4” computes the coefficients of the polynomial q1 := m, E Q[Xl, . . . ,X,, V;] = Ri[Ui]. In order to compute the parametrizations Ugly+,, , v!’ E O[X,, . . . ,X,_,, T] we have to use the same kind of techniques bined with the arguments
developed
(namely
truncated
Newton-Hensel
in [34] and applied
in [25,26].
iteration)
com-
Let us be more
exact. Let n-if1 5 k 5 n, let & be a generic linear combination of the variables & and Ui (one may use a new indeterminate for that) and iet rnx, E R,[Xk] and ??Q E Ri[Zk] be the minimal
polynomials
and zk in the Ri-algebra
of the Rj-linear
endomorphisms given by the images of & B, = Q[ K] = o[Xi, . . . , X,1/( f,, . . , fi ). The polynomials mx,
and mz, are manic and hence squarefree. We compute the coefficients of the polynomials mx, E Ri[Xk] and mz, E Ri[Zk] in the same way as before the coefficients of qi = m,. So, we have two manic squarefree polynomials mxk E Rj[Xk] and rnzk E R;[Zk]. Taking into account that by the choice of zk the variable Ui is a generic linear combination of & and zk and therefore separates the associated primes of the ideal (mx,, mz, ) in R;[Xk,&] = Ri[Xk, Ui] we can apply directly Lemma 26 in [34] in order to obtain the parametrization
associated to the variable xk. Doing the same for each of the variables involved, namely the variables Xn_i+, , . . . , X,, and putting together the corresponding straight-line programs, we obtain a procedure and a straight-line program & of the desired complexity which computes the output of Theorem 3 1. 0
3.3.
The recursion
Proposition 33. There exists a division-free (id6iL)O(‘) and non-scalar depth O((log, i-t/)
arithmetic network of non-scalar size log, Si) using parameters of logarithmic
height bounded by max{h, r/i, O((log, i + e) log, Si)} which takes as input _ a Noether normalization for the variety V:, _ a lifting point Pi, and _ a geometric solution of the lifting ,fiber VP, and produces as output _ a linear change of variables (Xl,. .,X,,) + (Yl(‘+“, . , Y,(‘+‘)) such that the new variables Y[‘+‘), . . . , Y,(‘+‘) are in Noether position with respect to E+l, _ a lifting point Pi+, for E+I, and _ a geometric solution for the lifting ftber VP,,,. Proof. The construction of the arithmetic network proceeds in three stages. In the first stage we apply the algorithm underlying Theorem 31. In the second stage we intersect algorithmically the variety K with the hypersurface V(f;+,) in order to produce first a
M. Giusti et al. IJournal of Pure and Applied Algebra 117&118
Noether normalization Then
we produce
of the variables
a linear
with respect to the variety
U;+l representing
form
Ri+l -+ Bi+l and a straight-line
integral ring extension analogous
to the one in the conclusion
algorithm
underlying
of Theorem
the proof of Proposition
take care of is that we are working
a primitive
277-317
309
K+l = E n V(fi+l). element
ui+l
program representing
of the
polynomials
31. For this purpose
we use the
14 in [25]. The only point we have to
now over the ground field Q and that we have to
take into account the heights of the parameters program. We remark that the straight-line
(1997)
program
of Q we introduce
in this straight-line
in question
has non-scalar
size (id6J)0(‘)
+ 6) with parameters
of logarithmic
height bounded
and non-scalar
depth O(log,(dGi)
by O(log2(dGi)
+ e). In the third and final stage we consider
the polynomial
JU’I . . ..h+~> := det the ideal Ii+1 = (fi,. . . , fi+, ). Let us observe that
which is a non-zero
divisor modulo
the polynomial
. . . , fi+l ) can be evaluated
J(fi
by a division-free
straight-line
of length O((i + 1)5 + L) and depth O(log,(i + 1) + S). Let ~1E a[& , . . . ,Xn-i_ I] be the constant term of the characteristic
program
polynomial
of
the homothety given by J := J(f, , . , fi+l ) modulo Ii+,. Since J represents a non-zero divisor modulo Zi+, we conclude that /A does not vanish. Furthermore, we observe that 1-1can be evaluated by a division-free straight-line program in O[Xl,. . . ,Xn_i_1] of length (i6iL)O(‘) and depth O(log, d6i + (log, i + t) log, Si) = O((log, i + t) log, S,), and so does the product p . p, where p = nE_,+, pi” is defined as in the statement of Theorem 3 1. Using
a correct test sequence (see [32]) we are able to find in sequential time and parallel time O((log, i + d) log, Si) a rational point Pi+1 E .Zn-i-l of logarithmic height bounded by O((log, i + C) log, Si) which satisfies (p.p)(Pi+l) # 0.
(i6L)“(”
Clearly,
P,+l
is a lifting
point
for the variety
F+l.
In order to obtain
a geometric
solution of VP,+,with primitive element ui+l induced by the linear form Ui+l we have to specialize in the point Pi+, the polynomials obtained as output of the second stage (see [25, Section 31). By this specialization
we obtain the binary
representation
of the co-
efficients of certain univariate polynomials in Ui+l which represent a geometric solution of the fiber VP,,, . Nevertheless, it might happen that the height of these coefficients is excessive. In order to control the height of these polynomials we make them primitive. This requires some integer greatest common divisor (gcd) computations which do not modify the asymptotic
time complexity
of our algorithm.
0
4. Lifting residues and division modulo a complete intersection This section is dedicated to the proof of Theorem trace formula for Gorenstein algebras given by complete
ideal
21. The outcome is a new intersection ideals. This trace
M. Giusti ec al. I Journal of Pure and Applied Algebra 117& 118 (1997) 277-317
310
formula does not make reference Our trace formula represents
anymore
to a given monomial
an expression
basis of the algebra.
which is “easy-to-evaluate”.
4.1. Truce and duality Trace formulas mination
theory.
appear in several recent papers treating problems Some of these papers
use a trace formula
in algorithmic
eli-
in order to compute
a
quotient appearing as the result of a division of a given polynomial modulo a given complete intersection ideal (see [22,34]). Other papers use trace formulas in order to design algorithms for geometric (or algebraic) solving of zero-dimensional Gorenstein algebras
given by complete
intersection
ideals [ 1,2, 141. The paper [49] uses a trace
formula to obtain an upper bound for the degrees in the Nullstellensatz. However, all these applications of trace formulas require the use of some generating family of monomials
of bounded degree which generate the given Gorenstein
algebra as
a vector space over a suitable field. As a consequence, such trace formulas provide just syntactical complexity or degree bounds and in particular no intrinsic upper complexity bound (as e.g., in Theorem 21) can be obtained in this way. In this subsection we introduce an alternative trace formula in order to obtain maximum benefit from the geometrically and algebraically well-suited features of Gorenstein algebras. Let us start with a sketch of the trace theory. For proofs we refer to [37, Appendices E and F]. Let R be a ring of polynomials over a given ground field (for our discussion the ground field may be assumed to be Q). Let K be the quotient field of R and let . , fn R[& , . . . ,X,] be the ring of n-variate polynomials with coefficients in R. Let fi, be a smooth regular sequence of polynomials in the ring R[Xl, . . . ,X,,] of degree at most d in the variables Consider
now the R-algebra
B := R[X, We assume
Xl, . . . , X,, generating
a radical ideal denoted by (f;,
B given as the quotient
of R[X,, . . . ,X,] by this ideal:
,...,x,ll(fl,...,h>.
that the morphism
R + B is an integral
ring extension
representing
Noether normalization of the variety V( fi , . . , fn) defined by the polynomials in a suitable affine space. Thus, B is a free R-module of rank bounded by the the variety V( fi,. ..,fn)(this estimation is very coarse but sufficient for our Moreover, the R-algebra B is Gorenstein and the following statements are this fact. We consider B* := HomR(B,R) BxB*
..,fn).
as a B-module
a
fi , . . . , fn degree of purpose). based on
by the scalar product
-+B*
which associates to any (b,z) in B x B* the R-linear map b.z : B +R defined by (b.z)(x) := z(bx) for any element x of B. Since the R-algebra B is Gorenstein, its dual B* is a free B-module of rank one. Any element a of B* which generates B* as B-module is called a trace of B. There are two relevant elements of B* that we denote by Tr and a. The first one, Tr, is
M. Giusti et al. IJournal of Pure and Applied Algebra 117& 118 (1997)
called the standard trace of B and it is defined in the following qb : B + B the R-linear image Tr(b)
map defined by multiplying
311
way: given b E B, let
by b any given element
under the map Tr is defined as the ordinary
qb of B (note that this definition
277-317
of B. The
trace of the endomorphism
makes sense since B is a free R-module).
In order to
introduce
c (which will be a trace of B in the above sense), we need some additional
notations.
For any element
g E R[Xl , . . .,X,1
we denote by S its image in B, i.e., the
residue class of g modulo the ideal (f,, . . . , fn). Let Y,, . . . ,Y, be new variables and let Y :=(Yl , . . .,Y,). Let 1 < j 2 n and let 4’ := A( Y,, ,Y,) be the polynomial of R[YI,..., Y,] obtained by substituting us consider the polynomial
in 4 the variables
.f,‘-L=~G(Yk-Xk)ER[~l>...>
&,Yl>...>
Xi,. . . , X, by Y,, . . . , K. Let
r,l,
k=l where the bk are polynomials at most (d - 1) (observe
belonging
that the
J!jk
to R[&, . . . , X,,, Yi, . . . , Y,] having total degree are not uniquely
J;,..., fn). Let us now consider the determinant can be written (non-uniquely) as A=
determined
A of the matrix
by the sequence (Ijk)isj,kln
which
Ca,l(x,,...,X,)b,(Y,,...,Y,)EW[~~,...,X,,Y,,...,Y,1, m
of R[Xl , . . . ,X,] and the b, elements of R[ Yl, . . . , K]. (Observe that it will not be necessary to find the polynomials a, and b, algebraically, we need just their existence for our argumentation.) The polynomial A is called a pseudo-
with the a,,, being elements
of the regular sequence (ft , . . . , fn). Observe that the polynomials a, and b, can (and will) be chosen to have degrees bounded by n(d - 1) in the variJacobian determinant
ables Xi,. .,X,, and Yi,. . . , K, respectively. Let c, E R[Xl,. .,X,1 be the polynomial we obtain by substituting in b, the variables Yi,. . . , Y, by Xl,. . . ,X,. For j the class of the Jacobian
determinant
J(fi,
. ..,fn)in B we have the identity
the image of the polynomial A in the residue class ring R[X,,. . . ,X,, , K] modulo the ideal (f, , . , fn, fiy, ,f,' ) 1sindependent of the particular choice Of the matrix This justifies the name “pseudo-Jacobian” for the polynomial A. With these notations there exists a unique trace rs E B* such that the following identity holds in B: Moreover, Y1,
(t!kj)l
The main property of the trace 0, known as “trace formula” (“Tate’s trace formula” [37, Appendix F], [38] being a special case of it) is the following statement: for any gER[Xl,...,X,] the identity
(10)
M. Giusti et al. IJournal of‘ Pure and Applied Algebra 117&118
312
holds true in B. Let us observe that the polynomial underlying
the identity
Problem 34 (Lifting degree
a(g . a,) . c, E R[Xl, . . . ,X,]
C,
(10) is of degree at most n(d - 1) in the variables
The main use of this trace formula consists
bitrary
(1997) 277-317
of a residual
in Xi,. . .,X,,
class).
in solving
the following
Given a polynomial
find a polynomial
Xi,. . . ,X,.
problem.
g E R[Xi, . . . ,X,]
of ar-
g1 E R[Xi, . . . ,X,J of degree at most
Xi,. . .,X,, such that j, = 3 holds in B.
n(d - 1) in the variables
As we have seen before, the trace formula
( 10) solves Problem
34 since it allows
us to choose for gi the polynomial
c
g1 :=
[email protected],). c,
(11)
m
However, advantage
defining the polynomial g1 by the formula (11) inhibits us from taking of any special “semantical” features of the R-algebra B: one “a priori” needs
all monomials of degree at most n(d - 1) for the description of the polynomials c, (and a,). Therefore, we replace the trace formula (10) by the following alternative one. Proposition 35 (Trace formula).
With the same notations as before, let us consider
the free R[Xl, . . . ,X,,]-module B[Xl , . . . ,X,,] given by extending scalars of B (this means we consider the tensor product B[X,, . . ,X,,] := B @R R[Xl,. . . ,X,1) and let us also consider the polynomial Al E R[X,, . . . ,X,,] given by AI :=~&+,EB[X m
,,...,
Then for any g E R[Xl , . . . ,X,]
X,]. the following identity holds true in R[X,, . . . ,X,,]:
Ca(Y.a,).c,=T;(J-‘S.A,). m (Here Tr := Tr C%IdR[x ,,,._, x,,l : B[X,, . . . ,X,] + R[X, , . . . ,X,]
is the standard trace ob-
tained from the standard trace Tr : B + R by extending scalars). Proof. Let Tr : B -+ R be the standard trace of the free R-module [37, Appendix Tr(J-‘j)
B. Let us recall from
F] that for any g E R [Xl,. . ,X,J the identity = o(j)
holds. From the R[ Y,, . . . , Y,]-linearity of the map 6 : B[Yl,..., deduce that any g E R[Xl , . . . ,X,J satisfies the identities %(?‘gA1(Y,,...,
K)) = C Tr(J-‘Sa, m
In other words we have in R[Y,, . . . , K] Tlr(J-‘gA,(Y~,...,~~))=
x+j.&).b, m
.b,)=
K] + R[Yl,. .., K] we
xT;(j-‘j&&b,. m
313
M. Giusti et al. I Journal of Pure and Applied Algebra 117& 118 (1997) 277-317
for any g E R[Xl , . . . ,X,1. Replacing
in this identity the variables
Yt , . . . , K by Xl,. . . , A?,,
we obtain the desired formula Tr(J-‘&I,)=
~f7((j.am),c,. m
0
One easily sees that for any h E R[X,, . . . , X,,, Y,, . trace of the image h of h in the R[ Yl, .
, X,], T?(h) is simply the standard
, Y,]-module
B[ Y,, . . . , G]. This observation
together with Proposition 35 represents our basic tool for the evaluation of formula (11) and hence for the solution of Problem 34. This is the content of the following considerations. Let us consider
the K-algebra
B’ = K @R B obtained
by localizing
B in the non-zero
elements of R. Fix a basis of the finite dimensional K-vector space B’. Let M., , . , Mxn be the matrices of the homotheties ylx, : B’ +B’ with respect to the given basis of B’ and let Tr denote the function which associates these conventions let gt be defined as 91 :=Tr(J(fl,...,
fn)(&,...,&)-’
to a given matrix its usual trace. With
.g(Mx I,..., Mx,)
.d(Mx,,...,Mx,,Xl,...,X,)).
One easily verifies that gr belongs
(12)
to R[Xl,. . .,X,1 and that jr = S holds in B.
4.2. A division step The lifting process presented in the last subsection is now applied to compute the quotient of two polynomials modulo a reduced complete intersection ideal. More precisely, let us consider a polynomial f E R[Xl , , . .,X,,] which is not a zero-divisor in 3 and another polynomial g E R[X, ,. . . ,&I such that the residue class y divides the residue class j in B. The following proposition shows how we can compute q E R[X, , . . . ,X,J for the division of j by f in B.
a lifting
quotient
Proposition 36 (Division step). Let notations and assumptions be the same as in the previous subsection and let D be the rank of B as free R-module, Let be given the following items as input: _ a straight-line program S of size L and depth e representing the polynomials .f,g and fi,...,fn; _ the matrices Mx, , . . . , Mx, describing the multiplication tensor of B with respect to the given basis of B’ = K @R B. Suppose that f is a non-zero divisor of B and that f divides g in B. Then there exists a division-free straight-line program r in K[Xl,. . . J,,] of size L(ndD)‘(‘) and non-scalar depth O(&log, D+log, n j which computes from the entries of the matrices Mx, , . . . , Mx, and the parameters of r’ a non-zero element 0 of R, and a polynomial q of R [XI,. . . ,X,,] such that 0 divides q in R [XI , , . . ,X,,] and such that qf = &j holds in B.
314
hf. Giusti et al. I Journal of Pure and Applied Algebra 117& 118 (1997) 277-317
Proof. In order to prove this result, let us observe that any basis of B as free R-module induces
a basis of B[Yl, . . . , &] as free R[Y,, . . . , Y,]-module.
matrix of the multiplication as well the multiplication
if Mx, is the
by Xi in B[Yl, . . . , X] with respect to the same basis. Next, and J(fi, . ..,fn)are not zero-divisors modulo (fi, . . . , fn),
since the polynomials
f
the following
are non-singular:
matrices
Moreover,
by Xi in B with respect to a given basis, Mx, represents
FI :=fW..,,...,Mc,),
Finally,
let us denote by Gi and Al the following
two matrices:
and
AI :=A(Mx,,...,Mx,Y,,...,Y,), where A is the pseudo-Jacobian
determinant
of fi, . . . , fn. Let us remark that the matri-
ces F,, J1 and Gi have entries in K while Al has entries in K[Yi, . . . , Y,J. From formula (12) of the previous subsection we deduce that q1 :=Tr(Jy’ . F,’ . Gl . Al(X,, . . . , X,,)) S (by Tr we is a polynomial of R [Xl,. . . , X,] which satisfies in B the identity qij= denote here the usual trace of matrices). Finally, let us transpose the adjoint matrices Fz := ‘Adj(Fl), The quotient
and
of F1 and J1:
Jz := ‘Adj(Ji).
q E R[X,, , . . , X,,] and the non-zero
constant
0 E R we are looking
for are
given as follows: - q:=Tr(F2.J2.G1.Al(Xl,...,Xn)), - 19:= det(F2).
det(J2).
Clearly, the polynomial
q1 is the quotient
of q/8, while gj=
over, q can be computed
by a division-free
straight-line
&j holds in B. More-
program
r
in K[&, . . . ,X,,]
from the entries of Mx,, . . . , Mx, and the parameters of r’ (note that for the computation of q1 we need divisions). The complexity bounds in the statement of Proposition 36 follow by reconstruction of the programs determinants involved (cf. Section 3.1).
evaluating
f, g, J(f,,
. . . , fn), A and the
Proof of Theorem 21. In order to prove Theorem 21 we follow the algorithm underlying the proof of Proposition 36. We just have to add just some comments conceming the matrices Mx,, . . , Mx,. Let us consider a linear form ,4,X, + . . +1,X, (with %iE Z) inducing
a primitive
element u of the zero-dimensional
Q-algebra
Q[Xi , . . . ,X,1/
(f,, ...,fn).In this case R will be the field Q. Let qu E Z[T] be the minimal polynomial of u and pi&
- vi(T), . . . , p,,X, - v,(T)
the parametrizations
of the variety with
M. Giusti et al. IJournal of Pure and Applied Algebra 117&118
respect to this primitive a discriminant.
element.
Let a be the leading coefficient of q,, and p = n:=,
where A4 is a matrix with integer
entries. Now, the matrices
cation tensor of U&Y,, . . .,X,,]/(f,,.
. . ,fn)
,X,]
the multipli-
can be written as
andf=f,+r
in Proposition
36 we obtain 0 E Q, 8 # 0 and
such that
Q. 1 -q..fn+1
E(fl,...,hz)
holds. Finally, multiplying integer a and a polynomial a:=
describing
Zli(X-‘M)
for 1
pI
matrix of qu has the form
Then the companion
Mx, = pi’
315
(1997) 277-317
cxNpM.OEZ,
by appropriate powers g,+l of the form gn+l
such that a - g,+r fn+r E (ft,.
:=
of CI and p we obtain
a non-zero
8p”qEz[x,,...,Xn]!
, fn) holds. The bounds
of the Theorem
21 are then
obtained from the bounds of Proposition 36. The bounds for the height of the parameters are obtained choosing an appropriate primitive element u such that qu and the parametrizations
have height equal to the minimal
V := V(,fi ,...,fn>.
height of the diophantine
variety
0
Acknowledgements L.M. Pardo wants to thank the l?cole Polytechnique
for the invitation
and hospitality
during his stay in the Fall of 1995, when this paper was conceived.
References [1] M.E. Alonso, E. Becker, M.-F. Roy and T. Wormann, Zeroes, multiplicities and idempotents for zerodimensional systems, in: L. Gonzalez and T. Recio, eds., Algorithms in Algebraic Geometry and Applications, Progress in Mathematics, Vol. 142 (Bid&user, Basel, 1996) 1-15. [2] E. Becker, J.P. Cardinal, M.-F. Roy and Z. Szafmniec, Multivariate Bezoutians, Kronecker symbol and Eisenbud-Levine formula, in: L. Gonzalez and T. Recio, Eds., Algorithms in Algebraic Geometry and Applications, Progress in Mathematics, Vol. 142 (Birkbluser, Basel, 1996) 79-104. [3] C. Berenstein and A. Yger, Effective Bezout identities in Q[Xt , ,X,,], Acta Math. 166 (1991) 69-120. [4] C. Berenstein and A. Yger, Une formule de Jacobi et ses consequences, Ann Sci. E.N.S., 4itme serie 24 (1991) 3633377. [5] C. Berenstein and A. Yger, Green currents and analytic continuation, preprint, Univ. of Maryland, 1995. [6] S.J. Berkowitz, On computing the determinant in small parallel time using a small number of processors, lnf. Proc. Lett. 18 (1984) 147-150. [7] A. Borodin, J. von zur Gathen and J. Hopcroft, Fast parallel matrix and GCD computations, Inform. and Control 52 (1982) 241-256
316
M. Giusti et al. IJournal of Pure and Applied Algebra 117&118
[8] J.-B. Bost, H. Gillet and C. Soul& Un analogue Paris, Serie I, Math. 312 (1991) 8455848.
arithmetique
[9] J.-B. Bost, H. Gillet and C. Soule, Heights of projective IHES, 1993.
du theoreme
(1997) 277-317
de B&out,
C.R. Acad. Sci.
varieties and positive Green forms, manuscript
[IO] D.W. Brownawell, Bounds for the degree in the Nullstellensatz, Ann. Math. 126 (1987) 5777591. [I l] J.F. Canny and I.Z. Emiris, Efficient incremental algorithms for the sparse resultant and the mixed volume, preprint, 1995. [12] L. Caniglia, A. Galligo and J. Heintz, Borne simplement exponentielle pour les degres darts le theoreme des zeros sur un corps de caracteristique quelconque, C.R. Acad. Sci. Paris, Serie I, Math. 307 (1988) 255-258. [13] L. Caniglia, A. Galligo and J. Heintz, Some new effectivity bounds in computational geometry, in: T. Mora, Ed., Proc. AAECC-6, Lecture Notes in Computer Science, Vol. 357 (1989) 131-152. [I41 J.P. Cardinal, Dualite et algorithmes itiratives pour la solution des systemes polynomiaux, These, U. de Rennes 1, 1993. [I51 A.L. Chistov and D.J. Grigor’ev, Subexponential time solving systems of algebraic equations, LOMI preprints E-9-83, E-10-83, Leningrad, 1983. [I61 L. Csanky, Fast parallel matrix inversion algorithms, SIAM J. Comp. 5 (1976) 6188623. [17] J.-P. Dedieu, Estimations for the separation number of a polynomial system, preprint, Univ. Paul Sabatier, Toulouse, 1995. [ 181J.-P. Dedieu, Approximate solutions of numerical problems, condition number analysis and condition number theorems, preprint, Univ. Paul Sabatier, Toulouse, 1995. [19] M. Elkadi, Bomes pour le degre et les hauteurs dans le probltme de division, Michigan Math. J. 40 (1993) 609-618. [20] I.Z. Emiris, On the complexity of sparse elimination, Report No. UCB/CSD-94/840, Univ. of California, 1994. [21] G. Faltings, Diophantine approximation on abelian varieties, Ann. Math. 133 (1991) 549-576. [22] N. Fitchas, M. Giusti and F. Smietanski, Sur la complexite du theoreme des zeros, in: M. Florenzano et al., Eds., Approximation and Optimization in the Caribbean II, Proc. 2nd Intemat. Conf. on and Optimization (Peter Lang Approximation and Optimization, La Habana, 1993, Approximation Verlag, Frankfurt, 1995) 274-329. [23] P. Gianni and T. Mora, Algebraic solution of systems of polynomial equations using Grobner bases, in: Proc. AAECC-5, LNCS, Vol. 356 (Springer, Berlin, 1989) 247-257. [24] M. Giusti and J. Heintz, La determination des points isoles et de la dimension d’une variete algebrique peut se faire en temps polynomial, in: D. Eisenbud and L. Robbiano, Eds., Computational Algebraic Geometry and Commutative Algebra, Proc. Cortona Conf. on Computational Algebraic Geometry and Commutative Algebra, Symposia Matematica, vol. XXXIV, lstituto Nazionale di Alta Matematica (Cambridge Univ. Press, Cambridge, 1993). [25] M. Gmsti, J. Heintz, J.E. Morais, J. Morgenstem and L.M. Pardo, Straight-line programs in geometric elimination theory, J. Pure Appl. Algebra, to appear. [26] M. Giusti, J. Heintz, J.E. Morais and L.M. Pardo, When polynomial equation systems can be “solved” fast? in: G. Cohen, M.Giusti and T. Mora, Eds., Proc. I Ith Intemat. Symp. Applied Algebra, Algebraic Algorithms and Error-Correcting Codes, AAECCl 1, Paris 1995, Lecture Notes in Computer Science, Vol. 948 (Springer, Berlin, 1995) 205-231. [27] M. Giusti, J. Heintz and J. Sabia, On the efficiency of effective Nullstellensatze, Comput. Complexity 3 (1993) 56-95. [28] J. Heintz, Fast quantifier elimination over algebraically closed fields, Theoret. Comput. Sci. 24 (1983) 239-277. [29] J. Heintz, T. Krick, A. Slissenko and P. Solemb, Searching for a shortest path surrounding semialgebraic obstacles in the plane (in russian): Teorija sloinosti vycislenij, 5, Zapiski naucnyh seminarov LOMI (Leningrad branch of the Mathematical Institute Steklov), Nauka, Leningrad, Vol. 192 (1991) 163-173. [30] J. Heintz, T. Krick, A. Slissenko and P. Solemo, Une borne inferieure pour la construction des chemins polygonaux dans [w”, preprint, Univ. de Limoges, 1992. [31] J. Heintz and J. Morgenstem, On the intrinsic complexity of elimination theory, J. Complexity 9 (1993) 471-498.
M. Giusti et al. IJournal of‘ Pure and Applied Algebra 117& 118 (1997) 277-317
317
[32] J. Heintz and C.P. Schnorr, Testing polynomials which are easy to compute, Proc. 12th Ann. ACM Symp. on Computing (1980) 262-268; also in Logic and Algorithmic. An Intemat. Symp. held in Honour of Ernst Specker, Monographie No. 30 de I’Enseignement de Mathematiques, Geneve (1982) 237-254. [33] J. Kollir, Sharp Effective Nullstellensatz, J. Amer. Math. Sot. 1 (1988) 963-975. [34] T. Krick and L.M. Pardo, A computational method for diophantine approximation, in: L. Gonzalez and T. Recio, eds., Algorithms in Algebraic Geometry and Applications, Progress in Mathematics, Vol. 142 (Birkhiuser, Basel, 1996) 193-254. [35] T. Krick and L.M. Pardo, Une approche informatique pour I’approximation diophantienne, C.R. Acad. Sci. Paris, t. 318, Serie I (5) (1994) 407-412. [36] T. Krick, J. Sabia and P. Solemo, On intrinsic bounds in the Nullstellensatz, Appl. Algebra Eng. Commun. Comput. (AAECC J.) 8 (1997) 1255134. [37] E. Kunz, Kihler Differentials, Advanced Lectures in Mathematics (Vieweg Verlag, Braunschweiglwiesbaden, 1986). [38] B. Iversen, Generic Local Structure of the Morphisms in Commutative Algebra, Lecture Notes in Computer Science, Vol. 310 (Springer, Berlin, 1973). [39] M. Mignotte, Mathematiques pour le Calcul Formel (Presses Universitaires de France, 1989). [40] J.L. Montana, J.E. Morais and L.M. Pardo, Lower hounds for arithmetic networks II: sum of Betti numbers, Appl. Algebra Eng. Commun. Comput. (AAECC J.) 7 (1) (1996) 41-51. [41] J.L. Montana and L.M. Pardo, Lower bounds for arithmetic networks, Appl. Algebra Eng. Commun. Comput. (AAECC J.) 4 (1993) l-24. [42] Yu.V. Nesterenko, Estimates for the order of zero o functions of a certain class and applications in the theory of transcendental numbers, Math. USSR Izvestija 11 (2) (1977) 239-270. [43] Yu.V. Nesterenko, On a measure of the algebraic independence of the values of certain functions, Math. USSR Sbornik 56 (1987) 5455567. [44] L.M. Pardo, How lower and upper complexity bounds meet in elimination theory, in: G. Cohen, M.Giusti and T. Mora, Eds., Proc. 11th Intemat. Symp. Applied Algebra, Algebraic Algorithms and ErrorCorrecting Codes, AAECC-I 1, Paris 1995, Lecture Notes in Computer Science, Vol. 948 (Springer, Berlin, 1995) 33-69. [45] P. Philippon, Criteres pour I’independence algebrique, Pub. Math. de I’IHFS 64 (1987) 5-52. [46] P. Philippon, Sur des hauteurs alternatives, I, Math. Ann. 289 (1991) 255-283. [47] P. Philippon, Sur des hauteurs alternatives, II, Ann. Inst. Fourier, Grenoble, 44 (2) (1994) 1043-1065. [48] P. Philippon, Sur des hauteurs alternatives, III, J. Math. Pures Appl. 74 (1995) 345-365. [49] J. Sabia and P. Solemo, Bounds for traces in complete intersections and degrees in the Nullstellensatz, Appl. Algebra in Eng. Commun. Comput. (AAECC J.) 6 (1996) 353-376. [50] T. Schneider, Einfiihrung in die Theorie der transzendenten Funktionen (Springer, Berlin, 1957). [51] M. Shub and S. Smale, Complexity of Bezout’s theorem I: Geometric aspects, J. Amer. Math. Sot. 6 (1993) 4599501. [52] M. Shub and S. Smale, Complexity of Bezout’s theorem II: Volumes and probabilities, in: F. Eyssette and A. Galligo, eds., Computational Algebraic Geometry, Progress in Mathematics, Vol. 109 (Birkhauser, Basel, 1993) 267-285. [53] M. Shub and S. Smale, Complexity of Bezout’s theorem III: condition number and packing, J. Complexity 9 (1993) 4-14. [54] M. Shub and S. Smale, Complexity of Bezout’s theorem V: polynomial time, Theoret. Comput. Sci. 133 (1994). [55] M. Shub and S. Smale, On the intractability of Nilbert’s Nullstellensatz and an algebraic version of NP # P?. IBM Research Report, Yorktown Heights (1994). [56] M. Shub and S. Smale, Complexity of Bezout’s theorem IV: probability of success, extensions, SIAM J. Numer. Anal., to appear. [57] C. SOL&, Personal communication. [58] V. Strassen, Vermeidung von Divisionen, Crepe J. Reine Angew. Math. 264 (1973) 184-202.