Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Mar
Giusti
UMS CNRS{Polyte
hnique MEDICIS
Laboratoire GAGE
E
ole Polyte
hnique
F-91128 Palaiseau
edex, Fran
e
Email: Mar
.Giustigage.polyte
hnique.fr
Joos Heintz
Depto. de Matem ati
as, Est. y Comp.
Fa
ultad de Cien
ias Universidad de Cantabria
E-39071 Santander, Spain and
Depto. de Matem ati
as
Fa
ultad de Cien
ias Exa
tas y Naturales
Universidad de Buenos Aires
Ciudad Universitaria, Pab. I
(1428) Buenos Aires, Argentina
Email: heintzhall.mates
o.uni
an.es
Abstra
t
This paper is devoted to the
omplexity analysis of
ertain uniformity
properties owned by all known symboli
methods of parametri
poly-
nomial equation solving (geometri
elimination). It is shown that any
parametri
elimination pro
edure whi
h is parsimonious with respe
t
to bran
hings and divisions must ne
essarily have a non-polynomial se-
quential time
omplexity, even if highly eÆ
ient data stru
tures (as e.g.
the arithmeti
ir
uit en
oding of polynomials) are used.
1 Introdu
tion
Origins, development and intera
tion of modern algebrai
geometry and
ommutative algebra may be
onsidered as one of the most illustra-
tive examples of histori
al diale
ti
s in mathemati
s. Still today, and
more than ever before, timeless idealism (in form of modern
ommuta-
Resear
h partially supported by the following Argentinian, Spanish, German and
Fren
h grants : UBA-CYT TW 80, PIP CONICET 4571/96, ANPCyT PICT 03-
00000-01593, DGICYT PB96{0671{C02{02, ARG 018/98 INF (BMBF), UMS 658
MEDICIS, ECOS A99E06, HF 1999{055
1
2 Giusti and Heintz
tive algebra) is bravely struggling whith se
ular materialism (in form of
omplexity issues in
omputational algebrai
geometry).
Krone
ker was doubtless the
reator of this eternal battle eld and its
rst war lord. In a similar way as Gauss did for
omputational num-
ber theory, Krone
ker laid intuitively the mathemati
al foundations of
modern
omputer algebra. He introdu
ed 1882 in [24℄ his famous "elim-
ination method" for polynomial equation systems and his "parametri
representation" of (equidimensional) algebrai
varieties. By the way, this
parametri
representation was until 10 years ago redis
overed again and
again. It entered in modern
omputer algebra as "Shape Lemma" (see
e.g. [36, 7, 10, 22℄). Using his elimination method in a highly skillful,
but unfortunately inimitable way, Krone
ker was able to state and to
prove a series of fundamental results on arbitrary algebrai
varieties. He
was able to dene in a pre
ise way the notion of dimension and to prove
a
orresponding dimension theorem for arbitrary algebrai
varieties over
an algebrai
ally
losed eld, to estimate the number of equations needed
to dene any algebrai
variety in aÆne or proje
tive spa
e and
ertainly
he knew already the spe
ial form of \Hilbert's Nullstellensatz".
Not everything that
ame to Krone
ker's mind was laid down by him
in a expli
it and written form. Nevertheless a
areful interpretation of
his work suggests his deep understanding of the general stru
ture of
algebrai
varieties.
A parti
ular result, proved by Krone
ker, says that any algebrai
vari-
ety
an be dened by nitely many equations. Later Hilbert generalized
this result to his seminal \Basissatz" introdu
ing for its proof a new,
non
onstru
tive method, far away from the traditional elimination-type
arguments used by Krone
ker and other
ontemporary mathemati
ians.
It took some time to
onvin
e the mathemati
al world that \mysti
s"
and \magi
ians" are able to produ
e (
orre
t) mathemati
al results, but
nally the new dis
ipline of
ommutative algebra be
ame legitimate.
Hilbert's dis
overy of the Basissatz was also the starting point for
a long and huge
on
i
t whi
h dominated a
onsiderable part of the
history of modern algebrai
geometry and whi
h did not
ome to an end
until today.
Classi
al algebrai
geometry is motivated by the need { or the wish {
to nd tools whi
h allow to \solve" or to \redu
e" (whatever this means)
systems of polynomial equations. This leads to the following questions:
These questions, impli
itly raised by the work of Hilbert and Ma
au-
lay, look very a
ademi
. However any attempt to answer them leads to
deep
onsequen
es.
Let us here outline just one possible way to answer these questions.
If we
onsider the \Hauptproblem der Idealtheorie" as a question of
pure theoreti
al
omputer s
ien
e, this problem turns out to be
ompu-
tationally intra
table, at least in worst
ase. More pre
isely, the \Haupt-
problem der Idealtheorie" turns out to be
omplete in exponential (mem-
ory) spa
e (see [33, 43, 8℄). On the other hand, almost all of the most
fundamental problems of
omputational
lassi
al algebrai
geometry are
proved to be solvable in polynomial spa
e (see e.g. [32, 9, 30, 29, 25℄).
Thus
omputational
omplexity is able to distinguish between geometry
and algebra and supports the viewpoint of Krone
ker (geometry) against
the viewpoint of Hilbert and Ma
aulay (algebra).
It is well known that Krone
ker's personality was highly
on
i
tive
for his time. It is less known how mu
h posthumous reje
tion Kro-
ne
ker's personality was able to produ
e. Hilbert's writings are eloquent
in this point (\die Krone
kers
he Verbotsdiktatur", see [21℄), whereas
Ma
aulay's atta
k against Krone
ker's work on elimination was a rather
well edu
ated one (see [28℄, Prefa
e). In this
ontext let us also remind
Andre Weil's \elimination of the elimination theory". Krone
ker's radi-
al spirit did not re
ognise limitations. His radi
alism was as universal
as his spirit was. Of
ourse he was right requiring that any mathemati-
al re
e
tion has to end up with nitely many and pra
ti
ally realizable
omputations whi
h settle the
on
rete (e.g. appli
ation) problem un-
der
onsideration. However de
laring the natural numbers as the only
mathemati
al obje
ts
reated by god for mankind and de
laring the
\rest"(real,
omplex numbers, innite
ardinals and ordinals) as devils
work, tempting humans to play with the innite, he demonstrated that
he was not able or willing to distinguish between syntaxis and seman-
ti
s. He was right to require that mathemati
al expressions (algebra)
have always to move within nitary limits, but he was wrong to ex
lude
re
e
tions about innite mathemati
al obje
ts (geometry) using a ni-
4 Giusti and Heintz
tarian language. Of
ourse his own work ended up to be godless enough
to guarantee him mathemati
al re
ognition by his worst adversaries.
Krone
ker's own formulation of his elimination method in [24℄ was
impre
ise and general enough to allow dierent interpretations of it,
omputationally eÆ
ient and ineÆ
ient ones. Possibly it was in his
mind to leave open the door for future
omplexity issues of his method.
Hilbert's and Ma
aulay's atta
ks against Krone
ker's method were
based on the
omputationally ineÆ
ient interpretation. They noti
ed
that under this interpretation Krone
ker's elimination method leads to
a hyperexponential swell of intermediate expression size and used this
observation for the promotion of their own, more \simple" and \mathe-
mati
al" point of view. Hilbert and Ma
aulay's position be
ame nally
predominant in the future development of algebrai
geometry and
om-
mutative algebra (see e.g. [40℄). Their ideas led to the modern,
omputer
implemented tool of Grobner basis algorithms for the symboli
resolu-
tion (simpli
ation) of polynomial equation systems. This tool was in-
trodu
ed in the sixties by B. Bu
hberger (and his s
hool) and represents
today the
ore of all
urrent
omputer algebra software pa
kages.
The dis
overy of ee
tive (aÆne) Nullstellensatze and their appli
a-
tion to the
omplexity analysis of Grobner basis algorithms for the so-
lution of geometri
al problems, represented at the end of the eighties
the turning point for a pro
ess whi
h led nally ba
k to Krone
ker's
original ideas. This pro
ess was not a
ons
ious motion with
lear goals,
but rather a slow emerging of mathemati
al insight and of algorithmi
design within the s
ope of Krone
ker's intuitions.
The rst step in this dire
tion was even made as part of a mathemat-
i
al proof and not of an algorithmi
design. In its standard interpre-
tation, Krone
ker's elimination method relies on an iterated use of re-
sultants of suitable univariate polynomials with parametri
oeÆ
ients.
In fa
t, a resultant is nothing but the
onstant term of the
hara
teris-
ti
polynomial of a suitable linear map determined by the polynomials
under
onsideration. Repla
ing in this version of Krone
ker's elimina-
tion algorithm the o
uring resultants just by the
onstant terms of the
orresponding minimal polynomials, one obtains an enormous redu
-
tion of degree, height and also of arithmeti
ir
uit (straight-line pro-
gram)
omplexity for the polynomials produ
ed during the pro
edure
(see e.g. [4, 5, 16℄). By the way, let us remark that Krone
ker applies
in his method the mentioned simpli
ation, however he omits to draw
these important
on
lusions from his argument.
Other important ingredients of the emerging new algorithmi
method
Krone
ker's bla
k boxes 5
were the parametri
representation of equidimensional algebrai
vari-
eties (a redis
overy of Krone
ker's old idea) and the arithmeti
ir
uit
(straight-line program) representation of polynomials (whi
h was nei-
ther out of the s
ope of Krone
ker's intuition nor
learly in
luded in the
main stream of his thinking). The
ombination of these ingredients pro-
du
ed a
onsiderable ee
t upon the worst
ase
omplexity of symboli
elimination pro
edures (see e.g. [12, 23, 14℄).
Nevertheless all this progress did not suÆ
e to allow the design of
pra
ti
ally eÆ
ient symboli
elimination algorithms and in parti
ular
of algorithms whi
h were able to
ompete in
omplexity aspe
ts with
their numeri
al
ounterparts. The traditional huge gap between sym-
boli
and numeri
polynomial equation solving methods remained open.
Whereas numeri
algorithms are very eÆ
ient with respe
t to the num-
ber of arithmeti
al operations they require, they
annot be eÆ
iently
used in
ase of parametri
, underdetermined, overdetermined or degen-
erate polynomial equation systems. Symboli
algorithms are free from
these restri
tions but they are also too ineÆ
ient for any reasonable use
in pra
ti
e. A way out of this dilemma be
ame apparent in [13, 11℄ by
a new interpretation of Newton's
lassi
al method. Interpreting New-
ton's approximation algorithm as a global pro
edure instead as a lo
al
one, allowed its use as a tool for data
ompression in the Krone
ker{like
elimination pro
edure of [14℄ ( whi
h relied on the arithmeti
ir
uit rep-
resentation of polynomials). It turns out that Newton's method is well
adapted to exa
t symboli
omputation if the
orre
t (seminumeri
al)
data stru
ture is used. In this way a new algorithmi
method nally
emerged. This method is based on a
ombination of Krone
ker's and
Newton's ideas, and is able to distinguish di
hotomi
ally between \well
behaved" and \badly behaved" polynomial equations systems. Moreover
this algorithmi
method is optimal for worst
ase (i.e. generi
) systems.
Unlike Grobner basis algorithms this new method avoids any signif-
i
ant
omputational over
ow during its exe
ution. Roughly speaking,
the new algorithms are always polynomial in the output size and even
polynomial in the input size, if the given polynomial equation system is
\well behaved".
This view of algorithmi
algebrai
geometry produ
ed also the follow-
ing new insight:
elimination polynomials are always \smart"(i.e. not easy and not
hard) to evaluate. How many variables they ever may
ontain, their eval-
uation
omplexity is always polynomial in their degree (whereas their
6 Giusti and Heintz
number of monomials may be exponential in the number of their vari-
ables).
Although this insight is beyond of the s
ope of Krone
ker's way of
thinking, his formulation of the elimination method indi
ates that he
might have intended to leave the door open for
omplexity issues.
A further development of this new algorithmi
method was able to
demonstrate its pra
ti
al eÆ
ien
y (see [27℄ and [31℄). Moreover it
turns out that the parameters whi
h dominate the
omplexity of the
new symboli
pro
edures determine also the eÆ
ien
y of their numeri
al
ounterparts (at least if aspe
ts of diophantine approximation are taken
into a
ount; see [34℄). The new symboli
algorithms (and their numeri-
al
ounterparts) have still a worst
ase
omplexity whi
h is exponential
in the (purely synta
ti
al) input length. One may ask whether this fa
t
is due to the algorithms and data stru
tures employed or whether this
is due to the intrinsi
nature of geometri
elimination.
The new results presented in this paper address this question. In
order to dis
uss this point, let us turn ba
k to our interpretation of
Krone
ker's ideas behind geometri
elimination. Krone
ker's elimina-
tion method (and theory) behaves well under spe
ialisation of the input
equations. In fa
t, at any moment of the pro
edure one may
onsider
the (parametri
) input equations as given by their
oeÆ
ients and these
oeÆ
ients may be
onsidered as purely algebrai
obje
ts, determined
only by their algebrai
relations. A given equation system may even be
\generalized", i.e. the
oeÆ
ients of the given input equations may be
repla
ed by indeterminates and the dis
ussion of the solvability of the
generalized systems answers all imaginable questions about the solvabil-
ity of the original system. This
on
ept of spe
ialisation{generalisation
reveals an idea of universality (or uniformity) behind Krone
ker's elimi-
nation theory. This philosophi
al idea of universality be
ame one of the
orner stones of modern algebrai
geometry and
ommutative algebra.
Sin
e we are (still) unable to think in a dierent way, we shall
onsider
the input equations of a parametri
elimination problem as fun
tions
whi
h may be
alled by their values in the variables to be eliminated
(bla
k-box representation) or simply as being given by their
oeÆ
ients
(formal representation).
Any elimination algorithm we are able to imagine today starts from
this kind of data. In other words, the bla
k box representation
onsti-
tutes today the most general way we may think the equations of our
input system to be given. Of
ourse ea
h evaluation of the input equa-
Krone
ker's bla
k boxes 7
tions has its
osts and these
osts may be measured by the size of a
division-free arithmeti
ir
uit whi
h represents the input equations.
In this paper we shall show that any suÆ
iently uniform elimination
pro
edure (whi
h avoids super
uous bran
hings) be
omes ne
essarily
exponential in worst
ase, if the input equations are given in bla
k-box
representation and if the required output is a
anoni
al elimination poly-
nomial ( a \resolvent" in the terminology of Krone
ker and Ma
aulay).
This is a general and provable fa
t for any symboli
as well as for any
numeri
al elimination pro
edure. We shall also show, that even in
ase
that the input equations are not given by a bla
k box, but by an expli
it
arithmeti
ir
uit, the same
on
lusion holds true for any suÆ
iently
universal and uniform elimination pro
edure whi
h is able to
ompute
eÆ
iently Zariski
losures and (generi
ally squarefree) parametri
great-
est
ommon divisors for
ir
uit represented algebrai
families of polyno-
mials.
Below we shall give a pre
ise denition of the meaning of a \suÆ
iently
universal and uniform elimination pro
edure" for the
ase of a
at family
of elimination problems. Su
h an elimination pro
edure will be
alled
parametri
.
If an elimination pro
edure is used in order to assign \
oordinates"
(i.e. in order to parametrise) suitable and easy-to-represent families of
algebrai
varieties the same
on
lusion holds again, without any further
restri
tion on the input representation (see [6℄ and [18℄). Summing
up, we may say: the vision of elimination theory initiated by Krone
ker
hides a
on
ept of universality and uniformity whi
h obstru
ts its general
eÆ
ien
y. The question what happens with
omplexity when we drop
this universality and uniformity requirement, ex
eeds the horizon of to-
days mathemati
al thinking and is equivalent to the question whether
P 6= NP holds over the
omplex numbers (in the sense of the BSS
omplexity model; see [2℄).
2 Parametri
elimination
2.1 Parametri
elimination pro
edures
The pro
edures (algorithms)
onsidered in this paper operate with essen-
tially division-free arithmeti
ir
uits (straight-line programs) as basi
data stru
tures for the representation of inputs and outputs. Su
h a
ir
uit depends on
ertain input nodes, labelled by indeterminates over
a given ground eld k (in the sequel we shall suppose that k is innite
8 Giusti and Heintz
and perfe
t with algebrai
losure k). These indeterminates are thought
to be subdivided into two disjoint sets representing the parameters and
variables of the given
ir
uit. The
ir
uit nodes of indegree zero whi
h
are not inputs are labelled by elements of k, whi
h are
alled the s
alars
of the
ir
uit (here "indegree" means the number of in
oming edges of
the
orresponding node). Internal nodes are labelled by arithmeti
oper-
ations (addition, subtra
tion, multipli
ation and division). The internal
nodes of the
ir
uit represent polynomials in the variables of the
ir
uit.
The
oeÆ
ients of these polynomials belong to the parameter eld K ,
generated over the ground eld k by the parameters of the
ir
uit. This
is a
hieved by allowing in a essentially division-free
ir
uit only divisions
whi
h involve elements of K . Thus essentially division-free
ir
uits do
not
ontain divisions involving intermediate results whi
h depend on
variables. A
ir
uit whi
h
ontains only divisions by nonzero elements
of k is
alled totally division-free. The output nodes of an essentially
division-free
ir
uit may o
ur labelled by sign marks of the form \= 0"
or \6= 0" or may remain unlabelled (by sign marks). Thus the given
ir
uit represents by means of its labelled output nodes a system of
polynomial equations and inequations whi
h determine in their turn a
lo
ally
losed set (i.e. an embedded aÆne variety) with respe
t to the
Zariski topology of the aÆne spa
e of parameter and variable instan
es.
The unlabelled output nodes of the given
ir
uit represent a polyno-
mial appli
ation (in fa
t a morphism of algebrai
varieties) whi
h maps
this lo
ally
losed set into a suitable aÆne spa
e. We shall interprete
the system of polynomial equations and inequations represented by the
ir
uit as a parametri
family of systems in the variables of the
ir
uit.
The
orresponding varieties
onstitute an parametri
family of varieties.
The same point of view is applied to the morphism determined by the
unlabelled output nodes of the
ir
uit. We shall
onsider this morphism
as a parametri
family of morphisms.
To a given essentially division-free arithmeti
ir
uit we may asso-
iate dierent
omplexity measures and models. In this paper we shall
be mainly
on
erned with sequential
omputing time, measured by the
size of the
ir
uit. O
asionally we will also refer to parallel time, mea-
sured by the depth of the
ir
uit. In our main
omplexity model is the
total one, where we take into a
ount all arithmeti
operations (addi-
tions, subtra
tions, multipli
ations and possibly o
uring divisions) at
unit
osts. For purely te
hni
al reasons we shall also
onsider two non-
s
alar
omplexity models, one over the ground eld k and the other one
over the parameter eld K . In the non-s
alar
omplexity model over
Krone
ker's bla
k boxes 9
K we
ount only the essential multipli
ations (i.e. multipli
ations be-
tween intermediate results whi
h a
tually involve variables and not only
parameters). This means that K -linear operations (i.e. additions and
multipli
ations by arbitrary elements of K ) are
ost free. Similarly, k-
linear operations are not
ounted in the non-s
alar model over k. For
more details about
omplexity measures and models we refer to [3℄.
Given an essentially division-free arithmeti
ir
uit as input, an elimi-
nation problem
onsists in the task of nding an essentially division-free
output
ir
uit whi
h des
ribes the Zariski
losure of the image of the
morphism determined by the input
ir
uit. The output
ir
uit and the
orresponding algebrai
variety are also
alled a solution of the given
elimination problem. We say that a given parameter point xes an in-
stan
e of the elimination problem under
onsideration. In this sense
a problem instan
e is des
ribed by an input and an output (solution)
instan
e.
In this paper we restri
t our attention to input
ir
uits whi
h are
totally division-free and
ontain only output nodes labelled by \=0" and
unlabelled output nodes. Mostly our output
ir
uits will also be totally
division-free and will
ontain only one output node, labelled by the mark
\=0". This output node will always represent a
anoni
al elimination
polynomial asso
iated to the elimination problem under
onsideration
(see Se
tion 2.2 for more details).
In
ase that our output
ir
uit
ontains divisions (depending only on
parameters but not on variables), we require to be able to perform these
divisions for any problem instan
e. In order to make this requirement
sound, we admit in our algorithmi
model
ertain limit pro
esses in the
spirit of de l'H^opital's rule (below we shall modelise these limit pro-
esses algebrai
ally, in terms of pla
es and valuations). The restri
tion
we impose on the possible divisions in an output
ir
uit represents a
rst fundamental geometri
uniformity requirement for our algorithmi
model.
An algorithm whi
h solves a given elimination problem may be
on-
sidered as a (geometri
) elimination pro
edure. However this simple
minded notion is too restri
tive for our purpose of showing lower
om-
plexity bounds for elimination problems. It is thinkable that there exists
for every individual elimination problem an eÆ
ient ad ho
algorithm,
but that there is no universal way to nd and to represent all these ad ho
pro
edures. Therefore, a geometri
elimination pro
edure in the sense of
this paper will satisfy
ertain uniformity and universality requirements
whi
h we are going to explain now.
10 Giusti and Heintz
We modelise our elimination pro
edures by families of arithmeti
net-
works (also
alled arithmeti
-boolean
ir
uits) whi
h solve entire
lasses of elimination problems of arbitrary input size (see [41℄, [42℄).
In this sense we shall require the universality of our geometri
elimina-
tion pro
edures. Moreover, we require that our elimination pro
edures
should be essentially division-free.
In a universal geometri
elimination pro
edure, bran
hings and di-
visions by intermediate results (that involve only parameters, but not
variables)
annot be avoided. From our elimination pro
edures we shall
require to be parsimonious with respe
t to bran
hings (and divisions).
In parti
ular we shall require that our elimination pro
edures do not in-
trodu
e bran
hings and divisions for the solution of a given elimination
problem when traditional algorithms do not demand this (an example of
su
h a situation is given by the
at families of elimination problems we
are going to
onsider in the sequel). This restri
tion represents a se
ond
fundamental uniformity requirement for our algorithmi
model.
We
all a universal elimination pro
edure parametri
if it satises our
rst and se
ond uniformity requirement, i.e. if the pro
edure does not
ontain bran
hings whi
h otherwise
ould be avoided and if all possibly
o
urring divisions
an be performed on all problem instan
es, in the
way we have explained before. In this paper we shall only
onsider
parametri
elimination pro
edures.
We
all a parametri
elimination pro
edure geometri
ally robust if it
produ
es for any input instan
e an output
ir
uit whi
h depends only
on the mathemati
al obje
ts \input equation system" and \input mor-
phism" but not on their
ir
uit representation. We shall apply this
notion only to elimination problems given by (geometri
ally or s
heme-
theoreti
ally)
at families of algebrai
varieties. This means informally
that a parametri
elimination pro
edure is geometri
ally robust if it pro-
du
es for
at families of problem instan
es \
ontinuous" solutions.
Of
ourse, our notion of geometri
robustness depends on the (geo-
metri
or s
heme-theoreti
al)
ontext, i.e. it is not the same for s
hemes
or varieties. In Se
tion 2.2 we shall explain our idea of geometri
robust-
ness in the typi
al situation of
at families of algebrai
varieties given
by redu
ed
omplete interse
tions.
Traditionally, the size of a system of polynomial equations (and in-
equations) is measured in purely extrinsi
, synta
ti
terms (e.g. number
of parameters and variables, degree of the input polynomials, size and
depth of the input
ir
uit et
). However, there exists a new generation of
symboli
and numeri
algorithms whi
h take also into a
ount intrinsi
,
Krone
ker's bla
k boxes 11
semanti
(e.g. geometri
or arithmeti
) invariants of the input equa-
tion system in order to measure the
omplexity of elimination pro
edure
under
onsideration more a
urately (see e.g. [13, 11, 15℄ and [38, 39℄).
In this paper we shall turn ba
k to the traditional point of view. In
Theorem 5 we shall show that, under our universality and uniformity re-
stri
tions, no parametri
elimination pro
edure whi
h in
ludes eÆ
ient
omputation of Zariski
losures and of generi
ally squarefree paramet-
ri
greatest
ommon divisors for
ir
uit represented algebrai
families of
polynomials, is able to solve an arbitrary elimination problem in polyno-
mial (sequential) time, if time is measured in terms of
ir
uit size and
input length is measured in synta
ti
al terms only.
Finally let us refer to the books [3℄, [35℄ and [37℄ as a general ba
k-
ground for notions of algebrai
omplexity theory and algebrai
geome-
try we are going to use in this paper.
Claim: For xed N and k, the
oeÆ
ient `k;j
an be written as a poly-
nomial expression of degree exa
tly k 1 in the index j . Moreover, this
polynomial expression for `k;j has integer
oeÆ
ients.
Proof of the Claim. We pro
eed by indu
tion on the index parameter k .
For k = 1 we have `1;j = 1 for any 0 j N and therefore `1;j is a
polynomial of degree k 1 = 0 in the index j .
Let 1 k N + 1. Assume indu
tively that `k;j is a polynomial
of degree exa
tly k 1 in the index j and that the
oeÆ
ients of this
polynomial are integers. We are now going to show that `k+1;j is a
polynomial of degree exa
tly k in j and that the
oeÆ
ients of this
polynomial are integers too. Observe that
X
`k+1;j = j1 j k
0 j1 <<jk N
jr 6=j for r=1;:::;k
!
X X
= j1 j k j j1 j k 1 :
0j1 <<jk N 0 j1 <<jk 1 N
jr 6=j for r=1;:::;k 1
holds. Sin
e the term
X
j1 jk
0 j1 <<jk N
Krone
ker's bla
k boxes 19
does not depend on j and sin
e by indu
tion hypothesis
X
`k;j = j1 j k 1
j1 <<jk 1 N
0
jr 6=j for r=1;:::;k 1
is a polynomial of degree exa
tly k 1 in j , we
on
lude that `k+1;j is
a polynomial of degree exa
tly k in j . Moreover, the
oeÆ
ients of this
polynomial are integers. This proves our
laim.
It is now easy to nish the proof of Lemma 1. By our
laim there
exist for arbitrary 1 k N + 1 integers
(0k) ; ;
(kk)1 with
(kk)1 6= 0
su
h that for any 0 j N the identity `k;j =
(0k) + +
(kk)1 j k 1
holds. Hen
e for arbitrary 0 k N there exist rational numbers
1(k) ; : : : ; (kk+1
)
(not depending on j ) su
h for any 0 j N the
ondition
j k = (1k) `1;j + + k(k+1
)
`k+1;j
is satised (here we use the
onvention 00 := 1). This implies for any
index 0 k N the polynomial identity
X n
Y
(1k) L1 + + (kk+1
)
Lk+1 = jk Ui[j℄i :
jN
0 i=1
X n
Y
Hen
e for any 0 kN the polynomial Pk := jk Ui[j℄i
0j N i=1
belongs to the Q -ve
tor spa
e generated by L1 ; : : : ; LN +1 .
On the other hand, we dedu
e from the nonsingularity of the
Vandermonde matrix j k 0k;jN that the polynomials P0 ; : : : ; PN are
Q -linearly independent. Therefore the Q -ve
tor spa
e generated by
L1; : : : ; LN +1 in Q [U1 ; : : : ; Un ℄ has dimension N + 1 = 2n . This implies
that L1 ; : : : ; L2n are Q -linearly independent.
With the notations of Se
tion 2.2, put now r := n + 1, T := Un+1 and
let us
onsider the following polynomials of Q [T; U1; : : : ; Un; X1 ; : : : ; Xn ℄:
G1 := X1 2 X1 ; : : : ; Gn := Xn 2 Xn and
n
X n
Y
F := 2i 1
Xi + T (1 + (Ui 1)Xi ): (3.3)
i=1 i=1
It is
lear from their denition that the polynomials G1 ; : : : ; Gn and
F
an be evaluated by a (non-invariant) totally division-free arithmeti
ir
uit of size O(n) in Q [T; U1 ; : : : ; Un ; X1 ; : : : ; Xn ℄. Observe that the
polynomials G1 ; : : : ; Gn do not depend on the T; U1; : : : ; Un and that
20 Giusti and Heintz
their degree is two. The polynomial F is of degree 2n + 1. More pre-
isely, we have degX1 ;:::;Xn F = n, degU1 ;:::;Un F = n, and degT F = 1.
Although the polynomial F
an be evaluated using O(n) arithmeti
op-
erations, the sparse representation of F , as a polynomial in the variables
T; U1; : : : ; Un ; X1 ; : : : ; Xn ,
ontains asymptoti
ally 2n non-zero mono-
mial terms.
Let us now verify that the polynomials G1 ; : : : ; Gn and F form a
at
family of elimination problems depending on the parameters T; U1 ; : : : ; Un .
The variety V := fG1 = 0; : : : ; Gn = 0g is nothing but the union
of 2n aÆne linear subspa
es of A n+1 A n , ea
h of them of the form
A n+1 f g, where is any point of the hyper
ube f0; 1gn. The
anoni
al
proje
tion A n+1 A n ! A n+1 indu
es a morphism : V ! A n+1
whi
h glues together the
anoni
al proje
tions A n+1 f g ! A n+1 for
any in f0; 1gn. Obviously the morphism is nite and generi
ally
unramied. In parti
ular has
onstant bres whi
h are all
anoni
ally
isomorphi
to the hyper
ube P f0; 1gn. Let (j1 ; : : : ; jn ) be an arbitrary
point of f0; 1g and let j := 1in ji 2i 1 be the integer 0 j < 2n
n
whose bit representation is jn jn 1 : : : j1 . One veries immediately the
identity
n
Y
F (T; U1; : : : ; Un ; j1 ; : : : ; jn ) = j + T Uiji :
i=1
Pn
Therefore for any point (t; u1 ; : : : ; un ; j1 ; : : : ; jn ) 2 V with j := 1 ji 2i 1
we have
n
Y
F (t; u1 ; : : : ; un ; j1 ; : : : ; jn ) = j + T uji i :
i=1
>From this observation we dedu
e easily that the polynomial
P 2 Q [T; U1; : : : ; Un ; Y ℄
we are looking for (as the elimination polynomial asso
iated to the
at
family G1 ; : : : ; Gn ; F ) is in fa
t
n
2Y1 n
Y
P= (Y (j + T Ui [j℄i )):
j =0 i=1
With the notations of the beginning of this se
tion, this polynomial has
the form
n X n n X n
P =Y2 + Ak Y 2 k Y 2 + (ak +T Lk )Y 2 k modulo T 2 ;
1k2n 1k2n
(3.4)
Krone
ker's bla
k boxes 21
P
with ak := 1j1 <:::<jk 2n 1 j1 : : : jk for 1 k 2n .
Suppose now that there is given a geometri
ally robust parametri
elimination pro
edure. This pro
edure produ
es from the input
ir
uit
an invariant arithmeti
ir
uit in Q [T; U1; : : : ; Un ; Y ℄, with s
alars
in Q [T; U1; : : : ; Un ℄, whi
h evaluates the polynomial P . Re
all that the
invarian
e of means that the s
alars of are invariant polynomials of
Q [T; U1 ; : : : ; Un ℄, say
1 ; : : : ;
N .
Let L( ) be the total and L( ) the non-s
alar size of the arithmeti
ir
uit over Q [T; U1; : : : ; Un℄. Without loss of generality we have
L( ) L( ) and N (L( ) + 3)2 .
Let Z1 ; : : : ; ZN be new indeterminates. From the graph stru
ture
of the
ir
uit we dedu
e that there exists for ea
h 1 k 2n a
polynomial Qk 2 Q [Z1 ; : : : ; ZN ℄ satisfying the
ondition
Qk (
1 ; : : : ;
N ) = Ak (3.5)
(see [3℄ for details). Let us now
onsider two arbitrary elements
u; u0 2 A n . Observe that the parameter points (0; u) and (0; u0 ) of
A n+1 are equivalent (in symbols: (0; u) (0; u0 )). From the invarian
e
of
1 ; : : : ;
N we dedu
e therefore that
j (0; u) =
j (0; u0 ) holds for
any 1 j N . This means that the polynomials
!1 =
1 (0; U1; : : : ; Un ); : : : ; !N :=
N (0; U1 ; : : : ; Un )
are independent from the variables U1 ; : : : ; Un and therefore elements
of Q . From identity (3.4) we
on
lude that the same is true for 1 :=
A1 (0; U1 ; : : : ; Un ); : : : ; 2n := A2n (0; U1; : : : ; Un ). We shall abbreviate
! := (!1 ; : : : ; !N ) and := (1 ; : : : ; 2n ).
Let us
onsider the morphisms of aÆne spa
es : A n+1 ! A N and
2n
: A N ! A given by := (
1 ; : : : ;
N ) and := (Qk )1k2n .
Observe that
Æ = (Qk (
; : : : ;
N )) k n = (Ak ) k n
1 1 2 1 2
Proof From [20, 23℄ we dedu
e that there exist in M n
orre
t test
sequen
es of length m := 6(2L + n + 1)2 for H2L;n and that su
h a
orre
t test sequen
e 2
an be found with probability of su
ess
(1 4LD 12(2L+n+1) ) by a random
hoi
e in M n .
Krone
ker's bla
k boxes 25
Let
:= (
1 ; : : : ;
m ) 2 kmn be a
orre
t test sequen
e for H L;n .
2
Let H1 ; H2 2 WL;n and suppose that
H1 (
1 ) = H2 (
1 ); : : : ; H1 (
m ) = H2 (
m )
holds. Observe that H := H2 H1 belongs to W2L;n . Therefore we
have H (
1 ) = 0; : : : ; H (
m ) = 0, and, sin
e
is a
orre
t test sequen
e
for H2L;n, we may
on
lude H = H2 H1 = 0.
Let now m := 6(2L + n + 1)2 . Then, by Lemma 3, we may x an
identi
ation sequen
e
:= (
1 ; : : : ;
m ) 2 kmn for the
omplexity
lass H = HL;n . Let S1 ; : : : ; Sm be new indeterminates denoting the
anoni
al
oordinate fun
tions of A m . We are going to
onsider the
(
) (
)
morphism L;n : WL;n ! A m , dened for H 2 WL;n by L;n (H ) :=
(
)
(H (
1 ); : : : ; H (
m )). Let SL;n be the Zariski
losure of the image of the
(
) (
) (
)
morphism L;n and let us abbreviate := L;n and S := SL;n .
Sin
e the variety W and the morphism : W ! S are k-denable,
we
on
lude that S is k-denable too. Moreover for any polynomial
H 2 W and any value 2 k we have
(H ) = (H (
1 ); : : : ; H (
m )) = (H )
and from this homogeneity property of we
on
lude that S is a
one
over the eld k. Therefore S is denable by homogeneous polynomials
belonging to k[S1 ; : : : ; Sm ℄. Sin
e the variety W is irredu
ible and S is
the
losure of the image of , we
on
lude that S is irredu
ible too.
By assumption
:= (
1 ; : : : ;
m ) is an identi
ation sequen
e for the
omplexity
lass H = HL;n. Therefore we
on
lude that : W ! S is
an inje
tive, dominant morphism. Hen
e is birational.
Let = (1 ; : : : ; m ) where 1 ; : : : ; m are suitable
oordinate fun
-
tions of the aÆne variety W . Let us
onsider W as a
losed subvariety
of the aÆne spa
e A D . We re
all that the points of A D
orrespond to
the polynomials of k[X1 ; : : : ; Xn ℄ of degree at most D and that the mor-
phism is dened by means of the evaluation of the polynomials of W
in the points
1 ; : : : ;
m . This implies that there exist linear forms in
the
oordinate ring of A D su
h that 1 ; : : : ; m are the restri
tions of
these linear forms to the variety W . From the inje
tivity of we dedu
e
that W \ f1 = 0; : : : ; m = 0g
ontains only the origin of A D . This
implies that the homogeneous map indu
es a nite morphism between
the proje
tive varieties asso
iated to the
ones W and S . In fa
t, the
standard proof of this
lassi
al result implies something more, namely
that also the morphism : W ! S is nite (see [37℄ I.5.3, Theorem
26 Giusti and Heintz
8 and proof of Theorem 7). In parti
ular, is a surje
tive,
losed map
and hen
e a homeomorphism with respe
t to the Zariski topologies of
W and S .
Thus we have shown the following result:
(
)
Sin
e the varieties WL;n and SL;n are irredu
ible, we
on
lude that
(
)
their
oordinate rings k[WL;n ℄ and k[SL;n ℄ are domains with fun
tion
(
) (
)
elds k(WL;n) and k(SL;n ). The morphism L;n indu
es an embedding
(
)
of the
oordinate ring k[SL;n ℄ into k[WL;n℄ (and the same is true for the
orresponding fun
tion elds). Moreover, the niteness of the morphism
(
) (
)
L;n means that k[WL;n℄ is an integral ring extension of k[SL;n ℄.
Disregarding the
omplexity aspe
t, Lemma 4 says that it is possible
to re
onstru
t the
oeÆ
ients of a polynomial H 2 HL;n from the values
of H in a given identi
ation sequen
e
, even if the sequen
e
is
short in
omparison with the degree of H . Lemma 4 says further that
this re
onstru
tion is rational (i.e. it uses only arithmeti
al operations)
and that possibly o
uring divisions may always be performed by limit
pro
esses in the spirit of de l'H^opital's rule. These pro
esses produ
e
only nite limits, be
ause the
oeÆ
ients of H are integrally dependent
on the values of H in the given identi
ation sequen
e
. In algebrai
terms, we may modelise these limit pro
esses by pla
es (
orresponding to
(
) (
)
valuation rings) whi
h map the fun
tion eld k(SL;n ) of the variety SL;n
into the set of values k [f1g and take only nite values on k[SL;n℄. The ( )
4 Con
lusions
It was fundamental for our argumentation in Se
tion 3 that our notion
of parametri
elimination pro
edure restri
ts severely the possibility of
bran
hings in the output
ir
uit or network. This suggests that any
polynomial time elimination algorithm (if there exists one) must have a
huge topologi
al
omplexity. Thus hypotheti
al eÆ
ien
y in geometri
elimination seems to imply
ompli
ated
asuisti
s.
Let us also mention that the proof method of Se
tion 3
ontributes
absolutely nothing to the elu
idation of the fundamental thesis of al-
gebrai
omplexity theory, whi
h says that geometri
elimination pro-
du
es elimination polynomials whi
h are intrinsi
ally hard to evaluate
(i.e. these polynomials need asymptoti
ally degree{mu
h sequential time
for their evaluation). A good example for this thesis is the Po
hhammer{
Wilkinson polynomial whi
h
an be obtained easily as the solution of a
suitable elimination problem but whi
h is
onje
turally hard to evaluate
(see e.g. [19, 1℄) Similarly no advan
e is obtained by our method with
respe
t to the question whether PC 6= NPC holds in the BSS
omplexity
model, see [2℄, Chapter7.
In fa
t our
ontribution
onsists only in the dis
overy of a very lim-
iting uniformity property present in all known elimination pro
edures.
This uniformity property inhibits the transformation of these elimina-
tion pro
edures into polynomial time algorithms.
In
on
lusion, when treating with algorithmi
elimination problems,
one should not expe
t too mu
h from the output: as observed already
by Hilbert, a strong limitation to spe
i
elimination problems (with ad-
ditional a priori information or additional semanti
al stru
ture) is ne
-
essary in order to avoid huge diÆ
ulties and one should also renoun
e to
ompute the whole elimination obje
t, if this is not required in advan
e.
In parti
ular,
anoni
al elimination polynomials
ontain the
omplete in-
formation about the elimination problem under
onsideration, they are
"
o{versal", and this makes ne
essarily intri
ate their representation in
34 Giusti and Heintz
any reasonable data{stru
ture (for more details see [6℄). A way out
of this dilemma should be found in analysing whi
h information about
elimination obje
ts is really relevant, avoiding in this way to struggle
with enormous obje
ts whi
h en
ode a lot of spurious knowledge.
A
knowlegment.
The authors wish to express their gratitude to Guillermo Matera and
Rosa Wa
hen
hauzer (Buenos Aires, Argentina) and to David Castro
and Luis Miguel Pardo (Santander, Spain) for many fruitful dis
ussions
and ideas. They are deeply indepted to Ja
ques Morgenstern (Ni
e,
Fran
e), who suggested this resear
h and made together with the authors
the rst and most fundamental steps toward the
omplexity model and
results of this paper.
Bibliography
[1℄ Blum, Cu
ker, Shub, and Smale. Algebrai
settings for the problem
\P = NP ?". In Renegar, Shub, and Smale, editors, The Mathemati
s of
6