LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES Managing Editor: PROFESSOR I.M. James, Mathematical Institute, 24-29 St.Giles, Oxford Already published in this series 1. 4. 5. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26.
27. 28. 29. 30. 31. 32. 33. 34. 35. 36.
General cohomology theory and K-theory, PETER HILTON. Algebraic topology: A student's guide, J.F. ADAMS. Commutative algebra, J.T. KNIGHT. Introduction to combinatory logic, J.R. HINDLEY, B. LERCHER and J.P. SELDIN. Integration and harmonic analysis on compact groups, R.E. EDWARDS. Elliptic functions and elliptic curves, PATRICK DU VAL. Numerical ranges II, F.F. BONSALL and J. DUNCAN. New developments in topology, G. SEGAL (ed.). Symposium on complex analysis Canterbury, 1973, J. CLUNIE and W.K. HAYMAN (eds.). Combinatorics, Proceedings of the British combinatorial conference 1973, T.P. McDONOUGH and V.C. MAVRON (eds.). Analytic theory of abelian varieties, H.P.F. SWINNERTONDYER. An introduction to topoligical groups, P.J. HIGGINS. Topics in finite groups, TERENCE M. GAGEN. Differentiable germs and catastrophes, THEODOR BROCKER and L. LANDER. A geometric approach to homology theory, S. BUONCRISTIANO, C.P. ROURKE and B.J. SANDERSON. Graph theory, coding theory and block designs, P.J. CAMERON and J.H. VAN LINT. Sheaf theory, B.R. TENNISON. Automatic continuity of linear operators, ALLAN M. SINCLAIR. Presentations of groups, D.L. JOHNSON. Parallelisms of complete designs, PETER J. CAMERON. The topology of Stiefel manifolds, I.M. JAMES. Lie groups and compact groups, J.F. PRICE. Transformation groups: Proceedings of the conference in the University of Newcastle upon Tyne, August 1976, CZES KOSNIOWSKI. Skew field constructions, P.M. COHN. Brownian motion, Hardy spaces and bounded mean oscillation, K.E. PETERSEN. Pontryagin duality and the structure of locally compact abelian groups, SIDNEY A. MORRIS. Interaction models, N.L. BIGGS. Continuous crossed products and type III von Neumann algebras, A. VAN DAELE. Uniform algebras and Jensen measures, T.W. GAMELIN. Permutation groups and combinatorial structures. N.L. BIGGS and A.T. WHITE. Representation theory of Lie groups, M.F. ATIYAH. Trace ideals and their applications, BARRY SIMON. Homological group theory, edited by C.T.C. WALL.
Partially Ordered Rings and Semi-Algebraic Geometry
Gregory W. Brumfiel
CAMBRIDGE UNIVERSITY PRESS CAMBRIDGE LONDON NEW YORK NEW ROCHELLE MELBOURNE SYDNEY
CAMBRIDGE UNIVERSITY PRESS Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo Cambridge University Press The Edinburgh Building, Cambridge CB2 8RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521228459 © Cambridge University Press 1979 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 1979 Re-issued in this digitally printed version 2007 A catalogue record for this publication is available from the British Library ISBN 978-0-521-22845-9 paperback
Contents
Page INTRODUCTION
1
CHAPTER I - PARTIALLY ORDERED RINGS 1.1. Definitions
32
1.2. Existence of Orders
33
1.3. Extension and Contraction of Orders
34
1.4. Simple refinements of orders
36
1.5. Remarks on the Categories (PORNN) and (PORCK)
37
1.6. Remarks on Integral Domains
39
1.7. Some Examples
40
CHAPTER II - HOMOMORPIIISMS AND CONVEX IDEALS 2.1. Convex Ideals and Quotient Rings
45
2.2. Convex Hulls
46
2.3. Maximal Convex Ideals and Prime Convex Ideals
49
2.4. Relation between Convex Ideals in
(A,'p) and
(A/I,'IV I) ... 52
2.5. Absolutely Convex Ideals
52
2.6. Semi-Noetherian Rings
,
56
2.7. Convex Ideals and Intersections of Orders
62
2.8. Some Examples
66
CHAPTER m
- LOCALIZATION
3.1. Partial Orders on Localized Rings
77
3.2. Sufficiency of Positive Multiplicative Sets
79
3.3. Refinements of an Order Induced by Certain Localizations .. 80 3.4. Convex Ideals in
(A/J3)
and
(A T ,^ T )
81
3.5. Concave Multiplicative Sets 3.6. The Shadow of 1
83 ;
84
3.7. Localization at a Prime Convex Ideal
87
3.8. Localization in (PORCK)
88
3.9. Applications of Localization, I - Some Properties of Convex Prime Ideals
89
3.10. Applications of Localization, H-Zero Divisors
91
3.11. Applications of Localization, m - Minimal Primes, Isolated Sets of Primes, and Associated Invariants 3.12. Operators on the Set of Orders on a Ring
93 96
CHAPTER IV - SOME CATEGORICAL NOTIONS 4.1. Fibre Products
101
4.2. Fibre Sums
102
4.3. Direct and Inverse Limits
103
4.4. Some Examples
104
CHAPTER V - THE PRIME CONVEX IDEAL SPECTRUM 5.1. The Zariski Topology Defined
106
5.2. Some Topological Properties 5.3. Irreducible Closed Sets in 5.4. Spec(A,'p)
107 Spec(A,'£)
107
as a Functor
5.5. Disconnectedness of
109
Spec(A,'}3)
109
5.6. The Structure Sheaf, I - A First Approximation on Basic Open Sets
112
5.7. The Structure Sheaf, II - The Sheaf Axioms for Basic Open Sets
113
5.8. The Structure Sheaf, HI - Definition
115
CHAPTER VI - POLYNOMIALS 6.1. Polynomials as Functions
118
6.2. Adjoining Roots
120
6.3. A Universal Bound on the Roots of Polynomials
123
6.4. A "Going-Up" Theorem for Semi-Integral Extensions
125
CHAPTER VII - ORDERED FIELDS 7.1. Basic Results
130
7.2. Function Theoretic Properties of Polynomials
132
7.3. Sturm's Theorem
135
7.4. Dedekind Cuts; Archimedean and Non-Archimedean Extensions. 137 7.5. Orders on Simple Field Extensions
140
7.6. Total Orders and Signed Places
144
7.7. Existence of Signed Places . . ..
148
CHAPTER VIII - AFFINE SEMT-ALGEBRAIC SETS 8.1. Introduction and Notation
i62
8.2. Some Properties of RHJ-Algebras
I68
8.3. Real Curves
,
178
8.4. Signed Places on Function Fields
184
8.5. Characterization of Non-Negative Functions
193
8.6. Derived Orders
196
8.7. A Preliminary Inverse Function Theorem
206
8.8. Algebraic Simple Points, Dimension, Codimension and Rank . 212 8.9. Stratification of Semi-Algebraic Sets
218
8.10. Krull Dimension
224
8.11. Orders on Function Fields
232
8.12. Discussion of Total Orders on
R(x,y)
240
5.13. Brief Discussion of Structure Sheaves
247
I - The rational structure sheaf
248
II - The semi-algebraic structure sheaf
252
IH - The smooth structure sheaf
262
APPENDIX.
268
The Tarski-Seidenberg Theorem
BIBLIOGRAPHY
273
LIST OF NOTATION
278
INDEX
279
Preface
This text represents an attempt to formulate foundations and rudimentary results of a type of geometry and topology in purely algebraic terms.
I feel
that the approach taken here is very natural and that it is only coincidental that the point of view I advocate did not emerge fifty years ago. The mathematics itself is most similar to elementary commutative algebra and algebraic geometry.
The level of difficulty is about like that of the
texts on commutative algebra by Zariski - Samuel or Atiyah-Macdonald.
Although,
strictly speaking, the text might be read without any previous knowledge of basic commutative algebra, essential motivation would probably be lacking. On the other hand, I see no reason why a student couldn't simultaneously read this text and some classical commutative algebra. In the final two chapters, I assume the reader is familiar with, or can read elsewhere, basic results of Artin-Schreier theory, Krull valuation theory, and algebraic geometry.
The basic algebra texts listed in the bibliography
as references [63] - [68] contain more than adequate background material in the appropriate sections.
The final two chapters of this text are, in fact,
somewhat independent of the first six chapters. I recommend that after looking at the introduction, the reader look through Chapters VII and VIII in order to gain motivation for the foundational material of Chapters I through VI.
Introduction
It is my hope that the methods developed in this text will lead to an interesting embedding of algebraic topology in a purely algebraic category, namely, some category of partially ordered rings. At the same time, the theory provides a convenient abstract setting for the theory of real semi-algebraic sets, quite analogous to commutative algebra as a setting for modern algebraic geometry. I might motivate the study of partially ordered rings (somewhat frivolously) as follows.
One observes that the integers, together with
their ordering, is an initial object for a lot of mathematics.
On the
one hand, consideration of order properties leads to the topology of the real line, then to Euclidean spaces, and eventually to abstract continuity and general point set topology.
On the other hand, consideration of
arithmetic properties leads to the abstract theory of rings, fields, ideals, and modules. Following either route, one can go too far.
Completely general
topological spaces and continuous maps are uninteresting. general rings and modules are uninteresting.
Completely
Thus the mainstream in
topology concentrates on nice spaces (for example, polyhedra and manifolds) and the mainstream in algebra concentrates on nice rings (for example, finitely generated rings over fields, subrings of the complex numbers and their homomorphic images.)
The two theories seem to intersect even-
tually in category theory and semi-simplicial homotopy theory. The topologists put back in some algebra and the algebraists put back in some topology.
On the other hand, algebraic geometry works best over
an algebraically closed field and such concepts as manifolds with boundary, homotopy of maps and mapping cones, which are extremely useful to topologists,
1
are not readily available in pure algebra.
A simple observation is that
such concepts are easily described by algebraic equalities and inequalities in real affine space, however.
Certainly all of geometry is deeply rooted
in the study of equalities and inequalities of functions on real affine space.
Even in point set topology real functions and inequalities play a key
role, for example, in the theory of paracompact spaces. The real number field is a formally real, real closed field. In fact, the Artin-Schreier theory of formally real fields is precisely an abstract algebraic treatment of inequalities in field theory.
The real closed fields
are the analogues of algebraically closed fields-they admit no proper algebraic extensions in which inequalities still make sense. It thus seems to me that a true understanding of the relations between algebraic geometry and topology must stem from a deeper understanding of real algebraic geometry, or, actually, semi-algebraic geometry.
Moreover,
real algebraic geometry should not be studied by attempting to extend classical algebraic geometry to non-algebraically closed ground fields, nor by regarding the real field as a field with an added structure of a topology.
Instead, the abstract algebraic treatment of inequalities
originated by Artin and Schreier should be extended from fields to (partially ordered) algebras, with real closed fields replacing the algebraically closed fields as ground fields.
It is obvious that such a
category of partially ordered algebras provides an abstract setting for semi-algebraic geometry (study of sets defined by finitely many real polynomial equalities and inequalities), and it seems plausible that such a category would allow a natural development of algebraic topology and homotopy theory. It is essential that the reader understand that algebraic topology (at least the
homotopy category of finite simplicial complexes and the
study of reasonable functors
on this category) is known to be completely
independent of topology, that is, independent of limits, continuity, the infinite arithmetic of open and closed sets; even the completeness of the real numbers is irrelevant.
The most highly developed reduction of homo-
topy theory to pure algebra is the semi-simplicial, or combinatorial,
approach, developed by D. M. Kan, J. C. Moore, M. M. Postnikov, and others in the 1950's.
The problem with this reduction (ignoring inefficiency) is
that it seems unmotivated without first developing the point set topology of finite simplicial complexes, which, in turn, is founded on the topology of the real line.
Also, differential topology seems unnatural in this setting.
My philosophy is that the derivation of homotopy theory from point set topology is anhistorical accident.
Basically, I regard the real
goal to be a mathematization of our experience^sensation, and perception of space, time, and matter.
This experience is inherently finite, but
involves counting, hence algebra, and order relations, hence inequalities. OUT immediate perception of boundaries of objects, and spatial and temporal order relations justifies a more structured approach than the mathematical reduction of all experience to simply counting small or big finite sets. In fact, it seems to me that a reasonable first approximation of our perception is the set theory of sets in affine space defined by finite collections of algebraic equalities and inequalities, along with the Boolean set theoretic operations of finite unions and intersections, and differences.
This sort of set theory has much in common with topology,
but is fundamentally very different.
Thus we will often use the language
of open sets, closed sets and so on, but it is to be understood that a set in the plane like
y > x
is not open because it is a union (necessarily
infinite) of open balls but rather because it is the set of points where the 2 single algebraic function
y-x
is positive.
In general, functions can be
replaced by their graphs, hence admissible functions from one semi-algebraic set to another can be thought of as certain types of semi-algebraic subsets of the product.
Thus morphisms also avoid the infinite definitions of point
set topology. According to the Hilbert Basis Theorem, any set in affine space over a field defined by algebraic
equalities
by a finite subset of the equations
f
of affine space defined by inequalities f (x,...xn) = 0
f (x^.-.x ) = 0 = 0.
is already defined
Over an ordered field, subsets x
gg( i'-- x n ) :L °
as wel1
as
equalities
are clearly not always representable by finitely many
equalities and inequalities.
It is perhaps this fact which has led to the
divergence of the fields of algebra and topology. Topology is a good example of a subject which produces answers of interest before the real problems are fully clarified.
As several examples
of such answers, which come up in more than one context in mathematics, I would list Lie theory, the theory of compact surfaces, the Bott periodicity theorem and K-theory, the classification of differentiable structures on spheres, the theory of cohomology operations (even cohomology theory itself), the computations of the classical group bordism rings, and the emerging classification of singularities of maps.
These subjects deal with topological
concepts, but also turn out to be related to problems in algebra and number theory.
Thus one feels certain that it is good stuff.
On the other hand, much of geometric topology is concerned with analyzing just what pathology can and cannot occur, using infinite definitions and constructions.
Thus one has space filling curves, but a very strong
regularity theorem about simple closed curves in the plane.
I regard this
as evidence that arbitrary continuous curves tend to be uninteresting, but it is not evidence that simple closed curves are interesting.
In fact, in
three space one has wild embedded arcs and spheres and their classification is not regarded as a mainstream problem.
Under the assumption of topological
local flatness, it is known that every n-sphere in topological ball. smooth n-sphere in n / 3
n+1
space bounds a
More important, however, is the question of whether a space bounds a smooth ball.
n+1
Provocatively, if
the answer is known to be yes, but the proof is harder than the
corresponding topological theorem. smooth copy of
S
in
IR
If
n = 3, it is possible that some
bounds a topological ball which is not diffeo-
morphic or piecewise linearly equivalent to the standard
4 D .
Pathology can involve morphisms, as do these examples, or absolute properties of spaces.
Thus topological manifolds of dimension one and two
are classified, manifolds of dimension three are known to possess unique piecewise linear and differentiable structures, but are not yet classified, while it is still unknown if manifolds of dimension four and higher can always be triangulated.
There is perhaps a widespread feeling that if attention is restricted to, say, differentiable manifolds, pathology disappears.
It is known
(Whitehead) that smooth manifolds admit a unique compatible piecewise linear structure and it is also known (Nash, Tognoli) that compact smooth manifolds are diffeomorphic to non-singular real algebraic varieties.
Also
(Milnor, Serre and many other contributors) any compact manifold admits only finitely many distinct differentiable structures. category allows too many morphisms.
But the differentiable
Any closed set in Euclidean space
can be realized as the zeros of a smooth
(C°°)
function.
The studies of
singularities, diffeomorphisms, flows, and foliations in recent years have produced many pathological phenomena, as well as regularity theorems under suitable hypotheses.
Another area in which there are strong regularity
theorems for the objects is the study of smooth compact Lie group actions on compact manifolds.
The compact Lie groups are algebraic groups and
Palais has extended the Nash-Tognoli theorem to an equivariant version which roughly says all compact Lie group actions on compact manifolds are algebraic, up to isomorphism. My own interest is exactly the reverse of this tradition of seeking regularity theorems in topological situations.
Instead, I advocate beginning
with algebra and working toward geometry, in an attempt to discover just what geometric phenomena are realizable by finite algebraic constructions. It is not so much a question of one type of mathematics being superior to another, but simply a question of how best to understand the dividing line between algebra and topology, and I think this dividing line should be approached from both directions.
The notion of inequalities is very close
to this dividing line, but essentially on the algebraic side.
From the
algebraic point of view it is more or less clear that the real algebraic numbers are just as useful as, or even preferable to, the real numbers. I will return to this philosophy from time to time in the course of this introduction. In any event, in this book, I first develop systematically an abstract theory of partially ordered rings.
The models I have in mind are rings of
real valued algebraic functions on certain semi-algebraic sets in affine
space and their quotients by various allowable ideals.
My axioms thus
reflect properties of these models, and aside from a certain amount of curiosity, the abstract theory interests me only insofar as it contributes to the eventual goals of better understanding semi-algebraic geometry and algebraic topology. The second half of the book is devoted to a systematic introduction to real semi-algebraic geometry via Artin-Schreier Theory and the language of
partially ordered rings.
For the most part, the actual results can be
found in the existing literature.
In particular, the papers by Dubois,
Efroymson, Lang, and Stengle, referred to in the bibliography are very similar in spirit to (and, in fact, influenced greatly) my philosophy. Also, there are excellent introductory accounts of Artin-Schreier theory in the algebra texts of Lang, Jacobson, and van der Waerden. What, then, is a partially ordered ring?
Generally, the definition
found in the literature is a ring, together with a subset of elements called positive such that sums and products of positive elements are positive and such that if
x
and
-x
most efficient terminology is to call positives as strictly positive.) squares are positive. unit.
are positive, then 0
x = 0.
(The
positive and to refer to non-zero
I make the further assumption that all
Also, of course, all rings are commutative with
Given such a set of positives, a partial order relation is defined
on the ring by
x >_ y
if
x- y
is positive.
The definition is purely
algebraic. The assumption that squares are positive is justified, first, because it is true in all the examples I want to fall within the scope of the theory and, secondly, because it seems to be a very useful assumption for proving analogues of the basic results in commutative algebra. It is easy to see that a ring admits such a partial order if and only if the following condition holds:
1 < i < n.
whenever
n 2 Z a. = 0, then each X i=l
2 a. = 0 , X
The set of all finite sums of squares is then an allowable set
of positives, in fact, clearly the smallest such.
Nilpotent elements are
decidedly permitted by this condition. The morphisms between partially ordered rings which are important are
the order preserving ring homomorphisms.
Kernels of such morphisms are
called convex ideals and are characterized by the property that if a sum of positive elements belongs to the ideal, then so does each summand.
This
gives a category, (POR). The category (POR) turns out to be not the best approximation to the ultimate goals.
A useful subcategory is the category (PORNN), partially
ordered rings with no nilpotent elements.
However, nilpotent elements
are actually useful, just as in modern algebraic geometry.
A compromise
is the intermediate category (PORCK), partially ordered rings with convex n killers. The added axiom is the following condition: whenever ( 2 P - ) x = °» i=l 1 p.
positive, then each
p.x = 0 ,
1 <_ i <.n«
I will indicate advantages of (PORCK). nx = 0, but (PORCK).
x i- 0.
Since
Throughout this introduction,
To begin, in (POR) one can have
n = 1+ ••• + 1, such pathology is avoided in
Secondly, in (PORCK) the associated primes of a convex ideal are
automatically convex, as are the isolated primary components.
Thirdly, a
ring of polynomials in finitely many indeterminates over a ring
A
in
(PORCK) can be regarded faithfully as a ring of A-valued functions on affine space over
A.
A more subtle advantage is that the defining condition
for (PORCK) makes sense if
x
takes values in a module over
A.
Thus, in
(PORCK), a ring is an admissible module over itself, in a certain useful sense. In any category of partially ordered rings, only ideals which are kernels of morphisms in the category should be considered at all.
Thus,
in (POR) one sees the convex ideals, in (PORNN), one sees the radical convex ideals convex ideals.
I = /F, and in (PORCK) one sees what I call absolutely Namely, I
positive, implies
p^x
is absolutely convex if
€= I,
( 2 P i )x E l , i=l
p^
1 < i < n.
Much of this first volume is concerned with the partially ordered analogues of basic results on ideals in commutative algebra.
Each such
result must be checked, but, as a rule, slight extensions of classical arguments work for convex ideals in partially ordered rings. we mention:
As examples,
1. Every proper convex ideal is contained in a maximal convex ideal. (Note maximal convex
±
convex maximal.)
2. Maximal convex ideals are prime. 3. The intersection of all prime convex ideals is the (convex) ideal of all nilpotent elements. 4. The intersection of all prime convex ideals containing a given convex ideal is the nil radical of the ideal. These results hold uniformly in the categories (POR) and (PORCK) since a convex prime ideal (2 p )
X
Q
is necessarily absolutely convex.
(2 p i )x 2 = 2 p ^
e Q, then
2
e Q, hence
(Proof:
p i x 2 e Q, hence
if
p i x € Q.)
5. In all the categories, residue partially ordered rings are defined as the cosets of a convex or absolutely convex ideal. cosets are those containing a positive element.
The positive
The usual fundamental
isomorphism lemmas and correspondences between ideals in residue constructions are established. 6. Localizations are defined in all the categories and the desired basic properties established.
Of importance here is the concept
of a concave multiplicative set, which, when it contains a positive element also contains all larger elements.
This is natural, since
if one wants to invert a strictly positive function on some set, then every larger function also has no zeros, hence might as well also be inverted.
Complements of prime convex ideals are concave
multiplicative sets, as is all elements
S(l), the shadow of
1, consisting of
s >_ 1.
7. Maximal convex ideals are characterized by the property that the associated
residue ring is a semi-field.
that for each non-zero elements
a
a, one has
ab _> 1
By semi-field I mean for some
b.
Such
(semi-units) are the analogues of units since they
belong to no proper convex ideal. 8. Certain categorical constructions such as fibre sums, fibre products, direct and inverse limits are carried out in the various categories. 9. The set
X
of prime convex ideals is given the Zariski "topology",
and basic functorial and "topological" properties established.
A
structure sheaf of partially ordered rings is constructed by means of localizations.
The stalks of the structure sheaf are the partially
ordered rings one obtains by localizing with respect to complements of prime convex ideals.
The ring of global sections is generally
larger than the original ring. of
1,
In fact, in (PORCK), the shadow
S(l) = {s _> 1}, consists of non-zero-divisors
and the global
sections of the structure sheaf is the ring one obtains by localizing with respect to
S(l). This is reasonable, since this localization
exactly inverts all elements which belong to no prime convex ideals, that is, "functions nowhere zero on
X".
(In (POR), S(l)
can have
zero divisors, a ring does not even inject into its global sections, in general, and the global sections cannot be described by a simple localization.) 10. A "universal bound" is obtained for roots of a monic polynomial with coefficients in a partially ordered ring. obtained for solutions of of even degree.
More generally, bounds are
f (x) <_ 0, where
f
is a monic polynomial
As a corollary, one obtains a going-up theorem for
prime convex ideals in what I call semi-integral extensions Namely, an element
x£ B
is semi-integral over
A
if
A C B.
f(x) <^ 0
for some monic polynomial of even degree, with coefficients in
A.
Unfortunately, the going-up theorem requires a mild hypothesis on the partial orders on
A
and the extension
B.
So the theorem does
not have the same applicability to, say, the real Nullstellensatz that the classical going-up theorem has to the Nullstellensatz over algebraically closed fields. 11. A theory of associated primes and isolated primary components is worked out, which is quite satisfactory for rings in (PORCK) satisfying the ascending chain condition for convex ideals.
However,
there is no general decomposition of convex ideals as intersections of primary convex ideals, even for Noetherian partially ordered rings. Here, as in the going-up theorem above, the gap between the abstract category (PORCK) and the specific study of finitely generated rings
over fields begins to widen. 12. Each subset
X
of a partially ordered ring
convex ideal
H(X)
ideal
(the absolute hull of
then
AH(X)
(the hull of
AH(i n )AH(l m ) C A H ( i n + m ) .
does not seem to hold.
A
belongs to a smallest
X) and a smallest absolutely convex X).
If
I C A
is any ideal,
The corresponding property for hulls
Thus in (PORCK) there is a natural graded
ring associated to the ideal
Relations between
A
and
I, namely
G(A)
© AH(I n )/AH(I n+1 ) = G(A). n^ 0
in (PORCK) parallel properties of
ordinary associated graded rings in commutative algebra. There are two fundamental and unavoidable reasons why partially ordered algebra is "harder" than commutative algebra and why real algebraic geometry is "harder" than algebraic geometry over algebraically closed fields. first is that principal ideals are not convex in general.
The
The smallest
convex ideal containing a given element or set of elements is rather complicated.
The second is that the statement that a given polynomial over an
ordered field has roots in some ordered extension field, is a non-trivial statement, requiring specific verification. the
entire theory.
These two difficulties permeate
The second especially is perhaps the main reason why
the abstract theory of partially ordered rings is not as useful in real algebraic geometry as abstract commutative algebra is in classical algebraic geometry. Here is a simple example which illustrates certain features of the theory.
If
R[X,Y,Z]
is partially ordered as a ring of real valued functions 2 2 on affine space, then the ideal (X + Y + 1 ) has no zeros, but is not convex. (Any convex ideal containing
more subtle is the ideal
2 2 X + Y +1 2 2
X +Y
2
is the ideal
2 2 X ,Y ,1.)
Slightly
(X + Y ) , which has zeros but for which the co-
dimension of the zero set is too big. 2
must also contain
The smallest convex ideal containing
(X ,Y ,XY), consisting of all functions which vanish
to second order on the Z-axis. To see this, observe that the functions ( X + Y ) 2 and ( X - Y ) 2 are positive and (X + Y ) 2 + (X - Y ) 2 = 2(X 2 + Y 2 ) . So 2 2 2 2 the smallest convex ideal containing X + Y must also contain X ,Y and 2 o (X + Y) , (X-Y) . Conversely, if a sum of positive functions is in the
10
ideal
(X ,Y ,XY), the summands must vanish to second order on the Z-axis.
One can even see that ring
2
(X ,Y ,XY)
2
IR[X,Y,Z]/(X ,Y ,XY)
is absolutely convex, so the residue
is in (PORCK) and has nilpotent elements.
Roughly, the goal of the abstract affine theory is to interpret partially ordered rings
A
as rings of "functions" on some set (the
prime or maximal-convex ideals of some order" on a subset.
A) modulo functions which "vanish to
That is, some jet is zero on the subset.
An
algebraic set corresponds to zeros of some convex radical ideal
I = /F.
The jets associated to
/J = I.
I
are the (absolutely) convex
One must contend not only with the rings partial orders on these rings. think of
A/I,
A/I,
A/J
with
but with various
In geometric language, it is natural to
I = /f, as a ring of functions on the zeros of
partially order accordingly.
I, and
In the general case, it is natural to relate
orders on
A/J, /f = I, to more refined orders on
regarding
A
the zeros of
J
A
itself, obtained by
as a ring of germs of functions defined on neighborhoods of I.
Neighborhoods in the space of prime or maximal convex
ideals are defined by inequalities.
In the right examples, neighborhoods
are always Zariski dense, which is why
A
can be regarded as a ring of
germs of functions defined "near" the zero set of an ideal.
The use of
inequalities thus enables one to define "strong" open sets and semi-algebraic sets in spaces of prime and maximal convex ideals.
In good cases the structure
sheaf soups up to these strong open sets, again using suitable localizations. This discussion brings me back to the philosophy of rejecting point set topology altogether.
Above I mentioned a Zariski "topology" on the
set of convex prime ideals.
Now, the open and closed sets in the Zariski
topology on the prime ideal spectrum of a Noethevian ring are accessible by finite set arithmetic from sets defined by algebraic equalities and nonequalities.
So, again, I regard viewing the Zariski topology as a classical
topology to be an historical accident.
In the partially ordered context,
one has the added freedom of defining sets by equalities, non-equalities, and inequalities, but one should stick with finite methods and not refer to weak and strong topologies, nor allow infinite processes. Continuing, it is
clear that once an affine theory is established,
11
one could globalize to define partially ordered preschemes.
A modern
viewpoint is that structures are conveniently organized by sheaves on spaces.
But one still begins with "a topological space" and then imposes
a structure sheaf of, say, smooth, analytic, or algebraic functions on open sets.
In the algebraic case this seems unnatural because, although
an algebraic function can be restricted to any open set, the natural domains of definition of algebraic functions are a more restricted class of sets. Also, in topology, analytic geometry and algebraic geometry, one generally comprehends the open sets in the underlying spaces in the first place in terms of inequalities or non-equalities involving functions of the structure sheaf.
That is, one sees these functions before one sees the open sets.
Thus, it would seem preferable to dispense altogether with the original topological space.
In the following paragraphs, I make some (vague) sug-
gestions about alternative foundations.
First, one might go backwards to
Weil's method of constructing abstract varieties. A.
Simply take some rings
(preferably finitely many, all Noetherian) and match up their local
structure sheaves along open neighborhoods. Spec(A.)
These open neighborhoods in
should also be finitely described by Boolean operations, equalities,
non-equalities, and inequalities. depend only on the ring
A.
Note the strong open sets in Spec(A0
and its partial order, not on some preassigned
topological space on which the ring
A.
is imposed.
Naturally, in the global case, one wants an invariant definition, not dependent on a particular finite affine cover.
A reasonable approach might
be to follow Grothendieck and identify a partially ordered prescheme with a set-valued functor on some category of partially ordered rings.
Another
approach which looks mildly interesting is to axiomatize a notion of pseudo-ring, where sums and products of elements are sometimes but not always defined and ordinary ring axioms hold for all defined expressions. The model is provided by pairs and
f
(f,U)
is a suitable function on
where
U
is open in some space
U.
The idea of working directly with "sheaves" in some axiomatic manner, bypassing
an initial topological point set, is also the starting point
of the theory of topoi.
12
For the purposes of algebraic geometry, semi-
algebraic geometry and algebraic topology, I think topos theory should be refined in two ways.
First, rings should be built into the basic definitions,
and secondly a purely finite theory should be developed.
These two modi-
fications are related. One possible approach to a more finite topology is to simply look for substitutes for the axiom that "arbitrary" unions of open sets are open.
Perhaps interesting alternatives do exist.
However, I think
that such an approach would suffer from lack of natural examples and with such an approach the appropriate notion of morphism would be obscure. Categorical products would probably also be awkward.
On the other
hand, in the theory of Grothendieck topoi, a very natural finiteness assumption does exist, namely the Noetherian assumption that every allowable cover
{U
•+ U}
has a finite subcover.
Thus one should
somehow blend set theory and the theory of Noetherian topoi. is where a structure ring or sheaf of rings is useful.
If one has
"rings of functions"
A(X)
define subsets of
in various ways, then one can associate to
X x Y
the ring
X
associated to sets
Here
A(X) <2> A(Y)
(®
X, which are used to
over a suitable ground ring)
and thus define various subsets in
X x Y.
Morphisms and more
general correspondences can be studied as suitable subsets of
X x Y.
Once open sets are defined, the associated Noetherian Grothendieck topos, where covering families mean essentially finite covers, can be studied.
Admittedly, in Chapter V, I consider old-fashioned
topological sheaves.
There, I study the prime convex ideal spectrum
of arbitrary partially ordered rings.
(Old habits die hard.)
In
Chapter VIH, however, I try to emphasize the significance of Noetherian topoi in semi-algebraic geometry. In scheme theory, non-equalities in the structure rings define the open sets.
In topology, the inequalities in the structure ring
define the open sets.
Note that if
R
is a real closed field, the
"interval topology" is generally terrible from the standard viewpoint, as it tends to be totally disconnected, non-separable, non-locally compact,
13
etc.
But if one pursues the above finite approach to topology, all this
pathology is irrelevant. To repeat the main point:
just because certain interesting functions,
say polynomials,or more general algebraic functions, satisfy conditions such as continuity or differentiability, it does not follow that all continuous or differentiable functions are interesting objects of study.
Over
any real closed field, topology makes sense as the classification of semialgebraic sets up to semi-algebraic isomorphism.
At a point of an affine
semi-algebraic set the germs of semi-algebraic functions can neatly be defined as those continuous germs which satisfy an algebraic relation with the coordinate functions.
Alternatively, the semi-algebraic function
germs are those germs with closed, bounded semi-algebraic graphs.
It
makes sense to say a germ is r-times continuously differentiable at a point of affine space,
1 < r < °° .
The
C°°
semi-algebraic germs are called
(germs of) Nash functions and correspond over any real closed field to formal power series solutions of algebraic equations.
Thus various classi-
fication problems of differential topology have analogues over any real closed field. In the second half of the book I restrict attention to finitely generated algebras over a totally ordered field.
The starting point is
the Artin-Schreier theory of ordered fields, which I think properly belongs to a more general theory of partially ordered rings.
Most of the results
can actually be found in recent and not so recent literature. Here is a rough outline.
Let
RfT^-.T^
be the polynomial ring.
Right from the beginning, two sorts of partial orders present themselves. On the arithmetic side, one can take as positive elements only the sums of squares, denoted
$
, or the polynomial combinations with only plus signs
of squares and finitely many additional polynomials $
w
[g-i • • • g ] • l g
A
theorem of Artin assures this latter is admissible as a
set of positives if and only if there is a point 1 £ i <_ k.
gi*««'»g^> denoted
(We assume
R
real closed.)
x £ Rn
The set
a partial order on 14
'P(V)
V
of polynomials nowhere negative on
RfT-^...^].
g i (x) > 0,
On the geometric side, let
be a Zariski dense set (that is, a polynomial vanishing on everywhere).
with
V C Rn
vanishes V
defines
Artin's solution of Hilbert's 1 7 t n problem can be interpreted as an algebraic relation between the orders
'])
and
£(Rn).
Namely, if
f
is
a nowhere negative polynomial, Artin proved that for suitable polynomials h, h^,
h f = Zh..
In other words, as a rational function, f
is a sum
of squares. Given a partial order to be the subset of those with
p
f C B x G B
not a zero divisor.
generally, define the derived order 'J*, which satisfy
px = q, for some
p,q G 'p
(-p ) = '$ (R n ). w d
Thus, Artin's theorem becomes
The motivation for introducing the derived order is the following. The geometric notion of a function being nowhere negative is, to a large extent, a birational notion, depending only on behavior on a dense open set.
Artin's theorem is quite reasonable, in fact, expected, when considered
from this viewpoint. abstractly yield f
Similarly, an inequality such as
f >_ 0.
However,
f
G'£
does give
f
>_ 0
does not
f G 'p , at least if
is not a zero divisor , and certainly if we seek a function theoretic
interpretation of our partially ordered rings, the deduction of from f >_ 0 is desirable. Consider
$ = ^[
g l
. .. g k ] ,
g±
TR]
{x G R |g-(x) > 0, 1 < i < k}, an open set. theorem reads of all
^ d = "^(U).
U, but decidedly g-(y) 21 °-
Of course,
.
Let
f >^ 0
\}{gl . .. gk>
A generalization of Artin's
^ (U) = $ (U), where
U
is the closure
If may not include certain degenerate points
y, where
Birationally, these degenerate points are lost from the
semi-algebraic set
W = {x G Rn|g.(x) >_ 0}.
Graph of
g
The degenerate points are also lost from the point of view of ideals. The real Nullstellensatz says that the maximal convex ideals of
R[T, ... T ] ,
15
relative to the weak order
in the usual manner of evaluating functions at ideal,
/F
x.
If
I
is any
'p -convex
consists of precisely the functions vanishing on the zeros of
Suppose now
'.p
'£ = ']3 [g-, . . . g-, ].
is replaced by
The maximal
n
ideals are the points
W = {x G R |g.(x) > 0, 1 <_ i £ k}.
arithmetic character.
However, the
points of
x €E R n
'$ , correspond precisely to the points
U, where
p-convex
This retains an
'£, maximal convex ideals are just the
U = {x|g.(x) > 0, 1 £ i _< k}
as above.
the appropriate strong Nullstellensatz also holds. algebraic.
The set
In both cases, U
is semi-
This is not trivial, but follows from the Tarski-Seidenberg
theorem, which we discuss below. Let order
Q C A = R[T, ... T ]
satisfy
Q = >/Q, and
Q
convex for some
-p C A.
The general theory in this first volume guarantees that the m finitely many associated primes of Q are also 'p-convex, Q = ,H Q..
Also, if
A/Q
is given its residue partial order
convex ideals will correspond to the maximal contain
Q.
'p/Q, the maximal
'])-convex ideals of
However, if one considers other orders such as
A
which ^d^f
CP/Q)^
C P J / Q ) > the maximal convex ideal spectrum may change--certain degenerate zeros of
Q
for
(Remark that even if
'pj.
will not be convex.
For
'JVQ = 0VQ)ci-)
In fact, Q
might not even be convex
'p = '£,, one does not necessarily have
example, when
maximal convex ideal spectrum of
£ = ' ^ ^ . .. g R ] (A, '$) is
Q
in
However, if some
W.
If all
g. £ Q,
"enough zeros" in
Q
are
(A/Q, CP,/Q) )
of
x
all
U
Q
as
is the set
'£, convex. Q
has
Specifically, U
in some neighborhood of an algebraic
'^-convex, the maximal convex ideals of Q
in
U, but the maximal convex ideals
are only those zeros
x
such that every neighborhood
contains an entire neighborhood of simple points.
g A £ Q, then
C£/Q) d = C£d/Q) •
coincides with the order
16
is
"the zeros of
of
in
Q
is prime, the
Q.
Continuing, if (A/Q, P J / Q )
will be
U = {x G R |g i (x) > 0}.
must contain all the zeros of simple zero of
Q
(A/Q, p/Q)
'p, convex unless
n
If, where
Q
W = {x G R |g-(x) >^ 0}
g. £ Q, then will not be
and n
before, and the maximal convex ideal spectrum of of zeros of
I.
If some
g. G Q, obviously
If 'p/Q
'£ [g.|g. £ Q]/Q» so the above discussion applies
with fewer A/Q
g.. In any case, Cp/Q),
consists of exactly the functions in
nowhere negative on the maximal convex ideals of
is a nice generalization of Artin's theorem.
(A/Q, CP/Q) d )•
This
The pictures below indicate
how the sets above can differ.
W ( g i ) , some
gj E Q
SpecflVQ) = * ) 0
P ^/Q)d = )
S ec
0
SpecCJ>d/Q) = •
Z(Q)
Another natural problem generalizing Artin's theorem is to find an algebraic characterization of functions non-negative on the set of all zeros of Q
in W, that is, the maximal convex ideals of
(A/Q, p / Q ) •
This is solved by a theorem of Stengle, which implies that such
f £ A/Q
are precisely those for which an equation
(f 2 n + p)f = q
(*)
holds, with
p, q ۥ$.
(mod Q)
In the language of our structure sheaf of partially
ordered rings on Spec(A/Q, -p/Q), this condition translates to the statement that
f
is "positive" in the partially ordered ring of "sections over the
basic open set D(f)".
Similar results hold for (A/Q, 'P^/Q), characterizing
the functions non-negative on all zeros of Q
in U.
The paragraphs above discuss "irreducible" affine semi-algebraic sets. A general closed, affine semi-algebraic set S
is a union of finitely many
sets, each defined by finitely many polynomial equalities and inequalities. By the affine coordinate ring of S, we mean the ring the polynomial ring by dividing by the ideal on
A(S), obtained from
I(S) of functions which vanish
S, together with the partial order '$(S) consisting of functions nowhere
negative on S. Our general theory allows us to identify convex ideal spectrum of
(A(S), 'P(S)).
S with the maximal
Moreover, two general results allow
us to reduce much of the study of arbitrary
S to the irreducible case.
17
First, the associated primes of any absolutely convex ideal are always convex.
Secondly, if
'£, and
'$„ are two orders on a ring and
'£, 0*.p?, then
prime ideal convex for
Q
is either
(This last result was found by A. Klapper.) two sets
Wj = {x|g i(x) >_ 0, 1 < i < r }
Q
'£, -convex or
is a
'^-convex.
As a corollary, the union of
and
W 2 = {x|h.(x) >_ 0, 1 < j < s}
can be identified with the maximal convex ideal spectrum of the order
Vgi]
nt
ideals
Mhj]'
i£ indeed
I(W,), I(W 2 )
V g i ] ' '^w[hj]
are non-trivial.
are orders
- Otherwise, the
In general, degenerate inequalities
on one variety are handled by passing to non-degenerate inequalities on subvarieties.
Finally, the theorems of Artin and Stengle characterizing
non-negative functions on certain sets can be used to give necessary and sufficient conditions for
f £ A(S)
are purely algebraic formulas for
to belong to
'P(S).
These conditions
f, like (*) above, expressed in terms
of the original finite collections of polynomials which define
S.
Just
as in algebra, where the ideal generated by a set of elements is more important than the specific basis, in our case the invariant notion of a polynomial being non-negative on a semi-algebraic set has more geometric significance than the particular defining equations and inequalities, and yet this invariant geometric notion is algebraically expressible in terms of the defining polynomials. In the end, perhaps the following is the neatest characterization of the affine coordinate rings (A, '.J3).
First, A
(A(S), 'p(S))
among all partially ordered rings
should be a reduced algebra of finite type over
Secondly, there should exist finitely many (0) = n p
and orders
'.jK C A . = A/P^
refinements of the weak order A -^IIA^
w
is the natural inclusion.
ij
are
(A-j/^i)
as
discussed above.
f = A H II1]). ,
The
I
P^
need not be distinct These basic building
The maximal convex ideals will
correspond to the points in the closure of the set of simple zeros Pi
at which all
a formula for
g-.(x) > 0.
(A(S),'^(S))
with
d
nor minimal, although the minimal primes do all occur. blocks
P. C A
which are derived orders of finite
*p. = ('£ [g. .]) , such that •*•
where
'^-convex primes
R.
A set of primes
{P^}
x
of
which leads to such
can be intrinsically described.
We also establish the basic results of dimension theory in our semi-
18
algebraic category.
That is, simple points do exist and the expected relations
between transcendence degree and chains of convex prime ideals hold. real closed field we prove an implicit function theorem. analogous to the classical case of real numbers: algebraic functions defining a germ singular, then
f
f:
if
inverse function theorem.
The statement is
f,,...,f
(R ,0) -*• (R ,0)
has a smooth algebraic inverse near
For any
with
0.
are smooth df(0)
non-
(This is the
Implicit function theorems are routine corollaries.)
Note that if one has an R-valued function germ, it makes sense to ask if it has derivatives.
One should not get carried away and try to study "all
differentiable functions."
There are quite nice relations between formal
algebraic derivations, abstract partially ordered rings, and the usual definitions of derivatives.
e-6
Once the implicit function theorem is available,
it is routine to give Whitney type stratifications of semi-algebraic sets into non-singular manifold-like strata, with any real closed field as ground field. The study of quotients Q
is
'$ absolutely convex
A/Q, where
Q f ft\ is more complicated.
£ = Cp [gi-'-giJ) • W 1 K (J
convexity of all associated primes of of the isolated primary components.
Q
Suppose
General theory yields the
and also the absolute convexity
However, embedded primary components
must be chosen carefully, before one can establish an absolutely convex primary decomposition of
Q.
Very crucial to the argument are (i) the
restriction to finitely generated algebras over fields, (ii) the restriction to absolutely convex ideals (category (PORCK) rather than (POR)), (iii) the specific form of order.
p
as the derived order of a finite extension of the weak
Dropping any of these conditions leads to convex ideals which cannot
be expressed as intersections of convex primary ideals, even for Noetherian ambient rings
B.
A quick example might be in order. isolated prime
(X)
component is also
and embedded prime
(X,Y).
(X ,XY) C R[X,Y], with The isolated primary
(X), and in pure algebra, one has a large choice for the
embedded component, say now, however,
Consider
(X 2 ,Y),
(X2, Y-cX),
(X 2 ,Y 2 ,XY), etc.
that among the order relations are
Any convex ideal containing
Y
primary component belonging to
must contain
X,
0 £ X,
0<^Y,
Suppose X<^Y.
So one must choose as
(X,Y), one which contains X , but no power
19
less than
Yn+1.
of
Y
In fact,
if
$ = (^w[X,Y,Yn-X])d.
(X2,XY) = (X) n (X 2 ,Y n+1 ,XY)
will work
If one included infinitely many generators for the inequalities, say 0 £ X,Y
and
0 <_ Y n - X, all
n >_ 1, then
X
becomes "infinitesimal".
There would clearly be no primary decomposition of it is absolutely convex. a power of
Y, hence
(X ,XY),
Because, any ideal with radical
X, if it were to be convex.
even though (X,Y) contains
This counterexample
shows where the usual proof of primary decomposition breaks down.
Basically,
principal ideals, which are used in the classical proofs, are not generally convex, and putting one element in a convex ideal forces
a lot of other
elements to be included. Order properties of
A/Q, Q i /Q, are also complicated.
As hinted at
earlier, the results which seem most natural in this direction relate various partial orders on A
A/Q
to orders on
A = RfX^—X^
obtained by regarding
as a ring of germs of functions on semi-algebraic neighborhoods of the
zeros of
Q.
This infinitesimal study of functions, along with primary
decomposition discussed above, is intimately connected with the relations between partial orders on rings and differential operators.
I expect
partially ordered rings with nilpotent elements to be useful in topological intersection theory, providing an alternative approach to transversality. A nice application of partially ordered rings with nilpotent elements is the following characterization of geometric simple points.
First note
that in the semi-algebraic category, points may be simple which do not
look simple t o a nalgebraist.
I n Figure (a), t h eequation
y 3 + 2x 2y - x 4 = 0
3 2 4 y + 2 x y - x = 0
y2-x2-x3
xy > 0
Figure (a)
20
Figure (b)
has a formal power series solution (convergent, of course, over the real numbers) at the origin.
By all rights, the origin should be a simple point.
In Figure (b) the double point is made nice by restriction to the semialgebraic domain
xy >^ 0.
Again, this branch in the first and third
quadrants has a formal power series description.
In general, suppose
is a maximal convex ideal of some affine coordinate ring assume is an integral domain. the order
'$ at
neighborhood of 'p
m, by taking m
f £ '.p
if
f
(B,'.j>) which we
We can also localize
is non-negative on some
in the maximal convex ideal spectrum of
will be an order if
points.
We can localize, B .
m
m
(B,p).
Here,
is in the closure of suitable algebraic simple
Our criterion is, then, that
m
is a geometric simple point if
the associated graded ring
G(B ) = m
AH(mnBj © n>0 AH(m n + 1 B m )
is a polynomial ring and the positive elements in the induced order are the polynomials which are non-negative in some neighborhood of the origin in the appropriate affine space. respect to the order
'$ C B .
in the order
x
'|3 , so
and
In Figure (b), 0 <_ y - x £ 2x
y
Here, the absolute hulls are taken with Thus, for example, in Figure (a), 0 _< y £ 2x are not linearly independent in
mB /AH(m B ) .
in the order '£m.
Non-singular boundary points of a semi-algebraic set can be defined by modifying the order requirement on the associated graded ring to a statement about germs at the origin in an affine half space.
In general, in partially
ordered algebra we have these two types of localization available, first, by inverting elements of a ring in the usual way, and, secondly, by leaving the ring alone and refining the order. In commutative algebra, it is fashionable to study all rings and claim rings with nilpotent elements arise naturally in geometry, say in specializations, intersections, and fibres of a morphism, as fibre products.
But the formation
of a fibre product requires agreement first on a suitable category.
If, say,
fibre products are to be taken in the category of all rings rather than in the category of rings without nilpotent elements, one must justify this by
21
finding convincing illustrations of its geometric usefulness.
In real
algebraic geometry, there are already restrictions on the allowable rings without nilpotent elements, namely , vanish.
non-trivial sums of squares cannot
I have discussed two methods of introducing nilpotent elements,
the categories (POR) and (PORCK).
But, for the most part, the arguments
given that (PORCK) is better have been based on purely algebraic properties. The above application of the graded ring construction (which exists in (PORCK)) to characterize simple points, is meant as geometric propaganda for (PORCK).
But one would like many such applications, say to intersections
and fibres of a morphism.
It is clearly inadequate to simply say "take
fibre products" since fibre products exist in both (POR) and (PORCK). What is required is close scrutiny of the geometry involved in these problems in order to decide which category is giving the "desired" answers.
It is,
of course, possible that even (PORCK) is not the "right" category for all purposes. Given a ring which admits a partial order, the set of all partial orders is an interesting structure.
The partial orders themselves are
partially ordered by refinement, arbitrary intersections of partial orders are partial orders, and Zorn's lemma implies any partial order admits maximal refinements. orders.
In an integral domain, all maximal orders are total
Also, in an integral domain, and in certain other situations, the
intersection of all maximal refinements of an order derived order
.p
In general, the orders
'£
is exactly the
']* which satisfy
'£ = '$^
correspond exactly to partial orders on the larger ring obtained by inverting all non-zero divisors. ring
RfT^-.T^
Thus, derived orders on the polynomial
correspond to all partial orders on the field
These derived partial orders on function-theoretically. follows.
R[T1...T ]
Define a family
= (x|each
of open subsets of
g.
as
g^,-->,g^ let
is locally non-negative at
It is a theorem that these
allow only non-empty
Rn
x}.
(Thus
V
is
1
the interior of the closure of the set
22
can be interpreted
For each finite collection of polynomials
V{g-i > • • • >g^ IK
positive.
V
R(T ;[ ...T n ).
U V
where all
gi
are strictly
are semi-algebraic sets.)
V, that is, g1 ,...,g,
such that
'$ [g, ...giJ
We is
a
partial order on
RfX^.-X ] .
Then there is a natural, bijective, refinement
preserving correspondence between derived orders on &
in the family of sets
Given a filter provided
&
and filters
'V. Total orders correspond to ultrafilters.
the associated order
f | >_ 0, some
R[T.....Tn]
-pf^)
f E'$(<£)
is defined by
V E 3*. The proof is a simple argument based on
Artin's work. In any event, there are tremendously many partial orders, even total orders, on the function field
R(X.....X ) .
quite analogous to valuations, or places.
In fact, these total orders are Krull, in his original work on
valuation theory, pointed out that given a totally ordered field subfield
F, one obtains immediately a valuation ring in
elements finite relative to infinitesimally close to contains
F.
E
E, consisting of
The maximal ideal consists of elements
0, relative to
F, and the residue field
F, is naturally ordered, and is Archimedean over
F.
E
A
Krull
proved a converse, which he attributed to earlier work of R. Baer. given a place on a field
and a
Namely,
with residue field an ordered field, then
one can lift this situation to an ordering of place by the construction above.
E, yielding the original
Lang generalized Krull's results and
established further relations between total orders and real places.
In
particular, he emphasized how the theory of real places can be applied to the study of real algebraic varieties, I go a little further and develop carefully the notion of signed place. A signed place is a place with values in a totally ordered field in which + °° and
- °° are distinguished in an arithmetically coherent fashion.
This is clearly the thing to study in real algebra.
A useful signed place
extension theorem is proved, which has applications to certain versions of the Nullstellensatz and to dimension theory.
Signed places are intimately
related to total orders and subfields. Geometrically, they correspond to studying a variety infinitesimally near a subvariety, and they also provide a convenient language for dealing with behavior at infinity in affine space. A somewhat valid objection to the place theoretic approach is that it has a non-constructive character (Zorn's lemma) at least as complicated as some reasoning about limits and continuity that I am trying to circumvent
23
in the first place.
However, I think this is more illusory than real.
the approach of Tarski
By
and Seidenberg to some of these same problems, the
algebraic foundations of the theory of algebraically closed and real closed fields can be established quite constructively, more or less by the methods of classical elimination theory.
This approach yields rather trivial
proofs of the basic results concerning varieties, once the initial logical inductive arguments are made, but has the disadvantage of obscuring the underlying geometry.
More precisely, the Tarski-Seidenberg theorem asserts
the existence of a decision procedure for answering any elementary (in the logical sense) question about affine semi-algebraic sets, and semi-algebraic functions.
Thus, if one can test a statement in some larger real closed
field, or even over the classical real numbers, the answer will be the same in the original real closed field as in this new field.
In this sense
transcendental methods will play the same role in semi-algebraic geometry over any real closed field as they have played in classical algebraic geometry over algebraically closed fields of characteristic zero.
In some
cases the shortest proof of a theorem may be transcendental, but algebraists should still be interested in conceptually nice, direct algebraic proofs. I would like to establish certain basic properties of semi-algebraic sets, including the result that an algebraic image of a semi-algebraic set is semi-algebraic by another method, namely by exploiting properties of integral or semi-integral extensions. over
A
if
f(x) £ 0
coefficients in
A.)
(Recall
x G B
is semi-integral
for some monic, even degree polynomial
f
with
I have not been very successful in this, but I
still believe it might be possible.
Semi-integral
elements are related
to signed places or total orders pretty much the same way that integral elements are related to places. and to
x G B A
semi-integral over
in any total order on
Namely, given A, then
B
x
A C B
is necessarily finite relative
extending the partial orders.
are affine coordinate rings of semi-algebraic sets (POR) morphism inducing
Y -• X, then
B
pull-backs of closed, bounded subsets of The rings
24
A,B
with partial orders,
X,Y
and
semi-integral over X
If
A -* B A
A,B is a
means the
are closed and bounded in
Y.
need not have the same transcendence degree over the ground field.
As a final remark on the birational interpretation of Artin's solution of Hilbert's 1 7 t n problem, consider a different category, that of smooth manifolds and smooth real valued functions.
Then Paul Cohen has shown me
that (i) there exist nowhere negative smooth functions on any manifold which are not finite sums of squares of smooth functions (in fact, the zero set can be a single point) and (ii) given any nowhere negative smooth h, g, both smooth and
h
not a zero divisor, such that
f, there are
h f = g .
The
zero divisors are, of course, the smooth functions which vanish on some open set.
If
M
is the manifold, '$ (M) C C°°(M)
the nowhere negative
functions, and
'£ C C°°(M)
CP ) = 'P(M) • w d
Note that because of the existence of smooth bump functions,
any partial order on
the sums of squares, this result again reads
C°°(M)
lies between
$
and
'£ (M).
Following the development of the affine theory of semi-algebraic sets, it would be interesting to develop the basic results about projective coordinates in the real domain.
Actually, there are some important additional
possibilities for homogeneous methods.
It is true enough that the zeros
of a real homogeneous ideal is naturally a subset of protective space. However, for forms of odd degree the set of points where the form is nonnegative is only sensible on the sphere, the double cover of protective space.
Thus, from the point of view of semi-algebraic geometry, spherical
homogeneous coordinates are better than projective coordinates.
A third
alternative is the hemisphere, since given any form, its values on the sphere are determined by the values on any hemisphere.
Also, the
usual manner of passing from an affine variety to a projective variety, thus including points at infinity, seems in the real domain to be most naturally expressed by passing from a plane to a hemisphere by radial projection.
Finally, from a topologists point of view, the hemisphere
is the natural domain for working with pairs of spaces. One of my goals is to develop such homotopy theoretical concepts as
K-theory, cohomology, cohomology operations, and techniques
for computing homotopy classes of maps in the framework of partially ordered algebras.
So far my ideas are more speculative than concrete, but
25
some discussion might be in order in this introduction, to encourage further interest. The idea of recovering properties of a space from a ring of functions is quite classical, of course. compact Hausdorff space
X
There are the well-known results that a
is the space of maximal ideals of the ring
C(X) of all continuous real valued functions on
X, and that there is a
natural correspondence between real vector bundles over generated projective
C(X)
modules.
X
and finitely
There are also definitions of cohomology
of commutative rings in the literature which give the Alexander-Spanier or Cech cohomology of C(X).
X
with real coefficients when applied to the ring
Few of these results emphasize order properties of
C(X). This is
perhaps because the non-negative functions are exactly the squares in C(X), so it is easy to disguise order agruments by making use of obvious properties of squares.
For example, a sum of squares in
unless the summands have a common zero. X, Y
C(X)
has no zeros in
Also, any ring map
X,
C(Y) -•C(X),
both compact Hausdorff spaces, is necessarily order preserving, and
gives rise to
X ~*Y.
For smaller rings, such as finitely generated rings
of functions, the order properties are not as close to the algebra. The definition of real K-theory seems to provide no problem.
First,
it is pretty clear that K-theory is more naturally associated to sheaves of rings, rather than to single rings.
So the natural thing to consider
is the Grothendieck group associated to sheaves of finitely generated projective modules over the structure sheaf of a semi-algebraic variety
X.
This works and the proof is already in the literature in slightly different form.
Namely, all one really has to do is localize affine coordinate
rings by inverting all polynomials with no zeros on the semi-algebraic set
X.
This gives enough continuous functions on
idempotent operators on free
C(X)
X
to approximate the
modules by idempotent operators on
free modules over this much smaller localized affine coordinate ring. Inverting functions with no zeros on
X
also has the effect of reducing
the ordinary maximal ideal spectrum exactly to the points of
X, and this is
perhaps regarded generally as the algebraic rationale for this localization. I regard the ordinary maximal ideals as irrelevant.
26
Besides, the localization
does not eliminate all unwanted prime ideals.
In the partially ordered
context, the maximal convex ideal spectrum and prime convex ideal spectrum are both regulated by the partial order on the ring, but are not further affected by the above localization.
The localization does occur naturally
in the construction of structure sheaves, however, and K-theory should involve the structure sheaf. Similarly, there are examples of affine coordinate rings for real curves whose ideal class groups change after this localization, the class group after localization being closer to the geometry.
If one defines an
ideal class group in the partially ordered category by considering only absolutely convex ideals and interpreting principal ideals as those ideals which are minimal absolutely convex ideals containing some single element, then, again, the localization above does not change anything.
In higher
dimensions, one has a similar discussion for divisors, and an algebraic formulation of
H 1 ( ,Z/2).
For curves, of course, H 1 ( ,Z/2)
and K-theory
coincide. Perhaps a harder invariant to recover neatly from a sheaf of partially ordered rings is the cohomology (with general coefficients) of a semi-algebraic variety.
An approach which probably works is to simply write down some
simplicial complex, or semi-simplicial complex, of the correct homotopy type. X
For example, Sullivan and De Ligne have suggested that given
defined by real polynomial equalities and inequalities, there is an
algorithm for paving affine space with, say, rectangles which along with their faces intersect the solution set all.
X
The nerve of the resulting cover of
the correct homotopy type.
in a contractible space, if at X
is a simplicial complex of
Alternatively, one could probably construct
a kind of semi-algebraic singular complex, or even a triangulation, using only constructive algebraic steps. not in the right spirit.
In some sense, these approaches are
On the other hand, they are related to the
general problem of understanding how the "topological" type (by this I mean classification
up to semi-algebraic isomorphism) or homotopy type
of a real semi-algebraic set changes, with changes in the parameters, the coefficients of the polynomial equalities and inequalities which define
27
the set.
Presumably, such changes can occur only along semi-algebraic
locii in the parameter space, defined over
Q, and this seems to me to
be the real reason that algebraic topology can be reduced to combinatorics and number theory. More direct algebraic approaches to cohomology might be based on sheaf theory or on algebraic versions of Alexander-Spanier cohomology or de Rham cohomology. on multiple products of near the diagonal".
The Alexander-Spanier method is based on functions X
with itself, modulo functions which "vanish
This is not so hard to formulate algebraically, using
tensor products, augmentation ideals and completions, or inverse limits. In fact, this approach already exists in algebraic geometry, and no doubt provides an algebraic computation of the cohomology with complex coefficients of a complex affine variety.
I believe the correct interpretation would
also work for real semi-algebraic varieties, with any coefficients. De Rham cohomology, on the other hand, is problematical.
It is true
enough that the complex cohomology of complex affine varieties whose singularities are not too bad, can be computed as the cohomology of the exterior algebra complex on the module of KShler differentials over the affine coordinate ring. ring
For example, the variety defined by
C[X,X~ ]
is
has coordinate
and the cohomology is generated by the constant 0-form
and the closed but not exact 1-form XY - 1 = 0
XY - 1 = 0
R*, not
dx/x.
1,
In the real domain, the variety
answer which is now wrong.
The point seems to be that in the complex domain,
if an algebraic form has any integral at all, it has an algebraic integral, while in the real domain, forms like
dx/x
on
R*
have integrals, but
outside the world of algebra. At first, I hoped to salvage de Rham cohomology for real semi-algebraic sets, perhaps by modifying the notion of module of differential forms in some appropriate way, or by justifying the formal adjunction of solutions of algebraic differential equations to the structure sheaves, whenever such solutions were formally consistent.
This latter viewpoint would be somewhat
similar to the two conceptions of, say, /2 , first as a limit of rational numbers and secondly as a formal solution of
28
X -2 = 0
in some ordered
field extension of the rationals. Some such interpretation of de Rham cohomology might work.
However,
I have come to take the philosophical position that it should not be expected to work and that it works as well as it does in the complex case only by luck or for some reason we do not yet understand.
Certainly, the notion
of algebraic differential equation seems close to the dividing line between algebra and geometry.
But I believe that the mechanism of modeling natural
phenomena on the continuous solutions of differential equations was never meant as more than an approximation of hopelessly complicated combinatorial situations.
The purely mathematical introduction of differential equations
into science reinforces the use of arbitrary analytic or even differentiable functions in geometry.
But the invariants of algebraic topology, such as
cohomology, are global combinatorial invariants, which should retain their significance long after the real numbers, at least most of them, are forgotten. Thus, there seems no reason to believe that de Rham cohomology, with its dependence on the completeness of the real numbers, can be interpreted naturally in a purely algebraic context over a real closed field. Another piece of mathematics close to the boundary between algebra and topology is Morse theory and the theory of singularities in general. Morse theory originated in analysis and is generally regarded as a tool of differential topology, but I believe it has a purely algebraic significance, which shows up most clearly in the well-known applications of Morse theory to the topology of real or complex varieties.
In these applications, one
begins with a variety and some real algebraic function.
Then one describes
information about the critical points, say second order behavior, by algebraic computations.
Then one says, the variety is a smooth manifold,
and smooth manifolds with smooth real functions with certain critical point behavior have such and such local and global geometric properties. seems to me that this appeal to analysis should be unnecessary.
It
The
critical point behavior of real algebraic functions on real semi-algebraic sets should be directly interpretable in abstract semi-algebraic geometry. In fact, this discussion
reaches to the very heart of the distinction
between algebraic and transcendental methods.
In differential topology,
29
one proves varieties equivalent by integrating vector fields, say, provided by Morse functions.
But if the varieties are algebraic, then Nash's
approximation theorem guarantees they are also algebraically equivalent, in a suitable way.
(Their sheaves of Nash functions are isomorphic.)
Thus one ought to investigate whether the transcendental methods are really essential or whether they serve merely as a crutch which has been used simply because the algebra is too hard. Certainly, semi-algebraic and smooth algebraic classification problems are extremely difficult.
A more modest goal would be to find an algebraic
approach to cohomology or homotopy theory, based on an abstract Morse theory. As supporting evidence, I cite the result of Harrison and others that signature type homomorphisms from the Witt ring of quadratic forms over a field to the integers correspond to possible orderings of the field.
Knebusch
has applied quadratic forms to the algebraic study of components of real curves. One of the most important techniques for studying real algebraic varieties is to regard them as the fixed points of an associated complex variety on which conjugation acts as an involution. technique is obviously the equality
C = H[i] . Now, among the fantastic
results of Artin-Schreier theory are that if then
R[i]
The germ of this
R
is any real closed field,
is algebraically closed, any algebraically closed field of
characteristic zero can be written
R[i]
for some real closed field
R
(although non-unique, in general, even up to isomorphism), and the only fields of finite codimension under their algebraic closure are the real closed fields.
Certainly, no analytical or topological result about
complex varieties can be proved without ultimately some appeal to properties of real numbers, precisely because the norm is defined in terms of the real numbers. Nonetheless, many very nice results on real algebraic varieties are established in the literature by making use of the existence of a complexification.
For example, (i) Whitney used the complexification to study
stratifications of real algebraic varieties by dimension, (ii) properties of real algebraic curves are often deduced from classical results about
30
complex curves, (iii) the usual proof that a real algebraic variety carries a fundamental
Z/2-homology class (there are an even number of sheets at the
singular set) makes use of the fundamental fied variety and conjugation.
Z-homology class of the complexi-
Sullivan has generalized these last results.
He shows that a point on a complex variety has a neighborhood which is a cone on a space of zero Euler characteristic and deduces that a point on a real variety has a neighborhood which is a cone on a space of even Euler characteristic. More in the direction of algebraic topology, Atiyah has shown how real K-theory can be efficiently studied via complex vector bundles with conjugation, and Sullivan has suggested studying the homotopy theory of a real algebraic variety by means of the canonical involution on the etale homotopy type of its complexification. I feel that this extensive use of the complexification may be attributable more to a lack of suitable algebraic foundations in the real case, rather than to any natural priority of the algebraically closed case.
Also,
semi-algebraic sets do not complexify, so a real theory is certainly necessary for their study.
In any event, efficient proofs of the results
described above, entirely within the realm of partially ordered algebra, offers a reasonable challenge to the theory. fications,
The study of strati-
and basic properties of one dimensional semi-algebraic sets
are in fact not difficult, using the machinery of this book.
31
I - Partially ordered rings
1.1.
Definitions By "ring" we always mean commutative ring with unit.
morphisms map unit elements to unit elements. this holds only in the trivial ring
We allow
All ring homo1 = 0 , but obviously
(0).
By a partially ordered ring, we mean a ring
A
and a subset
^ c A
which satisfies
(i) T> n (-?) = {0} (ii) ? (iii)
+
K ?
a2 G ^
and ?•?<=? for all
a e A.
By a morphism of partially ordered rings, mean a ring homomorphism
f:
A -»- B
with
f:
(A,p.) ->• ( B $ g ) , we
f ($.) C ^
We thus have a
category of partially ordered rings which we denote If
(A,$) G (POR),
we refer to
$
as an order on
elements of ty positive and elements of f In fact, the subset a £ b
s
]3 C A
if and only if
(POR).
= p - {0}
A -*- B
notation
v
b - a G |5.
strictly positive. A
by
It is easy to redefine partially
is a morphism in (POR) if
b-a G ^
We call
defines a partial order relation on
ordered rings and morphisms in terms of the relation f:
A.
a £ a'
is somewhat preferable to
<_.
implies
For example, f(a) £ f(a').
The
a <_ b, since the former
makes it more clear which partial order relation on
A
is involved.
Nonetheless, the latter is often notationally simpler, and we use it routinely. If we add axioms to the definition of partially ordered rings, we can define subcategories of (POR).
For example, one such axiom simply prohibits
nilpotent elements. (iv)
32
If
a 2 = 0 G A, then
a = 0.
This is a natural axiom for studying rings of real valued functions on a set. We denote the resulting subcategory (PORNN) C (POR),partially ordered rings with no nilpotents. A more subtle axiom is the following. (v)
(p1 + p 2 )x = 0 ,
If
If (A,'p)C (POR), a set C C A p,,p 2 G C.
p i G $,
x G A, then
is convex if P x + P 2
p 1 x = p 2 x = 0.
e
Thus axiom (v) says the annihilator of
C,
p^^ G <$ implies
x
is convex, all
x G A.
We denote the resulting subcategory (PORCK) c (POR), partially ordered rings with convex killers.
Roughly, the motivation for axiom (v) is that it is a
natural axiom for studying real valued functions on a set modulo those functions which vanish to some preassigned order on a subset. Axioms (i) and (iv) imply (v). Namely, if 2
p1x + p2x
2
= 0, hence
2
p^x = p 2 x
and by (iv), p ^ = p 2 x = 0.
2
(p, +p 2 )x = 0, then
= 0, by (i). But then
p 2 x 2 = p2x
Thus, (PORNN) C (PORCK) C (POR).
= 0
Since the
definition of morphism is unchanged, each subcategory is a full subcategory. A third axiom which might be interesting is (vi)
If
(p1 + p 2 ) x G | ,
pte?,
x G A, then
p ^ x G ? .
We refer to this subcategory as (PORPP), partially ordered rings with plenty of positives.
It is easy to see that (i) and (vi) imply (v). Thus
(PORPP) C (PORCK). Now, let we say that $f
A
be a ring. refines ^
If ^,^f
are orders on
(or that ? f
is stronger than $
than !£')• The set of orders on a fixed ring by refinement.
A, with $ C
A
or ^
weaker
is itself partially ordered
An arbitrary intersection of orders is again an order and the
union of a chain of orders (totally ordered subset of the set of orders) is an order.
Thus, any ring
A
which admits an order has a weakest order ^ w ,
and, by Zorn's lemma, any order ^
has maximal refinements.
These remarks
apply equally to the categories (POR), (PORNN), (PORCK) and (PORPP).
1.2.
Existence of Orders Let
A
be a ring.
33
Proposition 1.2.1. (a)
A
admits an order
^
if and only if the following condition
holds: n
2 2 a. = 0 i=l 1 (b)
A
admits an order
$
n 2 2 a. = 0 1 i=l
(c)
A
2 a. = 0, J
implies
with
1 < j < n.
(A,!f) e (PORNN)
implies
if and only if
a. = 0,
1 <_ j <_ n.
3
admits an order
y
n 2 ( 2 a7)x = 0 i=l
with
(A,?) e (PORCK)
if and only if
2 aTx = 0 ,
implies
1 < j < n.
•
3
The proof is a simple exercise.
In all three categories (POR), (PORNN),
(PORCK), the weakest order on a ring
A
is the same,
consisting of just the sums of squares. If
A
is a field, the three conditions (a), (b), (c) in the proposition
are obviously equivalent, and, in fact, are equivalent to the assumption that -1
is not a sum of squares in
A.
This is the usual definition of a formally
real field.
1.3.
Extensions and Contractions of Orders Let
on
A.
f:
A -*- B
We ask if there is an order on
that is, such that
$A C A
be a ring homomorphism and let B,
is a morphism in (POR).
f
be an order f CPA) £ P B>
Clearly, the answer is yes
if and only if the following condition holds:
n
2
^_j
2 f(p.)b* = 0, 1 x -
p. e p 1
The weakest such order on
34
A
B
b. G B, X
is then
implies
2 f(p.)b. = 0 , J
J
1 < j < n.
b. 6 B } ,
which we call the extension of More generally, if and
*p C A
f :
^A
A
-> B
are orders on the
that the family of maps
to
B. is a family of ring homomorphisms
A , there are always ideals
irf : A
•> B -»• B/J
J C B
such
satisfies a generalized extension
condition, which can be formulated as follows: k n ? S f II f (p..))b7Gj, . - v . . a.. v\yJ} 1 i=l j = l IJ
One could take ideals
{J}
p. . G!B , b.GB, r *ij a . .' 1 J IJ
J = B, for example.
of
B
implies F
n ? nf (p..)b7GJ, . . a.. *i;r 1 J 3=1 ij
More to the point, the family of all
with this property is closed under arbitrary inter-
sections, hence there is a smallest such ideal, I c B. ring
Then on the quotient
B = B/I, one can impose an order ty such that the maps
are simultaneously order preserving, Next, suppose order on
B.
fCP A) C*P . A D
f:
A •> B
Trf
a CP a )
C<
a
P>
H
a
^B c B
*p. c A
Obviously, there is such if and only if
$. = ^ , the weakest order on
irf :
A +
B
-
is a ring homomorphism and
One might ask if there is an order
namely, just take
Ki
A
is an
such that admits an order,
A.
This is not too satisfactory, but we next observe that the set. of orders
f. C A
with
fCP A )
i\
C<
PD
admits maximal elements, by Zorn's lemma. D
A.
In general, however, there will not be a maximum such order which we could call
f*pB. If
f:
A -> B
namely the order
^
the Gontraotton of If to
f:
B, then
f*(f*?B)
c
A -> B
?B-
If
or
(more precisely, f~ 1 CP B ) c A ) .
n A C A ^R
to the subring
is injective and
$ A C f*(f J P A ) .
is an order, then g*(f*^A)
is injective, there obviously is such a maximum,
f
:
$.
g:
is an order on
B •> C f
g*(f*?A) = (g ° ) * ? A
(g° f ^*^A
A, and sometimes denote it
Similarly, if
A -> B,
are defined
-
We call
c
Pg
A
is an order on
> whenever
Similarly, if
^ c ) = (g° £)*? c C A, when defined. Note that in any order ? c A we have
n
A
^*?o*
which extends B, then
are two homomorphisms and c
?B
^A C A
f^A
and either of
^c C C
is an order,
1 = I 2 G ^, hence
0<^l£2£3£---
35
If ever
n = 0 G A,
n G Z, then
implies the unique ring map ring
1=0
Z -*• A
and
A = (0). Thus
l / 0 £ A
is injective and order preserving.
The
with its unique order is the initial object in the categories (POR),
Z
(PORNN), (PORCK) and (PORPP).
1.4.
Simple Refinements of Orders Let
A
be a ring, f
C A
an order, and
$ c P1
a subset closed
under sums and products and containing all squares in
A.
Then ty is an
order. Suppose
Since
a
*P[a]
a G *pf -ty
and consider the set
G f , we see that, in fact, ^[a]
is the weakest refinement of
as a simple refinement of Given an order ty C A is, in fact, an order.
Proposition 1.4.1. is an order on (i)
A
Pi + P 2 a
(ii)
*p
is an order refining
containing
a.
$.
Clearly,
We will refer to
$. a £ *p, we ask when the set
and an element
The answer is easy.
If
(A,?) G (POR),
a G A, then
$[a] = {p 1 +p 2 a
if and only if both the conditions below hold: =
°>
q;.a+q2a = 0,
Pi G^>
im 1:i
P
-es
p x = p 2 a = 0.
q i G $, implies
q ; .a=q 2 a = 0.
Equivalent conditions are: (i) 1
fa n - p = {0}
(ii)'
?a n-!pa = {0}.
(A,"P) G (POR), we define certain subsets of
Given
UJ) and
36
Dp = {a e A|pa G !p
for some
p G ^+}
Dp + = {b G A|qb G ^ +
for some
q G <$*}.
Dp
are called the derived sets of
$
and
A.
$ , respectively.
In general, Dp will not be an order on A
and, in general, Dp - Dp t {0}.
However, there are important relations between the derived sets
Dp and Dp
and refinements of *p.
Proposition 1.4.2.
If (A,sp) G (POR)
and -a £ Dp, then p[a] is
an order on A.
Proposition 1.4.3. (a)
If (A,
implies ^5 is a maximal
order. (b)
If
Remark.
(A/P) G (POR)
and 'P is a maximal order, then 'PU(-Dp) = A.
Note the similarity of these results with the definition of a
total order on A, f U (-
Proof of 1.4.2.
We verify the two conditions of 1.4.1. +
is clearly equivalent to -a £ Dp . if
(q1 + q 2 )a = 0, then
+
But Dp c Dp.
(q1 + q 2 ) (-a) = 0. Hence
Condition (i)
As for condition (ii), -a £ Dp implies
qx = q 2 = 0.
•
Proof of 1.4.3(a).
If f U (-Dp+) = A
and a £
-a G Dp + .
Thus, condition (i) of 1.4.1 does not hold for a, hence *p[a] is not an order, and *p is maximal.
Proof of 1.4.3(b).
If
an order. By 1.4.1, we must have
1.5.
•
-a G Dp, hence ^ u (-Dp) = A.
D
Remarks on the Categories (PORNN) and (PORCK) We want to extend the discussions of sections 1.3 and 1.4 to (PORNN)
and (PORCK).
Section 1.3 applies without much modification to (PORNN).
This is because the assumption assumption about
(A,*P) G (PORNN)
is really a purely algebraic
A, and does not involve *p. The only change required
concerns the souped up extension condition for f : A •> B. restrict to ideals
J C B, with
J = /T, the nil radical.
One must
Section 1.4 also applies to (PORNN), but there are simplifications. For example,
Proposition 1.5.1. (a) p
+
P2a
If
=
0
(b)
(A/P) G (PORNN), implies
If
V
a G A, then
^[a]
is an order if and only if
p, = p 2 a = 0, that is, -a ^ Dp .
(A,15) G (PORNN), then
is a maximal order if and only if
+
P U (-Dp ) = A. Proof.
The point is, condition (ii) of 1.4.1 is automatically satisfied
in (PORNN).
In fact, this condition is the defining axiom 1.1.(v) for (PORCK),
and we showed in 1.1 that (PORNN) C (PORCK).
•
Despite the redundancy of condition (ii) of 1.4.1 in (PORCK), one does not quite have 1.5.1 in (PORCK). involves more than just that
This is because the assertion
*p[a]
is an order on
D C K CP + ) = {a G A ^ x = p 2 ax ± 0, some
-a f. Drv.Q5 )
Thus
if and only if
A.
1
Specifically, define
p i G
(p, +p o a)x = 0
Ll\
(A,15[a])G (PORCK
implies
p,x = p o ax = 0.
Z
1
Z
Proposition 1.5.2. (a)
If
(A,15) G (PORCK)
(A,15 [a]) G (PORCK) (b)
If
if and only if
and
a G A, then
if and only if
+
15 [a]
is an order with
-a £ D C K C P ) .
(A,15) e (PORCK), then
is a maximal (PORCK)-order on
A
15 u (-DCKCP+)) = A.
•
We leave the proof as an exercise. (A,15) G (PORNN), then
+
Also, we mention that if
+
D C K CP ) = #P , so 1.5.2 contains 1.5.1.
The extension conditions in 1.3 also require modification in (PORCK). Suppose order
^B
(A,*p.) G (PORCK), on
B with
A -» B
fCP A ) C ^^
n 2 (X f( Pi )br)x= 0, V±eVk>
38
f:
and
bi? x e B
a ring homomorphism. (B/P J G (PORCK)
implies
Then there is an
if and only if
f (p^bfx = 0 ,
l < j < n .
This is contained in a souped up condition concerning a family of maps f : A
•+ B,
(A ,p ) G (PORCK).
Namely, there is always a smallest ideal
I C B which satisfies:
n 2 n f (p )bfx = 0 , j=l ij ij
implies
The ring
¥ = B/I
the compositions
admits an order f 7rfa:
1 < i
with
(A , $ ) -+ (B,f)
(B,f) G (PORCK)
such that all
are (PORCK)-morphisms.
The concept of contraction of order to a subring is unchanged in (PORCK)..
If
(B,
and
AC B
is a subring, then
(A/p.) G (PORCK),
where
Remarks on Integral Domains If we restrict our attention to rings without zero divisors, that is,
to integral domains, the theory of orders begins to look very much like the classical theory of formally real fields.
Proposition 1.6.1.
(A/J3) G (POR),
Let
(a) The derived set
If C A
A
an integral domain.
is an order on
Then
A.
(b) Every maximal refinement of ty contains
Dp
and
Dp
is the
intersection of the maximal refinements of *$. (c) Every maximal refinement of ty is a total order on
A.
Proof.
and
(a) Suppose
p ^ p ^
p 1 p 2 (a 1 + a 2 ) G $
and
and sums.
with
p,p 2 ^ 0.
p x ,p 2 e
Thus
Dp
Since 'P C Dp, all squares belong to
then we would have implies
e $
pqa = 0.
pa <_ 0 £ qa Again, since 1
(b) Suppose f c p . 1
of
with A
Then
Then
p ^ t a ^ ) €
is closed under products Dp.
Finally, if
p , q € f . Thus
a G Dp n (-Dp) ,
qpa £ 0 £ pqa, which
is an integral domain, a = 0. Dp C Dp 1 .
If *pf
For the second statement, suppose
is a maximal refinement a £ Dp.
By 1.4.1,
39
y> [-a] so
is an order, and can be refined to a maximal order
s
p' . Thus
-a E*p f ,
a £ *p'. Namely, if *pf
(c) This is easy from 1.4.3(b). A = f1 U (-DP1) = ^
U (^'), so
is total.
We observed in 1.2 that a field if
E
2 -aI =I (-) a a > 0.
E
•
admits an order f C E, if and only
We claim for any such order vp = Dp.
is formally real.
0 < a < b, then
is a maximal order,
0 < £D a< -, since Thus if
pa = q,
It follows from $ = Dp
Namely, if
= (b-a)/ab = ab(b-a)/(ab) 2 > 0, and
(a - Dh
p,q > 0, then
1 a = (-)q > 0. p
that any partial order ^
on a field
E
is the intersection of the total orders which refine *p. Thus any element a E E
with
a £ *p and
some refinement.
-a ^ *p can be made either positive or negative in
In particular, this applies to the weak order p
=
2
{2 a |a. G E}. Let
(A,*p) be a partially ordered integral domain and let
be the inclusion of
A
in its field of fractions.
i:
A -* E
The following is now an
easy consequence of our definitions.
Proposition 1.6.2. (a) The extended order a,b E A, belongs to (b)
i$
i*p C E
is defined.
if and only if
abp
An element
E
a/b E E , p E sp.
i*i*
(c) The functions
i*,i*
between the set of orders *p on of all orders on
E.
to total orders on
define inclusion preserving bijections A
which satisfy *p = Dp, and the set
In particular, total orders on
A
correspond bijectively
E.
•
We will prove a more general result in a later chapter on localization.
1.7.
Some Examples (1) The ring 'L of integers admits a unique order, as does its
quotient field
Q.
If A
admits an order and if every element or its
negative is a sum of squares in
A, then
weak order, which is a total order.
40
A
admits a unique order, the
(2)
Let
(A,$) G (POR)
A-valued functions on order
and let
S
S , can be ordered.
$ w , and the affine order
p g = {£:
be a set.
A S , the ring of
Then
Two natural orders are the weak S -»• A|f(s) >_ 0, all
These orders can be contracted to any subring of
s G S}.
A , and the relation
between the orders is an interesting problem in many instances.
Examples
include the ring of continuous real valued functions on a topological space X, the ring of «UX 1 ,...,X n ]
C -real functions on a C°°-manifold
M, and the ring fnl B> J.
of polynomial functions on affine n-space
More generally, let
S
be a set and let
(A $
) , s G S, be a family
of partially ordered rings indexed by the elements of J! A -*• S s € S S
of sections of the projection
ways, as can any subring.
For example, X
S.
Then the ring
can be ordered in various
might be a topological space
equipped with a sheaf of partially ordered rings.
Then the ring of global
sections of the sheaf becomes a partially ordered ring in various ways. (3)
Since the square root of a nowhere negative continuous real
function is continuous on a space affine order on
C(X) coincide.
are exactly the squares. order structure.
X, we see that the weak order and the That is, the positive functions in
Thus the algebra structure of
Any ring homomorphism from
an order, e.g., C(Y), where
Y
C(X)
C(X)
C(X)
determines the
to any ring which admits
is a topological space, is necessarily order
preserving. If
X
is reasonable, C(X)
Namely, suppose Let
^.
f(x) < 0
x E X.
for at least one point
X •> IR be a nonnegative continuous function with
p:
p E 0
f ^ '£, that is
actually has only this one order
outside some small neighborhood of
-f G Dp .
x.
This proves that the order ty on
Then C(X)
p(x) > 0 +
pf G (-£ )
and
and
admits no simple
refinements.
(4)
Let
functions on
M M.
be a
C°°-manifold,
The affine order is
C°°(M)
the ring of
C°° real
^ M = {f G C°°(M) |f (X) :> 0
all
x G M}.
Just as in the case above, *&.. admits no refinements since there always M exist
C -functions on
M
positive at a point and vanishing outside any
small neighborhood of the point.
41
On the other hand, one cannot characterize C°°(M). a
C
For example, the function function.
x +y
on
One next might guess that
^M
R?
as the squares in is not the square of
*$„ = *p , the sums of squares.
This is still not correct. Note that the zero divisors
in
C°°(M) are exactly the functions
which vanish on some open set of
M.
Suppose
f G C°°(M)
satisfies
h2f =
2 g? , h,g. G C°°(M) and h not a zero divisor. Then f G $_,, X M i=l x as is easily checked. The set of such f defines a natural order *B,,, d refining ¥> , and it turns out that ty 4 ^ w , but "2>, = 'B... W W M u M I am indebted to Paul Cohen for proofs of these results. In fact, 2 2 h f = g
one can always write x G M.
for suitable
In the other direction, there are
C
f°r
f (x) >_ 0
h,g, if
a
^
nowhere negative functions
on the real line, which vanish only at the origin, but which are not sums of squares of
(5)
C
functions.
We consider the polynomial ring
functions on affine space for all
x G R , then
f
R^.
It is true that if
n = 1
and
is a sum of squares of polynomials.
However, if
Examples are due to Hilbert.
However, if we pass to rational
functions, then every positive polynomial is a sum of squares. Hilbert problem, solved by Artin. If
f G R[X. .. . X ] in
then for suitable R[X, . . . X ] order
and
^
.
and
f (x, .. . x ) > 0 in —
h, g i G R[X X . .. X J ,
h ± 0,
(6)
42
in
h2f =
2 g 2.
If ^
is the weak
is the order defined by positive functions on
2 h f =
x = (x1 ... x ) G R ^ , then x
(x, ... x ) G R ( n ) , in
for all
is an integral domain, of course.
c *£ , . since if
borhood of
This was a
We may rewrite the result as follows.
then Artin's result becomes in our notation D£
f(x) _> 0
f(X, X 2 ...X ) >_ 0, which are not sums
n >^ 2, there are polynomials of squares.
R[X, ... X ] , ordered as real
R^
n
.
^
(
. C D^^ .
R^
,
But also,
m
2 S g.
h
Since
and if
f(x) < 0
for some
vanishes identically on some open neighh
is a polynomial, this implies
h = 0.
Return to the general construction of Example (2), (A,^) G (POR),
g S
is a set.
Suppose
BC A
is a subring.
separates points of S, by identifying all
We may as well assume
s and s
1
B
if f(s) = f(s') G A,
f e B. Suppose
Y C S
is a subset such that
f G B, f | Y = 0 implies
f = 0. We call such a Y • B-Zariski dense. the weakest topology on S
for which all sets
Y
is, in fact, dense in Z = {s g(s) = 0} are
closed, g G B. A if
B-Zariski dense
gIY — °
Y
as a f u n c t i o n
(
determines an order on B, say P Y , by Y -• A ) .
Thus ^ g
gef
Y
is the affine order on B,
which we defined earlier. Clearly, if Y, and Y_ are two such sets, then implies $ cp . Y C Y 1 l Y Y 2 l More generally, let ^ be a family of B-Zariski dense subsets of S. Suppose and
&
is a filter, that is, Y ,Y 2 G^" implies
Y- G e^f, Y D Y~ implies 5
order on
4
B, and
Y G &.
5
g G$
4
if gl
J^ C J^ are filters in <&, then 1 ^
(7) Again suppose
Then £,— = U «* x^¥<
>• 0 for some Y. G &.
Y 1 n X^ G g? ^ i
is an
Clearly, if
^^ C $ . /j /
(A,*P) G (POR),
S a set,
BC A
a subring of
functions which separates points of S. Let Z C S be a subset and let I C B be the ideal of functions in B which vanish on Z. Then the ring is,in a natural way, a subring of the ring of functions
B/I
A , and
separates points. (8) We consider as a special case of (7) , A = IR, S = R ^ , affine n-space, and B = R[X ... X ] , the polynomials in n-variables. Z CR
(n
^,
I C R[X 1 ... X ] the ideal of polynomials vanishing on Z,
then R[X 1 ... X ]/I algebraic set,
is a ring of algebraic functions on some real
specifically the Zariski closure of Z in R
However, the affine order ^ 7 L
on
Given
.
on R[X, ... X ]/I very definitely depends in
Z itself, not just the Zariski closure. (9) For example, let K be a finite simplicial complex, with vertices
v..,...,v . Embed
|K| rectilinearly in £c n ^ be sending each vertex v^
43
to the i t h unit basis vector If the
e± = (0... 0,1,0... 0), and let
Z = |K| C ]R
X.
are interpreted as coordinate functions on R , then Z is n contained in the hyperplane 2 X. = 1. The Zariski closure of Z = IK| i=l 1 is the union of all affine linear subspaces of this hyperplane through k-simplexes of |K|
|K|, k _> 0. The ideal I C IR[X1 ... X n ] vanishing on n is generated by X X. - 1 and monomials X. •...•X. , one for x x i=l x o k (e- ,...,e- )
each set of vertices of
|K|.
the
X.
of
|K| which do not span a simplex
In the affine order on IR[X1 ... X ]/I, one has become the baracentric coordinate functions on
(10) Suppose
X
0 £ X± £ 1, and |K|.
is a compact Hausdorff space, {U.}, .
is a
finite cover of X and <£. : X -»- R is a subordinate partition of unity. n That is, 2 ^ ( x ) = 1, 0 £^>i(x) £ 1, all x e X, and suppO^) C Uiil The functions ip ,.. . ,^n generate over Z (or Q or R ) a ring of real valued functions on (or Q[XX ... X n ]
X, and this ring is a quotient of Z[X, . . . X ]
or R[X ... X R ] , depending on the coefficients).
The
ideal of relations is just the ideal of relations among the <£. as functions on 0 £ X i £ 1.
X.
One has a natural partial order, on this quotient, with
One might call such rings "partially ordered rings with a
positive partition of unity". (11)
It is easy to find rings with maximal orders which are not total. A = II A^ be a direct product of rings. Then any order i is a product n ^ , where ^ C A^ are orders. This is immediate
For example, let *$ on
A
from the equation ( a ^ d j ) = (0... a. ... 0) where
1. = (0 ... 1. ... 0) = 1..
if each 'JK C A^ on
A.
44
is maximal, but
A product order II ^. np^
is clearly maximal
will essentially never be total
II - Homomorphisms and convex ideals
2.1.
Convex Ideals and Quotient Rings Fix
(A,p) G (POR).
implies
p G C.
p i G C,
i= 1,2.
are convex.
Equivalently, C
An ideal
I C A
(POR) if and only if
in
A
Thus
with I
CCA
is oonvex if
is convex if
0 <_ p £ q,
p, + p~ G C,
q6C
p. G ^, implies
Clearly, arbitrary unions and intersections of convex sets
Proposition 2.1.1.
Proof.
A subset
I
First, if
is said to be convex if it is convex as a subset.
An ideal
I C A
is the kernel of a morphism in
is convex.
I = kernel(f: (A,
f(q) = 0, then
0 £ f(p) £ f(q) = 0
in
and
0 £p
A', hence
£q
f(p) = 0.
is convex.
Conversely, suppose
I
is convex.
Consider the projection
According to 1.3, we can impose an order on n
preserving if, whenever 1 £ j £ n.
A/I
so that
IT is order
2
2
p.a. G I,
IT: A -^ A/I.
2 p. Gtyt
a. G A, then
p.a. G I,
But this is immediate from the second characterization of
convexity above.
•
The weakest order on
A/I
such that
T*CP) = {p + l|p Gf}
fr is order preserving, namely
c
A/I
will be called the induced order, or quotient order, on A/I, and will be denoted
(A/I, ^ / I ) .
Let
f: (A/P) -* (A'/p1)
convex ideal kernel
be a morphism in (POR).
(f). Then the morphism
f
Let
I C A
be the
factors as a composition
(A,?)
45
where
f,
is, ?2
is a quotient projection,
f2
is a refinement of order (that
is the identity as a ring homomorphism), and
f~
is an inclusion
of a subring with the contracted order. Notice that because of the existence of refinements of order, a map can be an order preserving isomorphism of rings, without being an isomorphism in the category (POR). on
A, then
A
In general, if
will have fewer
A
is a ring and
*£'-convex ideals than
^ C
are orders
^-convex ideals.
We now consider the operations on convex ideals of (i) intersection, (ii) forming nil radicals, (iii) sums, (iv) products, and (v) quotients.
Proposition 2.1.2.
Fix
(A/p) e (POR).
Then
(a)
An arbitrary intersection of convex ideals is convex.
(b)
If
Proof.
I C A
vT = {a e A|a n e I, some n _> 1} is convex.
is convex, then
Statement (a) is trivial.
0 < a < b 6 y T .
0 £ a11 <_ b n
Then
(b-a) (bn~l + ab n ~ 2 + • • • + a""1) .
For statement (b), suppose
for all
n, since
b n G I, then
If
b n - a11 =
a n G I, hence
a G /f.
•
It turns out that sums, products, and quotients of convex ideals need not be convex.
2.2.
Examples will be given in 2.8.
Convex Hulls Let
(A/p) G (POR),
convex ideal of
A
X C A
a subset.
which contains
I
X.
We denote by
H(X)
the smallest
Thus
convex ideal X C I
It is possible that
H(X) = A, of course.
convex hiilly or simply hull, of Here is a construction of be the ordinary ideal of
H 0 (X) = J(X)
46
A
We refer to
H(X)
as the
X.
H(X). For any subset
generated by
Y.
Define
Y C A, let
J(Y)
H n + 1 (X) = J(H n (X) U {b G A|0 < b < a, a G Hn (X)}) .
Proposition 2.2.1.
Proof. ideals of
some
m.
H(X) =
|J H (X). n >_ 0 n
Clearly, H (X) C H,(X) C ••• A, hence
Then
U H (X) n > 0 n
x G H
is an increasing sequence of
is an ideal.
(X). This proves that
m+i
Finally, for all which contains
containing
m,
H
(X)
H (X). Thus
m
Suppose
0 < x < y, " "
U H (X) n >_ 0 n
y G H (X), m
is convex.
is obviously contained in any convex ideal
n
(J
H (X)
is the smallest convex ideal
_. °
X, as claimed.
•
The following "compactness" result is useful.
Proposition 2.2.2.
Let
there is a finite subset of
Proof.
(i)
Let
be any subset and let
X, say
{ x ^ - x ^ , such that
I = \y e Aly G H({x.••-x, }), some l I I K
X C I C H(X)
and
(ii)
ideal, then necessarily Obviously, if
y 1 >y 2
I
A.
Then
y G H({x1- • -\}) •
x. e Xl. J 1
is a convex subset of
a G A, then
X p X 2 C A, e
y G H(X).
If
Then
I
is an
I = H(X).
y e I,
H(X) C H(Y), hence if easily that if
X C A
l
>
tnen
a y e I.
But also, if
X c Y c A,
H(X X ) U H(X 2 ) C H(X X U X 2 ) .
It follows
y ]_ + y 2 G- I -
•
The next result is also extremely useful.
We characterize the nil
radical of a convex hull.
Proposition 2.2.3.
v4l(Z) = {zGAJO £ z
Let
Zcf
k <_ S P i z i i=l
be any set of positive elements.
for some integer
s > 0, z ^ Z
Then
and p ^ ^ } .
Although stated for positive subsets, 2.2.3 actually characterizes the nil radical of any convex hull. X
2
2
= {x |x G X} C $.
Then
Namely, if 2
X C A
X C */H(X ), hence
AUX)
is any subset, let c Mx*)>
since
v4l(X2)
47
is convex and coincides with its own nil radical. /H(X) = 4 K X 2 ) .
so
H(X 2 ) C H(X),
But also
Thus 2.2.3 has the following corollary.
Corollary 2.2.4.
Let
X C A
be any subset.
Then
k
/H(X) = {z e Alo < z
< 2 p.xf, some s > 0 , p. e f and x. e X}. i=1
Proof of 2.2.3.
Let
i i
I = {yGA|0£y
k
2r
<_ Z p . z . ) ,
1
I = /H(Z). and (iii) imply
It suffices to prove (i) I = /T.
I D H(Z) Now,
i=i I
x
is an ideal, (ii)
Because, obviously, /H(Z) D I D Z.
and (iii) implies
We want to prove
x
I
Then (i) and (ii)
I D Ai(Z).
(ii) and (iii) are trivial, so we will prove (i).
that if y e I, say k 2 (a 2 r p i )z i . Thus i=l
0 <_ y
2r
£
2 p ^ , and
hence for any
a G A, then
It is clear
0 £ (ay) 2 r <_
i=1
ay e I.
Finally, we must show that Then for suitable
is convex,
b e A
I
is closed under sums.
and any
Let
yi> v 2
e
**
s ^> 1,
r ^> 1,
Expanding the left-hand side of (*) gives
j=0
where the
J
c.
j=l
and
d.
J
are (positive) binomial coefficients.
In particular,
C Q = 1, so transfering the negative terms to the right-hand side of (*) gives
(**)
o <
(y1+y2)4sr r •^ c . v.y-i + y o -
j-o
48
}
1
2
li But
2s 2r 2s 2r 0 0 << yy 1 << SS pp.zz . and and 00 << yy 9 5 2 p ! z ! , l 2 j=l ^ : j=i 3 J 2s 2r Replacing y, ,y~ in (**) by these larger terms
y,,y9 G I, say
p.,p! G ty, z.,z! G Z.
results in an inequality which implies
2.3.
(A;p) G (POR) throughout this section.
is a prime convex ideal, if
with
D
Maximal Convex Ideals and Prime Convex Ideals We fix
or
y, + y~ G I.
b G Q.
Equivalently,
Q ^ A
Q
and
whenever
is convex and
A/Q
A convex ideal
ab G Q, either
Q c A
a G Q
is an integral domain,
1^0. A convex ideal
whenever
Q C Q',
Q C A Q'
is a maximal convex ideal if Q ^ A
a convex ideal, either
Q' = Q
or
Q
f
and
= A.
The concept prime convex ideal just means ordinary prime ideal which is convex.
On the other hand, a maximal convex ideal may not be maximal
in the family of ordinary ideals. We first establish the existence of maximal convex ideals. Proposition 2.3.1. a convex ideal, 1 £ I
Let
(A,^) G (POR), 1 i 0
(equivalently,
I 1 A).
in
Then
A. I
Let
IC A
be
is contained in at
least one maximal convex ideal. Proof.
The usual Zorn's lemma argument applies, since the family of
all convex ideals containing
I but not containing
1
is non-empty,
partially ordered by inclusion, and satisfies the chain condition.
In particular, since that any non-zero
•
I = (0) is always a convex ideal, we conclude
(A,^) G (POR) has maximal convex ideals.
Next we characterize maximal convex ideals by the quotient ring they define.
If
and for all
(A,$) G (POR), we say a G A,
(A,?)
a ^ 0, there exists
is a semi-field if b G A
with
1 £ ab.
1 7* 0
in
Note that
this definition depends on the order ty C A, and not just the algebraic structure.
49
A
Proposition 2.3.2.
If
(A/P) G (POR), the following conditions are
equivalent. (i)
(A/P)
(ii)
is a semi-field.
Every homomorphism of
(iii)
(A,*P)
in (POR) is zero or injective.
(0) C A are the only convex ideals,
(iv)
For all
a G A,
a ± 0, we have
1 G H(a).
(v)
For all
a G A,
a ± 0, we have
1 £ pa 2 , some
Proof.
The implications
are trivial.
(i) =» (ii) => (iii) =*• (iv) and (v) => (i)
Corollary 2.2.4 gives
In general, if
Q C A
(iv) => (v).
clarified
Q
(A/Q, $/Q)
is a maximal in (POR) is
in the next section, thus 2.3.2 has the following corollary.
Let
(A/p) G (POR),
Q C A
is a maximal convex ideal if and only if
Equivalently, for all with
Q
This is more or less obvious, but in any event will be
Corollary 2.3.3. Then
•
is a convex ideal, Q / A, then
convex ideal if and only if every homomorphism of zero or injective.
pef.
a G A,
a convex ideal, Q i A. (A/Q, ^/Q)
a £ Q, there should exist
b G A
is a semi-field. and
1 £ ab + q.
q6 Q •
Partially ordered fields are semi-fields and semi-
Proposition 2.3.4.
fields are integral domains.
Proof. a i 0
The inequality
implies
a 2 ± 0.
1 £ ab
If now
implies
ax = 0,
2 2 1 <_ a b , so in any semi-field,
a ^ 0, we would get
0 <_ x 2 £ a2b x = 0,
2 for some
b, hence
x
Corollary 2.3.5.
= 0
and thus
x = 0.
Maximal convex ideals are prime.
•
We will give a second proof of 2.3.5., also based on Proposition 2.2.3. First, if convex.
Q C A Suppose
is a maximal convex ideal, then ab G Q,
b g Q.
Then
1 G H(Q,b ) .
have k
b 0 < l < 2 p.q. + pb "" "" i l X 1
50
Q = T/Q, since
2
By
v^Q
is
2.2.3, we will
for suitable
p i ? p e ty, q. G Q.
0 < a2 <
Since
Q
is convex, a
Multiplying by
a
gives
2 (p.a2)q. + pa 2 b 2 e Q. i=l
G Q,
and, since
Q = /Q,
a e Q.
Thus
Q
is
prime.
•
Finally, we relate nilpotent elements and prime convex ideals.
Proposition 2.3.6.
Let
(A,$) G (POR).
/0 =
fl
Then
Q-
Q C A prime convex ideal
Proof.
Clearly
/o c Q
if
Q
is prime.
Conversely, suppose
Then the set of convex ideals not containing any power of since
(0)
is such an ideal.
We assert such an ideal, say
let
c £ Q.
integers
b £ Q,
n,m >^ 1.
is non-empty,
Zorn's lemma applies to give a convex ideal,
maximal in this set. be G Q,
a
a g /o.
Then
a11 G H(Q,b),
By 2.2.4, there are integers
Q, is prime. a m G H(Q,c) r,s >_ 1,
Otherwise,
for some
q i G Q,
p^ G $,
i = 1,2, with 0 < a 2 r < qx + p x b 2
0 < a 2 s < q2 + p 2 c 2 .
Multiplying these equations gives
0 < a2(X+S) < q € Q ,
which is a contradiction.
Corollary 2.3.7.
Let
I C A
be a convex ideal.
n P
Proof.
Then
p.
convex prime I C p
We apply 2.3.6 to the partially ordered ring
(A/I,
and use the proposition of the next section.
51
2.4.
Relation Between Convex Ideals in Fix
(A/p) E (POR), and let
I C A
(A,?)
and
(A/I, !J)/I)
be a convex ideal.
Proposition 2.4.1. (a)
There is a natural, bijective, inclusion preserving correspondence
between ^-convex ideals J/I
of
J
of
A
which contain
I
and
p/I-convex ideals
A/I.
(b)
For any such
J D I, there is a natural isomorphism in (POR),
(A/J, $/J) 3- (A/l/j/I,
(c)
The prime convex ideals of
A
which contain
bijectively to the prime convex ideals of ideals of
A
which contain
convex ideals of
Proof.
and
J/I
Given a ^-convex
is
J C A/I
'p/I-convex. is
correspond
and the maximal convex
correspond bijectively to the maximal
A/I.
order preserving morphism Thus
I
A/I
I
p/I
J D I, the identity on
A
induces an
(A/1, *p/I) •> (A/J, !p/J), with kernel Conversely, if
convex, then
IT: A -> A/I
J = TT'^J) C A
is
J/I.
is projection, ^-convex.
The
remaining details are equally simple.
2.5.
D
Absolutely Convex Ideals Although the category (PORCK) is a full subcategory of (POR), one
sees fewer ideals as kernels of morphisms in (PORCK). of the category (PORNN), where obviously an ideal
I c A,
is the kernel of a (PORNN)-morphism if and only if "P-convex.
The same is true (A,$) E
I = /f
and
(PORNN),
I
is
In this section we will investigate kernels of morphisms in
(PORCK) . Let convex if
(A;p) E (POR), 0 ^ SL ^_ b
absolutely convex if
52
and
I C A
an ideal.
bx E I
(p+q)x E I,
implies
We say that a x e I.
p,q E*p, implies
I
is absolutely
Equivalently, px E I
and
I
is
qx E I.
Proposition 2.5.1. (a) Absolutely convex ideals are convex. (b) Arbitrary intersections of absolutely convex ideals are absolutely convex. (c) Any convex ideal
I with
I = /f
is absolutely convex.
In
particular, convex prime and maximal ideals are absolutely convex. (d)
If y c $ »
is a refinement of order and
convex, then (e)
If
I
is £'-absolutely
is ^-absolutely convex.
(A,?) -> (A 1,^ 1 )
f:
I c A
absolutely convex, then
is a morphism in (POR) and I = f" (If) c A
I ! C A1
is
is absolutely convex.
Proof. (a)
Let
x = 1, the unit in
A.
(b) Obvious. (c)
If
0 ± a <_ b
hence
and
(ax) 2 e I.
bx e I, then Since
I = /f,
0 £ ax 2 ± bx 2 , hence
ax 2 e I,
axel.
(d), (e) Obvious.
Any subset
X C A
•
is contained in a smallest absolutely convex
ideal, namely the intersection of all absolutely convex ideals containing X.
We denote this smallest absolutely convex ideal
hull of
AH(X), the absolute
X.
It is easy to give a construction of AH(X)
as a countable union
of an increasing chain of ideals of A, just as was done in 2.2 for H(X).
Namely, define
AHQ(X) =
J(X)
AH n + 1 (X) = J(AHn (X) U {ax|0 < a < b, bx e AH
where
J(Y)
is the ordinary ideal generated by
Y.
Proposition 2.5.2. (a) AH(X) = U A H (X). n
53
(b)
If y G AH(X), then there exists a finite subset of {XJ-'-XJ^}, with
(c)
X,
y E AH({x 1 ---x k }).
/AH(X) = M(X).
(d) AH(X)AH(Y) C AH(XY)
where
XY = {xy|x e X, y G Y}.
Proof: (a), (b) Very similar to 2.2,1 and 2,2.2 (c) Clearly
H(X) C AH(X).
2.5.1(c), hence
Also, M(X)
AH(X) C Af(X).
(d) We prove by induction that
is absolutely convex by
Thus,
AHn (X)AH m (Y) C AH n + m (XY).
n = m = 0, this is clear.
If
bx G AH
If
(X), 0 £ a £ b, and
y £ AH (Y), we must show m
(ax)y £ AH (XY), since n+m
typical new generator of
AH n (X).
ax
is a
But, by induction, (bx)y =
b(xy) e A H n + m _ 1 ( X Y ) and 0 £ a £ b , hence
a(xy) e A H n + m ( X Y ) .
•
Proposition 2.5.2(d) allows us to construct graded rings naturally (PORCK).
in have AH(I
Namely, if
(A,$) G (PORCK)
2
3
A 3 AH(I) 3 AH(I ) 3 AH(I ) = > • • • .
and
I CA
is any ideal, we
By 2.5.2(d), AH(In )AH(I m ) C
) . We thus have a graded ring associated to
I,
GA = 0 AH(I n )/AH(l n+1 ) . n>0 This construction does not seem natural in
(POR) because the analogue
of 2.5.2(d) for hulls, rather than absolute hulls, is not true.
Our next result is that absolutely convex ideals are closed under the quotient operation, (I : J) = {b E AjbJ C I}.
Proposition 2.5.3. subset,
Let
(A,$) E (POR),
(I : Y) = {b E A|bY C I}. Then
54
(I : Y) = (I: AH(Y)).
an ideal, Y C A
any
I is absolutely convex if and only if
(I :Y) is absolutely convex for all subsets convex, then
IC A
YCA.
Moreover, if
I
is absolutely
Proof. if
Since
(I: {l}) = I, one direction is trivial.
I is absolutely convex,
0 <_ a <_ b
b(xy) G I for all y G Y. Hence Thus
a x G (I : Y ) , which proves
and bx € (I : Y ) , then
if
(I : Y)
is absolutely convex.
b*AH (Y) C I. It suffices to prove 0 <_ a <_ c,
bJ(Y) C I. Assume inductively
b'AH _(Y) C I, and, in particular,
ex G AH n (Y), it suffices to prove
c(bx) G I, hence
(bx)y =
a(xy) = (ax)y G I for all y G Y.
For the final statement, if bY c I, then that
Conversely,
b(ax) G I. But-, b(cx) =
a(bx) = b(ax) G I.
•
In 2.8 we will give examples of convex ideals which are not absolutely convex.
We will also give examples of sums and products of absolutely convex
ideals which are not absolutely convex. We now establish the relation between absolutely convex ideals and the category (PORCK).
The defining axiom 1.1.(v) for (PORCK) is exactly the
condition (0) absolutely convex.
That is, (p1 + p 2 )x = 0 ,
p^ G ^
implies
p^x = P2X = 0. In fact, we have
Proposition 2.5.4. (i) (ii) (iii) (iv) (v) (vi)
If
(A,^) G (POR), the following are equivalent.
(A/P) e (PORCK). (0) C A
is absolutely convex,
(0 : a) is absolutely convex for all a G A. (0 : a) is convex for all a G A. (0 : X) is absolutely convex for all subsets (0 : X)
Proof.
is convex for all subsets
We have just observed that
X c A.
(i) ^ (ii). Proposition 2.5.3
gives (ii) =* (iii) and (iii) =* (iv) is trivial. if
0 £ a <_ b
hence
and bx = 0, then
X C A.
(0 : x)
Now, (iv) =* (ii) because
convex implies
a G (0 : x ) ,
ax = 0. The equivalences (iii)** (v) and (iv) <=> (vi) follow from (0 : X) =
D (0 : a) . a G X
Remark: The proposition is true with
(0) replaced by
•
I, for any
55
ideal
I c A, if statement (i) is replaced by
(A/I, ^/I) G (PORCK).
The
proof is the same. As a consequence of the results above, we state Corollary 2.5.5. (a) I C A
Let
(A,$) G (PORCK)
[respectively
(A,q3) e (PORNN) ].
An ideal
is the kernel of a (PORCK)-morphism [respectively (PORNN)-morphism]
if and only if (b)
If
I
is absolutely convex [respectively
(A/P) G (PORCK)
and
Ic A
I = /F].
is absolutely convex [respectively
(A,$) G (PORNN),
I = / F ] , then there is a natural bijection between absolutely
convex ideals
of
J
A
A/I [respectively ideals
containing J
of
A,
I
and absolutely convex ideals
J = / j , containing
I
J/I
and ideals
of
J/I
of A/I, /J7T = J/I]. (c) The inclusion functors (PORNN) •* (PORCK) •* (POR) have adjoints (POR) -> (PORCK) •> (PORNN), defined by assigning to partially ordered quotients
(A,?) G (POR)
(A/AH(0), ?/AH(0)) G (PORCK)
and
the
(A//o, ^//o) G
(PORNN). Proof.
2.6.
Whatever remains to be proved will be left as an exercise.
•
Semi-Noetherian Rings It should not be surprising that various finiteness conditions on partially
ordered rings lead to structure theorems and results which cannot possibly be proved in general.
Since a partially ordered ring consists of a ring
A,
together with an order '£ C A, natural finiteness conditions can involve either the '^-convex ideal structure or the order 'p itself, as an extension of the weak order *pw c A.
We are ultimately interested in partially ordered
structures on finitely generated extensions of real closed fields. Chapters 7-8.)
(See
Such rings are, of course, Noetherian in the classical
sense, hence no finiteness assumptions are necessary on chains of convex ideals.
However, even for this restricted class of rings, many classical
results fail to generalize to the partially ordered context.
For example,
one cannot always decompose convex ideals in such a ring as intersections
56
of primary convex ideals. the order.
One needs additional finiteness conditions on
When one examines the classical proofs of primary decomposition
and other results which fail to generalize the point which causes the difficulty often involves the fact that in commutative algebra the multiples of a single element always form an ideal, while in partially ordered algebra the smallest convex ideal containing a given element may be quite large. On the other hand, chain conditions on convex ideals have certain interesting consequences, regardless if the underlying ring in the classical sense or if the order
'p
A
is Noetherian
satisfies extra conditions.
Some-
times these results are proved most easily in the classical context of Noetherian rings as applications of primary decomposition, rather than by exploiting the chain conditions directly.
Thus it actually gives one some
added insight into these classical results to investigate chain conditions in the partially ordered context, where certain techniques of proof are unavailable. It is the purpose of this section to initiate this study.
We work
entirely in the category (PORCK) because of our frequent use of the quotient construction
(I : X)
for ideals
I C A
and subsets
X C A.
However, 2.6.1 through 2.6.4 have obvious (POR) versions as well.
Definition 2.6.1.
A ring
(A,'£) e (PORCK)
is semi-floetherian if any
of the three equivalent conditions below hold: (i)
The absolutely convex ideals of
(A,'p)
satisfy the ascending
chain condition. (ii)
Any non-empty collection of absolutely convex ideals of
(A,}5)
contains a maximal element with respect to inclusion. (iii)
Every absolutely convex ideal I = AH(x1,...,x, )
I
of
(A,'.J5) may be written
for some choice of finitely many elements
x r . . . , x k e I.
As simple applications of the definition, we state some standard results.
Proposition 2.6.2.
If
(A,'.j>) e (PORCK)
is semi-Noetherian,
I C A
57
A 1 = A/I, then
is absolutely convex and any refinement
Proof.
'£' of
(A','-P!)
p/I.
This is immediate from 2.5.5 (b).
Let us call an absolutely convex ideal I.
is semi-Noetherian for
absolutely convex, implies
Proposition 2.6.3.
If
I = I.,
•
I
or
irreducible if
1 = 1 ^
i
I = I~«
(A,'£) G (PORCK)
is semi-Noetherian, then
every absolutely convex ideal is a finite intersection of irreducible absolutely convex ideals.
Proof.
An absolutely convex ideal maximal among those not so
expressible leads to an immediate contradiction.
Proposition 2.6.4. I = /F
(A/p) e (PORCK)
is semi-Noetherian and
is a radical convex ideal, then there is a unique expression
I = P.n...n p 1 K if
If
i ^ j.
where the
P. j
Moreover, the
containing Proof.
P.
are prime convex ideals and
are precisely the minimal prime ideals
In fact, one only needs the ascending chain condition for The proof is just as in the classical case.
Uniqueness is clear since if a prime ideal
P
P
contains some
D P! D P .
P^
for some
also shows that the taining
P. £ P. 1 J
I, which are necessarily convex.
radical convex ideals.
then
•
Thus if
i,j , hence
P.
P
contains
P^.-np^,
P ^ - . - n p ^ = p|n...np^, j = 1
and
Pj = P!.
then
The argument
which occur are exactly the minimal primes con-
I, which are thus necessarily convex.
(This result will be
generalized in section 3.9.) Existence of the stated decomposition is established as follows.
Each
radical convex ideal is a finite intersection of irreducible radical convex ideals, by the chain condition on radical convex ideals. is an irreducible radical convex ideal, then ab € I, let
a,b £ I.
{P^,}
containing
58
Let
{P }
is prime.
I = /F
Otherwise, let
be all the prime convex ideals containing
containing
a
and let
be those
Pa
b.
I = n P by 2.3.7. a oi
Then
I
But if
{PaiJ
be
those
I,
Pa
On the other hand, {P }= {P ,}U{P „} a a a
since
I = I1 n I"
ab G I, hence 1
a e I ,
I1 = n p , , I" = n P „ .
where
b e I", this contradicts the irreducibility of
Since
I.
•
Directly from 2.6.3, absolutely convex ideals in a semi-Noetherian ring can be understood in terms of irreducible absolutely convex ideals. However, irreducible absolutely convex ideals need not be primary, even if the ring is Noetherian.
Despite the lack of primary decomposition,
the associated primes of an absolutely convex ideal are accessible in any semi-Noetherian ring.
Proposition 2.6.5. I C A
Suppose
(A/p) G (PORCK)
an absolutely convex ideal.
is semi-Noetherian,
Then the following sets of primes
coincide: (i)
{P|P
prime and
P = (I : x ) , some
(ii)
{P|P
prime and
(I : x)
(iii)
{P|P
prime and
P = /I : x, some
(iv)
{P|P
a minimal prime containing
is P-primary, some
Passing to
x G A}
x G A} (I : x ) , some
Moreover, this set of associated -primes of
Proof.
x e A}
I
x G A}.
is non-empty and finite.
(A/I, 'p/I), we may assume
I = (0). Certainly,
Set (i) c Set (ii) c Set (iii) c Set (iv). Moreover, among all the absolutely (0 : y ) ,
convex ideals (0 : z) = P
is prime.
y i 0, choose a maximal one, say For suppose
abz = 0 ,
hence by maximality, (0 : z) = (0 : bz) Next, suppose ideals if
(0 : y)
c £ P,
P G Set (iv), say
with
(0 : z) = (0 : cz) C P.
bz ^ 0.
(0 : z) = P.
abz = 0 ,
Namely, if
This proves
Then
In fact, if
bz ^ 0, then
(0 : z) C (o : b z ) ,
(0 : x ) .
(0 : z)
Thus
(0 : bz) C P,
c £ P, hence
(0 : z)
Among the
be maximal.
bz t 0, then
We now claim
Then
Thus Set (i) i 0.
minimal over
Otherwise, bcz = 0, some
= (0 : z ) , contradicting
az = 0.
P
az = 0.
(0 : x) C (0 : y) C P, let
(0 : z) = (0 : bz).
hence
and
bz ^ 0.
(0 : z).
b G (0 : cz)
is prime, hence
(0 : z) = (0 : bz), hence
Set (iv) c Set (i).
Finally, we must prove there are only finitely many associated primes. Suppose the set of associated primes is there are infinitely many.
Let
{P } = {(0 : x )}
AH({xrv}) = AH(x
••• x
).
and suppose If
x,, / xrt ,
59
then
P
2.5.3.
^ P
K P = ( O : x ) D H ( O : x ) =
, but
Thus
Pa D Pa
f
i
, some
i.
3,
0
3^33
many
P Y
all
j , then
infinite chain
P
C Po
Proposition 2.6.6. I,J C A
PD D Po P ^ Pi
properly contain some C P
I.
Then
P.DP . P t %
some
PQ . Po
Suppose
J c P., some
Let
P.,
Conversely, if
xJ C I,
ideal of the form prime of
x £ I, then
(I : J) ^ I.
In
consists of all elements
x.J c I, hence
(I : J) t I.
j
J C (I : x)
is contained in some maximal
I.
•
If
(A/p) e (PORCK)
I.
P
Proof. Let
is semi-Noetherian and
containing
I
I, and
Suppose
(I : x) £ P, where
xy G I,
y <£ P.
Then
(I : x)
(I : z)
I,
{P.}, are
/F = n p^.
(I : y) c P
of 2.6.5, a maximal ideal among the
I c A
contains an associated
In particular, the minimal associated primes of
exactly the minimal primes containing
I.
1 <_ j <^ k, be the associated
(0).
J
is absolutely convex, then any prime
of
A
•
(I : y ) , which, by the proof of 2.6.5, is an associated
Proposition 2.6.7.
prime of
So infinitely
is semi-Noetherian,
j , if and only if
J C p. = (I : x . ) , then j
i, by the same argument.
(A,'])) e (PORCK)
belonging to some associated prime of
If
AH({xo}) = AH(x o •••x o ) . P Pi Pn
Iterating this process leads to an
particular, the set of zero divisors in
Proof.
Let
ou 1 , call it
c ..., which is impossible.
absolutely convex ideals.
primes of
by Proposition
Choose an index among the
a , such that for infinitely many If
K n p
since
with
be prime, hence will be an associated prime of
is some associated prime
I
I C p.
By the proof
(I : y) C (I : z) c P
will
contained in
•
P.
Although irreducible absolutely convex ideals of a semi-Noetherian ring are not necessarily primary, we can prove that an irreducible ideal has a unique maximal associated prime. ideal
I C A
and prime ideal
P,
I C p C A, define
I(P) = {x G A | xs G I, some
60
In fact, for any absolutely convex
s £ P}.
(The ideal where
I(P)
A, p .
is the kernel of the composite homomorphism
is the localization of
considered in the next chapter.) properties of the ideals
A
at
P.
A**A.p^A.p./IA. . ,
This point of view will be
The following proposition lists some (A,'£) e (PORCK),
I(P). We assume
I C A
absolutely convex.
Proposition 2.6.8. (a)
I(P) =
(b)
I c i(P) c P.
(c)
If
P
n
(I : s ) , hence
primes of
is absolutely convex.
is a minimal prime containing
in fact, I(P) Moreover, if
I(P)
(A,'p)
I, then
I(P)
is P-primary;
is the smallest P-primary ideal containing
is semi-Noetherian and
{P.}
I.
is the set of associated
I, then
(d)
Px C P 2
if and only if
(e)
I = n I(P.)-
I(P 2) c I(P 1 ).
In particular, if
I
is irreducible, it has a
unique maximal associated prime. (f)
The associated primes of of
Proof.
I
contained in
I(P.)
are exactly the associated primes
P..
(a) and (b) are trivial.
The result (c) is actually true in
arbitrary commutative rings, for arbitrary ideals
I.
This point will be
discussed again in the next chapter, so we postpone the proof of (c). If
?1 C P 2 , then certainly
I(P 2) C I(P X ), but y £ P2
implies
x G I(P)
Px £ P 2 .
x^ €
J
CP 2 )
Let
c
is equivalent to
by the proof of 2.6.5.
c
But
(I : x) £ P.
(I : x) Thus
I^)*
yGP1-P2,
!(?]_)•
Also, we deduce (e) since if But any proper ideal
I(P 2)
tnis
is
(I(P.) : x i ) = P ^
Then
x^Gl,
ridiculous, since clearly
x G n I(P.), then
(I : x) £ P., all
is contained in an associated prime of x e I.
If
j. I,
The second statement of (e) now
trivial unless there is a unique maximal
above show
P1=(I:x1).
This proves (d).
follows from (d), since the intersection formula
Finally, we verify (f).
Conversely, suppose
I = n I(P.)
is non-
P..
(I : x.) = P. c P., then arguments just as
Also, if
(I(P-) : x) = Q.
is prime, then
61
Q. C P..
We must show
Q. = AH(y, •••yv.). K
Then
Q.
xy.s. E I
•••
3
verify
s- £ P.. 3
Let
It is now easy to
3
ideals.
•
Note that 2.6.8(c), (d), (e) imply that if
associated primes, then
I
I
has no embedded
is an irredundant intersection of convex primary
A standard argument shows such a representation is unique.
Convex Ideals and Intersections of Orders Let
A
be a ring, '-P-j/K
is an order on for
for some
3 3
I.
(I : xs •••s,) = Q., using 2.5.3. IK J
Remark.
2.7.
is also an associated prime of
$v
A.
? 2 > and
orders on
!p-2
C S, as in Example (6) of 1.7.
n $
S, with values
That is, for
j = 1,2,
s € S.}.
is clearly the affine order corresponding to the subset
s e S.
functions
'£ 12 = ^
'#., '$„ could be affine orders defined by suitable
? 1 2 = {f e A | f(s) >. 0, all
Points
Then
could be a ring of functions on a set
•j). = {f 6 A | f(s) >_ 0, all
Then
A.
$u.
in an ordered ring, and S.^ S
two
We are interested in relations between the convex ideals
For example, A
subsets
c A
define
f G A
with
S.US^S,
s e SjUS^.
'p.-convex ideals, j = 1,2, namely, the ideal of f(s) = 0.
Thus points of
S1US2
define
£12
convex ideals. More generally, for any
A
and
•J^-convex or ^-convex, certainly j = 1,2, then
I. n I
I
'.p, and
'^ , if
is '^12-convex.
is '^12-convex.
I c A If
I.
is either is '^.-convex,
Both these remarks are equally obvious
for absolutely convex ideals. If
I C A
is
'^12
below that we can write convex ideal, j = 1,2.
convex and
I = /F
I = l^ n I , where
I. = /T.
is a .p.-radical
For non-radical ideals, even primary ideals, such
a decomposition is impossible in general.
62
is radical, then we will show
We will illustrate these phenomena
with examples in 2.8. The main positive result in this direction is the following very useful fact, discovered by Andrew Klapper.
Proposition 2.7.1 (Klapper). ideal, p 1 2 = ^
Proof.
n
Then
'P 2 «
Suppose
P
p
is
Suppose either
was neither
P C A
is a prime 'pi:?-convex
'^-convex or
'p.-convex nor i
0 £ x £ y (rel 'pj), 0 £ u £ v (rel 'P2), with
'^-convex.
'p0-convex. £
y,v e p,
x,u
0 < x 2 u 2 < y 2 u 2 < y 2 u 2 + x 2v 2
(rel $.)
0 £ x 2 u 2 £ x 2v 2 £ y 2 u 2 + x 2 v 2
(rel 'P ) .
£ P.
Choose
Then
and
Thus
0 <_ x 2 u 2 <_ y 2 u 2 + x 2 u 2
2 2 x u € P, contradicting
Corollary 2.7.2.
(rel ' P 1 2 ) .
y 2 u 2 + x 2 v 2 e p, hence
But
x,u £ P.
•
Any maximal
p --convex ideal is either a maximal
'p,-convex ideal or a maximal '^-convex ideal.
Proof.
Maximal convex ideals are prime.
Corollary 2.7.3. I = 1^. n i
where
Proof.
P a
I.
'p12-convex ideal
is a radical
I.
which are
Let
I
can be written
'p.-convex ideal, j = 1,2.
I = n P , the intersection taken over all
ideals containing those
Any radical
•
'pl9-convex prime
L
= n p ,, L = n p where the P , are 1 a 2 a a 'p.. -convex and the P „ are those P which are 1 a" a
f
£2 -convex.
•
Corollary 2.7.4.
If ^ . . . . ^ C A
•p n ••• Op,-convex prime ideal, then IK 1 £ j £ k.
If
P
is a maximal
^
P
are orders and is
P C A
'p.-convex for some J
n ... rvp-convex ideal, then
is a j,
P
is a
maximal 'p.-convex ideal, some j. If I is a radical 'Pj n ••• n P k - convex ideal, then I = L n - . - n i , where the I. are radical 'p.-convex ideals. IK j j
Proof.
Each statement follows by induction from the results above.
63
•
Proposition 2.7.1 has a very useful application to the following problem.
Given
can the order
(A,'.p) G (POR),
Q C A
'£ be refined, keeping
seems difficult.
a convex ideal, to what extent Q
convex?
In general, this problem
However, for prime ideals in integral domains we have the
following.
Proposition 2.7.5. be a subring and let exist total orders
Proof. p[-x]
Q C A
Suppose
be a
x G K,
f = p1 = p2
or
that
is either
with
K
a field.
'pnA-coiwex prime ideal. 'p, with
x,-x £'£.
Q
still
(q2 + q x )f = q 2 P x + q ^ .
f = (q 2 P 1 + q 1 p 2 )/(q 2 + q x ) •
Let
A C K
Then there
$ n A-convex.
Then from 1.6, both
Moreover, ']) = 'p [x] n^[-x]
K.
pi,q;. e p, then
QC A
(K, '£) G (POR)
"jT c K, refining
are orders on
P2-q2x,
Let
$ [x]
since if
Either
and
f = p , +q..x =
qx - q 2 = 0
and
We now apply 2.7.1 to deduce ('£ [-x] n A) -convex.
Cp[x])n A-convex or
Zorn's
lemma then completes the proof.
Corollary 2.7.6. Q C A
be a prime
refining
'p, with
Proof.
Let
d
(A,'p) G (POR),
^-convex ideal. Q
still
'^-convex if and only if
D'£
Q
is
Let 'p c A.
D'^-convex.
were studied in 1.6.
is the intersection of all total orders on
'£, so necessity of the condition is clear. D']5 C A
an integral domain.
Then there exist total orders
Properties of derived orders
particular, D'£
A
A
In
refining
Conversely, the derived orders
are exactly the orders contracted from orders on the field of
fractions of
A, hence sufficiency follows from 2.7.5.
•
The next result generalizes 2.7.1, but the first step in the proof is the argument used in 2.7.1.
Proposition 2.7.7. on
A.
Let
QC A
is not
64
A
be a primary
absolutely convex ideal]. P
Let
'.p^-convex, Q
Let
be a ring, 'j^, £ 2 > $ 1 2 = 'j^ n f2
orders
'.p, 2-convex ideal [resp. primary P = v^
be the (prime) radical.
i-s '^-.-convex [resp.
1.-
Then if
"£, absolutely convex].
Proof. Since
P
Suppose
0 <x
(rel
P 2-convex, choose
is not
'£•.)»
Y
we conclude
x
x u
Q-
We need to show
0 <_ u £ v (rel
Raising to a large even power, we may assume The argument for 2.7.1 shows
E
G Q
v E Q
and since
:p 2 ), v E P, and
Q
x E Q. u £ P.
0 < u,v (real -Pi)-
is primary, u
£ P,
€ Q.
We now observe
p = x 2 + 2ux + v 2 = (x + u ) 2 + (v2 - u 2 ) e •£
since the first term is obviously in in
'^2.
f^
and the second term obviously
Similarly,
q = (y-x) 2 + 2u(y-x) + v 2 = (y-x+u) 2 + (v2 - u 2 ) 6 $
Now
p + q = x + (y-x) + 2uy + 2v .
and since
Q
Again, since
is Q
'£
Since
-convex, we deduce
is primary, we conclude
x ,y,v E Q, we have
p,q E Q.
Thus
p + q E Q,
p - x - v = 2ux E Q.
x E Q.
The proposition for absolutely convex ideals is a slight extension of the argument. must show (rel 2
xz E Q.
'£12)>
2
Assume
x u z E Q
as
Choose
in the
and
If and
u,v
'.p,) and
as above.
proof of 2.7.1, and
2
x z E Q.
as is easily checked. finally
0 <_ x _< y (rel
if
Thus
xz E Q
as desired.
Q C A
is a primary
p,q E -p12 pz, qz E Q,
yz € Q, some
We have 2
2
2
0 £ x u 2
(y u + x v )z E Q.
We
£ y u + x v Hence
are also as above, then
(p+q)zEQ
2
2 pz - x z - v z = 2uxz E Q, and •
'^12-convex ideal and
v^ = P
'^2~convex, it seems difficult in general to relate
'•P2-convex ideals.
z E A.
In 2.8 we give an example of such a Q = Q-, n Q2> where
be written as an intersection
We give one final generalization of 2.7.1.
Q.
Q Q
is
is both to
'^
%
£, and
which cannot p.-convex.
Again, the proof is very
similar.
Proposition 2.7.8. ideals with H i
Let
A
= (0), '£ C A
be a ring, = A/I
I
C A
a finite collection of
orders, and set
$ = A n n $
with
65
respect to the obvious inclusion
A -* IT A .
'£-convex if and only if for some
a,
Proof: to be
We certainly know some
and
I C Q .
C A ), or '
Multiplying by a suitable square in all Ig
with
Ig £ Q.
Q/I
is
is
'$ -convex.
If for all these, Q/I
failed
(Note for those
y
with
I
x
Q
is
x
x a
£ Q. x
IQ - Q, we may assume
2 )y 2 a a
x ,y
G I
for
(rel * ) . Y
(A,'£) .
Since
$ = A n n $ ,
The right-hand side belongs to
'£-convex, we have a contradiction.
part of the proposition.
2.8.
e Q,
£ Q, both sides vanish.)
we have the same inequality in so if
y a
J
Then, just as in the proof of 2.7.1, we get for all y
0 i n X 2 < Z ( n a' ctfa1 a a
Q,
This proves the "only if"
The "if" part is trivial.
•
Some Examples (1)
The ring
and
Z, and
(2)
Partially order the ring
on the line. Then p
CQ
Q C A
'£ -convex, we get inequalities
0 < x 2 < y 2 (rel $ r a — a — Ja
(0)
I
Then a prime ideal
Let
(f) C £[T]
TL has a unique order.
The only convex ideals are
TL is a semi-field.
Z[T]
as a ring of real valued functions
f(T) be an irreducible polynomial which has real roots. is a convex prime ideal, since, if
must vanish at all roots of
f, hence
0 £ p £ q
and
q E (f),
p G (f).
In fact, (f) is a maximal convex ideal.
The only ideals of
properly containing (f) will also contain some integer If such an ideal is convex, it must contain
1.
Thus
m E Z,
Z[T]
m ^ 0.
Z[T]/(f), with the
quotient order, provides another example of a semi-field which is not a field. (3) If
C
Let
(A/P) G (POR).
is convex, then
In fact, 0 <_ a £ b convex, then
66
C2
C2
If
CCA
is a subset, let
is convex, since
implies
2 2 0 £ a x
is absolutely convex.
0 <_ a 5 b
2 2 £ b x , all
C^= {a€ A|a 2 e C}. implies
0 £ a
x e A, hence if
C
<_ b . is
Now suppose
IC A
is a convex ideal.
an absolutely convex ideal. a G A.
Also, if
yvy2
ordered ring
I*
is an ideal, hence
2
ay E I 2 , all
Certainly, if y G I , then (yx + y 2 ) 2 + ^
G 1%, then
(y, + y 2 ) G I 2 , because
hence
Then
I
( 0 ) 2 = {a G A|a
is convex.
= 0}
- y 2 ) 2 = 2y 2 + 2y 2 G I,
For example, in any partially
is an absolutely convex ideal.
We can say slightly more about I 2 . If y-^^ e *2> t n e n (vi + Y^) " 2 2 y^ - y 2 = 2 v i y 2 G *• *f either I is absolutely convex or 2 G A is invertible, then
(4)
We have
Let
y^y?
G
implies
(A,?) e (POR).
Define
K = {c G A | S C = 0, some
1 £ s}.
T
and
s G A,
K
We regard the convex ideals
are convex ideals, as is easily T
and
(A/P) G (PORCK), necessarily
(5) Order Z[X]
Suppose
f(X), g(X) G Z(X).
X
same time that
Letting
f(X)
is convex.
is divisible by
T = K = (0).
0 2
X .
shows that Thus
On the other hand, (2X,X )
Consider the quotient ring expressible as
as somewhat pathological.
0 £ f(X) <_ 2nX+X 2 g(X), some
approach
2X G (2X,X 2 ), but
0 £ 1 £ 2,
K
as a ring of real valued functions on the line.
'Consider the ideal (2X,X 2 ).
since
*•
n ^ 0}
Note that if
(2X,X )
e
y ^
T = {b e AJnb = 0, some integer
T C K, and both
checked.
*'2
n + eX, n G 2,
n = 0
n G Z,
and at the 2
f(X) G (2X,X )
and
is not absolutely convex
X £ (2X,X 2 ).
Z[X]/(2X,X ) . Elements are uniquely e = 0 or 1.
Since
2 2 2 2 _ 2 (n + eX) = n +2neX + e X = n ,
the quotient ring Z[X]/(2X,X ) has a unique non-zero convex ideal, consisting of the two elements and
{0,X}.
We have
{0,X} = T = K
(as defined in Example (4))
{0,X> = (0)^ (as defined in Example (3)). (6) We continue with the example in (5), (2X,X2) C Z[X]. 2
(2X,X ) order.
2
3
4
= (4X ,2X ,X ) C z[X]
is not convex since
0 £ X
2
£ 4X 2
Thus, products of convex ideals need not be convex.
H((2X,X2) ) = (X 2 ), which is, in fact, absolutely convex.
The ideal in any
We have
Namely, if
67
0 1 P 0 0 £q(X) at the origin. or
f
q(X)f(X) e ( X 2 ) , then
and Either
q
In any case, pf
second order at the origin, hence
q
and
f
vanish,
will also vanish to
pf e (X ) .
Z[x]/(2X,X 2 ), consider the ideal
In the quotient ring 2X = 0,
vanishes to second order
vanishes to second order, both
vanishes to second order.
Since
qf
2 <= ((0): (X)), but
((0): (X)).
1 £ ((0): (X)). Thus quotients of
convex ideals need not be convex.
(7)
Order
£[X,Y]
as a ring of real valued functions on the plane.
A typical non-convex ideal is the principal ideal Certainly
X 2 ,Y 2 e H(X 2 + Y 2 ) .
XY G H(x 2 + y 2 ) .
hence
(X + Y ) .
0 £ 2X 2 + 2Y 2 + XY <_ 4X 2 + 4Y 2 ,
But also
In fact, H ( X 2 + Y 2 ) = AH(X 2 + Y 2 ) = (X 2 ,Y 2 ,XY), the
absolutely convex ideal of functions vanishing to second order at the origin. The argument in Example (6) shows that Z[X,Y].
convex ideals of
But
2 2 (X ),(Y )
are also absolutely
H((X 2 ) + (Y2)) = (X 2 ,Y 2 ,XY), hence sums of
absolutely convex ideals need not be convex.
(8)
1R[X,Y,Z]/(Z2-X-Y)
Order the ring
on the piece of the surface that is, 0 £ x,y,z. morphism
z
= x+y
Evaluation at
R[X,Y,Z]/(Z2-X-Y) -> R ,
as a ring of real valued functions
lying in the first octant of
(0,0,0)
BS
\
defines an order preserving
with kernel
(X,Y,Z).
Thus
(X,Y,Z)
is
a maximal convex ideal. (X,Y,Z) 2 = (X 2 ,Y 2 ,Z 2 ,XY,XZ,YZ).
Consider the square and
0 <_ X, 0 £ Y, we see that
(X,Y,Z) 2
is not convex.
Since
Z2 = X + Y
In fact, H((X,Y,Z)2)
(X,Y,Z ) . This example shows that products of absolutely convex ideals need not be convex. The ideal
(X,Y,Z )
on our piece of surface example, if
c:
If we regard
is, in a real sense, exactly the ideal of functions S
[0,e] •*• S
is any smooth curve,
R[X,Y,Z]/(Z2-X-Y) 2
on the entire surface
68
which vanish to second order at the origin. c(0) = (0,0,0), then
as a ring of real valued functions (3}
z = x + y in Ii?lv J, then the functions
X,Y are
For
no longer positive.
There are curves on the surface through the origin
along which
Y
X
and
c(t) = (t,-t,0).
do not vanish to second order at
In fact, with this order
(X,Y,Z)
2
(0,0,0), say,
is convex.
This
discussion illustrates how the convex ideal structure of a ring of real valued functions on a space can reflect some of the geometry of the space.
(9)
The zero set
Z
topologically a circle.
Order the ring
ring of functions on this zero set. the ideal class group group
K Q (A).
1 = y 2 + x 2 (x 2 +l)
of the equation
C(A)
2
2
in
IR^2'
2
R[X,Y]/(1-Y -X (X +1)) = A
Now,
A
is
as a
is aDedekind domain, hence
is isomorphic to the (reduced) Grothendieck
This example has been studied by Evans [27],in connection
with the problem of relating
K Q (A)
Evans showed that the ideal
(Y-1,X)
contains elements not of order
to the real K-theory of the zero set C A
is not principal, hence
Thus, the natural map
2.
Z.
C(A)
K Q (A) •*• KO(Z)
is not injective. However, suppose in defining
C(A), we replace ideals by convex ideals
and principal ideals by ideals of the form = H(Y-l), which is "principal". X 2 G H(Y-l).
hence
Y2 -l = X 2 (X 2 +1)
Then
H((Y-1),X) 2)
Namely, (Y-1)(Y+1) = Y 2 -l E X 2 (X 2 +1) € A,
(Y-1,X) 2 C H(Y-l).
Thus
shows
H(a), a G A.
Y2 -l G (Y-1,X) 2 .
(Y-1)((Y+1)-(Y-1)) = 2(Y-1) G (Y-l,x) 2 .
Conversely, the same equality
Also
(Y-l) 2 G (Y-1,X) 2 , hence
We conclude
H(Y-l) = H((Y-1,X) 2 ),
as claimed. Actually, we prefer a (PORCK) version of
C(A), that is, absolutely
convex ideals should be used in the definition. K-theory is not really so relevant.
The correct (PORCK) analogue of
should really correspond "to a "real divisor 1
turn should be
H (Z, 2/2). The group
through projective modules. zation
Ag, where
zeros on
Z.
S C A
Evans
1
The relation with real
class group" of
KO(Z)
C(A)
Z, which in
should still be approached
method was to replace
A
by a locali-
is the multiplicative set of functions with no
We will see that this localization occurs quite naturally
in the category (PORCK).
(10)
Let
K
be a finite simplicial complex, with vertices
and associated barycentric coordinate functions
X ,...,Xn-
Let
v ,...,vn A (K)
69
be the partially ordered ring of real valued functions on over Z where
by the functions I(K)
K
generated
XQ,...,X . Thus,
A^(K) = Z[X Q . . . X n ]/I(K), n by polynomials 2 X. - 1 and X- •...•Xi
is the ideal generated
1=0
for all
i0 < ... < i r
such that
(v^ ,...,v^ ) do not span an r-simplex of or In particular, if K is an n-simplex A n , then
(See 1.7, Example (9).)
Thus
A~(K)
is an integral domain if and only if
A (K) we have the order relations
K
is a simplex.
K.
In
0 £ X i <_ 1.
We have the following assertions as exercises. A^(K) •> Z = ^ ( A 0 )
(i) The non-zero (POR)-morphisms exactly to the vertices
v^
0
A •> K
of
K.
(ii) The (POR)-automorphism group of % ( A r ) group
S
which acts on
A
exactly to the inclusions of r-simplices in If
K, L
is exactly the symmetry
by permuting vertices.
(iii) The surjective (POR)-morphisms
(iv)
correspond
A ^ K ) •* ^(A 1 *)
correspond
K.
are two finite complexes,
We thus obtain a faithful embedding of the category of finite simplicial complexes in the dual category of (POR).
(11)
In this example we study the ring of polynomials
rational numbers.
It is convenient to think of
say on the real line. think of
Q[X]
Q[X]
Q[X], Q
the
as a ring of functions,
For algebraic purposes it is slightly preferable to
as a ring of functions on
Q, the real algebraic numbers.
Before beginning the discussion, we indicate two directions of generalization.
First, Q
real closure dense in
F.
could be replaced by any totally ordered field The discussion below will apply to
F, in the interval topology.
examples, no element of
70
F
F, with
F[X], provided
(This is a tricky point.
F
is
In all
will be infinitesimally small relative to
F.
That is, all intervals around F.
0
in
F
will contain non-zero elements of
However, there are examples where some intervals of
elements of
F.)
F
contain no
Secondly, the number of variables could be increased.
This would require much more work.
In fact, the study of polynomial rings
in several variables over an ordered field is, in some sense, the final goal of the whole theory.
Polynomials in one variable is essentially
zero-dimensional algebraic geometry, which is trivial from a geometric point of view. All ideals in convex ideals rings
Q[X]
We will (i) characterize the
(f(X)), and (ii) characterize certain orders on the quotient
Q[X]/(f(X)),
be given.
are principal.
f i 0.
Since this is only an example, proofs will not
We fix the weak order ty C Q[X]
is known that
throughout this example.
It
$ w , the sums of squares, coincides with the affine order,
the polynomials which assume no negative values on the real line.
Assertion 1.
An ideal
irreducible factor of Assertion 2.
Suppose h.
(f(X)) C Q[X]
f(X) has real roots.
All convex ideals
(f(X))
is convex if and only if every
is convex.
(f(X)) C Q[X]
Factor
f, say
are absolutely convex.
k i f = h. 1
k r h^ , where the
are distinct irreducible polynomials, with real roots.
real roots of
h.
by
a-••
We denote the
By the Chinese Remainder Theorem, we have an
isomorphism of rings
i=1
(hj1)
From 1.7, Example (11), we know that any order on a direct product of rings is a direct product of orders. h-
Let
Thus we are reduced to studying
k i Q[X]/(h. ) ,
irreducible. g = g Q + g 1 h i + ••• + g ^ ,
If
g(X) e Q[X], write
j,
0 <_ j £ m, be least such that
Assertion 3. only if
j > k.
We have (whence
where
degree (g.)< degree
g. i 0.
g e ? w /(h i i ) C Q[X]/(h i i ), the weak order, if and g = 0) or
j < ki
is even and
g i (a i -) > 0
for all
71
real roots
a.,
of
h..
Assertion 3 can be interpreted geometrically. zero set of
h^.
Regard
neighborhoods of to
Z..
k.-th order on
Q[X]
Z. = {a--}
be the
as a ring of germs of functions, defined on
The ideal
Z..
Let
k i (h. )
is the ideal of germs which vanish g £ (h.1)
Assertion 3 says that a function
is
k
i Q[X]/(h. )
positive in the weak order on is locally non-negative near
if and only if the germ of
g
Z..
The geometric interpretation of the weak order suggests other orders on
k i Q[X]/(h i ) .
denote by
Each root
ou . , aT. .
1>S C Q[X]/(h i i ).
ou . has two sides on the real line, which we
Any non-empty subset
A function
g £ (hu1)
S C {a.. , aT.}
belongs to
$g
is locally non-negative on those sides of those roots the subset
k i The orders P s C Q[X]/(hi )
Q[X]. The maximal orders on
by single elements of h^. y>s
if the germ of a..
g
picked out by
S c {at. , aT.}.
Assertion 4. on
defines an order
k i Q[X]/(h. )
are all quotients of orders are precisely the orders defined
{a-. , aT.}, that is, by one side of a single root of
The maximal orders on
k i Q[X]/(h. )
are thus total orders.
All the orders
are (PORCK)-orders.
(12)
If we refine the order on
convex ideal changes. algebraic numbers.
For example, let
Y
is
*PY
in Example (11), the notion of
Y C Q
be any infinite subset of
(Infinite is equivalent to Zariski dense.)
1.7, Example (6), define Let
Q[X]
<$Y C Q[x]
by
f(X) e $ y
be the usual topological closure of
Y.
if
f(y) >_ 0, all
Then an ideal
convex if and only if every irreducible factor of
one real root in (13)
As in
h(X)
y € Y.
(h(X)) c Q[X] has at least
Y.
In several examples above we have studied polynomial rings with
orders defined in terms of values of functions on subsets of affine space. There is a more intrinsic, algebraic approach to these same orders. Let with C] I
(A/P) e (POR)
{l a }
= (0). Then the natural map
define a refinement of
72
and let
s
£
s
be a collection of ^-convex ideals, A -»• IT A/I
by ]5 = A Q ^1 Q)/I ) . a a
is injective, and we can
We are in the situation of
1.7, Example (2), where
A
is now interpreted as a "ring of functions" on
the index set {a}, with value at a The order ^ C A (i) All I
in the partially ordered ring
(A/I ,ty/I) .
can also be characterized as follows,
are ^-convex
(ii) ?/I a =F/I a CA/I a (iii) f
is the union of all orders on A
satisfying (i) and (ii).
As a specific example, let A = H[X,«*'X ] , with the weak order ^ . Each
y €= IR> ' defines an evaluation homomorphism
a maximal *$ -convex ideal.
(In fact, the I , y e Ec ^, are exactly the
maximal ^ -convex ideals of A.) Certainly, quotients
on R
(n)
f] f % I = (0), and the
y € R are all isomorphic to R, which has a unique order. The
A/I
order ty C A
A -*• IR, with kernel I
is thus the affine order, of polynomials nowhere negative
. (2X,X 2 )CZ[x]
(14) Reconsider the convex but not absolutely convex ideal of Example (5). The associated primes are is not convex.
(X) and
(2,X).
However, (2,X)
This illustrates an advantage of the category (PORCK), where
associated primes are always convex. (15) Consider R[X,Y]
and the family of orders p n = '£W[X,Y, 1-Y, Y n -X],
We have P n c $ n + 1 since
n^l.
(X2, XY)
The ideal 2
is !p -absolutely convex for all n. To see this, 2
write
(X , XY) = (X) n (X , XY, Y m ) .
and
2
m
(X , XY, Y )
Y n - X = Y n + 1 - X + Yn (l-Y) . Let p = U ^ .
It is relatively easy to check that
are $ -absolutely convex if m > n.
Thus
(X)
2
(X , XY) is
'^-absolutely convex, since a contradiction of this assertion would involve only finitely many elements, hence would already be a contradiction in some :p . The associated primes of observed that
(X , XY)
(X2, XY)
are
prime
(X,Y).
We have just
has an absolutely convex primary decomposition for
each of the orders '$> . However, (X , XY) for
(X) and
has no convex primary decomposition
$> since, first, the primary component corresponding to the minimal (X) is necessarily
(X) itself in any decomposition, and secondly,
73
any
'£-convex ideal with radical
0 < X < Y
m
(rel
(X,Y)
Y m , hence
contains some
'£) . 2 (X , XY)
The ideal
is actually irreducible among
'^-convex ideals. If
but
(X^, XY) i I
since
X, since
then clearly
X £ I- n l o . 1 z
^
/ i ^ = (X),
= (X). We must then have
It follows that
(X) C / T ^ ,
/I~ must be a finite intersection 2
(X, Y - c i ) , corresponding to points on the Y-axis.
of maximal convex ideals
(X2, XY) = Ij n i 2
But now Proposition 2.6.8(c), (e) would imply that
has a
primary convex decomposition, which we know is impossible by the above discussion.
(16)
(Andrew Klapper)
'$. C A, neither
j = 1,2, prime £ ..-convex nor
We give an example of a ring
"p.-convex ideals
'p2-convex.
Note
P., such that
A, two orders I = Px n P 2
is
I = /F , so Proposition 2.7.1
does not extend to radical ideals. A = R[X] x IR[Y].
Let Then
'J^ n $ 2 = ^
P 2 = (X) x H [ Y ] . (but not P,
n
?2
^
= iyX-1] x ^ ,
x -j^, the weak order on Then
P2
is
Let
similarly, P x n P 2
Geometrically, the ring
A
n P2
E(E-l) = 0,
is
$ -convex But
would imply
'^2-convex.
can be rewritten as a quotient of a
polynomial ring in four indeterminates, 1R[X,Y,E,F]. by the relations
Px
'^,-convex).
(X,Y) ^ p
is not
x ^[Y-l].
?x = H[X] x (Y),
.p -convex (but not
'^-.-convex since otherwise
(1,Y) e P 1 n p 2 .
A.
£2 = ^
?x n P 2 = (x) x (Y). Clearly
'^-convex) and is not
Let
F(F-l) = 0,
Namely, one divides
E + F = 1,
XF = YE = 0.
The
maximal convex ideals for the weak order then correspond to the points on the two lines in F = 1,
X = 0.
IR
defined by
The variable
parametrizes the second. P x = (E,X,Y),
(17) ideals convex,
Q X ,Q 2 Q2
F=0,
Y = 0
and
E = 0,
parametrizes the first line, while
In this notation
'^ = '^[X-E],
Y
£ 2 = £ w [Y-F],
P 2 = (F,X,Y).
We give an example of two orders
'^,-convex nor
74
X
E = 1,
with is
^
= v ^ = (X,Y)
such that
^-absolutely convex, but
.p2-convex.
Since
Q
'p., '.p2 on
Q
=
Qi
Qx n
is primary with
IR[X,Y], primary is
Q2
is
'^-absolutely ne
itner
T/Q = (X,Y), we see
that Proposition 2.7.1 does not extend to primary ideals. •p1 = ? W [ X , Y , Y 2 - X ] ,
We let 2
Q 2 = (X ,Y). convex. if
* 2 = '^[X,Y,X 2 -Y].
It is not difficult to check that
In fact, Q x
is
Let
Q1
is
^-absolutely convex, where
0 £ x,y,y -x}. Similarly,
•^ = {£ e H[X,Y] |f (x,y) ^ 0
is
Q2
2
),
'^-absolutely
^
= {f eH[X,Y] |f (x,y) >_ 0
'p'-absolutely convex, where
0 £ x,y,x 2 -y}.
if
Q^CX^
Q 1 n Q 2 = (X 2,Y2,XY)
But
2 is not
'PT-convex since
0 < X < Y
(rel
'$,) and is not
'po-convex since
2
0 £ Y £ X (rel -P2). (18) a
The example in (17) still leaves open the hope that if
'^ n ' ^
convex primary ideal in some ring
'^-convex and
'^-convex, then
where
£.-convex.
Q.
is
Q
A
with
P = v^Q
Q
is
both
Q = Q-^ n Q 2
might always be written
However, we now give an example where such a
decomposition is impossible. Let Clearly
A=H[X,Y], Q
is not
£ -convex since In fact, let
^
= $ W [Y,X-Y],
'^-convex since
0 £ Y £ X (rel
Q
is
Suppose
Then
'P(S)-convex.
(In fact, Q
0 £ f £ g
as functions on
(3f/3y)(0,0) = 0.
and
is
v = i + j
S
and
vectors along the
x
v
S, and both
w
lie in
and
y
and
Q
is not
*^(S) = and we will
g e (X,Y ) = Q.
Equiva-
f(0,0) = 0
and
£(0,0) = 0. f
and
g
g
i, j
axes, respectively. f
and
'£.. n $ -convex.
We must show
w = -i + j , where
and
Q=(X,Y2).
'^(S)-absolutely convex.)
(8g/9y)(O,O) = 0.
It is obvious that
Let
'^ n $ 2 c -p(S)
We consider directional derivatives of
and
is
S = {(x,y) e IR2 | 0 £ y, y 2 £ x 2 } , and let
lently, g(0,0) = 0
the vectors
'p^
'P 2 ). However, Q
0 £ Y £ -X (rel
{f e H[X,Y] |f (x,y) >_ 0, (x,y) e S}. show
? 2 = ? W [Y,-X-Y].
with respect to
are the usual unit Since
vanish at
0 £ f £ g
on
S,
(0,0), we must have
0 <^ (0,0) £ ^ | (0,0) 3v 3v and 0 <^ (0,0) £ ^ | (0,0) . 8w 9w But then (suppressing evaluation at (0,0))
75
- 9+
3x
+
3y — 3x
3y
3x
and Q <
3f _ -3f
which implies by adding that The ideal R[X,Y].
<
-3g
t
3g
=
-3g
2 -r— (0,0) = 0, as desired. oy
Q = (X,Y2) C R[X,Y]
is irreducible among all ideals of
Thus there is certainly no decomposition
is '£.-convex.
76
3f
Q = Q1
n
Q2
where Q.
Ill • Localization
3.1.
Partial Orders on Localized Rings Let
A
be a ring and let
T c A
be a multiplicative set, that is,
a subset closed under products and containing the unit denote the ring obtained by "inverting elements of
T".
1 £ A.
We let
A
That is,
A T = {a/t|a e A, t e T } / ~
where
a^tj ~
a
2/
t
if
2
there
is
t e T
with
t ( a 1 t 2 - a 2 t 1 ) = 0.
Sums
and products are defined by
tit 2
There is a canonical ring homomorphism such that
i T (t)
is invertible for all
i™:
t ^ T
A -*• A™,
i T(a) = a/1,
and, in fact, (A T ,i T )
is
the universal morphism in the category of rings with this property. We ask if partial order
A ^T
has a partial order, say on
AT
such that, first, i T :
preserving and, secondly, the triple
Proposition 3.1.1.
(A/P) G (POR), is there a natural
Define
^T
c
(A^ ^ T ) ,
AT
(A/£) -> (AT,?PT) iT
is order
is universal in some sense?
by
!PT = {a/t|ats2 e p , some
s e T}.
Then
(a) P T (b)
iT:
is an order on
A™.
(A,P) -* (A-., ^ )
is order preserving.
77
(c)
Given a morphism
(A,$) -* (A'/p1)
f:
is invertible in A
1
for all
in (POR) such that f(t)
t e T, then there is a unique
(A T /£ T ) -* (A 1 ,? 1 ), such that
morphism in (POR), g:
(A,?) - ^ > (AT,?T)
( commutes. (d) !PT = (i^) ^ , the weakest order on A~ such that
i T is order
preserving.
Proof.
We first show that our definition
a/t E $ T
does not depend
on the choice of representative, a/t, of the element of A,p. at't" = a'tt"
then
for some
t" E T.
We thus have, for any
If a/t ~ a'/f s E T,
Cat)(t')2 (t") 2 s2 = (a l t')t 2 (t") 2 s 2 .
ats 2 E ? , then
Thus, if (a)
a ' t ' ( s t t " ) 2 € $ , hence
a f /t' E ?
r
The equations in A
(alt2 + V l
5 1
^
s
l S2
=
show that if a ^ t j , a 2 t 2 E ? T , then Also, since
(a/t)
2
2
= a /t
2
2
Vl
S
l *2
a^t^
and a t
2
2
+
¥ 2 S2 * !
&2/t2, (a^tj) (a 2 /t 2 ) E ? 2
= a t -l
2
r
E $, we see that P T
contains all squares. Finally, suppose
a/t E ^
T
and -a/t E !PT.
ats 2 s 2 £ 0 £ ats 2 s 2 , hence
Then
^T C N 1
Say, ats 2 £ 0 £ ats 2 .
at(s s + ) 2 = 0 and a/t = O G i ^ .
Thus
is an or< er
^*
(b) is trivial. (c)
We already know that there is a unique such morphism
in the category of rings. check that
78
g
Namely, g(a/t) = f(a)f(t)~
is order preserving.
1
1
E A .
g: A™ -* A' We need to
Lemma 3.1.2. x,y
invertible.
(A 1 ,? 1 ) G (POR), and let
Let
Hence
A ! , with
in
0 ± y"1 ± x" 1.
Then
a/t G ^L,, then for some
The lemma implies our result, since if a t s 2e p .
0 < x f y
f (a)f (t)f (s) 2 G $'
s G T,
f (a)f (t)f (s)2(f (t)"1£ (s)" 1 ) 2 =
and
g(a/t) The lemma itself is easy, since
3.2.
y~
= y(y~ )
y
- x~
Sufficiency of Positive Multiplicative Sets In general, if
T c A
other than elements of
T
is a multiplicative set, there are many elements which are invertible in
T = (a G A|ab G T, some
then if
and
T
is exactly the set of elements of
ab G T, then
then
1
b/ab = a.' G A T
axs = ts G T
for some
A .
b G A> ,
A
invertible in
and, conversely, if
s G T.
In fact, if
Note that
T
A^.
First,
(a/1) (x/t) = 1/1 G A T , is a multiplicative
set.
Proposition 3.2.1. TCA
as above.
If
T
f
Let
(A/p) G (POR),
T C A
a multiplicative set
C A
is any multiplicative set with
T C T' C f,
there is a natural isomorphism in the category (POR)
defined by i T ^ r ( a / t ) = a/t.
Proof. clear that
It is routine that i T T , CPT) c ? T i -
morphism in (POR).
course.)
is a ring isomorphism.
It is not quite so clear that
However, since elements of
the universal property of -* (A T ,^ T )
iT T,
(AT,^T,)
is a morphism in (POR).
T
1
It is also
i~ T ,
is a
are invertible in
in (POR) implies that
i^ T ,:
AT,
(A^,^,
(This could also be proved directly, of E
79
Turning the argument around, we see that there are multiplicative sets smaller than
T with the same localization in (POR).
Proposition 3.2.2.
Let T C A be a multiplicative set
(A/P) e (POR).
Define T + = {t t G T n T 2 = {t2|t e T}. T c T+ c T
Then
are multiplicative sets and the natural maps
(A i 2, ? T 2 ) ~* (AT+, ? T +) ~* (AT, ^ T ) i li 11 are isomorphisms in (POR).
Proof.
The first assertion is trivial and the second follows from
Proposition 3.2.1 and the observation that
T C T + C T C (T 5 ).
D
Thus, when localizing in (POR), we can get by with positive multiplicative sets.
3.3.
Refinements of an Order Induced by Certain Localizations Suppose
i T : A -> A T
is injective, (A,!p) G (POR).
Then there is a
natural refinement of *£, namely, the contraction of *PT to A, that is, ^ c i*(iT)^p = f T n A = {a e A|as 2 G f, some
It is easy to see that *$„ n A = *p + n A = !B 2 IT I 3.2.2 of the preceding section. then
T
+
If A
= ^ , the strictly positive elements of A.
This is exactly the derived set Dv£
80
A, say from Proposition
is an integral domain and T = A - {0},
^T+ H A = {aGAlpaG'P
domains.
n
s G T} .
of $
some
In this case
p G "P+ } .
discussed in 1.6 for integral
More generally, if
(A,^) G (POR)
is arbitrary and
plicative set of non-zero divisors of We conclude that any order where
T C A
iT:
coincides with
$
is the multi-
A -• k^, is injective.
has a natural refinement
is the set of non-zero divisors.
der-ived order of
*p, = ^
We will call
n A,
^,
the
(hopefully remembering that the derived set
D^
*p , only in special cases).
Proposition 3.3.1.
Proof.
^
A, then
T
If
If
T C A
(A,$) € (POR), then
CPd)
= 1^.
is the set of non-zero divisors, i^:
A -• A T
the
canonical map, one has (as a special case of a more general principle) that (iT) i*(iT) $ = (iT) p C A T . i
*
l
i
*
i *
Remark.
Now intersect with
Note that if
A
A. Q.
of
D
is an integral domain
the following "simpler" definition of an order fractions
A.
i
A
(A,*P) G (POR), then
^/-Q^ °n the field of
does not work:
The condition doesn't always say the same thing for the equal elements and
as/bs
2.4.
of
A, . unless the order
Convex Ideals in The morphism
enables to
(A,P)
iT:
and.
^
is sufficiently strong, say
$ = ^
(A^ ? T )
(A,p) -> (AT, !f T ),
T C A
a multiplicative set,
us to define correspondences between ideals of
I C A, assign the ideal
a/b
I T = IA T C A^, and to
A
and
A™.
Namely,
J C A T , assign the ideal
i^(J)-C A. If *P
(A/P) G (POR)
convex, since
i^
and
J c AT
is
$T
is a (POR)-morphism.
convex, then
i^.1^) C A
is
The other correspondence also
preserves convex ideals. Proposition 3.4.1.
If
I c A
is ^-convex, then
IT C AT
Moreover, there is a natural isomorphism in (POR), induced by
is iT:
PT A
81
where
IT:
(A/p) -* (A/I, y>/I)
Proof.
One knows from commutative algebra that
a ring isomorphism
(A/I) ,„,. 25, A^/IA
with forming quotients." to be
is the quotient projection.
(Namely,
.
iT:
A -»• A T
induces
That is, "localization commutes
Acp/IA^
has the right "universal property"
(A/I) (T>jO Consider the commutative diagram
A/I
Following the arrows clockwise from be extended, first to order
$
that =
A, we see that the order
$/I C A/I, then to
c (A/I)^..
can also be extended going counterclockwise from
*PT C A,j, can be extended under is
IT
flVl).^
^ C A
A.
can
Thus the
It follows
TF, which can only happen if kernel (IT)
1$T- convex.
The equality of orders weakest extensions of
Remark.
CP/I)^,-^ = ? T / I T
follows since both are the
$ C A, under either i • TT or
The result that
1^
is
i ° iT»
*£„,-convex could also be proved by
direct computation.
Proposition 3.4.2. (a) spondence If
J
If
J C AT
J ^ i
is an ideal, then
(J)
is injective from ideals of
is a primary ideal of (b)
If
Q C A
disjoint from
Q «• QA T T
A ^ then
i^. (J)
is a primary ideal and
a primary ideal, v ^ H T = 0, correspondence
J = i~ (J)AT.
Thus the corre-
AT
to ideals of
is a primary ideal of
Q n T = 0, then
/
v Q A T = /QA T , and
AT>
A.
QAq, C A ?
1
Q = i~ (QA T ).
is
Thus the
is a bijection between all primary ideals of
and all primary ideals of
A.
A
The correspondence commutes
with the nil-radical operation and preserves inclusions, hence also provides a bijective correspondence between prime and maximal ideals of
82
A^. and
prime and maximal ideals of I C A, then
A, disjoint from
T.
(If
I n T t 0, any ideal
IA ? = A r )
Proof.
This is a result in commutative algebra and doesn't really
involve (POR).
Of course, we get a result in (POR) as a corollary by
replacing the word "ideal" in the proposition by "convex ideal."
This
is justified by Proposition 3.4.1 above.
3.5.
•
Concave Multiplicative Sets (A/p) e (POR).
Let p 6 S
implies
is convex.
q e S.
A subset
S C A
equivalently, S
is ooncave if
is concave if and only if
Y C A, we denote by
S(Y)
Thus, S(Y) = f l S a
containing
Y.
S S Q (Y)
Given any subset
the smallest concave multiplicative set containing
where the
S
a
run over all concave multiplicative sets a
We call
We construct
Then
A- S
Arbitrary unions and intersections of concave sets are concave.
We are interested in concave multiplicative sets.
Y.
0 £ p £ q,
S(Y)
S(Y)
the concave shadow (or, simply, shadow) of
explicitly as follows.
Y.
Let
o°°
is the smallest multiplicative set containing
is the empty product.)
Y.
(1 G S Q (Y)
Let
S n + 1 (Y) = S Q (S n (Y) U { q | 0 < p < q, p € S J Y ) } ) .
S m (Y)
Then each
is a multiplicative set, S m (Y) C S
is the concave shadow of
Proposition 3.5.1. Let in
(A/P) -> (A1 ,$•)
f: A
and
morphism
t E T, then g:
(
A
S(Ty
(Y)
and
S(Y) = U S n ( Y )
Y.
Let
T C A
be a multiplicative set
(A/£) e (POR).
be a morphism in (POR) such that wherever f(s)
is invertible in A?
^s(T)^ "* (
^')
such that
A1. the
0£t£s
Then there is a unique dia
gram
83
commutes.
Proof.
First we prove a Lemma.
(S(T)) 2 C S(T 2 ) C S(T + ) C S(T).
Lemma 3.5.2.
Proof.
(
m (T))
in
(S Q (T)) 2 = T 2 = S Q ( T 2 ) .
S Q (T) = T, so
and s
Only the first inclusion is not obvious.
2
2
C Sm(T ).
S m (T)
(Sm+1(T))
S
Any element of
2
C Sm+1(T ).
Thus
(S(T))
2
S(T) = U S (T) n n
Suppose, inductively,
m+1(T)
is a product of elements either
or larger than positive elements of 2
But
S m (T).
It follows that
2
C S(T ).
D
Returning to the proposition, from 3.2 we have canonical isomorphisms in (POR) (we suppress the symbols for the orders)
A
We now replace suitable
g:
( A S m' ^SfTp
S(T 2 ) ^ A S(T + ) ^ A S(T) '
by
^AS(T+V ^S(T+ )^
(AS(T+^> ^SfT+'P "*" (A* $'^ '
But
Siven
a n d look for a a
positive multiplicative
+
T ,
set
S(T + ) = {s|o £ t £ s, t € T + } ,
as one verifies easily by showing the set on the right to be a concave multiplicative set.
Thus, the hypothesis of the proposition clearly gives
( A s ( T + ) , ? S ( T + ) ) •+ (A',P f ), as desired.
g:
3.6.
The Shadow of Let
1
(A/p) G (POR).
set, we study
As an important example of a concave multiplicative
S(l) = {b € A|l £ b } , the shadow of
First, an element
84
D
y G A
is invertible in
1.
Acr .
if and only if
1 <_ xy
for some
x G A; that is, if and only if
1 G H(y).
In a commutative
ring, any element not contained in some ordinary maximal ideal is a unit. If
(A,*p) G (POR), an element
ideal if and only if making all such
y
y e A
1 G H(y).
is not contained in any maximal convex
Thus, passing to
Acri.
has the effect of
units.
Secondly, the natural map
icri>.:
A -• A c r i .
oil)
need not be injective.
Z>{i.)
In fact, kernel (i c r i .) = {a G Alab = 0, some
If
1 < b, then —
0 < a —
implies
a
= 0.
kernel(ig.,.), which one sees by
2
(a.^ + a 2 )
pathological.
G
a1,a2
2 expanding
a G kernel(i c r ,O ^HlJ
< a b , so —
Consequently, 2a1a2 = 0 for any
b, 1 < b} .
and
^
- a2) .
It is the ideal
K
We regard this kernel as somewhat
of
2.8, Example (4).
We now give an alternate characterization of semi-fieIds. Proposition 3.6.1. if and only if
igriy
such a case, Ag,..
Proof. A
Let A
(A/p) G (POR).
"*"Asfl1
is
in
Then
3ective
is the field of fractions
Semi-fields were defined in 2.3.
is an integral domain, so
paragraph above.
i,.,...:
and
A. .
If
i s a fielc
^sfl") of
^#
In
A.
(A,$)
is a semi-field,
Moreover, directly from the definition of semi-field and A
are invertible in
Thus, A S ( 1 ) = A ( o ) , a field.
Conversely, if
igQ-j = A ^ A g , .
then all non-zero elements of
A
is injective and
are invertible in
paragraph above, this says exactly that
We point out that any convex proper S(l).
is a semi-field
A -*• A« <-•.-. is injective by the second
the first paragraph above, all non-zero elements of AS(1).
(A,$)
(A/p)
If
gflV
S
Q ^ is a field, Bv tne
fi
is a semi-field.
ideal of
Thus, at least the primary convex ideals of
bijectively with primary convex ideals of
A
A
A
•
is disjoint from
(A,sp)
correspond
A
( g(iV ^ s f l ^ "
(A,?) G (POR), we define the Jacobson radical of
(A,?), R(A/p),
to be the intersection of all maximal convex ideals
85
n Q = maximal convex ideal
We have the following analog of a classical result.
Proposition 3.6.2.
Let
IC A
be a convex ideal.
Then the following
are equivalent. (i) (ii) (iii)
I C R(A,?) 1 + (a/b) is invertible in For all
Proof.
A
for all
a e I, b e S(l), there is an
a e I, b e S(l).
x e A
(ii) and (iii) are clearly equivalent.
with
1 £ x(b+a).
We will deduce the
equivalence of (i) and (ii) from the classical result in commutative algebra.
Namely, one first checks that
R(A,P) = R(Ag. ., ?
S
n
Q O
A
and
using the correspondence between maximal convex ideals of (A
S(l)'»sa)>-
ThuS
ICR
(A'?)
Secondly, one checks that in (POR).
^
and only if
(A s ( 1 ) , ? g ( 1 ) ) =
IAg ( 1 ) C R(Ag ( 1 ) , ? s ( 1 ) ) . (CAs ( 1 ) ) s ( 1 ) .<$ S (l)> s ( 1 ) )
In particular, those elements of Ag,_. not contained in any
maximal convex ideal are already invertible in Now, if
IA c r .
then the elements
1+ (a/b),
Conversely, if all
a/b e IA c r i ., Thus
1+ (a/b)
= b, for suitable
1 = b/b = q/b + pa 2 /b is impossible since
and
are not in any maximal convex
(i) =» (ii). are invertible, a e I, b e S(l), but
a £ Q, for some maximal ^-convex ideal 2.2.4, 1 < q + p a
A«,-v.
is contained in all maximal convex ideals of A c
ideal, hence are invertible.
86
(A/P) and
Q C A, then
q e Q,
1- (pa2/b) = q/b
pelf.
1 e H(Q,a). Computing in
is invertible in
q/b e Q A « Q N , a proper convex ideal.
By Corollary Acri. ,
A g . ^ , which
Thus (ii) =*• (i) .
•
3.7.
Localization at a Prime Convex Ideal Another important concave multiplicative set S
a prime convex ideal
Q, that is, S = A - Q.
localized partially ordered ring unique maximal ideal
Q A ^
More generally, if = n (A - Q )
A - U Q
We denote by
( A ^ ,ty( Q O the
(A. ~,ty.~ ) . A.^. is a local ring, whose
is in fact convex.
{Q }
is a set of convex prime ideals, then
is a concave multiplicative set.
ideals of the corresponding localization Q
is the complement of
The maximal convex
i^ffn \}> ^(fr\ T.0
are tne
A,r Q .,., (at least assuming there are no non-trivial inclusions If
A
Q
is an integral domain, then for all multiplicative sets
A^, is a subring of A.-., the field of fractions of A.
C Q).
T C A,
The following is
also an analog of a standard result in commutative algebra.
If
Proposition 3.7.1.
(A/j>) e (POR), A
an integral domain, then, in
convex ideal
Proof. if
(A s(1) , P S Q ) )
x/b G A.Q.
x/b = x Q /b 0 , (since
x
0/b0
c
(A(Q)> ^(Q))
belongs to all
b Q ^ Q.
Then
A,Q. C A
S(1) C A
'Q-
Conversely,
, we can write for each
1 E H({b Q}).
We may as well assume
Q,
b
e
= x Q b 0 /b 0 ), hence we can write
1
± *, PQ- bQ1=1
for suitable
pn e p vi
X
l
X
l
{Q.} , a finite subset of the maximal x 1
convex ideals.
which shows that
and
x/b G A gflV
Finally, suppose convex ideals
since
Q.
Then
1 £ b,
0 £ x/b
0 £ xbb
1 e H({b Q }), hence for suitable
in pQ
in
(km)> ^(Q)^
for a11 maximal
(A/p), for suitable
b~ £ Q.
Again,
G ?,
87
1 <
0 £ xbb 1
But now
3.8.
in
2 p b n = b1 ' i=l i i
0£xb(b')2
(A,^), hence
and
0 <_ x/b
in
(Acri^,1
Localization in (PORCD The results in this chapter extend to the category (PORCK), as we verify
in this section.
In fact, in (PORCK) certain desired results become true,
that are not true in (POR).
Proposition 3.8.1.
Proof. i' b i
is a
Suppose
a multiplicative set.
(a1/b]L + a 2 /b 2 ) (x/b) = 0 G A T ,
2 2 a b s i i i^^ b i s i^
by
assume
w e ma
y
as wel1
b^ = b 2 , by a similar trick. t G T
with
a,xt = a o xt = 0. 1
a^ G
and
Then
•
1 1
(A,$) G (POR),
an absolutely convex ideal.
implies
(A-,, ?„) G (PORCK).
1
1 1
Let
Replacing
Also, we may
Now, (A,!})) G (PORCK)
Thus, (a./b.)(x/b) = 0 e L
Z.
I C A
assume
a i /b i G S P T .
The hypothesis then implies that there
(a1 + a2)xt = 0 G A.
Proposition 3.8.2.
of
(A,$) G (PORCK), T C A
(A T ,? T ) G (PORCK).
Then
a
Let
T C A
IA T
a multiplicative set,
is an absolutely convex ideal
AT.
Proof.
Actually, we don't have to prove this, since we know that
IA T
is convex and that in (POR) there is an isomorphism
By Proposition 3.8.1, the left-hand side is in (PORCK), hence
so is the
right-hand side.
•
Proposition 3.8.3. set
S(l) = {b G A11 < b}
Let
(A,$) G (PORCK).
consists of non-zero divisors.
is always injective in (PORCK) and the order refinenent
Proof.
Then the concave multiplicative
p
bx = 0,
0 < 1 < b, then
L
n
• A -» A
admits a canonical (PORCK)
? s ( 1 ) = P s ( 1 ) n A.
If
Thus
l»x = x = 0.
Proposition 3.8.4. set and
I C A
Let
(A/p) G (PORCK),
an absolutely convex ideal.
S C S(l)
Then
a multiplicative
I = IAg n A.
Thus, there
are natural bijective correspondences between the absolutely convex ideals of
(A,?),
(A, ? s n A)
Proof. _________
and
(Ag, $ s ) .
In general, IA^C n A = {a e Alsa G I, some i
is absolutely convex, s a £ I,
s
0 _: 1 __ > "then
s G S}.
la = a £ I.
But if
I
This proves
IA S = I. The second statement follows, since by 3.8.2 convex, hence
I = IA g H A
Lemma 3.8.5.
by some
Let
2
q G Q,
q1 E I.
Then
ab G I,
a , we assume
I
is
ty
absolutely
absolutely convex.
I C A
•
absolutely convex with
/f = Q
is a primary ideal.
a £ /f = Q.
a G $.
p E p.
^s n A
(A/J3) G (POR),
Let
a maximal convex ideal.
Proof.
is
IAC
We must prove
b G I.
Replacing
a
Now, 1 G H(Q,a), hence by 2.2.3, 0 £ 1 £ q +pa,
Raising to some large power, we get
But now, (q1 +p ! a)b G I
and since
I
0 <_ 1 < q f * p f a ,
is absolutely convex,
1 • b = b G I.
•
We can apply Lemma 3.8.5 and Proposition 3.4.2 to define symbolic powers in (POR).
Let
(A,Q. , ^ ( - Q O
G
(A,!p) G (POR), (POR)
and
absolutely convex ideal
Definition 3.8.6.
QA^
Q C A
is a maximal convex ideal.
AH(Q n A. J
Q ^
a prime convex ideal.
c A
= AH(Q nA
^-absolutely convex, primary ideal, with
3.9.
and certainly
) n A. yQ
Then, Q ^
Then We have the
AH(QnA
C A
) = QA,
is a
= Q.
Applications of Localization, I - Some Properties of Convex Prime Ideals In this section we first give an alternate proof of the fundamental
result that maximal convex ideals are prime (Corollary 2.3.5).
The argument
is due to Andrew Klapper and is more elementary than the proof given in Chapter II, where we appealed to the characterization of the radical of a convex hull (Proposition 2.2.3).
We then use localization to give a
89
.
quick proof that the nil radical of a convex ideal is an intersection of convex prime ideals (Corollary 2.3.7).
Finally, we use localization to
study minimal primes. We begin by pointing out that the existence of maximal convex ideals in any
(A,'-p)
Next, if
I C A
(A/1, '$/!) ideals of Q C A
is a trivial Zorn's lemma argument (Proposition 2.3.1). is convex, then the construction of the residue ring
is elementary, as is the correspondence between A/I
and
'p-convex ideals of
is maximal convex, then
A
(A/Q, 'p/Q)
containing
p/1 -convex
I.
Thus, if
has no nonzero proper convex
ideals.
Proposition 3.9.1. "^-convex ideal, then
Proof. (Klapper)
(A,1^) e (POR)
If
A/Q
Q C A
is an integral domain.
(A,'£)
domain.
First notice that if
is a maximal
Thus
By the paragraph above, we
that if
Q
0 f p e '|5, then
is prime.
are reduced to showing
has no nonzero proper convex ideals, then
(0 : p) = (0). Namely, if also
and
A
(0 : p) C A
0 <_ x <_ y, then
is an integral is convex, hence
0 £ xp £ yp, so if
yp = 0,
xp = 0. Now, suppose
principal ideals P 1 ,p 2 ,q 1 ,q 2 ^ f P
ab = 0, (a), (b)
with
a,b / 0.
a
and
b
are not units.
The
cannot be convex, hence we can find elements
p x + p 2 G (a), q x + q 2 G (b) , but
Multiplying, (p, +p 2 )(q, + q 2 ) (0 : p) = 0,
Then
=
0* hence
p.q- = 0.
p i £ (a), q.. £ (b) .
But this contradicts
p G 'J>.
•
Our next observation is that if set, then the construction of
(A/J5) G (POR),
(A^/p^) E (POR)
S C A
a multiplicative
in 3.1 is elementary, as are
the results in 3.4 concerning the correspondence between prime
'£c-convex o
ideals of
S, and the
Ag
and prime
f-convex ideals of
A
disjoint from
permutability of localization and residue ring constructions. just proved, we know there exist prime
'pg-convex ideals of
enables us to shorten certain arguments. Proposition 3.9.2.
Let
A g , and this
For example, we can reprove 2.^.1.
(A,'£) G (POR),
I C A
a convex ideal.
/T = n p , the intersection taken over all prime convex ideals
90
By the result
P
Then
D I.
Proof.
Passing to
(A/I, *-p/I), we may assume
trivial step is showing that if ideal
P,
f £ P.
ideal of {f }
n,
f
£ 0
all
Consider the localization
A-^ corresponds to a prime
I = (0). The non-
n, then for some prime convex (A,,, 'P~).
'^-convex ideal of
Any prime A
'p--convex
disjoint from
as desired.
•
The use of localization also aids in the study of minimal convex primes.
If
(A,^) e (POR)
of multiplicative sets S(l), the shadow of any such prime
S
S C A
I
is a convex ideal, consider the family
with
S n I = 0.
1, are such sets.
is contained in a maximal
'$„ -convex ideal containing
P , with Thus
and
I C P Q C A- SQ
S Q C A - PQ
and
Since our original
S .
Consider
gives a prime
IA g
(A - P Q ) = 0.
could have been
have proved the following
(POR)
and
Zorn's lemma easily implies that
(such primes do exist since
in
S
For example, {1}
(Ag , '£s ) . Any 'p-convex ideal A g /IA g
By maximality of A-P
t 0).
SQ,
for any prime
S Q = A - PQ.
P 3 I, we
extension of a familiar result in com-
mutative algebra.
Proposition 3.9.3.
If
(A,'.p) e (POR)
ideal, then the minimal prime ideals of 'p-convex.
3.10.
Any prime ideal containing
A I
and
I C A
containing
is a I
£-convex
are always
contains a minimal such prime.
Applications of Localization, II - Zero Divisors Our next application of localization will be to zero divisors.
The
results are most satisfactory if attention is restricted to the category (PORCK). If
A
is any ring, let
and let
N
denote the non-zero divisors.
algebra is that
D
D
denote the set of zero divisors of
A
A standard result in commutative
is a union of prime ideals.
More precisely,
D
is the
union of those prime ideals which are minimal over ideals of the form (0 : a ) , a e A.
Since
(0:1) = (0), the minimal primes of
The proof of these assertions is as follows. union of the ideals
A
are contained in
Certainly
(0 : a ) , so if we prove that whenever
P
D
is the
is a prime
91
D.
minimal over
(0 : a ) , then
P C D, we are finished.
maximal multiplicative set disjoint from x £ P-D
(0 : a).
But If
S = A- P
is a
P £ D, choose
and observe that
(J S x is a multiplicative set strictly n>0 S, but disjoint from (0 : a ) . This contradiction proves P C D.
larger than
Using the results that if and that if
(A,'£) e (POR), any minimal prime is convex
(A,'p) G (PORCK), all ideals
minimal primes over
(0: a)
are convex, hence so are
(0 : a ) , we see that we have proved the following.
Proposition 3.10.1. (a) then
If
(A,'p) G (POR)
and
P
is a minimal convex prime ideal,
P C D. (b)
If
(A,'p) e (PORCK), then
ideals minimal over ideals
D
(0 : a ) ,
is the union of all convex prime
a £ A.
•
For the sake of argument, we give two alternate proofs that in (PORCK),
D
is a union of convex prime ideals.
These arguments are
analogous to standard arguments for commutative rings. First, if fact, if
0 <_ a £ b
then so is such that let
I
(A,'£) G (PORCK), then the subset
a.
and
be = 0, then
ac = 0, so if
For example, (0)
be a maximal such ideal.
Otherwise, let
ab G I, y f. D.
0 £ y 2 n £ q" + br"
is such an ideal.
Then
a,b £ I.
I = /F.
Choose
By 2.2.4, we can write
for suitable
n _> 1,
convex set, so
x £ D
xy £ D
and hence either
We claim
0 £ x
q f ,q" G I, But
is a zero divisor,
(A/p) G (PORCK).
2n
I
and
£ q' + ar f ,
r',rM G A. and
or
y £ D. D
I C A
is prime.
x g D
I C D
The second proof also exploits the fact that
D
Then
is a
is convex if
This is, of course, equivalent to the statement that
is concave.
sense that if
b
In
By Zorn's lemma,
x G AH(I,a),
0 £ ( x y ) 2 n £ q = (q1 + ar 1 )(q"+ br") G I.
N = A- D
is convex.
Now consider the family of absolutely convex ideals
I C D.
y G AH(I,b),
D C A
N
is also a saturated multiplicative set, in the
st e N, then
s G N
and
t G N.
We digress a moment to establish a lemma of independent interest. This lemma clarifies in some abstract sense how suitable localizations of partially ordered rings will have a "good" maximal ideal spectrum.
92
Lemma 3.10.2. set.
Let
(A,£) e (POR),
Then any maximal ideal in
Proof, prime ideal
Q = PA~
for some ordinary
P C A, maximal among those disjoint from q G P, hence
q £ P, the ideal
a' e A, then
s
P
and
P+ (q)
= p" + a"q
Remark.
Q
S.
Let
0 < q < p G P.
are convex.
intersects
for some
2 2? 0 £ s = p" + a"q £ p M + a"p a contradiction.
A~
a concave multiplicative
is '.pg-convex.
If Q C A,, is maximal, we know
We will show If
Ag
S C A
p" e P
S.
s = p 1 + a'q,
If
and
p' € P,
1
a" = (a ) G A.
But
? and since
S
is concave, p" + a"p
e P n S, •
As a corollary of 3.10.2 we deduce that all semi-units in
are actually units, since the semi-units belong to no maximal convex
ideal, hence, by 3.10.2, to no ordinary maximal ideal.
This result could
also be verified by direct calculation. We now complete our second alternate proof that if D
is a union of convex prime ideals.
By the above discussion, this
follows immediately from the next lemma, by taking
Lemma 3.10.3.
If
(A,'p) e (POR), then
multiplicative set if and only if
Proof.
A- S
Ag.
S c A
S = N = A- D.
is a saturated* concave,
is a union of convex prime ideals.
The if statement is trivial.
and concave, consider
(A,'p) e (PORCK),
Conversely, if
S
is saturated
By the Remark above, the set of non-units in
Ag
is the union of the maximal convex ideals, hence the set of elements
of
A
which are non-units in Ag
ideals which are disjoint from
S.
is the union of those prime convex But the assumption that
simply says that the set of elements of A exactly 3.11.
S
is saturated
which are non-units in
is
Ag
A- S.
•
Applications of Localization, III - Minimal Primes, Isolated Sets of Primes, and Associated Invariants. Suppose
A
is a ring, P C A
a minimal prime ideal.
result in commutative algebra is that associated to
P
A standard
there is a minimum
93
P-primary ideal, Q = kernel (A -• A . p . ) . decomposition, then
P
If the ideal
(0)
has a primary
is necessarily an associated prime of
(0)
and
Q
is necessarily the associated primary component. If
(A,']5) e (POR), then any minimal prime
Q = kernel (A -*• A , p . ) .
P
is convex, as is
This remark establishes part (c) of Proposition 2.6.8,
which was postponed in Chapter II. In the interest of completeness, we prove the classical result. if
Q
1
is P-primary and
s £ P.
Thus
1
x G Q , and
xys = 0, some we must show
x E Q = kernel (A -• A . p O , then
f
s £ P.
Since also
P C I/Q.
set disjoint from
Q C Q .
Let
Secondly, if
(0). But if
Then
yep
and
is a multiplicative set disjoint from If
A
is a ring and
xy G Q,
sy £ P, we conclude
S = A - P.
S
(0) = n Q.
If
P
Js
s
component
Q
associated to
A
and
P
Finally,
is a maximal multiplicative
y11 £ Q
all
U Sy11 n>0
n, then
is a finite intersection of primary
We then have
is a minimal prime of
u £ P, then
x e Q.
S,
(0) = n Q^ A g c A .
this intersection reduces to the intersection of those Q.AC = A c .
for some
(0), contradiction.
ideals, then for any multiplicative set
since otherwise
xs = 0
First
Q.AC
with
kernel (A -*• A c ) = s
Of course, Q . n s = 0,
0
Q--
J
Q.ns = 0
S = A - P, then only the primary
can be disjoint from
S.
We state these results in our partially ordered context in the following
Proposition 3.11.1.
If
(A/P) G (POR), [resp. (PORCK)], I c A
convex ideal [resp. absolutely convex ideal] and prime over
I, then
P
is a minimal convex
Q = kernel (A -• A, p ./IA, .) is the smallest P-primary
convex [resp. absoltely convex] ideal containing
I.
Moreover, if
a decomposition as a finite intersection of primary ideals, then primary component belonging to
Remark.
Q
has is the
P.
•
If
(A,'p) e (POR)
and
PC A
convex prime ideal, consider the family of convex ideals This family is closed under the operations
(and also, of course, I, n I ) .
94
I
We can apply 3.11.1 to construct convex primary ideals associated
to certain sums and products.
/f = P.
is a
Given any convex
I
is a fixed such that
H(I 1 + l~) I
with
and
H(I,I 2 ),
/T = P, Proposition
3.11.1 gives the minimal P-primary convex ideal containing kernel (A -• A, J I A , J take
I = H(I, + I 2 )
, which we abbreviate
or
H(I,I 2 ), if
I(P). For example, we could
/T7 = P.
an analogue in (POR) of the symbolic powers of kernel (A -• A. p ./H(P n )A. p O .
I, specifically,
This construction includes P, by taking
P^ ^ =
As an exercise in the use of the explicit
construction of a hull, following:
H(X) = (j H.(X), given in 2.2, we state the , . , . , . i>0 1 I = H ( P ( n V m ) ) , then I(P) = P ( n + m ) .
If
There is a completely analogous construction for absolutely convex ideals, using absolute hulls, A H ( I 1 + I 2 ) ,
AH(I 1 I 2 ), etc.
As final applications of localization, we discuss the semi-Noetherian case.
If
(A,'£) E
(PORCK)
any multiplicative set
is semi-Noetherian, then so is
Suppose
is a multiplicative set, I C A Then the associated primes of P C A
Proof.
(A,'£) E (PORCK)
I = (0). If
I Ac
in
Ac
are exactly those primes
I
disjoint from
P = (0 : x) C A
is prime and disjoint from
PA g = (0 : x)A g = (0 : x/1) C A g .
prime of
in
Ag.
P = AH(y1...y,)> with
Conversely, suppose
(A/p)
P
exactly the ideal
PA C
S.
I, then the ideal {x E A|xy E I
some
S, then it is
PA g
is an associated
PA g = (0 : x/s) C A g . s. E S.
is an associated prime of
is semi-Noetherian, I C A
associated prime of
Thus
xy.s. = 0 E A, for suitable
(0 : xs-...s,) = P C A, hence
If
S.
(A g/IAg , '-Pg/IAg) = ((A/I) g , CP/I) S ), we may assume
easy to check (0)
is semi-Noetherian, S C A
an absolutely convex ideal disjoint from
is an associated prime of
Since
for
S c A, by 3.4.2(a).
Proposition 3.11.2.
where
(Ac, $ c )
Let
Then (0)
in
absolutely convex, P C A
A.
an
I (P) = kernel (A -* A-.p./IA,p.) y £ P}
considered in 2.6.
•
is
According
to 2.6.8(e), I = n I(P.), the intersection taken over the finitely many associated primes of
I.
by 3.11.2 and 2.6.8,
IA g = n IAgCPJVg), the intersection taken over those
associated primes and
Pi
If now
of
I
S C A
disjoint from
I(P-) = kernel (A -• A C /I(P.)A C ). 1
O
is a multiplicative set, we have
1
S.
Easily, IAgCPJVg) = I(P i )A g
Thus we have proved the following.
O
95
Proposition 3.11.3.
If
absolutely convex, S C A
(A/£) E (PORCK)
is semi-Noetherian, I C A
a multiplicative set, then
kernel (A -• A^/IAg) =
n I(P.), the intersection taken over those associated primes disjoint from
Remark. of
I
of
of
I
S.
•
As a special case of 3.11.3, let
and choose
I
P.
f. E A
not contained in
with P.
only associated primes of
f. E P. - P
Let I
f = ITf.
disjoint from
P
be an associated prime
for each associated prime n
and let S
S = {f }
_.
P.
Then the
are those contained in
P.
(J (I : f*) • Since (I : f) c (I : f 2 ) C . . n>0 is an increasing chain of absolutely convex ideals, we must have I(P) = I(P) = kernel (A -> A /IA f ) =
Thus
1
(I : f n ) , for suitably large
n.
In the classical Noetherian case the ideals the intersection of the primary components of of associated primes contained in of an -isolated set
E
P.
I
I(P)
belonging to the set
is a union of
Ep.
of associated primes, meaning
E
contains all asso-
The most general such
Thus for any isolated set of associated primes
the intersection of all primary components of E
Ep
This is a special case of the notion
ciated primes smaller than any one of its elements. E
are interpreted as
I
E,
belonging to primes of
is invariant and is characterized as
Q I(P)« In our semi-Noetherian PGE case we do not have primary decomposition in general, but the results above and in 2.6 show how many of the classical results involving isolated sets of associated primes do extend.
3.12.
Operators on the Set of Orders on a Ring Let
to
(A,£) £ (POR).
'$ a refined order
potent operators
We are interested in operators
Aj>.
( A $ = Ap)
A
which assign
In particular, functorial operators and idemare natural objects of study.
The motivation is ultimately this:
We seek to interpret partially
ordered rings as "rings of functions" in some generalized sense. formulas involving a "function"
96
f £ A
ought to imply
f
Certain
positive.
Thus
if
f
is positive and
a set, then
f
so should be '.p W
(A,'.p)
is a ring of suitably valued functions on
should be positive.
f.
But if the order
If 'p
p
and
(l + p)f
are positive,
is too small, say a weak order
or a finitely generated refinement
'p [g-,*#*gv], such function theoretic W 1 K
results will not follow by simple algebraic manipulation. natural algebraic extensions of
Thus we seek
'p, which more closely capture the behavior
of functions. We present four such operators here.
The first three are related to
localization. The operator
^.
If
(A/P) e (POR), (PORCK), or (PORNN), and
NCA
is the multiplicative set of non-zero divisors , then we have seen that iN:
(A,'p) -* (A N /P N)
'£d = A n !pN p
of
'p.
is injective and induces a natural refinement Specifically, f G '£
not a zero divisor.
The operator shadow of
$ .
(A,'p)
Let
p,q G 'p,
originally.
The operator
'£, is idempotent,
(A,'£) G (PORCK) or (PORNN).
Then
S(l), the
1, is a multiplicative set, contained in the non-zero divisors. iaf,^'b(.lj
We thus have an injection
fs
pf = q, for some
(A/p ) e (POR), (PORCK), or (PORNN),
We have
whichever category held
if
t h e r e f i n e m e n t of
'£
g i v e n by
(A,'£) -• (A c r i ., f c n j , and we denote by oil) oil)
Anf
Thus
f e f
if
( l + p ) f = q,
e
P,q ' P We have
(A,$ ) G (PORCK) or (PORNN), accordingly.
the absolutely convex ideals of verified that
operator
'pg
Given
'£ S
(A,'£)
and
(A/p )
is idempotent, 'P = CP ) • S S c
is a functorial operator in
In 3.8 we saw that
are identical.
Certainly
fc c 'p,. S Q
AN)
f = q/p,
p,q E'p,
From the point of view of functions, this gives information about p.
On the other hand, if
gives a globally defined formula for the function 'p
The
(PORCK) or (PORNN).
f E '£,, we can write formally (or in
least off the zero set of
We also
f.
f G 'pg, then
f, at f = -j^-
Thus functions in
are "non-negative" in rather a strong way. Suppose
with
A
is a ring, and
n p^ = (0). Then
A
Pi C A
is a finite collection of primes
has no nilpotent elements and
A
has only
97
finitely many minimal primes, say Moreover,
Hp
= (0)
and
Up
P , which are included among the
P..
is exactly the set of zero divisors of
A.
Proposition 3.12.1. In the situation of the paragraph above, assume either an order
'£ C A
with all
Pi
or assume orders
'.p. C A^
A -> II A^.
$ d = A H nCP a ) •
Then
Proof:
and set
convex, and set
'•]) = A n n tp^.
There are two points here.
then 'J5
p/P
Pi
than
and
£ C A
will be strong enough to agree with 'p, C IICP )
not a zero divisor,
suppose given
(A/P ,'£/P ) •
f £ A
a
G
P , since
h f = q (mod P ) , where
and since E h
a
tp C A H n 1^i. £ PQ p
A .
Still
h
pf = q,
p £ P .
£ P
p,q *='$,
Conversely,
and q (mod P )£ '£
Adding these equations gives
for any minimal
(A^'-P^,
n p - p , we may assume ijta 1 A, hz f = q £'£, since a a
n p. - p , and we may assume equality in i a and
Secondly, if
on
is clear since an equation
i^a
(0) = H p
'£
can
A n nCP ) •
can be reduced modulo
with
$,
is defined using all the
Multiplying each such equation by a square in h
(A,'p), then
will generally be a much weaker order than
The inclusion p
Pa
Consider the inclusion
First given
be computed by passing to the integral domains there are more
(A^'^) = (A/P^'p/P
P o , we conclude P
(Eh^)f = E q a G 'p
f £$,. a
•
The geometric significance of 3.12.1 will come out in Chapter V M , when we relate derived orders with positivity conditions on certain "non-degenerate" subsets of real varieties.
The non-minimal primes
correspond to proper subvarieties (hence lower dimensional or "degenerate") of the irreducible components of a variety.
The operator
$ .
Let
(A/p) G (PORNN)
canonical injection where the Consider those elements p,q G'.p. order
Such
f
{P }
and let
A -> n A/P
be the
are the minimal convex primes of
f G A, such that for some
A.
n
(f + p ) f = q,
n > 0,
form a set closed under products and contained in the
A H ITCP/P ) .
It is not clear under what conditions this set will
be closed under sums. In any event, define
n^^ >^ 0,
98
p.,q i G'p}.
Then
'£
!p
= {£ G A|f = E f ^
is an order and
2n. (fi + Pi) f i
:p C rp .
=
If
^i*
A
£or
some
has only
finitely many minimal primes U p
P , then the zero divisors
This can be used to show
The operator
$
Rewrite
is functorial on
(f
+p)f = q
positivity condition on the shadow of
'J>d = A H IICP/Pa)
f
as
f = q/(f
n
+ p) .
off the zero set of
(f n + p)f = q £ |
i(f) E *-P «(£"»•
are exactly
In this case
f.
In fact, if
S(f)
e
(A/-P) -+ ( cff) »'
can be seen to be equivalent to A^-*
ought to be a ring
f.
Q> hence each summand
these summands are in on
k, all
'^.
is
A
It is not difficult to see that any 'p-convex prime ideal Q C A k 2m +Vi)£± = <\i> P ^ i e ?• T h e n '^-convex. Suppose X f± € Q, {f± k k 2nd S ( n (£. + p O J f x J j=l i=i 1
'$ c'$
This provides a strong
i = ig/-£<»:
This is reasonable since
of functions off the zeros of
A
(A/p) e (PORNN).
f, we have the localization
and the condition
C A.
of
is
k 2n^ II (f. +p.)f- G Q, since 3 i=l x
It follows that some
f. G Q, hence by induction
f i e Q.
The operator ) G (PORNN)
y .
As a last example of an operator, consider those
so that the Jacobson radical vanishes,
R(A,?) =
f|
Q = (0) .
Q = maximal convex ideal Then n p : max. Q ^
(A,?) -
n (A/Q, ' max. Q
is injective and we define
B = A n ( n 'P/Q) . max. Q That is, '.J5
consists of elements of
A
positive at each maximal ideal.
The analogy with functions is clear, and, in fact, in a later chapter we will see that for finitely generated integral domains over real closed fields
A = R[x,'--x ] , with finitely generated extensions '$ = *P,t[g, • • -gv] in w i K
of the weak order on
A, the orders
'^
and
'£
coincide* and consist of
exactly those functions nowhere negative on the semi-algebraic set defined
99
by the relations In fact,
'P
f^x^'-x ) = 0 G A
= *J5m = {£|(f
2n
and inequalities
+ p)f = q, some
p,q G 'p,
g.(x ---x ) >_0.
n _> 0}.
This is
essentially a theorem of Stengle [22], and generalizes in several directions the work of Artin on representation of positive functions [4 ] . It is easy to check that all and that
*p m
is an idempotent operator
Finally, suppose relating the operators
(a) CP»
'^-maximal convex ideals are
n
T)
CPm) m
=
f • m
'p!, "p" are two orders on *p , 'p , p ,
'P -convex
A.
We state some formulas
above and intersections.
s
(b) CPf n ' P M ) p
('£' n ' P M ) d
The inclusions (a), (b) are trivial, as is If
f e T
p f ,p M
d
n f » ,
let
p'f = q',
not zero divisors.
Then
p"f = q",
pSq'Gf,
c
'Pd n '^ d •
p",q" € J>»,
(pfp") 2f = p'q'Cp") 2 = p"q M (P') 2 G 'P' n '^
Thus f'd n •pjj c fpi n -p") d . It is not quite as easy to formulate a result for the operator since even if the Jacobson radicals of clear when the Jacobson radical of
1
'p
'£' n $"
and
tp" vanish, it is not
vanishes.
However, in the
algebro-geometrical situations to be studied later, our rings will have the property that for each maximal convex ideal a fixed real closed field
'p ,
(A/p)
Q,- A/Q
R, therefore with a unique order.
is
In the
reduced case, (A/p) G (PORNN), the Jacobson radicals will vanish and (A,'P )
identifies with a ring of R-valued functions (ordered by func-
tional values) on the set of maximal convex ideals of such orders
'£','£" C A
the maximal convex ideals of
exactly the union of the maximal convex ideals of and clearly
100
$'
n
'P") m = '£'m n '£"• m
'P'
(A,p). p
1
n 'p"
and
For two will be
'p!f by 2.7.2,
IV • Some categorical notions
4.1.
Fibre Products Proposition 4.1.1.
The categories (POR), (PORCK), (PORNN) admit
arbitrary fibre products.
Proof.
The construction is identical in all three categories.
We
first show that direct products exist. Let fa:
(Aa, p a ) G (POR).
(C,^c) -* (Aa, $ a )
a unique (POR)-morphism diagram commutes for all
Consider
(A,?) = (II A a , II? a ).
If
is a family of morphisms in (POR), then there is f:
(C,^c) -• (A/p)
such that the following *
a:
(A,
Thus, (II A a , n f a )
is the direct product of the family
More generally, suppose
ga:
morphisms over the fixed base (Aa, $ a )
over
(B, *PB)
contraction of the order
(Aa, P a ) -> (B, *Pg) (
(B, ]5fi) .
is the subring II^ .
(Aa, $ a ) . is a family of
Then the fibre product of the n A a c IlAa, together with the
B That is,
II
(A , y )
is the ring
together with the order
n
4,2.
Fibre Sums Proposition 4.2.1.
The categories (PORNN), (PORCK) and (POR) admit
finite fibre sums.
Proof.
We first make the construction in (POR).
The fibre sums in
(PORNN) and (PORCK) require a slight modification of the construction. Let
(C/Pc) •+ (A,$A)
We begin with the ring
and
(C/Pc) -* (B/PR)
A ® B. C
We would like to impose an order on
such that the natural ring homomorhisms order preserving.
be morphisms in (POR).
A -> A ® B C
and
B -* A ® B C
A ® B C are
In general, however, this cannot be done.
Instead, we construct the smallest ideal
R C A ® B C
which satisfies the
following condition: k 2 .2 (Pi ® q±)xl G R, i= 1
Pi
G '?A , q. G P R , x. G A ® B C
4.2.2. implies
(p. ® q-)*- G R,
1 < j £ k.
This is exactly the generalized extension condition of 1.3 for the pair of maps
A-»A<8>B, C
is then the ring
B -• A ® B.
The fibre sum
(A/BJ
c
A
® (B,^D) B (C,?c)
(A ® B)/R, together with the order C
k
? = {^2
? (p. (8) q.)x*) |p. G ? A , q. G ^
x. G A ® B}
1— 1
where
L
IT: A ® B -• (A ® B)/R C C We verify that
is the projection.
(A ® B/R, !p) C
Suppose given (POR)-morphisms
f:
has the desired universal property. (A,!PA) -»> (D,?D)
such that the diagram below commutes.
102
in (POR)
and
g:
(B,?fi) -* (D,
Then there is a unique ring homomorphism' f ® g: A ® B-*D such that f is A-^A® B-*D and g c
c
k
? B -* A ® B -> D. Let I = kernel (f ® g). Suppose 2 (p. ® q.)x7 G I, x C i=l * p. G ( P A , q i e ? B , Xj_ G A ® B. Since f ® g(( P j ® q.)x2.) = f (p..)g(q..) (f ® g(x. ) ) 2 e ^ D is
1 < j < k, we deduce
(p. ® q.)x. e I, 1 < j < k. Thus,
~ "~ 3 3 3 — — factors A ® B -» (A ® B)/R -» D. Clearly, (A ® B)/R •* D C C C hence
((A ® B)/R, $ ) = ( A ® ? . ) A
c
®
R C I, and f ® g
is order preserving,
(B;f R ).
(c;pc)
In the category (PORNN), the ideal
R c A ® B must satisfy R = v^R, C There is a smallest such R, and the rest of the proof is
as well as 4.2.2. unchanged.
In the category (PORCK), the condition 4.2.2 is replaced by
R, p i e f /
q.
1=1
G p B > x. ,y e A ® B,
1
1
1
,
,
4.2.3. implies
(p. ® q.)x?ye R,
Again, there is a smallest such
1 < j < k.
R, and the rest of the proof is the
same.
•
4.3.
Direct and Inverse Limits Let
I
be a directed, partially ordered set, with relation
{(A., ? - ) } - F T
be a family of partially ordered rings.
Proposition 4.3.1. k < Jj —
in
<_. Let
I, such that '
the direct limit
Suppose given morphisms & g..
j3
= Id
lim (A.,?.) X 1
g, .:
(\>?v) "*" (A->?-)>
and 5 g,5.g , =5 g ., if m < k < Jj . kj mk mj — —
Then
exists in all categories (POR), (PORCK),
(PORNN).
~T* Proof.
The construction is the same in all categories.
Begin with the
disjoint union, k G I,
V A.. Identify a. G A. and a. G A. if there is a x 1 3 i G I x 1 i,j < k, with g.,(a.) = g.v(a.) G A v . The resulting ring is the —
IK
ring theoretic direct limit
1
J K J
Impose an order ty on A in
A
of all the $. C A..
K
lim A. = A. by taking as positive elements the images
It is easy to see that
(A/P) = lim (A i ,P i ),
in the sense of universal property.
•
103
Proposition 4.3.2.
Suppose given morphisms
f., :
(A.,p.) -> (A, , P V ) , J J K K f. . = Id and f. f.. = f. , if m < k <J i. Then 33 km jk jm' — — JK
k < j in I, such that — the inverse limit
lim ( A ^ . ) exists in all categories (POR), (PORCK) and
(PORNN). Proof.
Again the construction is the same in all three categories.
The answer is the usual ring theoretic inverse limit
jim
i C O A .
together with the contracted order
•
4.4.
Some Examples (1)
Let C = Z[X], ordered as a ring of functions on the line. Let
A = B = Z, with the unique order. by
ct(X) = 1, 3(X) = -1. Thus
at the points
Define
a: Z[X] -* Z
and 3: 2[X] -»- Z
a, 3 are simply evaluation homomorphisms
+ 1, - 1, respectively.
In the category of commutative rings with unit, 'L ® Z = Z / 2 , but in Z[X] (POR), Z ® Z = (0). Geometrically, this is preferable, since the fibre Z[X] sum should be a ring of functions on the intersection of the two subspaces {+ 1} , { - 1} , which is empty. (2)
Let
(A,$) G (POR),
(A,
(A/I, ?/I) and
(A/I, $/I) ®
I,J C A
convex ideals.
(A/P) -* (A/J, ^ / J ) .
We have projections
Then
(A/J,
(A,?) There are obvious (PORCK) and (PORNN) analogs. Note that in this example the fibre sum is (0) convex ideal of (3)
Let
a convex ideal. 104
A
which contains both
(A,$A )G (POR), Then
I
and
(A,^) -* (B/PR)
if there is no maximal
J. a (POR)-morphism and
I C A
(A/I,
Again, there are obvious (PORCK) and (PORNN) analogs. (4)
If
(A,SPA), (B,$fi) G (POR), then their direct sum in (POR) is the
fibre sum over the unique ring maps sum
Z -• A, "£ -*• B.
We denote the direct
(A,PA) ® (B,? B ). As an example, let A = Q[X]/(X 2 ),
B = Q[Y]/(Y 2 ), ordered as rings of
germs of functions at 0, module those vanishing to second order at 0. In the category of rings, A ® B = Q[X,Y]/(X 2 ,Y 2 ). 2
2
2
Moreover, an element
However, in (POR),
H(X ,Y ) = (X ,Y2,XY)
A ® B s Q[X,Y]/(X ,Y ,XY), since
2
2
by 2.8, Example 7.
a + b X + cY will be strictly positive if and only if
a > 0 G Q. So, the direct sum is the ring of germs of functions at modulo those which vanish to second order at (5) We continue with
A = Q[X]/(X 2 ),
(0,0) e ^
(0,0).
B = Q[Y]/(Y 2 ), but with refined
orders induced by functions nowhere negative for all sufficiently small positive real arguments. As a ring, the direct sum in (POR) will still be Q[X,Y]/(X 2 ,Y 2 ,XY).
However, now an element
a + b X + cY will be positive
if and only if a > 0 or a = 0 and b,c >_ 0. by ordering (6)
Q[X,Y]
This is the order induced
as a ring of functions on the first quadrant in R
Let (A/p) G (POR),
Ic A
(2) .
a convex ideal. There are various
inverse systems associated to I. One has the (POR)-inverse system (A/H(I n ), $ / H ( I n ) ) n > 1 , and the (PORCK) inverse system If
(A/AH(I n ), s £/AH(I n )) n > r
I is a prime convex ideal, we have the symbolic powers
and we can form
[k/I^n\ $/I
)
n > 1
I
of 3.8.6,
- E a ch of these inverse systems will
have a partially ordered inverse limit. (7) and
Let (A,?) G (POR),
QC A
S(a) is the shadow of a, then
Thus there is a natural (POR)-morphism
a prime convex ideal. S(a) c A - Q, since
If a £ Q
A - Q is concave. A
(A g ( a ) , ? s ( a ) ) "> ( (Q)> *(Q))"
We will see in the next chapter that the rings
(A cr .,tyQf. ) ,
a £ Q, form a natural directed system, and that there is an induced isomorphism in (POR),
^(A
s ( a )
, ? s ( a ) ) =V(A ( Q ) . ? ( Q ) ) -
105
\
V • The prime convex ideal spectrum
5.1.
The Zariski Topology Defined Let
(A/p) e (POR).
ideals of A
We denote by Spec(A,$)
the set of prime convex
and by Spec (A/£) the subset of maximal convex ideals.
We define a topology on Spec(A,p), called the Zccrishi topology, in the usual way.
The closed sets are
Z(I) = {Q|Q€E Spec(A,$), I C Q}
where
IC A
is a convex ideal.
In fact, if X C A
is any subset, let
Z(X) = {Q|Q e SpecCA,^), X C Q} .
Proposition 5.1.1. (a)
Z(X) = Z(H(X)) = Z(/H(X))
(b) 0 Z ( J ) = z ( 2 ! ) a a a a
for an
for any subset
y collection of ideals
I C A. a
for an
(c)
Z(I ) U Z(I 2 ) = Z(I 0 I 2 )
(d)
Spec (A,?) = Z((0)) and 0 = Z(A).
(e)
If X,Y C A
Proof.
X C A.
X Pair
are subsets, Z(X) c Z(Y)
o£
ideals
I X ,I 2 C A.
if and only if Y C /H(X).
The proofs are routine, in view of the Corollary 2.3.7, which
asserts vE(xT
H
Q•
D
QGZ(X) As a basis for the open sets of Spec(A,*J3) we have the distinguished open sets
106
D(a), a £ A, defined as follows:
D(a) = {Q|Q e Spec (A,?), a g Q}.
That the
{D(a)}
.
do form a basis follows from the formulas
D(a)f| D(b) = D(ab)
Spec (A,?) - Z(X) = U D(a) . a G X
We give
5.2.
Specm(A,![)) C Spec(A,$)
Some Topological Properties Proposition 5.2.1.
Q
the subspace topology.
A point
is a maximal convex ideal.
Spec(A/P)
TQ.
is
Q e Spec(A,*p)
Thus
Spec(A,^)
is closed if and only if is not
That is, given two points
T 1 , in general.
P , Q G Spec(A,p), at least
one is not contained in the closure of the other.
Proof.
For any
V C Spec(A,^), the Zariski closure of
V = Z( H
V
is given by
Q) •
Q G V This proves the first statement, and also the last since Q e {?}
implies
P C Q
Proposition 5.2.2. In particular,
Proof. all the
ai
and
P e {Q}
Q C P.
•
The basic open sets s
D(l) = Spec(A, £)
and
D(a) C Spec(A/£)
are compact.
is compact.
Suppose
D(a) C U D(a^). Then any convex prime ideal containing i must also contain a. By Corollary 2.3.7 , a G >/H({ai>). By
Proposition 2.2.2, there is a finite subset a G i/H({a1,...,av}).
Thus
D(a) C
U
{a.,...,a,}
of
{a.}
with
D(a.), as desired.
•
5.3. Irreducible Closed Sets in Recall that a closed subset of a topological space is irreducible if it is not the union of two proper closed subsets.
Proposition 5.3.1. there is a unique
Q G V
If
V C Spec(A,!p)
with
is closed and irreducible, then
V = {Q}.
107
Proof.
Uniqueness of
preceding section.
Q
(Specifically, Spec(A/£)
We prove existence of is a convex ideal.
Q
I
{b}.
as follows.
is Let
It suffices to prove that
Suppose, then, ab e I, u
is a consequence of Proposition 5.2.1 of the
Any prime
P
a £ I,
b £ I.
containing
I
T .) V = Z(I), where
I
is prime.
Consider the sets
contains
I = /f
a
or
b.
I U {a},
Thus
Z(I) = Z(I U {a}) U Z(I u {b}) .
On the other hand, since containing
I
but not
I = /F, a.
Thus
a £ I
implies there is a prime
Z(I u {a}) C Z ( I ) , and similarly,
Z(I u {b}) C z(I), which contradicts the irreducibility of
Corollary 5.3.2.
Spec(A,£)
/(0)
V.
•
is irreducible if and only if
a unique minimal convex prime ideal irreducible if and only if
P'
P.
A
Equivalently, Spec(A,$)
has is
is prime.
Consider the family of convex ideals
• I C A
such that
I = /F.
Note
that this family of ideals satisfies the ascending chain condition (every increasing chain the closed sets of decreasing chain
I, c I
C ...
Spec(A,'p) Z, D Z« ^ ...
is eventually constant) if and only if satisfy the descending chain condition (every is eventually constant.)
thesis it follows that every closed subset of of irreducible closed sets.
Spec(A,tp)
Under this hypois a finite union
Translating this in terms of ideals of
A
we
have
Corollary 5.3.3.
Suppose the family of convex ideals of
A
which
coincide with their nil radical satisfies the ascending chain condition. Let
J C A
P C A
be any proper convex ideal.
containing
Then among the convex prime ideals
J, there are finitely many minimal such
P, say
P..,...,P,,
and
/r = n p, •
108
a
5.4. Spec (A ,'ft) as a Functor If
f:
(A,£) •+ (A1 ,^1 )
convex prime ideal, then f
induces a function
is a morphism in (POR) and Q ! c A1 (Q!) C A
Q = f
f*:
X C
is a convex prime ideal.
Spec(A',£') -* Spec(A,£).
continuous in the Zariski topologies since
is a Thus
Clearly, f* is
(f*)~ Z(X) = Z(f(X)), any
A.
As special cases, we have morphisms I C A
is a convex ideal, and
iT:
IT: (A,$) -* (A/I, $/I), where
(A,'])) -• (A^/p-p), where
T C A
is a
multiplicative set. In the case of a projection, homeomorphism from
TT*: Spec (A/I,ty/I)-* Spec (A ,^)
Spec(A/I, $/I)
to the closed subset
Z(I) of
is a Spec(A/p)
This follows from 2.4. In the case of a localization, homeomorphism from
Spec(AT,'£T)
This follows from 3.4. If T set (that is, if T
i*: SpecCA™,'^) -* Spec(A,'p)
is a
to the subspace
Q D(t) of Spec(A,'p). t€T is finitely generated as a multiplicative
is the smallest multiplicative set containing, say,
t p . . . , ^ ) , then k
(1 D(t) = 0 D(t.), t6T
i=l
and this intersection is an open set in Spec(A,'£). We also point out that because the convex prime ideals of A from
T
coincide with those disjoint from the shadow
the natural map
5.5.
A
(AT,:pT) -» ( opry^S(Tp
disjoint
S(T), it follows that
induces a homeomorphism
Disconnectedness of Spec(A,'ft) In the category (POR), there is not quite as close a relation between
disconnectedness of Spec(A,'£) product
and decomposition of A
as a nontrivial
A 1 x A 2 , as there is in the category of commutative rings. We do
have the following semi-analogue. Proposition 5.5.1.
Spec(A,'^)
is disconnected if and only if there
109
2 Q
i>e2
are elements 1 2 HCe^,
n
'^
C A
* with
e
1 £ e, + e 2 ,
=
Under the hypothesis, it is clear that
Z(e 2 ) = 0, and
Z^)
t 0,
°>
e
i £
e
i »
Z(I i ) t 0,
i = 1,2, then
1 G H(I 1 + I 2)
By 2.2.4, we can write
Spec(A/p) = Z{e^) U Z(e 2 ),
i = 1,2.
Conversely, if Spec(A,'p) = Z(I X ) U Z(I 2 ),
elements.
i'e2
i = 1,2.
Proof. Z(e x )
G
Z(I 1 )
and
n
Z(I 2 ) = 0, and
I1 n I2
1 £ ej + e 2 ,
consists of nilpotent
e^ G I i n p .
(ej e p n = 0,
If
we get 2n-1 2n-l n n < (ej + ej) = a^e') + a 2 (ej) , 1 = lZn
with
a i G!p,
hence
i = 1,2.
1 ^ H(ei),
If
e i = a i ( e p n G!p.
Set
i = 1,2.
Also, e ^
1 £ e1 + e 2
1 £ H(ep,
and
A -> Ag.,.
and an order
$
S
Q^
C
Recall Acfli
Spec(A,'^) = S p e c ( A s ( 1 ) / p s ( 1 ) ) .
Proposition 5.5.2.
1'A2C
A
in
(A,^) G (PORCK), then
e . , e . e f C A,
Ag,.y
Spec(A;p) for
suitable
is disconnected subrings
i
In 5.5.1 we proved that if
are elements Now,
A
S(1)'
Proof.
If
( A S Q ) » ^ S ( 1 ) ^ ~ ( A i^i) x (A2'^2^
if and only if A
e| e I.,
(A/p) G (PORCK), we can push this result somewhat further.
that by 3.8.4 we have an injection with
= 0,
Since
u
Spec(A,$)
1 £ e, + e 2 = u,
is invertible and
is disconnected, there
e^ £ e^ ,
e.e2 = 0,
1 ^ H(e i ).
1 = e,/u + e2/u = ej + e'.
Consider the natural projection
The kernel since gives
110
of
TT is
AH(eJ) n AH(e^).
e!e* = 0, we have
But if
e!x' = elx1 = 0.
(0: e!) = (0: AH(e!)).
Thus
x
1
xf G AH(eJ) n AH(e^), then,
This follows from 2.5.3, which
= (ej + e^)x! = 0, and
IT is injective.
If
b ! ,c' £ A S . 1 . , then
7T(b'e^ + c'ep
Thus,
cr = c f e ^ + c l e ^ , and hence
b 1 = b'e^ + b f e^,
= (b',c») e Ag (1) /AH(eJ) x Ag (1) /AH(e») .
IT is surjective. b',c' e f p s ( 1 ) , then
Next, if Thus,
TT~
b'e' + c' e[ e ? s ( 1 ) , since
is order preserving and
Finally, 1 £ H(e i ) C A
TT is an isomorphism in (PORCK).
implies
1 £ AH(e i ) c A, since
Then, also, 1 £ AH(ej) C A g ^ . , since A n AH(e i )A s . 1 . = AH(e i ).
e!G
Ai^e^
AH(e£) = AH(ei)Ag
Thus the factors
= /AH(e7)
and
Ag^./AHCep
are nonzero.
D
5.6. The Structure Sheaf, I - A First Approximation on Basic Open Sets We will eventually impose on rings.
The classical idea is this:
a E A, we know "a hence
Spec(A/p)
a ^ 0 e A/Q.
is never zero".
on a basic open set That is, if
Q e D(a).
set
{al}i>0-
a
Specifically, let If
D(a) C Spec(A,*P),
Q G D ( a ) , then
So we localize and invert the element
But in our partially ordered ring if
a sheaf of partially ordered
A
S(a)
over D(a).
we know more than just
a £ Q
be the shadow of the multiplicative
b G S(a), Q e D(a), then
is positive, since then
a
a £ Q,
b £ Q.
<_ b, some
S(a) = {b|0 £ a
harder in the general case, but since
The proof is easy if
D(a) = D(a )
and
n}.
It is not much
(A cr 2^>$Cr * O D(^a j
(A cr \>'#Qr O >
we mav
as w
^ll assume
a
is positive, (see 3.2 for details).
Thus it is natural to begin by trying to invert all elements of over the basic open set
D(a). That is, to the basic open set
we assign the localized (POR) this assigns
(Ag^^g^^)
Lemma 5.6.1.
If
(A c . .,'£c, O £>laJ ^laJ to Spec (A,?).
D C a ^ = D(a 2 ),
a
^
b^a j
-,,a2
G
Since
S(a)
D(a) C Spec(A,'p),
Spec(A,'^) = D(l),
A, then there is a canonical
isomorphism in (POR),
Proof.
DCa^^) = D(a 2 )
primes containing
a«
implies that the primes containing
coincide.
Thus
y/H(a,) = Ai(a^).
Since
a^^ a
and the a
i> o
e
P»
we have by 2.2.3
111
0 £ a^1 <_ pa 2 0 < a r 2 < Hqa — 2 — I
for suitable integers
r.^,^ > 0
since now all elements of
S(a 9 )
and
p,q e:p.
The proposition follows
are invertible in
In fact, this argument really shows that if
" = *D(a2);D(ai):
defined by so that
and vice versa.
r
DCa^) C D(a 2 ), then
0 < a"^1 < pa 2 , and there is a canonical morphism in — * —
a, G ,/H(ao), hence l z (POR),
A
(A
r
p b > (Pa?)
^.a i
S(a 2 )'*S(a 2 ) ) -* ( A S(a 1 )'*S(a l ) ) - *"
r
e Ag,
1#
. , where
a € A, b >^ a^
These morphisms
,
,
,
and
is
pa2 > a ^ ,
are our
candidates for the presheaf restriction maps corresponding to the inclusions D(a,) C D(a 2 ).
of open sets
Lemma 5.6.2.
If
D(a Q ) C
D(a 1 ) C D(a 2 ), then the diagram below
commutes
where
Proof.
This is easy using the universal property of the localizations
(A cr ^y^cr 0 o^a^j b^a.j
If
Q c A
(rather than the explicit formulae for
is a convex prime ideal,
the concave multiplicative set There is thus a natural map
112
A - Q
a ^ Q, then
contains
Q G D(a)
•
and
a, hence also S(a).
D
Lemma 5.6.3.
The induced map
is an isomorphism in (POR).
•
We leave the proof to the reader, with only the reminder that one must not forget to check that the inverse ring isomorphism is order preserving. Thus, from Lemma 5.6.3 we expect the stalk of our structure sheaf at the point
Q G Spec(A,]5)
to be the usual local ring
A. Q.
with the natural
order „( Q r
5.7. The Structure Sheaf, IT - The Sheaf Axioms for Basic Open Sets We formulate the "first sheaf axiom" for basic open sets in the following.
Lemma 5. 7.1 (a) . Suppose
x/b e A
and
Lemma 5.7.1(b). is injective.
(A,$) G (POR), *>
^^ a S(aa])
From 5. 7.1 (a), the map
$cr on
. C Acr .
l
Secondly, D(a.) C D(a)
the element Since
x/b
means that
and integers
r. > 0.
.) n A C , ..
That is,
= x
Moreover, since
p >^ 1,
U l
D(a.)» 1
a^ G v/H(a), hence
0 £ a^1 £ p^
We can replace all the
p.
by
in these inequalities.
A c r a >. , we assume •H J
*D(a),D(a.) ( x / b )
D(a) =
K
p = 1 + P j + ••• + p k G
Then
is compact, hence there is a finite with
a , say, a,,...,av
p. Gtp
a.
( A g ( a ) f ? s ( a ) ) - H ( A g ^ ,» s(gfl) )
. = (npc,
fcr
G •£.
is the contraction of the natural product
First, D(a)
a
for some
all
a
IlA c r v a S(a a )
Proof 5.7.1.(a). subset of the
D(a) = U D(a ) , a,
(x/b) = 0 e A
We claim moreover that
the natural order order
Let
b > a s , some integer ~~
P S / b P S ' Cbps > a s p s > a ^ S , hence a s p s >_ a s , hence
bp s >^ a s
s > 0.
Then
bp s e S(a.).) and we can write
113
x/b = xp s /bp s e A s ( a ) .
To assert that
means that there exists a
*D(a)^
c. > a. 1 , with
r
0 < a
integer s
x/b = xp /bp
s
But now,
= 0 G Acr . o ^aj
Proof 5.7.l(b). above.
i -
a
for suitable
X X
xps c = 0, and since
q. G •£, and some *
c G S(a), we have proved
as desired,
The proof is a slight modification of the argument
xp S /bp s >_ 0
i1
with
in all
Ag,
.,
1 £ i £ k, then there exist elements
in
A.
If
xbp2s(cp2^0
xbp S (c') ^ 0
0 £ q! , we see that But
q.c. = c
aGv / H({a 1 1 ,.. . ,a. k } ) , Ik
implies
We continue with the same notation.
If C
J
D(a) = U D(a.) x i=l
k t. k < 2 q.a. < 2 X x ~ i=l ~ i=l
r > 0.
11
F
l—i
hence
(xp s/bps ) = 0 e A ^ ^
xp s c. = 0 G A.
l—i
We now use the fact that
^
in
0 < a r ' <_ S q l q . 1 ^ xbp S (c f )
A, hence also
xp s /bp s >_ 0
c' G S(a), so we conclude
in
2 q ! ( c ! ) 2 = c'
>_ 0
A g . ..
D
The "second sheaf axiom" for basic open sets would say this: D(a) = U D(a ) ot o^
and suppose
x / b G Acr . ot ot, o (_a I
fx /b ) = ip
if> (Note
Ot p
such that
Ot
DCa
, for all
Then there is an element
ct, 8
x/b G A c r . o (_B-J
p
} (x/b)
Suppose
have the property that
(x /b ) G A
D(a a o ) = D(a ) n D(aD).)
in A.
= x a /b a G A s ( a ^
for all a.
This assertion does not seem to be true in general.
However, the failure
of the second sheaf axiom does not really affect the way one defines the structure sheaf.
What it does affect is the evaluation of the global sections.
We will return to this point in the next section. What we can show is that in the category (PORCK), the second sheaf axiom for basic open sets is true.
Lemma 5.7.2. a G iB, and let x /b
a a S AS(a )
Let
(A/p) G (PORCK), D(a) G Spec(A,'f)
D(a) = U D(a ) a a have the
Specifically
be a cover,
114
Suppose the elements
Pr°P«rtX * hat *D(a ))D(a a . ) a
W),D(aaJ(VVeAS(aaq) (X p
G •£.
a
a p
a
01 p
a basic open set,
f o r a11 a
C
W
=
a p
> 6 - Then there exists
x/b6Ag(a) with » D ( a ) > D ( a a ) W W = * A 6 A S ( a a )
f
° r a 1 1 "•
Proof.
The proof is essentially the usual proof for commutative rings. k First, D(a) is compact, so D(a) = U D(a.), for some finite subset of the
a .
If
x/b G X
and
x/b
restricts to
x./b.,
^laJ
CX
1
have the same images in all
A~,
1 < i < k, then
x /b
,
1
..
But
01 Ot
D(a)= U
D(aa.),
and we already know from the "first sheaf axiom", Lemma 5.7.1, that an element in
A
. *-acr
x/b
Thus
is uniquely determined by its restrictions to the
r
must restrict to
x /b , all
a.
Acr
.. ctai
This reduces our problem to
the case of a finite cover. From the hypothesis, there are elements such that
x i - b . - c i . - x . * b i - c i . = 0, all r
0 £ ( a ^ . ) 1 ^ £ c^.
c^. G A,
i,j. Since
(A,$) G (PORCK),
r
x^b. ( a . a . ) - x . b i ( a . a . ) = 0. Replacing x i /b- by x ^ . / b ^ T , we may assume x.b. = x.b.. W em a y also assume that 0 < as . < b., l < i < k. 1331 — 1 — 1 — — Since the Pi G £
D(a i )
cover
D(a), a G /H{a ••-a,} , and we can find positives
such that
t 0
In
A , . b(^aj
1 a•
we have
1 = b/b =
k Pi 2 (-=--)b. .
^_^
D
1
Consider the element
\c
If
This proves that
vn. . n .
.(x/b) = x./b. G A c r
.,
x/b = V
1 < j < k, as desired.
5.8. The Structure Sheaf, IH - Definition Let
X = Spec(A,'p)
with the Zariski topology.
€L, of partially ordered rings on If
U C x
X.
is open, consider the product
jects naturally to
U C X. A section
s:
II ( A r m ,'£r m ) , which pro-
QGU W j U ->• n A QGU
x
n/bn
G
ArnV
and elements
We will define a sheaf
w
Q G U , is continuous if there exists a cover x /b G A c r ., such that for all ot ot o^a 1
WJ
say
s(Q) =
U = U D(a )
Q G D(a ) , ?(x /b ) = ot ot ot
115
D
is the natural map.
Definition 5.8.1.
Define
U,^) = {s: U ^
(T(U,^X), ^(U,#x)) e (POR) by
II A rn . |s = continuous section}
^
Qeu x
x
Lemma 5.8.2. U y:
CQ}
Qeu
Let V C u C X be open sets. There is a canonical map
(r(U,^x),^(U,^)) -> (r(V,^x),'j3(V,^x)) induced by the obvious projection
s: U -• n A.o. , s(Q) = xo/bo G A r o v ^ ^ ^J Q€U ^ J
Proof. Suppose
is continuous
»
say with respect to a cover U = U D(a ), and elements x /b e Acr .. a a a a ^*-aa^ Cover V, V = U D(a'). Since V C U, we have V = U D(a a'). Then the P a, 3 a ^ 3 sL : V -• n A. . is defined locally by the elements v W QGV
section
^D(a ) , D ( a a ' ) ( W (X
d
*^U V ^
= s
eA
S(aaL)'
ThuS
S
I V ^ continuous and we define
CX p
p
lv G ^W,0y)>
^y y
Proposition 5.8.3. The ^x),!p(V,^))
is obviously order preserving.
(r(U, (^) ,'jJ (U,^)) and ^
^
y:
form a sheaf on X.
Proof. The proof is routine. The first sheaf axiom is a consequence of Lemma 5.7.1.
The second sheaf axiom is a consequence of the definitions. D
Remark 1. The significance of the failure of the second sheaf axiom for basic open sets is that we cannot easily describe global sections T(X,^X),
X = Spec(A,^), or more generally
r(D(a),^x),
D(a) C Spec(A,'p)
a basic open set. Lemma 5.7.1 implies that
116
(AS(-a) ^ s ( a ) ) "* (r(D(a) ,0£ $ (D(a) ,^))
is
injective and that However, F(D(a),^Y ) A
'p g, . = A g , . n p ( D ( a ) , ^ ) may be bigger than
is the contracted order.
A c r .. b ^aj
In any case, one shows in
the usual way that there is a natural sheaf isomorphism
Remark 2.
It is perhaps good that
r"(X,/?Y)
is larger than
A
The natural map
(A,'£) -*- (Ac,,>. ,'.pcr,O oil) o^lj
In such case, T ( X , ^ )
Ac ^
is not always injective in (POR) .
may contain information about
(A,'p)
lost in the
localization In the category (PORCK), the second sheaf axiom for basic open sets, Lemma 5.7.2, does allow computation of the global sections of the structure sheaf.
We have
Proposition 5.8.4.
Let
(A/p) e (PORCK),
sections of the structure sheaf
X = Spec(A,'p).
€L over the basic open set
Then the D(a) C x
A
is the ring
F(D(a),^ x) = (A g . •x/^Sr ^ '
is the ring
T ( X , ^ ) = (Ag ( 1 ) ^ s ( 1 ) ) •
In
Particular»
tne
global sections D
The Zariski topology on the prime convex ideal spectrum of a partially ordered ring has the following property not shared by unordered rings. A finite union of basic open sets is still a basic open set. The proof is easy since D(f) = D(f 2 ) and D(IIf?).
U D(f?) = D(E f?), which is symmetrical with
H D(f?) =
If, further, our ring is Noetherian (more generally, if the radical
convex ideals satisfy the ascending chain condition),then every Zariski open set is a basic open set.
117
VI • Polynomials
6.1.
Polynomials as Functions Let
over
A
A.
be a ring, A | X ... X ]
Let
A ^
f e A [ X 1 ... X n ] If
(A/p)
A X , from for all
X
x e X,
1
If
to
(A, ]))
denote affine n-space over
defines a function
f:
A ^
A.
-» A
n-variables
Each polynomial by evaluation.
is a partially ordered ring, then so is the ring of functions A, for any set
X.
f (x) e '$. Note
belongs to
(PORCK)
injective?
Namely, define
A C AX
or
A very natural question is: -* A A
the polynomial ring in
'£x c A X
by
(the constants) and
(PORNN), then so does
f e$
^x
x
if
extends '£.
X
(A , £ x ) .
when is the ring homomorphism
A[X.^ ... X n ]
The following at least gives an easy sufficient condition.
Proposition 6.1.1.
If the ring
A
is
z (n) A f X ^ . .X^]-• A
Z-torsion free, then
is injective.
Proof.
We first establish the result for
a Q + a x X + • • • + adX
is a polynomial such that
in particular, f(0) = a
=0.
Next, let
equations in the coefficients
a.,...,a.
2 ad = 0
118
o
+---+da, = d
Suppose
f(k) = 0
k = l,2,...,d.
a. + a_ 1 2
da.. + d a 1 2
n = 1.
f(X) =
for all We obtain
k € Z C A. d
Since
2 d\ x, • • • x, \ x X 2
det
Xx
2
2
••• Xx d 2
x) n
, xx \ n
xx
2
n
n (x - x ) , l
... x x d n
we deduce that for some non-zero integer
/ i,
m,
ma. = 0 ,
1 j< i <_ d.
Namely,
i, •••, i \ 2
2, 2 , ..-, 2 d
m = det
\ d , d 2 , -.., d d /
Thus
a i = 0, If
1 < i < d , and hence
f(X) = 0.
n > 1, we use induction.
A[x1...xn_1][xn]
C
Corollary 6.1.2.
If
A
is
A
Z-torsion free and
B c A
is any subring,
(n) then the homomorphism
...XJ^A"
A[X, . . . X ] -•• A
is injective.
In particular
is injective.
Corollary 6.1.3.
If
(A/£) e (PORCK), then
A
is
Z-torsion free,
A (n)
hence
A[X X . .. X^-* A
is injective.
From Corollary 6.1.3 we see that if
D (A,'p) G (PORCK), then there is a
natural (PORCK) order on A[X, ... X ] , namely the contraction of the order A (n) y , . on A to A[Xj ... X n ] . We denote this contracted order by the same symbol
(A[XX ... X j , $
Next suppose
(R)
).
(A,£A) C (B,'PB), both in (PORCK).
Suppose as an
119
algebra
B
is finitely generated over
A[T.. ... T ] where
be the polynomial ring in
projection
n-indeterminates, so that
Let
B=A[T,...T ]/I
A[T.. ... T ]
is order preserving, where
extending
p..
p
is the
In particular, I C A[T, ... T ]
'^-absolutely convex.
Proof.
A typical element of
G A[T X ... T n ] .
Remark.
The image in
p B
is is
Ep.g.(T)
In words, the proposition shows that if
pure polynomial extension of
A
(B,p R ).
p
D
(A,'PA) C (B,'pg)
(B,pR)
In the case
a totally ordered integral domain over
That is, first the order
p. G p ,
is obtained from a
on
A
a totally ordered field,
A, this procedure can be reversed.
A[T, ... T ]
can be refined to a total order,
This assertion will be established in Chapter VIII.
We conclude this section with an example.
Consider the ring Z[T],
ordered as a ring of real valued functions on the real line. is a convex ideal. expressions
Let
n + eT,
A
the order on
A.
6.2.
.
A = Z[T]/(2T,T ) .
e = 0 or 1,
vanishes identically on functions on
is
by first dividing by a convex ideal and
then, possibly, refining the order.
inducing
where
Ep.g^(y) e 'PR.
a finitely generated algebraic extension, then
B
... y ] , y. G B.
(A,p A ), (B,pR) G (PORCK) , then the
If
( A p ^ . . . T n ] ,'p) -* (B,'PB)
weakest order on
gi
B = A[y
I = {f e A p ^ ... T j | f iyy . . y j = 0 G B}.
Proposition 6.1.4.
is
A, say
A^
n G Z.
Elements of
A
The polynomial
, so we cannot regard
On the other hand, A[X]
Then
(2T,T )
have unique f(X) = TX+ TX
A[X]
G A[X]
as a ring of A-valued
does admit an order, extending
Adjoining Roots Let
(A,'p) G (POR),
a G A,
We ask if there is an order on
A -• A[T]/(T2-a)
the obvious inclusion.
A[T]/(T - a ) , extending
'p. The answer is
similar to well-known results for (totally) ordered fields, and extends the results of 1.4.
120
Recall from Chapter I the notation
Dp = {aGA|pa G'p
some A.
pG'j) }. We established that - a £ D'£ implies ']>[a] is an order on
We now refine this result. Proposition 6.2.1. (a)
If -a £ Dp, then
B = A[T]/(T2-a)
can be ordered extending
p,
and if B can be ordered extending '£, then '£Ja] is an order on A. (b)
If either
A
has no nilpotent elements or a E A
divisor, then the existence of an order on B to '$ [a] (c)
is not a zero
extending ty is equivalent
being an order on A. If
(A,'£) e (PORNN), '£[a] is an order on A and a 2 zero divisor, then B = A[T]/(T -a) has no nilpotent elements. (d) If A is an integral domain, then either 2 A[T]/(T + a) admits an order extending $.
is not a
A[T]/(T -a) or
Proof. (a) The second statement is obvious since if B can be ordered extending '£, then contains
a = T
is positive in B, hence the contraction of the order to A
£[a]. For the first statement, note that elements of the ring
have unique representatives
b + cT,
b,c e A.
B
If B can't be ordered extending
£, then we have a nontrivial relation
k
i=l
with
2 X
X
X
p i e 'j5 C A. Thus, in A, k k 0 = ( 2 Pi b^) + ( 2 p ^ J a . i=l i=l
(**)
We will prove
k 2 2 p.c. ^ 0 i=l X x
in A.
2 2 Assuming the contrary, then p.c. = 0 and p.b. = 0 , 1 £ j <_ k. Also k 0 = 2 2 p . b . c , b y (*) . On t h e o t h e r h a n d , 0 < p . ( b . + c . ) 2 = p . ( b ? + c ? ) + 2 p b . c . • _ -i
i l l
so the terms
j
2p.b-c.
are all positive.
j
J
J
j
J
j
j
Since their sum vanishes, 2p.b.c. = 0,
121
J
1 £ j £ k. (b)
This contradicts the assumption that (*) is a nontrivial relation. If
'£[a]
is an order and if there is a relation (*) in
then from (**), p.b. = 0, hence
p.b. = 0, if
2
2
Also, p.c.a = 0, hence
p. (b. + c.T)
A
2
A[T]/(T 2 -a),
has no nilpotent elements.
2
= p. (b. -f-c.a) + 2p.b-cT = 0
and relation
(*) is trivial. If we assume that
is an order on A and a is not a zero divisor, 2 2 2 then a relation (*) in B = A[T]/(T -a) implies p.b. = 0 and p.c. = 0, 1 £ j £ H- The argument now proceeds just as in the proof of (a). (c) b
( b + c T ) 2 = b 2 + c 2 a+2bcT = 0 G B, then the assumptions imply
If
= c a = 0, hence (d)
Hence if
'p[a]
If
A
b = ca = 0
and
c = 0.
is an integral domain, then - a £ Dp
a ^ 0, either
admits total refinements, either
or
DjJCA
a £ DJ5.
^ [a]
or
is an order on
A.
Alternatively, since 'p
£[-a]
is an order on
A.
These
two assertions have already been discussed in 1.6. Thus (a) or (b) implies (d). •
Remark.
It does not seem clear whether there is a reasonable
analogue of parts (a) and (b) of the proposition. (A,'£) e (PORCK),
One would hope that if
(*]5+), then one can adjoin a square root of
-a £ D
That is, there ought to exist a
(PORCK)
order on
However, imitating the proof of (a) in the
(PORCK)
a.
A[T]/(T - a ) , extending 'p.
(PORCK)
case does not lead to a
simple relation like (**), but rather to a matrix equation
(***)
k
where
r=
2 i=1
2 2 p. (b. + c.a) i
i
The usefulness of (***)
Proposition 6.2.2. f(T) e A[T] orders on
Proof.
122
i
k
and
s =
Z i=1
2p.b-c., I
I
i
p. e '£, I
ii
is not apparent.
Let
(A,'£) 6 (POR),
A
an integral domain.
an irreducible polynomial of odd degree.
A[T]/(f(T))
p. ,c. ,x,y e A.
Let
Then there exist
extending 'p.
This is a familiar result in the Artin-Schreier theory of
ordered fields. K
First, we assume
be the field of fractions of
we proceed to show
'p A.
K[T]/(f(T))
is a total order on Then
K
A, and we let
is also totally ordered, and
can be ordered, extending the order on
K.
If not, there would be a relation
- 1 = Zg?(T)+h(T)f(T) e K[T]
where degree ( g ^ then
g.^
cannot all be constant, because
h = 0, and we would contradict the fact that
degree than of
< degree (f). The
(Igi(T)) _< 2n-2, and is even, hence
n-2.
h(T)
6.3.
is ordered.
Now,
has odd degree, no greater
By induction we have a contradiction since some irreducible factor
also has odd degree.
•
A Universal Bound on the Roots of Polynomials Let
where
h(T)
K
©
A
be a ring.
For simplicity in what follows, we assume
CCA,
is the field of rational numbers.
Proposition 6.3.1.
f(T) = T 2 n + a T 2 "" 1 + ••• + a
Let
monic polynomial of even degree.
be a
Then there are universal polynomials
3 + (u 1 ... u 2 n ) , 3~(u]L ... u 2 n ) e
€ A[T]
and polynomials
h!(T),
such that
T-3+(a
... a ? j + n
2 ht(T) 2 = £(T) i=l
and
k
3 (a ... a _ ) - T + l zn
Corollary 6.3.2.
If
2 h.(T) 2 = f(T) . i
I i=1
(A/p) e (POR) ,
Q C A
is a monic polynomial of even degree, and if
and
f (a) < 0,
f (T) = T 2 n +
a^2*'1*
a e A, then
Proof.
Substitute
a
for
T
in the polynomial equalities of the
proposition.
•
123
t^T
Corollary 6.3.3.
Same assumptions as above.
~ + ••• + b 2 n - l
a monic
a e A,
be
Let
g(T) = T
polynomial of odd degree.
If
+
g(a) = 0,
then
Moreover, if
a e '•]> and
g(a) <^ 0, then
0 < a < 3 + (b 1 ,...,b 2n _ 1 ,0)
Proof.
Apply Corollary 6.3.2 to
Proof of Proposition 6.3.1.
f(T) = Tg(T).
•
The proof is by induction on
n.
If
n = 1,
f(T) = T 2 + a 1 T + a 2
Thus
and
Assume the proposition for polynomials of degree
2i < 2n.
Let
X = T + (a 1 /2n), and write
f(T) = g(X) = X 2 n + b 2 X 2 n - 2 + b 3 X 2 n - 3 + . . . + b 2 n
where the
b.
are polynomials in the
ai
with rational coefficients.
~±, . n fb2-l^vn-2 hjCX) = X + (^—2—J X
so that
124
Let
h- (X)
= X
+ b2X
and Y 2n ~ 2 +^ bbKXXY 2 n " 3 +^ fu b b
= XX
r^2-l-) 2 W 2 n-4 +Abb, XXY 2n-5
l 4 " ll~2~J J Jx
3
5
Now apply induction to the right-hand side to find
...b J )
Then, since
+
Z h±(X)2 j==2 2 J
2n
3 , h.
such that
= g(X) - h*(X) 2 .
X = T + (a../2n) , we set
hJ(T * 2i) and check the desired formulas.
Remark.
The assumption
ordered ring zation.
If
A
we have
ma > 0
•
J}C A
is not too important.
Z C A, and one can include
implies
a>0,
m G Z ,
Q
In any partially
by a simple locali-
a G A , even this is unnecessary.
For, one can go through the proof above, "clearing denominators". event, one obtains bounds in of
6.4.
f(T), where
m
A
for some integral multiple
ma
In any
of a root
a
depends only on the degree of f(T).
A "Going-Up Theorem" for Semi-Integral Extensions We apply the result of 6.3 to prove an analogue in
theorem in commutative algebra. Then each prime ideal of
A
if the convex prime ideals of B.
A C B,
is the intersection with
If we now assume an order
ideals of
Specifically, let
'pfi C B, extending A
(POR)
A
B
of a well-known integral over
A.
of a prime ideal of B.
'£A C A, it is natural to ask
are intersections of
A
with convex prime
In this generality, the answer is surely no, because one could
have a very weak order on
A
and a strong order on
B.
Thus it is philosophi-
cally sound to make some additional assumption on the orders
'.p. and '£R.
In categories of partially ordered rings, there is perhaps a more natural notion than integral extension, at least for certain purposes.
If
(A,'^A) C (B,'^g
125
we say that
B
is semi-integral over
B.
A
if for all
2n
1
b e B
there is a monic,
+ a 1 T " + • • • + a. , a. e A, with i zn 1 Clearly, integral implies semi-integral.
even degree polynomial in
2n
f(T) = T
We first characterize semi-integral elements in several ways. simplicity, assume
C
For
J}C A.
Proposition 6.4.1. of
f(b) < 0 —
Let
(A/p.) c (C,'Pr) . Then the following subsets
coincide. (i)
{b|a' < b < a"
some
a', a" e A}
(ii)
{b|-a < b < a
some
a e ^A}
some
a e PA>
(iii) (iv) (v)
2
{b|0£b £a 2
{b|f(b ) <_ 0
some monic polynomial
(b|f(b) <_ 0
some monic, even degree
Moreover, if subring of
B C C
f(T) e A[T]} f(T) e A[T]}
is the subset so described, then
B
is a convex
C.
Proof. (i) C (ii)
we have
a' < b < a", then
0 <_ ( a " - % ) 2 , so
-a < b < a
(ii) C (iii)
with
a n < (a") 2 + l. M
2
f
Thus, if
2
a = 1+ (a ) + ( a ) .
if - a < b < a, then
a 2 - b 2 = (a-b) (b+a) _> 0, hence
b 2 < a2.
(iii) C (iv)
(iv) C (v)
If
f(T) = T - a
Obvious,
(v) C (i)
Finally, B
b 2 £ a, let
This is exactly Corollary 6.3.2.
is a subring because (i) is obviously closed under sums and
(iii) is closed under products and negatives.
Also, (iii) (or (ii) or (i))
is clearly convex.
Remark.
The extremely simple characterizations (i), (ii), (iii) of
semi-integral elements make the original definition (v) seem unnatural. However, geometrically, semi-integral extensions are quite analogous to
126
•
the integral extensions of algebraic geometry, as we will see in later chapters. Also, in some situations, it is exactly the characterizations (iv) or (v) which are easiest to verify.
We now turn to our going-up theorem.
The proposition below is surely
capable of some improvement and clarification.
We will present a different
approach for integral domains in the next chapter.
Proposition 6.4.2. (a)
Let
Assume
(A,'£A) c (B,'£B),
Then for any convex prime
c
Q B
A. semi-integral over
A,
fA = A n ^ g .
P C A, there exists a convex prime
Q C B
with
A n Q = P. (b)
Suppose
prime of and
B
}>A/Po
clear.)
C P PQ.
above
coincide on
c A
are convex primes, and
If
'^./P, then
Q C B
P C A If
prime and
A/P Q .
(Only
C B C£ B /Q 0 )
is a convex n
(A/PQ)
$A/PQ = C£ R ^A/P o ) C C^B/QO)
Q, C B
above
P.^
with
is
Q Q C Q1«
is a maximal convex prime, P = Q n A, and
C-Pg/Q)
n
A/P =
is a maximal convex prime.
A C B
Q C B
Q
Assume further that the orders
Then there exists a prime
(c)
(d)
P
is an integral ring extension, P C A
a convex prime above
P, then
Q
a maximal convex
is maximal convex.
Proof. (a)
Let
S = A- P, a multiplicative set.
is a semi-integral extension and
Then
"pR n A^ = '.pA . s S S
(Ac, 'pA ) C (B c , £ R ) b b S S
Thus we may assume
P C A
is a maximal convex ideal. We now claim
/H(PB) c B
is a proper convex ideal.
In fact, yfa(PB)n A = P.
denote the set {y | y G P C A } . To see this, let (for purposes of this proof) P Then vft(W) = {z|0 < z 2 n < Z p ^ ? , j>± e ^ R , y± E P}. Since (B,$
is semi-integral over
(A,'PA), l e t 0 < p . < a . , A
in
£ A = '^B H A .
If
—
1 —
z e A, then convexity of
Finally, with our assumption that it is only necessary to choose
Q C B
P C A
a . G A.
Then
0 < z
P
n
—
1 1
implies
< Ea.y. —
11
z e P.
is a maximal convex prime ideal,
any prime convex ideal containing
127
/H(PB).
(A/PQ, '^A/po)c (B/Qo>'£{
(b) Apply (a) to the semi-integral extension (c) Apply (b) to deduce a contradiction if
P C P C A,
P,
convex
prime. (d) A/P C B/Q field. and
If
is an integral extension and
S = S(l) C A/P
(B/Q) s }> '^R/Q)
is the shadow of
is integral over is a
(A/P) g .
Thus
(A/P, P A /P)
1, then (B/Q) g
is a semi-
(A/P) g
is a field
is a field, hence
semi-field (see Proposition 3.6.1).
D
We conclude with a counter-example to a "going-down" property for semiintegral extensions. Let
Let
R be a real closed field
A = R[X,Y],
B = R[X,Y,Z].
3
S = {(x,y,z) G R 10 <_ x,y,z, z <_ 1, xz <_ y ) . We partially order
a ring of functions on
S, and contract this order to
is certainly a semi-integral extension since
prime ideals
(Y) = P Q C p^ = (X,Y) C A
and
QQ
by a trivial argument.
) C
Then
Q x = (X,Y) C B. QQ C Q1
0 £ XZ £ Y (rel 'pg)
Then
with
did exist, P Q = (Y) C Q Q C Q 1 = (X,Y) But
as
0 £ Z £ 1 (rel '£„). Consider
Q x H A = P 1 ? but there is no 'J>B convex prime Namely, if such a
A.
B
implies
Q Q H A = PQ. gives
(Y)
Q Q = (Y),
is not
'po- convex. The primes
Q1 C B
Geometrically, P the origin and
128
Q,
and
^0>^i
c A
are
convex for '£g, $
corresponds to the (positive) X-axis, P, corresponds to the (positive) Z-axis.
respectively. corresponds to
A potential
Qo C Q
lying over
PQ
would correspond to a 2-dimensional surface in
S
containing some of the Z-axis and intersecting the XY-plane in some of the X-axis.
The only possible such surface is the XZ-plane which has only a
1-dimensional intersection with
S.
129
VII • Ordered fields
7.1.
Basic Results In the first few sections of this chapter, we recall the basic results
of Artin-Schreier on ordered fields.
The last few sections develop the
notion of signed place, which seems very natural in real algebraic geometry. A signed place of a field is a place with values in a totally ordered field and for which we distinguish fashion.
+ °° and
- °° in an arithmetically coherent
In the next chapter, we apply these ideas to establish the basic
results of affine semi-algebraic geometry.
Our organization follows expo-
sitions of Lang [13], [66], up to a point.
In particular, Lang emphasized
the usefulness of real places in real algebraic geometry. To begin, then, recall that a field of squares in
E.
partial orders on
E
is real if
-1
is not a sum
From Chapter I, this is equivalent to the existence of E.
If
ACE
is any ring which has
E
as field of
fractions, then partial orders on E coincide bijectively with derived partial orders on
A
CP ='£<»)•
order refinements.
In fact, if
are refinements of
p.
A real field R
is real.
R
Any partial order
'|J on
a, - a £ p, then both
E
'p[a]
admits total and
is real closed if no proper algebraic extension of
A simple Zorn's lemma argument shows that any real
in a real closed
p[-a]
R, algebraic over
E.
E
embeds
A more precise understanding of
this situation stems from the following, which we established in Chapter VI.
Lemma 7.1.1.
Let
E
be a totally ordered field.
Let
f(x) e E[X]
be an irreducible polynomial, either of odd degree or a quadratic with
a - 4b > 0.
Let
a
be a root of
f.
Then
(hence totally) ordered, extending the order on
130
E(cQ E.
X + aX+b
can be partially D
Proposition 7.1.2.
The following conditions are equivalent on a field
(i)
R
real closed
(ii)
R
real, every element or its negative is a square in
polynomial of odd degree has a root in (iii)
R
not algebraically closed, but
R:
R, and every
R.
R[/^F]
algebraically closed. D
We omit the proof, which can be found in many elementary texts [64], [65]> [66], [67]. The main points are that Lemma 7.1.1 provides the necessity of the conditions in (ii), and Galois theory and the solvability of 2-groups gives the implication (ii) =*• (iii). The following corollaries of 7.1.1 and 7.1.2 are routine, the first two by Zorn's lemma, the third directly from 7.1.2 (ii).
Corollary 7.1.3.
Let
E
be a totally ordered field.
algebraically extended to a real closed field the original order on
E.
Then
E
can be
E, whose order contracts to
That is, each positive element of
E
is a square
in I".
•
Corollary 7.1.4. istic
0.
Let
be an algebraically closed field of character-
Then there are real closed fields
Corollary 7.1.5. Then
K
Let
E
be a real field
If, the algebraic closure of
Remark 7.1.6.
Thus the
R
in
with
E C R,
R
of
R
K = R[/Tj.
E.
D
K = C, the complex
uncountably many pairwise non-isomorphio real IR/with
C
isomorphic (algebraically) to
R
of 7.1.4 is highly non-unique.
Remark 7.1.7.
•
real closed.
R, is a real closure of
Artin and Schreier showed that if
numbers, then there are closed subfields
E
R C K
Among other results of Artin-Schreier theory is the
fantastic characterization of real closed fields as the only fields of finite codimension under their algebraic closure.
We will not use this
result.
131
T
7.2. Function Theoretic Properties of Polynomials If
E
is a totally ordered field, any polynomial
be interpreted as a function
f:
E -*• E.
f(T) G E[T]
can
We will use the usual symbols
(a,b), [a,b], [a,b), (a,b], for open, closed, and half open intervals in We also allow
a = - °° ,
b = + ° ° , with the obvious meaning.
Similarly, it
is clear what me mean by
f
on an interval and by
positive or non-negative on an interval.
a € E , we define
f
|a| € E
E.
(strictly) increasing or (strictly) decreasing
by
|a| = ± a,
|a| >_ 0.
If
The following is proved
in general just as it is for real numbers.
Lemma 7.2.1(a). ordered field.
If
f (T) = a Q T n + a^ 1 1 " 1 + • • • + a R G E[T],
Let x,
b E E,
E
a totally
|x| £ |b|, then
|f(x)| < 2 |a. Mbl""1 . i=0 (b) If a € E
is a root of
f, then n
|a| <_ I |a. |/1a. | . i=0
then
Corollary 7.2.2.
Polynomials are "continuous".
Corollary 7.2.3.
If
f
at
p GE
T = 0
•
is a monic polynomial of odd degree,
is negative on some interval
some interval
If
f(T) £ E[T]
D
x
(- «», a)
and
f
is positive on
(b,«>).
•
is positive, the polynomial
and positive if
T > 1 + p.
T -p
is certainly negative
From this fact, 7.1.2 and 7.2.3 we
obtain the "if" direction in the following characterization of real closed fields.
Proposition 7.2.4.
A totally ordered field
and only if for every polynomial
f(T) E R[T]
f(a) < 0 < f(b), there is a root
c E R
132
of
R
and f
is real closed if a,b G R
between
a
with and
b.
Proof:
The algebraic closure of a real closed
R
is
R[/T].
We
thus can factor
where the
r. E R
are the real roots and the
pairs of non-real roots. b
It is then clear that if
must lie on opposite sides of a real root
Remark: that if
a. _+ $. *^T
r.
are the conjugate
f(a) < 0 < f(b), a
of odd multiplicity
and
m..
•
The existence of this simple factorization also easily implies
f(T) G R[T]
f(T) = gj(T) + g*(T)
is nowhere negative on for suitable
g][ ,g 2
R,
R
real closed, then
e R[T].
Next we consider the geometric behavior of a polynomial function near a point
x = a, over any totally ordered field
f(x)-f(a) =
E.
f(x)
We can write
.,(a) (x-a) 1 .
I
Using the estimates of Lemma 7.2.1, one can verify the following.
Proposition 7.2.5.
Let
f(T) G E[T],
E
m ^ 1
be smallest such that
f ^ C a ) ^ 0,
a £ E.
in
such that:
E
(i) If
m
even, f
(a) > 0, then
(a - £, a + e ) , decreasing on (ii) If
m
odd, f
(a - e, a + e ) , negative on Moreover, on these positive when
f
a, the derivative
We obtain statements for the two cases when f(T)
by
-f(T).
f
(a,a+e).
is increasing on
and positive on
is increasing, negative when
e > 0
is non-negative on
and increasing on
f(x) - f(a)
(a-e,a)
e-intervals around
Then there is
f(x) - f(a)
(a-e,a)
^(a) > 0, then
totally ordered, and let
ff
(a,a+e). of
f
is
is decreasing.
fW(a) < 0
O
by replacing
We draw the four possibilities for local behavior below.
Combining this local behavior of polynomials with the characterization 7.2.4 of real closed fields, we can deduce certain global results about the behavior of polynomials over real closed fields.
133
i,
f ( m ) (a) > 0
m
fW(a) > 0
odd,
f• > 0
f' > 0 /
x= a
< 0
f1 > 0,
m
even,
vf' < 0
f ( m ) (a) < 0
m
odd,
f^(a)
< 0
Figure 7.2.6
Proposition 7.2.7. a < b
in
If
is real closed, f(T) G R [ T ] ,
R
R, then the derivative
More generally, for any
of
f
has a root
f(T) G R[T], there exists
*ir^
Proof:
f'
c,
c G (a,b)
f(a) = f(b) = 0, a < c < b. with
- f(b) -f(a)
The second statement (Mean Value Theorem) follows from the first
(Rolle's Theorem) in the usual way.
To prove Rolle's theorem, one first
assumes
a
f(T)
has no roots between
root closer to
a
if necessary.
and
b, by replacing
By 7.2.3, f
b
by another
does not change sign on
(a,b).
Then a short case-by-case consideration of the local possibilities for near
a
f'(a+e)
and and
b, drawn in Fig. 7.2.6, implies that for suitable f' (b - e)
Corollary 7.2.8.
134
have opposite sign.
Let
R
f
e > 0,
Now 7.2.4 completes the proof. •
be real closed, f(T) e R[T].
Then
f
is
monotonic on any interval [a,b],
closed interval the endpoints
a,b
(a,b) f
not containing roots of
f'.
On any
assumes maximum and minimum values, either at
or at roots of
f'
(a,b).
in
•
7.3. Sturm's Theorem Let
E
be a totally ordered field, f(T) G E[T]
polynomial.
A priori, the number, or even the existence, of roots of
in a real closure of
E
E
any real closed field extending
Proposition 7.3.1. E.
Let
for counting the real roots of
E.
uniqueness of the real closure of
real closures of
f
might depend on the choice of real closure.
theorem provides an algorithm in
over
an irreducible
E
Sturm's f(T)
in
We first show how this implies the
E.
be a totally ordered field,
R,,R~
Then there exists a unique homomorphism
two
R.^ -> R 2
E.
Proof: gives
Uniqueness follows from existence, since, first, symmetry
R 2 •> R,
over
over
E, since any
E.
But
R, •> R,
R,
admits no non-trivial endomorphisms
maps squares to squares, hence preserves order,
but also induces a permutation of the roots in
R,
of any polynomial over
E, which are, of course, finite in number. Sturm's theorem implies that if f(T) G E[T], then E(a) -> R 2 a
o
< a
l
<
over
''' <
a
k
f
has roots in
E.
Let
^"^
a
a £ R-,, with minimal polynomial
R2-
a Q ---a^ G R^
-+i " a i
=
2 "Y'+i"
Thus we obtain algebraic maps be all the roots of The
tneorem of tne
f, and let
primitive
element implies find
E(a ••• a,, y, • • • y v ) = E(3), and by the above, we can O K 1 K E(3) -> R~. This gives by restriction E(a ••• a,) -*- R o , keeping Z O K Z
the roots of
f
in order.
by symmetry, since other hand, if extension
f
There is clearly a unique such map over
has no more roots in
0 < y G E(a • •• ou )
E(3,y) -*• R2> hence
Zorn's lemma now gives
and
R2
y = y
E(a Q ••• a^) -»- R 2
than G R
in
R,.
E,
On the
then we get an
is order preserving.
R, ->• R ?.
•
135
We turn now to Sturm's algorithm. £(T) £ E[T], or more generally, any f 1 = f»
and define
deg(f i _ 1 ).
f^
i >_ 2
f
by
(f,ff) = 1 , we get a sequence
Since
f^, let
(f Q(x) ^
sequence g.-f. - f.
fi ,
of all the
of
f,
[a,a,), (a , b ] . and
x',x"
E [T] , d e g C ^ ) <
£ Q ,f 1 ,...,£ k
a non-zero constant. f..
of Let
For each
x G [a,b],
From the defining equations
f.
of
fj, in the interval
fQ = f,
f^
1
=
is a non-zero constant, we see that if
i >_ 1, then R
e
be the number of sign changes in the
(x) , ... ,f, (x)}.
In some real closure
x' < x"
5(x)
and the fact that
is a root of
and on
f,
be two points which are not roots of any
not a root of any
a
(f,f!) = 1, let
with
f±_2 = g±_1^i_1 - f ±
polynomials o£ decreasing degree with a,b €= E
Beginning with an irreducible
(a)f.+1 (a) < 0. ct-^ < • • • < a n
E, let (a,b).
Then
6(x)
is constant on 6(x f )
We ask the relation between
are separated by a single
f = f Q , the sequences defining
f
6(x )
a..
and
be all the roots
If M
6(x )
a.
(ou,a-
),
6(x M ), if
and
is not a root
may differ at
{... f.^Cx'^f.Cx'^f^Cx') ...} and {... f^Cx'O.fiCx'^f^Cx") ...}. But even if
f^
changes sign at its root
a., so that
f^(x')
and
f i(x")
have opposite sign, one checks that the number of sign changes, reading across the two sequences, is invariant, 6(x') = 6(x"). is a root of
f = f Q , then the irreducibility of
sign across f
a.
but
f
does not.
Comparing
M
On the other hand, if f
implies xl
{fo(x' ),f1(x ),---}, we see, in fact, 6(x ) - 6(x") = 1.
{+>+t •••}•)
changes
x
{f Q ( )>?iC ')»•••}
f
7.2.6 shows the two sign sequences are
f
with
(The local picture
{+,-,•••}, {-,-,•••} or
{-,+,•••},
We conclude then, with the notation as above:
Proposition 7.3.2.
The number of roots of
hence is independent of the real closure
Since we can first compute all the all roots of all
fi
by
R
of
f
in
(a,b)
E
chosen.
f., and then choose
is
6(a) - 6(b), D
(a,b)
We state a consequence of 7.3.2.
Proposition 7.3.3.
containing
7.2.1(b), we can count all the real roots of
by this algorithm.
136
a.
For each
n >^ 1, there is a finite number of
f
systems of finitely many inequalities g.ji
e
sucn tnat
Qt^i'•'"tjj]* n
given any totally ordered field
n
f(T) = T + a 1 T " + • • • + a n G E[T] roots), then systems
Si
f
E
with
E
(or, more generally, f(T)
has real roots over
holds in
S. = {g..(t-...t ) > 0} , with
E
and irreducible with no multiple
if and only if at least one of the
t^ = a^, that is, for some
i, all
g..(a ...a ) > 0.
•
Remark 7.3.4. set of systems
The Sturm algorithm itself would give a rather inefficient
S^.
Of course, 7.3.3 could be improved to a statement con-
cerning the precise number of real roots in a given interval
(a,b), again
using nothing but the Sturm algorithm. Despite the totally elementary character of the proof of Sturm's theorem, (which is basically no harder than the statement itself), it seems reasonable to believe that added insight into 7.3.3 could be derived from Galois theory. The coefficients
a.^ of
f(T) G E[T]
are the elementary symmetric functions
of all the roots in
R[/-T], and somehow the signs in
combinations of the
a.^ should determine not only the real roots and their
E
of various rational
location, but also the location of the real and imaginary parts of the complex roots and other information about real subfields of the splitting field of
f.
The determination of whether a quadratic has real roots is trivial: T +aT+b
has real roots if
r
d = a - 4b > 0.
d = -(4p + 27q ) , we have one or three real roots.
d < 0
or
d > 0
T + p T + q, with
accordingly as the cubic has
It turns out that a quartic
two real roots if the discriminat p < 0
For the cubic
2
d < 0
T + p T + qT+ r
and four real roots if
has
d > 0,
8pr- 2p 3 - 9q 2 > 0.
and
7.4. Dedekind Cuts; Archimedean and Non-Archimedean Extensions Let
E
such that if and if D
be a totally ordered field. b G D,
a < b, then
D = E, we write
is rational if
D
D = + <» .
a G D.
By a out we mean a subset If
D = 0, we write
D = - °° ,
All other cuts are called finite.
has a maximal element or
E-D
D C E
A cut
has a minimal element
137
b G E.
We write
D «* b
two cuts rational at f
D » D
E C F
a e F
D
is rational at
b, one contains
if the two cuts
Let Each
if
D, D
1
b
the other does not.
D
of
E
D E(a)
of
E
algebraic if
a €= E, the real closure of
Proposition 7.4.1. ordered fields
E.
We write
differ by at most a single element of F
E.
extending
E.
by
D E (a) = {a e E|a < a
(i) If
Note there are exactly
be totally ordered fields, the order on
defines a cut
We call a cut
b.
in
F} .
D = +_ °° or if
D »
D
E (°0
for some
All other cuts are called transcendental.
The following are equivalent for an extension of
E C F:
a G F
(ii) D E (a)
and
Dg(a)
is rational, then
is finite for all
(iii) For all
e > 0
Proof: (i) =* (ii)
in
F
a € E.
a 6 F.
there exists
Suppose some
e > 0
in
D £ (3) = +_ °° •
E
Then
with
e > e.
3 £ E, but
D E (l/3) » 0. (ii) =* (iii) elements of
E.
Suppose
Then
(iii) =* (i)
e > 0
in
F
was smaller than all positive
D (1/e) = + °° . D E (a) » a G E, a £ E.
Suppose
is positive, but is smaller than all positives of conclusion holds for
If
a < a, then
E.
If
I/a- a
a < a, the same
I/a-a.
D
Extensions satisfying the conditions of 7.4.1 will be called Archimedean.
Lemma 7.4.2. (a) If over (b) If
E C F
and
F
is algebraic over
E, then
F
is Archimedean
E. E C F
is Archimedean and
F C K
is Archimedean, then
E C K
is Archimedean.
Proof:
138
The first results follows from the
bounds on algebraic quantities
given in 7.2.1.
The second result is clear from any of the conditions in 7.4.1.
Remark 7.4.3.
© C IR, the real numbers over the rationals, contains all
Archimedean extensions of
Q
and
IR admits-no proper Archimedean extension.
IR is unique with this latter property. properly containing Dedekind cuts of Q
E
E
is a totally ordered field
IR, then one cannot make sense out of arithmetic of
E.
For example, the smallest cut
cannot be distinguished from
sequences in
If
2D, whatever
2D
D
of
means.
E
containing
Also, Cauchy
are not generally interesting, for example, if
E
has
uncountable transcendence degree over IR.
Remark 7.4.4. over
Although the real closure of
E, it does not follow that
E
is order dense in
certain topological reasoning precarious. of
real closure of
IR(t).
F.
E.
This fact makes
R(t)[X]
which are strictly
IR(t), yet which have roots and change sign over the (See 8.2 for specific examples.)
We now point out how an extension valuation ring in
E, is Archimedean
For example, given any ordering
IR(t), there are irreducible polynomials in
positive as functions on
E, say
E C F
of ordered fields defines a
This construction and its uses goes back to Krull [12].
We define A E = {a e F|DE(a)
If
3 £ A £ , then
is finite} .
D E (3) = +_ °° , hence
D E(l/3) » 0, so
A£
is a valuation
ring.
Proposition 7.4.5. maximal ideal is
residue field over
A£
is a totally ordered valuation ring in
Q c = {a £ F|Dc(a) » 0}, and E E
A £ = A £ /Q E
Qc E
F.
is a convex ideal.
is totally ordered, contains
E
The
The
and is Archimedean
E.
D
This result is checked routinely, so we omit the details. Note that given over
E.
If
E C F, we have
E C K C F
and
K
A£ = F
if and only if
is Archimedean over
E, then
F
is Archimedean
A n = kv. E
Lemma implies that any
K C F, Archimedean over
Zorn's
K
E, is contained in a subfield
139
•
maximal with this property.
We say
maximal Archimedean over itself in F
implies
E
in
F.
F
If
E
In particular, E
and
A£
F
if
E
is
Archimedean closed in
F.
is Archimedean closed in
is algebraic over
Ap/Qp = A p
closed in
is Archimedean closed in
algebraically closed in
Proposition 7.4.6. field
E
E.
F, then the residue
Conversely, if
is algebraic over
E, then
E
E
is algebraically
is Archimedean closed
F.
Proof: over
E.
Suppose
Then
t £ Ap
ECE[t]
is Archimedean over
and the image
and
t
in
Ap
Q E n E[t] = (0), hence
is transcendental E(t) C A £ .
Thus
E(t)
E, contradiction.
Conversely, if
t €= F - E
Q £ fl E(t) = (0), hence
A£
and
E(t)
is Archimedean over
E, then
contains the transcendental extension
E(t).
D
7.5. Orders on Simple Field Extensions We will classify all total orders on a simple field extension E(x), extending a fixed total order on If E
and
x
is algebraic over
E(x)
E.
Let
E(x)
in
f(t)
not separated by any element of
E
E(x)
However, let
E(x) if
E(x 1 )
and
in
and
(see 7.4.4).
be two roots of
E(x 2 )
coincide in
in
R.
E,
Dg(x).
R, we don't distinguish the orders on x
to
x.^
x 2 , respectively.
and
R
over
However, E
E, by the proof of
But this is impossible, by the same proof.
has a real closure, which contains
Thus orderings on
140
R
Of course, if the two
We now turn to orders on the transcendental extension
way.
f(t). This is
are distinct, any order isomorphism between them over
uniqueness of real closure.
E(t)
correspond
Thus, there may exist several
f (t)
would extend to a non-identical endomorphism of
ordering on
E(x)
may have two roots (or more) in
*pX2
E(x 2 )
The orderings on
R, which define the same cut of
induced by the two maps taking E(x 1 )
R.
E.
f(t) € E[t], then
R, that is, to the real roots of
not completely obvious, because
embeddings of
be the real closure of
E, with minimal polynomial
have the same real closure
to embeddings of
fields
R
E(t)
E(t). Any
R(t), in a natural
are seen to correspond bijectively to orderings
on
R(t). Recall total orderings on
Proposition 7.5.1.
R(t)
A total order
correspond to orderings of R[t].
p
on
R[t], extending
D = D R (t)
closed, is uniquely determined by the cut
of
R.
R,
R
real
The possibilities
are: Case (i)
D
infinite.
If
D = oo , ']5 = ' ^ = {f G R[t] |f((b,~)) > 0, some
If
D = - oo, p = $_m=
Case (ii)
D
rational, say
p = -p
If
Then:
a £ D
^= ^a Case (iii)
=
D
•p = p
{£ G R[t] | f ((- oo,b)) > 0, some
D *» a G R.
in
aGD
^f G R^ lf ^ a " e » a^ > °> some
transcendental.
Moreover, R(t)
b G R} .
(that is, a < t
e > 0
in
in
R(t)), then
R} .
R(t)), then
e>
°
in R
-^ *
Then
= {f G R[t] | there exists
Proof:
If
= {f G R[t] |f ((a,a + e))> 0, some
(that is, t < a
bGR}
is Archimedean over
R
a,b G R, a < D < b, such that
if and only if
First one checks the orders
^m, '$ „, '-P
D
+,
f((a,b)) > 0} .
is transcendental
'-P
, '^D
are, in
fact, total orders on R(t). Secondly, the case with
D
infinite is dealt with by replacing
t
by
1/t,
D(l/t) » 0, hence rational. Thirdly, suppose
D » a, as in case (ii), with
which we may assume monic, factor
f(t)
where
a. ± /-T 3-
(Recall
r.
are the real roots and R(/TF)
if and only if
over
is the algebraic closure of r. < a 1 —
in
R, if and only if
R.
a < t.
Given
f(t) G R[tJ,
Thus
are the non-real roots. R.)
Obviously, (t-r.) G £
(t - r.) G '$ 1
.
Thus
'p = '£
3-,+
3.,+
141
as asserted.
If
if and only if
t < a
in
'£, then
(t-r.) G'.]}
a,—
We sketch a second proof that
'$ = '$
if
+
a < t,
D = D R (t) » a.
f(t) = f(a) + f ^ (a)(t-a) + ••• + (f^(a)/n!)(t-a) n =
Begin by writing
(f^(a)/m!)(t-a) m (l + terms in (t-a)), for some In the abstract order
'^ on
small relative to
Thus the 'p-sign of
f^
R.
R[t], t - a
m > 0, with
f^
is positive, but infinitesimally f(t)
'£ = '$
, as asserted.
The case
is simply the sign in
t < a
We point out that the replacement of
R
t
is analogous.
by
1/t, converting behavior
of functions "at +_ «>" into behavior "to the right or left of formulated in terms of coefficients. n
a Q t + • • • + a n , then then
f 6'|5 T
if
n
f G ^
is even, a
> 0
odd,
D = D R (t)
and
closer to
D
f
R(t), then the polynomial some real field containing between
a
f
(a,b)
D(t) = - oo ,
One simply
R. and
By Sturm's algorithm, f
(t-ri) G $
r i G D, if and only if
A R C R(t)
in the case
R.
so we may assume
t G AR.
f.)
If
D
f(t)
QR C AR
is a non-trivial, convex prime ideal of D = D D (t) » a.
had
'£
on
and
b
Thus
in
'p = '$~.
transcendental can be
(t-r^ G $
Obviously,
D
-
be the valuation ring associated to
Conversely, assume If
a
(Simply
would therefore have
II (t-r.)((t-a.)2+ 3?). x i,j 3 ^
is infinite or rational, then
Archimedean over
(a,b).
b, contradicting our choices.
f(t) =
Finally, let
f G R[t], choose
would have a real root between
based on the factorization
if and only if
For each
in the abstract order
'p = '$
142
If
r
has constant sign on
A simpler proof that
a G R, hence
< 0.
than all the roots of
the opposite sign as its values on
D = D R (t)
R.
O
is transcendental.
choose
b
a
in
t .
such that the function
R
n
aQ > 0
f(t) =
n
a < D < b
a root in
or
0" is easily
D(t) = + oo , and
O
Next, suppose
a
Namely, if
if and only if
-oo
divides through by
some
(a) ^ 0.
(a). By the discussion of local behavior of functions of 7.2, we
deduce
If
r. < a,
(t-r.) G'p 1
of
if and only if
A R ± R(t)
and
A R f R(t). Either
R(t) t
or
is the maximal ideal, then
R C R(t). is non1/t G A R , Q R fl R[t]
R[t]. Necessarily, then, Q R n R[t] = ((t-a]), D
Corollary 7.5.2. Suppose
§!>•••»§!<
e
R
Let t
t ]
R(t)
are
be partially ordered, R
finitely many elements positive in the order.
Then there is an interval of values
Proof:
real closed.
(a,b)
in
R
with
gi((a,b)) > 0, 1 £ i £ k.
Extend the partial order to a total order and consider the
classification of total orders directly.
Corollary 7.5.3. Suppose on
E,
E
E
is order dense in
f(t) G E[t].
Proof:
Then
be totally ordered with real closure R
f(t)
and
f(t) < 0.
f(t)
f
is non-negative as a function
2 2 h f = Zg.
By 7.5.2, f(t)
(a,b) C R, hence
R.
is a sum of squares in E(t).
If there is no equation
be ordered with interval
Let
•
in
Eft], then
Eft]
can
is strictly negative on some
is negative at some point of
E, which
contradicts our assumption.
Remark. f(t)
•
It is known that if
f(t)
is actually a sum of squares in
is a sum of squares in
E(t), then
Eft]. See [ 4 ] , [ 5 ] , [ 6 ] for
detailed discussions of this point.
Corollary 7.5.4. (a) Let
E
be totally ordered with real closure
is ordered, non-Archimedean over residue field
A = A£/QE
R, and suppose
E, with valuation ring
A p C E(t).
is then a simple algebraic extension of
E(t) The
E,
A = E(a). (b) If in
t 6 A E , let
E(t). Then
Eft]
HEft],
g(t)€E[t]
with degree
g i (t) < degree
and only if
g(a) > 0
Proof:
f(t)
A = E[t]/(f(t)), and the ordering on
corresponds to a choice of a root of (c) If
Eft]
(f(t)) = Q
in
and
irreducible, A
f(t) > 0
induced by that of
f(t), a £ R.
g(t) = f(t)m (g Q(t) +g x (t)f(t) + ••• + g k (t)f k (t)),
f(t), g Q (t) i 0, then
g(t) > 0
in
E(t)
if
R.
Assertion (b) implies (a) since either
is a principal ideal domain, the localized ring
t
or
1/t £ A £ .
Eft].^^..
Since
is the
143
valuation ring
Ag, and
Q E = (f (t))E[t] , f , _ . .
This proves (b). Finally,
(c) is more or less obvious.
•
7.6. Total Orders and Signed Places Just as valuation rings in fields correspond to places, we will show that totally ordered valuation rings with convex maximal ideal correspond to a special type of place, in which we distinguish place on a field
F
+ °° and
immediately yields a total order of
totally ordered valuation ring in
- °° .
A signed
F, in fact, a
F, with convex maximal ideal, and con-
versely, such a valuation ring yields an equivalence class of signed places. Let to
A
A
be a totally ordered field.
We adjoin symbols
- ©o
and extend the operations of addition, subtraction, multiplication,
and division between
°°, - <» and elements
then justify the notation
Addition
a E A
- °°.)
00+00=00 (_ 00) + (_ 00) = _ 00 a
+ 00 = 00
a + (- 00) = _ 00
Subtraction
00 - (- 00) = 0 0 r_ 00} - 00 = - 00
a
- 00 = _ 00
a - ( - 00) = ° o
Multiplication
°° • °° = (- °°)• (- °°) = °° 00 • (_ 00) = _ 00
ifi
a > 0
I - 00 ifi
a < 0
r
a • (-00)
144
oo and
oo 00
r-00 = j u 00
if
a > 0
if
a < 0
as follows.
(These rules
Division
a/°o = a/-00= 0
o°/a = (l/a)« 00
0* ( ± ° ° ) , ± °°/0,
the four possibilities
a ^ 0
if
a / 0
(l/a)-(- 00)
- 00/a =
The symbols
if
00 + (_ 00),
00 - 00 ,
± 00/+ 00 are undefined.
(- 00) - (- 00)
and all of
Of course, we want the usual
commutative and associative rules to hold, in all defined expressions. that we do not distinguish the signs of both
0. If
a/- °°, a ^ 0, but call
These elements can be distinguished by "inverting". K
function and
a/°° and
is a field by a signed place, with the values in A, we mean a
p:
K->A, ± °°, such that
p(x+y) = p(x)+p(y),
p(l) = 1, whenever the terms are all defined.
guarantees
p ^ 0
and
the ring
A
Let
p:
= {x G K|p(x) G A}
The last condition simply
K -• A, ± 00 be a signed place on is a valuation ring.
Moreover, K
ordered if the strictly positive elements are defined to be p(l/x) = °° , or
p(x) > 0
the maximal ideal
Proof: p:
Q
in
C A
A}.
If the order on
products
K
x, -x G p . xy
and
becomes a place on
K.
It remains to show that
x + y, for
The possibilities for
Then
is totally
'£ = {xEK|p(x)= °°,
is restricted to
First, if we simply suppress the distinction between K -* A, 00
K.
A ,
is convex.
Thus
Secondly it is clear that for all non-zero one of
p(xy) = p(x)p(y),
p ^ °° .
Proposition 7.6.1.
- °° ,
Note
A
is a valuation ring.
x G K,
'£
+ °° and
x
G '.p+, and exactly
is closed under sums and
x,y G ']> .
p(x), p(y)
are covered by (i) p(x),p(y) > 0,
(ii) p(x) = 00 , p (y) > 0, (iii) p(i) = °° , p(y) > 0, (iv) p(x) = °o , p(y) = 00 , (v) p(-) = 00 , p(i) = 00 , (vi) p(x) = 00 , p(I) =00. it is trivial to check directly that evaluate either p(xy) ^ 0.
If
p(xy)
or
p(xy) < 0
This contradiction shows
p(l/xy) or
xy G'|) . directly.
In case (vi), one cannot By symmetry, we may assume
p(xy) = - «> , then
p(xy) > 0
or
i n all cases except (vi)
p(x) = p(xy y) = p(xy)p(-) =
p(xy) = °° .
The six cases for sums
145
are all trivial except (v). But
x + y = xy(— + — ) , x y by product cases, and case (iv) for sums. Finally, Q homomorphism
C A
A
is convex as the kernel of an order preserving
-*• A.
A signed place
•
p:
K -• A, ± °° determines a total order on
given a valuation ring ideal
Specifically, if
in a totally ordered field
1/x £ Q, then we set
and
p(x) = - °° if
and
p(xy) = p(x)p(y)
with in
A
Q C A, one recovers a signed place
that since
A
x
is negative.
This shows that if
Verification of
ap^
Here
valuation ring of
and
p^:
p i (K)
p..
valuation rings in
K -• A^, ± °° ,
Remark 7.6.2.
g:
K
P-^A
A
is all of
K.
maximal ideal, one sets becomes the "addition" in
of
K,
K
G Q.
Thus
) C A ^ where A
C K is the
and totally ordered A signed place
p
of
K
K.
correspond to a third concept, that of
A valuation is a function
If
A C K
T = K*/A*, T
and
A - (0) = {x G K|v(x) :> 0}
A* = A - Q. 1 E K*
and
v:
K* -* V, where
K* = K - (0), with
is a valuation ring with
The natural projection
valuation, v(x) = v(-x).
146
0 <_ 1 £ 1 + (y/x)
In commutative algebra, equivalence classes of places on
and valuation rings in
y/x E A.
K
with convex maximal ideal.
v(x+y) >_ min(v(x) ,v(y)).
if
1
p-^ (K), ± ° ° ^ p 2 ( K ) , ± °° such
a totally ordered (additive) abelian group and
x <_ y
x + y = x(l + (y/x))
We then see that there is a bijective correspondence
a (Krull) valuation of
and
Note
i = 1,2, are equivalent if
means the subfield
is trivial if its valuation ring
K
K
p(x+y) = p(x) +p(y)
l/(x+y) = (1/x) (1 + (y/x))"
between equivalence classes of signed places of
a field
is positive in
p(x) = °° , p(y) = °° , then
there is an order preserving isomorphism p2 =
x
A = A/Q.
This is perhaps the trickiest verification.
We say signed places
that
K -* A, ± °° where
p(x) = °° if
Conversely,
consists of a routine case-by-case discussion.
1 + (y/x) £ Q
p(x+y) = °° .
p:
K.
K, with convex maximal
is a valuation ring, we may always write
y/x E A.
A, hence
so this case is covered
"0" 6 F .
v: K* -*• V
Q CA
We order
the
K* T
by
is then a valuation
Q - (0) = {x € K|v(x) > 0}.
is
v(xy) =v(x)+v(y)
The multiplication in
becomes
T
For any
Now, if
K
is totally ordered and
A C K
maximal ideal, then the sign homomorphism a(x) = + 1 to
v:
if
x
K* -> T
the
x.
hence
v(x-) > 0.
x^ positive and
a(x i ) = + 1
and
v(Ex i ) > 0, then
x
is negative, is related
v(x i ) >_ 0,
If
a(x i ) = + 1, and
x- € Q.
v(x i ) > 0
In fact, we can state if
for any
But then
1 <_ E(x./x )
in
x i G K. Q, then
Namely, if v(l/x ) >^ 0,
A, which contradicts
Q.
cr(-l) = - 1 ,
K, and a sign homomorphism
cr(xy) = a(x)a(y)
one recovers a total order on is a valuation such that then the valuation ring such a pair
if
x , did not belong to
Conversely, given a field with
a(x) = - 1
Ex. £ Q, then
v(E(x./x )) > 0.
K* -»• {± 1}, defined by
Translating, this condition simply says if
of minimum value, say
convexity of
a:
in the following manner.
v(Ex-) > 0, then x- £ A,
is positive and
is a valuation ring with convex
(v,a)
the elements
K
and
by
$
cf(x-) = + 1 A
of
v
a(x+y) = 1
+
a(x) = a(y) = 1,
and
v(Ex-) > 0
implies
has convex maximal ideal K, since
not distinguished by
Although the residue fields
K* -* {± 1}
= {x £ K* |o*(x) = + 1}.
a split valuation on
+ x, - x
if
a:
A/Q = A
a
Q.
If
v:
K* •> r
v(x.) > 0, We call
serves to distinguish
v.
and the domain
K
of such valuations
are certainly restricted (they are totally ordered), no general restrictions are implied on the value group
T
of a split valuation.
We have seen in 7.4 that if subfield, then pE:
place and
E1
is
E
Archimedean over pp
Let, then, p:
A = A £ /Q E . E.
p| p
is a
A E C K, hence a signed
Moreover, p E = p £ l if
ECE1 C K p
We may assume
Consider the family of subfields is trivial.
For example, Q C K
p
E C K is such a
E.
Proposition 7.6.3. (a) A
K
is
Zorn's lemma implies that there exist maximal such subfields
and we fix such an
on
E C K.
K •> A, ± °° be a signed place.
such that the restriction
E C K
We show now that any signed place
for some subfield
surjective, that is, p(A ) = A.
subfield.
is totally ordered and
defines an (ordered) valuation ring
K -+ A, ± °° where
is equivalent to
K
is algebraic over its subfield
p(E) C A.
147
E C K,
(b) The place (c) E
p
is equivalent to
is Archimedean closed in
pE<
K.
Proof: (a) Let
t £ A - E, so
p(t) £ A.
homomorphism by the choice of (b) We prove
= A£.
A
the induced order on p(E) C A.
If
E, hence
p:
p(t)
The signed place
x £ A , then
p(x) £ A
for some
for some
of
PE = p
p:
K -> A, ± °° orders
is algebraic over
e E E. Thus, x €= A £ .
e *= E, hence
and
!
E , hence
1
E
in
A
and
x £ A£, xE A .
C K were Archimedean over
E = E', by the maximality
E.
•
is a signed place on to the order on
K
K
Suppose
A
and
L ->• K, ± °° is a signed place on
q:
induced by
p.
is totally ordered, p:
Then
L, which induces the same order on
po q: L
as
a signed place on
K
and surjective, then L.
(If
f
q:
q (a) G A
CK
L
qf:
K.
•
L -*• K, ± °° is q1
If
is non-trivial
L -> A, ± °° will never be a signed place
"changes sign" and
q' (b) = <» , then
Nonetheless, if we suppress the distinction between is a place on
L, relative
L •> K, ± oo .
relative to some new order on r = poq':
K -* A, ± °°
L -> A, ± °° is a signed place
The proof is routine. On the other hand, suppose
on
K
p(E) by (a), hence
Conversely, if
p(-e) < p(x) < p(e)
would be trivial on
Proposition 7.6.4.
on
is a non-trivial
is algebraic over p(E).
(c) This is a consequence of (b) since if E, then
E[t] -> A
E C K of course coincides with the induced order on
p(-e) < p(x) < p(e) -e < x < e
Then
+ °° and
and we will see in the next section that
L
r(ab) t r(a)r(b).)
- °° , po q1 : L-*-A,°° can be ordered
so as to convert any real valued place into a signed place.
7.7. Existence of Signed Places We have seen in previous sections how signed places of a field closely with total orders on
K
together with a subfield
E C K.
K
In this section
we show how, beginning with less data, signed places can be constructed.
148
correspond
The
first main result, which is due to Krull [12] (and in a different form to R. Baer, according to Krull) is that if of
K
and
A
refined to a signed place, p : (K,:p)
K •> A, °° is an ordinary place
is a real field, then given any total ordering of f
if
p:
K •*- A, ± °° .
is a partially ordered field, A C K
a subring, P C A
p:
on
(A ,Qp)
(that is, A C A , Q
and maximal ideal of to refine
'p
ideals of tf w
n
A
A
that is,
A = A/Q
Let
p:
Suppose are
A
is a real field.
£ w n A - convex, where
Suppose
Ex. = 0,
x i /x Q ^ A.
1 + 2 p(x./x ) 1 i^O originally. Thus
Then
in
say
x. € K. 2
Since
Q.
G Q.
Finally, let
If some Thus
'£
Let
K
with valuation
is real and all prime
is the weak order on
x
K
(so
have minimum value for
= 1 + 2 (x^x ) i^O
2
=0
in
p,
A, hence
0 =
x. = 0
be a prime ideal of
P, that is, p A, p . = P.
first part of the proof, A p
A. p . C K
Let
H A - convex in
A.
P
f
n A
= 0 €= A
is a contradiction.
A standard result of valuation
is a valuation ring with maximal The
pair
A p , with residue field
is a real field.
is
A.
A p = A. p ./P.
with valuation ring
part of the proof that
x^ €= Q, it follows that
( 1 / X Q ) 2( E X ? ) = E(xi/xQ)2 =
But, again, 1 + Ep(x./xQ)
P C Q
K
If we show all
l/xQ 6 A , hence
valuation ring and maximal ideal in
$
Then
K
x. £ Q, then the term of minimum value,
theory is that the localized ring
of the place of
will be Archimedean over the
is real, we must have had all
A
is real. 2 Ex. £ Q, x. €= K.
x Q , is not in
ideal
can be arranged
of the weak order on A ) ,
E(x./x 0 )
A.
'j$ n A - convex.
1 + 2 (x^/x ) i^O
K
K
Next suppose is
P
is the valuation ring
K -»• A, °° be a place of a field
1
Q
'$ n A-convex
K -* A, ± °° with center
p ) . Moreover, the induced order on
= ("B ) , the derived order w ^
Proof:
a
A/P.
Lemma 7.7.1. (A,Q).
where
and the residue field
fraction field of
ring
H A = P
can be
The second main result is that
prime ideal, then there exist signed places A
A, p
Since
A
(A/P, Q/P) A.
is a
Thus, by the
is the residue field
(A. p .,P), we deduce from the second -convex in
A r D . , hence is also
(In other terms, we are using the specialization chain
K -> A , °°+ A, °° corresponding to the pair of valuation rings
A Ck,^
149
in
K.
But see the Remark following the proof of the next proposition.)
Proposition 7.7.2. (Krull) with valuation ring
(A,Q).
p H A - convex ideal, and
Let
Let
']) C K
$" = 'p/Q
be any total order refinement of p! 3 $
refinements
on
K
Moreover, all prime ideals
Proof:
Let
Q~
pf:
and by
A = A/Q
such that P CA
a:
Let
K,
Q
a
p"' D f
Then there exist total order 'p1 n A - convex
is
p'/Q=Tf-
and
p' n A - convex.
1/x G K
K* -> {± 1}, then refine
p:
with
x G Q.
K -> A, °° to a
'JT1 C A, by p 1 (x) = p(x), x E A
K •> A, ± °° , relative to The order
f
CK
will be the order induced
p'. A*
denote
x = yz, with
product
A - Q, the units in
z E A*.
x-j^-.-'x, ,
A subset
x^ G s,
possibly
a:
then there is a relation xi
define
occur.)
Define
a(y) G {± i}
are fovoed on us if orders on
p
x,y G K*
are associates
'p -independent if no finite
'p - independent subsets
SCR*
although
S H A* = 0.
S •> {± 1}
a(z) £ {± 1}
be an arbitrary function. x. G S , p £ j ) ,
to be the sign of
z
If
y G K* - S,
z G A*.
(Possibly
(A, p"1)
in
and
cr(y)a(x1) • •-a(xk) = a(z). The point is, these choices is to refine
p
and if we are to have
p'/Q = 'p!
as
A.
We first check that yx| .. • x^ = p'z an associate of choice of
S.
1
with
a:
Thus
K* •* {± 1}
x! G S,
pp'/y
and it follows that
p. !
verify that
p'
Verification of
Gp,
is well defined. z
1
G A * , then
This can occur only if f
pz = p z , cr(y)
p
1
f
z = (p'/p)z ,
If also x-^ • -\*[- • - ^
{x^} = {x!}
(p'/p) G A * , hence
pf:
a(z)=a(z'),
K •> A, ± °° and we only need
is a signed place, relative to the total order f
p ' O y ) = p (x)p'(y)
is
by our
is independent of the choices made.
At this point we have a function
150
is
v x ^ — x ^ = pz, with
by 1
We say
x^ ^ x., is an associate of an element of 'p.
S = 0, and certainly
In any event, let
A.
S C K*
By Zorn's lemma, there exist maximal
no
K, with
the induced order.
A. Q
are
be a place of a field
be a partial order on
p" on
p 1 (y) = a(y) • °°, y E Q" 1 .
Let if
c
K -> A, «
denote the subset of elements
We will define a function signed place
p:
•
'p' on
is quite simple and verification of
A.
p'(x+y) = p 1 (x) +p f (y) Finally, let
is also easy if use is made of the identity
P CA
signed place
p : K -> A, ± °° to
The positive elements in in
A. p . C K.
A p = A, p ./P.
Restriction of the
induces a signed place
A p -* A, ± °° .
are clearly the images of the positive elements P c A, p >.
|5f n A. p . -convex.
is
•
The last paragraph in the proof can be established more directly.
In fact, if
(A,Q)
f = '£d
Q
and
is a valuation ring, '£ C A
-^-convex, then any ideal
Namely, suppose is
Ap
A, p . C K
But this just says
Remark:
Q
be a prime ideal,
1
x+y = x ( l + — ) .
0 <^ x <_ y,
:p-convex,
y/x £ Q.
yz £ I. Thus
Since
is any partial order with
I CA
is '^-absolutely convex.
£ = .p^,
x/y £ A, hence
0 <_ 1 <_ y/x, and since
x = y(x/y)
and
xz = y(x/y)z £ I.
Example:
Let
E
be a totally ordered field, We obtain the f-adic place
irreducible polynomial. A = E[x]/(f(x)). If
'£ = p
If
f(x)
has real roots over
is the weak order on
of a single element each ordering of
K = E(x), f(x) £ E[x]
{f(x)}.
K, a maximal
p:
K •> A, °° , where
E, we obtain orderings of
A.
-p-independent set will consist
Thus there are exactly two orderings on
A, determined by whether
an
f(x) £ K
K
over
is made positive or negative.
We turn now to our second main existence theorem, the construction of signed places on a field
Lemma 7.7.3. P C A
a prime
'£' D'$
on
K
Proof:
Let
K
with preassigned center on a subring
(K, £)
be a partially ordered field, A C K
'£ n A - convex ideal. such that
P
A C K.
is
a subring
Then there exist total order refinements
tp1 H A - c o n v e x .
We use the basic result of
A.
Klapper, Proposition 2.7.1, that
if a prime ideal is convex for an intersection of two orders, it is convex for one of them separately. p[x]
and
since if y £•£.
'p[-x]
x £ K,
are (derived) orders on !
!
y = p + qx = p -q x, Thus our prime ideal
^P[-x] n A-convex.
If
x, -x £ p , then both simple refinements K.
Moreover,
P>p'>q,q' ^'P> then P CA
is either
p[x] H^[-x] = '$
(q+q')y = pq' + p f q, hence
'.p[x] O A - convex or
Zorn's lemma now implies 7.7.3.
•
151
Proposition 7.7.4. subring, order
PCA
a
Let
(K,'£)
be a totally ordered field,
'$ n A - convex prime ideal
1
Q , such that
Af = A'/Q',
We may as well assume
by passing to /HPA[x]
or
assuming
1
A C A ,
(A,p. , PA, .) /HPA[l/x]
x > 0.
m
A
a^jb.
e
111 1
P,
m <_ n
since
Thus
is 1
"
'$ n A - convex.
+ ...+b]Ji)
we must have
1 -b
b! = b./(l-b Q ).
a > 0, or the term n
1 < a +a,x+ ••• +a x n . — o 1 n
But now
n
A
bQ < 1
and
Q
f
in the total order
where
A, = A[x]
P, . Then
A[x] If
Archimedean over defined by
x.
x
A. If
D
or
n.
(A',Q')
in
K
A[x]
because if
A[l/x].
Zorn's lemma now implies the
with the properties
Q' H A = P
A1 = A'/Q 1
Archimedean over
If
x
is transcendental, consider the cut of
is a transcendental cut of A.
If
D
is an algebraic cut of
k^ = A[l/x]
D = °° .
rather than
b! = b . / ( l - b Q ) .
x = x
A[x]
is is
A,D = D»(x),
A, then from 7.5 we conclude
were improper, we would have
x m £ bjx"1"1 + ••• + b ^ ,
to
P, C A.,
A, certainly the field
would not be a semi-field unless
/HPA[l/x]
(A,P)
Now, say, A,/P, = A[x], where
is algebraic over
then we return and work with
152
P, C A, = A[xJ, say, with
is a totally ordered semi-field since
is Archimedean over
from 7.5,
or
and
is least possible,
Return to the inductive step, where we passed from
maximal convex.
A(x)
n
A
$
£ - convex.
A = A/P.
modulo
Then
1 < a + a,x + • • • + a 1 x n " 1 + a (b'x n~ 1+. • .+b'x nJs~ m ), — o 1 n-1 n^ n m
It remains to show that we may construct
(A,,P,)
+
could simply be omitted from r J
is a local ring, P 1 H A = P.
existence of a valuation ring
1 <_ b
least possible.
Now we simply choose any maximal convex ideal Since
and
Finally, since
a x n
P,
x £ K, we claim either
is a positive unit in
which is a contradiction of minimality of
P C P1#
n A, and residue field
First, we lose no generality by
and
We must have
in
with
(A, '£A) .
1 £ a + a, x + ••• + a x
x (l - b Q ) £ bj^x " + ••• + b m . P
Given
is a proper ideal.
with
A' C K
is a local ring with maximal ideal
if necessary.
Then suppose
b 1 /x + .. .+ b m / x m
x
P = Q
1
-p1 = p H A'/Q', Archimedean ordered over
Proof:
a
A = A. p ./PA, . , with induced
'|5. = tp n A. p ./PA. . . Then there exists a valuation ring
p - convex maximal ideal
ra
A C K
A, then again But if
D = °° ,
A[x]. This is justified
1 <^ b
+ b,/x + ••• + b /x
But then from 6.3, x
would be
bounded by an element of
A, so we could not have
D.(x) = °° . This completes
the proof of 7.7.4.
•
Combining 7.7.3 and 7.7A, field, and
P C A
is a
K
refines
(K,£)
K •> A , ± °° such that the induced total
p:
'£, p
is finite on
in knowing if first the order
A
A
of
'$ n A/P
with center A/P.
on
For example, if
(A,P)
A/P
ring over a totally ordered ground field
A/P.
pf = q + r g
in
so must be A/P.
f.
A'
is
K
constructed lifting
is already a valuation ring and maximal We will see in the
A = R[X1«-.X ]
next chapter that this is also the case if
the order on
and
can be refined to a total
K, Proposition 7.7.2 says this is the case.
field of rational functions.
P
One might be interested
order in an arbitrary way, then a total order on
ideal in
K, then
f
Archimedean over the fraction field
this choice.
is a partially ordered
•£ n A - convex prime ideal of a subring of
we can find a signed place order on
we see that if
R
is a polynomial
K = ( ]/'* x n )
R, and
X
is
tne
However, in general, one cannot first assign
A typical difficulty is that one could have an equation A, with
p>q>r £ P.
Thus, if
However, this relation between
g f
is made positive in and
g
A,
disappears in
The geometry of this problem is interesting, and will be clarified
by results in the next chapter. Our proof of 7.7.4 is directly modeled after the standard proof of Chevalley of the place extension theorem in algebra. f
additional conclusion that the residue field over
A = A. ./PA,p..
x
can be constructed algebraic
A
Now, in our context, this cannot always be done.
difference is, a simple extension unless
is algebraic over
A[x]
of a field
A, whereas
A[x]
Specific examples arise as follows.
of
K
will be trivial.
Then
K
in two indeterminates over
Q,
P = (x,y).
x
A. A = Q, the rationals,
'£ the total order on
is Archimedean over
Secondly, suppose
The
can never be a field
First, suppose
P = (0), K = C O ) , IT = 3.14159... GIR, with IR.
A
can be a semi-field if
defines a transcendental cut of the ordered field
by the embedding in
That proof has the
A = Q[x,y]
induced
Q, hence any signed place is a polynomial ring
The extension field
contain enough elements so that even the weak order on
K
K
forces
K
could p(x/y) = T\ £ IR,
153
where
p:
even if
K -*IR, ± °° is any signed place with center K
is smaller, say
P
on
A.
Thirdly,
K = ©(x,y), our method of proof of the signed
place extension theorem involves first choosing a total order on (x,y) C Q[x,y]
is convex.
so that
If this choice is made "carelessly", it is possible
p(x/y) = TT G IR, for the resulting signed place on
that again
K
In a positive direction, we can at least say this.
K.
Suppose
A = R[x,---x ]
is a finitely generated integral domain over an ordered ground field K = R(x 1 *-*x n )
is the fraction field of
to a partial order
field
A
1
(So
P
tr.deg. R (A/P), since otherwise
p
would necessarily be the trivial place.
will be algebraic over
P C A
tr.deg. R(A 1) =
with center
must have
A, the fraction field of
A/P.
actually be a finite algebraic extension of R, of transcendence degree over
R
Moreover, Af
[68 , Vol. II, Chapter VI, §14, Theorem 31], the residue field
over
is
Then from classical valuation theory, the residue K
A'
p
is a convex prime relative
on
Thus
of any place
P C A
tr.deg.D(A/P) = tr.deg.n(A) - 1. K K
'p C K, with
a minimal non-zero prime.)
A,
R,
will
A, that is, a function field
one less than that of
K.
We next give some applications of the signed place extension theorem to the notion of semi-integral elements introduced in 6.4. Proposition 7.7.5. a subring.
Let
(K/-P)
be a partially ordered field, A C K
Then the following subrings of
K
coincide
(i) The ring of elements semi-integral over
A.
(ii) The ring of elements bounded by elements of refinements of
]5 on
A
in all total order
K.
(iii) The ring of elements finite in all signed places of
K
finite on
and refining the order '£. (iv) The intersection of all valuation rings of
K
containing
A
with
'^-convex maximal ideal.
Proof:
(i) C (ii)
bounded by elements of
since semi-integral elements over A
in the order '$.
(ii) C (iii) since a signed place induces an order.
154
A
are already
A,
(iii) C (iv) since if an element was not in some such valuation ring containing
A, it would be infinite in the associated place and in any signed
place refinement. (iv) C (i) since if an element 2
x G K
z
then the ideal
Ai(l/x )k[l/x ] C A[l/x ]
be an equation
n
1 <_ a,/x
is not semi-integral over
2
is proper.
+ • • • + a /x , or
x
n
Otherwise, there would
<_ a,x n "
Thus, there exist by 7.7.3 and 7.7.4 valuation rings A[l/x 2 ]CA',
with
l/x2GQ«.
x 2 £ A1
Thus
hence
A,
+ • • • + an,
(A',Q')
in
K,
Q
a. € A.
1
xfEA1.
convex, D
The following corollary was proved more generally in 6.4.
Corollary 7.7.6. '•P = '^
a
Let
derived order,
(B,'J>) be a partially ordered integral domain,
A C B
over
(A, '.p H A ) .
exist
'$ - convex prime ideals
Proof: defined by
a subring, such that
Then for each
Let
K
Q C B
with
(Since
'.p = 'Pd
on
where
B C A , where
Q
C A
P
the "order
B, the notation is justified.) p
of
K, with center
is the valuation ring of
p.
Let
P C A.
Q = B H Q ,
is the maximal ideal.
a local ring, with closure of
elements
B, f C K
D
P
Proposition 7.7.7.
(A',Qf)
A
P C A, there
Q H A = P.
denote the field of fractions of
'-p C B.
is semi-integral
f H A - convex prime ideal
Using 7.7.3, 7.7.4, construct a signed place By 7.7.5
(B,'.p)
A
in
with
Q1
x
Let
(K/p)
be a partially ordered field,
'|5 - convex maximal ideal K
P C A.
Then the semi-integral
is the intersection of all valuation rings in
£-convex, A C A 1 , and
P = Q 1 n A.
which satisfy a semi-integral equation
A C K
x
n
K ,.
Moreover, the + b,x
+ ••• + b-
l with
bi £ P
are exactly the elements in the intersection of the above
maximal ideals Proof: 2
Q' C A 1 .
Suppose
x G K
2
2
, l/x )A[l/x ] C A[l/x ] with since
bQ G P , A
< 0
zn —
a i e A, hence
is a local ring.
is not semi-integral over is proper.
Otherwise
A.
Then
1 <_ b Q + a^x2
x 2 n <_ a|x 2 n " 2 + • • • + a R , where
+ • • • + a n /x 2 n ,
a! = a./(l-bQ) € A,
(Strictly speaking, we should first refine the
155
order '$ to insure that f
(A',Q )
valuation ring
1-b with
is positive.) 2
1
The last statement follows since if and
p:
(P, 1/x2) C Q « .
A[l/x ] C A ,
does not belong to all valuation rings over x
Thus, we can construct a
A n
with maximal ideal over
+ b ^ n ~ 1 + ••• +^2n
K •> A, ± °° is a signed place finite on
p(x) £ A
since
x
Conversely, if
is semi-integral over
x
This shows
A
and zero on
A, hence
p(x
b
-°>
) <_ 0
i
x P.
G
P
P, then and
p(x) = 0.
satisfies no such equation, then one argues as above that
2
2
/HPA[l/x ] C A[l/x ]
is proper.
Q1 H A = P
constructed with
(A f,Q !)
Thus a valuation ring 1/x2 e A 1 , hence
and
can be
x f. Q 1 .
D
Our final results in this section will concern chains of prime ideals in an integral domain containing of
P C P- C ••• C P C A.
In particular, if
K
is a field
A, we are interested in specialization chains of signed places
K, K -»• A
with
pi
o , ± °° -*• A ,1 , ± oo -»••..-»- A n , ± ° °
finite on
the composition ideal of
and
P^
the center of
K -* A. , ± °° . If
p i , then
VV
A
(A-,Q.)
A n C A n l C ••• C A Q
and
p^
on
A.
Here
p^
denotes
is the valuation ring and maximal Q Q C Q 1 C •. . C Q^.
In fact,
VcQ.r
The following general result provides some added insight into the close relationship between valuation rings and totally ordered integral domains. Proposition 7.7.8. Let I,J C A
(A,p) be a totally ordered ring.
are two convex ideals, either
I C J
or
Then, if
J C I. All convex ideals
are absolutely convex.
Proof:
Suppose
is convex, no element elements of
yGJ
x E I-J.
We may assume
can be greater than
x.
Thus
J, in particular for all positive elements.
y G I, hence
positive, hence
Next let
x > 0. y < x
Since
I
Since
J
for all is convex,
J C I.
Also, if
156
I <j- J, say
I
is convex, 0 < x < y 0 < xz < yz
hence
and
yz G I, then we may assume
xz G I and
I
(K/p) be a partially ordered field,
z
is absolutely convex.
A C K a subring and
•
p
i>
° -
i
n>
-
^
n A
order refinement of form a chain, say
convex
"
P r i m e ideals of
f, keeping all
Pi
A.
When is there a total
convex?
P Q C P 1 C • • • C P^.
Necessarily, the
Also, all
Pi
P./Pi C A / P ^
must
i < j , are
convex for a total order refinement of But now if
'P/P-. Thus P./P. must be CP/P-) -convex. •*J -*i d Pv/P. 3 P./P- are CP/P-) -convex ideals of
i < j < k, then
K
A/P i
and
l
3
l
i ^
P^Pj = (P k /P i )/(P j /P i ) C ( A / P ^ / C P ^ ) = A/P j
is convex for a
total order refinement of
C£/P.) /(P./P.) C A / P . . If we abuse notation 1 d 3 1 J slightly and write this last order as CP/P-) /p-> t n e n w © deduce P v/P. C A/P. 1 K d 3 J J is necessarily (CP/P-) /P-) -convex. Continuing, we find the following 1 d J d
necessary conditions for the existence of a total order refinement of which keeps all
(*)
If
P. C A
i < j , then —
(K,'J>)
convex:
P./P. CA/P. ji
is
i
Proposition 7.7.9.
( • • • ( f f / P J . / P , ) -1• - / P .1) °
d
d
d
- convex.
In the notation above, the conditions (*) are
necessary and sufficient for the existence of a total order refinement $' D'P
on
K
Proof: orders on
such that all
Let K.
P.CA
x E K, with
x, - x £ ' p .
As in the proof of 7.7.3,
generally that if
if
'£, '£', tp
that conditions (*) hold for
p ! HA-convex.
are
Let
'P'=1>[x],
']>' 0 $ " = f.
are orders on
A
with
'£ M =1>[-x]
as
Now we will prove f
'P n •£" = $
'$, then conditions (*) hold for
and such
'£' or for '£".
The proposition then follows from Zorn's lemma. We distinguish two cases, case (I) if case (II) if
P
is both
conditions (*) hold for with prime ideals
'£' and
P
'^"-convex.
CA
'£'. In case (II), we will pass to the ring
P i /P Q = P±, and orders
to the chain of prime ideals $'
or
'£" on
number of prime ideals on
A = A/P Q .
"^"-convex, and
In case (I) we will prove
f' = C P 7 P O ) , , and
Then we will establish conditions (*) for the order
for either
is not
P^
A
P.,
1 £ i <_n.
tp"1 n^»»
A/P Q = A
f" = CPff/Po) • on
A
relative
The assertion that (*) holds
then follows in case (II) by induction on the (We do not assert
'JF1 n ff" = (p/P )
as orders
If this last assertion were true, case (II) would be easy.)
157
Consider first case (I). Fix an equation u £ PQ.
First, we check that
P.j/P0
are all
makes sense because 2.7.1 implies if
P
r
'£'
convex.
Suppose
Pj/P Q
0 < u < v (rel 'p") , v G P , CP'/P0) -convex.
is not
J5
Note this
convex, then
P o
T
o was not
M
Cp'/P0) -convex.
is
There would then be
an equation 0 £ ax +z 1 £ ay + z 2
ae-jj 1 ,
with z k G P Q ,
a £ PQ,
(rel -p')
y^p.,
x
£ P^
Combined with
0 £ u £ v
(rel -p"), we obtain as in the proof of 2.7.1
0 £ u 2 ( a x + z x ) 2 £ u 2 (ay+z 2 ) 2 + v 2 ( a x + z x ) 2
But
(rel $)
v G P Q , z^ G P Q , so this equation simplifies to
0 £ u2a2x2 £ u2a2y2 + w
w G PQ.
with
in the ring x G p^
But
(rel 'P)
u,a ^ P Q , and we deduce therefore
A/P Q .
But
P^/PQ
0 <_ x 2 <_ y
p
is assumed
C^/ o) -convex and
(rel Cp/PQ) )
y G p^
hence
which is our contradiction.
Next, we verify the "second level" conditions of (*), namely that if then
P./P-
is
(Cp'/P ) /P-) -convex.
part of the proof above.
x £ P./P .
This makes sense because of the first
Assuming our desired conclusion false, we would obtain
0 < ax+
an inequality in
A/P Q , with
ay
a £ Pi,
a G'p'/P Q , z^ G P i /P Q ,
From this we obtain an inequality in
0 £ b(ax+z 1 ) + zj £ b(ay+z2 ) + z^
where
b f- P ,
b G«p',
z£ G p
0 £u £v
A
P
j/p0'
A,
(rel $•)
(rel 'p")
and its image in
A/P Q .)
Now
to obtain as in the proof of 2.7.1
0 £ u 2 b 2 (ax+z 1 ) 2 £ u 2 b 2 (ay+z 2 ) 2 + w
158
y ^
(By slight abuse of notation, we are
using the same symbol for an element of we combine this with
i <_ j ,
(rel 'P)
with
w £ P .
Since
u,b f- P , we have
0 <_ a 2 x 2 <_ a 2 y 2 + w'
w1 £
where
in the ring
P
i/p0
c A
/ p 0 > since z^ €
A/P.. i
P
(relCP/PQ) )
j/Po-
From this follows
0 < x 2 < y2
(rel(Cp/P ) /P.) ) ° d x d
But
and
y6p./P. ji
P./P. j i
is assumed
(more precisely, (*) easily implies this), hence original choice of
(Cp/P ) /P.),-convex ° d x d
x G P./P^, contradicting our
x.
The general technique involved here should now be clear. is not
((•••(Cp'/P ) /Pi) "-/P ) /P.) °d d d d
inequality in
A/P.
convex, m < i, we unravel a suitable
and get a (complicated) inequality
is combined with the inequality
0 <_ u <^ v
obtain a complicated inequality
(rel p)
again to get an inequality
(rel $") , in
A.
(rel '£')
v £ PQ,
Finally,
P./pi
given by (*).
m
d
in
A.
u £ PQ
This
to
this is unraveled
rel ((• • • (('p/P ) /P.) **•/?) /P.)
°d-Ld the assumed convexity of
P./p-
Assuming
1
which contradicts
d
We will leave the actual details
of this induction to the reader. We next consider case II ,
is both '.p1 and ^"-convex. We
when P
need to establish conditions (*) for the order relative to the chain of prime ideals that each P i /
p o
is
Cp7P0)
p
-j/
p 0
n
CP7PQ)
C ... C
p
p
n/ 0-
CP"/P0)
on A/PQ,
We f i r s t argue
n C-P"/Po) - convex, or, what is the same by
2.7.1, that each P ^ P Q is either
Cp7PQ) -convex or
(^M/Po) -convex.
Assuming otherwise, we would have inequalities 0 _ < a x + z 1 £ a y + z 2 ( r e l '£') 0 <^ bu + w : £ bv + w2 ( r e l
where w ^ z ^ P Q , y,v e p ^
x ,u
^ Pi,
a^1])1,
a £ PQ ,
b G f ,
b £ PQ.
From this (ax+z 1 )2 (bu+w 1 )2 <_ (bu+w x )2 (ay+z 2) 2 + (ax+z x) 2 (bv+w 2 ) 2
or,
(rel $
simplifying,
159
0 <_ a 2 b 2 x 2 u 2 <_ a 2 b 2 (u 2 y 2 + x 2 v 2 ) + t (rel
f o r some
t Gp . o
Then
a,b £ P ' o
0 £ x V <_ u y in t h e r i n g
z
i mrp l i e s
+x V
A/P , which i m p l i e s s i n c e
convex, t h a t e i t h e r
x G p.
or
~~
J
n
(CCP'/P )
VT,
y , v £ P.
and P . / P
is
Cp/P )
u G p . , contradiction.
The second l e v e l v e r i f i c a t i o n i < j , P./P . is 1
(rel
i n c a s e ( I I ) c o n s i s t s o f showing t h a t f o r
CP"/P ) ) / P1 - ) - c o n v e x .
° d
°
d
d
Assuming o t h e r w i s e , we
g e t in A/PQ 0 < ax+ z. < ay + z o —
where
z K
G P., 1
i—
y G p . , J
(relCB'/P )
z
x ^ P . , J
n
O f j
a ^ P . , 1
C^"/P ) )
a G (-^'/P
o&
) O (p"/P) . O J O(J
This
inequality becomes two inequalities in A
where
z^,z£ G P Q ,
+ z» < b ' ( a y + z 2 ) + z^
( r e l '])')
+ zy < b"(ay+z 2 ) + z'^
(rel ^")
b ' , b " £ PQ,
b 1 G^p», b " Gtp».
Then we g e t
(rel
in A < (b') 2 (b") 2 (ax+z 1 ) 4 < 2(b') 2 (b") 2 (ax+z 1 ) 2 (ay+z 2 ) 2 + t
with
t Gp
This g i v e s i n A/PQ
0 — < a4x4 — < 2a4x2y2 + u
where
u Gp
s i n c e z, G P^.
F i n a l l y , in A/P.
0 < x 4 <_ 2x 2 y 2
since
a ^ P..
imply
x G p.
But y G P.
(relflVP o )^ )
(rel(CP/PQ)d/Pi)d)
and convexity conditions (*) for
P./P.
which is our contradiction.
Again, we leave the rest of the details of case (II) to the reader. The general step is to show that
P./P. is ((• • • ((CP'/PJ J
160
1
° d
n CT/PJ °
1/PJ ...PJ /P,) d
i
d
m
d
d
convex for
m < i.
P ) /PJ
Assuming otherwise ultimately contradicts the assumed /PJ
•*• d
d
'"/?
m
•*• d
) /P-) d
1
- c o n v e x i t y of
P./P..
d
3
So e n d s o u r p r o o f
1
of 7 . 7 . 9 .
D
Remark:
The geometric significance of 7.7.9 for function fields over a
fixed real closed ground field P C R[X^...X ]
R
is roughly as follows.
Convex prime ideals
will have d-dimensional zero set in affine space
d = tr.deg.R(R[X1-..X ]/P). local dimension less than
However, P
d.
R n , where
may also have "degenerate" zeros, of
Passing to derived orders has the effect of
eliminating these degenerate zeros from the maximal convex ideal spectrum of R[X,,,,X ]/P.
Similarly, total orders on
R[X,...X ]/P
can be interpreted
as orders obtained by evaluating functions on (infinitesimally) small open sets in the non-degenerate part of the zeros of
P.
in 7.7.9 that assure a chain of convex prime ideals
Thus the conditions P.
can be kept convex
in a total order amount to adherence conditions on the non-degenerate zeros of the P_
All this will be clarified by the results in the next chapter.
Proposition 7.7.10. a subring
P
C p
Suppose
C ••• C P
the conditions (*) of 7.7.9. places of
(K,'p)
a chain of
Q^ H A = P^.
± °° -*• A 1
Proof:
±<x>->... -> A n
K -> A ^ ± °° denoted by
Moreover, each residue field
over the field of fractions of
First, choose a total order on
Archimedean over
(A^Qp
P]/P Q
c A
/PQ
such that
A^ = A^/Q^
A/P .
pQ:
K, keeping all
A C Ai
is Archimedean
P.
We now have the convex prime
and with
A^
convex.
K •> A Q , ± °° , with center
and we apply 7.7.4 again to get a suitable signed place with center
i °°,
A/P^
Now apply 7.7.4 to get a signed place A
p n A-convex prime ideals satisfying
K,
with valuation rings of
and
A C K
Then there is a specialization chain of signed
K -* A o
and
is a partially ordered field,
Archimedean over
PQ C A
P^/P Q CA/P
CA
A Q •* h^, ± °° , A/?^.
Continuing
by induction gives the desired specialization chain.
•
161
VIII - Affine semi-algebraic sets
8.1.
Introduction and Notation Throughout this chapter we fix a real closed ground field
E C R
n
of affine n-space over
R
is a semi-algebvaio set if
the smallest Boolean ring of subsets of
R
U(g) = {x G Rn|g(x) > 0>,
R. E
A subset belongs to
which contains the sets
g G RtX^'-Xj .
We also define sets
W(g) = {x G R n|g(x) > 0} Z(g) = {x G R n|g(x) = 0} .
If
g. G R[X •••X ]
is a finite collection of polynomials, define
u{ gi } = nu(g.) w{ gi} = n w( gi ) z{g.} = nz(g.) . Any semi-algebraic set sections of
(*)
Zf s
E C R
and
can be expressed as a finite union of inter-
U's
E = u (Z{fik> n u { g j k } ) ,
for finite collections of polynomials is of course highly non-unique. properties of the sets tations (*) .
162
E
f-v> S-v*
Such a representation of
We are primarily interested in "geometric"
themselves, not in the particular represen-
E
Finite unions of the Their complements in
R
n
U{g^}
will be called open semi-algebraic sets.
will be called closed semi-algebraic sets.
typical closed semi-algebraic set
S = U si ,
S
A
can be represented
where
(**)
f..} n w { g . k } .
Of course, equalities and
f = 0
could be avoided altogether by writing
f > 0
f £ 0, but this is psychologically less natural.
Lemma 8.1.1.
If x G U{g i >, then there is
g G RfX^.X^
such that
x 6u(g) Cu{g.}.
Proof:
The estimates of 7.2 can be used to find a ball around the
point
x = (x.....x ) G R n , contained in
U{g.}.
where
B(x,e) = {y| lly-xll < e} ,
= £(y.-x.) .
given by
2
Hy-xll
That is, B(x,e) C u { g . } , A suitable
Rn.
topology on
•
U(g)
form a base for a
But we now emphasize that semi-algebraic geometry is
precisely not concerned with this strong topology. ring of semi-algebraic subsets of
R
Instead, only the Boolean
is relevant.
If
E CR
is semi-
algebraic, then we also get a natural Boolean ring of subsets of by intersecting
closure
is then
g x > £ (y) = e - E ( 7 j - X j ) .
Of course, Lemma 8.1.1 states that the sets
If
g
2
E, F ¥~
of
E
with other semi-algebraic sets.
are semi-algebraic and F
in
E
F° = {x e E | exists
F"E = {x € E | for all
Of course, F°
E, simply
and
F
in the topology on
E
F C E, define the interior
F°
and
as follows.
e > 0
such that
y G E, lly-xll < e
e > 0, there exists
yEF
implies
such that
yEF}
lly-xll < e}.
are the ordinary toplogical interior and closure of with base the sets
U(g) H E .
F
But, for our purposes,
the point is that it follows easily from the Tarski-Seidenberg theorem, to be
163
discussed below, that we will simply write
F° F°
and and
F
are also semi-algebraic.
(If
E = Rn,
F.)
We now make the following confusing (but crucial for understanding the differences between algebra and topology) definition. F C E F
A semi-algebraic subset
will be called an open, semi-algebraic subset of
will be called a closed, semi-algebraic subset of
commas should not be ignored, at least for now.
E
of
U{g.} n E
E.
if
if
F = Fp
F = F .
and The
It is easy to check that
the complements of the open, semi-algebraic subsets of semi-algebraic subsets of
E
E
are the closed,
It is also easy to check that finite unions
are open, semi-algebraic and (hence) finite unions of
Z{f.} <^w{g.} H E
are closed, semialgebraic subsets of
E.
In other words,
open semi-algebraic subsets are open, semi-algebraic subsets.
In fact, the
converse is also true, which we now state as
Unproved Proposition 8.1.2. open semi-algebraic subsets of
Open, semi-algebraic subsets of
E
are
E.
D
Thus, the comma can be ignored when talking about open or closed semialgebraic subsets. difficult to prove.
However3 this foundational point seems to be rather In particular, it does not follow directly from the
Tarski-Seidenberg principle, as does the fact that
F
and
F
are indeed
semi-algebraic. The paragraphs above are a bit of a joke, written in order to emphasize the subtlety of
8.1.2.
We will develop in this chapter an extensive theory
of semi-algebraic sets without using 8.1.2.
This will require some care
with words, or, at least, with commas.
It would be nice to have a simple
proof of 8.1.2 right at the beginning.
On the other hand, all the results
we will prove in order to circumvent 8.1.2 are results we would want anyway. The Tarski-Seidenberg theorem, [56], [57], [62], is sometimes regarded as a theorem in logic.
It asserts the existence of a decision procedure
for deciding the truth of any elementary sentence, built up from finitely many inequalities
P.(x....x ) > 0, where the
P.
are polynomials over a real
closed field, using basic logical connectives "and", "or", "not", and quantifiers "exists
164
x.", "for all
x.".
The decision procedure amounts to checking
whether or not certain polynomial inequalities involving the coefficients of the
Pi
hold.
For example, the sentence "there exists
2
x G R
such that
2
ax + bx + c = 0"
is equivalent to
"b - 4ac >_ 0".
Also, Sturm's theorem in 7.3
is a special case of such a decision procedure. The proof of the Tarski-Seidenberg theorem is not difficult. Cohen's proof [62] in an appendix.
We will give
From the algebraists point of view, what
is involved is just an argument making use of (1) induction on degrees of polynomials and number of variables and (2) explicit calculations in polynomial rings involving partial derivatives and division algorithms.
In other words,
elimination theory. The applications of the theorem are rather striking. sharp distinction between two types of application.
We will make a
The first type, which
is almost a reformulation of the theorem itself, says that any set defined in terms of semi-algebraic sets by an elementary sentence is still a semialgebraic set.
For example, this includes closures and interiors mentioned
above, images of semi-algebraic sets under polynomial mappings and other frequently used constructions.
R n -> H m ,
(However, the theorem gives little
insight into the question of whether a set is, say, open.)
It is amusing that
even the simplest special cases of this type (say the projection to zeros in
R
R
of the
of a single polynomial, or the closure of the set where a
single polynomial is strictly positive) are no easier to analyze than the whole theorem.
Thus the Tarski-Seidenberg theorem is a very efficient tool.
We emphasize that this type of application actually provides a proof that the asserted set is semi-algebraic, simultaneously for all real closed fields, in fact, an elementary proof.
The reason is, any single elementary sentence
is just a special case of the theorem.
The more subtle application is this.
Given an elementary sentence, suppose it can be checked in one real closed field where it makes sense and
is true.
For example, it might be checked for
the classical real numbers by transcendental methods (use of completeness, possibility of integration, etc.)
Then the sentence must be true for any
real closed field to which it applies, because there exists a decision procedure which is independent of the field in which it is carried out. Thus, one might say that this method of application amounts to proving by non-
165
elementary methods in one special case that an elementary proof of the statement in general does exist.
It certainly might be very tedious, if not
physically impossible, to actually work out this elementary proof. In this book we absolutely and unequivocally refuse to give proofs of this second type. fields.
Every result is proved uniformly for all real closed ground
Our philosophical objection to transcendental proofs is that they may
logically prove a result but they do not explain it, except for the special case of real numbers.
Also, one of our central themes is that the real numbers are
totally irrelevant in algebraic topology, so it would not do to rely on them at some point in our chain of reasoning.
Finally, there is already a respect-
able tradition in this century of finding non-transcendental proofs of purely algebraic results concerning algebraically closed fields.
We think real closed
fields deserve (at least) equal time and effort. We do not at all mean that only elementary proofs are acceptable. we use Dedekind cuts, total orders, and signed places repeatedly.
In fact,
The point is,
in the form we use these concepts they apply uniformly to all real closed fields. One advantage to developing such techniques is precisely that one is not tied down to "elementary sentences".
There is rich non-elementary theory to be
studied for arbitrary real closed fields, and even if a statement turns out to be equivalent
to
an elementary statement, it may be unnatural to dwell
on this fact, and even worse to be forced to depend upon it. There are obviously aesthetic questions involved in this discussion.
We
admit that many of our proofs are long and could be replaced.by the single phrase "Tarski-Seidenberg and true for real numbers". effort is worthwhile.
However, we feel the
In fact, we do not use the Tarski-Seidenberg theorem at
all, until 8.7. We now change tacks somewhat and indicate a more invariant approach to semi-algebraic geometry. R-algebra of finite type.
Suppose
(A,p) £ (PORNN), and assume
Define a set
X = X(A,'£)
A
is an
by
X(A/p) = H6m ( p 0 R ) ((A,tp),(R^ w )).
(Of course, 'C
166
is the only order on
R.)
We have the canonical adjoint
homomorphism from the elements
A
to the ring
R
of
R-valued functions on
f £ 'p define functions nowhere negative on U{g i >, W{g.}, Z{g.} C X
the subsets
X.
X, and If
g. E A,
are defined in the obvious way. Interiors
and closures of subsets of X are also defined just as above, essentially using the topology X with base the collection of sets U(g), g^A. (The U(g), g e A, are invariants of (A,'£) and replace the e-balls in the earlier definition.) If we choose a specific presentation I CRpC^.-.X ]
A = R[X,...X ]/I, where
is some ideal (necessarily radical and
n
gets identified with a subset of the zeros of
I, Z(I) C R .
X = X(A/p) = {x G Z(I) |g(x) ^ 0 all
Obviously, X
is a closed subset of
Rn
$ -convex), then
X
Specifically,
g € tp}.
in the topological sense. y
Now, the problem is, the homomorphism In fact, X
could be empty.
Definition.
RHJ-algebra (real, Hilbert-Jacobson) vif
is an R-algebra of finite type P C A , and
f £P
and
need not be injective.
Thus, we make the following definition.
(A,'£) is an
ideal
A -• R
A
(A,'£) £ (PORNN), and for each prime 'p-convex
g £ P, there exists
x E X
with
f(x) = 0
for all
g(x) t 0.
Corollary 8.1.3.
If
(a) the homomorphism (b) the set
X
(A,'£) is an A -* R
RHJ-algebra, then
is injective
is identified with the maximal convex ideal spectrum of
(A,?)
(c) for any subset
J C A, we have
vtt(J/p) = {£ e A|f(x) = 0 all
where
Z(J) C X
is the set of zeros of
xC-Z(J)},
J.
Proof: (a) The intersection of all '^-convex prime ideals of (A,'p) G (PORNN).
167
A
is
(0), since
(b) Each all
e
f '£.
x £ X
corresponds to a surjection
Thus kernel (p) C A
the definition of
contain
/H(J,$) J.
f €= A
is already maximal, then
"jj-convex maximal ideals containing
which vanish on all zeros of
J
in
P.
Thus
P
is /H(J,$)
X.
D
RHJ-algebra is attractive, there is still a
Without some control on the order
p C A, the resulting subsets
of affine space (relative to a presentation
can be rather chaotic. an
P
P
x £ X.
Although the notion of
X = X(A,p)
Conversely,
Again, the definition and (b) says each p-convex prime
is the set of
p(£) >^ 0,
is the intersection of the prime '^-convex ideals which
the intersection of all
problem.
with
RHJ-algebra states that every '^-convex prime ideal
corresponds to some (c)
A -*• R
is a maximal, '^-convex ideal.
is contained in such a maximal ideal. Thus if P
p:
A = R[XX ...X n ]/I)
The control needed in order to guarantee that such
X(A,'£) is semi-algebraic (necessarily closed, semi-algebraic) roughly
amounts to finiteness conditions on orders naturally associated to •£. Our study of semi-algebraic geometry will essentially amount, then, to identifying a large class of
RHJ-algebras
(A/p) with
X(A,'p) semi-algebraic and,
conversely, given a closed, semi-algebraic set RHJ-algebras easy if
E
(A,'p) with
E = X(A,'£).
E C R n , construct natural
This last turns out to be fairly
is closed semi-algebraic, but not so easy if
E
is only closed,
semi-algebraic.
8.2. Some Properties of RHJ-Algebras In this section we will construct many
RHJ-algebras, assuming the
following basic real Nullstellensatz, which will be proved in 8.4.
Proposition 8.2.1. is an
The polynomial ring with the weak order
(R[X^.. .Xn] ,'p
RHJ-algebra.
Of course, X(R[X,...X ] ,'£ ) = R n . l n w
•
Once we have this one
RHJ-algebra,
there are natural constructions of others. We indicate several such constructions in the propositions below.
168
Proposition 8.2.2. 'P(X) = { f G A|f (x) :> 0
If
(A/-P)
all
is an
RHJ-algebra
x G X}, then for any order 1
$ O p ' C ^ ( X ) , we have that
(A,? )
In particular, the radical
is an
X = X(A/P), and f
CA
with
RHJ-algebra and
X(A/p») = X(A/P).
'p-convex ideals and the radical
'.p1-convex ideals
coincide.
Proof:
Any
'£'-convex ideal is also 'p-convex.
obviously define
'-P(X)-convex ideals, hence also
Since the points
x ۥ X
f
'p -convex ideals, the
proposition follows easily.
Corollary 8.2.3.
•
If
(A/p)
and
(A/PJ m are RHJ-algebras. X(A,!Pm), and $m
Proof: P
=
'-P(x)
is an RHJ-algebra, then
(A/p ) , (A,'P ) ,
We have X = X(A,tp) = X(A/PJ s = X(A,:pj p =
The orders
and
it
'£ , 'p , :p were defined in 3.12. r s rp rm eas Y t 0 check that for an RHJ-algebra
is
By definition, J ' (A/p), we have
• Proposition 8.2.4.
If
(A,'p.)
are
RHJ-algebras, 1 £ i £ k, then
(A, H'p.) is an RHJ-algebra and X(A, rvp.) =iUX(A,!p.). l I l Proof:
The results of 2.1 show that any
'.p.-convex for some
j.
The result follows easily.
Proposition 8.2.5.
Proof: P C A
Any prime
is a prime
(A,'p )
Proof:
is an
Write
(A,'p)
is an
RHJ-algebra and
'^/I-convex ideal of
If
A
•
RHJ-algebra and
'.p-convex ideal containing
Corollary 8.2.6. then
If
(A/I, (]5/I) is an
'^-convex ideal, then
H *p. -convex prime ideal is
A/I
I CA
is a radical,
X(A/I, -p/I) =
is of the form
P/I, where
I.
D
is a reduced, real R-algebra of finite type,
RHJ-algebra.
A = R f X ^ . .X n ]/I, where
I CRj^.-.X^
is a radical,
169
'P -convex ideal.
(This is exactly what the hypotheses on
A
mean.)
Now
apply 8.2.1 and 8.2.5.
•
Proposition 8.2.7. such that
A
is an
is finitely generated over n
(A/£ A ), where
over
(B,£B)
If
£ A = A '-Pg, then
X(A,'£A) = image (X(B,$B))
R
and
(A,'£A)
A C B
(B,'£ )
is an
under the natural map
induced by the inclusion
Proof:
RHJ-algebra,
a subring
is semi-integral
RHJ-algebra and
Spec(B,'^B) -• Spec(A/pA)
A C B.
This is immediate from the going up theorem for semi-integral
extensions proved in 6.4.
Proposition 8.2.8. let
f CA
Let
A
(0) = O p . ,
Then
Proof:
(For example, the
(A.,'p.)
Next, suppose g £ Q^.
P i C Q, and g(x) t 0. Thus
A
is an
(A,'£)
Then
g £ Q.
Then
(Ai,']5i)
Qi Find
A
contains some
is an
is an
x G X(A,'£) x
Proposition 8.2.9. Pi C A
given orders
Let
A
by 3.11.)
A -*• R P-.)
under the inclusion
(A^p^) = is an
RHJ-algebra
contains some
This implies
Q^ C A.
is a
^-convex prime
so that
x
P. .
X(A,'p) =
'.{K-convex
is a zero of
Q
and
g(x) t 0.
The converse is equally routine, using A
contains some
P^.
•
be a reduced R-algebra of finite type,
and define an .order
(0) = n p . , '|5 C A
by
Suppose
'|5 = A f l l f - ,
A -• n A. . Then X(A/p) = U X(A. ,'p.) and (A,tp) l l l i RHJ-algebra if and only if each (A^^K) is an RHJ-algebra.
170
(In
Q C A , with
Q i = Q / P ^ and
be a finite collection of primes with !p. C A . = A/P.
Let
Moreover, (A/P)
is a zero of
RHJ-algebra.
the fact that any prime ideal of
could be the minimal primes
RHJ-algebra and
corresponds to a
x e X(Ai>^i),
P.
'p-convex
RHJ-algebra.
The kernel of any homomorphism
fact, any prime ideal of
prime
be any finite collection of
X(A/P) = U X(A.;j5.).
if and only if each
an
P. C A
A, which are convex for any order on
(A/Pi, '-P/Pp.
let
be a reduced, real R-algebra of finite type,
be an order, and let
primes with of
•
is
Proof: for some
By Proposition 2.7.8, if
i,
P. C Q
and
Q/P.
is
Q C A
is any
$.-convex.
'^-convex prime, then
All parts of the proposition
above follow routinely from this fact.
Proposition 8.2.10. and
'£ = '£ [g.] w j
is an
A
is a reduced, real R-algebra of finite type
is a finite refinement of the weak order on
A, then
(A/p)
RHJ-algebra.
Proof: refinement
Let
P. C A
'p [g.] = 'p. I w j
Of course, some •}K.
If
•
Now, A^
g.
be the minimal primes
will also be an order on
may be
0
in
A.
B
.
is a domain, unless
g
1
§m+1
and go on to
steps, we have a domain
B
= A.,
where
*-Pw(B) C B
is the weak order.
Applying 8.2.7, (Ai,tpi)
(A,p)
RHJ-algebra.
Remark.
g
. = B [/g
. ] , m > 0. B .
The first
, £'J> [!•••••§ ] C A . .
instead.
B = lim B , integral over
RHJ-algebra. is an
B
g. ^ 0, consider the
is already a square in
Bm[>/gm+2]
R, and real, say be Proposition 6.2.1.
g. = g. (mod P . ) . j j i
and these can be omitted in studying
time this happens, an easy computation shows we can skip
The finite
A., where l
is an integral domain, and assuming all
finite sequence of integral extensions Each
A. = A/P..
Thus,
After finitely many A., of finite type over
Moreover, f. = 'p [g.] = A. n •£ (B), Applying 8.2.6, (B,$w(B)) is an RHJ-algebra.
is an
Finally, by 8.2.8, •
The proof of the basic Nullstellensatz 8.2.1, to be given in
8.4, will, in fact, yield directly that any order of type
finite real integral domain over R gives an
'£ [g-] w 3
RHJ-algebra.
be used instead of 8.2.7 in the proof of 8.2.10.
on a
This could then
On the other hand, both
proofs use the device of adjoining square roots. If
(A/P)
is an
X(A,'p) C X(A,? w ). Now, if
RHJ-algebra, then so is
In fact, X(A/£) = {x € X(A,? w ) |g(x) > 0, all
A = R[X1...X ]/J, where
^-convex ideal, then n
Z{f.} C R . 1
then
If
(A/p w ), by 8.2.6, and
X(A,'^w)
£ = ^ [g.] W
J = (f/) C R p ^ . J ]
g
is a radical,
is identified with the real algebraic set
is a finite refinement of the weak order on
J
X(A,'£) = Z{f.} Hw{g.}. 1 3
More generally, if
!p = H •£K
where each
171
A,
$
= 'P w [g- k ]
is a finite refinement of the weak order, 1 £ k £ m, then by
8.2.4, X(A,p) = U z{f i > n W{g. k >.
Thus orders of this type give rise to
closed semi-algebraic sets. We now want to prove that any closed semi-algebraic set
(**)
s = us.,
can be represented as A
a quotient of
RfX^.-X^. S.
Trivially, I(S) = n I(S i ), S.,
is an order on Let
P.
A ^ , since the
We now have
k(S/) = R[X X . . .X J/ICS^ g.v
g-k
A i a = R[Xr..Xn]/Pia.
I(S.)
Then
f±a
in
P.
A(S.)-
= » w [g i k l
P i a C P.0
implies
and thus we have an inclusion
P.
CP.»
A(S)
occur among the
hold, even
are the minimal primes of
RHJ-algebra and
Proof:
we have
I(S.), we know
If
$ = A(S) n
n
^
C A ( S ) , then
(A(S),^)
is
X(A(S),'^5) = S C R n .
X(A(S),tp) =
X(A(S),'^) C S.
I(S.)
is a zero of
as an
A. -ideal, since to ik ia
U
*(\a>Via)'
Conversely, since
P. , some
a.
g., (x) > 0. -
If
Since clearly
x
Of course, the order
is an
XCA^,^)
I(Si) = C\ P ^ , any zero
x G S., then
Thus
(A(S),'P)
is clearly
x
c S
'^ -convex
•
y
p C A(S)
constructed above depends on the
i>
of
S. C U X(A. ,'£. ) , and l a ia la
S C X(A(S)/P).
172
= P.^,
1
Propositions 8.2.9 and 8.2.10 guarantee that
RHJ-algebra, with
P.
i i j.
Proposition 8.2.11. an
is an order
RfX^^.. .X n]/I(Si ) = A(S^).
10L
that
IK
R|X . . .X ] , so that
are convex for any order on
It is possible that some relations i, the
$ T[g.,] W
may be zero in
Note that all the minimal primes of
but since for fixed
S., hence 1
Of course, some
P.a/I(S^)
S.
is a ring of functions
are non-negative on
I (S) = H I(S.) = flflP.
A(S) -> II A. . P. .
is a radical, f -convex ideal
denote the minimal primes over Let
denote the ideal of
IK
A(S.)«
I(S i ) = n PiQj. on
I(S)
(A,'p), with
is a ring of R-valued functions on
1
1J
RHJ-algebra
I(S) CR[X 1 ...X n ]
Then
f.. £ I(S.), and the
1
for a suitable
Let
A(S) = R[X,...X ]/I(S)
since
on
s i = z{f..} nw{g ik }
S = X(A,'.p)
functions which vanish on
S C R ,
representation (**)
of
{f GA(S)|f(x) > 0
all
RHJ-algebra. S C Rn.
S, and is a rather weak order. xGS}.
We refer to
'p(S) =
Then by 8.2.2, (A(S)/P(S))
(A(S) ,'£(S))
The fact that it is an
Let
is also an
as the affine coordinate ring of
RHJ-algebra is geometrically very satisfying.
However, a useful question is, how does one go about deciding algebraically, in terms of a representation (**) to
'P(S)?
f G'-P(X')
for some
n >_ 0,
this implies
f G A(S)
(A',p'), with
RHJ-algebras
if and only if there is an equation
p,q G ^ 1 .
belongs
X' = X(A',p'), (f 2 n + p)f = q,
(The "if" part is trivial.)
'£' = $(X f ) = '£' , where
defined in 3.12. f
S, whether a given
One answer is provided by a theorem of Stengle, which asserts
that for a certain class of we have
of
'£' and
In particular,
'£' are the operators on orders
In fact, it also shows that in this case the collection of
satisfying such a formula is closed under sums, thus simplifying the con-
struction of
p1
in these cases.
The most general class of orders satisfying
Stengle's theorem is obscure, but it includes at least all is a reduced R-algebra of finite type, and
f CA
(A/p), where
A
is a finite refinement of
the weak order. In particular, reconsider the closed semi-algebraic set by (**)
above.
R[X r ..X n ]/P i £ x ,
We constructed an inclusion P.a
prime.
S = X(A(S),A(S) H n 'Picp-
A(S) -• II k^a>
We had the orders
^
S
described
where
A^a =
= 1>w[gikl C A.^
and
As a consequence of Stengle's theorem, we can
now state
Corollary 8.2.12. $(S) = A(S) n nCP- ) , where .
ia
CPia) = {f e
AicJ(f2n+
p
p)f=q, some n > 0 , p , q 6 y .
D
This result characterizes algebraically the collection of polynomials f G R[X 1 ...X n ]
non-negative on the closed semi-algebraic set
S C Rn .
We
will prove Stengle's theorem in 8.5. Note that recovering
S
from the
(A. ,'p. )
an algebraic set from its irreducible components. setting, where orders
f^a
is analogous to recovering In this semi-algebraic
are carried along with the integral domains
173
A i a , or with the prime ideals
P.^ C R[X^...X ] , we can have the same
occurring several times with different orders
'£ [g-k]
corresponds to several patches on the variety
Z(P.^)
W
^g'k^
n
Z
( P icP J
P
ia5 P j0 J ° r
a11
contained in
e< uivalentl
l
belonging to
S.
Remark.
with
P a t c h e s on
The orders
$ [g.,] W
on
A.
IK
so we can find
V w [ g i k | g i k £ Pia] •
x G X(\a$ia)
hence the patches
^ia*
z p
( j^)
= R[X,...X ]/P.
XCc
Let
x
n
g =
also
g(x) ^ 0.
)
contain points
are strict, g i k (x) > 0, for all tJie c o n
d
ition
g^k >. °
is
g ik -
Then
g £
P^,
ik ^ P ia
with
W{g.^} H Z ( f
considered above
10c
II g
ik
of the form
nw( gl )
can be replaced by
e
This
We can even have proper inclusions
y» Z(?j^ $ Z ( P i c P '
— — — —
g i k (x) :> 0
A*/y"
P^a
S.
s=
g
on
It follows that x
where the inequalities
gik ^ P ^ .
superfluous on
z
g i k 0 0 > 0,
Of course, if
P
( £a)-
We have indicated above how certain operations on orders defining RHJ-algebras yield other
RHJ-algebras.
Certain other operators, which one
might hope do the same, turn out to be more subtle. are
RHJ-algebras
A
is
( /P[g])>
(A,'p)
such that some simple refinement of the order '^,
not an RHJ-algebra.
i: A -* A[t]
where
indeterminate over
We will show that there
Also, (A[t],iJ>)
is the inclusion of
A
need not be an RHJ-algebra,
in the polynomial ring in one
A.
The basic example, which is a good source of counterexamples, is (A,p) = (R[x,y], ? w [x] n ? w [ y ] ) . Suppose
h(x,y)
Thus, X(A,?) = {(x,y) G R 2 |x > 0
is nonnegative on
X(A,'^).
Since
h
or
y > 0}.
cannot change sign
across either the positive x-axis or the positive y-axis, it follows that
174
h(x,y)
is divisible by an even power of
for some
r < 0,
{(x,0)|x < r},
h
x
-x,-y
y.
It then follows that
is nonnegative on a neighborhood in
{(0,y)|y < r}
on the negative
A somewhat surprising consequence is that is an order.
and
This follows since for any
x-
and
R
of the two rays
y-axes, respectively.
$[-x,-y] = CPw[x] n •£ [y])[-x,-y] Si >•••>&]< *= P> the elements
are simultaneously positive on a non-empty open set in
sets in
R
R .
are Zariski-dense, relative to the ring of polynomials
g.,
(Open R[X,...X ] ,
because of the finite Taylor expansion of a polynomial about any point. holds
for any real closed
This
R, since the estimates of 7.2 show that the
formal algebraic partial derivatives of a polynomial agree with the "limit" definition of partial derivatives.) It is unclear whether or not
(A,'.]S')
2
X(A,'^') = {(x,y) G R |x < 0 < y}. be an
RHJ-algebra.
cannot be an
is an
RHJ-algebra.
It is then obvious that
(A/£) = (R[x,y], ?
RHJ-algebra.
W
M
n
above concerning functions
and
g
vanishes on
cannot
(A[t], i*£)
The point is, one can find an algebraic surface {(x,y,t)|x,y < 0 < t}
X(A[t],iJJ) = {(x,y,t)|x >^ 0
ijp
(A,1])11)
'£w[y])> we argue that
x-y plane to precisely the open third quadrant
is
In any case,
RHJ-algebras.
in three space in the octant
I(S) C A [ t ]
'£" = y [-y] = P[-x,-y].
In either case, we conclude simple refinements of
RHJ-algebras need not be With this same
p f = $[-x],
Let
y >_ 0}.
h(x,y)
convex. S, then
or
which projects on the
{(x,y)|x,y < 0}.
Now
But the discussion in the paragraph
non-negative on
The reason is, if f
S
X(A,'p)
implies the ideal
0 £ f(x,y,t) <_ g(x,y,t) (rel
must also vanish on at least a non-empty open
175
subset of
S, above a region near the (sufficiently) negative
x- or
y-axis.
In general, open subsets of varieties need not be Zariski-dense relative to the coordinate ring of functions
A(S), because of "degenerate zeros".
in this case, such an open subset of the hypersurface hence
S
But
will be Zariski-dense,
f e I(S). The ring
(A[t],i^)
is also easily identified with the fibre sum over
(A,?) ® (R[t],'£w ) . Thus direct sums of R in general.
In case
(A,'£ [g.])
RHJ-algebras need not be
and
(B,'£ [h.])
as finite refinements of weak orders, then
If
S
and
A(S) ® A(T) R
T
V
h
j
] )
RHJ-algebras,
RHJ-algebras obtained
(A,'p [g.]) ® (B,'p [h.]) =
w 1 R w j w i l l b e a n R H J - a l g e b r a w i t h X ( A & B , '£ [ g . ® 1 , R
(A & W B , 'p [ g . ® 1 , 1 ® h . ] ) R 1 J X(A, ? w [ g i ] ) x X(B,
are
R,
Wl ® 1
h.]) ^
'
are closed semi-algebraic sets, so is
is a ring of R-valued functions on
S x T.
S x T, and Thus
A(S) & A(T) R
is
a reduced, real
R-algebra, in fact, A(S) ® A(T) = A ( S x T ) . If A(S) and R A(T) are integral domains, so is A(S) &> A(T). (This is most easily proved R geometrically, using the irreducibility of the Zariski closures S and T of
S
and
T.)
Since a homomorphism
A(S) & A(T) •> R is just a pair of R homomorphisms A(S) •> R, A(T) -*• R, it is clear that X(A(S)® A(T) ,$ (S) ®fP(T)) = R R S x T. However, in general even £(S) ® '£ (T) will be too weak an order and R will have convex prime ideals with no zeros in
S x T, as in the above example.
Contracted orders are also hard to work with, in general. if the order
p = 'p [xy - 1] C R[x,y]
the weak order.
is contracted to
R[x], one obtains
Geometrically, it is hard to reconcile
X(R[x,y] ,']$) C R ,
which lies in the half plane X(R[x],'|5 ) .
x > 0
with the entire x-axis, which is
Presumably, a contraction of an
an RHJ-algebra.
For example,
RHJ-algebra need not even be
On the other hand, 8.2.7 states that contractions are very
well behaved for semi-integral extensions. Suppose and an order where
Rn.
176
is any subset.
Then we can define
p(Y) = {f e A(Y)|f(y) ^ 0
Y = Z(I(Y))
to see that in
Y C Rn
is the Zariski closure of
X(A(Y) ,'p(Y)) = 7, where
Namely, any
all
f
Y
non-negative on
y e Y}. Y
in
A(Y) = R[X][.. .Xn]/I(Y)
Of course, I (Y) R .
= I (Y),
Also, it is easy
is the topologieal closure of Y
is also non-negative on
Y.
Y
If
Y
is semi-algebraic, then so is
Y
by Tarski-Seidenberg. n
Y C R ,
we will prove that for all closed, semi-algebraic RHJ-algebra.
semi-integral extensions A C B
f e A,
i*:
with
(B,p)
an
RHJ-algebra with
X(B,p) -* X(A, 'p n A ) .
Zariski-dense in f
closure
i*(X(B/P) C X(A/P w ).
Y = X(A(Y) ,p(Y)) -• X(A, A H']5(Y)).
is non-negative on
Pa C B
A -> B
Thus
X = X(B,'|5), then
is injective, no function
That is, i*(X(B,'£)) f 6 'j5 H A
is
if and
A
-• B a
A a = A/i" (P a ). We have
and a diagram of inclusions
i • IIB a
is the contraction of an order on n B , namely, n '£(X(B /-P ))>
QJ
f C B, but
(X
'^
OC
f
= $/P . (X
(We could also contract IItp ,
CC
OL
could be definitely weaker than
the Noether Normalization Lemma, we can obtain of some pure polynomial extension
A , and then contracting
n(A
the only point at which we might lose pure polynomial extensions
B
A [X,...X, ]
AHf by first contracting the \i\), \
of
A .
= (VV> to
Thus we can study t0
A.
form
ViY^
for some subset
that
(A , Aa n>]3(Y1))
Perhaps surprisingly,
In fact, by induction we could
RHJ-algebra order on
Y x C X(A ,']$w) x R.
was an
VXr"\]'
RHJ-algebras by contracting is in the
A -• A [X,...X, ] . K^ a <x l
a finite, real domain, and with an
By
as an integral extension
X
H 1 ^ (X ))
'JHXCB^/p )).)
attack this extension by attacking the case of one variable A
is
X(A, '£ H A ) = i*(X(B,p)).
be the minimal primes, B a = B/P a,
X(B,'P) = U X(B ,']) ) , where
then to
f CB
i*(X(B/j$)), or, equivalently, on the topological
I f CB
i:
i*(X(B/|3)) C X(A/£ ) .
IIA a
and get
is an
RHJ-algebra, under
'£ = '^(X), where
Since
inclusions of finite integral domains
since
an
X(A,'p ) . Moreover, it is obvious that
only if
The order
(A(Y),'^(Y))
A C A(Y), inducing
f ^ 0, can vanish on
Let
Y
is any extension of reduced, finite R-algebras, and
an order with consider
(A(Y),'^(Y))
So far, we have established this only (a) for closed semi-algebraic
sets and (b) for images of sets
If
Ultimately,
RHJ-algebra, with
A
-•A [X,], with A [ X ^ , of the
We would then like to conclude X(A , A^ H']5(Y1)) = T T Q ^ T ,
177
where
IT: X(A^, ^^)
R -• X(A ,^ w)
x
is projection.
However, it is not clear
how generally this holds, and even in the semi-algebraic case, we need the Tarski-Seidenberg theorem to prove it. We conclude this section by pointing out that we have established an invariant criterion that
(B,'p) G (POR)
is the affine coordinate ring of some
closed semi-algebraic set in affine space over R-algebra of finite type. Pi C B B
i
= B
with
/ p i»
R.
Namely, B
must be a reduced
Also, there must be finitely many p-convex primes
(0) = O P i , and finite refinements of the weak orders
with
^
= B n
nC-Pi) > under the inclusion
B^IIB^
^
C Bi,
(Strictly
speaking, we must still prove the basic real Nullstellensatz, 8.2.1, and 'P(X(B. /£•)) = CP-) • 1 1 p these tasks in the next three sections.)
We proceed now to
establish Stengle's theorem that
8.3. Real Curves Let
K = R(x,...x )
closed ground field
be a real field, finitely generated over our real
R, with tr. deg. D(K) = 1. K
fractions of the integral domain P C R[X 1 -..X n ] zeros in
is a prime
Thus
K
is the field of
A = R[x-,...x ] = R[X,...X ]/P, where
'^-convex ideal.
We will prove that
n
R , that is, there exist homomorphisms over
R,
R C K, with
the valuation ring of elements finite relative to place of
K,
K
A C A R , where R.
must be Archimedean closed in
K
R.
(Or, since
for any order on
can be AR C K
K
over
tr. degR(K) = 1 , K
is
The associated signed
p R , is then R-valued since a non-trivial place of
lowers the transcendence degree over
has
p: A -• R.
Actually, we will prove something stronger, namely that totally ordered non-Arehimedean over
P
R R
non-Archimedean over
R.
Thus the residue field Thus we obtain
pR:
A n /Q n is algebraic over R by results of 7.6.) K K R. K -• R, ± °° and, by restriction, p: A -• A R
Proposition 8.3.1.
Let
K = R(x1-..x )
one variable over a real closed field Let
x £ K - R, so that
the contracted order
178
K
R.
Let
is algebraic over
'$ n R(x).
be a real function field in ']) C K
be any total order.
R(x), and let
'£ C R(x)
Then (using the notation of 7.5):
denote
Case (i) a >b
If '£ = '$m [resp. 'p J , we can find
[resp. a < b ] , both orders
f
and "J5 a,—
a,+ orders on K with Case (ii)
b G R
on R(x) extend to total
A C AR.
If "jT = $c
such that for all
+
[resp. $c
] , c G R, we can find
[resp.
a G (c-e,c)], both orders
a G (c,c+e)
!P on R(x) extend to total orders on K with a, ~ Case (iii)
If f = £ D where
such that for all a G ( a 1 , a " ) , both orders total orders on K with
c = 0, by replacing
'p
v j>J .i
be the minimal polynomials for f^ . (x) G R[x].
a 1 , a" G R,
R(x), say
Suppose
y
and
m .l«y ij
i j
x.,
1 <^ j _< n, with
g.(x),
K, the polynomial
(«(x),j3(x)) c R(x)
The last
aG
(c, c + e)
[respectively
G i (a(a)), Gi(3(a))
have
e > 0
has real roots over
{G Q (a(x)),.. .Gk(a(x)) = h(x)}
f-signs in
in case (ii) or
R(x). We now appeal [resp.
a' < D < a"
in
(c - e, c ) , depending on whether
a G (a 1 ,a")], then the R-signs of all
agree with the
We also make our choices so that
G(y)
is an interval containing (at least)
The terms in the sequences
case (iii)] so that if 'B c,-
G Q (y),... ,G k (y).
h(x) ^ 0 G R(x).
to the results of 7.5 to choose
terms
K = R(x,y). Let
m.-l J-»i ,J
I J
{Go(j3(x)),...,Gk(j3(x)) = h(x)}
or
on R(x) extend to
••• + g m (x)
+
G Q (y) = G(y), G ^ y ) = ^
by our assumption on
one of these roots.
•IT = *B c,+
, '$
a f < D < a",
Consider the Sturm sequence for the irreducible polynomial
term is a constant
and
K
y G K so that
m. j j J J
(R(x), f).
A C AD.
x by 1/x.
0 = G(y) = g o ( x ) y m + g 1 ( x ) y m - 1
Now,
and
A C AD. K
In cases (ii) and (iii), choose
over
'J5
Case (i) occurs only if D = D R (x) = ± °° . We can deduce case (i)
from case (ii) with
G(y)
e > 0 in R
D = D R ( x ) , the cut of R defined by the
order 'p C K , is transcendental, then we can find
Proof:
such that for all
-f-signs
a(a) < 3(a)
in
of the R
G i (a(x)), G i (3(x)).
and so as to insure that
179
a
is not a root of any of the numerators or denominators of the rational
functions R(x).
a(x), 3(x) , or the coefficients of the polynomials
G.(y)
over
Finally, and this is the most important restriction, we insure that the
conditions on a guarantee that coefficients
a
is not a root of any of the leading
f n .(x) of the minimal polynomials
F.
for the elements
x.,
1 £ j <_n. With all this accomplished, we now start over with either order '$ a, + or '£
on
R(x), and consider the Sturm algorithm for real roots of
in the interval
(a(x), 3(x)) G R ( x ) .
place defined by our new order 'p a,+ pCG^cxCx))) = p(Gi(a(a)))
and
If p:
R(x) •* R, ± »
or 'p a,-
on
G(y)
is the signed
R(x), then, by our choices,
p(G/(3(x))) = pCG^GCa))) G R
since
p(x-a) = 0.
The signs of terms which occur in our new Sturm algorithm thus coincide exactly $*-Sturm algorithm for G(y).
with the signs which occurred in the original
Thus G(y) has real roots over R(x) in (a(x), 3(x)) with respect to either order •£ or '£ , and this order on R(x) extends to K. a, + a, — Finally, since 0 = £ .(x)xJ +••• +f (x) G K, f± . (x) G R[x] and ^(a ) t °> fu, n j
i1:
Remark.
follows trivially that
x.JKG A D C K, hence
Because of our careful choice of
additional conclusions can be drawn.
in the proof above,
Note that the signed place
associated to our order '$ or 'ft on a, + a, ~* sequence
a G R
A = R[x1J...x Jn CA DK. D .
p: R(x) + R, ± °°
R(x) has the property that the
p(GQ(y)),...,p(GR(y)) = p(h(x)) = h(a) t 0, is actually a Sturm
sequence over
R
for the polynomial
G(a,y) = Eg.(a)ym " 1 .
This is not
completely obvious because construction of a Sturm sequence involves division, so our choice of
a
coefficients of the the polynomial
avoiding roots of numerators or denominators of the G.(y)
G(a,y)
over
is essential here. R
has no multiple roots.
little further care we could arrange that roots in
R
as
G(y)
or *jT, and that the G(y).
has over
In any event, since
G(a,y)
In fact, with a
has the same number of
R(x) with either of the orders '£
At these points
180
in
R
, '$
R-roots of
G(a,y) are "close" to the R(x)-roots of 2 (a,b) G R on the curve G(x,y) = 0, we have
(d/dy)G(a,b) ^ 0. Thus we have a special type of simple point: G = 0
h(a) £ 0,
crosses the line
x = a
transversally.
the curve
,
Examples.
Consider the curves below:
xy - 1 = 0
y -x = 0
(b)
(a)
y 2 -x 3 +x 2 = 0
(d)
(c)
(e)
(f)
y 6 -2xy 3 +x 2 -x 3 = 0
(g)
In each case we have an irreducible ring
G(x,y) = 0, an affine coordinate
A = R[x,y]/(G(x,y)), and a function field
K = R(x,y).
We study the
181
behavior near
x = 0, to illustrate certain aspects of the proof above.
Example (a) illustrates that p:
R(x,y) •> R, ± «
curve.
If
with
(R(x), '£n
(R(x), fQ^J
)
R(x,y) may admit a signed place
p(x) = 0, but with no points is extended to
(0,b) G R
R(x,y), then
on the
p(y) = °° and if
is extended, p(y)
Example (b) has the origin as a simple point, but only the order (R(x), :j3
)
extends to
cross the line
x = 0
R(x,y)
since
> 0.
transversally, but does cross
The curve does not x = a
transversally
a G (0,e).
for
3 2 In example (c) , G(x,y) = y -y -x
specializations a>0,
x = a, small
y - y - a
has three real roots over
(R(x), !pQ
a.
If
a < 0,
y -y
has one real root, and if
-a
has three real roots,
a = 0,
y
-y
has a double
root and a simple root. Example (d) has a zero at the origin, but D R (x) « 0.
total order with
Since
R(x,y)
does not support a
G(0,0) = 0, we get a homomorphism
p: A = R[x,y]/(G(x,y)) •> R, hence (x,y) C A is f -convex. However, (x,y) C A is not ('Cw ) -convex. If it were, the signed place extension d theorem of 7.7 would imply the existence of a total order on
K = R(x,y),
with signed place
Note specifically
that
2
pR:
K -»• R, ± °° extending
2
x (x-l) = y , hence
(x,y)
is not
x-1 G ('£ ) w d
Cpw) -convex.
or
p:
A •*- R.
1 < x rel ($ ) . w d
The conclusion
relative to any total orders on
x-1 £ ($w)
K = R(x,y)
we must have
Clearly, then
also shows that 1 <^ x, which
agrees with the picture. In both examples (e) and (f), R(x,y)
supports total orders at the
origin, but the origin is a singular point. extends to
R(x,y)
does not extend. in two ways to
in two ways (with
In example (e) , (R(x), 'pQ + )
y > 0
In example (f), both orders
or
y < 0 ) , but
'pQ + , 'pQ _
on
(R(x), 'pQ _) R(x)
extend
R(x,y).
In example (g), (R(x), 'pQ + )
extends to
R(x,y)
this case, it turns out that there is no element of over
in two ways. R(x)
In
between the two
real roots of
G(x,y)
y = f(x)/g(x)
lies between the two infinitesimal pieces of the curve
182
)
(R(x), 'pn ) . This behavior is reflected by the u, +
but only one real root over
if
x = y
(R(x), 'pQ + ) .
(Geometrically, no graph of
G
(x,y) = 0
in
at
(0,0).)
(R(x), 'pQ + )
The Sturm algorithm for any interval
will yield either
2
roots or
(a(x), 3(x)), depending on whether or not
R(x)
roots for
0 G (a(x) , 3(x)).
when we specialize to small positive values This example shows how
0
(a(x), 3(x)) G(x,y)
in
Of course,
x = a, these two roots separate.
is not order dense in its real closure (see 7.4.4).
The proof of 8.3.1 leads to a rather nice picture of the real closure of the field
R(x)
with any order
Also, all the orders
$
a,i
$.
If
look the same.
The points in the real closure of polynomials
G(y)
over
x
Thus we look at only
(R(x), 'Pn
)
by
1/x.
pn and u, +
are roots of irreducible
R(x). We can think of these roots as infinitesimal
"connected" pieces of plane curves x-side.
f = 'p+oo, we replace
G(x,y) = 0, near the y-axis on the positive
One point is greater than another if its graph is above that of the
other, infinitesimally near the y-axis.
A point is finite relative to
its graph meets the y-axis, and is infinite relative to asymptotic to the y-axis. relative to
R
if
R if its graph is
The elements which are infinitesimally small
are those graphs through the origin.
The addition and multi-
plication corresponds to addition and multiplication of functions on The case of the transcendental order First, the real closure of
R
(R(x), $ D )
'$„ c R(x)
(0,e) C R.
is slightly simpler.
is actually Archimedean over
R.
We visualize points as infinitesimal pieces of graphs of plane curves
183
I Cut D I
I "crossing" the cut at
x » D.
graph can be extended to cuts of
R.)
closure of
(The "values" of the function defined by the
x = D.
In the case of
These "values" are necessarily transcendental
'£n, R
is actually order dense in the real
(R(x), £ D ) .
The notion of "connected" piece of a curve, together with rigorous formulations and proofs of the statements in the preceding two paragraphs, would require some effort to develop. intuitive discussions only.
After
Thus, we mean these paragraphs to be
all the machinery of this Chapter VIII
is developed, for semi-algebraic sets of all dimensions, it would still be a respectable project to investigate curves carefully.
This study would
include a detailed analysis of singular points, especially branches at a singular point, and a classification of orders on any function field in one variable, extending the results of 7.5 (which is the special case
R(x)).
8.4. Signed Places on Function Fields In this section we establish the existence of nice signed places on real function fields, in order to study prime convex ideals for various orders on integral domains of finite type over P C R[X 1 ---X n ]
be a
R.
A.
Thus,
K
K = R(x 1 -..x n ), the fraction field
is a real function field.
Proposition 8.4.1. (Lang) with
Let
'^-convex prime ideal in the polynomial ring, let
A = R[x 1 -.-x n ] = R[X 1 -..X n]/P, and let of
We begin with the weak order.
If
K = R(x 1 -..x n)
is a real function field
tr.deg.D(K) = r, then there exists a sequence of signed places over R
184
R,
K - > K 1 , ± ° ° ^ K 2 , ± oo -» * •. -*K r ,± oo, with each tr.deg.D(K.) = r - i . K 1
(Thus, K = R.) r
ring of the composite where
p^:
A = R[x,...x ] .
hence
IC
Moreover, if
a function field with A. C K l
is the valuation
K -*• K.^, ± oo , then we can arrange that
In particular
p | : A -*• R
(a-^-.a ) = (p ( x ^ ,.. . ,p (x )) £ R
n
A CA.,
is a homomorphism,
is a zero of the prime ideal
P C R[Xr..Xn].
Proof:
The proof is by induction on
follows from 8.3.1. of
K
over
K,
F CT K1
Thus
r > 1, assume
R, so that
R(x^...x ) . of
If
K
the real closure of
r = 1, the proposition
{x,...x }
1
p:
K
K
and let
L
C F
Let
p 1 = p| :
denote the real closure Kf = F(x r...x n ) C r.
Let
tr.deg.D(K,) = r-1
with
F, such that
K -• F, ± «> . since
Remark.
J
J
—
However, when we
K
K
2
~*K
± ©o is dropped.
l'±O°
i s a t
least
K
2
-• K^
using
place if the distinction K
, ± oo relative to this "new" order on
Continuing, we work backwards until all
reordered so that all
a
K _^
Using Krull's Theorem 7.7.2, we then reorder
and recover a signed place .
r-1.
K, -• K^, ± °° , a new order
K -* K.^, ± °° -••••-• R, ± oo , then we order
K
—
•
ultimately arrive at
between
is a
The proposition now follows
may need to be chosen, as in the proof of 8.3.1.
Now
p1
1^.
In continuing the proof to obtain
, -* R, ± °° .
is
of transcendence degree
(See [68], Vol. II, Chapter VI, §14, Theorem 31.) by induction applied to
The image of
±
R
p(x.) G F
F.
p,(x.) = x., 1 < j < r-1.
K l
K, is actually a function field over
K,
r
R(xr . . x r l ) .
-* F, ± oo t over
1
on
is a transcendence base
is a function field in one variable over the real closed field
finite, r <_ i <_ n.
Thus
If
is algebraic over the pure transcendental extension
Choose any total order on
Using 8.3.1 we find
subfield
r.
K., including
K = K , are
K^ -• K^ +1 , ± °° are signed places.
Our proposition has many consequences.
The first is the following "weak"
weak Nullstellensatz.
Corollary 8.4.2. ideals
The maximal
(X^ - a^,...,X - a )
'.J3 -convex ideals of
corresponding to points of
R[X ...X ]
are the
n
R .
185
Proof:
In fact, any
f -convex ideal w
P
'•]) -convex ideal
and if
there is a chain of prime
I
is contained in a prime
r = tr.deg.n(R[X., .. .X ]/P), 8.4.1 shows that K i n '^-convex ideals
(Xx - a 1 ,...,X n - a n ) C R [ X 1 . . . X n ] .
P = PQ C p
C . .. C p^ =
Namely, P^^ = kernel (pi| :
A -» IO) , in
the notation of 8.4.1.
Proposition 8.4.3. transcendence degree
r
Let
K = RCxj.-.x )
over
R.
Suppose
be a real function field of
finitely many elements positive in some total order on signed places over
R,
p:
g. i- 0, are
gi>«-«>g^ ^ K, K.
Then there exist
K ~* R, ± °° , with rank(p) = r,
p(x.),p(g.) E R,
p( g j ) > 0.
and
Proof:
Apply 8.4.1 to the real function field
E = R(x.,y.,z.) l
j
3
where
y]-zy Yfr-i. Remark.
n
An alternate proof of 8.4.3 could be given as follows.
as in 8.3.1, first any order extending one of the orders
!|5 C K 'p
+
If
r = 1,
is chosen, the replaced by an order
C R(x^) C K, by a careful choice of
a E R.
This choice can be made exactly as in the proof of 8.3.1, so as to preserve the signs of the relative to 8.4.1
R.
g.
and keep the
g.
finite, but not infinitesimally small
This technique is now combined with the inductive proof of
to give 8.4.3. Using either method of proof of 8.4.3, the conclusion can be strengthened
to include an approximation statement. defines a signed place The new order on
K
pR:
Precisely, the original order on
K -* A, ± °° , where
A
is Archimedean over
corresponds to an R-valued signed place
The conclusion of 8.4.3 can then be extended to say that the approximate the
p R (g-) E A, ± °° as closely as desired.
this just means
p(g.)
p:
K
p R (g.) = ± °° ,
can be made large.)
The next corollary constitutes a "weak" strong Nullstellensatz.
Corollary 8.4.4. g ^ P.
186
Let
P C R^.-.X ]
Then there exist zeros of
P,
be a prime n
a E R , with
'p^-convex ideal,
g(a) ^ 0.
R.
K -• R, ± °° .
p(gO ^
(If
K
Proof:
g / 0 £
R ( x r . . x n ) = K, the fraction field of
Choose a total order on
K
R [ X r . -XJ/P.
and apply 8.4.3 to the single element
g
or
-g,
whichever is positive.
Remark.
•
Corollary 8.4.4 is exactly the basic Nullstellensatz stated
previously as Proposition 8.2.1. established.
For example, if
n
{a e R |h(a) = 0 all of
a €= z(I)}.
all
Thus all the consequences of 8.2.1 are now
I CR[x i ...X n ]
h E I}, then
is any ideal, Z(I) =
*ti(I/Pw) = {f € R [ x r . . X j |f (a) = 0
If we combine this result with the characterization 2.2.4
vfa(I,pw), we conclude that the ideal of functions which vanish on
consists of exactly those
f
such that
Pi,q. E Rp^.-.X ] , h. E I.
Z(I)
S
f + Z p . = Z q . h . , for some
s _> 1,
This result, or similar versions, is known as
the Dubois-Risler Nullstellensatz in the recent literature, [14], [16], [17]. We point out that our proposition 8.4.3 as well as Corollary 8.4.4 date back to the early work of Artin, at least in essentially equivalent forms.
The
other ingredient of the Dubois-Risler Nullstellensatz is then the characterization of the radical of a hull for general partially ordered rings, given in 2.2.4. In fact, little extra work is now required to characterize those functions f E RfXj.-.X ]
which vanish on a basic closed semi-algebraic set of the form
Z(I) H w{g.} C R n . V = $„[g-] w j
is
First, assume
an order on
is an
RHJ-algebra,
f E A
which vanish on
Z(I) H W{g.}
is a radical ideal so that
and assume
J C I.
D Z(I) Hw{g,}
f
s >_ 1
...g
g X
l
X
= r
E(k2 J
l
and polynomials
allow the "empty product"
f E /H(I,^) C A.
vanishes on
1
q
g 3
...g 1
J
h,k,y,z, with
of the
g.
) 2 q
J
y GI
l'"\
and
z E J.
I(W{g^».
Certainly,
'^[gj]
(We
here.)
Secondly, it is not hard to see that the smallest such ideal exactly
(A/P)
and the functions
are exactly the functions
f E R[X....X ] , we conclude
Then
if and only if there is a formula
f2s+Zh2 1 l*-' 1 r
for some
A = R[X,...X ]/J in
X(A,$) = Z(J) Hw{g,}
In terms of polynomials Z(I) n w { g . }
J C R[X1 ...X R]
is an order on
J
is
R [ X r . .X n]/I(W{g.}).
187
If
J
is any radical ideal with this property and
vanishes on if
X ( R [ X r . .Xj/J, '^w[g. ]) , hence
J = I(W{g.}), then
h G I(W{g.}), then
h = 0 eRp^.J^/J.
Z(I) n W{g.} = Z(I + J) n W{g.}, and
The two paragraphs above reduce the determination of to the determination of all
I(W{g.}).
j , then, of course, I(W{g.}) = 0.
Even if
n
R , in a given case we may know an ideal
R[X,...X ]/J.
J C I
W{g.} with
Moreover,
J C I + J. I(Z(I) nw{g.})
x G Rn
If there exists
h
with
g.(x) > 0,
has no interior in '$ [g.] w j
an order on
It would then be unnecessary to actually determine
I(W{g.}).
Next we state a weaker form of 8.4.3 which is as useful for most purposes. Corollary 8.4.5 •P [gi'**gv]
a
(Artin)
A = R[x....x ]
be a real integral domain
finitely generated refinement of the weak order on
Then there is a homomorphism
Proof:
Let
Extend
p:
'^w[g^]
A -• R
with
p(g.) > 0 ,
to a total order on
A
g. ^ 0.
1 < i < k.
and apply 8.4.3.
In particular, this result applies to the polynomial ring In fact, we deduce the following, also due to Artin.
A,
•
R[X,...X ] .
(Part (b) is Hilbert's
1 7 t h problem.)
Corollary 8.4.6. (a)
g. G R[X 1 ...X n ]
If
is an order on
R[X 1...X n ]
are nonzero polynomials, 1 <_ i •< k, then a G Rn
if and only if there exists
with
'^[g^] g i (a) > 0,
1 < i
The polynomials
^, J
Proof:
188
some
nowhere negative as functions on
h,h. 6 R [ X .L . . . Xn J , h ji 0 } . J
The "only if" statement in (a) is 8.4.5,
true because the set nonempty.
f G R[X^...X ]
The "if" statement is
U{g.} = {a|g.(a) > 0, 1 <_ i <_ k}
is Zariski dense if
Similarly, a function
f(a) < 0 around
and
a.
f G ('|5 ) w d
h f = Eh. , then
Conversely, if
h
is clearly nowhere negative since if
would vanish identically in some ball
f £ (:pw) , then
'^w[-f]
is an order on
R[Xr..X
and (a) applies.
Q
Remark 8.4.7.
The "if" statement of 8.4.6(a) is not true for arbitrary
finite real domains p:
A -• R
point
with
A = R[X ...X ]/P.
p(g i ) > 0
That is, the existence of a homomorphism
does not imply
a = (pCXp,...,p(X n ))
$ [g^]
is an order on
could be a "degenerate" zero of
Example (d) of 8.3.
The argument of 8.4.6(a) shows that if
is Zariski dense in
X(A,'£ ) , then W
Z(P) C R n , the zeros of * w [gil
c A
is an order
is an order on that
h
M O G A ,
A, since
Remark 8.4.8.
A is a domain.
as in
U{gi> n X(A/}5w) Here
X(A,'B) = W
Conversely, if or
^[g^-h]
In either case, 8.4.5 guarantees U{g i> n X(A,'£w).
Thus
U{gi>
is
C Q> ) C CP ) C cp ) = (B ) =?(R n W WS Wp W m W(J In the next section we will prove a
Consider the orders R[X^...X n ].
theorem of Stengle which implies distinct orders here, f
w
C (tp ) w s
V
('£ ) = '£(R n). Thus, there are only three w P C'^(R n ). These inclusions are definitely
For example, Hilbert [ 1 ] gave an example of a strictly positive f(X,,X2), which is not a sum of squares.
polynomial in two variables, proof that
(£)
= *p(Rn)
W P polynomial belongs to
hand, by definition, term of
^[g^h]
The
X(A,'p ) .
on the polynomial ring
proper.
then either
does not vanish identically on
Zariski dense in
is an order.
U{gt> = {b G R n |g ± (b) > 0}.
P, and
and
'^1T[g.] C A W 1
P
A.
f
(p ) . w s Q3 ) w s
Thus
'B / (? ) , if n = 2. On the other w w 9 9 = {f|(l + Eh.)f = Z g . } . Immediately, the homogeneous 1 3
of lowest degree must be a sum of squares.
Hilbert's example
f(X,,X2)
But now if, say,
is made homogeneous, F(X ,X.,X 2 ), so that
f(X 1 ,X 2 ) = F(1,X 1 ,X 2 ), then hence cannot belong to
Our
will show, in fact, that any strictly positive
F(XQ,X;L,X2)
Cp ) . w
s
Thus
cannot be a sum of squares and
Cp ) w
f f-P(Rn)
if
n = 3.
s
The last result in this section is a signed place perturbation theorem for function fields.
The statement and proof are similar to 8.4.1, but
instead of trying to make elements finite relative to a discrete rank
r
189
signed place, we assume certain elements infinitesimally small relative to one signed place, then try to keep them that way relative to a discrete rank r
signed place.
This result will be the basis of our study of derived orders
in 8.6 and dimension theory in 8.10.
It provides a tool for studying the
geometry of a semi-algebraic set near a point, or more generally near a semi-algebraic subset.
Proposition 8.4.9. R
with
p':
Let
tr.deg.R(A) = r.
A = R[x,...x ]
Let
K
be the fraction field of
K ""* A, ± °° is a signed place over
p'(x^) = 0.
(Necessarily
p
1
R, with
signed place
Proof: p1
hence
p:
If
A
A
K -• R, ± °° with
and suppose
Archimedian over
= p R , relative to the order on
This follows easily from results of 7.6.) r
be a finite integral domain over
and
induced by p 1 .
K
Then there exists a discrete rank
p(x^) = 0.
r = 1, there is nothing to prove since we assume
is already discrete rank
R
1
and R-valued.
If
p'Cx^) = 0,
r >_ 2, we prove the
proposition by induction. By the Noether Normalization Lemma we can find of the
x., say
p'(u.) = 0.
u,,...,u , such that
A
r
linear combinations
is integral over
Thus we may as well assume that
A
R ^ . - . u ].
Certainly
is already integral over
R[xr..xr]. Let x
.,
f.(x
.) = Ea..(x, ...x )x
j > 0, with coefficients in
polynomials, hence applying polynomial
pf
. = 0
be the minimal polynomial for
R[x,...x ] .
we get that
Za i -(0...0)T 1 £ R[T].
Choose
The
0 £ R
0 < e G R
than the absolute value of all other roots
{a,}
f.(x
.)
is a root of the nontriviat such that
of the
L
be the algebraic extension of
j > 0, and replace on
K
defined by
namely, take
p
hence replacing rank
r
place of
190
1
A p'
extends to R C L.
K
L
by
p:
L.
is less
/e 2 - x z + . ,
y. = / e 2 - x 2 + . - e.
In fact, the place
p
1
The order
extends to
The residue field is Archimedian over
does not change our hypotheses.
K = L.
L,
R,
Also, any discrete
L -• R, ± °° restricts to a discrete rank
Thus we may assume
over
IJ
obtained by adjoining
R[x^,y.], where
= pn,
signed place K.
by
K
e
Ea..(0...OjT1
K
Let
are monic
r
signed
R.
With this assumption, given any signed place with
p(x i ) = 0 ,
1 <_ i £ r, then necessarily also
The reason is that interval
Let
0 G R
(-£, e)
Let
F
F C r
is the only root of the
p(x + .) e (-e, e)
and
be the real closure of
y
x
x.
by
-x.
F(x )
Dp(x r) = D
of
F(x ,y) of
F.
in the
in
K. p1.
R(x 1 -..x r 1 )
over
of
K.
Write
F(x ,y)
is a
F, totally ordered as a subfield x^, 1 <_ i <_ r, are infinitesimally
We may assume the
x^
positive, be replacing
induces an order on the simple transcendental
This order is completely determined by the cut
F, as in 7.5.1.
regarded as a subset of D
2
if necessary.
The order or extension
R.
x +. < e
R(x,...x ) . Then
With respect to this total order, the
small relative to the subfield
Ea^^. (0- . .OjT1 2
R
j > 0.
relative to the order defined by
algebraic over
function field in one variable r.
K
K -*• R, ± °° over p(x + .) = 0,
since now
be the real closure of the subfield
K = R(x,...x ,y), with
of
p:
F.
We have
Now, D
0 < D, that is, 0 E D
if
D
may be algebraic or transcendental.
is algebraic, we will not yet perturb the order on
F(x ,y).
If
D
transcendental, we appeal to Proposition 8.3.1 to choose an element 0 < 3 < D, so that if
F(x ) r
is reordered over
as in 7.5.1, then this new order extends to we still have
xr
is
F
infinitesimally small relative to
Since R C F.
is
3 6F,
with either order
F(x ,y).
If
!p , p,+
0 < 3 < D> Of course, the
x., i < r, are also still infinitesimally small relative to R since F has not changed. We now consider the signed place (possibly) new order. F-valued.
We have forced
on
F(x ,y), with respect to our
pp
to be non-trivial, thus
In fact, we have forced the restriction
to be non-trivial.
pp:
(Specifically, p p (g(x 1 ...x r )) = 0
is the minimal polynomial for trivial on
pp
R(x,...x
) , hence
3 ^ F
over
pp:
R[x,...x
pp
is
R(Xj. . .xr,y) = K + F, ± °°
where i]«)
gCxj...x r_1,3) = 0 But
p
is still
K -» F, ± °o is a discrete rank 1 signed
place. The image degree
r-1
p p (K) = K' C F over
R.
is thus a function field of transcendence
In fact, the proof of this result in, say, [68, vol.11,
Chap. VI, §14, Theorem 31] shows more. closure of
A = R[xi ]
in
K.
Then
Let 1
K
A' = R[x.,z.]
denote the integral
is the field of fractions of
A'/P',
191
P1 C A 1
where
is some minimal prime ideal. 1
by the inclusion integral over
K
A.
C F, the
z.
With the order on
are finite relative to f
Thus we can write
K'
induced
R, since f
z.
is
x| = x.(mod P 1 )
A /P' = R[x!,z!] C K , where
and
z! = z. - p (z.)(mod P'). By induction, applied to the signed place
pR:
K1 -> A 1 , ±°o,
signed place
A1
K1 -*• R, ± °° , with
q:
We now compose altering
pp
Archimedean over
pp
and
q.
r-1
q(xp = 0.
K -* K', ± °° -* R, ± °° , and reorder
K
K!
induced by
pp
We end up with our desired rank
Remark.
Here is an extension of
r
is not the same as that induced signed place
p:
K -> R, ± °°.
In the notation of 8.4.9, suppose given finitely
many elements
g. ± 0, which are positive in the order on
g. G A ,
p1 :
K -*- A, ± °° .
statement that the discrete rank
r
g.
are also positive in the order induced by the
signed place
p:
is the composition of the first gj
K
Then we may add to the conclusion the
K -* 1^, ± «>->... -> K p r ^Cg^) t °» where
and moreover, we may arrange that
If
•
8.4.9 which is useful in studying
geometry near a point.
induced by
by
on infinte elements, as in 7.7.2, to take into account the
fact that the order on by
q,
R, there is a discrete rank
r-1
g.
P r j_ •
±°°->R, ± ° ° , K
~*K r y
±
°°
maps in the chain.
£ ( x r . . x n ) C A = R [ x r . . x n ] , then already
are various ways to keep the
^
p( g j ) t 0.
There
positive relative to our new order on
K.
For example, using the technique of the proof of 8.2.10, one can adjoin certain P'(g-)
/gT =e
to
K, then replace
-^R>
0 < c.
there is only a single
so that
g
{x^.-.x^,}.
R[x,...x , Jg7 - £ . ] , where
This technique works whether or not
Thus we can concentrate on the case
just try to arrange that
formula for
by
(See also the proof of Proposition 8.6.5 in a
later section for this technique.) g. €= (xj.-.x^.
A
p g.
,(g-) £ 0.
Setting
g = Ilg., we may assume
Now, we just add the dummy generator
g
to our
A = R[x,...x ,g]. Then we might as well rename the generators becomes
x^, the first element of the transcendence base
If one inspects the proof above, one finds
so that at the crucial inductive step constructing we obviously have
192
g. G (x-^.-.x^, and
p F (g) = g / 0.
pp:
g = x^ & R(x 1-. .x r _ 1 ) F(x ,y) ~* F, ± °° ,
Our extension of 8.4.9 is thus proved by
induction.
(Note that, again, if
r = 1
there is nothing to be proved.)
8.5. Characterization of Non-Negative Functions In this section we prove the theorem of Stengle [22], characterizing the functions nowhere negative on a closed semi-algebraic set. Begin with a reduced, real R-algebra
A
of finite type, and a finite
refinement of the weak order ty = p [g.]. w (Stengle).
Proposition 8.5.1 X(A,'p)
j
An element
f E A
if and only if there is an equation
(f
n
is nowhere negative on
+ p)f = q, some
n :> 0,
p.q eVProof,
The "if" statement is trivial.
We prove the "only if" statement A 1 = A[t]
by applying the Nullstellensatze of 8.2 and 8.4 to the ring polynomials in one indeterminate over Let on
A
f
= f [g.] C A'.
We know
extends to an order on
simply a homomorphism follows that
f
x:
!
X(A ,^ )
A1
A -* R
of
A. '.|5! is an order on
by 6.1.
A homomorphism
together with a value
is identified with
Af
since any order xf:
A1 -> R
x'(t) E R.
X(A/p) x R.
is
It
(This we also dis-
cussed in 8.2.) Suppose
f E A
is nonnegative on
t2+ f
on all zeros of
f E /H(t z + f/p1) C A 1 .
in
X(A',^').
X(A,'p).
Then
f €ACA'
vanishes
Thus, by the Nullstellensatz,
From 2.2, we obtain an equation
P(x,t) = Q(x,t)(t2 + f(x))
for some
n > 1,
functions on
P(x,t) E:p'.
We are regarding elements of
G(x,t) E A' can be uniquely written
G (x,t ) + tG,(x,t ) . Applying this decomposition to P(x,t), Q(x,t)
by
F(x,t2) = 0 E A' a E A.
as
X(A,'p), X(A',^'), respectively.
Now any element
any
A,A'
2
P Q (x,t ), QQ(x,t ) , respectively. implies
F(x,t) = 0 E A1
Thus we obtain a relation in
and thus
G(x,t) =
(*), we can replace But any identity F(x,a) = 0 E A,
A,
193
f 2n (x) + P Q(x, -£(x)) = 0.
(**)
Now, we investigate
P Q (x, -f(x))
more closely.
Since
P(x,t) ^ ^ ' ,
we can find a formula
P(x,t) = EPj(x,t) gj (x),
Specifically, the
gT j
gj £'j5CA.
are finite products of the
g., where 3
'p = '£ [g.] C A. w j
Thus
P Q (x,t 2 ) = (Pj )0 (x,t 2 ) + t 2 P^ 1 (x,t 2 ))g J (x)
by writing
PT(x,t) = P T
by
) +tP
-,(x,t ) , expanding, and using uniqueness 2
of the decomposition t
ft(x,t
P(x,t) = P Q (x,t ) + tP,(x,t ) .
Replacing
t
Thus (**)
f n (x) + p(x) = q(x)f(x)
gives an equation
We obtain 8.5.1 by multiplying this last by
CoroHary 8.5.2. '-P = '-P [g • ] w j
a
and
f)g(x)) j
p(x),q(x) G'^.
f(x).
•
be a reduced R-algebra of finite type,
Suppose
g GA
A,
vanishes on
f £ A
a function
Z(f) H X(A,f) , so 0 £ g2n <_ qf (rel
Then, there is an inequality
q£'|J.
Proof. = A[t]
A
X(A,']>).
g G v^H({f}, $) C A.
for some
A
Let
with
finite refinement of the weak order on
nowhere negative on
1
t
- f(x), we find
PQ ((x, -f(x)) = (Z P x f)g(x)) -f(x)(E f(x)(E Pj(# 1x Pj() 0 (x, -f)gj(x))
that
by
This is a consequence of the proof of 8.5.1. again, and
'£' = ? w [g k l
c
1
A .
Then
We introduce
g e >^M({t^ + f}, f1) C A 1 .
Going through the proof of 8.5.1 8.5.] yields an equation
g
+ p = qf,
p,q
which is exactly our assertion.
Remark. the hypothesis
Corollary 8.5.2 should be compared with 2.2.3. f £'|5
of 2.2.3 by the weaker hypothesis
f
We have replaced non-negative on
X(A,']>), and have established the same conclusion in both 8.5.2 and 2.2.3.
194
Corollary 8.5.3. '£
=
a
'P [§•] w 3
function
Let
A
be a reduced R-algebra of finite type,
finite refinement of the weak order on X(A/p).
strictly positive on
1+p = (l+q)f, some
Proof. setting
a
p,q ^='p.
But now
p'
(l+q')f = p f ,
is also strictly positive on f
8.5.2 again gives
f €= A
Then there is an equation
Corollary 8.5.2 gives equations
g = 1.
A, and
1+p" = q"p ,
p",q" £'p.
p! ,q' e 'P X(A,'.p).
by
Applying
1
Now check that
(1+q )(l+q")f =
1 +p' +p".
D
Note that the equation e
CA
*s(i)
Let
1+p = (l+q)f
A
be a reduced, real R-algebra of finite type, fp = 'p [g.] w j A.
is an RHJ-algebra and that the functions coincide with
the orders
consists of all functions
(Y - X 3 )
(p/I)
A
I
C A/I
n >_ 1, is
p,q G ])}.
A = R[X,Y],
wil1
2
f G A
I = (Y - X ) , then
R 2.
congruent to
(Consider
3f/3X
f £ (*p /I)
then
(l+p)f = q, some
t
f £ A
P
er
is nowhere X
modulo
at the origin.)
('P /I)
is easily seen to be non-
C (p/I)
by Stengle's theorem.
is a reduced real R-algebra of finite type and
the ring of functions $ = 'Pw[t5]>
X
P
r0
= CP/I)p> since, first, 'P/I ° P / I , which gives
2(1) C x, which gives
with
be a
3
C A/I
On the other
of the functions
a finite refinement of the weak order, then the order all
By 8.5.1, fly I)
Z(I) C X.
Z(I)
f (X,Y)
Suppose
!p -convex by 8.2.2, and
In general, ']) /I C CP/I) p
X.
(A,'.p)
X = X(A,'p)
1
are defined.
Z(I) C R , but no polynomial
Cp/I)
nonnegative on
non-negative on
Cp /I) , and, secondly, any
negative on If
Then
can be non-negative on all of
We do have c
f G A/I
For example, if
negative on
f £ A
consists of the restrictions to
non-negative on all of inclusion.
CP /I)
a
We have now established that
+p)f = q, some
^-convex ideal.
Cp/I) , 'p / I ,
'p /I
n
•£ = {f G A|(f
is a radical
hand,
f = 1+p/l+q
s(i)-
finite refinement of the weak order on
I CA
can be written
A
'$ = 'p [g.]
is
'P C A, consisting of
p,q G'p, is not a geometric invariant of
on the set
X(A,'p).
For example, if
A = R[t] ,
^^s"
On the other hand, if
f
is strictly positive on
X(A,'p), then 8.5.3
195
guarantees that
f G'j5 .
contraction to
A
Also, recall that
of the order
'pcfl. C A
characterized as the ring of functions on all
f G A
with no zeros on
X.
C f n
.
Now, A c f 1 . can be invariantly
X = X(A,'p)
f
pG'P,
is perhaps a more natural geometric invariant of '£ by
£ ( X ) , then
$(X)cn-\
(A(X) c r i ., $ ( X ) c r i O
than the ring A
X
( )CM^ X.
is
Ag^.
A.
If
invariantly
Recall from Chapter V
Stengle's theorem has a nice interpretation
in terms of the structure sheaf associated to X
The ring
is the ring of global section of the structure
sheaf associated to (A(X), '^(X)).
nowhere negative on
X, then
g G A.
X c
characterized as the functions nowhere negative on that
obtained by inverting
is nowhere zero on
and by 2.2.4, 1+p = fg, some
we replace the order
is defined as the
This is immediate from 8.5.3, or directly
from the Nullstellensatz, since if 1 G H({f}, p)
'p C A
if and only if
(A/p).
Namely, f G A
is
f G ' P g ^ C Ag, f . . That is, f
be "positive" in the ring of sections over the basic open set
should
D(f) C X.
Stengle's theorem also generalizes Artin's result 8.4.6(bJ, in the case A = R[X X .. .Xn] , p = £ w . nowhere negative on Eh^, some
h^}.
Namely, it is trivial from 8.5.1 that if
n
R , then
$ ww)) . f G ($
f
is
2
($ w : f) = {h G R[X 1.. .XR] | h 2 f =
Let
It would perhaps be interesting to characterize the zeros
Z(pw : f) = Z(/H(Cpw: f) , ?J) C Rn.
By 2.2.3, if ZCPw : f) = 0 , then
and conversely.
By Stengle's theorem
Z('^ ' f) C Z(f).
s t include all zeros
x
Z($ w : f)
that
mu
of
of lowest degree in the Taylor expansion of
f f
f G Cp w ) g ,
It is also easy to see
such that the homogeneous term about
x
is not a sum of squares.
8.6. Derived Orders Let
A
be a reduced, real R-algebra of finite type, 'J5 = '-P.Jg-] w j
finite refinement of the weak order have identified
X
X = X(A,'^3).
some
n >_ 1, p,q ۥ]>}.
Also, if orders
196
I C A
In previous sections we
with the maximal convex ideals of
that the functions nowhere negative on
X
(A,'jJ) and established
coincide with
']3 = {f G A| (f n +p)f = q ,
In this section we study the derived order
is a radical
a
'^^
of 3,12.
p-convex ideal, we investigate briefly the
(1> / I ) d and, when defined, the orders
'^/I
and
C^d/I) •
All of
these orders are geometric invariants of the rings of functions X(A,p),
p,q G p
and
p
CP/I)^-
*4>,/1, when defined.
zation of
']3^ consists of those
not a zero divisor.
include, of course, the orders
on
Our study of
An equation
pf = q,
nowhere negative on where
f
with
pf = q,
'£,, ^ = ^^[g.]
(Hp.), = nCPO
Note that since
! d
we will also have described
U(-f) C x
f GA
will
Also, we will then have a characteri-
Q
f
A/I
X(A/I, 3 V I ) , respectively.
Recall that the derived order some
A,
by 3.12,
! d
'£, for a finite intersection of orders of type
p,q £'.p,
p
X = X(A,'|5).
is negative.
not a zero divisor, does not imply
The function
p
could vanish on the set
We need to distinguish between "degenerate"
and "non-degenerate" points of our semi-algebraic sets. In general, let
(B/p)
is degenerate if for some and
be any RHJ-algebra, f,h G B,
y e U(f).
h(y) = 0, all
h
f
points of
X,.
X
by
with
(We will write
X = X(B,'^).
We say
not a zero divisor, we have h(U(f)) = 0.)
definition of degenerate point depends only on particular order
X = X(B/P).
X
and
x G X x G U(f)
Note that the
B, not on the
We denote the set of non-degenerate
The subscript refers to "derived" or "dense" for
reasons which will appear shortly. Given any point as follows. U C U' I
=
For each open
implies U
I(U)
h i (U(f i )) = 0. hence
I
U(f) C X If I in
x £ X
we can associate an ideal of degeneracy U,
x £ U, let
f
I(U ) C l ( U ) .
Since
I(U) = {h G B|h(U) = 0}.
Now, if
U(f) C U{f.}
^,...,11,, with
is chosen by 8.1.1, then
Choosing an affine embedding
as a small open ball in
R n , intersected with
is an integral domain, it is clear that
For general
the minimal primes.
B,
x G Xd
if and only if
This is equivalent to
I
Recall that the
order
CP.
We claim that
I x
I A = ( 0 ) C B/P . J
x £ X,
h.^ e I(U(f)),
x
if and only if is Zariski dense
I x C U p . , where
P-
if and only if
J
From 8 . 2 , Z ( P . ) = X ( B / P , , tp/p ) J J J
x ^ U ^ ) ,
X.
C P . ( some
argument in commutative algebra. '$ C B.
Clearly
X C R n , we can interpret
= (0), which just says that any open neighborhood of X.
C B
is Noetherian, the ideal
is finitely generated, say by
= I(U(f)).
B
B
I
P. C B
are
j , by a well-known
are convex for any x e Z(P.)
and
J
and
X(B/P) = U X ( B / P . , p / P . ) i
197
The "if" part of the claim is easy, since if some a neighborhood
U(f) of
x G X, then
U(f) H Z ( P . ) , contradicting 3
on
I C p., x J U(f)
f G p., 3
since
f( x) > 0.
hf = 0 G B.
h G B/P.
Thus
x G Z(P.)-
vanished on
is non-zero, but vanishes
I = (0) C B / P . . x 3
Then any element
h £ P.
Conversely,
h G n ^
P. - p. i 3
Finally, if
suppose vanishes on
hj £ P.
vanished
on a neighborhood
U(f) n Z(P-), choose h 9 G O p. - p and consider 3 I ± j . i 3 We have h £ P., h(Z(P.,)) = 0 , i / h, hence h(U(f)) = 0,
h = h1h2. since
U(f) = U (Z(P.) n u ( f ) ) . x i
prove
the following.
X(B,1>) , =
Proposition 8.6.1.
Proposition 8.6.2, (0) = n p
with where
U X(B/P.,?/P.) . d P i minimal
Suppose
and suppose
B -•IIB.
The arguments in these last two paragraphs
p. C B
is a finite collection of primes
'^ C B i = B/P^ are orders.
is the natural inclusion, then
1
Proof: U X(B., p . ) . ^ l l
x G X(B, '^) , a
If
x G X(B., '^.) , then 3 3 d
but
x £ X(B., 15.), let
X(B,'.|>) , = U d Pj minimal
U(-f) .
if and only if
f G B,
f(x) < 0,
Thus
U m
3 d
I C P., some minimal x 3
and, again, any
The proof that
just like the last step of the proof of 8.6.1, with the the
3
If
P.. 3
I C P., x 3
f (mod P.) G -jj. C B., then
X(B., '^.) i i
x G X(B., f.) . J j
X(B., ']>.) .
X(B, %]>) =
I C P., just as in the proof of 8.6.1. ^ J
xGU(-f)CX(B,f)-X(B.J.)C 3 3 vanishes on
If '£ = B nl!']).,
The starting point is 8.2.9, which asserts that We know
D
hG
n P. -P. . ^ 1 3
x G X(B., '^.) 3 3d
is
X(B., 'p.) replacing
Z(P i ).
•
Proposition 8.6.1 seems more natural than 8.6.2, since it applies to any RHJ-algebra.
But 8.6.2 is better since it applies more directly to the
situation of 8.2.11, where we showed how any closed semi-algebraic set is of the form
X(B, •]>) .
We next show how non-degenerate points behave with respect to intersections of orders.
Proposition 8.6.3. i £ i £ k, and suppose
198
Suppose
P C B
(B, 'JK)
are RHJ-algebras,
is a '^-convex prime ideal.
']$ = n-j^,
Let ]3. ,
denote
those
p.
'p..)
Then
1
so that
P
is
X(B/P, $/P)
Proof:
a
'£., convex.
(Thus if
P = (0), we get all the
= U X(B/P, ?.,/P) . i1 -1 d
First note that by 2.7 we know some
'p.,
exist.
If
x G X(B/P, .p. ,/P) , but some f G B, f £ P, vanished on a neighborhood U 1 d of x in X(B/P, $/P), then f also vanishes on U nx(B/P, f.,/P). This contradiction shows
U X(B/P, *.p.,/P) C X(B/P, 'P/P) . 1 d d i1
Conversely, if
Let P
f. , £ P is not
vanish on a neighborhood of
'£. ,,-convex, where the
follows that f.,, 4 P
P
x,
x £ X(B/P, p.,/P) 1 d
vanish identically on
Since
'.p.,, are the other orders among the H(P, 'p.M) C B.
if.
all
U(h. ,) C\ X(B/P, 'p.,/P).
is properly contained in the hull
From 8.1.1, choose f
x G X(B/P, 'p/P) , assume d
'p., it
Thus let
Z(P) n X(B, '£.„) = X(B/P, p/P) n X(B, 'p.,,).
x G U(h) G n U(h i f )
vanishes on the neighborhood
U(h)
and let of
x
f = II f±1- II f i M £ P.
in
X(B/P, '.p/P)
Then
since
u(h) = u [u(h) n x(B, 'p.,) n z(P)] U U [u(h) n X(B, £ in ) n z(P)] . i1
x
Remark 8.6.4.
•
i"
The study of derived orders of the type we are interested
in has already been reduced to the case of integral domains in Proposition 3.12.1. Thus, if
B
is a ring,
P. C B
a finite collection of primes with
and we have either
(1) an order
'p. C B. = B/P.
$ = B nil p., then
i
l
l
'P, = B n n('p.) d x
and
f C B l
with all
P.
convex, or
'P, = B O IlCp/P.) a
i d
(0) = n P^ (2) orders
in case (1) and
in case (2).
d
The next result is the central result of this section, and gives the relation between non-degenerate points and derived orders in a crucial special case.
Proposition 8.6.5. domain over on
A,
R.
g i t 0.
Moreover,
Let
Let
']3 = 'Pw[g-]
Then
(A, £ d )
p , = { f G A| f (x) > 0 a
A
•
—
be a real, finitely generated integral be a finite refinement of the weak order
is an RHJ-algebra, with all
x G X(A, '£),}.
coincides with the following subsets of
a
X(A, 'Pd) = X(A, '£) d .
Finally, X(A, 'p), also a
X = X(A, 'P w ), (the irreducible real
algebraic variety associated to A ) : 199
(i) {x G X |exists total orders p D p
(ii) {x G X|for all
g e A
with
on
A
g(x) > 0,
with
!p[g]
x
'^-convex}
is an order on A}.
Before giving the proof, we discuss some applications. the Proposition gives us a class of RHJ-algebras not at all obvious that
Corollary 8.6.6. 'P = V [g-1 Vd
a
If
B
for which it is
is a closed semi-algebraic set.
is any reduced real R-algebra of finite type,
finite refinement of the weak order, X = X(B/p), then
= {f G B|f(x) _> 0
Proof:
X(A/pj)
(A/p^)
Let
all
x G X d >.
P. C B
be the minimal primes.
and by Remark 8 . 6 . 4 , ' p , = B n i I C P / P . ) . a 3d
But
By 8.6.1, X d =UX(B/P., 'P/P.) ,
' p / P . = 'P [Wg . | 1g . £1 P . ] , and by J J
8.6.5, f G CJ5/P.) if and only if f is non-negative on J d Thus f G *.p if and only if f is non-negative on X,.
Next, let with
S
S i = Z{f ik >
in n-variables. primes above
n
W{g ik > C R , as in 8.1, where
Then
= R
I(S.)«
i a ia We also have the order p.
P.
FrOm
S 2Al
-
S = X(A(S), 'p), where
among the
we have
>
the
occur.
are polynomials are the minimal
= '$ [g., ^ P. ] W
IK.
delusion
p = A(S) nil']). .
On each ring
A
P
on
XCt
A(S) -• n A i a
and
There may be repetition
P. , but a given prime occurs at most once among the
We can collect similar terms, and denote by
S = U S^ ,
ia
ICt
[ x i " 'XJ/Pia'
we have
•
f i k , g.fc
I(S) =Hi(s.) = f l O p . , where
1
ia
X(B/P., '-P/P.) . J J d
be any closed semi-algebraic set, and represent n
1
A
First note that
P. , fixed i.
the various primes which
= [R[X,...X ]/P , put the order
'P
= O'p. , the
intersection taken over those indices with A(S) -• IlA
and
'P = A(S) H n -p .
ot
P. = P . We still have inclusion ia a Denote by P o the minimal primes of A(S).
ot
Necessarily, they all occur among the
Moreover,
p
P . a
$o = H ' p . o , so from 3.12, CPJ |i
ip
From Remark 8.6.4, "p, = A(S) nn(']5o)d a P
= n C-P-o) • 1
p d
refinement of the weak order on the integral domain
negative on
X(A O , 'P1 • o) . P P d
iK
200
ik
Q]• p
p. o 1
is a finite
P
A o , namely, 'p. o = P
'P [g-vlg-i ^
P
w
Now
Each
d
P
8.6.5 describes
CP-Q)
ip ^
!p
as the functions in
Finally, from 8.6.2,
a
S, = X(A(S), 'P) , = d
A
p
nowhere
U X(A O , *pj . From 8.6.3 (with Pg minimal P P d U Pg minimal
P = (0)), we conclude
U X(A O , p. ) , and moreover, pd , C A(S) P ^ = Pg P XP d
of functions
S, a =
consists precisely
f nowhere negative on S,. We summarize these arguments in
the following.
Corollary 8.6.7.
Let S be a closed semi-algebraic set, 'p C A(S) a
specific order of the form considered in 8.2.11 with
S = X(A(S), 'p). Then
for any order p' with p C p ' C'P(S), we have that
(A(S), 'Pp
is an
RHJ-algebra and S d = X(A(S), p ^ ) .
Proof: 1
p ,
The argument above gives the result for p' = p. For any such
p d C ^ Cp(S)d-
on
But if f eip(S) d , trivially
f is nowhere negative
S d . Thus ? ( S ) d Ctp(S d ) =1> d .
We remark that if '£" C $
•
is a weaker order with
S = X(A(S), '£"), then
it is not clear when '£" = 'p(Sd ). In fact, we do not assert that if (A,'p) is any RHJ-algebra, then necessarily
(A,'p,)
is also an RHJ-algebra. Our
arguments definitely use properties of orders of specific types. Next let (A/p) be an RHJ-algebra with '$ = '^^[g^ of the weak order.
a finite refinement
Let I C A be a 'p-convex, radical ideal
X = X(A,'p).
Corollary 8.6.8 (a) I is '^j-convex if and only if each minimal prime I
is p d -convex if and only if P^^ = I(Z(Pi) n x d )
P. C A
over
if and only if
I = I(Z(I) H x d ) . (b) If I is £ d -convex, then X(A/I, *Pd/I) = Z(I) to
Z(I) ^ Xj
n
Xd -
Moreover, ^ d / I C A/I
of functions
X(A/I, Cp/I)d) = X(A/I, 'P/I) d-
functions
is an RHJ-algebra with
consists of the restrictions
f ^ A nowhere negative on X-..
(c) For any radical 'p-convex with
(A/I, $ d /I)
I C A,
(A/I, CP/I)d)
Moreover, CP/I) d
is an RHJ-algebra,
consists of the
f ^ A/I nowhere negative on X(A/I, $/I)j.
(d) Op ) = 'pd C A, hence for any 'p-convex radical
I C A,
CP/i)d = CPp/i)d = C^/i) p ) d . 201
Proof:
(a), (b), (c) are routine restatements of various results.
first statement in (d) holds, since by 8.5.1 and 8.6.5 CPd)
have
Tne
= '-Pd-
last
P
art o f
d
( ) holds because CP/I) d
finite refinement of the weak order, so p p /I C CP/I) , so we have
c
Q}/I) d C CP p /I) d
=
and we
always
is still a
Also, by 8.5.1,
(CP/I)p) •
•
c
(1*d/I)
A/I, when
I C A
'$£-convex. We state the following without proof.
(A/I, CPd/I) )
Proposition 8.6.9. (Z(I) H X,) .
Moreover, f £ (;£,/I)
is an RHJ-algebra, with
if and only if
f
I = P C A
too hard. 8.6.5.
a prime
*£d-convex ideal.
X(A/I, CPd/I) ) =
is non-negative on
Using techniques above, one can reduce to the case and
Pd>
£/I C A/I
(CP/I) ) •
An interesting order not covered by 8.6.8 is is
C:
£
The
A
an integral domain
The last two assertions are not
But the RHJ-property seems to be a fairly strenuous extension of
We will provide the necessary ingredient for this extension in 8.10.
We now return to the proof of Proposition 8.6.5.
We follow the notation
in the statement of that proposition.
Proof of 8.6.5: we must have
f
First, if
non-negative on
neighborhood of a point of Thus, points of
X(A, '$),
are exactly the points of
-p on
A.
h(U(f)) = 0.
If Then
x
p G £,
Otherwise, h
h ^ O , then
would vanish on a
X(A,'$),, contradicting the definition of are
$d-convex. x
h f £ 0
on
By 7.7.3, the '^-convex points
f ,h G A ,
X(A, $ ) .
h j- 0,
By 8.4.3, f
'£, that is, -f
We have now proved
X(A, *#)J-
is convex for some total order refinement
x £ X(A, ']5)d, choose 2
h 2 f = p,
say
X(A,*.J3) ,.
for which
in any total refinement of Pd -convex.
f G •£
e>
^d-
Since
f (x) > 0
and
is therefore negative -f(x) < 0,
x
is not
X(A,v£d) = X(A,'J))d> and have established the
characterization (i) of this set. Next, suppose find a signed place Here, K
f £#d. p:
Then
'£[-f]
is
K -• R, ± °° , with
is the fraction field of
A.
an order on
By 8.4.3, we
p(A) C R , p(g i ) > 0,
But then
which is convex for the total order refinement of
202
A.
p
gives a point '£ defined by
p(-f) > 0. x £ X(A,'p) p.
Thus
x
is
'^-convex and
f(x) < 0.
This proves that
x e x ( A / P ) d } , since
X(A/P) d = X(A,$ d ).
Third, if
and
x ^ X(A,'£)d
orders for which we see that whenever
x
£[g]
'^[^1
is convex.
A
In particular, U{g^,g}
g
is positive in all total
Since some such order exists, refining •£, A.
is an order
x £ X(A,'|5)d, we can find
an order.
g(x) > 0, then
is an order on
c
p d = {f G A|f(x) _> 0, all
Conversely, note that by Remark 8.4.7,
U{f i> C X(A,'.pw)
h,g e A,
h ? 0,
is Zariski dense.
g(x) > 0, and
is not Zariski dense in
X(A/Pw)
This establishes characterization (ii) of
Finally, we come to the hard part of the theorem. P C A with
is a
open, hence
of
X(A,'^)
is closed.
is not
X(A,']3)... We must show that if x G Z(P) n X(A,'£)d, X(A,'£)
is obviously
X(A,'^) ,, which will imply by Tarski-Seidenberg that
contains functions in
A
(This is not clear yet since the definition as quantified variables.
finitely many distinct ideals of degeneracy we could conclude that
I
C A
If we knew only x G X(A,1^),
occur, for
X(A/.p) , is semi-algebraic without the new characterization.)
But even granting all this, we would still only know that semi-algebraic.
*])[g]
In a later section we will give another
is a semi-algebraic subset.
X(A,p),
hence
Note that the set of degenerate points of
characterization of X(A,'.J3)j
h(X(A/P) n U(g)) = 0.
'^-convex prime ideal, g £ P, then there exists
g(x) ? 0.
Now if
X(A,'p)d
was closed,
If unproved Proposition 8.1.2 were available, then we would
have a proof that
is an
(A>Pd)
RHJ-algebra by the constructions in 8.2.
However, we must make do without 8.1.2 and give a direct proof of this part of 8.6.5. Assuming '£, with a domain
P
P C A
is
']5 convex, we can find a total order
still convex, by 7.7.3. B, integral over
By adjoining
/gT
to
A, with a total order extending
'^' on
A, refining
A, we can construct '£' on
A.
(We
use the method described in the proof of 8.2.10 and adjoin no more of the than we need to insure that the weak order on
B
contracts to
By the going up theorem 6.4.2, we can find a convex prime order on
B, lying over
any total order on Let that on
K B.
B
P C A.
The reason for passing to
will refine
'P [g.]
be the fraction field of Write
on
'-P = '-P [g«]
Q C B B
/gT on A.)
for this total
is because now
A.
B, with the total order induced by
B = R[x,...x ] , and suppose
x. = x. (mod Q ) , 1 £ i <_ s,
203
gives a transcendence base for of
K
and let
field of
E
B/Q
B/Q
over
be the real closure of R(x.....xn), with
is
x
R.
Let
RCx^.-x ) , ...,x
transcendental extension
R(x ...x ) . Consider
F C r, with convex ideal
Q C B
was an order on
A.
K C r be a real closure in
T. The fraction
algebraic over the pure B = R[x,...x ] C E(x +-,---x ) =
for our total order '£'. (Originally, '])'
We extended it to
B, K
and
T, and we continue to call
the extension '£'.) By the signed place existence theorem 7.7.4, we can find a signed place 1
p : F -• A, ± °° , inducing the order 'p1, finite on and with p1
A
Archimedean over
R(x1-..x ) . Since
is also trivial on the algebraic extension
p'(E). over
Since E.
A
is Archimedean over
Since the
polynomials
f.
It follows that 1 _< i £ n.
x +.
f
p (x s+ .) G E.
R [ x r . . x n ] , then
E
with
Rfx^-.x ]
p (x. - £.) = 0
f.(xs+.) G Q.
and with
p'(xi) = ^ £ E, all
and also, of course, if
Q = (h.)
1
Let
F = E(x +-,---x )
r = tr.deg c(F).
J
over
(Thus
E
is the fraction
r+s = tr.deg (B).)
c
By
K
the signed place perturbation theorem 8.4.9, there exists a discrete rank signed place over
E, p:
p(x i - Z,/) = 0, and Let
Q1 C Q
rings
F -* E, ± «>
p(h.) = 0.
C ... C Q
(inducing a new order on
Certainly, p
is finite on
E
I X S +1* ' < x n J
in
are all convex for the total order defined by Q 1 C Q 2 C ... C Q r .
Also, Q C Q r
function field over
E
with
since
F
'
assoc
p.
Let
p(h^) = 0.
tr.deg^Fp = r - i .
iated
204
F ) , with
E[x s + 1 -..x n ].
tr.degR(B/Q^) = s+r-i.
t0
P-
The
Qj. C *i
Q i = Q- n B, so that Each
B±/q± = F±
is a
This follows by repeated
application of [68 , vol. II, Chapter VI, §14, Theorem 31]. below we deduce that
r
denote the chain of maximal ideals in the valuation
B ^ B ^ - ' O ^ D
degrees.)
c
p'(h.) = 0.
E[x. -£.,h.]. 1
RCx^.-x ) ,
R(x.....x ) , j > 0, there are
We can then write
Certainly, the function field field of
is trivial on
R(x1-..x ) , it is also Archimedean
f
Thus we have
p
Q C B
E, and we identify
are algebraic over
with coefficients in
B with center 1
From the diagram
(The integers indicate transcendence
R[x r ..x n ]/ Q i
R[x r ..x s ]
In particular, tr.degR(B/Q r) = tr.degR(B/Q) = s, and since Q = Qr.
Note this implies that if g £ Q,
We now restrict the signed place K = R(x 1 ...x n ) r
p(g) ± 0.
p: F -* E, ± °° to K C F . (Recall
is the fraction field of B.) We then have a discrete rank
signed place
p: K -* E, ± °° over the ground field
B = R[x 1 -..x n ], with
kernel(p|B) = Q C B, hence with
Again, the image field tr.degR(L) = s. rank
g G B, then
Q C Q^, we have
If g G B,
s signed place
We then compose
L = p(K) C E
p
R C K, finite on tr.degR(p(B)) = s.
is a function field over
R, with
g £ Q, then by 8.4.3 we can find a discrete
q: L -* R, ± °° , finite on B/Q C L, with
and q, K -* L, ± °° -• R, ± °° , and change
q(g) ^ 0 € R.
p
elements as in 7.7.2 to get a new order on K and a discrete rank q op:
place
on infinite r+s
signed
K -* R, ± °° .
We have thus constructed a strictly increasing chain of convex prime ideals
Q x C Q 2 C • • • C Q r = Q C Q r + 1 C • .. C Q ^ C B, with
g £ Q
. We now contract to our original ring
The (new) total order on B of
B. The ideals
increasing chain A/P
+
= R.
P^ = Q. H A
A C B of Proposition 8.6.5.
'£ = *Pw[gi]
are then all
on A by construction
'pd=convex and form a strictly
P 1 C P 2 C ... C P r = P C P r + 1 C .•• C P r + g C A, with
If g € A , g £ P, then
g £ Q + , hence with
still refines
B/Q r + s = R,
g £ P + . Thus, P +
g £ B,
g £ Q, and we may assume
corresponds to a point
x £ Z(P) n X(A,^^
g(x) i 0, completing the proof of 8.6.5.
D
For later reference we state a corollary of this proof of 8.6,5.
205
Proposition 8.6.10.
Let
A
be a real finite integral domain, '$ = '$ [g.] C A
a finite refinement of the weak order, P C A r = tr.degR(A) - tr.deg R (A/P), and let
K
a '£,-convex prime ideal.
Let
denote the fraction field of
A.
Then there is a sequence of discrete, rank 1 signed places P
P
l
P
2
I* ~ inducing a total order on where
p
order refinement of
r>
*• K
is the composition
chain of prime ideals
r
refining p:
(0) C p
-
»
'£, all finite on
K ~* K , ± °° .
C • • • C p
A, with
P = kernel(p| ) ,
In particular, we obtain a
= P C A , all convex for a total
$.
D
8.7. A Preliminary Inverse Function Theorem In order to make sense of differential topology over an arbitrary real closed field, it is imperative to investigate purely algebraic versions of the inverse function theorem.
The result proved in this section is a rather
special case of a better algebraic inverse function theorem, but is strong enough to provide a good picture of a real algebraic variety near an algebraic simple point.
This application will be given in the next section.
This, in
turn, will be used to stratify arbitrary closed, semi-algebraic sets and prove that any such is the maximal convex ideal spectrum of an RHJ-algebra.
Also
crucial for this discussion will be the work on derived orders in the previous section.
(We refer again to 8.1 for the distinction between closed, semi-
algebraic sets and closed semi-algebraic sets.) Before stating the main theorem of this section, we digress a bit in order to put in perspective a consequence of the Tarski-Seidenberg theorem which we seem to require at this point. is a polynomial and n
The result is that if n
B = {(x....x^ € R |a. £ x^ <_ b^}
is a closed bounded
(or a closed ball), then the set of values {t e R|t= f (x),
rectangle in
R
some
is a semi-algebraic subset of the line
x £ B}
fCx^.-x^
R.
This is trivially a consequence of the Tarski-Seidenberg theorem (see the Appendix), but it does not seem any easier to prove directly than the
206
Tarski-Seidenberg theorem itself. Let F in
F = {(x 1...xn,y) e R is the graph of Rn+ .
f
We can reformulate the result as follows.
x R = R
over
The values of
via the map
n
f
n+1
1 (xr . .xj £ B , y =
f(x 1 ...x n )}.
Thus,
B, and is obviously a closed semi-algebraic set on
B
are obtained by projecting
F
onto
R,
R n x R -> R.
y:
As discussed in 8.1, the Tarski-Seidenberg theorem easily can be used to prove certain sets are semi-algebraic, but it does not give much information on whether sets are closed or open, without further work.
On the other hand,
our going-up theorem for semi-integral extensions provides a nice tool for concluding that certain sets are closed. 8.2.7, where we showed that if
A C B
This was formalized in Proposition
is a semi-integral extension of finitely p C B, then image (X(B,'p)) =
generated R-algebras relative to an RHJ-order X(A,£ H A ) , under the projection is always closed. image(X(B,'P))
If
X(B,'£)
X(B,$W) -+X(A/£ W ).
Of course, X(A/P H A )
is semi-algebraic, Tarski-Seidenberg implies
is also semi-algebraic.
As an example, we can deduce that a polynomial
f(x,...x )
maximum value on any bounded closed semi-algebraic set the graph of
f
over
S, say
F C R
closed semi-algebraic set and the Moreover, (A(F),$(F)) We project
F
n
x R = R
n+
RHJ-algebra
is semi-integral over
.
Then
n
S C R . F
(A(F)/£(F))
assumes a Consider
is a bounded has
R, hence also over
X(A(F),^(F)) = R[y] C A ( F ) .
onto the y-axis and get a bounded closed, (comma) semi-algebraic
subset of the line.
But in dimension one, the distinction between closed,
semi-algebraic and closed semi-algebraic obviously is unnecessary, both notions simply corresponding to finitely many closed intervals (including single points and closed rays).
Thus, bounded, closed, semi-algebraic
implies a maximum element in dimension one. Now, as another way of applying the Taski-Seidenberg theorem, one could draw the same conclusion about maximum values directly from the fact that it is true in the case of the real numbers.
But this is a "transcendental proof",
whereas we have just given a "purely algebraic proof". We now state an inverse function theorem.
Proposition 8.7.1.
Suppose
Y ...Y
£R[X,...X ] .
We regard the
207
Rn
as functions on R
n
n
-• R .
Assume
and we regard the n-tuple
Y(0) = 0
B(0,e) C R n
ball
(b) Given
as a ma
e > 0
of radius
Then:
such that the map
Y
restricted to the closed
e at the origin is injective.
e > 0, there exists
6 > 0
B(0,<5) Cy(B(0,e)).
such that
(c) From (a), (b), near the origin the coordinate functions functions of the
Y..
—
i
X^
are
These functions are in fact algebraic functions in
the sense that for suitable polynomials 1
P
and assume that the derivative matrix
is non-singular. (a) There exists
Y = (Y1--«Yn)
f^, we have
f i ( Y 1 ..« Y »X^)
E
°>
n
— (d) Let
A(x) = ((3Y./3X.)(x)), the derivative of
Y
Then sufficiently near the origin the inverse function tiable with derivative
A
, in the sense that if
Proof:
x e Rn.
X = X(y)
is differen-
y = Yfx ). oo
X(y) - X ( y 0 ) - A " 1 ( x 0 ) ( y - y 0 ) H X ~ X0 H
lim
y -• V
at
=
The point to be made at the outset is that once we have the maximum
value property of polynomials on closed bounded semi-algebraic sets (which we have just established purely algebraically) one can write out word for word one o f the standard proofs of the inverse function theorem for real numbers entirely in elementary algebraic terms. Part ( a ) , the local injectivity of Y necessary to find
e > 0 and a constant
is very easy since it is simply c > 0
such that for all x,x f £ B(O,e),
we have
Hx-x'll <_ c||Y(x) -Y(x')|| .
By a linear change of coordinates, we m a y assume that is the identity matrix, that i s , Y
208
i
= X
i
+
Cterms
of d e
g r e e 1 2)«
((3Y./3X-)(0)) = A(0)
Now using the estimates of 7.2 we can, in f a c t , find x,x
!
e BCO,^)
e-^ > 0
so t h a t for
we have
\ llY(x) - Y(x')ll <_ l l x - x f l l <_ 2||Y(x) -
(Note that the first inequality will guarantee that the inverse function X = X(y)
is continuous near
0.)
We next prove (b), the local surjectivity of llxll < e < e, , then
if that
A(x)
P(x)=
llxll = e 2 , then the function
X
e o
2
If 2
= liY(x) - yll
£~ < £i
so small
Ilyll < •—• , consider
on
B(0,e o ).
If
Z
P(x) >_ (e 2 /4) 2 > Ilyll2 = P(0). Thus
llY(x)ll > -| , hence P(x)
If
2 (Y. (x) - y.) x 1 i=l
Directly from (*),
Let us now assume x G B(0,£ 2 ).
is non-singular for all
the function
B(0,e 2 )-
-|- < ||Y(x)ll.
Y.
does not assume a minimum value on the boundary of B(0,e2)
is an
interior point at which
P(x Q )
is
minimum, then
0 = (3P/3Xj)(xo) = E2(Y i (x 0 ) - y ^ ( O Y . / S X ^ (xQ))
for all
1 < j
singular, we deduce Since R(Y 1 -..Y n )
That is, A ( X O ) ( Y ( X Q ) - y Q ) = 0.
Since
A(x Q )
is non-
Y(x ) = y , as desired.
A(0) = ((8Y./3X.)(0))
is nonsingular, it is easy to see that
has transcendence degree
is an algebraic extension.
n
over
R.
Thus
R(Y1...Yji) C R(X][...X )
This proves (c).
Finally, to prove (d), apply the linear isomorphism
A(x )
to the
limit statement of (d), transforming (d) to
lim y
Ao(X(y) -X(yo)) - (y-y o l |fy - yo|| y
o
which does hold by continuity of
=
±
v -> v y y o X(y)
A 0 (x-x 0 ) - (Y(x) - Y ( x 0 ) ) f x - x o | = || x - x o || I y-yo || "
and differentiability of
Y(x).
We discuss further the differentiability of the inverse function
X ( y y lim — — £-0
+£,...y) - X.(y —
D
X = X(y).
Once we have (d), it is easy to see that the partial derivatives
(3X./3Y.)(y ) = J X °
'
y ) —
209
exist and, in fact, that the matrix yQ = Y(x Q ).
where
((3X./3Y.)(yQ))
Thus, Cramer's rule expresses the
coincides with (8X./3Y.)(y )
det(A(x ) ) .
Continuing this discussion, it can be shown that the infinitely differentiable functions of x Q = X(y Q )
computable in terms of
X. - x . , where
x
y.
X. - x.
= (y,...y ) .
Y^-y^
(il) ,..*..(y
Alternatively, the total differential
dY
as
The constant
((3Y i /8X.)(X Q ))(X-x Q ).
as power series in the
CY1-y1)V...C Y n- yn) « will then be
dY:
The higher derivatives are formally
= (x.....x ) , y
i
function
are actually
Y. - y. = Y.(X,...X )-Y.(x.....x )
terms vanish and the linear terms will be solve for the
X.(y)
as follows.
First, rewrite the polynomials polynomials in
as
(8Y i /8X.)(x Q ),
specific rational algebraic functions of the polynomials with denominator
A(x ) ~ ,
Then formally
The coefficient of
, O^/BY 1 ) (yp) , I = Ci r ..!„).
can be regarded as a polynomial
R n x R n -*• R n x R n , linear on the second factor at each point of
the first factor, and, moreover, dY has non-singular differential at (0,0). 2 The second derivatives (3 Y./9X.8X,) then occur as part of the first derivative i 3 K of
dY.
Applying the general discussion of first derivatives to
the inverse function of of the
Y
dY
shows that
is twice differentable, and leads to a computation
(3 2 X./8Y-3Y V ). 1
J
K
Here is perhaps a more algebraic approach to the derivatives The derivation to a derivation
Z/dY±: D±:
R[Y1...Yji] ->R(Y 1 ...Y n ) CR(X 1 ...X n )
R [ X r . .XR] -» R(Xj .. .X n) .
a o (Y)X m + a 1 (Y)X m " 1 + ... +a m (Y) = 0
(3X./3Y.).
extends uniquely
Specifically, if
is the minimal polynomial for
f.(Y,X.) = X.
over
R[Y,...Y ] , then we must have 0 = D i (f j (Y,X j ))
= ( O a o / 8 Y . ) X m + . . . +(3a m /3Y i )) + ( O f ./3X.) (Y,X.))Di(Xj) ,
and this last equation can be solved for
D.(X-).
Using the inverse function theorem 8.7.1, we can easily establish the following implicit function theorems.
Proposition 8.7.2.
210
Suppose
Yj ,... ,Y R e R[XX .. . X j ,
k < n,
Y^O) = 0 ,
and suppose the vectors
d Y ^ O ) = ( ( B Y ^ B x p (0),.. ., ( S Y ^ B X ^ (0)) ,
are linearly independent.
Reordering the variables if necessary, assume that
{dY.(O), 1 < i < k; dX.(O), k < j < n}
i
—
—
3
0 G R k , the equations
Then
F
Define
Y^x) = c ^
is locally
F: 1-1
span
—
uniquely define the coordinates
Proof:
1 < i < k,
1 < i < k,
x-^.-.x^
Rn
Rn.
by
Then, sufficiently near
c± G
x-(x 1 ...x k ,x k + 1 ...x n )
as functions of
x
ic+]/--xn
F(x) = ^(x),...,Y k (x),X R + 1 (x),...,X n (x)).
and onto by 8.7.1.
The result then follows easily.
n-k
Surfaces
Remark.
i i i k-
Surfaces
Y..(x) = c±,
D
Proposition 8.7.1 and the discussion of derivatives above also
gives the tangent plane of the surfaces (8Yi/8X.)(O) = 0 ,
Y i (x) = c^.
j > k, then the tangent plane to
at the origin is the coordinate plane general, the tangent plane of
x =
(^
Y^x) = 0
For example, if Y i (x) = 0 ,
••• = x ^ = 0 } = R
1 <_ i <_ k, x {0}.
In
at the origin is the annihilator
with respect to the usual scalar product on
Rn
of the k-vectors
(dY^)(0),
1 <_ i £ k, dY± = (3Yi/8X1,.. • ^ / a ^ ) .
Proposition 8.7.3.
Suppose
Yl.. .YR G R[X X . . . X j , k > n,
and suppose the vectors
dY.(O),
1 £ i £ k
variables if necessary, assume that
span
R n.
like the standard inclusion, mapping ||x||
Reordering the
dY^^ (0),. .. ,dY (0)
after smooth algebraic change of coordinates the map to
Y ^ O ) = (0) ,
Y:
span
Rn .
R n ~> R
Then looks
(x 1 ...x n , 0...0), for
sufficiently small.
211
Proof: (0 ... 0,x
Define
G ( x r . .x n ,x n + 1 .. .xk) = (Yj (x1.. .xn) ,... ,\(*v
... x v ) , so that K
n+1
G:
R k -• R k
+
• -\))
is locally 1-1 and onto near
0
by 8.7.1.
D
8.8.
Algebraic Simple Points, Dimension, Codimension, and Rank We continue our study of the polynomial ring
real closed field
R.
n-vector of functions. of
P
to be
oodimension of df. G A .
If
f e R[X], we have P C R[X]
P
A = R[X]/P.
df = (9f/9X 1 ,...,3f/3Xn ), an
Thus
Suppose
dim(P)
It is routine to check that P, using the formula
is an invariant of the
P = (f 1 ...f s ).
to be the maximum number of
of generators of
space of
R
codim(P)
is independent of the choice
d(hf) = f dh + h d f = hdf G A n
derivations
D:
A -* K, where
K
P.
Such a
D
is determined by
precisely the linear conditions words, we have
codim(P)
codim(P) < codim(Q).
212
K
of the
is the fraction field of D:
f € p.
R[X] -* K
A.
which vanish
DX = (DXX,...,DXn) G K n , subject to
0 = Df = df-DX, if
linear conditions on
We will use the fact that if
if
dim(P) + codim(P) = n.
can also be characterized as the dimension over
This coincides with the space of R-derivations on
Define the
A-linearly independent vectors
A standard result in commutative algebra is that In fact, dim(P)
over a
is a prime ideal, define the dimension
dim(P) = tr.degR(R[X]/P).
finite integral domain
n
If
R[X] = RfX^.-X^]
P C Q, then
f = f(X) G P.
DX, so
In other
n = dim(P) +codim(P).
dim(Q) < dim(P), hence
It is also clear that (Sf^/SX.), where r x r
at
x
P = (f ...f ) .
is the rank over
Let
submatrix of maximal rank over Z(P) C R n
Let
codim(P)
to be
denote the zeros of
V E R[X]
be the determinant of some
A.
V £ P, that is, V £ 0 G A.
P.
Thus If
x E Z(P), define the rank of
K ok .
Suppose
g i ^ 0 E A,
A = R[X]/P,
and
X
affine embedding of
Then there exist zeros
with
rank (P) = codim(P) x
x E Z(P)
with D
X(A,*p ) . This definition refers to a particular
X(A,$ ) . However, from commutative algebra, algebraic
local ring associated to the point
x
and let
is a simple point if and only if
If we refine the order on
A, say to
m
C A
figures below illustrate what might occur.
Let
A
be the
be the maximal ideal.
dimD(m /m ) = dim (A) = tr.deg D (A). K X X K $ = '£ [g^] , we are not especially
interested in all the algebraic simple points belonging to
'£
is an order on
will be called
simple points can be invariantly characterized as follows.
=
A oA •
In particular, rank x(P) = codim(P).
x € X(A,'l> ) w
algebraic simple points of
x
—
P C R [ X ] is a prime and '^ ='^w[gi]
1 < i < k.
V(x) ^ 0.
The points
Then
P
From the Nullstellensatz 8.4.3 we obtain the following.
Proposition 8.8.1.
g.(x) > 0
of the matrix
rank (P) = rankD (-5—" (x)). Obviously, rank (P) < rankA(—^-) = X
codim(P).
A
In both cases
X(A,'£).
The
X(A,'-p), where
'P [g] > contains degenerate points which are algebraic simple points of g >. 0
Z(P)
Figure (a)
Figure (b)
X(A,$ ) . We will define the algebraic simple points of
X(A,p)
to be the
non-degenerate algebraic simple points, that is, simple points of belonging to
X(A,'£,).
X(A,£ )
Propositions 8.8.1 and 8.6.5 show that these simple
213
points are, in fact, dense in X(A,£ d ). and
g.(x) > 0, then
'£[g-]
Namely, if x G X(A,£ d) = X(A,£) d
is an order on A and 8.8.1 guarantees simple
points exist in the neighborhood
x G U{g-} C X(A,'£),.
Now, however, we should check that an algebraic simple point of X(A,'p ) is automatically non-degenerate.
This requires a somewhat careful argument.
In fact, we will base the proof on the inverse function theorem.
The following
is the central result of this section.
x G R n , with
Let g,f ,... ,f r G R[X] ,
Proposition 8.8.2.
f^x) = 0
df i (x) linearly independent, g(x) > 0, and assume the neighborhood x
is sufficiently small.
In fact, P Moreover,
Then the ideal
I(Z{fi>
n
U(g)) = P
is the unique minimal prime of the ideal P
is convex for any order
is prime.
(^...f ) , with
£ = p [g.] C R[X], with
at
U(g) of x
vanishing of the
f., 1 < i < r. l — —
A = R[X]/P
in the variety
Z{f^}
g. are positive in the order
for
•£ [g-]. This proves the last part of the proposition, w 3
If A
Proof:
A = R[X1...X ]/P.
that
P
df.(x)
zeros.
f^ near
'•P
'P [g-]
214
x and P
a
If A
R and
f 1 ...f r G P,
Rank considerations show
The zeros of P near
Thus every neighborhood of x
Corollary 8.8.4. =
By assumption, there exists
linearly independent.
we are in the situation of 8.8.2.
is convex
x G x^.
is necessarily a minimal prime of (^...f ) , and since
the zeros of the
P
is a real finite integral domain over
is a simple point, then
r = codim(P), with
U(g) C u { g . } . l
$(Z{f.} fiu(g)) C A, hence
x G X = X(A,'J)w)
Write
is a ring of
defined by the
If g-(x) > 0, then choose 3
The
Corollary 8.8.3.
In practice,
can be thought of as small balls centered
x. The first part of the proposition says that
functions on any neighborhood of x
x G Z(P).
g.(x) > 0.
Before indicating the proof of 8.8.2, we give applications. the "small" neighborhoods
U(g) of
x G Z(P),
x thus coincide with
is exactly the ideal vanishing on these is Zariski dense in X, as desired. •
is a real finite integral domain over R,
finite refinement of the weak order on A, g. ^ 0, then the
non-degenerate set simple points
Proof:
X(A,'£),
x £ X(A,£ )
is exactly the closure of the set of algebraic with
g i (x) > 0.
We know from 8.8.1 and 8.6.5 that any neighborhood of a non-
degenerate point contains such simple points.
Conversely, 8.8.2 guarantees that
a function which vanishes on a neighborhood of such a simple point is already 0
in
A.
•
Corollary 8.8.5. X(A,£) d
With the same assumptions as in 8.8.4, the set
is a closed, semi-algebraic set.
Proof: —————
If
A = R[X]/P,
P = (f.), then the simple points of I
are the points in the open, semi-algebraic set
X(A,$ ) w
{x|rank((8f./8X.) (x)) =codim P}.
(In fact, by looking at the possible subdeterminants, this is even an open semi-algebraic set.)
U{g.}, then take closure.
We intersect with
Tarski-
Seidenberg guarantees that the closure of a semi-algebraic set is semialgebraic.
•
Proof of 8.8.2. result„
Denote by
First, part of the proposition is a purely algebraic
Ax
the local ring obtained by localizing
and dividing by the ideal generated by the
fi#
R[X]
The assumption
at
x
df^Cx)
independent implies that the graded ring associated to the maximal ideal m
C A
is a polynomial ring, in particular, a domain.
a domain, or equivalently, (f1...f ) i r this says precisely that of
(f^.,^)
in
x
Therefore,
is a prime ideal in
R[X] . x
A
In turn,
is a zero of precisely one minimal prime
R[X], Geometrically, Z(P)
a Zariski open neighborhood of
x £ Rn .
and
Z(f io ..f r )
is
P
coincide in
For details of this argument, see
[63, Chapter 11], or texts on algebraic geometry. Now we must look more closely at small semi-algebraic neighborhoods of We must show that if
h
vanishes on
Throughout the argument we may assume assume of
P
df,(y)...df (y) and
Z{fi> H u ( g ) , U(g)
g(x) > 0, then
as small as desired.
h e P.
Thus, we
are independent, y £ U(g), and we assume the zeros
(^...f^) coincide in U(g) o
Suppose otherwise,
that is, P C l(Z{fi>
n
U(g)) = H p
where the
215
P.
x.
are convex prime ideals, strictly containing
P.
say
and by 8.8.1 we can find
x G Z(P ) . Then
y G Z(P Q ) n u ( g ) independent.
U
Z(P j ), such that
Since
we know that near Thus, f
codim(PQ) > codim(P)
G P
Z{£.} n u ( g ) = U (Z(P.) nu(g)) i j J y
the zeros of
df,(y),...,df (y). 1
f
o o
G
P
° *
=
(f,...f )
I
z
f
^ ^ i^
nu
f
y £ Z(P.), J
and
j t 0,
y, yet
df (y) 1
(Alternatively, if we choose by
f fJ
without changing
( g ) ) » which certainly vanishes on
€= n
f
yet still has differential independent of
p.,
j^o J
df Q (y).
Z{f i )
P .
is
° f
G P ,
coincide with the zeros of
Z{f.}, near
j,
are linearly
r
f Q (y) = 1» then we can replace f
x G Z(P-)> some
df Q (y) ,df: (y),. .. ,dfp(y)
vanishes identically on
independent of
We know
near
Now, y,
From this,
df,(y),...,df (y).)
we will derive our contradiction, by appeal to the implicit function theorem 8.7.2. By translation and linear change of coordinates, we may assume
and
f i = x i + (terms of degree •> 2) G R[X],
theorem states that near graph of a map y
i
= y
0
R n " r •* R r
i^ x r+l* ''Xn-^
are
the surface
f^ = ••• = f
The implicit function =0
• • -\h
where
algebraic functions of the last
n-r
variables.
Moreover, the tangent plane of the surface is the plane equivalently, (8yi/8x.) (0), function with have
df Q (0)
l < i < r ,
independent of
(8f /dx.)(0) £ 0, some o j
shows that
fQ
fQ
x, = ... = x
r+1 <_ j <_ n.
If
df i (0) = dx i (0),
j > r+l, say —
calculation of derivatives of t,0...0)
is given by the
R n = R r x R n ~ r , { y r . 'YT>\+1
in
smootn
1 < i < r.
y = 0 G R
j = r+1.
along the curve
= 0, or
fQ G R [ X ]
is a
1 < i < r, we must
But then a simple (y^(t,0...0),...,yr(t,0...0),
cannot vanish identically on the surface
• The last part of Proposition 8.8.2 states roughly that a prime convex if it has enough zeros. prime, •£ = 'p [g.] C R[X] P
is
'^-convex if
P
Precisely, one can say that if
has algebraic simple zeros
x G R
with
is
P C R[X]
a finite refinement of the weak order, n
P
is
g. £ P, then g i (x) > 0.
We will now establish another criterion for convexity with this same flavor. The result can be found in [21].
It could be derived from the implicit function
theorem and existence of simple zeros, but we will follow the proof of which avoids the implicit function theorem.
216
[21],
Proposition 8.8.6. domain over
R,
Let
U(g) C R
n
A = R[x,...x ] = R[X,...X ]/P a non-empty open set,
be a transcendence base over
R
R[x 1...x d] -»A.
defined by '£ [g] C A
and let
Then
P
is convex for
Proof:
First assume
y6R ,
z£z
of
P.
1
and
are
R[x 1 -..x d ] -•A -•A/P i ,
P
is
Choose
A.
Let
R[xr..xd].
d+1 < j < n . such that 2m=1
$ [g] C A
11
f(T) = I !
is
Suppose not, that is,
f = Hf i # f = 0.
is an order, let
Then
in the kernel
Then
f
vanishes on
This contradiction
t G A
R(x 1 ...x d ) C K, where
a. (x, .. .x-JT
We can write
1
a m ( y r . . y d ) + 0,
be a primitive K
is the fraction
be the minimal polynomial for
t
1
Xj
= 2?;J ( b ^ (xy . . x ^ / c ^ ( x r . .x d ))t G K, y = Cy1--->rn)
By the Nullstellensatz we can find c ij (y 1 ...y d ) + 0, and
i-ai(y1.. . y ^ t 1 " 1 t 0.
Thus
t(y 1...y d ) G R
G
Z(P) H U(g)
3f / 3t (yx . . . y d )
=
is a simple root of the
f (yx,... ,yd,T) 6 R [ T ] ,
polynomial Suppose
(z-^... zd )
will have a root f(z-...z,,T) id
x
is very near
near
T = t(y,...y ) . in
•
Tnen
f
(z^ • • • z d > T )
If
d+1 < j < n, define — — z = (z.....z ) £ R .
z G U(g) C R .
We are finished
z G Z ( P ) C R n.
the kernel of
f(xr..xd,T).
R
by the rational formula above and consider
Watching our estimates carefully, we may assume if we prove
in
(y^'-'y^)
tCy^.-y ) , by a simple estimate argument showing
changes sign near
z. = x. (z, .. .z ,,T)
Now,
Now, P
£ i GR[ X;L ...x d ]
f i i 0, and let
element for the field extension
over
contains a disc
'£w[g] -convex after all.
Conversely, if
field of
(equivalently,
W
TT(Z(P) flU(g)), which contains a disc, hence shows
CR[X]
'B[g]-convex primes.
1
tr.degR(R[x]/Pi) < d.
^[g]
contains a disc.
P = I(Z(P) n u ( g ) ) .
P CI(Z(P) nU(g)) = H P . , where the
x-^.-.x^ E A
e > 0 G R.
TT(Z(P) HU(g))
\[g]-convex if and only if
Let
be the projection
TT(Z(P) nij(g)) d
B(y,e) = {z G R | llz-yll <_ e}, some
P CPi
TT: Z(P) -• R
is an order), if and only if d
g £ P.
be a finite integral
Define
f(z 1 ...zd,x) = 0
R[x,...xd,T] -* R[x,...xd,t] = B C A p:
implies
R[x1...x d,T] "• R p
factors through
by
p C x ^ = z±, p:
a little further care with estimates will insure
B -* R.
is precisely p(T) = x.
Since
Then
c^. (y^. . .yd) f- 0,
c..(z,...Zj) j- 0
and thus
217
p
extends to
p:
A "• R,
p(x 1 ...x n ) = (z1 ...z ji ).
Thus
z^Z(P)
as
desired.
D
8.9. Stratification of Semi-Algebraic Sets In this section we give procedures for stratifying semi-algebraic sets, that is to say, for breaking arbitrary semi-algebraic sets up into "simple" pieces.
We should emphasize that we deal only with the most elementary
aspects of stratification questions.
For example, we do not consider the
Whitney-Thorn regularity conditions, nor questions of equisingularity. Actually, our primary goal in this section is to show that if
F C Rn
any closed, semi-algebraic set, I(F) = {f e R [ X r . . X j | f ( x ) = 0, all A(F) = R[X r ..X n ]/I(F), fl(F) = (g e A(F)|g(x) >_ 0, all (A(F),!P(F))
is x e F},
x e F } , then
is an RHJ-algebra (with, of course, X(A(F) ,'£(F)) = F.)
Strati-
fication considerations arise naturally in the proof of this result. E C Rn
Let
be any semi-algebraic set, and represent
E. = Z { f . . } n u { g I
IJ
Then the
I = U ^ P^
},f..,g GR[X ij iK l
and
n
X]. n
I(E) = 1(1") = n 1(1..)
are the minimal primes of
g i k £ P^ . hiQ e
iK
I(E^).
Let
I C R
n
E
as
E = U E.,
be the closure of
E.
Write I (Ei) = l(E/) = H p ^ , where Then
f. . E P. , all
j,a, and
This last holds since if, say, g i o E P^ , we could choose
Pi3-Pia-
Then
g i Q h i o e I(E i ), but
gio
is strictly positive on
pjFCX
E^, hence Let not have
h^o£
I(E i ), contradiction.
A. = R[X. ...X J1/P. , 'B. L 1 n / ia y i a ia
X(A. , tp. ) C E*., in general, because the XOc
10t
zeros far away from
1
ia' ^ia1 *
Proof:
Then
S
S.
= X(A. , 'JK ) , the non-degenerate points of
Moreover
> I^i^ J I^i " s i a ) •
We use 8.8.4, which characterizes
set of algebraic simple zeros
218
could have degenerate
However, we do have the following.
Let
iaC f i*
g.v IK.
U{g ik >.
Proposition 8.9.1. X(A
=r $ L 6[g.,1 C A. . Now, we definitely do 3 w ikJ la
x
of
P. , with lOt
S.
In
as the closure of the
g., (x) > 0. IK
Call this last
set
V. a , so
thus clearly Let
S. a = V. a.
V i a C E. = l{£^}
gi
n U{g ik >. Thus
= n gik.
Then
g^A^ x
Via = (
algebraic simple zeros
A
E
minimal primes of the inclusion
n u
to be
A
^i* ia^
with
vanishes on We have
and
S.a = V i a C E.. denote the set of
A. = 2 det(A ) 2 , where ia p p
x £ Z(P i )
We are now in Fat City.
{A }
Z(P. a) C Z{f..}
r x
where
g^OO
Via
> °>
as
Moreover,
is the set of above.
1". - V^ , hence on
I(E) = 1(1) =
r
r = codim(P. ) . la
/ O G A i a , that is, g i -A i a £ P i a .
ia^ia)
gj_#A^
it is obvious that
dvnens'ion of
Let
(3h./3X,), and let j K
it is clear that
gj_'A^ae15
Since
IL - S^ .
H p. . i,a i a
•
Define the
max dim(P. ) . This is the maximum dimension of the
I (E), which all must occur among the
A(E) = A (I) "• II A i a i,a
S = U s.ia C U E .I = L
with
f.. €= P.^ we have
P i a = (h.) C R [ x i . . . X n ] .
submatrices of Let
Since
and the order
Then,* by J 8 . 6 . 5 ,> each
X(A ia ,(!p ia ) d ) = S. a
and
P. .
A(E) H n($ i ) .
v(A.i a » ,('£• w10LJ)£>)
<#±a)^ = V(S±a).
We consider Let
i s an RHJ-algebra, & >
By 8.2.9,(A(E),A(E)
is an RHJ-algebra, with X(A(f),A(l) H (tp. ) ) = S.
Obviously,
nnflj.^)
A(E) nn(tp. ) =
icx ^
the functions in Consider
A(E)
nowhere negative on
E - S.
By Tarski-Seidenberg
Clearly, F - S = U I. - U S. i 1 i,a l a But
I(f.) C i(f. - U s . ) . I :£ i a ia
must properly contain some < dim(E.), hence
ia ^
S. E -S
is a semi-algebraic set.
C U ( Ix. - U Si.a ) , thus i a
dim(¥- S) < max dim (IT. -US. ) 1 ~ a ia
In fact, by 8.9.1, each minimal prime of P. . Thus, we have strict inequality
dim(E - S) < max dim(E.) = dim(E).
X(A/p) d C R n , where
dim(E. - U S . )
We can now repeat the
whole process above, beginning with the semi-algebraic set many steps we succeed in writing
1(1". - U S . ) I a la
E - S.
After finitely
E C
A = R[X1...X ]/P,
P
prime, and where
$ C A
is a
finite refinement of the weak order. As consequences, we have proved the following two results.
Proposition 8.9.2. (A(F),$(F))
F C Rn
be a closed, semi-algebraic set.
Then
is an RHJ-algebra.
Proposition 8.9.3. type.
Let
In order that
Let
•
(A,$) €= (PORNN), with
(A,«p) — (A(F),$(F))
A and R-algebra of finite
for some closed, semi-algebraic set
F, it is necessary and sufficient that there exist (1) finitely many primes
219
P. C A %
C A
with
i
(0) = n p .
= A p
/ -[»
and (2) finite refinements of the weak order
such that
$ = A H Il(^i) , under the inclusion
A -^IIA^ D
Proposition 8.9.3 should be compared with the results of 8.2, especially 8.2.11, 8.2.12, and the last paragraph of 8.2. '^
Working with derived orders
instead of ty , we can now deal with all closed, semi-algebraic sets, not
just the closed semi-algebraic sets. If we combine the Tarski-Seidenberg theorem, the going-up theorem for semi-integral extensions, and 8.9.2 we can prove the following.
Proposition 8.9.4. inducing set.
cp:
Then
R
n
m
-> R .
cp(S) C R
Proof:
m
Let
cp*:
Suppose
R[Y.....Ym] -•R[X1...X ]
S C R
The image
A(cp(S)) ->A(S), where
cp(S)
is a closed, bounded, semi-algebraic
is semi-algebraic by Tarski-Seidenberg and The homomorphism cp* induces an inclusion
A(S) = R [ X r . .Xn]/I(S)
Moreover, ^(cp(S)) = A(cp(S)) H ^ ( S ) .
Since
and
A(S)
is certainly semi-integral over A(cp(S)).
RHJ-algebra, hence by 8.2.7, so is
Remark.
A(cp(S)) = R ^ . . .Yj/IfcpCS)) .
is semi-integral over By 8.9.2, (A(S),$(S))
(A(cp(S)) ,$(cp(S))) , and
Thus, cp(S) C R m
X(A(cp(S)),'^(cp(S))).
be a homomorphism,
is a closed, bounded, semi-algebraic set.
bounded by simple estimate arguments.
A(S)
n
R, is an
cp(S) =
is closed.
D
As an immediate corollary of 8.9.4, we get that any polynomial
function on a closed, bounded, semi-algebraic set assumes maximum values.
This
generalizes the discussion at the beginning of 8.7, where we proved this result for bounded, closed semi-algebraic sets. In many ways the non-degenerate sets integral domain over
R,
$ = $ [g-] C A
X(A,-.p)d, where
A
is a finite
a finite refinement of the weak order,
w j are more natural "building blocks" for semi-algebraic sets than the Although
X(A,'p),
X(A,*£).
is not a manifold, as we have seen it is the closure of a
d-manifold, where
Specifically, it is the closure of the
set
x
V
d = tr.deg D (A). K of algebraic simple points
of
X(A,'B )
w implicit function theorem of 8.7 guarantees that that is, locally like
220
R .
Thus, X(A,$),
with
g• (x) > 0, and our
j V
is an algebraic d-manifold,
is a sort of closed "d-manifold with
boundary and singularities". contractions to fractions
If
K
E
of
A
Secondly, these orders
'£, C A
are exactly the
of finite refinements of the weak order on the field of
A.
Thus we can view
X(A,$,)
as an "affine model" for
is a semi-algebraic set, the procedure above for obtaining
E~ = UX(Ai,>|5i) , where
Pi C A ( E )
are primes,
A i = A(E)/P i , and
^
C A^^
a finite refinement of the weak order, definitely depends on a specific presentation of
E = UE-,
E. = Z{f..} n u { g . v } .
1
1
1J
It is also not really a
lK
stratification of E, since the pieces
X(A.,$.) can overlap, and, in fact, 1 -1- d
can overlap on more than their "boundaries", as in Figure (a). Even if they overlap of
only on boundaries, we may end up with a very unnatural decomposition
E, as in Figure (b).
it Figure (b)
We can rectify these problems somewhat by replacing the orders
(•£.) 1
by finite intersections of such orders.
Specifically, let
{P }
C A. d
1
index the
distinct primes which occur among the P.. We assume our decomposition E = U X(A.,'£.) is irredundant, in the sense that no term can be omitted. 1 -1- d On
A
a
= A(E)/P , we impose the order a
n p p
'£• = '£ . i a
Then we have
1
(•£ ) = fl(p.) 1 d d RHJ-algebras
and
X(A ,$a ) = u X(A.,$.) . i : L ^ d p = p d i a
(A , ('£ ) )
are geometric invariants of
= (A(E),A(E) if
a / 3.
A(E)
The primes
E,
E = U X(A ,$ ) ,
dim
) , and
The proof is not hard.
certainly minimal primes of primes of
Now, we claim that the
P
of maximal dimension are
A(E), hence are characterized as the minimal
of maximal dimension.
which is a geometric invariant.
Now
Moreover, X(A ,'£ ) = (E n X(A $ )) , i Y d "Y w d 1" - U X(A $ ) has strictly lower
221
dimension than invariant.
E.
Thus, by induction, the whole decomposition of
E
is
This discussion indicates that very nice basic building blocks
for semi-algebraic sets are affine models of orders on function fields where
'£
is a finite intersection of finite refinements of the weak order.
We can also use the ideals of degeneracy stratifications.
By definition, I
=
lim
= I(E n U )
x G E" C Rn.
if
U
Write
minimal primes of
C A(E),
I
x G E, to study
I(E fiU), where
smaller and smaller open neighborhoods of I
'£ C K,
x.
U
parametrizes
Of course, the limit stabilizes
is small enough, say a small ball centered at
I x = I (I H U ) = H P
where the
P i x C A(E)
are
I .
Proposition 8.9.5. (a) over
E". Each (b)
the
Only finitely many distinct prime ideals
£
a
If
C A
a
P. = I , ix y
for suitable
E = U X(A ,$ )
y
near
P.
P ot
The subset
is an irredundant representation of
P
= kernel (A(E) -*• A )
X(A ,•£ )
C E
is a minimal prime of
Proof:
which occur are exactly the prime, y G E . x G E
such
I . x
P. .
x
in
In fact, we can find
P.
is
y G I" O U
H z(P. )
y G U,
XX
such that
E H n
X
entirely of simple zeros of
P. .
such that
Then
I Pix,
y f. Z(P. ) , ~J X
= Z(P. ) O IL XX
X
The finiteness of the set of all
nl)-convex,
XX
0
j- P. , and an open set
I(U
If there will exist algebraic simple
XX 1X
I
is characterized as those
For example, since each
we know that arbitrarily near
P.
E", where
The proof consists of reviewing the various results of 8.6, 8.7,
8.8, and this section.
zeros of
varies
x.
prime ideals which occur as ideals of degeneracy,
that
x
are finite intersections of finite refinements of the weak
order, then the primes
(c)
occur, as
consists
X
= P. . x G E, follows from (b) and (c).
These two statements can be proved readily with all the machinery at hand. For example, suppose Consider the
X(Ag,«£g)
E H U , and since
222
I
= P
is prime, and assume
which intersect
U H X(Ag,$g)
E H U.
P = I(E H U ) , y G U. Since
is Zariski dense in
P
vanishes on
X(Ag,$ w ), we have
P C p
But also, some
X(A,# ) i
of
P
in
will contain a whole neighborhood of simple zeros
T d
E H U, so
P
C p , hence
P
= P.
We leave the rest of the details
of (b) and (c) to the reader.
•
The ideals of degeneracy can be used to define local notions of rank and dimension.
Namely, if
I
= H p. , define
d i m ( E ) = max dim(P. ) A
J.A
rankx(E) = rank x (I x ) .
(Rankx (I) = rank{dh.(x)}, where J fications of
E
I = (h.)> x G Z(I).) J
based on rank and dimension.
{x £ E|dimx(E) >^ k} an open subset of
is a closed subset of
E.
E
We can define strati-
Note that for any integer
k,
and
is
{x £ E|rankx(E) _> ^}
We do not quite want to begin a stratification with
all points of maximal rank, since this set will have (possibly rather singular) boundary points due to inequalities defining
E.
to be the "interior points of maximal rank".
These are obtained as follows.
Take the irreducible components take only those points
y
X(A ,'£ )
Cf
However, we can define
of least dimension.
E.
We will have
but will terminate.
Then
which are simple points, lying on a unique
X(A ,'£ ) , and for which an entire neighborhood of to
E
I(E-E^ -*) 3 I(E)»
so the
y
in
Z(P )
belongs
process can be iterated,
Using the implicit function theorem of 8.7,
E^ *
is
a manifold.
In the figure above, E^ ^
consists of the circle minus the vertex of
the triangle.
Then at the next stage
the triangle.
Next, E ^
triangle and finally
E
E ^
we get the interior points of
consists of the interiors of the edges of the is the set of vertices.
223
(x 2 +y2 )(y2-x-l) = 0
y 2 -x 3 + x 2
Figure (a)
Figure (b)
We point out that by defining rank locally, our stratification by rank is not the same as that of Whitney [44] for algebraic varieties
Z(I) C R n.
In our stratification, the origin has rank 2 in both varieties above, hence is the first stratum. origin has rank
0
On the other hand, in Whitney's stratification, the
in Figure (a) and the curve has rank 1, whereas in
Figure (b), the origin has rank 2 and the curve rank 1. 2 the polynomials
x(y
(In Figure (b),
2 - x - 1)
and
y(y - x - 1)
vanish on the variety,
but have independent differentials at the origin.) We can also find a manifold stratification of
E
by dimension,
of E E D E. Q . 3 E.j. D ••• . We begin with the pieces X(A ,'£ ) of greatest dimension, and take for the "interior" simple points, lying on a unique
X(A ,$ ) .
Then
dim(E-E, 0 .)< dim(E), and the
stratification continues inductively.
8.10. Krull Dimension Let
A
be an integral domain, •]} C A
an order.
By a weak %-oha-in of
prime ideals we mean a strictly increasing sequence of '^-convex primes (0) C P 1 C ... C p^ C A.
The length of the chain is
r.
We define the
weak Krull d-imens'ion, dim (A,$) , to be the maximum length of such a chain. In complete generality, this notion is probably uninteresting. we make the drastically simplifying assumption that
224
A
Even if
is finitely generated
over a ground field example, if then
R[T]
R[T]
R, the order
•£
is ordered with
can make things complicated.
T
For
infinitely large relative to
is a semi-field, with weak dimension zero.
R,
This pathology is
caused by lack of finiteness conditions on the order. Suppose, then, that we begin with a finite real domain R
real closed, and an order
the weak order. Thus
If
Q Cp
*jp = *J3 [g.] w 3
primes of length
any '^-convex prime, we can apply 8.4.3 to '^-convex primes above
Q,
Q C p
=£
l
R,
tr.degR(A/P) < tr.deg R (A/Q).
But from 8.4.3 we have chains of ^-convex
dim (A,'£) = tr.deg(A).
r, thus
over
obtained by finitely extending
are prime ideals, then
dim (A,'£) <^ r = tr.deg(A).
A
Q C A
is
(A/Q, :J3/Q) and obtain chains of
C ••• C p
=£
In fact, if
=£
C A, where
s = tr.deg(A/Q).
s
By analogy with classical Krull dimension theory, at least two other questions come to mind. •^-convex primes below
First, given such a Q,
(0) C Q
Q C A, can one find a chain of
C • • • C Q
= Q?
More generally, given
Q C p C A , both $-convex, can one find a chain of $-convex primes between and
P,
Q = P Q £ P x J •**
c
P t = P, where
Q
t = tr.deg(A/Q)-tr.deg(A/P)?
The answer turns out to be "yes", but before discussing this further, we want to argue that there are other, perhaps more natural, notions of dimension.
Again, we begin with
(A,'p) , '£ = $ [g.] w j
the weak order on the finite domain
A.
Define a strong 'p-chai-n of prime
ideals to be a strictly increasing chain some total order refinement
'p ^ $ , all
a finite refinement of
(0) C p P-
are
C ••• C p «p-convex.
9
such that for
Note that a weak
•jj-chain corresponds to a sequence of (non-injective) epimorphisms of domains (A,p) •+ (A/P1, tp/Pp -*...-> (A/Pr, tp/Pr).
From Proposition 7.7.9, a strong
#-chain requires much more restrictive convexity hypotheses on the primes
P^,
and by 7.7.10, a strong «j)-chain corresponds to a sequence of signed places K -• Kj^, ± «>-• ... "* K r , ± °° over where
K
dim (A,$)
is the fraction field of
R, finite on A.
P i = kernel (A -• K i ) ,
We define the strong Krull dimension
to be the maximum length of a strong $-chain.
actually produces strong '^-chains in tr.deg(A).
A, with
Of course, 8.4.3
A, hence we still have
dims(A,'.p) =
But what is more important, we have available the signed place
perturbation theorem 8.4.9 and its consequence Proposition 8.6.10 which gives more delicate information immediately.
225
Proposition 8.10.1.
Let
Q C A
there exists a strong 'p-chain
be a
(0) C P
'^,-convex prime
C • • • C P
= Q C • • • C P
'=f=-
' =F
=p
•$ = $ [g-]. Then
'=fz
^
C A, where
^
s = tr.deg(A) - tr.deg(A/Q) = codim(Q).
Proof:
The chain below
Q
K -» K-j^, ± oo-> ... -^ K g , ± «>.
comes from 8.6.10, which gives signed places
Then
Kg
is still a function field, to which
we apply 8.4.3, and extend the strong $-chain above Actually, a little more is required if some that the total order one finally obtains on one constructs an
integral extension
A^B
necessary in order that the weak order on Q
lifts to a convex prime of
field of
K
B
Q.
g. G Q
in order to conclude
actually extends $.
Specifically,
by adjoining as many
fgT
contracts to
The prime
$ C A.
as
B, and we apply 8.6.10 and 8.4.3 to the fraction
B, then restrict to
K.
This argument was actually used in the proof
of 8.6.10.
•
The reason strong dimension is a more natural concept than weak dimension is that strong dimension is a looal concept, giving geometric information about a semi-algebraic set infinitesimally near a point. prime x
of 8.10.1 is a point
Q
is a non-degenerate point
then any !pd-convex prime of
x
in
P.
in
x E X(A,$),. is
Then
P.
'^,-convexity of
Moreover, if
p[g]-convex.
X(A,'p)j, the prime ideals
have lots of zeros. of
P
x E X(A/p).
Specifically, suppose the
g(x) > 0,
says gGA,
Thus in any neighborhood
U
of the strong *p-chain of 8.10.1
Specifically, P. = I(Z(P.) n U ) , we can find simple zeros
U, and so on.
The strong chain of primes going down from
correspond to a chain of subvarieties going up from the point l,2,...,r.
Q
Q
x, of dimension
In other words, i/nfinitesimally near* x, we can move about on the
semi-algebraic set
X(A,'£)
with
r-degrees of freedom.
By way of contrast, we reconsider weak dimension.
The result which
allows the "desired" conclusion is the following.
Unproved Proposition 8.10.2. Then there exists a fpj-convex prime tr.deg(A/Q) = 1.
226
Suppose
Q C A
Q 1 , with
is '^-convex, $ = •£ [g.].
Q1 C Q
and
tr.deg(A/Q') -
Corollary 8.10.3. in A.
Suppose
Pi
C ... C p . _ C A
(0) C p^ C •. • C P^ C A
Then there exists a ^-convex refinement
of
r = tr.deg(A).
maximal length
Proof:
is any $ -convex chain
It suffices to insert an appropriate chain between any two
•^-convex primes we may assume
P C Q.
Passing to
(A,$) , where
A = A/P,
= tp/P = $ w [g.]» j Q1 C Q
P = (0). Apply 8.10.2 to find a $,-convex prime
dimension one greater than that of
Q.
of
Then apply 8.6.10 to go down from
Q'.
•
A proof of Proposition 8.10.2 can be extracted from [21]. Here is a rough outline of the geometry involved. R
By extending the ground field from
to an appropriate transcendental extension, we may assume that
dimensional, that is, a point. of points.
Now, we know the variety
X(A,$)
For example, if tr.deg(A) = r, near a simple point
Q
is zero
has lots X = X(A,'p)
looks like affine r-space smoothly embedded in some higher dimensional affine space point
Q.
Rn.
We consider sections of
X
by
n-r+1-planes through the
The implicit function theorem guarantees that many of these sections
will be 1-dimensional semi-algebraic sets.
(Simply take the
general position with respect to the tangent r-plane of The difficulty is, the point Y, and
X
n-r+1
plane in
at a simple point.)
Q may be an isolated point of such a section
Y may not be algebraically irreducible.
However, the method of [21]*
is essentially to argue that generically these sections irreducible 1-dimensional sets through
Q, although
Y
are algebraically
Q may indeed be a
degenerate point. Such a set corresponds to the desired prime
Q1 C Q .
227
This
Qf
£,-convex since, by construction, Qf has sufficiently many zeros in
is
V Note the proof of this dimension theorem 8.10.2 uses completely different concepts than those required for the study of strong dimension.
This is because
8.10.2 really is not a local geometric result at all, but a global property of semi-algebraic sets. not be described by
Near a degenerate point
Q, the variety
X(A,$)
will
r-independent parameters; although globally we can pass
a curve, then a surface containing the curve, and so on, up to the r-fold itself, through the point
Q.
Also, we point out that the commutative algebra
analogue of the dimension theorem 8.10.2 can be proved using integral extensions and various going-up and going-down theorems for prime ideals.
In our real
setting this method seems to break down because of the special hypotheses needed in 6.4, especially in Proposition 6.4.2(b). The analog of 8.10.3 for strong chains does not follow routinely from 8.10.1. A
and
The reason is that given Q/P
a
P CQ CA
C£j/P) -convex prime of
strong ^-chains between
P
and
Q
with
P
a ^.-convex prime of
A/P, we cannot immediately construct
by passing to
(A/P, ($j/P) ) . The
finiteness condition on the order is lost by this process.
What is needed
instead is a more vigorous version of the signed place perturbation theorem 8.4.9, which was the basis of 8.10.1.
Proposition 8.10.4. over
R, and
:p' C A
small relative to
ideals of where that on
A.
if
induced by
r
Suppose
(0) = Pi
and
and all
K = R(x-,-..x ) P^.
r
x.
are infinitesimally are
P r i m e $'-convex
signed place
p:
is the fraction field of
are'^-convex, where
$ CA
K -*R, ± °° ,
A, such
is the total order
p.
The proof is by induction on
is quite small.
Let
we may also assume that the
r, there being nothing to prove
r. = tr.deg(A/P^.), so that
As in the proof of 8.4.9, we may assume
228
is a finite integral domain
C P " S '** S p # o ^ 1 i =£ =£ x s
Then there exists a discrete rank
p(x.) = 0
Proof:
A = R[x,...x ]
is a total order such that all
R.
r = tr.deg(A)
A
Suppose
A
r=r
is integral over
> r, > ••• > r .
R[x,...x ] , and
{x, = xv(mod Pi.), 1 < k < r.}, give a transcendence
base for If
A/P^., all r
o
- r, = 1, then the place l
the total order rank
1
A
£'
on
K
A
1
C K
= R[x.,y.], with
and the subfield
x.
and
y.
P^ • of
Af
fraction field of a quotient of
the convexity of the
Write
R[x,...x
i j K i r i , K
F
CF
relative to
R, the
A
A'
to
q:
+1]-..[xn],
Also, K1 is the
A'/P- )> hence by induction, K ! -* R, ± °° , preserving
and for
generate the prime
f. v (x v ) 3 > K- "•
is integral over
x, K
over
Eb ik (O...O)T 1
x, K
over
P- Ac j j
in
AQ . b j
R[x,...x ] . l r
Let
f. ,, with J>^
•£' C K.
Let
r
denote
F. C r be the real closure of R(x,...x ) , J i rj C ••• ^F-i- Since all x. are infinitesimally small x.
are certainly finite relative to
R(x1...x
) . In j p! = p f x .. (This Kv. i. . .x r j 3 i
A
of the place
p!(x i )
R(x 1 ...x p ) .
r
Thus elements of It follows that the
R(x 1 ...x r .), in the appropriate residue field,
with some
xa. . G F.. f. ,
is a root of the non-trivial polynomials
for
then gives
K, and let
are all algebraic over
k > r
K
r - r., > 2. I — o k > r., let
be a monic polynomial equation for
to our integral dependence relation for
if for
of
is the multiplicative set of non-zero elements
is the center in
and we can identify
Pj
Thus we may assume
are infinitesimally small relative to
p'-tx^)
R, we
$'-convex primes
follows easily from Proposition 7.7.4 and results in 7.6.) P^.
tr.deg(K') =r-l.
K, we can write
by 7.7.8.
1
fact, P^. C A 3
with
in
R[x1...x ] . We have our total order
a real closure of
is a discrete
Composition and reordering
{f. ,} J K ' k>r-
0 = Ea. ., (x ...x )f. ,
so that
j > 2.
S. C A
k > r = r , then o
coefficients in
A
R
represent the minimal polynomial for
] , then j
1
If
If
) C K
over
signed place
A/Pi. = R [ x r - . x r ][x r
R[x1...x, .] .
K1
(in fact, of
K -• K ! , ± °° -*• R, ± °° .
p:
f. , = f. (xv) £ P| x j,K i,k K j
of
r-1
P! ,
K -* A, ± °° defined by
infinitesimally small relative to
A', and they necessarily form a chain in
we can find a discrete rank
if necessary.
R(x,...x
is the integral closure of
can lift all the '£' -convex primes
the desired
x,...x
p', .: R(,x1...xriJ
place, with image a function field
Moreover, if 1
j , by simply rearranging
in
Applying the place K,
p!
k > r, we get that
Eai.-^(C1.. • C r )T 1
over
F..
Also,
)x11
we let
0
0 = Eb.v (x, .. .x be an integral dependence relation IK i r K R[x.,...x ] , then 0 is also a root of the non-trivial polynomials 1 r
over
R.
229
Let R
0 < e £ R
and let
over
0 < e. € F. 3 3
F..
to a field
where
L
by adjoining the elements
1 <^ j <_s,
1 £ i <_ n, and
e. < x...
We now extend
/ e 2 - x2., and
£• ., 3 >i
K
/e? - f 2 , , 3 3 >K
K
r < k £ n.
over
Ea. ., (£,...£ )T 1 i-JK i r
be less than all other roots of the
x 1 £ F., we may also assume
Since
Eb., (0. . .OJT1
be less than all other roots of the
(Note these elements belong to
B = [x.,x.-£. ., / e 2 - x 2 - e, 1 1 j,1 K small relative /e? - f? , J1. Since the f. , and x. - s£ . . are infinitesimally 7 3 3>k j,k I j,i 2 2 to F., hence also relative to R, and since e. < x. , all these generators
T
of
so
L
B
is real.)
We replace
A
by the ring
are infinitesimally small relative to
The places
p!
are still defined on
L
R. since
R^.^x
J
center of
p!
on
B
is a prime
all our hypotheses hold for
B.
restricts to a discrete rank
r
K = L.
) C L.
l
The advantage in this:
Q^.
which contracts to
Since any discrete rank place of
The
3 P^. C A. r
place
Thus p:
K, we may as well assume
if we now construct any total order
L -• R, ± °°
A = B, '£ C K
such that the {x-...x } are infinitesimally small relative to R and the {f. , |r. < k < r} are infinitesimally small relative to R(xn...x ) , then J »^ J 3 necessarily all
x. 1
are infinitesimally small relative to
are infinitesimally small relative to generate
Pi.'Ag,, where
we have degree
E C r F, C E. 1
over
recall, f
S. = R[x,...x r ] . Thus if
be the real closure of
E, which can be written
R[x x .. • x r _ 1 ] ,
(x ) £ P^
x .
pn f
.
If
D
E
so that
D £ (x )
is the place
E(x ,y) r
or
Since ...x ) E(f
of transcendence
, z ) , where, i ,r
1
on
K.
p
D = DEOr)
x of
over E
is non-trivial, hence
We are then finished by induction, We propose to reorder
E(x ,y)
is algebraic and so that all our hypotheses on
K C E(x ,y) are preserved in the reordering. It is better to work with the transcendental element
230
r - r 1 >_ 2,
represents a minimal polynomial for
as in the second paragraph of this proof. over
E(x ,x
is algebraic, the place
is discrete rank
f. , K 3 >
$ , we still have that the
x i = xi(mod P i ) . We consider the cut
defined by
p.
R(x 1 ...x r p .
Consider the function field
= £-
and all
R C x , . . ^ ^ ) . As we observed, {f. ,} 1 X 3 J » k k >r-
pDr . on K relative to this (new) order K^X]^. . .X r ) center of p. on A is Pj_.c A. J J Let
R
(x r) G E(f 1 ^,z)
= E(x , y ) , since
f,
R(x 1 ...x r ) C E.
Thus let
sign of
f,
is infinitesimally small relative to the (large) field D 1 = D £ (F 1
if necessary.
so that if
E(fn
)
E(f1
R(x,...x
We may assume
0 < D', by changing
We appeal to 8.3.1 to choose
is reordered over
new order extends to relative to
).
r
E
with
, z ) . Of course, 3
D c (f 1
3 £ E,
) = 3,
0 < 3 < D',
then this
is still infinitesimally small
).
We now consider our hypotheses on the elements
x.,
i <_ r
and
f. ,,
1
r. < k <_r. xr
Since the order on
and the
f.^,
1 < j < s.
Pp (x^) = C-, •
nas not
i
c^(£, •!•••£•,
ficients of
(x )
l,r of
pp
on
But
f ^
i)
= f ^ r ( x r ) = E c ^ . . . x ^ x * , say, -JPp (Xp) 1 * where
0 = Ec.(C, i---Ci
changed if
that is,not all f..
is unchanged, we need only worry about
p f , we get
and if we apply the place G F
E
J ,K
i < r.
This polynomial is non-trivial,
vanish, because, by definition, the coef-
are non-zero modulo
P-
r
C A
and
P^
i1
A.
We see that when we made our finite extension from
K
to
graphs back, we should have added two more elements, /6 - x r y,6 G F 1 ,
Cx
-xx < y < q
r
contains only the one root
r
< 6 < Cx
t>1
of
conditions indicated, the elements small relative to
R.)
r
G
F
Ec i (C 1 1"-^>1 /6 - x
and
E(f;.
to
Moreover, if one writes out the polynomials
i«
and
^T
1
= °-
in
Fj
(With the
particular, x r
,z) = E(x r ,y), we have
is still infinitesimally small relative f.
= f.
i < r, have not moved at all, and
xr
(x ) , 1 < j < s,
r
j,r
one sees that they are still infinitesimally small relative to
on
(y,6)
are infinitesimally
j,r
our reordering.
/x r - y, where
But there is no problem doing this retroactively, and
Pp (xr)
r
r
/x ^~y
=
^i
In
L, four para-
+ x x , and the interval
the conclusion is, even with our new order or
R.
is the center
ij
—
F., since
has not moved relative to
x^
F^^ 3 F., by
This establishes our original hypotheses for our new order
E(x r ,y), now with
D £ (x r )
algebraic, and hence completes the proof
of 8.10.4.
•
We can now prove the strong chain analogue of Proposition 8.10.3. Proposition 8.10.5. £,,[g-] w J
is
a
Suppose
A
is a finite integral domain over
finite refinement of the weak order.
R
and
Then any strong '£-chain
231
of prime ideals
P^
£ * * * £ ^i
(0) C P 1 C .. . C P^ C A
Proof:
g..
where
of maximal length
r = tr.deg(A) .
The proof consists of arguments used previously.
adjoining suitable ho
can be refined to a strong (£-chain
^ A
/gT
to
First, by
A, we are reduced to the case where there are
Secondly, by extending the ground field, we are reduced to the case P.1: C A s
is zero dimensional, that is, A/P^
= R.
Both these reductions
s
are just as in the proof of 8.6.10.
We then apply Proposition 8.10.4 to get
a maximal strong $-chain refinement of the P^
below j
discrete rank
fraction field of
pf:
signed place
r - rg
A, r
P- . x s
K -* K 1 , ± °° , where
= tr.deg(A/P^ ) , and
Kf
This gives a
K
is the
is a finite algebraic
extension of the fraction field of rank r signed place «B-chain above P, . s Remark:
If
A
p":
1
K
A/P- . Finally, 8.4.3 gives a discrete x s -* R, ± °° , which completes the strong D
is a finite integral domain, $ = $ 7[gi] c A, and
a prime $J-convex ideal, then the RHJ-property of rather easily from 8.10.5. if
g E A,
follows
In fact, in 8.10.5 one can add the conclusion that
g £= P^ , then the refinement chain can be constructed with
Of course, P x
(A/P, (pj/P) )
P C A
is zero dimensional, hence is a point
is highly non-degenerate.
g jE P .
x E X(A,$), and moreover,
This argument would complete the proof of Proposition
8.6.9 which we omitted previously.
8.11. Orders on Function Fields Let Write
K
be a real function field over the real closed ground field
K = R(x,...x ) , A = R[x,...x ] .
X = X(A,$ ) C R n w
Thus
is an algebraic variety and
algebraic variety of non-degenerate points of
where
£ W (K)
is an order on
is the weak order on A
if and only if
All non-empty open sets
232
K.
A
is an integral domain,
X, = X(A,($ ) ) a w jj X.
Note that
We know that if
is the (semi)($ ) = A n $ w d
g i ^ 0, then
U{g i> H X d ^ 0, where
U{g.} n x,
R.
are Zariski dense in
(K) ,
'^[g^]
U{gi> = {x E x|gi(x) > 0}. X^.
We will find it convenient to forget the degenerate points of and work in
X^.
Thus, the notation
U{g^}
now means
Let
V
Vlg^} ^ ^
(In
Xd,
W{g i }-U{g i >
% [g^]
V{g^} =
has no interior.) V{g.}.
be the collection of non-empty sets of the form if and only if
altogether,
{x £ Xd|g.(x) > 0}.
We will also find it convenient to work with the larger open sets U{gi> = W{g i> C X d .
X
is an order on
A.
The
Vfg^}
Thus
are regular
o_
open sets, that is, V{g.} = V{g.}, since, in fact, V{g.} = U{g.}. Seidenberg, the
V{g.}
are open, semi-algebraic sets.
in any semi-algebraic set) the the
U{g.}.
That is, each
finite union of
Vs.
V{gi>
V
is a finite union of
U's
much the same statement as Unproved Proposition 8.1.2. fact, although there are strong indications that the
By a prefilter
V
£ &.
U
is a
In fact, it is pretty We will not use this V{g.}
provide a more
we mean a non-empty subset of ~//, closed under
A prefilter is a filter if
V G ^ ,
V C V ' G ^
implies
A filter is an ultrafilter if it is not properly contained in any
other filter in
X
Note that
This is the property of the U{g.}
and each
U{g.}.
^^- ^
finite intersections.
In affine space (hence
form a base for the same "topology" as
This is certainly not obvious.
natural base than the
By Tarski-
V{g i> C v ( h )
V{g i >
if and only if
hCVCg^) _> 0.
which makes them more convenient than the
for the purposes of this section.
The set of filters in
V
is partially ordered by inclusion, an arbitrary
intersection of filters is a filter (every filter contains a union of a chain of filters is a filter. an ultrafilter.
Each prefilter
^ C ^
X^ = V(l)), and
Thus every filter is contained in
is contained in a smallest filter
d Let if
$ C A
g. E •£.
only if
«^('p)
is a prefilter since
Define ^tp) C f tp [g.]
by
V{g.} e^"($)
is an order on
A
if and
V{g.} £ ^ , which implies the desired finite intersection property
for ^ p ) . If Let
be any partial order.
&CV
^
C $ 2 , then
an order on
2
be an arbitrary prefilter in
if there is V Cv(g).
p
V £&
Since the sets A.
with V
g(V) :> 0.
V.
Define
<$(&) C A
by
Equivalently, this says
are Zariski dense, it is clear that
In fact, '$(0) is a derived order, since if
f g = h
233
f ± 0, and
h(V) :> 0, then
g must be non-negative on
would vanish on the Zariski dense set ^
c
Also
C ^ , then yi&fi 'p(^>)-
V.
Otherwise, f
U{g.} ^ U ( - g ) , where
$ W
=
V = V{g.}.
'£(«£§)» where ^
If
is the filter
generated by &". We will study the compositions p ( ^ W ) and
&^&(${&)')
and ^^(0)).
First, $
are completely obvious from the definitions.
Proposition 8.11.1. (a)
If
«^C /' is any prefilter, then ^ ( ^ ) = ^
(b)
If <3* is an ultrafilter, then '$Q?)
c
A
C X
is a total order. Also,
if $ C A is a total order, then &($) C "#" is an ultrafilter. (c)
is any order, then pG^Op)) = P d -
If 1) C A
Proof: (a) Since $££) = '£G^), we may as well assume know
3FC£T(:p@r)).
Conversely, if
g.(V.) ^ 0 , V. e ^ T Thus
Then
^ is a filter.
g^^ € $ 0 9 , so that
V. CV(g.)
and since &
We
V{gi > £&($(&)),
let
is a filter, V(g.) e ^
V{g i ) G J T and (b) Suppose f = #(£}
admits a proper refinement, $ C'£[g].
and since ,F is an ultrafilter, 3~= g'&ig]).
Then But now
, contradiction. Secondly, assume
'^ C A
properly contained in a filter
is a total order and suppose (
S. Let
? C^C^QJ)) Ctp(®, we have fl = ?G?).
&- &($') is
V = V{g.} G ^ - ^ T . Since But
g± e ? ( ^ ) , hence
V{ g i >
contradiction. (c) Finally, ^ c'p(^(^)) and $(^f(^)) ).
Also, 'p, = fl !p
total order refinements of $. c
is a derived order, so
where the intersection is taken over the Write $ n
Then ^ p ) C ^ , fl(^CP)) ' P ( ^ )
cn
= $C^T)
where
<2" = &{$ ) .
' ^ C ^ ) = ?d.
•
We have thus established a natural, bijective, refinement preserving correspondence between orders on the function field as derived orders on
234
A) and filters in the family
K
(which are the same
V of subsets of
X,.
This result makes it clear how an arbitrary order on a function field
K
is related to the particular orders defined by evaluating polynomial functions X C Rn
on Zariski dense subsets of any affine model
for
K.
Total orders
correspond to evaluating functions on infinitesimally small open subsets, that is, germs of open sets. Of course, the totality of all orders on of
X.
K
is a birational invariant
However, with respect to the particular model
X, we can distinguish
orders obtained by evaluating functions on bounded sets and orders obtained at infinity. writing element
A = Rfx^.-.x ] , we say a filter V{g^}
unbounded. If
With reference to the embedding
^"C V n
which is a bounded subset of
If &
X C R , which amounts to
R .
is hounded if it contains an Otherwise, we say &
has no bounded refinement, we say
V{g t } C Ty we have the closure
is a filter, by the center of &
&
V{g i> = U{gi>
is
is at infinity. in
Xd.
If
^C
V
we mean
n v{g.} ex, x d .}e^ If
^ = ^ ' C ^ w [ g i ] ) , then
center of
&
C(0) = V{ g i >.
could be empty, even if &
In particular, C { M £ W ) ) = X d . is bounded.
The
In the case of filters
at infinity, homogeneous coordinates and hemispherical models of semi-algebraic sets makes these notions quite analogous to the centers of places in algebraic geometry over algebraically closed fields.
This is especially true for total
orders, because of the intimate relations between total orders and real places. In the real case, we have the added geometry provided by partial orders on function fields.
Proposition 8.11.2. &CV
of subsets of
order refinement £P =^*CP»), then
If $ C K
X d , then
'p! D $
with
C ^ " ) = x.
is an order with associated filter
x € C(^j x
if and only if there is a total
a 'p1 -convex maximal ideal of
In general, if
&
A.
If
is an ultrafilter, C(^)
consists of one point or is empty, and is always empty if
&
is at infinity.
235
Proof:
The main point here is that if
g £ A,
necessarily positive in any total order on ideal.
K
for which
$ [-g] C A
in
X H W(-g)
of
X H W(-g).
if
is an order and
g(x) > 0
x G V(g), then
and even if
g(x) = 0 ,
The proposition is then proved as follows. family of elements
g £ G A, all
e > 0 G R, with
This is easy using the technique of 8.1.1. #[§£> all
e
]
is
an
exists with
x
'£[g ] .
f
£ -convex.
Otherwise, x GV(-g.)> some
Then
j , and
g.
If
x
g
is
is a convex
x
For
is clearly not
is a degenerate point
X E Cf^O, we choose a
g (x) > 0,
The assumption
order, clearly with center
any total order refinement of
•£' 3'£
x
This assertion can be deduced from either 8.6.5 or 8.11.1.
example, if
'p'
x E V(g), then
{x}.
n v ( g ) = {x}. x G Cffi
implies
We can then take for
Conversely, if a total refinement x G V { g . } = O V*(gi)
all
would be negative in
$'.
g-,,...,g^
The other statements of the proposition are also easily deduced from the remark in the first paragraph of the proof.
Remark.
•
If our ground field is R, the real numbers, then a compactness
argument implies that the center of a bounded filter is never empty.
Remark.
If the order
$ C A
is finitely constructed using the operations
of finite extension, finite intersection, and the operators of 3.12, then
C(^('^)) C X,
will be semi-algebraic.
£ , !p , # , $ ,
In general, however,
this will not be expected.
Remark. • *
If $ C K = R C x , . . ^ ) in
the associated signed place, with center
{(a,...a )}
A
if and only if
is a total order and Archimedean over
pD: K
K -» A, ± °°
R, then &ty)
has
P R (^ i ) = a^ G R C A.
It is pretty easy to see intuitively (but not necessarily easy to prove) what all orders are like on function fields in one variable.
236
An affine model,
with degenerate points excised, will look like a smooth curve, with finitely many singular points. A branch at
x
At each point some even number of branches comes in.
is a connected component (see 8.13) of
a very small open neighborhood of
x.
U - {x}
where
U
is
The total orders correspond precisely
to the branches at all affine points, together with a finite number of branches at infinity.
The individual sets
V{gi>
are finite unions of
open intervals on the curve, with disjoint closures. In 7.7 we studied places if
A
is real and
p:
K •> A, °° of real fields
K.
In particular,
A
is given any total order, then by 7.7.2
p
can be refined
to a signed place, p:
K •*• A, ± °° , thus inducing an order on
K
compatible
with the order on
To conclude this section, we study a finite variant
A.
of this problem of lifting orders in the case of function fields. Suppose A = A/P, and '3 C A P
A
is a real finite integral domain, P C A
$ CA
such that
will be
a total order.
TTCP) = $"> where
a prime ideal,
We ask when there exists a total order IT: A -• A
'^-convex, hence necessarily
P
is the projection. is
($ ) -convex. w d
In particular,
This condition
will be included in our proposition below. Let
X = X(A,'£w),
X = X(AyB ) , X . C X W
X^ C X
the non-degenerate points, and similarly
the non-degenerate points.
We first prove a lemma
Q
of independent interest.
237
Proposition 8.11.3. X = X(A,£ w ), in
X
E C X
Suppose
A
is a real finite integral domain
a semi-algebraic subset.
Then
(that is, I (E) = (0) C A) if and only if
E
is Zariski dense
E H ^
has non-empty
interior.
Proof: dense. some
The "if" statement is clear since open sets in
Write E^^
E = U E^
Ei = Zff^} n u { g i R} .
is Zariski dense.
Thus, f i . = 0 e A
I(X-X,) j- (0), we must have
Proposition 8.11.4.
^{g.^} n X,
If
P C A
then there exists a total order all finite sets in
{f } C $,
and
is Zariski dense,
E_L = U(g i k ) H X.
Since
non-empty as claimed.
is prime, f C A
$ C A
E
are Zariski
with
f. £ 0, the set
= A/P
IT (•£) = $
•
a total order,
if and only if for
U{f\} H x"d n x d
has interior
X,. d
Proof:
First a comment on notation.
Also, we write
U{f;[} C x,
F = {f G A | ? ^'F, ? 7* 0}. is
Then if
X,
U{? i> C x , if Suppose
$
We have
X CX
f± e A,
Tr(fi) = 7 ^
= $ W [F]
Obviously, fr(t|J1) = $ C A . *
CP-i) -convex. A d
TT(^) = '^5, since
is already a total order.
order and
f C A P
is
exists, certainly
IT: A ->A. Now, let
is an order and
P CA
From 7.7.3, we can choose a
£ ^ ^
total order
P
A
total order refinement '|5 C A
such that
c
from
is 'J3-convex.
Then also
Conversely, if our desired
F O p , hence
•£, = '£W[F] C A
is an
(#..) -convex.
We now prove the hypotheses of the proposition are sufficient. prove
$.. C A
is an order and
order, then some
^ [£^]
P
is
If
'J3, is not an
is not an order, for finitely many
is equivalent to
U{fi> H X d = 0, so
U{f-} n X,.) i (X
'p, l
If
($,) -convex.
is an order, but
U{? i ) n ^ P
n X d = 0.
is not
We must
f^ C- F. (Note
This
u^}
H x"d
('p..) -convex, then again Id
there is a finite set of (A, e p j f i l ) ^
f. G F with P not ('£ [f-]) -convex. By 8.6.5, 1 W 1 (J is an RHJ-algebra, with X(A, ( ^ [ ^ 1 ) ^ = U{f i> H ^ . (Here,
the bar denotes closure.)
But if
U{fi> H X d n f ^
is exactly the ideal of functions which vanishes on C U{f i } H X d H X = U{fi> n X d n Z(P). convex.
238
has interior in U{f.} n X, H X,
But in this case, P
would be
X*d, then
P
Finally, we prove the hypotheses are necessary.
Assuming
P
is
(•£ {f.}) W
convex, we know that
P = I(U{fi> H X d O Z(P)).
Zariski dense in
Since
X.
f = IT£.
That is, U{f i ) H X
vanishes on
f £ P, we must have
8.11.3 to
l
U{f i ) H x ^ f l x
A, we conclude that
Zariski dense in
U{fi> H x d H x"
1
a
X.
i
d
Applying
has interior in
x"d, as
desired.
•
Corollary 8.11.5. if
X
In the situation above, if
X = X(A,'J>w) = X d , that is,
has no degenerate points, then any total order
total order on
Proof:
'£ on
A
lifts to a
A.
The point is if
X = X,, then a
I
Gf,
I
f 0, then
U{?i> n X, + 0.
If
U{?.} n x . O X = U{f\} n l , I a d ! d
For example, if
D
A = R[X,...X ] , the polynomial ring, then 8.11.5 applies.
However, the conclusion admits a trivial proof in this case. the fraction field of
A = A/P.
give the polynomial ring small relative to C A[X,...X ] . of any
A
is
U{f.} H X, - U{f.} H X,
l
and
H X
A.
Then
A[X.....X ]
tp
A
induces a total order on
any total order with
denote A
(X. - C O
We then restrict this order to the subring
We see this order lifts
f G R[X][...Xn]
Let
and we infinitesimally
R[X,...X ]
'£, by writing out the Taylor series
in powers of the
(X± - q ) . Here
q
= X^mod P) G A.
The geometry of 8.11.4 is roughly illustrated by the picture below, where
we have orders
X = X,
for simplicity.
*& C A, we can think of
subsets of
X,.
Using the ultrafilter interpretation of total 'jf as
picking out infinitesimal ly small open
If these small sets do not contain enough points in
there can be no ultrafilter of sets in
Xd
giving an order on
A
Xd,
lifting $".
239
In the figure, orders
*p
"centered" in the right half of
X,
will lift,
while orders centered in the left half of
X will not. d We will state a slight generalization of 8.11.4.
Proposition 8.11.6. Suppose $
[g-1 *- A 3
w
c
A.
is a real finite integral domain,
a finite refinement of the weak order and
'•Pw[g-1-convex ideal. $ w [g-] j
A
Suppose
£ C A = A/P
Then there exists a total order
= t, i f and only if for all finite sets n
f^g.} nx(A,!D w[g.]) d X(A,-B w [g j ]) d
Proof:
P C A
a prime
is a total order refining !p C A
with
{?.} C $",
and
g.l
f\ ^ 0, the set
has interior in
First, 8.11.3 generalizes routinely to sets
Then the proof of 8.11.6 is exactly like that of 8.11.4.
X = X(A,£ [g.]). w 3 The reason one
must work with the closure of may belong to
P, that is,
U{f.,g.} H X(A;J [g-]) is that some g. 1 w J 3 d 3 g. = 0. If, in fact, all g . f- P, then one
can just require that
U{f.,g.} H X ( A ^ [g.]) meets X(A,'£ [g.]) in 1 W W 3 3 d 3 d a s e t with non-empty i n t e r i o r . For e x a m p l e , i f g . f- P and X(A,$ [ g - ] ) = w 3 3 d X(A,$ w [g.]), this always holds, giving a generalization of 8.11.5. D
8.12.
Discussion of Total Orders on Suppose given a total order
$
R(x,y) on
R(x,y), the rational function field
in two variables over the real closed field
R.
Let us assume that
'£
is
2 centered at the origin in pD: R
where
R(x,y)
R , that is to say, P
R
M
= P R (y) = 0 G R C A,
R(x,y) -* A, ± °° is the signed place associated to our order on
and the subfield
R CR(x,y).
The results of the preceding section show that the order describable by infinitesimal behavior of functions On the other hand, the order pR:
R(x,y) -* A, ± °° .
f £ R[x,y]
is necessarily near the origin.
is also described by the signed place
In this section we will reconcile these two descriptions
by comparing invariants of
240
$
*£
pR
(the residue field
A, value group
r, rank
of
Pn, etc.) with the particular geometric behavior of functions
origin which determines whether
f
is positive or negative (rel
f
near the
'£). We do
not attempt to state a theorem, and our discussion is meant to be enlightening, not rigorous.
Our discussion of signed places on
R(x,y)
is a watered down
version of the discussion of valuations on function fields of two variables given by Zariski in his work on resolution of singularities of surfaces. There are two possibilities for the residue field p •
R(x,y) -• A. ± oo #
Namely, A
Archimedean ordered over A ^ R. over
R, and
cut of in
There is then
R
R
A
R, or
converging to
T.
may be very large.
converge to have
is isomorphic to
0.)
R(t). Let
and let
(a!,aV)
This means
is either smaller than some {i}
A
R.
t = f(x,y)/g(x,y) G R(x,y), with
t
of the signed place
is either a function field in one variable,
is algebraic over
defined by
A
Assume first that R(t)
T = D R (t)
Archimedean
be the transcendental
be a nested family of intervals
a! < x < aV
and every element of
a!^ or larger than some
aV.
R
(The index set
Also,it does not follow that the differences
aV - a|
$ C R(x,y), we
In any event, with respect to the order
a! < t < a'.'. Consider the neighborhood in
R
defined by
U i j £ = {(x,y)|0 < x 2 + y 2 < e , a!g2(x,y) < f(x,y)g(x,y) < aV g 2 (x,y)}.
We decompose these
U.
into their "connected components".
are non-empty and have the origin in their closure. and
g(x,y)
vanish at the origin.)
These components
(In particular, f(x,y)
Moreover, the behavior (number and
approximate location) of these components stablizes for sufficiently small and
a!,a'.'
close enough to
T.
The order
one of the component families of the in the next section, 8.13.) if
h(x,y)
U.
$ .
is then determined by selecting (Connected components are studied
Specifically, h(x,y) G R[x,y]
is positive in
$
assumes only positive values on the selected component, infinitesi-
mal ly close to the origin.
We interpret this selection of a component of
as selecting a branch at the origin of the curve larger field
e
R ( T ) . If
A = R(x,u),
u
f(x,y) = Tg(x,y)
algebraic over
U.
over the
R ( T ) , this choice is
just the choice of one side of a real root of the minimal polynomial for
241
u.
As a concrete example, let
R
be the real algebraic numbers
component of a
Given any polynomial h
h(x,y), we can find numbers
has no zeros in the intersection of the cone
disc
0 < x
2
+y
2
< e.
of this region, hence In case h(x,y)
ry < x < sy
U.
e > 0
so that
and the puctured
Then by requiring, say, 0 < y, we single out a component h(x,y)
will have constant sign in this component.
A ^ R, the value group
is the order to which
r
h(x,y)
branch of the transcendental curve first non-vanishing derivative of Note that if
r < TT < s,
t = x/y,
is always
Z.
The value in
Z
of
vanishes at the origin along our chosen
f(x,y) = Tg(x,y). h(x,y)
That is, we find the
on the branch at the origin.
R = ]R , the real numbers,this example cannot exist since
]R
admits no Archimedean extensions. We now turn to the R-valued signed places are four possibilities for the value group
Case I.
r
non-Archimedean.
(m,n) < (m',n')
Case II.
if
r
Case IH(a).
m < m
1
T
242
There
r = Z 1
m = m ,
x
Z, ordered lexicographically,
n < n1.
That is, F = Z.
Archimedean, non-discrete, and containing two incomWe can then assume
an irrational real number.
F
R(x,y) •> R,± °° .
r.
Archimedean and discrete.
mensurable elements.
Case HI(b).
or if
Then
pn: K
F
F = {n+mx|n,m £ z ) , where
is ordered as an additive subgroup of
Archimedean, non-discrete, but a subgroup of
T
is
IR .
Q, the
rational numbers. the F
p.
The prime integers fall into two classes, p i ? q., where
occur with arbitrarily high powers in denominators of elements of
and where the powers of
say by
b..
q.
occuring in such denominators are bounded,
Then
a 1
= p
•••
r Pi
3, 3_ q / '•' *j . 0 < a. < oo, o < 3. < b. } .
We now analyze orders on
R(x,y)
which yield signed places
p:
R(x,y)-*R,± °°
with these value groups.
Case I.
Here we have a rank 2 valuation.
defines an algebraic cut of the subfield there is an algebraic function f.(x) G R[x]
with
irreducible in
f(x,y)
R(x) C R(x,y).
y,
In other words,
f(x,y) = fQ (x) +f 1 (x)y+--« +f m (x)y m ,
infinitesimally small relative to
R(x), and
R[x,y].
The curve
f(x,y) = 0
discussion in 8.11 or 8.3.
V
One of the variables, say
/
thus has branches through the origin, as in our The order
$ C R(x,y)
picks out not only a real
\ \ \ \ \ \
root of in
R
2
f(x,y) over
R(x) (which is a branch of the curve
f(x,y) = 0
at the origin, as indicated by solid lines in the figures), but also
a side of this real root in the real closure of
R(x) (as indicated by the
normal arrows to our branches in the figures). With these choices made, the £-sign of a function determined as follows.
If
g(0,0) > 0, then
g
g(x,y) G R[x,y]
is positive.
If
is
g(0,0) = 0,
243
but
g(x,y)
does not vanish on our branch
x
g( >y))> then to the origin.
g
(that is, f(x,y)
has the same sign as its values on Finally, if
small normal curve to the values of
C
g
g
vanishes on
does not divide
C, infinitesimally close
C, then
g
will not vanish on a
C, near the origin, and the sign of
g
is the sign of
on such a normal curve, on our preferred side of
C.
(In
fact, this last "test" actually covers all cases.) The value group
V
is
determined by first writing is, h h
x
z
0 =
2 a..x y , i+j=l l j
h
f ) . The invariant
vanishes at the origin, along
r = Z,
The value
g = f h, where
is not divisible by
Case II.
Z-
A = R.
(m,n)
assigned to
g
does not vanish on
C
n
is (that
is then the order to which
C.
In this case, there is an "analytic curve"
a.. £ R, with non-trivial real "branches" at the origin, 1J
and the $-sign of a function
g
selected branch of the curve.
is the sign of the values of
g
on a
Now, what does this mean in light of the
fact that convergence may not even be sensible, and even if it is sensible, a. .x1y-) = h(x,y)
2
may not converge?
The answer is provided by looking
13
i+j=l
a t t h e honest curves
0 = h r ( x , y ) , where
hr(x,y) =
r I
i
The point is the infinitesimal behavior of branches of origin will stabilize.
precisely because
g
h(x,y)
h (x,y) = 0
at the
g(x,y)
on these branches, near the origin.
is positive or negative will be stable.
It is
is not a polynomial, that our branch for
0 = h
differs slightly from that for the branches chosen for all
r.
the (stable) order to which
g
branches of the curves
r -• °° .
•*
We can coherently select branches in the limit, as
r -> °° , and measure the sign of The decision of whether
i
a^.x y J , as
i+j=l
0 = h .
0 = h , and no The value in
g(x,y) ^ 0 r = Z
can vanish on
assigned to
g
is
vanishes at the origin on the selected (In other words, the degree of the first
non-zero term in some power series.) As a concrete example, consider
2 n!x . Given g(x,y), n=l we look at the behavior of g(x,y) infinitesimally close to the origin, along the honest curves y = 2 n!x . As r gets larger, we must look nearer and n=l
244
h(x,y) = y -
nearer the origin, but "convergence" of
Case III (a). one example here.
h(x,y)
is irrelevant.
Y - {m+nT|m,n £ Z, T irrational}, Intuitively, the '£-sign of
A = R.
g(x,y) £ R[x,y]
yielding this value group is computed by restricting dental curve", for example, r./s-
converging to
T
y = x .
We give only
g(x,y)
for an order to a "transcen-
What one really does is find rationals
and then restrict
g(x,y)
to the curves
y
= x
,
or more precisely, to coherently chosen branches of this family of curves.
Case m ( b ) .
F Cj},
A = R.
Again, we give only one simple example.
We might have a series representation (mi/m) < (m2/n2) < ••• ^ F.
y = x *
2
* +x
2
+ ..., where
This representation is purely formal.
finitely truncated formulas define
The
honest algebraic curves with "stable"
infinitesimal behavior near the origin.
We test
g(x,y)
by restricting to
suitably selected branches. It is clear that the "general" total order on
R(x,y) is complicated.
This is consistent with our set theoretical characterization in the preceding section in terms of ultrafilters of certain open subsets. algebraist should only tolerate the orders of type I valuations on fields of functions in
r
variables).
In general, the
(discrete, rank
r
These are more in
line with our philosophy of finite algebraic computability.
The other types
of orders are perhaps interesting to analysts. For example, consider a first order differential equation Q(x,y)dy = 0,
P , Q G R[x,y],
Q(0,0) t 0.
P(x,y)dx +
We ask what interpretation can
245
be made, for an arbitrary real closed field say through (0,0)?
R, of the "solution curve",
Although we do not expect an honest curve, we do have a
procedure for deciding if a polynomial
f(x,y)
is infinitesimally positive
or negative at the origin, along our "phantom" curve.
One approach is to
00
just take the formal power series solution
order, as in Case II above.
y =
2 a.x i=l X
and construct an
However, it is much better to just use the
differential equation itself to decide if a polynomial ought to be positive or negative along this curve, say in the positive x-direction, near the origin. Specifically, first look at (df/dt)(0), where
know
is the evaluation of
f
f(0,0).
y = y(t)
If
f(0,0) = 0, then we want to
is the "phantom" curve and
on the curve
t > 0.
f = f(t,y(t))
This doesn't make sense, of
course, but the result of applying the chain rule does make sense,
Sf <«) - If (o.o) - If (0,0)>$$. since the differential equation says computation of order.
If
(df/dt)(0)
dy/dt = - P(t,y)/Q(t,y).
is positive, then
algebraic solution
is positive in our total
(d 2 f/dt 2 )(0), and so on until we
(df/dt)(0) = 0, we compute
finally reach a decision.
f
If this
Of course, if the differential equation has an
f(x,y) = 0, we are in Case I above, rather than Case II.
This sort of interpretation of differential equations seems quite reasonable and worth further study.
Equations of higher order and behavior
near singular points are topics to be investigated.
Of course, one wants to
make sense not just of the germ of a solution curve at an initial point, but also the continuation of the curve. one wants to associate an order
c
$ x
values of
xQ
in some interval.
o
That is, given an initial point,
R[x,y], with
D R (x) = x , for all K o
This collection of orders will play the
role of the solution curve of the differential equation through the initial point.
246
8.13.
Brief Discussion of Structure Sheaves The material in this section is partly in the form of an outline of a
discussion to be worked out in detail elsewhere.
On the other hand, the
ideas are quite fundamental for our program of algebraizing topology. Let
S
be a closed, semi-algebraic set, identified with the maximal (A,'£) = (A(S) ,»B (S)).
convex ideal spectrum of an affine coordinate ring
Recall from 8.9.3 that this situation is intrinsically characterized as follows. A
is a reduced R-algebra of finite type.
primes
P. C A
$• = 'P [g-J I
w
with on
A -* ITAi.
(0) = Hp., and finite refinements of the weak order
A- = A/P.
ij
The set
There are finitely many '^-convex
such that
*J3 = A HII0&.) , under the inclusion •*• d
1
1
S
is the union of the sets
S i = X(Ai,('Bi) ) , each of
which is the closure of the set of algebraic simple points variety
X(A.,# )
with
g.-(x) > 0.
non-empty strong open set We will refer to such
x
of the real
In the semi-algebraic sets
U{h.} = {y £ S-|h.(y) > 0}
S., every
is Zariski dense.
as -irreducible components of
S. = X(A.('J$.) )
S.
In Chapter V, we constructed rather generally a structure sheaf relative to the Zariski topology on
Spec(A,$).
The global sections turned out to be
the ring
(A CM >., ']) en J
than
Because of the Nullstellensatz, this amounts to inverting all
1.
functions
f £ A
obtained from
with no zeros on
S.
A
by inverting all elements greater
Also because of the Nullstellensatz
we can simplify the discussion by restricting this sheaf to the maximal convex ideals
X(A,'£) = S.
Zariski open set elements of on
A
D(f) C S
x £ S.
f £ A, the ring of sections over the basic
is
(Acr,~ , '£C/-.pO> obtained by inverting all
with no zeros in
D(f). Thus elements of are
are functions nowhere
D(f). The stalks of this sheaf are the local rings
(Ax>'£x)>
Elements of
neighborhoods of
x
Ax
^cr^
those functions in
Acr4~ Ag .f .
D(f). The elements of
negative on
If
may be regarded as germs of functions on Zariski open
and such a germ belongs to
where negative on a Zariski open neighborhood of On the semi-algebraic set
S
£x
if and only if it is no-
x.
we also have the "strong topology", that
is to say, the collection of open, semi-algebraic subsets
U C S.
We would
like to study "sheaves" for this strong topology, but classical sheaf theory
247
is intimately tied to the infinite procedures of point set topology, so we should proceed with some caution.
Perhaps the most natural finiteness con-
dition to impose is that we seek sheaves for the Grothendieok topology on the set of open subsets {IK -* U}
are allowed.
U Cs
in which only finite covering families
When discussing open, semi-algebraic sets and
finite open coverings, it is obviously very convenient to assume Unproved Proposition 8.1.2. of any sets
U C S
This assures us that, essentially, finite open covers
just amount to writing
U{f.} C S.
U
as a finite union of basic open
On the other hand, all the propositions proved in this
section are independent of Unproved Proposition 8.1.2. We have in mind three sheaves of rings, in fact, associated to any semi-algebraic set
E C Rn.
We call these sheaves the rational structure
sheaf, the semi-algebraic structure sheaf, and the smooth structure sheaf. Each structure sheaf corresponds to a category of morphisms between semialgebraic sets,
although we do not study these morphisms here.
These
three categories can be interpreted, within real algebra, as delineating the three subjects, algebraic geometry, algebraic topology, and differential topology. (I)
Let
E C Rn
be a semi-algebraic set, A(E) = A(!) = R f X ^ . ,Xn]/I(E)
the affine coordinate ring of embedding of
E
E.
Throughout this section the particular
in affine space can be suppressed.
invariantly as some dense subset
We can think of
E
n
E = UE.,
E. = Z{f..} U { g . v } C X (A (!) ,'£ (!)) ,
1
of the maximal ideal space of an RHJ-algebra
1
1J
(A,'£)
of a certain type.
IK
None-
theless, it will sometimes be convenient to refer to the distance between points in embedding. form.
Rn,
||x - y||. Thus we do not go out of our way to avoid an affine
The reader can reformulate for himself all statements in invariant
For example, if we say a subset
S C E
then this can be reformulated by saying that all
h £ A
are bounded as functions on
is closed and bounded in S C X(A,'£)
R ,
is closed and that
S.
The rational structure sheaf is very similar to the structure sheaf for the Zariski topology.
We first define it, then prove a few propositions
which enable us to compute rings of sections, in some sense. is (relatively) open, we define
248
I(U) = {f
E
If
A = A(E)|f(U) = 0}.
U C E Then
A/I(U)
is a ring of functions on
U
and we define
A(U)
to be the
localization of this ring obtained by inverting all functions with no zeros on
U.
We define an order
negative on
U.
•£ (U) C A ( U ) , consisting of all functions nowhere
We obviously have a presheaf of partially ordered rings
for our Grothendieck topology and the rational structure sheaf of the sheaf associated
E
is
to this presheaf.
In order to make computations, let us first consider the case where E = S = X(A,*0), where
A
is a finite integral domain over
R, and
is a derived order of a finite refinement of the weak order. $ = 0#,T[g.]) w 1 d Any non-empty basic open set U = U{h.} C s is Zariski dense, hence A(U) = A.
Proposition 8.13.1. order
$' = (:p[h-]), J d
an RHJ-algebra.
C
A
Let
coincides with
Secondly,
D(h) Cx(A,$»), where
U = U{h.} C S , as above.
U CX(A,'|5)
h=nh..
Then, first, the
('£w [g-,h.]) , hence ! J d
Proof:
C
C&Jg^V )d
m
i
j
(CVgi] ) d [ h j ]
w
j
I
since
h. > 0, J ~
II h. ^ 0 J
means all
}
= d
m _> 0}.
*[hj]V
BUt
is nowhere negative on
J d by Proposition 8.6.5.
tp [g.,h.]
It is completely obvious that
D(h) CX(A,'£ ! ),
£ f, rel •£' , some
also, i t is easy to argue that any f ^ 0B[h.]) U(g.,h.}, hence belongs to
which
Finally, (A(U),fl(U)) = (Ag ( h ) ^ s f h ) ^
S(h) = {f <= A|0 £ h
Certainly
is
coincides with the Zariski open set
is the ring of sections in the Zariski structure sheaf over obtained by inverting
(A,$')
d
U{h.} CX(A,$) h- > 0. J
coincides with
D(h) C X C A
The last statement follows from
the Nullstellensatz.
•
In the case under consideration here, a function
f G A(U)
of the
presheaf is completely determined by its restriction to any non-empty smaller open set
V C u.
components".
In particular, this presheaf has no respect for "topological
The effect of the sheafification is exactly to allow independent
rational functions over the components of each More precisely, let us say
U
is connected if
union of two open semi-algebraic subsets. a finite cover of
U
U. U
is not the disjoint
Equivalently, if
by open sets, then for any two indices
{U. -• U} i,j
is
there is
249
a chain
i = i , i1,...,i
= j
Suppose we have functions assume all
g. > 0, and I —
A(U), since
Zg._ > 0
on
such that
U.
f-/g- ^ A(U.)« g. > 0 I
on
U.. I
U = UU..
n U.
Since Then
Suppose
£ 0,
f-/g- = f-g-/g-> we may f/g = If./£g1 1
( f ^ ) \u
i,j.
Then if
A, since
U. n u.
(f/g)I,,
then
U. n u . f- 0, we have is Zariski dense.
= f-/g-> all
sheaf defined by the
belongs to
= C^/g^lu nu ' j i j in the integral domain
n u
i all
k = l,2,...,n.
f-g. = f-g-
It follows that if
U
is connected,
i, hence the section of the rational structure
f-/g- ^E A(U-)
already comes from the presheaf element
f/g e A(U). On the other hand, we will prove below that any semi-algebraic set is a finite union of connected semi-algebraic subsets.
Thus, if an open
is not connected, a section of the rational structure sheaf over of a section over each connected component of
U.
U C E
U
consists
Thus we have proved the
following.
Proposition 8.13.2.
If
U C S
is (relatively) open, S = X(A,$)
ducible and non-degenerate as above, and U
U = U u^
is the decomposition of
into disjoint, (relatively) open connected components, then the ring of
sections of the rational structure sheaf over
If
E
S. = X(Ai, Bi)
and each
A^
is a domain with
finite refinement of the weak order. A -• II A i I(U) =
U is
is any semi-algebraic set, we can write
<
A
irre-
and
$(E) = A n n ^ i -
n p. U ns. £ 0 J 3
= A/I(U)
since if
If
If
p^
A
O
E = U s^, where the derived order of a
A = A(E), we have the inclusion
U C E
is (relatively) open, then
U n s. £ 0, then 3
with inclusion
II(A(U.) /p(U.)) •
I(U n S.) = P.. J J
Let
"* II A/P., the product taken over those
j
with U n s. £ 0. Proposition 8.13.3.
If
U = U{h.} n E
the presheaf sections (A(U) ,$(U)) the localization of such that (That is, $
250
(Ao>'£ )
are obtained as follows.
$
C A oo
is the order
consists of the functions nowhere negative on
E, then
(A(U),'B(U))
obtained by inverting all functions
Z{f} n u { h . } n f C f - E , where I
is a basic open set in
U.)
is
f £ AQ A
nn(D.[h.]) . 031^
In particular,
if
E = E
is closed,
then
(A(U) ,$ (U))
is the ring of sections of the
Zariski structure sheaf associated to the RHJ-algebra Zariski basic open set
Proof: each
D(h) C X ( A Q , ^ o ) , where
h = II h^.
This is essentially the same as 8.13.1.
tp.[h^]
c
A-
(A ,£ ) , over the
Note that from 8.6,
is an order giving an RHJ-algebra, precisely if
V{\} n S. t 0.
D
In general, in the reducible case, it is not true that if
U C E
connected, then every section of the rational structure sheaf over comes from the ring
A(U).
of the connected plane set U 2 , U,
is
U
already
In the figure below, the affine coordinate ring
E
is
and the rational functions
A = R[X,Y]. y^AQjp,
But
E
is covered by
0 €= A(U.),
together to give a globally defined rational function on
U, ,
j = 2,3, fit
E.
A precise
description of the global sections of the rational structure sheaf seems somewhat complicated.
Remark.
As a general rule, when sheafifying a presheaf with respect
to anon-topological sort of Grothendieck topology, the description of sections in terms of elements of stalks at each point which locally fit together is not appropriate.
However, it turns out that the rational structure sheaf defined
above actually coincides (on semi-algebraic open sets) with the classical topological sheafification of the presheaf defined by the
(A(U) ,'£
251
The stalks are clearly the rings the localization of Here
I
=
X
I(U)
lim
A/I
I(U)
(A ,'£ ) , where, if
defined by inverting functions
is the ideal of degeneracy at
x
x^E, g
A
with
g(x) f 0.
(which coincides with
~G%
if
U
is sufficiently small).
a function at
x E E.
s E A y y
U
s
£ A , for all
with elements X
for all
y G u . x
'£ C A x x
A
is a germ of
are the germs which
x.
Suppose given an element has a neighborhood
Thus an element of
The positive elements
are locally non-negative near
f /g
x £ U, such that each
E A(U ) X
X
which restrict to
X
Now, U
not compact, so we cannot assert that finitely many of the
cover
However, let
f
f /g
l^gl'
f
V
= D(g ) C E
a
§ree
on a
those irreducible components of
E
with
U is U. g
i 0.
Moreover, if two functions
u^ n u*2, then they will
U(S. - Z(g 1 ) - Z(g2 ))
where the
S. n \J^ n U 2 ± 0.
is possible to replace the data defined by the fx-/gx-
V .
semi-algebraic open set
also agree on the much larger set
functions
U
denote the Zariski open set defined by
is defined on the much large set
2^ g 2
x
Such data exactly describes a section over
of the classical sheaf associated to the presheaf on question.
Thus
is
f /g
€E U
are
In this way, it
by finitely many
defined on larger neighborhoods of the form
This says exactly that the sheafification of our presheaf
S.
UHvXi
(A(U),$(U))
HS^. with
respect to the classical topology and the finite Grothendieck topology actually coincide.
II.
We next define the sheaf of semi-algebraic functions on a semi-
algebraic set
E C Rn.
Again, by "sheaf" we mean sheaf for the Noetherian-
Grothendieck topology of finite, open, semi-algebraic covers of open subsets of
E.
We first define semi-algebraic functions and establish many basic
properties. Let
f:
E -* R
consider the graph of
Definition.
be a function. f,
Then
E x R
is semi-algebraic, and we can
F C E x R.
The function
f:
E -*• R
is a continuous (semi-algebraic)
function if the graph
F C E x R
semi-algebraic subset
S C E, which is closed and bounded in
252
is a semi-algebraic set, and if for each R n , the graph
of
f
over
R" X R =
R
S,
n+1
F n (S x R ) , is a closed, bounded, semi-algebraic set in
.
Note the definition involves both local and global properties of the graph of
f.
with graphs (f^f^: <
If we are given two such functions F, C E x R
""l ^ F l^ n ^2 ^ F 2 ^ ' S C E, then R
n7T
E •> R
and
F 2 C E x R, consider the graph
f2 :
F, x F 2
E + R x R, that is, {(x.fjCx) ,f2(x)) |x G E } C E X R X R.
^1(x,y1,y2) = (x,yx)
n
and
f •
F,
and
wnicn
TT2(x,y1,y2) = (x,y 2 ), then obviously
i s semi-algebraic.
x
is certainly bounded. 9 i s Fz
If
S
E -*- R, of
If
Fj x F 2 =
is closed, bounded, and
Also, F, x F 9
is
z
is closed in
x R x R , which is directly seen from the formula F
2 ^
2
n
^
S X R
^'
since
sum and
F, x F o = IT" (F, H ( S X R ) ) 1 l l S l Product are polynomial maps R x R -> R,
we conclude from this paragraph and 8.9.4 the following result, by projecting F, X F - J C E X R X R - ^ E X R ,
Proposition 8.13.4.
usin
g
sum or
product
R x R -> R.
Sums and products of continuous semi-algebraic
functions are continuous semi-algebraic functions.
•
As another immediate corollary of 8.9.4, we get a generalization of the remark following 8.9.4.
Proposition 8.13.5.
A continuous semi-algebraic function assumes a
maximum value on any closed, bounded, semi-algebraic set.
•
We have chosen to define continuous functions without the usual and
S's.
The
e-6
e's
definition turns out to be equivalent, although this
is not completely trivial.
Since the
e-6
definition is more convenient
for certain arguments, we will now establish this equivalence.
First, we
need a lemma which guarantees that the local part of our definition of continuous function really carries some information.
The following
proposition is a sort of "curve selection lemma".
Proposition 8.13.6. x £ E
Let
E
a point in the closure of
be a non-discrete semi-algebraic set, E.
Then there exist closed, bounded,
253
one-dimensional, non-degenerate semi-algebraic sets
C = C,, with
x e C C E U {x}.
Proof:
First, we may assume
finite union of such sets.
E = Z{f.} HU{g.}
Then if
E = US.
into irreducible components, we have
since any
A
is a domain, and
the weak order.
$
In fact, the
$, that is, we may assume
is a
is the decomposition of
E
E = U(S. H E ) , and therefore we have,
say, x £ S , H E . This allows us to assume, in fact, that where
E
E = U{g.}CS =X(A,
the derived order of a finite refinement of g.
will actually be subsumed into the order
•$ = ($ [g-]) C A. w J d
Also, we may as well assume
Now we are in a position to apply the signed place perturbation theorem, 8.4.9 (or its extension in the Remark following the proof of 8.4.9) to the domain refining
A = R[x 1> ..x ] .
We know that there is a total order
0 £ R , is 'B'-convex.
order and find a strong chain of 'B-convex ideals (x,...x_) C A, where in
tr.degD(A/P.) = r - i. K l
8.4.9 allows us to assume The set of zeros of
g. £ P r P
semi-algebraic set, x = 0 a discrete subset of
C • .. C P
=
is thus a one-dimensional
is a non-degenerate point, and the . ) . Also, Z(P
Nonetheless, we obtain our "curve"
secting
Z(P
degenerate points of
0 C p
1.
points.
.) C s
We may thus perturb this
Moreover, the Remark following
, x € Z(P _.) C s
Z(P
A,
(x, .. .x^) C A, in
'B = (ftw [g-]) , such that the maximal ideal J d
corresponding to the point
'£' on
g.
vanish on
.)
may contain some degenerate
C
of the proposition by inter-
with a closed ball sufficiently small to exclude all Z(P
.)
and all zeros of the
g., except
x = 0.
•
Next, we need to know that a semi-algebraic function satisfies a "polynomial equation" with the coordinate functions.
Proposition 8.13.7.
If
E C Rn
a function such that the graph of is a polynomial
cp:
E -* R
cp, F C E x R, is semi-algebraic, then there
P(x;....x ,y) such that
x = ( x r . . x n ) € E.
254
is a semi-algebraic set and
P(x,...x , cp(x;....xn)) = 0
for all
Proof:
Express
E = U Z{£. -} H u { g . , } . If i 1] IK Z{£..} U { g . , } , then at least one f. . occurs. ij IK 13 n
(x,t),
a whole interval of values be the graph of a function.
Proposition 8.13.8. function.
Then
graph of
f,
0 < e G R
f
Otherwise
cp(x), and thus
We can now set
P = FI ( 2 f. . ) . i j 1J
E
be a semi-algebraic set,
would contain
F
would not •
f:
E -* R
a
is a continuous semi-algebraic function if and only if the
F C E x R, is a semi-algebraic set and if for all
there exists
Proof: bounded set.
0 < 6 = 6(x,e) G R, such that
First, we assume the
e-6
property.
x G E,
|jx-y|| < 6,
Suppose
yGE
is closed, bounded. since
S
If
is closed.
t = f(x), and
F
a closed,
We must show that
t / f(x), the
near
(x,t)
e-6
FH(SxR) x G S,
property easily implies
and this gives a contradiction.
(x,t) G F H ( S x R ) , as desired.
We will prove that
F H (sxR)
The result is obvious if
is a domain
is bounded by induction on
dim(S) = 0, since then
Also, we may assume
A
S
(x,t) G F H ( S x R ) C R " X R, then certainly
Now, if
there are no points of
points.
S C E,
(We will frequently drop the adjective semi-algebraic, it being
understood that all sets must be semi-algebraic.)
where
F
|f(x) - f(y)| < e.
implies
Thus
(x,cp(x))G
near
Let
t
x € E,
•£ C A
S
f
is a finite set of
is "irreducible", that is, S = X(A,'B) , d
a finite refinement of the weak order.
is justified because from 8.9 any closed ducible pieces and if
S
dim(S).
is unbounded on
S
This
is a finite union of such irreS, then
f
would be unbounded on
one of these pieces. Suppose where
f
satisfies the polynomial equation m
P(x,t) = p Q (x)t + ••• + p m ( t ) .
We may assume
since otherwise we could just drop this term from is a closed, bounded subset of lower dimension. for all
xf G s 1 .
Consider
0 = P(x,...x , f(x,...x ) ) , P Q (x) / 0 G A P(x,t).
Thus
= A(S), Z(p (x)) = S'C S
We may thus assume
U = {x G S|f(x) < b}.
Then
S - U
f(x') < b G R ,
is a closed, 2
bounded, semi-algebraic set, disjoint from all zeros of assumes a minimal, non-zero value on of
P(x,t), for
S - U.
x G S - U , henc * a bound on
P o (x).
Thus
(pQ(x))
This puts a bound on all zeros f(x), for
x G S-U.
As , f
255
is
also bounded by definition on
U, we have
Next, we go back and assume Suppose for some y = y(6) E E EQ EQ
x G E,
with
f
0 < e £ R
|x-y| < 6
= {y E E| |f(x) - f ( y ) | >^ e}.
but
f
bounded on
S.
continuous and prove the and any
e-6
0 < 6 £ R, there is a point
|f(x) - f (y) | _> e.
Since the graph
F
of
Then f
x E EQ, where is semi-algebraic,
is indeed semi-algebraic, as the image under projection
a semi-algebraic subset of
F.
f
is continuous, F H ( S x R ) = F Q
In general, if
FQ C R n + 1
E x R -* E
of
By Proposition 8.13.6, choose a closed,
bounded, one-dimensional, non-degenerate Since
property.
S
with
x E S C E Q U { x } C E.
is closed, bounded.
is a closed, bounded set, and
closed, the squared distance function
d'.
F Q -• R, defined by
B C Rn+1
is
dg(y) =
min ||y-b|| , obviously has the e-6 property. (For fixed y, |y-b| is b E B large for b ^ B far from y and does assume a minimum on any closed, bounded subset of
F
B, hence definitely assumes an absolute minimum on
as above, x E E
the t-axis above
our fixed point, and
x = (x^-.x^.
B.)
We take
B = {(x,t) |f(x)-t| >_ e}, part of
In this case, d g
is easily computed
explicitly:
fz(x!-x,) 2 if (x'...xM ,t (X....XJ if
256
(X-...X-) = (X....X1 .
In particular, it is clear that the graph of and, as already observed, has the
e-6
d R , above
property.
Thus
continuous by the first part of the present proof. a minimum value on establishes the
is semi-algebraic,
2 dfi:
F Q •> R
is
It must therefore assume
F , which it obviously does not do.
e-6
F ,
This contradiction
property for our original continuous function
f.
•
We have some immediate corollaries of 8.13.8.
Corollary 8.13.9. and
U C R
If
f:
E -* R
is a continuous semi-algebraic function
is an open, semi-algebraic set, then
open, semi-algebraic in
E.
Corollary 8.13.10.
If
z €= E
is (relatively) D
E
is a connected semi-algebraic set, f:
a continuous semi-algebraic function, then there exists
f" (U) C E
with
x,y G E, and
E -*> R
f(x) < t < f(y) *= R,
f(z) = t.
•
The following result will be useful in the last part of this section.
Proposition 8.13.11.
Suppose
semi-algebraic subset such that semi-algebraic function.
E
E Cf ,
fQ
Suppose
fQ:
E Q -* R
f:
and which has the
e-6
E "* R.
a
is a continuous
Then there is at most one extension of
continuous, semi-algebraic function which extends
is a semi-algebraic set, E Q C E
fQ
to a
Moreover, any function
property on
E
f
is, in fact, a semi-
algebraic function.
Proof:
The uniqueness of
continuous functions.
f
is obvious from the
Also, given an extension
property, it is clear that the graph of in
E x R
of the graph of
fQ,
f,
f
of
e-6 f
property of with the
F C E X R, is just the closure
F Q C E Q X R C E X R C E" X R.
We made use above of the distance function to a closed set say
dR:
R n "* R,
d R (y) =
e-6
min |y-b|. b £ B always continuous, semi-algebraic.
D
B C Rn,
We now want to prove that
dg
257
is
Proposition 8.13.12. :
R
n
-*• R
Proof:
If
B C Rn
is closed, semi-algebraic, then
is continuous, semi-algebraic.
The
e-6
property of
dn
is immediate from the triangle inequality
D
for distances.
Thus we only need to prove that the graph of
is semi-algebraic.
dg
Rn x R
in
To see this, begin with the subset
D = { ( x , b , ||x-b||)|x G R n, b G B} C Rn x B x R . D
is obviously semi-algebraic, hence by Tarski-Seidenberg, so is the image
of
D
TT: R n x B x R -• R n x R.
under projection
Now, ir(D) C R n x [0,co),
and the subset
D1 = { ( x , t ) | x G R n , 0 < t e R ,
is also semi-algebraic.
This again follows from Tarski-Seidenberg since
is defined by an elementary sentence. now be described as
[ 0 , t ] n ^(D) = 0}
1
(F
dg :
Note that the graph of
!
n
- D ) U (B x {0}) C R
x R.
Thus
dg
D'
R n -• R
can
is a semi-
algebraic function.
•
As a corollary of this argument and 8.13.8, we can prove that continuous semi-algebraic functions are uniformly continuous on closed, bounded sets.
Corollary 8.13.13. and
f:
S -* R
there is a
Proof:
If
S C Rn
is a closed, bounded semi-algebraic set,
a continuous semi-algebraic function, then for all
0 < 6 G R
such that if
For each
x,y G S
x G S, the set
S
—————
and
= S
X,£
0 < e £ R,
||x-y| < 6, then
= {y G si If(x)-f(y)I > e} is X
I
—
a closed, bounded semi-algebraic set, by Corollary 8.13.9 (possibly but the set of such
x G S
with
S
obviously any modulus of continuity points
x.
= 0 6
so is
258
Then
BQ
It is not hard to prove that
AS = {(x,x)|x G s} C s x S, say using the
= 0,
e-value at these
Anyway, the rest of this proof goes through even if some
B = B Q C s x s.
S
is open and can be discarded, since
will work for our
B Q = {(x,y)|x G S, y G Sx> C S x S.
Let
|f(x)-f(y)| < e.
e-6
S
=0.)
is semi-algebraic, hence S H B = 0, where
property of
f.
Thus, the
distance function 6'
on
AS.
dR:
AS -• R
assumes a minimum, strictly positive value
We are here computing distances in
no greater than the distance from if
0 < 6 < 6', then
for our given
e.
x
to
S
R n x Rn.
in
Obviously, 6 !
R , for any
<5 is a uniform modulus of continuity for
(If
B = 0, we define
is (relatively) open in
E, we define
C°(U)
with
f -Li fin = fi.
^
=
U = Uu.
Remark:
E.
If
U C E
'p°(U) C C°(U)
the
It is obvious that this presheaf is a
is a finite open cover of
(U)
S •
U, and
^*ln fin ' then there is a unique
Moreover, f
on
to be the ring (by 8.13.4) of
continuous R-valued semi-algebraic functions on subset of nowhere negative functions.
f
Thus,
d g = + °° .)
We now define the sheaf of semi-algebraic functions on
sheaf, that is, if
x £ S.
is
if and only if all
U
and
f £ C (U)
f i G C°(U i) with
£i ^ ^ ( U p .
Semi-algebraic functions are not infinitely collatable in
general.
Graph of
f
U = (-1, +1)
In the figure, f
is a function on an open interval which is locally but not
globally semi-algebraic.
Remark:
Very little work is required to extend the notion of continuous
R-valued semi-algebraic functions to vector valued functions where
E C Rn.
First, the graph
one can give either the
e-6
F C E x Rm
f. :
E -* R,
f = (^•••fm)
must be semi-algebraic.
Then
S C E.
With either definition,
continuous exactly if all the coordinate functions
1 <^ i <^ m, are continuous.
equivalent, by 8.13.8.
E -* R m ,
definition or the closed, bounded property of
the graph over any closed, bounded subset one proves
f:
Thus, the two definitions are again
It is also easy to prove that compositions
E ~> E' -*• E"
259
of continuous, semi-algebraic functions are continuous, semi-algebraic, where E C Rn,
E1 C Rn ? ,
if and only if
E" C R n", and that
go f G C°(E), for all
f:
E -+E 1
is semi-algebraic continuous,
g G C°(E').
We end this subsection with a proof of the fact that any semi-algebraic set E
E
has only finitely many components.
is connected if
open and closed in
E E.
Recall that a semi-algebraic set
has no proper semi-algebraic subsets which are both We want to prove that any
E
is a finite union of
connected semi-algebraic subsets which are both open and closed in
E.
These
subsets are obviously maximal connected subsets, which can then be called the connected components of
E.
Although we have notions of open sets, closed sets, connected sets, and so on, we must be careful not to assume facts from general topology which use infinite techniques inappropriate in our case.
For example, we cannot
routinely make use of the union of all connected semi-algebraic subsets containing a point
x G E
(which is how connected components are constructed
"topologically") because there is no reason a priori that this union is semialgebraic. If
E
E 1 ,...,E m> then we
is any union of finitely many connected sets
can consider maximal connected subsets of the form
are routinely shown to be both open and closed in components of
E.
sets, we can write
E^ U ... U E• . x x l k
These
E, hence are the connected
If an E exists which is not such a finite union of connected E = Uj U uj
are open and closed in
E.
where
l^ n uj = 0
Then, say, U-,
splits
producing an infinite strictly decreasing chain and closed subsets of
E.
The sets
and both
l^
and
U|
U-, = U 2 ^ Ui, and so on, U^ ^ U 2
V i = LL - U. + ,
c
U- ^ •••
of open
then give an infinite
collection of pairwise disjoint, open and closed, semi-algebraic subsets of
E.
We refer to this process as a splitting process. Clearly, we need only consider E of the form Z{fi> n U { g . } C R n , where 1 _< j _< k, say. Then E is the image of the real variety V = Z{f.,y.g. I 3 j - 1} C R n x R k = R n + k , under projection and we adjoin new variables since if
260
V. C E,
IT: R n + k -+ R n , where
Xi***^-
Tnus
it:
f^g. e
R^.-.xJ,
suffices to study varieties,
i > 1, are disjoint, open and closed subsets, then so are
IT
(V^)
c
assume
V.
V
If some
V
is not a finite union of connected sets, we may
is irreducible and of least dimension with this property.
that any O-dimensional semi-algebraic set is finite, so Let
VQ C v
singular set. where
E^V
be the algebraic simple points and
Then
is connected.
closed subsets.
Each
We split
E.V
V = l^ U UJ
is either in
continues with, say, Uj = U*2 U U£.
U,
Those
Now, the conclusion we want is that
VQ
dim(V) > 1.
EV = V - V Q
dim(EV) < dim(V), so we can write
Note, if
U|
in
into disjoint, open and
or
Ui.
Z^V C ^
This splitting process
are either in
V^
U1
less than
contains all
as above.
l£J, then
of the
splitting process.
U£.
V\ ,
i >_ 1, which
U1 C V Q
is one such set.
U*2.
E^V
If the
U^ = U 2 U U', then we
If not, we must start over and go back and split
V^ = U^.
and some in k
or
R .
UJ, but at least we have are in
U2
actually contains infinitely many
splitting process continues infinitely, starting with get all our
the
EV = Z^V U ... U E k V,
pairwise disjoint (relatively) open, semi-algebraic subsets are actually dosed
Note
Another case to consider is if some
Then we haven't even found
in whichever of
l^, U'
EiV
V 1 , but there are
begins the infinite
It follows, then, that sooner or later the splitting
process gives infinitely many disjoint, relatively open in
V\ C v , which are semi-algebraic, pairwise Rn.
V , and closed in
We are now in a position to use Whitney's proof [44] of the finiteness of the number of components. V
and a point
cuts the manifold
x
G V,
V..
and
xQ.
Let
g:
V -• R
transversal ly at x
be the function
x .
V^
connected sets. arranged that
In fact, we can assume that
g(x) = ||y-x| 2 . V
is a closed, semi-algebraic set in
value at, say, x^ G v^.
Then
x^ G V , hence
V
of lower dimension.
V
If
I(V) = (f±)
= {x1 G v|rank(df^(x),dg(x)) = r}. R n,
g| v
assumes a minimum
is not a finite union of
On the other hand, by our choice of x Q f- V , hence
||y- x || centered at
lies in the tangent plane to V..
codim(V) = r, we consider the subvariety
Since each
We bhoose a point
so that the sphere of radius
the normal vector to this sphere at at
dim(V-) = dim(V ) > 0, since, in fact,
is homogeneous, in the sense that it is a manifold.
y GR y
We know
y
and
is a proper subvariety of
x Q , we have V, necessarily
This contradiction proves our desired result.
261
Proposition 8.13.14.
Every semi-algebraic set is a finite union of pair-
wise disjoint, connected, open and closed, semi-algebraic subsets.
Remark: of
E
A further result is that if
which contains
x
is the set of
semi-algebraic "path" p:
•
x G E, then the connected component
y £ E
[0,1] -> E with
such that there is a continuous,
p(0) = x, p(l) = y, where
[0,1] C R
is the unit interval. We will not prove this result here, although the results of this chapter are sufficient for constructing a proof.
Essentially, one
must look closely at global stratification and local geometry near a point. Note that if our ground field is the field of real numbers, it is not obvious from the considerations leading to 8.13.14 that a semi-algebraically connected set is topologically connected.
However, since
[0,1] C K
is both
topologically and semi-algebraically connected, the fact that components are path components in general, does imply this fact about real numbers.
IE.
Finally we discuss briefly smooth semi-algebraic functions.
simplicity, we restrict our attention to open subsets
n
U C R.
For
We begin by
giving a little more structure to the graph of a continuous semi-algebraic function
f:
U -• R.
Proposition 8.13.15. and suppose
P(x,...x ,y)
P(x,f(x)) = 0 ,
x £ U.
(b)
If
P^
f:
U -* R
is continuous, semi-algebraic
is a polynomial of least degree in
Let
irreducible factors in (a) The
Suppose
P =
p # #p 1 --- r
be the
are distinct, that is, P
P± = P/Pi
and
P
into
is square free.
U, and
P^CxjfCx)) = 0
P.
vanishing on the graph of
IL
are
if x *= U^.
(P^) C RfX-^.-.X ,Y] are convex prime ideals.
(d) dimension (U - U U\) < dimension(U) = n.
262
factorization of
U\ = {x e uli^U.f(x)) * 0}, then the
(c) The principal ideals
If some
such that
RfX^-.X ,Y], then:
non-empty and disjoint open subsets of
Proof:
y
divided
In particular,
UCUJj..
P, we would find a polynomial of lower degree
f by dividing by
P^.
Similarly, if some
u\ = 0,
then
Pi
and
P.
would vanish on the graph of
have lots of zeros in
R
(P.) .
that the algebraic simple zeros of IJL. of
u\ f
are obviously disjoint, over
U-.
follows now from
8.8.6, since the
Moreover, the other results of 8.8 imply P^
are dense on the graph of
f
over
In particular, no polynomial can vanish on any open subset of the graph f
over
IL
Suppose graph of
f
unless
P^
U - U U. over
V.
divides it.
contained an open set Then
But all
P-
vanish on
dense in
P. U
Let
F C R
denote the
is convex and must have
n, which is therefore principal, say
(Q).
F, and applying the arguments of the paragraph above
Q, we would deduce that all
since the
V.
I(F) CR[X,...X ,Y]
some associated prime of dimension
to
The
obviously must vanish on the graph of
The convexity of the ideals (P.)
f.
P. G (Q). As this is clearly impossible
are distinct irreducible polynomials, we conclude
Uu.
is
and (d) follows.
•
We now want to define the subring entiable semi-algebraic functions
f:
C 1 (U) C C°(U) U -• R.
of continuously differ-
We first assume
f G C°(U),
then assume the limits f (x, . .. ,x.+e,.. .xn) - f (x, .. . x j (3f/3x.)(x) = 1
exist in
R
the
property.
£-6
x G u
for all
lim e - 0
x G u
and define functions
(3f/8x i ):
U -* R
We can apply the chain rule to our relation
with
0 = P(x,f(x)),
of lowest degree and this gives
0 = 3P(x,f(x))/3xi
= (3P/3Xi)(x,£(x)) + ((3P/3y)(x,f(x))((3f/3xi)(x))
Since
(3P/3y)(x,f(x) ^ 0
where
P = P^-.P
is dense in
U.
U, and because no
as in 8.13.14, we deduce that The function
since it can be written 8.13.11,
on
3f/3x i:
U -• R
Sf/Sx^
Pi
divides
3P/3y
V = { x G u | (3P/3y) (x,f (x)) i 0}
is obviously semi-algebraic over
(-(3P/3xi)/(3P/3y))(x,f(x)) also belongs to
if
x G V.
Thus by
C°(U).
263
V,
We refer to such functions The set of such
C (U) C C°(U)
U -* R
as
is a subring.
r
by requiring that r,
f
C -semi-algebraic functions. By iterating the procedure,
f G C r (U) C C ^ ^ U ) C ••• C C°(U),
C -semi-algebraic functions
we can define
including
f:
have continuous partial derivatives of order up to and
1 < r < °° .
hence we have a sheaf of
Clearly,
Cr-functions are finitely collatable,
Cr-functions, defined on (variable) open subsets
U C Rn. There is no difficulty whatever now in extending the inverse function theorem, Proposition 8.7.1, to the case of a (Rn,0) -* (Rn,0) injectivity of
C^-map
with non-singular derivative Y
Y = (Y....Y ) :
( O Y - ^ X . ) (0)) .
The local
is proved by standard estimate arguments using the
hypothesis of differentiability.
The local surjectivity of
Y
is proved just
as in the earlier proof of 8.7.1, using the minimum value property of continuous semi-algebraic functions on closed, bounded sets. In fact, we will sketch another proof, similar to a standard proof in the classical case of real numbers.
Beginning with any
x 1 ...x n
y, ...y
near
0, and
C -functions
(3y./9x.)(O) = id, one shows that
of the
y.. ,x_,... ,x
C -coordinate system
x., with
is a
C
y(0) = 0
and
coordinate system.
This argument uses "completeness" in the form of the intermediate value property on intervals for continuous functions.
Since an interval is
connected (in our sense), we have the intermediate value property from 8.13.10 at our disposal.
The inverse function theorem is then proved by
iterating this substitution procedure. Note that even if originally the functions and the the
y"2»»«»»y
will be
y\
x^
are the standard coordinate
are polynomials, at the second step of this proof,
will not generally be polynomials in
C -semi-algebraic functions of
y, ,X2»**«>x , but
y,,x~,...,x . Thus, this proof
would not have been feasible in the special case dealt with in 8.7. The
C°° theory is quite different from the
example, if
U C R
is open, r < °° , then
Specific examples are easy. if
x _> 0
x £ 0.
264
and
Then
f(x) = 0 f-g = 0.
if
Let
Cr
C r (U)
f,g G C ((-1,+1))
x £ 0,
g(x) = 0
if
theory, 1 £ r < °° .
For
is not an integral domain. be defined by x >_ 0
and
f(x) = x
g(x) = x 4
if
If U
is connected, then
if U = U U.
U
is an integral domain.
is the decomposition of
U
r
components, then obviously on
C (U)
(In general,
into disjoint open connected
r
C (U) = IT C (U.)»
0 £ r £ ° ° .) The C°°-functions
are known as Nash functions and have been widely studied in the case
of ordinary real numbers, [30] through [42]. following.
We use the notation n
algebraic functions at
x £ R ,
Proposition 8.13.16.
Cx
for the ring of germs of
C°° =
lim
P(x,f(x)) s o ,
(b) The function C (U) "* C
U
a connected, open,
f £ C (U). Then:
(a) There is an irreducible polynomial such that
C°° semi-
C°°(U).
0 ^ U ^ Rn,
Suppose
semi-algebraic set, and suppose
The basic result here is the
P(X 1...X n ,Y) G R ^ . - . X ^ Y ]
x G U.
f
is determined by its germ
[f] E C~, that is,
is injective.
(c) The germ
{f} =
2 K.—U
[f] £ C^
2 2JX
is determined by its formal power series
(i ) !
(i ) !
-L
Tl
• ~"iC
3
i 1
that is, C* "*-R[[X . ,.Xn]] (d) The power series
in (°)X1
'"
X
R
n
tE x i•' t X n ] ] >
C/A._ • • • O-A. n
is injective. {f}
is a formal solution of the equation
P ( x r . . X n , {£}) = 0. (e) Given an irreducible polynomial series
{f}
PCX^.-X ,Y) and a formal power
P(X 1 ...X n , {£}) = 0, then there is a germ
such that
[f] G C~
with underlying power series {£}.
Sketch of Proof:
Let
P(x,f(x)) = 0 ,
P = P^-.P
8.13.15 into distinct irreducible factors, UL C u P/P.
is non-zero.
factored as in
the open subset where
To prove (a), we must prove that no
x E U
belongs to
U. n U.,
i ^ j , because it then follows from connectedness that there is
only one
U..
If
We may as well prove
0 G u" , then we can find (xQ,f(x Q)) ^ 0.
Near
O ^ U . HU., x G u,
i^j.
arbitrarily near
(x Q ,f(x o )), the graph of
f
0
such that
coincides with
265
the zeros of
P^
The non-vanishing of
compute all partial derivatives the coefficients of
(3P1/8y) (xQ,f ( X Q ) )
(8 f/8x )(x Q ),
I = (i,...ijj, in terms of
P.^, by simply iterating the chain rule and using the
identity
0 = P^XjfCx)), near
variable
x, then
xQ.
For example, if there is only one
0
= fr tvf<xo» + IT cvf<xo» f
0
• i & (v f < x o» + 2 £ & (xo'f(xo» f
and so on.
In fact, these expressions are just the coefficients (up to constant
factor) of powers of
(x - x Q )
obtained by formally computing 00
where
allows us to
f(x)
is replaced by the power series
P^(x,f(x)),
no
k
(l/k!)f^ ; (x )(x-x ) .
2
k=0
The
°
multivariable analog is also true, of course. Now, {f},
we want to assert that the formal power series of
f
at the origin.
as in parts (c), (d) of the proposition, is a formal solution of
P,(X, ...X , {f}) = 0. finite expression by continuity of
But each coefficient in this formal expansion is a
which, by continuity of all P,
9 f/8x , and also of course
and its derivatives, can be evaluated as
1
lim
x
the corresponding coefficients of the formal computation about
of
0 -*°
x
^U,.
But we know these coefficients vanish, hence so do the coefficients at the origin.
This proves (d).
The polynomial ring and power series ring are integral domains R[X 1 -..X n ] C R[[x l ...X n ]]. over
Thus, if some power series
f
is algebraic
R[X 1 -..X n ], there is a unique irreducible polynomial (up to factor in
R[X . . . X J ) ,
P e R [ X r . . X n ] [ Y ] , with
P(X 1 ...X n , f ) s 0.
conclude from the paragraphs above that we cannot have
We therefore
0 *= U. n u.
This proves (a). But also, we conclude that if the power series then the irreducible polynomial This means that the function
P(X 1-..X n,Y)
i^j.
{f} = 0,
is just the polynomial
Y.
f = 0, and (b) and (c) are proved.
Part (e) is quite interesting.
266
if
However, we prefer to defer the proof
until we make a more detailed study of power series rings in a subsequent chapter.
•
Remark:
The
C
structure sheaf is analogous to the rational structure
sheaf, in the sense that a function
f:
U -* R
function in each germ, that is, [f] £ C , all C (U).
In other words,
C
which is a
C -semi-algebraic
x G U, is in fact already in
functions are infinitely collatable.
267
Appendix
THE TARSKI-SEIDENBERG THEOREM
In this appendix we give Paul Cohen's proof [62] of the Tarski-Seidenberg theorem [56], [57], which we state in the following geometric form.
Proposition A.I.
Let
E C Rn
the projection onto the last
(n-1)
be a semi-algebraic set, coordinates.
Then
ir: R n -> R11"1
TTE C R
is semi-
algebraic.
Actually, the proof provides an algorithm for writing down semi-algebraic set, in terms of the polynomials used to define
TTE as a E.
However,
we will leave to the reader the details of keeping track of this algorithm through the various steps of the proof. Proposition A.I can be reformulated in logical language as "elimination of quantifiers".
Specifically, by a polynomial relation
we mean a finite sentence built up from basic relations where
p
is a polynomial with coefficients in
tives "and", "or", "not". such that the sentence algebraic set in the type 3 G
or
R
p(x, ...x ,t.....t ) > 0,
R, using the logical connec-
Thus, the set of points
A(x.....x ,t,...t )
(x_...xn,t,...t ) £ R n + m
is true, is just a general semi-
- B y an elementary sentence, we mean a sentence of
(Q.^) • • • (Qnx n)A(x,.. .xn ,t, .. .t ) , where the
V, and
A
A(x . . .x jt^. . .t )
is a polynomial relation.
R m |(Q 1 x 1 )•••(Q n x n )A(x 1 ...x n ,t 1 ...t m )}.
Qi
Consider the set
are quantifiers, ^C t i--- t m )
Another form of the Tarski-Seidenberg
theorem is that this set is semi-algebraic.
Proposition A. 2.
268
There is a polynomial relation
B(tj...t )
(i-n
an algorithm for producing
B), such that
BC^.. .tm) •(Q 1x 1 ) • • • ( Q ^ ) •
A(x 1 ...x n) t 1 ...tJ. Since
(Vxn)A
is equivalent to
~(3x n )(~A), it is clear by induction
that Proposition A.2 needs to be proved only for is what we actually prove below. m
R |(3x 1 )A(x 1 ,t 1 -..t m )} set
n = 1
In this case, the set
is just the projection to
{(Xpt-L..-tm) E R m +
\k(xlitl.. -t m )}.
R
m
and
Q^ = 3.
This
{(t,...t ) £ of the semi-algebraic
We see therefore that Propositions
A.I and A.2 are equivalent. As an example of the sort of sets described by elementary sentences, we mention the topological closure of a semi-algebraic set E = { ( t r . . t m ) eR m |(Vx o )(3x 1 )-..(3x m )(x o = 0 or m o i
E C Rm,
^...xJGE
and
. y w < *o}In order to prove Proposition A.2, we need to introduce temporarily a new notion.
Consider functions
values in
Note that the set of polynomials in one variable of degree
<_ d
R.
is an affine space over
R
f
with domain
of dimension
(a . . . a ^ ) , the coefficients of a polynomial
a
subset
X C R
and
d+1, with coordinates
q = a x + ••• +a^.
Thus, it
makes sense to talk about semi-algebraic subsets of the space of polynomials of degree
<_ d.
Definition A.3. all
D
The function
f:
and all semi-algebraic subsets
Rm|q(f(x)) G E}
X -»- R
is psenri-algebraic if for
E C R, the set
+
{(q,x) G R
x
is semi-algebraic.
Note that the definition guarantees that the domain of a psemi-algebraic function is a semi-algebraic set. to whether
t > 0, t = 0, or
If
t < 0.
rather simple, it is easy to see that and
X = +1,0,-1
algebraic.
the sets
t £ R, let
sign(t) = +1,0,-1
according
Since semi-algebraic subsets of f
R
is psemi-algebraic if for all
{(q,x) e R d + 1 x Rm|sign q(f(x)) = X}
are d
are semi-
It is also clear that polynomials are psemi-algebraic.
We will
need to know that differences of psemi-algebraic functions are psemi-algebraic. However, it is not much harder to prove the following.
Lemma A.4.
If
Si»'##>Sic
are
psemi-algebraic functions of
m
variables
269
and
f
is a polynomial in
k
x
psemi-algebraic function of
Proof:
ar
+
Now
q(f ( y ^ . .yk)) = Q(y 1 ...y k ) =
Q^ = Qi( v]/--y k p
only on the coefficients of
f.
xR m |sign q(f(g (x),...
{(q,x) £ R
ficients are polynomials in the coefficients
it follows that the set
is a
x = ( i--- m )-
e semi-algebraic.
* *' + Q e , where the
f (g-^Cx), ... ,gk(x))
x
We must show that the sets
•••»g k 00)) = A} Q o yf
variables, then
are
polynomials whose coef-
(a^.a^)
Therefore, since
of
g k (x)
q, depending
is psemi-algebraic,
{(q,y1-•.yk_1,x)|sign Q ( y r • - y k _ r g k(x) ) = A}
is
semi-algebraic, hence can be expressed as a Boolean combination of sets of the form
sign F(a Q . . .a^y.^.. .y^-^x) = u, where
similar reasoning, the set
x
x
k i( )>Sk( ))
=
is
^
Continuing this type of argument, we
{(q,x)|sign QCg-^Cx),.. . ,gk(x)) = A}
is semi-
algebraic, as desired.
£j_(b) < £ 2 (b) < • • •, defined.
+
*•• + b n
be a polynomial with real roots
b = (b Q ...b n ).
Of course, the
£^
are not everywhere
The main result needed to prove Proposition A.2 is the following.
Proposition A.5.
Proof: n = 1.
D
p(x) = b Q x n +b^x 1 '
Let
The
^(b)
are psemi-algebraic functions of
We argue by induction on
Since the derivative
(nbQ,(n-l)b1,...,b
p'(x)
The data
has degree
_< n-1, with coefficients
^), we have by induction that the roots of
p'(x), say
sign(bQ),...,sign(bn),sign pCnjCb)),...,sign p(nk(b))
the number of roots of the intervals
of
p(x)
as well as locating each
and on
This is so because C~°°>Tli)> (iiv*00)*
an(
p(x)
p(x)
270
^(b)
in one
^ because the
at
sign(b-)
determine the
J
± °°. {(q,b) G R d + 1 x R n + 1 1 sign qC^Cb)) = \}
We must show that the sets semi-algebraic.
determines
is monotonic between any two roots
IK behavior of
b = (b Q ...b n ).
(- °°,ri-.(b)), (n- (b) ,n-+ 1 (b)) , (nk(b),°°), or among the points
n-^Cb) ,.. . ,n k(b) . p'(x)
£. (b)
b.
n, the Proposition being trivial if
n-j^b) < n 2 (b) < ••• < Tik(v<j (b), are psemi-algebraic functions of
of
By
{(q,y.....yk 2,x)|sign Q(y,.-.yk 2»
semialgebraic.
eventually conclude that
is a polynomial.
{(q,y1-..yk_2,x)|sign F(a Q ...a d ,y 1 ...y k _ 2 ,g k _ 1 (x),x)
is semi-algebraic, and it follows that g
F
We can divide
q
by
p
and write
are
q(x) = s(x)p(x) +r(x),
where
degree(r) < degree(p).
Then
q(^(b)) = r ^ C b ) ) , hence
{(q,b)|sign qCqCb)) = A} = {(q,b)|sign r(q(b)) = X}. ficients of
r
are polynomials in the coefficients
b =
(b0---bn)
of
r(x)
of
q
and
p.
b
psemi-algebraic functions of
= 0, b, ^ 0, and so on.
(a,b).
signCr^Cb) -Yj(a,b)), sign p(y^(a,b)). a psemi-algebraic function of
E $ = {(a,b) e R sign r(^i(b))
is constant on
(a,b).
x R n+1 |(a,b)
(Strictly speaking, there
contains $
$
E$.
sign(b.), sign p(r|.(b)),
By Lemma A.4, r ^ O ) - Yj(a,b)
is
Therefore, any one of the finitely
$}.
Therefore, the set
is determined by
which interval
^ ( b ) , or whether
We claim that
{(q,b)|sign r(^i(b))=
$, we observe first that
(- °° ,ru (b)), (n- (b) >r\., (b)), or 1 3 j+1
^ ( b ) = n-(b), some
Y-(a,b)
of
r(x)
j.
lie in this interval.
^(b),
or
whether
a b
^ ( b ) = Y-( > )>
(- 00 ,Y 1 ), (Yj>Yj+ 1)>
is constant on the intervals
K
some
3-
Thirdly,
p(Y-(a,b))
(-°° ,Y-i(a>b)), (Y- (a,b) >Y- + 1 C a > b )) > o r
determine the interval
(ru (b),°°)
Secondly, we know from"
is monotonic on this interval, hence the values of sign
which contains
bQ ^ 0
E,, hence is semi-algebraic.
sign rC^Cb))
which of the roots
p(x)
r(x), namely if
realizes sign data
is a finite union of some of the
we know from
y. < • • ' < Y m
$ , for this sign data determines a semi-algebraic set
d+1
To see that
and
This causes no problem.)
Consider the sign data consisting of
many possibilities,
a = (a o>..a,)
Thus, by induction, the roots
are cases to l?e distinguished in the formula for or if
Moreover, the coef-
But
(Ym>°°~) > arid
(Ym(a,b),°°) s
ign
our
r
(x)
claim is
established.
D
Proof of Proposition A.2: Seidenberg theorem. polynomials
It clearly suffices to show that given finitely many
^Cx-j^.-.x^
sign f.(x,...x ) = A.} in
Xj
signs \^3 the set
and
is semi-algebraic.
(x 2 ...x n ).
f^, say
^•(x2-•-xn)»
Consider the sign data
^ i -, say
£^ <_ ^
are
orderings, which partitions
R ~
f-
as polynomials
(x 2 ...x n ).
The real
psemi-algebraic functions of
sign(^i- - ^ . , . , ) , which orders the
£ * * * £ £t* n
{(x2 ...xn) G R n" | (aXj)
Regard the
with coefficients which are polynomials in
roots of the
roots
We can now complete the proof of the Tarski-
There are finitely many possible
into disjoint, semi-algebraic subsets.
271
If
(x 2...x n)
is fixed, the
C - 0 0 , ^ ) , ( £ . , £ . , ) , (£,«). sign data
sign ^ ( £ ^ - 1 ) , sign f j ^ — — 2 sign
is obvious that the truth of
Hence
have constant sign on the intervals
(In fact, these signs are computed by the
one could use the sign data
the data
f^
sign f ^ - " * ^ ) ,
3+
)* sign
f
i^t+1^'
(3 f ^ B x ^ (£.), f
x
x
k > 0. See 7.2.5.)
(3x^) sign i( ;j/-' n)
sign f^.),
sign f ^
{(x 2 -. .xR) | (3x1)sign f ^ x ^ . - x ^ = Xi>
Alternatively,
=
\
It
depends only on
,£ j+ 1 ) , sign
f^,-).
is semi-algebraic.
•
We can now use elimination of quantifiers to eliminate the notion of psemi-algebraic functions.
Proposition A.6.
A function
if and only if the graph of
Proof: function set.
f
f,
f:
X -> R,
X C R m , is psemi-algebraic
F C R m x R, i s a semi-algebraic set.
A very special case of the definition of a psemi-algebraic shows that the set
But this is Conversely, if
polynomial relation {(q>x)I(3t)(A(x,t) Proposition A.2.
272
This observation is due to Efroymson [36].
{(t,x)|f(x) -t = 0}
is a semi-algebraic
F. F
is semi-algebraic, let
A(x,t). and
Then
F = {(x,t)|A(x,t)}
for some
{(q,x)|sign q(f(x)) = X} =
sign q(t) = X]
which is semi-algebraic by •
Bibliography
I. Hubert's 17 t h Problem 1. D. Hilbert, Ober die DarsteXlung definiter Formen als Summe von Formenquadraten, Math. Ann. 32 (1888), 342-350. 2. E. Landau, tlber die Darstellungen definiter binarer Formen durch Quadrate, Math. Ann. 57 (1903), 53-64. 3. E. Landau, Uber die Darstellung definiter Funktionen durch Quadrate, Math. Ann. 62 (1906), 272-285. 4. E. Artin, Uber die Zerlegung definiter Funktionen in Quadrate, Abh. Math. Sem. Univ. Hamburg 5 (1927), 100-115. 5. J.W.S. Cassels, On the representation of rational functions as sums of squares, Acta Arith. IX (1964), 79-82. 6. A. Pfister, Zur Darstellung definiter Funktionen als Summe von Quadraten, Invent. Math. 4 (1967), 229-237. 7. D.W. Dubois, Note on Artin's solution of Hilbert's 17th problem, Bull. Amer. Math. Soc. No. 4, 73 (1967), 540-541. 8. T.S. Motzkin, The Arithmetic-Geometric Inequality, and, Algebraic Inequalities, Inequalities, Vol. 1 (O.Shisha, Ed.) Academic Press, New York, 1967, 205-254 and 199-203.
H. Ordered Fields and Real Varieties 9. E. Artin and 0. Schreier, Algebraische Konstruktion reeller Korper, Abh. Math. Sem. Univ. Hamburg 5 (1926), 85-99. 10. E. Artin and 0. Schreier, Eine Kennzeichnung der reell abgeschlossenen Korper, Abh. Math. Sem. Univ. Hamburg 5 (1927), 225-231. 11. R. Baer, Uber nicht-Archimedisch geordnete Korper, S.-B.
273
Heidelberger Akad. Wiss. Math.-Natur. Kl. 8. Abh. 1927 (Beitrage zur Algebra. L.). 12. W. Krull, Allgemeine Bewertungstheorie, J. Reine Angew. Math. 167 (1931), 160-196. 13. S. Lang, The theory of real places, Ann. of Matn. 57 (1953), 378-391. 14. D. W. Dubois, A Nullstellensatz for ordered fields, Ark. Mat. 8 (1969), 111-114. 15. H. Gross and P. Hafner, tiber die Eindeutigkeit des reellen Abschlusses eines angeordneten Korpers, Comment. Math. Helv. 44 (1969), 491-494. 16. J.J. Risler, Une caracterisation des ideaux des varietes algebriques reelles, C.R. Acad. Sci. Paris 271 (1970), 1171-1173. 17. D.W. Dubois and G. Efroymson, Algebraic theory of real varieties, I., Studies and Essays presented ot Y.W. Chen on his sixtieth birthday, Taiwan, 1970), 107-135. 18. P. Ribenboim, Le theoreme des zeros pour les corps ordonnes, Seminaire d'Algebre et Theorie des Nombres, Dubreil-Pisot, 24^ annee, 1970-71, exp. 17. 19. M. Knebusch, On the uniqueness of real closures and the existence of real places, Comment. Math. Helv. 4 7 (1972), 260-269. 20. G. Efroymson, Local reality on algebraic varieties, J. Algebra 29 (1974), 133-142. 21. D.W. Dubois and G. Efroymson, A dimension theorem for real primes, Canad. J. Math. 26 (1974), 108-114. 22. G. Stengle, A Nullstellensatz and a Positivstellensatz in semialgebraic geometry, Math. Ann. 207 (1974), 87-97. 23. M. Knebusch, On algebraic curves over real closed fields. I., Math. Z. 150 (1976), 49-70. 24. M. Knebusch, On algebraic curves over real closed fields. II., Math. Z. 189 (1976), 189-205.
m . Rings of Functions and Algebraic Topology 25. R. Swan, Vector bundles and projective modules, Trans. Amer. Math. Soc. 105 (1962), 264-277.
274
26. C. E. Watts, Alexander-Spanier cohomology and rings of continuous functions, Proc. Nat Acad. Sci. 54 (1965), 1027-1028. 27. E.G. Evans, Jr., Projective modules as fiber bundles, Proc. Amer. Math. Soc. 27 (1971), 623-626. 28. K. L0nsted, Vector bundles over finite CW complexes are algebraic, Proc. Amer. Math. Soc. 38 (1973), 27-31. 29. R. Swan, Topological examples of projective modules, Trans. Amer. Math. Soc. 230 (1977), 201-234.
IV. Real Analytic-Algebraic Functions 30. J. Nash, Real algebraic manifolds, Ann. of Math. 56 (1952), 405-421. 31. M. Artin and B. Mazur, On periodic points, Ann. of Math. 81 (1965), 82-99. 32. K. L0nsted, An algebraization of vector bundles on compact manifolds, J. Pure Appl. Algebra 2 (1972), 193-207. 33. J.J. Risler, Un theoreme des zeros en geometrie analytique reelle, C.R. Acad. Sci. Paris 274 (1972), 1488-1490. 34. R. Palais, Equivariant and real algebraic differential topology, Part I. Smoothness categories and Nash manifolds. Notes, Brandeis University 19 72. 35. J.J. Risler, Sur l'anneau des fonctions de Nash globales, C.R. Acad. Sci. Paris 276 (1973), 1513-1516. 36. G. Efroymson, A Nullstellensatz for Nash rings, Pacific J. Math. 54 (1974), 101-112. 37. J.J. Risler, Resultats recents sur les fonctions de Nash, Seminaire Pierre Lelong (Analyse) Annee 1974/75, Lecture Notes in Mathematics No. 524, Springer-Verlag, 19 76. 38. J.J. Risler, Sur l'anneau des fonctions de Nash globales, Ann. Sci. Ecole Norm. Sup. 8 (1975), 365-378. 39. J. Bochnak and J.J. Risler, Le thereme des zeros pour les varietes analytique reeles de dimension 2, Ann Sci. Ecole Norm. Sup. 8 (1975), 353-364. 40. T. Mostowski, Some properties of the ring of Nash functions, Ann.Scuola Norm. Sup. Pisa IK (1976), 245-266. 41. G. Efroymson, Substitution in Nash functions, Pacific J. Math. 63 (1976), 137-145.
275
42. J. Bochnak, Sur la factorialite des anneaux des fonctions de Nash, Comment. Math. Helv. 52 (1977), 211-218. V. Topology of Semi-Algebraic Sets 43. O.A. Oleinik, Estimates of the Betti numbers of real algebraic hypersurfaces, Rec. Math. (Mat. Sb.) N.S. 28(70), (1951), 635-640. 44. H. Whitney, Elementary structure of real algebraic varieties, Ann. of Math. 66 (1957), 545-556. 45. J. Milnor, On the Betti numbers of real varieties, Proc. Amer. Math. Soc. 15 (1964), 275-280. 46. S. Lojasiewicz, Triangulations of semi-analytic sets, Ann. Scuola Norm. Sup. Pisa 18 (1964), 449-474. 47. R. Thorn, Sur l'homologie des varietes algebrique reeles, Differential and Combinatorial Topology (Morse Symposium), Princeton University Press, 1965. 48. H. Hironaka, Triangulations of algebraic sets, Proc. Amer. Math. S o c , Symp. in Pure Math., 29 (1975), 165-185. 49. J. Bochnak, Quelques proprietes quantitatives des ensembles semi-algebrique, Ann. Scuola Norm. Sup. Pisa 2 (1975). 483-495
VI. Witt Rings and Reality 50. D.K. Harrison, Witt Rings, Lecture Notes, Dept. Math., Univ. of Kentucky, Lexington, KY, 1970. 51. J. Leicht and F. Lorenz, Die Primideale des Wittschen Ringes, Invent. Math. 10 (1970), 82-88. 52. M. Knebusch, Real closures of semi-local rings and extensions of real places, Bull. Amer. Math. Soc. 79 (1973), 78-81. 53. M. Knebusch, Real closures of commutative rings I, J. Reine Angew. Math. 274/275 (1975), 61-89. 54. M. Knebusch, Real closures of commutative rings II, J. Reine Angew. Math. 286/287 (1976), 278-213. 55. T.Y. Lam, Ten lectures on quadratic forms over fields, Conference on Quadratic Forms, 19 76, Queen's Papers in Pure and Appl. Math. No. 46, Queen's University, Kingston, Ont., Canada, 1977.
276
VII. Real Algebra and Logic 56. A. Tarski, A decision method for elementary algebra and geometry, 2 n d ed., revised, Berkeley and Los Angeles, 19 51. 57. A. Seidenberg, A new decision method for elementary algebra, Ann. of Math. 60 (1954), 365-374. 58. A. Robinson, On ordered fields and definite functions, Math. Ann. 130 (1955), 257-271. 59. A. Robinson, Further remarks on ordered fields and definite functions, Math. Ann. 130 (1956), 405-409. 60. A. Robinson, Introduction to Model Theory and the Meta Mathematics of Algebra, North-Holland Publishing Co., Amsterdam, 1965. 61. G. Kreisel and J.L. Krivine, Elements de logique mathematique, theorie des modeles, Dunod., Paris, 19 67. 62. P.J. Cohen, Decision procedures for real and p-adic fields, Comment. Pure Appl. Math. 22 (1969), 131-151.
VUT. Basic Algebra Texts 63. M.F. Atiyah and I.G. Macdonald, Introduction to Commutative Algebra, Addison-Wesley, 1969. 64. N. Jacobson, Lectures in Abstract Algebra, Vols. 1-3, Van Nostrand, 1951, 1953, 1964. 65. N. Jacobson, Basic Algebra I, W.H. Freeman, 19 74. 66. S. Lang, Algebra, Addison-Wesley, 1965. 67. B.L. van der Waerden, Modern Algebra, Vols. 1,2, Frederick Ungar Publishing Co., 19 53. 68. 0. Zariski and P. Samuel, Commutative Algebra, Vols. 1,2, Van Nostrand, 1958, 1960.
277
Notation
(Introduction not included)
Page 32 32 32 33 35 35 36
Page Spec(A,?)
106
m
S ec fA 13 pec t >+Z(X) D(a)
106
r(u, 0 )
116
T>(u, ^ x )
116
106 106
36 36 a,+ '
38
a
42, 81
u{ gi>
162
45
w{ gi >
162
H(X)
46 ^
^&J
162
AH(X) GA
53
Fg
54
-
(I:J)
54
163 163
60
X(A,P)
66
(A(S), P(
77
166 173 X(B,15)d 197
500
83
T
R(A,13)
84 85
vV^gjL^ {g }
i(n)
ftQ
89,Qc; 95
i
^
233
97 98
233
^P ( ^"")
99 (A,£ A )
®
(B/P B )
102 102
233
233 235
C°(U) C°(U) r
259
C (U)
264
C°°(U) C"(U)
264
103 T~
* * 104
278
Index
(Introduction not included)
absolute hull, 53, 54 affine coordinate ring, 173, 201, 219 affine order, 41 Archimedean closed subfield, 140 Archimedean extension, 138, 139 Artin, E., 42, 100, 122, 130, 131, 187 188, 196 associated primes, 59, 60, 61, 73, 95, 96 bounded filter, 235 center of a filter, 235 closed semi-algebraic set, 163, 172, 201 closed, semi-algebraic set, 164, 215, 219, 220, 258 closure of a set, 163 codimension, 212, 213 Cohen, P., 42, 165, 268 concave multiplicative set, 83, 93 connected semi-algebraic set, 249, 257, 262 continuous sections of a sheaf, 115, 116, 117 continuous semi-algebraic function, 252, 253, 255, 256, 257 contraction of an order, 35 convex set, 33 convex hull, 46, 47 Dedekind cuts, 137, 138 degenerate points, 197 derived order, 39, 42, 64, 81, 97, 98, 188, 199, 200 derived set, 36, 37, 38, 39 dimension, 212, 219 direct limit, 103, 105 Dubois, D., 187 Efroymson, G., 216, 227, 272 elementary sentence, 268 extension of an order, 35, 40, 120, 121, 122, 130 Fat City, 219 fibre product, 101 fibre sum, 102, 104, 105 filter, 43, 233 filter at infinity, 235 formally real field, 34, 149
Hilbert, D., 189 Hilbert 17th problem, 42, 188 ideal absolutely convex, 52, 53, 54, 55, 88, 89, 156 convex, 45, 46, 52, 81 maximal convex, 49, 50, 63, 87, 93 167, 185 minimal prime, 58, 60, 91, 92, 108 primary convex, 57, 64, 74, 75, 82, 89, 94 prime convex, 49, 51, 63, 66, 74, 90, 93, 127, 186 Jacobson radical, 85, 86, 99 Krull dimension, 224, 225 Krull, F., 139, 149, 150 Krull valuation, 146 Lang, S., 130, 184 localized order, 77, 79, 80, 81, 82, 111, 112, 113 maximal order, 33, 37, 38, 39, 44 morphism, 32 Nash functions, 265 nil radical, 37, 46, 47, 48, 51, 90, 167 Noetherian-Grothendieck topology, 248, 252 non-degenerate points, 197, 198, 199, 201, 202, 215, 218 open semi-algebraic set, 163, 164 open, semi-algebraic set, 164 order, 32 partially ordered ring, 32 partition of unity, 44 polynomial relation, 268 (POR), 32, 34 (PORCK), 33, 34, 55, 56, 88, 110, 114, 117, 119 (PORNN), 33, 34, 56 (PORPP), 33 prefilter, 233, 234 product order, 44
279
psemi-algebraic function, 269, 270 272 quotient order, 45, 52, 56, 58, 81, 201, 202 rank, 213 real closed field, 130, 131, 132, 134 135 refinement of an order, 33, 80, 81 96, 97, 150, 151, 157, 189 RHJ algebra, 167, 168, 169, 170, 171, 199, 200, 201, 202, 219 Risler, J. J., 187 Schreier, 0., 122, 130, 131 semi-algebraic set, 162, 268 semi-field, 49, 50, 85 semi-integral extension, 126, 127 154, 155, 170 semi-Noetherian ring, 57, 58, 95, 96 shadow of a set, 83, 84 shadow of 1, 84, 88 sheaf of partially ordered rings, 41, 112, 249 signed place, 145, 148, 161, 184, 186 190, 206, 228, 242 simple point, 213, 215 simple refinement, 36, 37 simplicial complex, 43, 69 split valuation, 147
280
Stengle, G., 100, 173, 189, 193, 195, 196 strong 'j^chain, 225, 226, 228, 231 strong topology, 247 structure sheaf rational, 248, 250 semi-algebraic, 248, 252, 259 smooth, 248, 262, 265 Zariski, 112, 115, 116, 249 Sturm algorithm, 135, 136 symbolic powers, 89, 95 Tarski-Seidenberg theorem, 163, 164, 165, 203, 206, 207, 215, 233, 258, 268, 271 total order, 37, 64, 141, 143, 145, 156, 178, 234, 238 ultrafilter, 233, 234 valuation, 146, 242 valuation ring, 139, 149, 152 weak order, 33, 168 weak $-chain, 224, 227 Whitney, H., 261 Zariski dense set, 43, 238 Zariski, 0., 244 Zariski topology, 106, 107, 108, 109, 110 zero divisors, 42, 60, 91, 92