Lecture Notes in Mathematics 1834: Editors: J.-M. Morel, Cachan F. Takens, Groningen B. Teissier, Paris
Lecture Notes in Mathematics 1834: Editors: J.-M. Morel, Cachan F. Takens, Groningen B. Teissier, Paris
Lecture Notes in Mathematics 1834: Editors: J.-M. Morel, Cachan F. Takens, Groningen B. Teissier, Paris
Editors:
J.–M. Morel, Cachan
F. Takens, Groningen
B. Teissier, Paris
3
Berlin
Heidelberg
New York
Hong Kong
London
Milan
Paris
Tokyo
Yosef Yomdin
Georges Comte
13
Authors
Yosef Yomdin
Department of Mathematics
Weizmann Institute of Science
Rehovot 76100, Israel
e-mail: [email protected]
Georges Comte
Laboratoire J. A. Dieudonné
UMR CNRS 6621
Université de Nice Sophia-Antipolis
28, avenue de Valrose
06108 Nice Cedex 2, France
e-mail: [email protected]
Mathematics Subject Classification (2000): 28A75, 14Q20, 14P10, 26B5, 26B15, 32S15
ISSN 0075-8434
ISBN 3-540-20612-4 Springer-Verlag Berlin Heidelberg New York
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specif ically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microf ilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.
Springer-Verlag is a part of Springer Science + Business Media
http://www.springeronline.com
This book presents results and methods developed during quite a long period
of time and many people helped in this work. We would like to thank Y.
Kannai for pointing out, in the very beginning of the work on quantitative
transversality, the relevance of metric entropy. Since 1983 M. Gromov encou-
raged this research and helped us in many fruitful discussions of qualitative
transversality and Semialgebraic Geometry in Dynamics and Analysis. We
would like to thank him especially for suggesting a problem of quantitative
Kupka-Smale, for his contribution to C k -reparametrization of semialgebraic
sets and applications to dynamics, for providing a central (for this book) refe-
rence to Multidimensional Variations and to books of Vitushkin and Ivanov
and for encouraging writing preliminary texts, which were used in this book.
This book would not have be written without the help and encouragement
of J. -J. Risler and B. Teissier and numerous fruitful discussions with them
during all the long period of the book’s preparation. It is a pleasure to thank
M. Giusti, J. -P. Henry and M. Merle for their invitation to give a course
on Metric Semialgebraic Geometry at École Polytechnique in 1985-86, and
again M. Merle for his remarks during lectures given at the University of Nice
- Sophia Antipolis in 1999-2000, and P. Milman for fruitful discussions and
for an invitation to the University of Toronto, where the preliminary text has
been written and typed. We would like to thank D. Trotman, who read and
corrected the text, and also indicated precious references. We would like to
thank M. Briskin, Y. Elichai, J.-P. Françoise, G. Loeper and N. Roytvarf for
their help and contribution. Many of the results and methods presented in
this book have been obtained in a long collaboration with them.
Last not least, many thanks belong to the University of Beer-Sheva, to the
Max Planck Institut für Mathematik, to IHES, to the University of Toronto
and to our home Universities - the University of Nice-Sophia Antipolis and
the Weizmann Institute of Science, where many of the results and methods
in this book have been developed, and to the Israel Science Foundation and
the Minerva Foundation, which have supported the research of one of the
authors for many years.
Table of Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V
2 Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Multidimensional Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
1 Introduction and Content
1.1 Motivations
This book deals with several related topics:
– Geometry of real semialgebraic and tame sets (i.e. sets definable in some o-
minimal structure on (R, +, .)), with the stress on “metric” characteristics:
lengths, volumes in different dimensions, curvatures etc...
– Behaviour of these characteristics under polynomial mappings.
– Integral geometry, especially the so-called Vitushkin variations, with the
stress on applications to semialgebraic and tame sets.
– Geometry of critical and near critical values of differentiable mappings.
Some fractal geometry naturally arising in this context.
Below we give a short description of each of these topics, their mutual
dependance and logical order. Motivation for the type of question asked in
this book, comes from several different sources: the main ones are Differen-
tial Topology, Singularity Theory, Smooth Dynamics, Control, Robotics and
Numerical Analysis.
One of the main analytic results, underlying most basic constructions of
Differential Analysis, Differential Topology, Differential Geometry, Differen-
tial Dynamics, Singularity Theory, as well as nonlinear Numerical Analysis,
is the so-called Sard (or Morse-Sard, or Morse-Morse-Sard, see [Morse 1,2],
[Mors], [Sar 1-3]) theorem. It asserts that the set of critical values of a suffi-
ciently smooth mapping has measure zero. Mostly this theorem appears as an
assumption, that the set Y (c) of solutions of an equation f (x) = c is a smooth
submanifold of the domain of x, for almost any value c of the right hand side
(in the semialgebraic or tame case, an asymptotic version of the Morse-Sard
theorem (see [Rab], [Kur-Orr-Sim]) shows that f induces a fibration over the
connected components of the “good values” c).
Another typical appearence of the Morse-Sard theorem is in the form of
various transversality statements: by a small perturbation of the data we can
always achieve a situation where all the submanifolds of interest intersect
one another in a transversal way. Fig. 1.1 shows a non-transversal (a) and
transversal (b, c) intersections of two plane curves.
Fig. 1.1.
Chapter 7 contains the main results of this book, as far as the tame
(semialgebraic) sets and mappings are concerned. Basically they have the
following form: assuming that the size of the differential Df of f is bounded
(in one sense or another) on a set A, we estimate variations (and hence metric
entropy) of the image f (A) (Theorems 7.1 and 7.2). We deduce from the result
the quantitative Morse-Sard theorem in the polynomial case (Theorem 7.5).
In particular, we obtain, as a special case, Theorem 1.6 below.
Chapter 8 continues the line of Chapter 7, with somewhat more special re-
sults, related to “quantitative transversality” on one side, and to the behavior
of mappings on more complicated singularities.
has its endpoint e1 fixed at the origin, and the endpoint of b2 is fixed at the
second endpoint e2 of b1 . Both b1 and b2 can rotate freely at e1 and e2 .
O1 , O2 and O3 denote the obstacles, and the initial placement Z1 and
the desired target placement Z2 are shown on the picture. Taking as free
parameters the angles ϕ1 and ϕ2 shown in Fig. 1.2, we get the space P of
parameters as the square [0, 2π] × [0, 2π] in R2 (or rather a torus T 2 – this
more accurate topological representation sometimes helps).
Fig. 1.2.
Fig. 1.3.
1.3 The Motion Planing Problem in Robotics as an Example 9
Fig. 1.4.
Now one sees immediately that the solution exists, since Z1 and Z2 be-
long to the same connected component of F P . Three of the “control (or
configuration) trajectories” joining Z1 and Z2 are shown in Fig. 1.4, and the
corresponding evolution of the manipulator is given in Fig. 1.5.
This figure shows one of the three solutions, represented on Fig. 1.4,
namely ρ1 . It consists of 4 rotations (3 of them are consecutive, illustrated
by arcs 1, 1 , 2 and 3).
Thus the main difficulty in solving the motion planning problem consists
in the construction of the free configuration space. This construction is non-
trivial already in the first example considered. In the second example the free
configuration space F P is fairly complicated: it looks like a spiralled worm-
10 1 Introduction and Content
Fig. 1.5.
hole in the three-dimensional cube, and we do not show it here. However, the
solution turns out to exist, and is shown in Fig. 1.6.
Fig. 1.6.
Now the basic fact is that if each part of the system B and each obsta-
cle O are semialgebraic (i.e., representable by a finite number of polynomial
equations, inequalities and set-theoretic operations), then the free configura-
tion space F P is semialgebraic, and can be computed effectively from B and
O.
There exists also an effective procedure to decide whether two given points
belong to the same connected component of a given semialgebraic set. Con-
sequently, for semialgebraic data (which is a very natural assumption) the
motion planning problem can be effectively solved. See [Sch-Sha] for details.
Important remark. “Effectively” does not mean “efficiently”! The com-
plexity of the algorithms, based on the direct approach as above and using
1.3 The Motion Planing Problem in Robotics as an Example 11
However, one can develop practical algorithms, based on a high order ap-
proximation and on hierarchy of singularities, before the theoretical founda-
tions have been completed. Simple empirical procedures in most cases provide
a reasonable answer to the problems above. As far as the motion planning
is concerned, such an algorithm has been developed and initially tested (see
[Eli-Yom 1-3]). It is based on an approximation of the free space F P on a
certain grid, while at each gridpoint a semialgebraic representation as above
is used. However, the hierarchy of the allowed degenerations at each grid-
point is restricted in such a way that the overall complexity of computations
remains strictly bounded. More degenerate situations are treated (within the
prescribed accuracy) simply by an appropriate subdivision of the grid. The
efficiency of this algorithm confirms (in very limited cases, as of today) our
theoretical expectations.
Of course, the discussion above is applicable not only to the Motion Plan-
ning problem. A combination of a high order approximation of the data, its
further structuring and organization along the hierarchy of singularities in the
problem, and its analytic processing, present a powerful computational ap-
proach in many important problems. The “Quantitative Singularity Theory”
will form a theoretical basis of this approach. We discuss it in more detail
(including, in particular, some specific implementations) in Section 10.1.4 of
Chapter 10 below.
This is the place to say that we do claim that various theorems in semi-
algebraic geometry given below are (or may be) useful in motion planning
and other applications, but only when combined with a clever data approx-
imation, with an analysis of the hierarchy of singularities of the problem,
and with an appropriate scheme of numerical computations. It is not the
purpose of this book to develop these methods (see however [Eli-Yom 1-5],
[Bri-Yom 6], [Bri-Eli-Yom], [Bic-Yom], [Wie-Yom], [Yom 24], [Y-E-B-S] and
Section 10.1.4 of Chapter 10 below). Consequently, all the examples of “ap-
plications” given in this book are pure illustrations of mathematical results,
and any attempt of their straightforward application in computations is in
our opinion completely hopeless.
After this warning we return to the description of our approach to the
example of a motion planning problem.
There are two general and well-known principles in real semialgebraic
geometry (although their specific implementation can be rather nontrivial or
impossible).
The first says that any reasonable operation with semialgebraic data leads
to a semialgebraic “output”, with a combinatorial complexity (i.e. the degrees
of the polynomials and the set theoretic formula in a representation – below
we call these data the diagram of the set) depending only on that of the input.
The second principle claims that any reasonable metric characteristic of a
semialgebraic set of a given combinatorial complexity inside a ball of a given
1.3 The Motion Planing Problem in Robotics as an Example 13
radius, can be bounded in terms of the complexity and the radius of the ball.
A good part of this book is devoted to various specific manifestations of these
general principles.
The simplest, but rather useful, example where both these principles work,
is the following result (see Theorem 4.12, Chapter 4 below; see also [Den-Kur],
[Har], [Kur], [Tei 1], [Yom 1,5], and [D’Ac-Kur] for an explicit value of the
bound K(D) of Theorem 1.1).
Theorem 1.1. Let A ⊆ Rn be a semialgebraic set with a given diagram D.
Then for the ball BR of radius R, centered at the origin of Rn , any two points
z1 and z2 , belonging to the same connected component of A ∩ BR , can be
joined inside A ∩ BR by a semialgebraic curve , such that the diagram of
depends only on D, and the length of does not exceed K(D) · R, with the
constant K(D) depending only on D.
Corollary 1.2. If a solution to a motion planning problem with semialge-
braic data exists, it can be given by a semialgebraic path in the parameter
space, whose complexity and length depend only on the combinatorial com-
plexity of the data.
In this book we mostly study not just semialgebraic sets, but rather their
behavior under polynomial (or, more generally, semialgebraic – i.e. those with
a semialgebraic graph) mappings. In the context of a motion planning prob-
lem, an important and highly nontrivial such mapping appears very naturally.
This is the so-called “kinematic mapping” ϕ of the manipulator; it associates
to any given values of the control parameters the position of the “tooling de-
vice” (or of a prescribed point, or of any prescribed part of the manipulator).
The kinematic mapping of a manipulator (as described above) is always
semialgebraic, assuming that the controls are properly parametrized. Now,
the main practical problem that appears in the programming of industrial
robots, is the so-called “inverse kinematic problem”:
For a given trajectory s of the tooling device in the workspace, find a
corresponding trajectory σ in the space of controls (i.e. such that s = ϕ(σ),
where ϕ, as above, is the kinematic mapping). The initial motion-planning
is a part of the inverse kinematic problem, since the manipulator in the pro-
cess of motion is naturally assumed to avoid collisions. The inverse kinematic
problem is usually redundant, since the number of the degrees of freedom
of the manipulator is normally chosen to be bigger than the dimension of the
configuration space of the tooling device (to provide a flexibility in program-
ming).
There are various approaches to the inverse kinematic problem, mostly
dealing with one or another way to eliminate the above-mentionned redun-
dancy. One of these approaches is given in [Sha-Yom], together with some
literature on the subject. Usually the redundancy is eliminated by introduc-
ing a certain distribution in the parameter space, transversal to the fibers
of the kinematic mapping (and of a complementary dimension). Any motion
14 1 Introduction and Content
of the tooling device can be now lifted to the parameter space, in a locally
unique way.
Without additional restrictions this lifting is not semialgebraic. Moreover,
it is not easy to combine this lifting with the requirement of the collision
avoidance. We consider a combination of the redundancy elimination with the
semialgebraic motion planning (as shortly presented below) a very important
problem.
The following results provide a semialgebraic solution of bounded com-
plexity to the inverse kinematic problem:
Theorem 1.3. (see Theorem 4.10 below) Let f : A → B be a semial-
gebraic mapping between two semialgebraic sets. Then for any semialgebraic
curve s in f (A) ⊆ B there exists a semialgebraic curve σ in A, such that
f (σ) = s. The diagram of σ depends only on the diagrams of f , A, B and s.
Corollary 1.4. For any semialgebraic trajectory s of the tooling device in
its workspace, there exists a semialgebraic control trajectory σ, such that
s = ϕ(σ). If a solution without collisions exists, σ can be chosen to be non-
colliding. The combinatorial complexity and the length of σ are bounded in
terms of the complexity of the data.
In fact, Theorem 1.3 is a special case of the following general “covering the-
orem” (see Theorem 4.10, Chapter 4 below):
Theorem 1.5. For f : A → B as above, and for S a semialgebraic set in
f (A) ⊆ B, there exists a semialgebraic set Σ ⊆ A, with dimΣ = dimS, such
that f (Σ) = S. The diagram of Σ (and hence the bounds on its geometry)
depends only on the diagrams of f , A, B and S.
Obviously, this result has a natural interpretation in terms of control of a
manipulator, whose tooling device has to cover a surface S in the workspace.
The next topic, which is central for this book, is the geometry of critical
and near-critical values of semialgebraic (and later smooth) mappings. The
near-critical points of f are those where the differential Df is “almost singu-
lar” (in an appropriate sense – see below). The near critical values are values
of f at the near-critical points.
As applied to the kinematic mappings ϕ of a manipulator, these notions
become quite relevant: near-critical points of ϕ are those, where some of the
controls do not affect the position of the tooling device. Near critical values
are the positions which we can get with such “bad” control. There are obvious
reasons (especially as the dynamics of the motion is incorporated) to avoid
such positions.
First of all, let us see what the critical set and the critical image look like
in the first example above (with the “tooling device” just the endpoint of the
manipulator). One can easily see that critical positions of the manipulator
correspond exactly to ϕ2 = 0 and ϕ2 = π (see Fig. 1.7).
In both these configurations the controls ϕ1 and ϕ2 do not affect the
distance of the endpoint from the origin. For ϕ2 near 0 or near π an easy
1.3 The Motion Planing Problem in Robotics as an Example 15
Fig. 1.7.
computation gives for the distance r of the endpoint from the origin (assuming
|b1 | > |b2 |):
r ∼ |b1 | + |b2 | − cϕ22 ,
or
r ∼ |b1 | − |b2 | − c (π − ϕ2 )2 .
Hence the set of near-critical points of the kinematic mappings, where
∂r ∂r γ γ
| | γ ( ≡ 0), consists of two strips, |ϕ2 | and |π − ϕ2 | .
∂ϕ2 ∂ϕ1 2c 2c
The corresponding set of critical values consists of two rings,
γ2
|b1 | + |b2 | − r |b1 | + |b2 |
4c
and
γ2
|b1 | − |b2 | r |b1 | − |b2 | +
4c
(see Fig. 1.7). Notice that the singularities of the kinematic mappings φ in
this example are of the “fold” type, according to the Whitney classification
(see [Whi 3], [Boa], [Gol-Gui]). After an appropriate coordinate change it can
be locally written in the form
y = x
1 1
y2 = x22 .
Fig. 1.8.
The “tooling device” once more is the endpoint e of b, and the kinematic
mapping associates to the control parameters (position of the frame F on
the ellipse E and the position of the bar b inside F ) the endpoint e on
the plane. Direct computations here are somewhat more involved. However,
this example is well studied in Singularity Theory (see [Gol-Gui]). The set
of critical values here is the curve Γ , shown in Fig. 1.8. It consists of the
endpoint positions, as the distance of e from F is equal to the curvature
radius of the ellipse E at F .
All the points on this curve Γ are folds, except the four vertices, at which
the kinematic mapping has a “cusp” singularity, according to Whitney’s clas-
sification ([Whi 3]). In a properly chosen system of local coordinates it can
be written as y = x
1 1
y2 = x32 − x1 x2 .
As the γ-near-critical values are concerned, so here they form a strip around
Γ of width of order γ 2 near the fold-points. At cusps the situation is more
complicated.
In all these examples we see that the γ-critical values of φ form a “small”
set: as γ tends to zero, the area of ∆(φ) tends to zero. This is a general fact,
and in this book we study the “size” of near-critical values in detail. However,
the “area” is not convenient to measure this size. Instead, we bound the metric
entropy of near critical values. For a compact X, the -entropy M (, X) is
the minimal number of -balls that cover X. Let us give a couple of reasons
why metric entropy is better for our purposes. Other reasons are given in
Section 1.3 and scattered all over the book.
To keep this introduction to a reasonable size, we do not give here formal
definitions of the metric entropy, near-critical values etc., but rather short
explanations of these notions. Accurate definitions of all the notions related
1.3 The Motion Planing Problem in Robotics as an Example 17
to metric entropy are given in Chapter 2. For near critical points and values
this is done in Chapters 6 and 7.
First of all, in the motion planning context we would like to avoid not only
the set of critical values, but a certain neighborhood. The fact that “area” of
a set is small does not imply restrictions on its δ-neighborhood (it can consist
of a dense collection of curves of very small length, etc...). Ultimately, the set
of rational points in Rn has Lebesgue measure 0, while its δ-neighborhood
for any δ > 0 is all the space Rn .
On the contrary, metric entropy is stable with respect to taking neighbor-
hoods: if certain balls of radius cover X, the balls of radius + δ, centred
at the same points, cover the δ-neighborhood Xδ .
Another reason is that in many cases we would like to restrict ourselves
to a certain grid and to find “good points” in this grid. Once more, a small
Lebesgue measure of a “bad” set does not guarantee that we can find good
points in any prescribed grid (take once more all the rational points).
It is easy to show (see Chapter 2 below) that if metric entropy of a bad
set is small, then in any sufficiently dense grid most of the points are good.
Thus metric entropy is a stronger and more convenient geometric invariant
for our purposes.
The third reason to work with it is that it is also much more natural
for the problems treated in this book. For semialgebraic sets and mappings
the behavior of the metric entropy reflects in a very transparent way their
geometry in different dimensions. For smooth mappings, robustness of the
entropy allows for a direct application of semialgebraic results via polynomial
approximation. Moreover, for mappings of finite smoothness, metric entropy
(and a related notion of block or entropy-dimension) turns out to be the
correct geometric invariant: sets of critical values of C k mappings can be
characterized in such terms.
Let us make a more accurate definition of γ-critical sets and values: x is a
γ-critical point of f if the differential Df(x) maps the unit ball into an ellipsoid
with the smallest semiaxis γ. The set of γ-critical points of f is denoted
by Σ(γ, f ), and the set of γ-critical values of f is ∆(γ, f ) = f (Σ(γ, f )).
Now we are ready to state one of our main results:
Theorem 1.6. (Theorem 7.5, Chapter 7) Let f : A → Rm be a
semialgebraic mapping of two semialgebraic sets, with diam(A) = R and
rank(Df|A ) q and the norm of Df (x) bounded by 1 for any x in A. Then
for each γ ≥ 0, ≥ 0. the -entropy of the set of γ-critical values ∆(γ, f )
satisfies
R R
M (, ∆(γ, f )) C0 + C1 ( )q−1 + C2 γ( )q ,
where the constants C0 , C1 , C2 depend only on the diagrams of f and A. In
particular, the q-dimensional Lebesgue measure of ∆(γ, f ) does not exceed
C3 Rq γ.
18 1 Introduction and Content
for x in Σ does not exceed γ, and hence the last integral is bounded by
γ.length(s) K(n, d) · γ · r. Therefore, each ∆i can be covered by at most
K(n, d) intervals of length γr, and ∆ = ∪∆i can be covered by N1 (n, d) ·
K(n, d) = N (n, d) such intervals.
In the proof proposed above, a possible bound for N1 (n, d) can be pro-
duced, via Bézout’s Theorem. Following Theorem 4.9, Chapter 4, we have:
N1 (n, d) d(2d − 1)n−1 (because the polynomial ||grad(f )||2 − γ 2 is of degree
2(d − 1)). The difficult point concerns in fact the obtaining of a bound for
K(n, d).
We can modify a little bit the proof of Theorem 1.8 in the following way.
We consider s a semialgebraic set of Σ of dimension 1, not necessarily
connected, (s being 0-dimensional in the case ∆ is itself 0-dimensional), such
that f (s) = ∆. Such a set s is given by Theorem 4.10, and in this construction
the diagram of s depends only on n and d. Let us consider now a connected
component ∆i of ∆ which is not a point. We can find a semialgebraic set
si ⊂ s such that f (si ) = ∆i and such that the sj are disjoint. The same
arguments as in the proof of Theorem 1.8 show that the total length of ∆i
i
is less than γ · length(si ) γ · length(s). But we have by Lemma 4.13:
i
length(s) K (n, d) · r. Considering that the number of connected compo-
nents of ∆ which are points is less than the number of connected components
of Σ, which, in turn, is less than d(2d − 1)n−1 , we obtain that we can cover
∆ by d(2d − 1)n−1 + K (n, d) intervals of length γ.r.
This bound is better than the one given in the proof of Theorem 1.8,
because the product has been replaced by a sum. Nevertheless, the point is
again to evaluate K (n, d).
Remark. In [D’Ac-Kur], it is shown that a possible bound for K(n, d) is:
k
maxx∈Brn dg(x) − dP (x) Rk (g) ,
r
1
where Rk (g) = maxdk g · rk is the remainder term in the Taylor formula.
k!
Hence, the critical points of g are at most γ0 -critical for P , where
k
γ0 = Rk (g), i.e. Σ(g, 0) ⊆ Σ(P, γ0 , r). Hence ∆(g, 0) = g(Σ(g, 0, r)) ⊆
r
g(Σ(P, γ0 , r)). Finally, since |g − P | Rk (g), g(Σ(P, γ0 , r)) is contained in a
Rk (g)-neighborhood of P (Σ, γ0 , r)) = ∆(P, γ0 , r).
Now by Theorem 1.8, ∆(P, γ0 , r) can be covered by at most N (n, k−1) in-
tervals of length γ0 r = k·Rk (g), and hence by k·N (n, k−1) intervals of length
Rk (g). Thus the Rk (g)-neighborhood of ∆(P, γ0 , r), and hence ∆(g, 0, r) can
be covered by the same number of intervals of length 3Rk (g), or by triple the
number of Rk (g)-intervals. We proved the following result:
Theorem 1.9. Let g : Brn → R be a k times differentiable function. Then
the set ∆(g, 0, r) of the critical values of g on the ball Brn can be covered by
at most N2 (n, k) intervals of length Rk (g), where
1 n
m(∆) lim · C · ( )n/k = lim C1−n/k = 0 for < 1 .
→0 →0 k
Concluding this section, let us discuss again the problem of finding the
explicit (and optimal) constants in the inequalities above, and throughout the
book. In principle, an application of the methods of Chapter 4 below allows
one to get an explicit estimate for each of the “algebraic” constants. Let us
illustrate our principle with the constant N (n, d) of Theorem 1.8, although
the following lines do not give a general proof, but rather a description of a
general philosophy (see [D’Ac-Kur] for a rigourous proof of the obtaining of
this bound):
we approximate the boundary of Σ(f, γ, r) by a smooth semialgebraic set
Z = {p = ||grad(f )||2 − γ 2 = δ}. Then we fix a generic linear form on Rn
and define the curve S of all the critical points of on Z ∩ {f = t}. Assuming
that Z ∩ {f = t} is smooth and compact, we obtain explicit equations for S.
Performing a linear change of variables in Rn , we can assume that (x) = xn .
Hence S consists of the points x in Z where the vector (0, · · · , 0, 1) is a linear
combination of grad(p)(x) and grad(f )(x) . This condition is given by the
equations:
∂f ∂p ∂p ∂f
− = 0, i = 2, · · · , n − 1,
∂x1 ∂xi ∂x1 ∂xi
each of degree (2d − 3)(d − 1), the equation defining Z being of degree 2d − 2.
We find, using Corollary 4.9, that the number of connected components
k(n, d) of S in generic hyperplane section is less than:
However, in most of the results in this book we do not give such explicit es-
timates. The reason on one side is that their producing is rather lengthy. On
the other side, in all the applications in Smooth Analysis, considered in this
book, the explicit estimates of the constants are not crucial. So just present-
ing explicit estimates of numerous constants below would not be especially
instructive.
On the other hand, in applications in Numerical Analysis, discussed in
Section 10.1.4, Chapter 10, accurate estimates of the algebraic constants are
crucial. Producing such accurate estimates is not an easy task. In some cases
even the asymptotics is not known. We consider this as an important open
problem. On these questions, we can at least refer to [And-Brö-Rui], [Hei-
Rec-Roy], [Hei-Roy-Sol 1,2,3,4], [Ren 1, 2, 3] (and of course to references
22 1 Introduction and Content
Fig. 2.1.
(or digitally memorize A with accuracy ). Thus, the behavior of M (, A) for
various reflects not only “massiveness” of the set A, but also its geometry
in X. A detailed study of metric entropy can be found in [Kol-Tih]. See also
[Lor 2] and [Tri].
A subset Z of X is called an -net for A, if for any y ∈ A there is z ∈ Z
with d(z, y) . Consequently, M (, A) coincides with the minimal number of
elements in -nets for A (these elements being the centres of covering balls).
On the other hand, call the set W ⊂ A -separated, if for any distinct
w1 , w2 ∈ W , d(w1 , w2 ) > . Denoting by M (, A) the maximal number of
elements in -separated subsets of A, we have easily: M (2, A) M (, A)
M (, A). Indeed, if x1 , . . . xM (2,A) are in a 2-separated subset of A, every
-ball containing xj does not contain xk , for k
= j and thus one needs at least
M (2, A) -balls to cover A, and if x1 , . . . , xM (,A) are -separated points in
A, by definition of M (, A), there exists no y ∈ A such that d(y, xj ) >
for all j ∈ {1, . . . , M (, A)}, thus the balls B(xj , ) cover A, showing that
M (, A) is less than M (, A).
M (2, A2 ) M (, A1 ).
2 Entropy 25
The behavior of metric entropy under mappings with known metric prop-
erties can be easily described:
(8) Let f : X → Y be an Hölderian mapping, i.e. such that there exist two
reals K > 0 and α such that, for all x, y ∈ X,
α
dY (f (x); f (y)) K dX (x; y) .
Notation. For a given > 0, define η() as follows: 2η() is the supremum,
over all pairs of balls B1 , B2 of radius in X, of the diameter of Σ(B1 , B2 ).
In general, η() measures the “nondegeneracy” of the action of the param-
eter t on X. Our main example is the following: X = Y = Rn , ft (x) = t + x.
Then clearly η() = 2, since the set of t for which t + B1 ∩ B2
= ∅ is a ball
of radius 2 in Rn .
Proposition 2.2. Let A1 , A2 ⊂ X. Then for any > 0 and ξ = η(2), we
have:
26 2 Entropy
Proof. We cover the -neighborhoods A1, and A2, by M (, A1 ) and M (, A2 )
2-balls Bi and Bj , respectively. Then the set of t ∈ Y for which ft (A1, ) in-
tersects A2, is contained in the union Σ(Bi , Bj ). But each of these sets is
i,j
contained in some ball of radius ξ = η(2) in Y, by definition of η(2). Thus
one needs less than M (, A1 ).M (, A2 ) ξ-balls to cover Σ(A1, , A2, ).
Using (7) above we can give a more “effective” version of this corollary:
Corollary 2.4. Let A1 , A2 , K1 , K2 , α, β be as above. Then for any > 0,
in any 2-separated set in Rn containing more than K1 .K2 ( 2 )α+β elements,
there is a t such that (t + A1, ) ∩ A2, = ∅.
Proof. By (7), any 2-separated set of Σ = Σ(A1, , A2, ) does not contain
more than M (, Σ) points. The proof of corollary 2.3 shows that we have
2
M (, Σ) K1 .K2 ( )α+β .
Hence if a 2-separated set of Σ contains more than K1 .K2 ( 2 )α+β elements,
we have a contradiction.
Notice also that in principle the statement of Corollary 2.4 allows defi-
nite verification by computations with bounded accuracy and time. Indeed,
we must check only a finite number of points t in the net, and for each ver-
ify the fact of nonintersecting, say, the 2 -neighborhood of A1 and A2 . But
it is enough to make computations with accuracy 3 to establish this fact
definitively.
(10) For subsets in Rn , we can compare M (, .) with the Hausdorff measure.
For more details about Hausdorff measures, see [Fed 2] or [Fal]. Remark
that the usual β-dimensional Hausdorff measure Hβ is defined in the same
way, but with the Bi ’s being arbitrary sets with diameter less than , in the
above definition. When A is an (Hm , m)-rectifiable subset of Rn , we have
Hm = S m ([Fed 2], 3.2.26).
Notice that S 0 = M (, A), but S 0 is equal to the number of points in A.
For any set A ∈ Rn , and for a suitable constant c(n), c(n)S n (A) is equal to
the Lebesgue measure of A.
Proposition 2.6. For any bounded A ⊂ Rn , and β ≥ 0, we have:
S β (A) lim inf β M (, A).
→0
→0 →0
i=0 i=0
28 2 Entropy
∞
∞
∞
β−γ riγ riβ β−α riα ,
i=0 i=0 i=0
log M (, A)
dime (A) = lim sup .
→0 log( 1 )
1
Thus dime (A) is the infinum of β for which M (, A) ( )β , for sufficiently
small .
2 Entropy 29
This inequality follows from Proposition 2.6: let β be a real such that
1
M (, A) ( )β (for small ). Proposition 2.6 allows us to write:
Thus dimH (A) β, and finally dimH (A) dime (A).
As an exercise, let us compute these two dimensions for the classical
Cantor set C 13 ⊂ [0; 1]. This set is obtained by removing from [0; 1] the in-
terval [ 13 ; 23 ]: one obtains two intervals C 1 and C 2 . We then proceed in the
same way with these two intervals, and we construct a sequence C i1 ,i2 ,...,in ,
ik ∈ {1, 2}, k ∈ {1, . . . , n}, of intervals such that diam(C i1 ,i2 ,...,in ) = ( 13 )n ,
C i1 ,i2 ,...,in ,in+1 ⊂ C i1 ,i2 ,...,in , dH (C i1 ,i2 ,...,in ,1 ; C i1 ,i2 ,...,in ,2 ) = ( 13 )n+1 . We de-
fine C 13 as the set consisting of all the points C i1 ,i2 ,...,in (see Fig. 2.2).
n∈N
Fig. 2.2.
log(2)
dimH (C 13 ) =
log(3)
log(2)
( and similarly dimH (C k1 ) = )
log(k)
(for a rigorous proof see [Fal]).
1 1
Let us now compute dime (C 13 ). For n+1
< , we have M (,
2.3 2.3n
C 13 ) = 2n+1 , hence:
For A = {1, 12 , 13 , . . .}, dime (A) = 1/2 (see below), while the Hausdorff
dimension of this countable set is of course 0.
The entropy dimension of sets in R may be defined in many different ways
(see [Tri]).
Theorem 2.9. (A. S. Besicovitch, S. J. Taylor) Let A ⊂ [a; b] be a
closed subset.
(i) If m(A) > 0, then dime (A) = 1
∞
(ii) Let m(A) be zero, i.e A = [a; b] \ Vi , with Vi open disjoint intervals
i=1
∞
and αi = b − a, where αi is the length of Vi . Then:
i=1 ∞
dime (A) = inf{β; αiβ < ∞}.
i=1
1 1 1
For example, for A = {1, 2a , 3a , . . .}, αi ∼ ia+1 ,
and hence dime (A) =
1
.
a+1
This theorem allows us to compute again dime (C 13 ): for 2n−1 i < 2n ,
∞ ∞
1 1 2 i−1
αi = ( )n , thus αiβ = ( ) , and this sum is convergent if and
3 i=1
3 i=1 3β
log(2)
only if β > dime (A) = .
log(3)
The following construction ([Yom 13]) generalizes Theorem 2.9 to higher
dimensions:
2 Entropy 31
proved in [Com 1]) and thus dimH (∆(f )) nk . Our Theorem 1.10 above im-
plies immediately a much stronger result: dime (∆(f )) nk (again this is the
sharpest bound by [Com 1]). In particular the set {1, 21a , 31a , . . . , 0} cannot be
the set of critical values of f , if k > n(a + 1), while the Morse-Sard theorem
gives no restrictions for countable sets to be sets of critical values.
However, the necessary and sufficient conditions for a given set to be the
set of critical values, are given just in terms of Vβ :
Theorem 2.12. ([Yom 13], [Bat-Nor]) The compact set A ⊂ R is con-
tained in ∆(f ) for some C k -smooth function f : B n → R if and only if
V nk (A) < ∞.
The rest of this book presents many examples of computation (or esti-
mation) of M (, A). Usually we are interested not only in the asymptotic
behavior of M (, A) as → 0, but in estimating M (, A) for any . The
following example illustrates the problems which can arise here.
Let A be a compact surface in R3 . For → 0, M (, A) ∼ c.H2 (A).( 1 )2 ,
with c some absolute constant. However, this expression does not bound
M (, A) for relatively big: indeed, taking our surface to be very “thin” and
1
“long”, we can get H2 (A) → 0, but M (, A) ∼ length(A) .
On the other side, taking the surface with a fixed number of connected
components and with the area and the “length” tending to zero, we see that
the number of connected components of A should also enter the upper bound
expression (see Fig. 2.3).
Indeed the correct bound has the following form (see [Iva 1], [Leo-Mel]
and theorems 3.5 and 3.6 below):
32 2 Entropy
Fig. 2.3.
1 1
M (, A) Ṽ0 (A) + C1 Ṽ1 (A) + C2 Ṽ2 (A) 2 ,
where Ṽ0 (A) is the number of connected components of A, Ṽ2 (A) = H2 (A),
and Ṽ1 (A) = (k1 + k2 ) dH2 , where k1 and k2 are the absolute values of
A
the mean curvatures of A.
3 Multidimensional Variations
Our main goal is to bound M (, A) for a given A. As the last example
of Chapter 2 suggests, the likely form of the required upper bound is (for
A ⊂ Rn ):
n
1
M (, A) C(n) Vi (A)( )i ,
i=0
Let Gkn denote the space of all the k-dimensional linear subspaces in Rn .
We have on the orthogonal group On (R) of Rn a unique invariant probability
measure. Taking the image of this Haar measure under the action of On (R)
on Gkn , we obtain the standard probability measure γk,n (denoted dP for
simplicity) on Gkn . This measure is of course invariant under the action of
On (R) on Gkn .
Let now Ḡkn denote the space of all the k-dimensional affine subspaces in
R . Representing elements P̄ of Ḡn−k
n
n by pairs (x, P ) ∈ Rn × Gn−k
n , where
x ∈ P , and P̄ = P̄x is the k-dimensional affine subspace of Rn , orthogonal to
P at x, we have the standard measure on Ḡn−kn : γ̄n−k,n = m ⊗ γk,n (m being
Here the coefficient c(n, i) is choosen in such a way that Vi (Qi ) = 1, where
Qi = [0, 1]i is the unit i-dimensional cube in Rn . In the above notations we
can also represent Vi (A) as follows:
Vi (A) = c(n, i) V0 (A ∩ P̄x ) dx dP or as:
P ∈Gin x∈P
Vi (A) = c(n, i) V0 (A ∩ πP−1 (x)) dx dP,
P ∈Gin x∈P
Fig. 3.1.
Remark. Of course when A is a tame set, it is easy to see that the function
P̄ → V0 (A ∩ P̄ ) is measurable. More generally, if A is a closed set of Rn , this
function is still measurable (see [Vit 1,2], [Zer]).
3 Multidimensional Variations 35
It is an exercise to compute the constant c(n, i, j); for instance when i = n−1
and j = 1 we have:
Vn−1 (A) = c(n, n − 1, 1) Vn−2 (A ∩ P̄ ) dP̄ .
P̄ ∈Ḡn−1
n
(n − 2) · On−1
And finally c(n, n − 1, 1) = 2 ·
2.In−1 .(n − 1) · On−3
(9) The variations of different orders of a set are independent ([Vit 1,2], 22,
Theorem 1): given any numbers ρ > 0 and 0 Ai + ∞, i = 0, . . . , n,
with A0 an integer and An < ρn , one can construct a closed set A lying
in the cube [0, ρ]n such that:
Vi (A) = Ai i = 1, . . . , n.
1
We denote On the n-volume of the n-unit sphere S n . Let us recall that
n
ξ∈S n
Ψ (ξ) dξ = α1
... αn
Ψ (ϕ(α1 , . . . , αn )) sinj−1 (αj ) dα1 . . . dαn , for Ψ :
j=1
S n → R, where ξ = ϕ(α1 , . . . , αn ) are the spherical coordinates.
π (n+1)/2
In particular, we get: On = On−1 α=0 sinn−1 (α) dα = Γ2π ((n+1)/2)
.
3 Multidimensional Variations 37
(10) The variations Vi (A), defined as the mean value of the number of con-
nected components of the slices P̄ ∩ A are constructed in the same way
as a lot of invariants in integral geometry. For instance, instead of the
number V0 (A ∩ P̄ ) of connected components of P̄ ∩ A, one may consider
the Euler-Poincaré characteristic χ(P̄ ∩ A), and the mean value:
Λi (A) = c(n, i) χ(A ∩ P̄ ) dP̄
P̄ ∈Ḡn−i
n
c(n, i, j)c(n − j, i − j) V0 (A ∩ Q̄, B ∩ Q̄) dQ̄dP̄ .
P̄ ∈Ḡn−j
n
i−j
P̄ ⊃Q̄∈Ḡn−j
Proof. The inequality is immediate for i = 0, and since the Vi ’s are defined
by integration of V0 , it follows for Vi , i > 0.
Remarks. Notice that the expected lower bound c(n) for the sum of all the
variations of sets A in B1 is necessarily less than 1, because on one hand, by
(9) the reals Vi (A), i
= 0, may be as small as we want, and Vi (A, B1 ) Vi (A)
by Proposition 3.3, and because on the other hand, 1 is the minimal number
of connected components of a nonempty set (i.e. V0 (A) ≥ 1)! Of course,
the important point of the theorem is that the centre of the ball Br lies
in the set A: as mentioned in property (9) above, one can construct a set
of arbitrary small variations in B1 , with no connected component inside B1
(V0 (A, B1 ) = 0). This set will certainly not contain the center of B1 but will
necessarily be very close to the boundary of B1 .
The proof of this theorem is rather tricky and complicated, so let us give
the main idea of the proof.
First of all, by the homogeneity property (property (6) above) of vari-
ations it is enough to prove (∗) for r = 1. Assume now for simplicity that
n = 2 and A is a tame set (see Definiion 4.17).
Let A0 be the connected component of A containing x, the center of the
ball B with radius 1.
If A0 ⊂ B, then already V0 (A, B) ≥ 1 (Fig. 3.2a).
If A0 ∩ (R2 \ B)
= ∅, we can find a tame (and thus rectifiable) curve Γ in
A, connecting x with the boundary of B. The length of Γ ∩ Br is at least r,
for any r ∈ [0, 1]. Hence by a classical integral geometry argument, the affine
3 Multidimensional Variations 39
Fig. 3.2.
Considering this draft of proof of Theorem 3.4, we see that the sum of
the variations of A in B is in fact related to the depth of embedding of A in
B, i.e. the real ρ = sup d(x, R2 \ B). One can actually prove (see [Vit 1,2],
x∈A
[Zer]) that, for any ball B of Rn (not necessarily centered at a point lying in
A) and of radius at most 1, we have:
n
Vi (A, B) ≥ c(n)ρn . (∗∗)
i=0
n
Vi (A, B1 ) ≥ c(n), (∗ ∗ ∗)
i=0
40 3 Multidimensional Variations
Fig. 3.3.
y
This implies that ξ ∈ Tρ (θ). If ω is the angle between ξ and we
||y||
have (see Fig. 3.3) sin(θ) = ||y|| cos(ω) (⇐⇒ y ∈ P̄ ), and d(y, P̄ ∩ S) =
cos(θ) − ||y|| sin(ω), thus ω = ω(θ, ρ) is given by:
cos(θ + ω)
d(y, P̄ ∩ S) = = ρ.
cos(ω)
3 Multidimensional Variations 41
y
Now by symmetry, Tρ (θ) contains the sphere of center and radius ω =
||y||
ω(θ, ρ).
Let K(s,||y||) be the connected component of A containing y in the closure
of B||y|| \ Bs for (s < ||y||) and K̃(θ,ρ) the image of K(sin(θ), sin(θ) ) by the
cos(ω(θ,ρ))
z sin(θ)
map z → , arccos ; K̃(θ,ρ) is connected in S × [0, ω(θ, ρ)].
||z|| ||z||
Furthermore, if (ξ , ω ) ∈ K̃(θ,ρ) then Tρ (θ) contains the sphere of center
sin(θ)
ξ and radius ω ; indeed (ξ , ω ) ∈ K̃(θ,ρ) ⇐⇒ ξ ∈ A, and then the
cos(ω )
cos(θ + ω )
sphere of center ξ and radius ω is contained in Tρ (θ), with ρ = ,
cos(ω )
but now ρ ≥ ρ implies Tρ (θ) ⊂ Tρ (θ).
Notice that y may be choosen such that K(s,||y||) , the connected compo-
nent of A containing y in the closure of B||y|| \Bs , hits the boundary of Bs , for
all s ∈ [0, ||y||]. Consequently we easily check that K(sin(θ), sin(θ) ) contains
cos(ω(θ,ρ))
sin(θ)
points of norm s, for all s ∈ [sin(θ), ||y|| = ].
cos(ω(θ, ρ))
Finally we have proved that there exists a connected set K (θ,ρ) in S ×
[0, ω(θ, ρ)], such that for all (ξ , ω ) in K̃(θ,ρ) , Tρ (θ) contains the sphere of
y
center ξ and radius ω , and that there exist ξ0 and ξω(θ,ρ) = with
||y||
(ξω(θ,ρ) , ω(θ, ρ)), (ξ0 , 0) ∈ K̃(θ,ρ) . We conclude that Tρ (θ) necessarily contains
a ball of radius ω(θ, ρ), and thus that the measure A(ρ, θ) of the set of
hyperplanes P(sin(θ),ξ) of Rn , with ξ ∈ Tρ (θ), is such that (see Footnote 1 on
Page 36):
ω(θ,ρ)
1
A(ρ, θ) ≥ sinn−2 (ν) dν, ()
I n−1 ν=0
π
with I n−1 = sinn−1 (ν) dν.
ν=0
Now we have by the inductive formula for relative variations:
n
n
Vi (A, B) = Vi (A, B)
i=0 i=1
n
= c(n, i, 1) Vi−1 (A ∩ P̄ , B ∩ P̄ ) dP̄ .
i=1 P̄ ∈Ḡn−1
n
n−1
where δA (P̄ ) is the depth of embedding of A ∩ P̄ in B ∩ P̄ .
π2 cos(θ)
n−1
We have: δA (P̄ ) dP̄ = ρn−1 d[−A(ρ, θ)]d[sin(θ)],
P̄ ∈Ḡn−1
n θ=0 ρ=0
and after an integration by parts:
π
2
cos(θ)
n−1
δA (P̄ ) dP̄ = cos(θ)A(ρ, θ) d[ρn−1 ]dθ.
P̄ ∈Ḡn−1
n θ=0 ρ=0
π π
2 −θ
1 2 cosn−1 (θ + ν) sinn−2 (ν)
cos(θ) dνdθ
I n−1 θ=0 ν=0 cosn−1 (ν)
π π
2 −ν
1 2 sinn−2 (ν)
= cosn−1 (θ + ν) cos(θ) dθdν
I n−1 ν=0 cosn−1 (ν) θ=0
π π
1 2 sinn−2 (ν) 2
= cosn−1 (ζ) cos(ζ − ν) dζdν
I n−1 ν=0 cosn−1 (ν) ζ=ν
π π
1 2 sinn−2 (ν) 2
= cosn (ζ) cos(ν) + cosn−1 (ζ) sin(ζ) sin(ν) dζdν
I n−1 ν=0 cosn−1 (ν) ζ=ν
π π
1 2 sinn−1 (ν) 2 1
≥ cosn−1 (ζ) sin(ζ) dζdν = .
I n−1 ν=0 cosn−1 (ν) ζ=ν n2 I n−1
n
µ(n)c(n − 1)
Finally, we have proved: Vi (A, B) ≥ = c(n).
i=0
n2 I n−1
3 Multidimensional Variations 43
Thus:
2n 1
n
M (, A) = q Vi (A)
c(n) i=1 i
n
1 2n
C(n) V i (A), where C(n) = .
i=1
i c(n)
Now, by Sard’s theorem, for all l ∈ RP2 there exists Ωl in πl (A) with m(A \
for all x ∈ Ωl , S = A∩Px is a smooth curve (of Px ). We ob-
Ωl ) = 0, such that
1
tain V0 (S) k(s) ds, where k is the absolute value of the curvature.
2π s∈S
(Notice that for S convex we can integrate
the
curvature
itself, and we have an
1
equality). Thus V1 (A) c(3, 1) k(s) ds dx dl. Now
2π l∈RP2 x∈l s∈A∩Px
1
k(s) = k(y, l), where k(y, l) is the mean curvature of A at y ∈ A in the
sin(θ)
direction of the plane passing through y and orthogonal to l (Fig. 3.4).
Fig. 3.4.
dx
But ds = dy. Therefore:
sin(θ)
1
V1 (A) c(3, 1) k(y, l) dy dl
2π l∈RP2 y∈A
1
= c(3, 1) k(y, l) dl dy.
2π y∈A l∈RP2
Finally, k(y, l) dl c (k1 (y) + k2 (y)), by the Euler formula, where
l∈RP2
k1 (y) and k2 (y) are the absolute values of the main curvatures of A at y ∈ A.
Thus we’ve proved the following.
Theorem 3.6. ([Leo-Mel]) For A a compact C 2 surface in R3 ,
3 Multidimensional Variations 45
V1 (A) C (k1 (y) + k2 (y)) dy.
y∈A
Via theorem 3.4 this proves also the formula for M (, A), given in the end
of Chapter 2. We mention also the following result: if Ak −−−→ A in the
n→∞
Hausdorff metric, and Vi (Ak ) are uniformly bounded, i = 0, 1, . . . , n, then
lim Vi (Ak ) ≥ Vi (A) ([Iva 1], theorem II.6.1).
k→∞
of A are semialgebraic, and the diagram of these sets depends only on the
diagram of A (see [Ben-Ris], [Cos], [Har 3], [Loj]).
Proposition 4.4. Any semialgebraic set admits semialgebraic stratification
(i.e. a partition into smooth submanifolds (Ai )i∈I of Rn - the strata, that are
semialgebraic sets of Rn , such that the family (Ai )i∈I is locally finite and
verifies the following property, called the frontier property: if Āi Aj
= ∅
then Aj ⊂ Āi . See Fig. 4.1 for an example of strafication), with the number
of strata and their diagrams depending only on the diagram of the initial set
(see [Loj]). The same is true for triangulations (see [Cos], [Har 3], [Loj]).
Fig. 4.1.
Remark. The above inequality is not true in general for isolated, but possibly
degenerate, solutions of (∗). For instance, let us consider the following system:
f1 (x1 , . . . , xn−1 ) = 0 deg d
.. ..
. .
fn−1 (x1 , . . . , xn−1 ) = 0 deg d
and let this system have dn−1 nondegenerate solutions. Now consider the
system in Rn :
g = f12 + . . . + fn−1
2
= 0 deg 2d
xn = 0 deg 1
.. ..
. .
xn = 0 deg 1
Clearly this system has dn−1 isolated (of course degenerate) solutions. But
the product of degrees is 2d < dn−1 , for d sufficiently big and n ≥ 3.
The reason is that a degenerate isolated zero of a real system (∗) can lie
on a higher-dimensional component of zeros of the complexification of (∗).
Definition 4.7. For any finite sequence D of integers of diagram type D =
1
p
n, p, j1 , . . . , jp , (dij ) i=1,...,p define B0 (D) as di (di − 1)n−1 , where di =
j=1,...,ji
2 i=1
ji
0 (A) as the infimum of B0 (D) over
dij . For a semialgebraic set A, define B
j=1
the diagrams D of all representations of A.
Theorem 4.8. For any semialgebraic set A ⊂ Rn , the number of bounded
0 (A), is bounded by B
connected components of A, B 0 (A).
Proof. We follow the proof of [Mil], [War].
Obviously, it is enough to prove that the number of connected components
q
1 q
of A = {pj 0}, deg pj = dj , is at most d(d − 1)
> n−1
,d= dj .
≥ 2
j=1 j=1
50 4 Semialgebraic and Tame Sets
We may assume that only the inequalities ≥ define A, and hence that A
is closed. Indeed, let us choose a point xα in each (not necessarily bounded)
connected components Aα of A. The number of connected component Aα ,
and hence of xα is finite (see e.g. [Mil], [Whi 2]).
If one of the inequalities defining A has the form {pj > 0}, let us denote
minα (pj (xα )) = δ > 0. Hence if we replace in the definition of A this
δ
inequality by {pj − ≥ 0}, we obtain a new set A ⊂ A. Of course any
2
connected components of A lies in exactly one connected component of A,
0 (A ) ≥ B
and all the points xα are still in A , therefore B 0 (A).
We may assume that each component of A has a nonempty interior. In-
q
deed, A = {pj ≥ 0} is closed, hence the minimal distance between the
j=1
components Aα of A (inside a ball B containing all the bounded compo-
ρ
nents of A) is ρ > 0. Let U be the open -neighborhood of A, and let
3
ξ = max min pj (x). We have ξ < 0, because the continuous func-
x∈B\U 1 j q
p−η = 0 deg d
∂p
= 0 deg d − 1
∂x2
∂p
= 0 deg d − 1
∂x3
.. ..
. .
∂p
= 0 deg d − 1
∂xn
By proposition 4.6, the number of critical points is at most d(d − 1)n−1 , and
therefore:
B 0 (Z) 1 d(d − 1)n−1 .
0 (A) B
2
Corollary 4.9. Let A ⊂ Rn be a semialgebraic set with diagram n, p, j1 , . . . ,
jp , (dij ) i=1,...,p . Then:
j=1,...,ji
Proof. For the first bound, we add to the inequalities defining A the inequal-
n
ity r2 − x2i ≥ 0 of degree 2, and for the second bound we substitute n −
i=1
1
p
variables in the equations by the others. The bound di (di − 1)−1 does
2 i=1
52 4 Semialgebraic and Tame Sets
not depend on the radius of the ball, but only the degrees, thus it also bounds
the number of connected components of A ∩ P itself.
1. C ⊂ Aξ , the ξ-neighborhood of A.
2. The Hausdorff distance between f (C) and f (A) is at most Kξ, with K a
Lipschitz constant of f on Brn .
3. dim(C) m.
4. Any -dimensional plane in Rn intersects C in at most ν connected com-
ponents, where ν is explicit in terms of m, n, , dk and d¯ij ,, assuming that
p q
A is given as {pij ≥ 0} and deg(pj ) = d¯ij .
i=1 j=1
Fig. 4.2.
Let us compare the statement of theorem 4.12 with another metric result
on semialgebraic sets.
Theorem 4.14. ([Har 3], [Loj]) For any compact semialgebraic set, there
are constants K and α > 0, such that any x and y belonging to the same
component of A, can be joined in A by a curve of length K||x − y||α .
Consider, for example, curves of degree 2 in the plane. For any such
curve,the biggest exponent α in this theorem is 1: take x tending to y in
A; the ratio dA (x, y)/||x − y||α , where dA (x, y) is the distance from x to y
in A, is bounded only for α 1. Furthermore, for any curve of degree 2 in
the plane, there obviously exists a constant K such that any x and y in A
can be joined by a curve in A of length K||x − y||. But the constant K
strongly depends on the concrete curve, as it is illustrated in Fig. 4.3. This
is in contrast with the conclusion of Theorem 4.12: any two points on any
parabola in Br2 can be joined by a curve of length C · r.
4 Semialgebraic and Tame Sets 55
Fig. 4.3.
The fact that all the coefficients in our inequalities depend only on degrees,
and not on a specific choice of the polynomials, is essential in our approach
(see for instance the proof of Corollary 4.9).
More generally we say that a closed set A has the Whitney property (with
exponent α), if for each a ∈ A there exists a neighbourhood U of a and two
positive constants K and α such that any points x and y in U can be joined
by a curve of length K.||x − y||α .
It has been proved that each closed subanalytic set has the Whitney prop-
erty in [Sta] (see also [Kur], in which the author proves that any subanalytic
set A admits an analytic stratification such that each stratum has the Whit-
ney property with α = 1). For questions concerning the Whitney property of
the geodesic distance see [Kur-Orr], and for questions concerning the suban-
alyticity or the regularity of the sub-Riemannian dictance see [Agr1], [Agr2]
for instance.
The property mentioned in corollary 4.9 is classically called the Gabrielov
property (see [Gab 4]). More precisely, we will say that a set A has the local
Gabrielov property if for any a ∈ A ⊂ Rn there exists a neighbourhood U of
a and an integer B(= B(a, U )) such that for any -dimensional affine plane
P of Ḡn , the number of connected components of U ∩ A ∩ P is bounded by B.
If we can take U = Rn in the above definition, we will say that A has the
global Gabrielov property
The Corollary 4.9 says that any semialgebraic set has the Gabrielov prop-
erty (with explicit bound B depending only on and the degrees of the
polynomials in the definitions of A).
However not only semialgebraic sets, but a very large class of sets has the
Gabrielov property. Let us give two definitions:
Definition 4.15. ([Dri-Mil]) An analytic-geometric category C is the datum
for each real analytic manifold M of the collection C(M ) of sets of M such that
the five following conditions are satisfied (for each real analytic manifold N ):
56 4 Semialgebraic and Tame Sets
Let us note that the subanalytic sets of any real analytic manifold M
(ie sets which locally are projections of relatively compact semianalytic sets.
The term subanalytic was introduced by Hironaka in [Hir 1], but the notion
has been first considered by Thom in [Tho 2] and L ojasiewicz [Loj]. See for
instance [Bie-Mil 5] or [Den-Sta]) form an analytic-geometric category, which
is the smallest analytic-geometric category.
Because in AG3 we allow analytic functions, and AG5 concerns only
bounded sets of C(R), the behaviour at infinity of sets in analytic-geometric
categories is not controlled, and these sets are not as globally nice as semial-
gebraic sets. For globally nice sets we have the following categories:
Definition 4.16. ([Dri], [Dri-Mil]) A structure on the real field (R, +, .)
is a sequence (Sn )n∈N such that for each n ∈ N:
S1 Sn is a boolean algebra of subsets of Rn , with Rn ∈ Sn .
S2 Sn contains the diagonal {(x1 , . . . , xn ) ∈ Rn ; xi = xj } for 1 i < j n.
S3 If A ∈ Sn , then A × R and R × A belong to Sn+1 .
S4 If A ∈ Sn+1 then π(A) ∈ Sn , where π : Rn+1 → Rn is the standard
projection.
S5 S3 contains the graphs of addition and multiplication.
The structure is said to be o-minimal if it satisfies the following additional
axiom:
S6 (o-minimal axiom) S1 consists of the finite unions of intervals of all
kinds.
This proof is short, but it uses two deep results: the existence of Whitney
stratifications of a proper tame map and the first isotopy lemma of Thom-
Mather. Of course, as mentionned in the proof, these results do not require
the Gabrielov property (at least for planes
= Rn ). For the semialgebraic
case, we may refer to [Har 2], that do not requires the firts isotopy lemma
and provides a method without integration of vector fields.
In what follows we will use the global Gabrielov property.
5 Variations of Semialgebraic and Tame Sets
where 1πP (A) is the indicator function of the projection πP (A) of A onto P .
But since A ⊂ B, for any P ∈ Gin , 1πP (A) 1πP (B) , and we can continue our
inequalities:
Vi (A) c(n, i)B0,n−i (A) 1πP (B) (x) dx dP
P ∈Gin x∈P
c(n, i)B0,n−i (A) V0 (B ∩ Px ) dx dP = B0,n−i (A)Vi (B),
P ∈Gin x∈P
Remark. Of course no inequality of this type holds for nontame sets: e.g.,
we can take as A a curve of infinite length in the unit square B in R2 and
√ π
then V1 (A) = ∞, V1 (B) c(2, 1) 2 = √ .
2
We find again 4.13, as a corollary of 5.1:
Corollary 5.2. Let A ⊂ Rn be a tame set of dimension . Then for any ball
Brn of radius r in Rn ,
and the classical Cauchy-Crofton formula ([Fed 1], 5.22 or [Fed 2], 2.10.15)
gives:
V ol (A) = V (A) = c(n, ) V ol(sP ) dP.
P ∈Gn
5 Variations of Semialgebraic and Tame Sets 61
Instead of considering bounded tame sets, one can consider germs of tame
sets (at the origin, for simplicity). Let us denote the germ of A by A0 .
Note that the projection of such a germ onto a given -dimensional vector
plane P is not well defined (for instance the projection (in some chart) onto
R2 of the germ at a point (a, 0) ∈ P1 × R2 of the blowing-up of the plane
R2 at the origin depends on the chosen representation of the germ. Following
the terminology of Thom, this mapping is not “sans éclatement”). However,
the projection of A0 is well defined for generic directions, the transverse
directions, namely projections onto planes P such that P ⊥ ∩ C 0 A0 = {0},
where C 0 A0 is the tangent cone of A0 at the origin (the tangent cone of the
germ A0 is the semi-cone of Rn consisting of all limit secants to A0 at the
origin).
It follows that a generic projection πP (A0 ) of the germ A0 splits into a
finite number of germs of tame domains k1P , . . . , kνPP , such that A0 covers
the germ kjP with multiplicity eP j ∈ N. With the same arguments as in 4.18
(the generic local topological triviality provided by the first isotopy lemma
of Thom) one can show that the numbers νP and eP j are uniformly bounded
as P ranges over all elements of Gn (in fact one can easily see that νP nP
and ePj Ej B0,n− (A), for any bounded set A that represents A0 ).
P
νP
Θ (σP ) = eP P
j Θ (kj ),
j=1
Notice that the measure that occurs in this formula is the probability mea-
sure dP on Gn , and not c(n, ) dP as in the Cauchy-Crofton formula for the
volume.
Actually, one has a more general result:
Theorem 5.3. ([Com 2], [Com 3]) Let A0 be a germ at the origin of a
tame set of dimension in Rn , G ⊂ Gn a tame set, G < On (R) a group
acting transitively on G and µ a G-invariant measure on G. Suppose that
62 5 Variations of Semialgebraic and Tame Sets
Remarks.
• These results illustrate a general principle: global data on algebraic sets,
such as the number of connected components in a plane subspace of the
ambient space, uniform bounds for these numbers, Betti numbers, etc...
are related to the degree, whereas the localization of these data, or local
invariants, are related to the multiplicity.
• It is a direct consequence of Corollary 5.2 and of the definition of the
density, that a semialgebraic subset A ⊂ Rn has its density bounded by its
complexity, namely:
Θ (A0 ) c(n, k)B0,n− .
However, one can obtain a better bound for the density of A in terms of
complexity, by using the Cauchy-Crofton formula for the density (Theo-
rem 5.3):
Θ (A0 ) B0,n− .
Remark. We can find a better bound for Θ (A0 ), say Θ (A0 ) β0,n− , where
β0,n− has the same nature as B0,n− , the multiplicity mij of the polynomi-
als pij replacing the degrees dij . We just have to prove the local version of
Bezout’s theorem, which only involves the multiplicity mij .
5 Variations of Semialgebraic and Tame Sets 63
the last inequality being a consequence of 5.2. But the last term of this sum
is bounded by:
n−1
r r
C(n) · B(A) ( )j n · C(n) · B(A)(1 + ( )n−1 ),
j=0
where B(A) = maxj∈{1,...,n−1} B0,n−j (A) · V olj (B1j ) , which proves the last
inequality with C = n · C(n) · B(A).
If we know a priori that our semialgebraic set has dimension less than n,
we get nontrivial bounds on its -entropy without any specific information,
except its algebraic complexity:
Corollary 5.7. Let A ⊂ Rn be a tame set of dimension < n. Then for any
ball Brn ⊂ Rn :
r r
M (, A ∩ Brn ) C(n) B0,n−j (A)V olj (B1j )( )j C[( ) + 1],
j=0
64 5 Variations of Semialgebraic and Tame Sets
It follows that:
k
Vi (A)
M
k
Vi (A, Bj )
≥ ≥ M.c(n), (by Theorem 3.4).
i=0
ηi j=1 i=0
ηi
1 k
c(n, i).V oli (B1i )
.α(n) .
c(n) ηk
i=0
Remarks.
5 Variations of Semialgebraic and Tame Sets 65
• Proposition 5.8 implies that when M (, B) is big, k and the degrees dij
cannot be too small.
• If for all > 0, there exists a semialgebraic set A of dimension k such
that B ⊂ (A ) , and if the complexity of A is uniformly bounded, then
dime (B) k.
The properties of the -entropy allow us easily to get bounds also for the
volume of η-neighborhoods of tame sets.
Theorem 5.9. Let A be a tame set of dimension < n. Then for any ball
Brn ⊂ Rn and for any η > 0, the volume of the η-neighborhood of A ∩ Brn is
bounded as follows:
V oln (A ∩ Brn )η C (n) B0,n−j (A) · rj · η n−j C · [r η n− + η n ],
j=0
But ∂A and ∂B are tame sets of dimension less than n − 1 (with diagrams
depending only on D(A) and D(B), in the semialgebraic case), hence by
Theorem 5.9, we have:
Also taking the δ-neighborhoods for tame sets does not change variations
too much:
Theorem 5.11. Let A ⊂ Brn be a tame set. Then, for all i ∈ {1, . . . , n}:
i−1 + δ i ),
Vi (Aδ ) Vi (A) + C(δr V0 (Aδ ) V0 (A),
Let us estimate the inner integral. Let πP denote the orthogonal projection
onto P . Of course, we have π(A) ⊂ π(Aδ ) ⊂ π(A)δ . By theorem 5.9:
We have:
V0 (Aδ ∩ Px ) dx =
x∈P
V0 (Aδ ∩ Px ) dx + V0 (Aδ ∩ Px ) dx.
x∈πP (Aδ )\πP (A) x∈πP (A)
Since Aδ is tame, we have the global Gabrielov property: there exists C such
that V0 (Aδ ∩ Px ) C , and consequently the first integral is bounded by:
C C(δri−1 + δ i ).
Now:
5 Variations of Semialgebraic and Tame Sets 67
V0 (Aδ ∩Px ) dx = V0 (Aδ ∩Px ) dx+ V0 (Aδ ∩Px ) dx.
x∈πP (A) x∈∆δ x∈πP (A)\∆δ
But dim(∆) < i (by Sard’s theorem), hence: V oli (∆δ ) C [δri−1 + δ i ], by
theorem 5.8, and once more:
V0 (Aδ ∩ Px ) dx C C (δri−1 + δ i ).
x∈∆δ
Furthermore:
V0 (Aδ ∩ Px ) dx V0 (A ∩ Px ) dx
x∈πP (A)\∆δ x∈πP (A)\∆δ
V0 (A ∩ Px ) dx.
x∈πP (A)
i−1 + δ i ).
Vi (Aδ ) Vi (A) + C(δr
Of course, nothing of this type holds for non-tame sets: for A = Q ∩ [0; 1],
Aδ = [−δ; 1 + δ], and V1 (Aδ ) = 1 + 2δ, V1 (A) = 0.
As an easy consequence of Theorem 5.11 we obtain the following extension
of Theorem 5.10 to lower variations:
Theorem 5.13. Let A, B ⊂ Brn be tame sets. Then if dH (A, B) = δ,
Vi (B) + δri−1 + δ i
Vi (A) C
Vi (A) + δri−1 + δ i ,
Vi (B) C
where C depends only on constants of the type B0,j , and thus only on D(A)
and D(B) when A and B are semialgebraic. In particular, if dH (A, B) → 0,
i (B) and lim sup Vi (B) CV
then lim sup Vi (A) CV i (A).
Fig. 5.1.
Then:
In fact, for tame sets, -entropy and variations are equivalent in the fol-
lowing sense:
Theorem 5.14. For A a bounded tame set of dimension , we have:
1 i
1
C1 Vi (A).( ) M (, A) C2 Vi (A).( )i ,
i=0
i=0
where C1 and C2 depend only on constants of the type B0,j , and thus only
on D(A) in the semialgebraic case.
Proof. The upper bound is given in 3.5; it remains to prove the lower bound.
q
We have: A ⊂ Bj , with q = M (, A) and Bj some balls of radius .
j=1
Hence, by the property (7) of variations,
q
Vi (A) Vi (Bj ∩ A).
j=1
But by Theorem 5.1: Vi (Bj ∩ A) c(n, i).B0,n−i (A ∩ Bj ).V oli (B1i ).i = C.i ,
hence:
1 1
Vi (A) qC.i or M (, A) = q ≥ Vi (A). i .
C
Adding all these inequalities for i = 0, . . . , , we obtain:
1
M (, A) ≥ C1 Vi (A). i .
i=1
A = [0; 1] ∩ Q, the set of rational points in [0; 1], V0 (A) = ∞, but M (, A) =
1
M (, [0; 1]) = [ ] + 1.
Another property of variations, which distinguish tame sets among more
complicated ones, is the following:
Theorem 5.15. Let A be a tame set. Then for each i ≥ 0 and 0 j i:
where C depends only on constants of the type B0,k , and thus only on D(A)
in the semialgebraic case.
Proof. By the inductive formula for variations (property 8 of Chapter 3), we
have:
Vi (A) = c(n, i, j) Vi−j (A ∩ P̄ ) dP̄ .
P̄ ∈Ḡn−j
n
c(n, i, j).B0,i−j (A ∩ P̄ ).Vi−j (A) V0 (P̄ ∩ A) dP̄
P̄ ∈Ḡn−j
n
c(n, i, j)
= .B0,i−j (A ∩ P̄ ).Vi−j (A).Vj (A) = C.Vi−j (A).Vj (A).
c(n, j)
Remark. Theorem 7.6 below generalizes this property for mappings of semi-
algebraic sets.
Corollary 5.16. Let A ⊂ Brn be a tame set. Then for any i = 0, 1, . . . , n and
0 j i,
C
Vj (A) ≥ i−j · Vi (A).
r
i−j .
Proof. By Theorem 5.1, Vi−j (A) C.r
Remark. Theorem 5.15 and Corollary 5.16 show that there exist strong cor-
relations among the variations of a tame set, or semialgebraic sets of fixed
complexity. In general, variations are independent (property (9) of Chap-
ter 3). Let us give here an example, contradicting Corollary 5.16, for sets
of unbounded complexity. Let S be a curve in B12 ⊂ R2 of infinite length.
Let à be the set in R3 defined by A = (S × [0; 1]) ∪ (B12 × ({0} ∪ {1})).
Then one can easily check that any plane cuts A in a connected set. Hence
V1 (A) V1 ([0; 1]3 ) = 1, while V2 (A) V ol2 ([0; 1]3 ) = ∞.
70 5 Variations of Semialgebraic and Tame Sets
n−q
r q
r
Vq+1 (A) ≥ C q+1 .( ( )i )−1 .[M (, A) − Ci ( )i ].
i=0
i=0
n
1 i
q
r i n
1
M (, A) Vi (A)( ) Ci ( ) + Vi (A)( )i .
i=0
i=0
i=q+1
Hence
n
1 q
r
Vi (A)( )i ≥ M (, A) − Ci ( )i .
i=q+1
i=0
n
1 n
1
Vi (A)( )i CVq+1 (A) ri−q−1 .( )i
i=q+1
i=q+1
1 n
r
= CVq+1 (A). ( )i
rq+1 i=q+1
r q+1 1 r i
n−q
= CVq+1 (A).( ) . q+1 ( ).
r i=0
n−q
r q
r
Vq+1 (A) ≥ C q+1 ( ( )i )−1 (p − Ci ( )i )
i=0
i=0
n
q
r
≥ C n−q−1
.(p − Ci ( )i ) > 0.
r i=0
5 Variations of Semialgebraic and Tame Sets 71
In particular, dim(A) ≥ q + 1.
The following example shows that in general the degree n in the expres-
sion cannot be improved.
Let A be the ball Brn itself and q < n. Then for small, M (, A) ∼ ( r )n ,
n
and the expression in Theorem 5.17 gives Vq+1 (Brn ) ≥ C n−q−1 M (, A) =
r
Crq+1 , the sharp bound, up to coefficients.
However, if we know a priori that dim(A) n, we get the following:
Corollary 5.19. If under the assumptions of Corollary 5.17 we have in ad-
dition = dim(A), q + 1 n, then:
q
r
Vq+1 (A) ≥ C −q−1
(p − Ci ( )i ) > 0.
r i=0
In particular, for = q + 1,
q
r
Vq+1 (A) ≥ C q+1 (p − Ci ( )i ) > 0.
i=0
by the inequality:
1
M (, A) Vi (A)( )i .
i=0
Remark. Results of Theorem 5.17 and Corollaries 5.18 and 5.19 can be con-
sidered as the generalization of the following fact: if the polynomial vanishes
at “too many” points, then it vanishes identically. Indeed, these results can be
formulated as follows: if the semialgebraic set A contains more “separated”
points than is prescribed for the sets of dimension q and of the same com-
plexity, than its dimension is at least q + 1 and the (q + 1)-volume is not too
small.
Theorem 5.20. Let A ⊂ Rn be a tame set. Then for each i = 0, 1, . . . , n,
there exists a tame set Ci ⊂ A, with dim(Ci ) = i and Vi (Ci ) ≥ λVi (A).
When A is semialgebraic, Ci is also semialgebraic and the diagram of Ci and
λ depend only on the diagram of A.
Proof. We have:
Vi (A) = c(n, i) V0 (A ∩ Px ) dx dP
P ∈Gin x∈P
72 5 Variations of Semialgebraic and Tame Sets
c(n, i)B0,n−i V oli (πP (A)) dP
P ∈Gin
In particular, for some P ∈ Gin , V oli (πP (A)) ≥ λVi (A). Now by Theorem
4.10 (this result being valid for tame sets, because one can stratify a tame
map), there is an i-dimensional set C ⊂ A of large enough i-dimensional
measure, such that dim(C) = i, and πP (C) = πP (A). Thus we get:
Vi (C) = V oli (C) ≥ V oli (πP (C)) = V oli (πP (A)) ≥ λVi (A).
Of course, nothing of this sort is true for non-tame sets: take A to be the
set of rational points in [0; 1].
Remark. The property given by Theorem 5.21 is much more precise than
the similar conclusion one can obtain by comparison of “global volumes”.
Indeed, the volume of an η-neighborhood of A ⊂ B1n , dim(A) = , is of order
η n− . Comparing this with the volume of the ball of radius r, V oln (B1n )rn ,
1
we obtain only V oln (B1n )rn η n− , or r Cη 1− n , instead of r Cη
Theorem 5.20 may be generalized in the following way: if a tame set B
lies in the η-neighborhood of an -dimensional tame set, then the variations
Vj (B), j > , are bounded in terms of lower variations of B and η.
Theorem 5.22. Let A ⊂ Brn be a tame set, dim(A) = < n. Let η > 0 be
fixed. Then if another tame set B ⊂ Brn is contained in the η-neighborhood
Aη of A, its variations satisfy the following inequality:
for any j > Vj (B) C̄ · V0 (B) · η j + C̄ · [Vi (B) + ηri−1 + η i ]η j−i ,
i=1
5 Variations of Semialgebraic and Tame Sets 73
where C̄ depends only on constants of the type B0,j , and consequently, only
on the diagram of A and B, in the semialgebraic case.
Proof. If B ⊂ Aη , then also B ⊂ (A ∩ Bη )η . Now by Theorem 5.11,
i−1 + η i ).
Vi (Bη ) C · Vi (B) + C(ηr
Vj (Aη ) C1 Vi (A)η j−i .
i=0
1
V olj πP (A) η cη j M (η, πP (A)) C1 η j Vi (A)( )i = C1 Vi (A)η j−i .
i=0
η i=0
Vj (B) C Vj (A ∩ Bη )η C C1 Vi (A ∩ Bη )η j−i
i=0
C C1 C 0 (B)η j
[Vi (B) + ηri−1 + η i ]η j−i + C C1 CV
i=1
= C2 Vi (B)η j−i + C3 ri−1 η j−i+1 + C4 η j .
i=0 i=0
In the next chapter we describe variations of the images and the preimages
of polynomial mappings (generally speaking, on semialgebraic sets). This de-
scription is given in terms of (as usual) “algebraic complexity” of the sets and
mapping involved, and it requires some metric information on the mapping.
Usually this information concerns upper or lower bounds for the first differen-
tial (but our approach allows one to encounter also more delicate properties:
e.g., the presence of higher Thom-Boardman singularities, as in Theorem 8.10
below).
The bounds on variations, obtained in this chapter, being restated in term
of -entropy, form a part of our main results in the algebraic category: the
quantitative Morse-Sard theorem and the quantitative transversality theo-
rem.
First of all, to describe quantitatively the behavior of the differential of
a considered mapping, we need some results from exterior algebra (for com-
plements see, for instance, [Bou], Algèbre, Chap. 3), and for the sake of
non-familiar readers we develop below some very basic results in this area.
Of course the well-informed reader can immediately go to Lemma 6.2.
n
Let iR = Rn . . . Rn denote, as usual, the tensor product of
Rn , . . . , Rn (i times). We recall that this vector space is formally defined as the
quotient E/F , where E is the vector space of real valued functions on (Rn )i
that vanish outside finite sets (1[{(v1 ,...,vi )}] being denoted by v1 ⊗ . . . ⊗ vi )
and F is the subspace of E generated by elements of the type:
and
v1 ⊗ . . . ⊗ vj−1 ⊗ λx ⊗ vj+1 ⊗ . . . ⊗ vi
−λ(v1 ⊗ . . . ⊗ vj−1 ⊗ x ⊗ vj+1 ⊗ . . . ⊗ vi ),
for all x, y ∈ R , all λ ∈ R and all j ∈ {1, . . . , i}.
n
We also denote by v1 ⊗ . . . ⊗ vi the class of v1 ⊗ . . . ⊗ vi in i Rn .
Now let us consider the subspace G of i Rn generated by all elements
of the type:
v1 ⊗ . . . ⊗ vj−1 ⊗ x ⊗ x ⊗ vj+2 ⊗ . . . ⊗ vi .
n n
The vector space i R /G, denoted by i R , is by
definition the i-th
exterior product of Rn . The class of v1 ⊗ . . . ⊗ vi in i Rn is denoted by
v1 ∧ . . . ∧ vi .
Of course we have:
v1 ⊗ . . . ⊗ vj−1 ⊗ x ⊗ y ⊗ vj+2 ⊗ . . . ⊗ vi
+v1 ⊗ . . . ⊗ vj−1 ⊗ y ⊗ x ⊗ vj+2 ⊗ . . . ⊗ vi
= v1 ⊗ . . . ⊗ vj−1 ⊗ (x + y) ⊗ (y + x) ⊗ vj+2 ⊗ . . . ⊗ vi
−(v1 ⊗ . . . ⊗ vj−1 ⊗ x ⊗ x ⊗ vj+2 ⊗ . . . ⊗ vi )
−(v1 ⊗ . . . ⊗ vj−1 ⊗ y ⊗ y ⊗ vj+2 ⊗ . . . ⊗ vi ) ∈ G,
and in particular:
(w, w ) = ξI .ξI .
I∈Λ(n,i)
We can naturally define a product ∧ : i Rn × j Rn → i+j Rn , by the
equality: (v1 ∧ . . . ∧ vi ) ∧ (v1 ∧ . . . ∧ vj ) = v1 ∧ . . . ∧ vi ∧ v1 ∧ . . . ∧ vj .
6 Some Exterior Algebra 77
• L(ej ) = λj (L)ej , for j ∈ {1, . . . , q}, and L(ej ) = 0, for j > q (if n > m).
• The norm wi (L) = ||Li || = max ||Li (w)|| is attained on the homo-
||w||=1,w∈∧i Rn
geneous w = e1 ∧. . .∧ei , and hence is equal to λ0 (L) . . . λi (L). In particular,
this norm is equal to the maximal expansion of the i-dimensional volume
by L.
Proof. The statements of this proposition are well known and their proof is
an exercise in linear algebra. For sake of completeness we give it here.
Consider the symetric bilinear nonnegative form ω on Rn given by
x21 x2k
+ . . . + 1.
λ21 λ2k
Hence, the numbers λ1 , . . . , λk are the semiaxes of the ellipsoid L(B1n ), which
is given by
x21 x2
2 + . . . + k2 1, xk+1 = . . . = xm = 0.
λ1 λk
This proves the first part of the proposition.
Now let us define the following symmetric bilinear form ωi on i Rn :
ωi (w1 , w2 ) = Li (w1 ), Li (w2 ) .
Li (ej1 ∧ . . . ∧ eji ) = λj1 . . . λji (ej1 ∧ . . . ∧ eji ), and (eI , eJ ) = δI,J .
i
wi (L1 + L2 ) k(n, m, i) wj (L1 )wi−j (L2 ), i = 0, . . . , min(n, m).
j=0
Proof. Let w = v1 ∧ . . . ∧ vi ∈ i Rn be the element on which the norm
wi (L1 + L2 ) = ||L1 + L2 || is attained. Then:
Hence:
wi (L1 + L2 ) = ||(L1 + L2 )i (w)|| || ∧ L1 (vj )||.|| ∧ L2 (vk )||
j∈I k∈I¯
I⊂{1,...,i}
i
||(L1 )|I| ||.||(L2 )i−|I| || k(n, m, i) wj (L1 )wi−j (L2 ).
I⊂{1,...,i} j=0
The next lemma will be used in the proof of the “quantitative transver-
sality” theorem below. A general structure of transversality results is the
following: we have a mapping F : N × T → M , where T is the space of
parameters. We assume that the parameters act non-degenerately, i.e. that
F|{∗}×T : {∗} × T → M is non degenerate. The desirable conclusion is that
for a generic value of parameters t0 ∈ T , ft = FN ×{t0 } : N × {t0 } → M is in
some sense nondegenerate. The following lemma gives a linear version of this
statement; Rq in this lemma serving both as T and M .
Consider the product Rp × Rq , and let
π1 : Rp × Rq → Rp , π2 : Rp × Rq → Rq
i
π2 ) k.wi (L−1 )
wi ( i−j (L ),
wj (L).w
j=0
i = 1, . . . , D = min q, dim(P ) . Here the constant k depends only on p, q,
dim(T).
Remark. In the special case, where T = ker(L), we obtain dim(T) = p,
is not zero). Thus if L is
π2 ) k.wD (L−1 )wD (L ) (since only w0 (L)
wD (
degenerate, also π2 : T → R is degenerate. This last statement is used in
q
i
π2 ) wi (L
wi ( −1 − L ◦ π
)wi (L 1 ) k.wi (L−1 ) i−j (L ),
wj (L)w
j=0
wi (πP ◦ L).(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) wi (L).(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) .
Hence:
ω(P ) wi (L).[(πker(πP ◦L) )∗ ωker(πP ◦L) ] ∧ [(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ]
= wi (L).ωRn .
Integrating over Gim , we obtain the left-hand side inequality.
We recall (see Proposition 6.1) that we denote by e1 , . . . , ek , ek+1 , . . . , en
an orthonormal basis of Rn such that ker(L) =< ek+1 , . . . , en >, e1 =
L(e1 ) L(ek )
, . . . , ek = is a part of an orthonormal basis of Rm , and
λ1 (L) λk (L)
λ1 (L) ≥ . . . ≥ λk (L) > 0 are the positive eigenvalues of the nonnegative
symmetric form t LL.
Now on the other hand, we have:
(πP ◦ L)∗ ωP ≥ wi (L).JL (P ).(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) , (∗)
where JL (P ) = J(π:<e1 ,...,ei >→P ) and where J(π:<e ,...,e >→P ) is the Jacobian
1 i
of the projection onto P of the i-dimensional subspace of Rn spanned by
e1 , . . . , ei .
For linearity reasons, to prove (∗) it suffices to prove that the inequality
(∗) holds for every (e1 , . . . , ei ), 1 , . . . , i ∈ {1, . . . , k}, because of course we
have ker⊥ (πP ◦ L) ⊂< e1 , . . . , ek >.
The following equality holds:
[(πP ◦ L)∗ ωP ](e1 , . . . , ei ) = ωP (λ1 (L).πP (e1 ), . . . , λi (L).πP (ei ))
where = +1 or −1.
But we have also:
82 6 Some Exterior Algebra
[(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ](e1 , . . . , ei ) = .| det(πker⊥ (πP ◦L) )|. (2)
λ1 (L) . . . λi (L) · J(π:<e ,...,e >→P ) · [(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ](e1 , . . . , ei ),
1 i
for all 1 , . . . , i ∈ {1, . . . , k}. Now considering that there exists a constant k
such that:
[(πP ◦ L)∗ ωP ] = k · [(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ],
we obtain that:
k · [(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ] ≥
∗
λ1 (L) . . . λi (L) · J(π:<e ,...,e >→P ) .[(πker⊥ (πP ◦L) ) ωker⊥ (πP ◦L) ],
1 i
ω(P ) ≥ wi (L) · JL (P ) · [(πker(πP ◦L) )∗ ωker(πP ◦L) ] ∧ [(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ]
= wi (L) · JL (P ) · ωRn ,
and integrating over Gim , we deduce that:
ω(P ) dP ≥ wi (L) · JL (P ) dP · ωRn ,
P ∈Gim P ∈Gim
7 Behaviour of Variations
under Polynomial Mappings
where N (f|A , y) is (f −1 ({y})∩A), the number of points of the set f −1 ({y})∩
A and wn (Df(x) ) is defined in the preceding chapter. In particular, the char-
acteristic function of f (A), 1f (A) (y) N (f|A , y), and thus we obtain the fol-
lowing inequality, showing that the n-dimensional volume under f is bounded
by the integration of the Jacobian:
Hn (f (A)) wn (Df(x) ) dHn (x).
x∈A
It follows that when A is a critical set for f (for all x ∈ A, Df(x) = 0),
then the n-volume of f (A) is null.
However, when n > m the situation is much more complicated; e.g. the
function f of Whitney (see [Whi 1]) is C n−1 smooth on B1n , but f (A) = [0; 1],
although A is a connected set of critical points of f (see also the references
of [Bat] and [Nor] for examples of functions with dense critical values set).
In the case n > m, we have the so-called coarea formula:
Hn−m (A ∩ f −1 ({y})) dHm (y) = wm (Df(x) ) dHn (x),
y∈M x∈A
which says that, for A a subset of critical points, and Hm -almost all critical
values y of f , the fiber A ∩ f −1 ({y}) is Hn−m -null.
Our first result in this chapter shows that for polynomial mappings of
semialgebraic sets of different dimensions, but of fixed algebraic complexity,
the formula of integration with the Jacobian is still valid in some sense, for
variations.
Theorem 7.1. Let f : Rn → Rm be a polynomial mapping, f = (f1 , . . . , fm ),
deg(fj ) = d¯j , for j ∈ {1, . . . , m}. Let A ⊂ Rn be a semialgebraic subset.
Assume that at each x ∈ A and for any i-dimensional affine space P̄ ⊂ Rn
tangent1 to A at x, wi (Df(x) |P̄ ) γ. Then:
with the constant C depending only on d¯1 , . . . , d¯m , m, n and D(A), the dia-
gram of A.
Remark. We do not know whether or not the constant C in the above
inequality really depends on the degrees d¯1 , . . . , d¯m .
Proof. By definition,
Vi (f (A)) = c(m, i) V0 (f (A) ∩ P̄x ) dx dP.
P ∈Gim x∈P
Obviously we have: V0 (f (A) ∩ P̄x ) V0 (A ∩ f −1 (P̄x )). Let us estimate the
integer V0 (A ∩ f −1 (P̄x )) with Corollary 4.9.
There exist L1 , . . . , Li : Rm → R, i affine linear forms in general position
such that P̄x = L−1 −1
1 ({0}) ∩ . . . ∩ Li ({0}). The algebraic set f
−1
(Px )) is
i
thus given by Bj , where Bj = {x ∈ Rn ; (Lj ◦ f )(x) = 0}. The degree
j=1
of Lj ◦ f is less than µ = maxj∈{1,...,i} d¯j . Now A ∩ f −1 (P̄x )) is given by
p jk i
( Akj Bj ), where Akj = {x ∈ Rn ; pkj (x) ? 0} and ? is one of the
k=1 j=1 j=1
symbols: >, ≥, and deg(pkj ) = d¯kj . By Corollary 4.9, we obtain:
1 ¯
p jk
V0 (A∩f −1 (P̄x )) C = (dk +iµ+2)(d¯k +iµ+1)n−1 , with d¯k = d¯kj .
2 j=1
k=1
We thus have:
Vi (f (A)) c(m, i).C V oli ([πP ◦ f ](A)) dP, (∗)
P ∈Gim
1
The affine space P̄ ⊂ Rn is tangent to A at x ∈ A if every direction ν of P̄ is a
tangential direction of A, that is to say there exists a smooth curve α in A with
α(0) = x and α (0) = ν ∈ P̄ .
7 Behaviour of Variations under Polynomial Mappings 85
V oli ([πP ◦ f ](A)) γV oli (C) = γVi (C) γ.C .Vi (A),
and finally
Vi (f (A)) c(m, i).C .C .γ.Vi (A) dP = C.γ.Vi (A).
P ∈Gim
The proof above is based on Theorem 4.10, which gives no explicit bound
on the complexity of the covering set C. Below we prove a little bit weaker
result, which however uses Exercise 4.11 instead of Theorem 4.10, and hence
gives an explicit bound for the coefficient in the inequality. We also use a
stronger, but easier to verify, condition wi (Df(x) ) γ, instead of requiring
this inequality only on i-dimensional subspaces tangent to A.
Theorem 7.2. Let f : Rn → Rm be a polynomial mapping, f = (f1 , . . . , fm ),
deg(fj ) = d¯j , for j ∈ {1, . . . , m}. Let A ⊂ Rn be a bounded semialgebraic
subset. Assume that at each x ∈ A, wi (Df(x) ) γ. Then:
Vi (f (A)) C.γ.V i (A),
where
p
= 1 c(m, i).
C (dk + iµ + 2)(dk + iµ + 1)n−1 κ,
4
k=1
jk
µ = maxj∈{1,...,i} d¯j , dk = dkj ,
j=1
86 7 Behaviour of Variations under Polynomial Mappings
p jk
and κ only depends on dk , d¯j , n, i, assuming that A is given by ( Akj ),
k=1 j=1
where Akj = {x ∈ Rn ; pkj (x) ? 0} and ? is one of the symbols: >, ≥, and
deg(pk j) = dkj .
Proof. We proceed exactly as in the proof of theorem 7.1. We obtain the
inequality (∗):
Vi (f (A)) c(m, i).C V oli ([πP̄ ◦ f ](A)) dP, (∗)
P ∈Gim
1
p
where C = (dk + iµ + 2)(dk + iµ + 1)n−1 , µ = maxj∈{1,...,i} d¯j and dk =
2
k=1
jk
dkj . To estimate V oli (πP (f (A))) we use Exercise 4.11, applied to the
j=1
polynomial πP ◦ f : Rn → P ∈ Gim , instead of Theorem 4.10. It provides
for any ξ > 0 a semialgebraic set C, such that C is contained in the ξ-
neighborhood of A, dH (πP (f (C); πP (f (A)))) Kξ, for all P ∈ Gim , where K
is a Lipschitz constant of f on a bounded subset of Rn containing A, and
dim(C) i. We obtain by Theorem 5.9:
And since we can assume that wi (πP ◦ f )|C 0 γ + η(ξ), with η(ξ) arbitrarily
small as ξ → 0, we obtain by the area formula:
V oli (πP (f (A))) = lim V oli (πP (f (C))) γ.V oli (C) = γ.Vi (C).
ξ→0
Remark. In the rest of this chapter we prove the results requiring “covering”
subsets C, using Theorem 4.10 to simplify the arguments. However, in any
of these results, Theorem 4.11 could be used exactly as in the proof above.
This allows us to obtain the explicit expressions for the coefficients below.
Definition 7.3. Let Λ = (λ1 , . . . , λq ), with λ1 ≥ λ2 ≥ . . . ≥ λq ≥ 0, be
given.
7 Behaviour of Variations under Polynomial Mappings 87
n
Vi (∆(f, Λ, B(a,r) )) C.λ0 λ1 . . . λi .ri ,
n
Vi (∆(f, Λ, B(a,r) )) C .λ0 λ1 . . . λi .ri .
q
1
M (, ∆(f, Λ, A)) C λ0 . . . λi · Vi (A) · ( )i ,
i=0
q
r
n
M (, ∆(f, Λ, B(a,r) )) C λ0 . . . λi · ( )i .
i=0
88 7 Behaviour of Variations under Polynomial Mappings
Now let us consider the set ∆0f = {x ∈ Rn ; Df(x) = 0}. Clearly ∆0f =
n
∆(f, Λ, B(a,r) ), for Λ = (0, . . . , 0). Hence we get:
Corollary 7.6. For f : Rn → Rm a polynomial mapping, with f =
(f1 , . . . , fm ), deg(fj ) = dj .
M (, ∆0f ) C,
where C = C(n, d1 , . . . , dm ) depends only on d1 , . . . , dm , n and m.
Clearly this result recovers the simple fact that the critical values of rank
0 of a polynomial mapping form a finite set with a number of points bounded
by the degrees (once we know that one can stratify a semialgebraic set with
a number of connected strata depending only on the degrees, this fact is
obvious).
Let us consider now ∆f = ∆(f, Λ, B(a,r)n
), for Λ = ( , . . . , ). We can
r r
now prove an extended version of Theorem 1.8:
Theorem 1.8. (extended) Let f : Rn → Rm be a polynomial map-
ping, with f = (f1 , . . . , fm ), deg(fj ) = dj . Then for any > 0 the set
∆0f can be covered by N (n, m, d1 , . . . , dm ) balls of radius . The constant
N (n, m, d1 , . . . , dm ) depends only on d1 , . . . , dm , n and m.
In particular the set ∆0f of rank-0 critical values of f|Brn contains at most
N (n, m, d1 , . . . , dm ) points, for any r > 0, hence the number of critical values
of f is bounded by N (n, m, d1 , . . . , dm ).
Proof. The proof follows directly from Theorem 7.5.
Remarks.
• We notice that, as a consequence of Theorem 7.5 and its smooth counter-
part, some “rigidity” results for smooth functions can be obtained: if the
function has inside a given ball “too many” critical points, with the values
at these points “too separated”, its high order derivatives must be “big”.
We do not touch these questions here (see [Yom 1], Theorem 3.9, Corollary
3.10 and Theorem 3.11.)
• Obviously Theorem 7.1, Theorem 7.2, Corollary 7.4, Theorem 7.5, Corol-
lary 7.6 and Theorem 1.8 are true for f : Rn → Rm a mapping definable in
some o-minimal structure (i.e. a mapping whose graph is a definable set in
some o-minimal structure). The constants that appear in these statements
depending only on the set A and the mapping f (the notions of degree and
diagram making no more sense in o-minimal structures).
Theorem 7.2 gives an upper bound for the variations of the images of
the polynomial mapping, in terms of the uniform upper bounds for the first
derivatives of the mapping.
7 Behaviour of Variations under Polynomial Mappings 89
The following example shows that the uniform lower bounds for the
derivatives do not allow one to bound variations of the image from below.
Let A ⊂ R2 be the rectangle [0; δ] × [0; 1] and f = π1 : (x, y) → x (see
Fig. 7.1).
Fig. 7.1.
· Vs−i (A) · 1
=C ·C Vi (f (A)) = C · Vi (f (A)) · Vs−i (A)
c(m, i)
1 C C
Vj (A) Vj (Cj ) Vi (Cj )Vj−i (Cj ) Vi (A)Vj−i (A).
λ λ λ
(The last inequality being a consequence of Theorem 5.1.)
In fact, Theorem 7.7 can be formulated in a stronger form, if we introduce
the notion of variations of functions and mappings, following more or less
[Iva 1].
Definition 7.8. Let A ⊂ Rn be a compact subset, and let f : A → Rm be
a continuous mapping. For i = 0, 1, . . . , n and = 0, 1, . . . , m, the variation
Vi, (f ) is defined as follows:
Vi, (f ) = c(i, , n, m) Vi (f −1 (P̄ )) dP̄ ,
P̄ ∈Ḡm−
m
There is an important special case, where the equality holds in (3) and (5):
V0, (f ) = V (f (A)),
Vi, (f ) =
c(i, i, , n, m) V (f (A ∩ P̄ )) dP̄ .
P̄ ∈Ḡn−i
n
c(0, , n, m) q.V0 (f (A) ∩ P̄ ) dP̄ = qV (f (A)).
P̄ ∈Ḡm−
m
There is one special situation, where the variations are reduced to the
usual integral-geometric invariant and hence can be computed in “closed
form”. Namely, for i + = n and f : Rn → Rm smooth and semialgebraic,
we have:
Vi, (f ) = c(i, , n, m) Vi (f −1 (P̄ )) dP̄
P̄ ∈Ḡm−
m
= c(i, , n, m) V oli (f −1 (P̄ )) dP̄ ,
P̄ ∈Ḡm−
m
7 Behaviour of Variations under Polynomial Mappings 93
where B1n is as usual the unit ball centered at the origin of Rn , and Vi is the
volume of the unit i-dimensional ball.
Remark. If rank(L) = i, dim(L(B1n )) = i, and the Cauchy-Crofton formula
for the volume gives:
1
ωi (L) = c(m, i) V oli πP (L(B1n )) dP
Vi .c(m, i) P ∈Gim
we have:
ω(P ) dP = ωi (L).ωRn .
P ∈Gim
[(πP ◦ L)∗ ωP ] = λ1 (πP ◦ L) . . . λi (πP ◦ L)(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) .
Of course we suppose that rank(L) ≥ i, and thus that for generic P ∈ Gim ,
rank(πP ◦ L) = i (if it is not the case, each side of the equality we want to
prove is 0).
Now λ1 (πP ◦ L) . . . λi (πP ◦ L).V oli (B1i ) = V oli ([πP ◦ L](B1n )), and
1
ω(P )dP = [(πker(πP ◦L) )∗ ωker(πP ◦L) ] ∧ [(πker⊥ (πP ◦L) )∗ ωker⊥ (πP ◦L) ]×
P ∈Gmi V i
94 7 Behaviour of Variations under Polynomial Mappings
V oli ([πP ◦ L](B1n ))dP
P ∈Gim
= ωi (L).ωRn .
Proof. By definition
Vs−i,i (f ) = c(s − i, i, n, m) Vs−i (f −1 (P̄ )) dP̄
P̄ ∈Ḡm−i
m
= c(s − i, i, n, m) Vs−i ([πP ◦ f ]−1 (ξ)) dξ dP.
P ∈Gim ξ∈P
For all P ∈ almost all the points ξ ∈ P are regular values of (πP ◦ f )|A0 ,
Gim ,
by the classical Sard theorem, hence [πP ◦ f ]−1 (ξ) = YP,ξ is empty or is a
smooth submanifold in A0 of dimension s − i. We obtain:
Vs−i,i (f ) = c(s − i, i, n, m) V ols−i (YP,ξ ) dξ dP
P ∈Gim ξ∈P
= c(s − i, i, n, m) ωYP,ξ dξ dP = c(s − i, i, n, m)
P ∈Gim ξ∈P YP,ξ
[(πker(πP ◦Df|A0 (x) ) )∗ ωker(πP ◦Df|A0 (x) ) ] ∧ [(πP ◦ f|A0 )∗ ωP ] dP
P ∈Gim x∈A0
= c(s − i, i, n, m) [ ωP (x)] dP
P ∈Gim x∈A0
= c(s − i, i, n, m) [ ωP (x) dP ]
x∈A0 P ∈Gim
7 Behaviour of Variations under Polynomial Mappings 95
= c(s − i, i, n, m) ωi (Df|A0 (x) ) ωA0
x∈A0
= c(s − i, i, n, m) ωi (Df|A0 (x) ) dHs (x).
x∈A
Now we turn back to semialgebraic sets and their polynomial mappings.
The following theorem shows that, as above, variations of semialgebraic sets
and mappings, are equivalent, up to coefficients, to their “integral-geometric”
invariants - average volume of projections.
Theorem 7.13. Let A ⊂ Rn be a bounded semialgebraic set, f = (f1 , . . . ,
fm ) : A ⊂ Rn → Rm be a polynomial mapping, with deg(fj ) = dj . Then:
(1) For i = 0, . . . , n; = 0, . . . , m: Vi, (f ) C1 .Vi (A).V (f (A)).
(2) For A ⊂ Brn1 , f (A) ⊂ Brm2 : Vi, (f ) C2 .r1i .r2 .
(3) Vi,m (f ) C3 .Vm (f (A)).Vi (A).
(4) V0, (f ) C4 .V (f (A)).
(5) Vi, (f ) C5 . V (f (A ∩ P )) dP .
P ∈Gn−
n
Vi (f −1 (P̄ )) B0,n−i (f −1 (P̄ )).Vi (A) B0,n−i (f −1 (P̄ )).Vi (A).V0 (f (A) ∩ P̄ ).
Now, we have shown in the proof of Theorem 7.1 that there exists C , a
constant depending only on D(A), such that: V0 (f −1 (P̄ )) C . It follows
that:
Vi, (f ) c(i, , n, m).C .Vi (A) V0 (f (A) ∩ P̄ ) dP̄
P̄ ∈Ḡm−
m
Remark. Theorem 7.12.(1), Theorem 7.11 and Lemma 7.8, or directly Corol-
lary 7.14 give again Theorem 7.7, for a semialgebraic set of dimension s:
C
Vi (f (A)) ≥ wi (Df|A0 (x) ) dHs (x).
Vs−i (A) x∈A
From Theorem 7.2 and corollary 7.14 (or Theorem 7.7), we have, for A
a semialgebraic set of Rn of dimension s and f : Rn → Rm a polynomial
mapping such that wi (Df(x) ) γ on A:
C
. wi (Df|A0 (x) ) dHs (x) Vi (f (A)) C .γ.Vi (A). (∗∗)
σi .Vs−i (A) x∈A
Fig. 7.2.
But by assumption:
Vs−i (f −1 (P̄ )) dP̄ ≥ K χ̄[f (A)] (P ) dP̄ ,
P̄ ∈Ḡm−i
m P̄ ∈Ḡm−i
m
where χ̄[f (A)] (P ) = 1 when P ∩f (A)
= ∅ and χ̄[f (A)] (P ) = 0 when P ∩f (A) =
∅. We also have:
Vi (f (A)) = c(m, i) V0 (f (A) ∩ P̄ ) dP̄
P̄ ∈Ḡm−i
m
c(m, i).B0,m−i (f (A)) χ̄[f (A)] (P ) dP̄ .
P̄ ∈Ḡm−i
m
Finally, we get:
c(m, i).B0,m−i (f (A))
Vi (f (A)) wi (Df|A0 (x) ) dHs (x).
C1 .c(s − i, i, n, m).K x∈A
Under similar assumptions we can give also the lower bounds for the vari-
ations of the image in terms of the uniform lower bound of wi (Df|A0 (x) ), as a
consequence of the inequality in the remark (∗∗) that follows Corollary 7.14.
Theorem 7.16. Let A ⊂ Rn be a semialgebraic set of dimension s,
f = (f1 , . . . , fm ) : Rn → Rm be a polynomial mapping, with deg(fj ) = dj ,
satisfying wi (Df|A0 (x) ) ≥ Γ (i s), then:
C.Γ
V ols (A) Vi (f (A)) C .γ.Vi (A).
Vs−i (A)
98 7 Behaviour of Variations under Polynomial Mappings
V ols (A)
Remark. The assumption ≥ K means that the sections of A in
Vi (A)
the directions orthogonal to its maximal i-dimensional section, are at least
K-big.
K has the physical dimension of Vs−i (A), but of course, if Vs−i (A) is
big, K may be small: A = [0; δ] × [0; 1], s = 2, i = s − i = 1, V1 (A) ≥ 1,
V ol2 (A) = δ.
From Theorem 7.16 we obtain:
Corollary 7.17. Let A ⊂ Brn ⊂ Rn be a semialgebraic set of dimension s,
f = (f1 , . . . , fm ) : Rn → Rm be a polynomial mapping, with deg(fj ) = dj ,
satisfying wi (Df|A0 (x) ) ≥ Γ > 0 (i s), then for any ball Bδm ⊂ Rm :
c i s−i
V ols (f −1 (Bδm ) ∩ A) δ .r ,
Γ
where c is a constant depending only on D(A), m, n, d1 , . . . , dm .
Proof. From Theorem 7.16 we have
σi i (f(A)),
V ols (A) Vs−i (A).V
C.Γ
r
8 Quantitative Transversality
and Cuspidal Values
Σ(f, Λ, δ) =
{(x, t) ∈ Brn × Brm ; λi (Dx f(x,t) ) λi , ∀i = 1, . . . , m and ||f (x, t)|| δ}.
We denote also by ∆(f, Λ, δ) ⊂ Brn the set π2 (Σ(f, Λ, δ)), where the
projection π2 : Rn × Rm → Rm is the standard projection.
Thus ∆(f, Λ, δ) consists of those parameters t ∈ Brm for which there exists
x ∈ Brn with λi (Dx f(x,t) ) λi , i = 1, . . . , m, and ft (x) ∈ Bδm ⊂ Rm .
Theorem 8.1. With the notations above, we have, for s = 0, 1, . . . , m:
c.M s
Vs (∆(f, Λ, δ)) λ0 . . . λj rj δ s−j ,
ρm . . . ρm−s+1 j=0
s
−1
ws (π2 |Tx C 0 ) k.ws ([Dt f(x,t) ] ) wj (D(f|C 0 )(x,t) ).ws−j (Dx f(x,t) )
j=0
k s
wj (D(f|C 0 )(x,t) ).λ0 . . . λs−j ,
ρm . . . ρm−s+1 j=0
k s
λ0 . . . λs−j .C3 .M.Vs−j (C).Vj (f (C)),
ρm . . . ρm−s+1 j=0
k.C3 .M.c.c
s
V ols ((πP ◦ π2 )(Σ(f, Λ, δ))) λ0 . . . λs−j rs−j δ j .
ρm . . . ρm−s+1 j=0
As we can expect, the bound of Theorem 8.1 shows that for λi and δ
small, variations of the set of “bad” parameters are small.
We can consider now some special cases of this theorem.
Corollary 8.2. With the same notations as above,
c.M
Vs (∆(f, Λ, 0)) λ0 . . . λs rs .
ρm . . . ρm−s+1
This formula is very similar to the bound given in Corollary 7.4. In fact
this Corollary is a special case of Theorem 8.1. Indeed, for a given polynomial
mapping f : Rn → Rm , let us consider the mapping F : Brn × Brm → Rm
defined by F (x, t) = f (x) + t. We have ρ1 = . . . = ρm = 1, and from Theorem
8.1 we obtain:
Vs (∆(F, Λ, 0)) c.M.λ0 . . . λs rs .
But ∆(F, Λ, 0) is the set of values f (x) = t such that x ∈ Brn , λi (Df(x) ) λi ,
for i = 0, . . . , m, and thus is ∆(f, Λ, Brn ), in the notation of Corollary 7.4.
Let us return to the general case. If t ∈ / ∆(f, Λ, δ), then for any x ∈ Brn
such that ||f (x, t)|| δ, at least one of the quantities λi (Dx f(x,t) ) is greater
than λi .
If we want to guarantee that for each i, λi (Dx f(x,t) ) ≥ λi , we can proceed
as follows:
Assume that everywhere on Brn × Brm , we have:
c.M i−1 s
Vs (∆(f, Λi , δ)) [ K j rj δ s−j + K i−1 λij−i+1 rj δ s−j ].
ρm . . . ρm−s+1 j=0 j=i
¯ Λ, δ),
Adding all these expressions, for i from 1 to m, we get a bound for ∆(f,
as follows:
¯ Λ, δ)) C 1 · (λ1 + . . . + λs ) + C 2 · δ.
Vs (∆(f, s s
¯ Λ, δ)
Thus for λ1 , . . . , λm and δ tending to 0 all the variations of the set ∆(f,
tend to 0. Consequently, for most of values of the parameters t, if ||f (x, t)|| δ,
then for each i, λi (Dx f(x,t) ) ≥ λi .
Now we consider another generalization of Theorem 8.1. Let A ⊂ Rm and
define Σ(f, Λ, A, δ) as the set:
Σ(f, Λ, A, δ) =
c.M.M (δ, A)
s
Vs ∆(f, Λ, A, δ) λ0 . . . λj rj .δ s−j ,
ρm . . . ρm−s+1 j=0
m
1 1 s
s
.M (δ, A)
C.M ( ) λ0 . . . λj rj .δ s−j ,
s=0
ρm . . . ρm−s+1 j=0
bounds wj (Df ) on B n ×
where C depends only on n, m, d1 , . . . , dm , and M r
m
Br for j = 0, . . . , m.
Substituting here = δ, we get:
Corollary 8.5. With the notations above:
m
r
M , ∆(f, Λ, A, ) C · C · M · M (, A) λ0 . . . λj ( )j ,
j=0
where K bounds from above σs−j (Dg(x)|T ), for all x ∈ Σ(f, g, ΛΛ , δ), all
T ∈ Gsn and all j ∈ {0, . . . , s} (σs−j being defined in Theorem 7.7.)
Proof. As in the proof of Theorem 8.1, it suffices to bound the s-volume of
the projections πP (∆(f, g, ΛΛ , δ)) ⊂ P , where P is a s-dimensional linear
subspace of Rm . Let P be a fixed plane in Gsm . By Theorem 4.10 there exists
a semialgebraic set C ⊂ Σ(f, g, ΛΛ , δ), such that D(C) depends only on
D(Σ(f, g, ΛΛ , δ)),
and
dim(C) = dim(πP (∆(f, g, ΛΛ , δ))) = s
2
The convention here is that for j>q, λj = 0, for j>min(n-q,m) , λj = 0, and for
j> q , ρj = 1.
8 Quantitative Transversality and Cuspidal Values 105
(we consider only variations Vs (∆(f, g, ΛΛ , δ)) for s dim(∆(f, g, ΛΛ , δ)).)
By the area formula, we have:
V ols (πP (∆(f, g, ΛΛ , δ))) ws ([πP ◦ Df(x) ]|Tx C 0 ) dHs (x)
x∈C 0
ws (Df(x)|Tx C 0 ) dHs (x).
x∈C 0
s
ws (Df(x)|Tx C 0 ) k(s, m, s) wj (L1 ).ws−j (L2 ).
j=0
Now by assumption, wj (L1 ) λ0 λ1 . . . λj , for all j ∈ {0, . . . , min(n − q, m)},
ws−j (Dg(x) )
and ws−j (L2 ) λ0 λ1 . . . λs−j λ0 λ1 . . . λs−j , for all j ∈ {0, . . . ,
ρ0 ρ1 . . . ρs−j
q}. We thus have:
s
λ0 λ1 . . . λs−j
ws (Df(x)|Tx C 0 ) k(s, m, s) λ0 λ1 . . . λj ws−j (Dg(x) )
j=0
ρ0 ρ1 . . . ρs−j
V ols (πP (∆(f, g, ΛΛ , δ))) ws (Df(x)|Tx C 0 ) dHs (x)
x∈C 0
s
λ0 λ1 . . . λs−j
k(s, m, s) λ0 λ1 . . . λj ws−j (Dg(x) ) dHs (x)
j=0
ρ0 ρ1 . . . ρs−j x∈C 0
s
λ0 λ1 . . . λs−j
σs−j (g|C 0 ).C.k(s, m, s) ×
j=0
ρ0 ρ1 . . . ρs−j
First of all, we see that Corollary 7.4 is a special case of Theorem 8.7.
Indeed, for g : Brn → Brn being the identity mapping, δ = r, we have
ker(Dg(x) ) = 0, λ0 = λmin(n−n,m) = 1, ρ0 = . . . = ρn = 1,K = 1, and
∆(f, g, ΛΛ , r) = ∆(f, Λ, Brn ), hence the only term remains:
106 8 Quantitative Transversality and Cuspidal Values
n−q
1
Vs (T δ ) K.Ĉ rj .δ s−j .
j=s−q
ρ0 ρ1 . . . ρs−j
of type Σ 1,s (f ), for all s ∈ {1, . . . , n − 1}. For instance, for δ = 0 and
λn−s = . . . = λn−1 = 0, we obtain Σ 1,s .
As usual we denote by ∆1 (f, Λ, δ) the image f (Σ 1 (f, Λ, δ)).
Theorem 8.10. For s n − 1, we have (with the convention (See Footnote 2
on Page 104 ):
M
Vs (∆1 (f, Λ, δ)) Ĉ.K[ .λ0 . . . λs−1 .rs−1 .δ + λ0 . . . λs .rs ],
γ
Ĉ.K
and for s = n : Vn (∆1 (f, Λ, δ)) .λ0 . . . λn−1 .rn−1 .δ,
γ
where K bounds above σs−j (DJ(x)| T ), for all x ∈ Σ 1 (f, Λ, δ), all T ∈ Gsn ,
and all j ∈ {s − 1, s}, and where M is an upper bound for
λ1 (Df(x)|(ker(DJ(x) ))⊥ ),
i=q
r
M (, ∆(f, Λ, Brn )) c λ0 λ1 . . . λi ( )i ,
i=0
p−1 t=1
1 j (1 − t)p−1 p
ϕ(x) = f (x) − Pp (x) = D ϕ(0) xj + D ϕ(tx) .xp dt.
j=0
j! t=0 (p − 1)!
Thus:
t=1
(1 − t)p−1 p
||f (x) − Pp (x)|| = || D ϕ(tx) .xp dt||
t=0 (p − 1)!
t=1
(1 − t)p−1
||Dp ϕ(tx) ||.||x||p dt
t=0 (p − 1)!
t=1
(1 − t)p−1
= ||Dp f(tx) − Dp f(0) ||.||x||p dt
t=0 (p − 1)!
t=1
(1 − t)p−1
.K.||tx − 0||α .||x||p dt,
t=0 (p − 1)!
and finally, we obtain:
K k
||f (x) − Pp (x)|| r = Rk (f ) . (1)
p!
The same argument for Dϕ gives us:
K p p
||Df(x) − DPp(x) || rk−1 = Rk (f ) . (2)
(p − 1)! r r
p
Hence if we denote by λi the numbers λi + , for i ≥ 1 and λ0 = 1, and put
r
Λ be (λ1 , . . . , λq ), by inequality (2), we have Σ(f, Λ, Brn ) ⊂ Σ(Pp , Λ , Brn ).
Now ∆(f, Λ, Brn ) ⊂ f Σ(Pp , Λ , Brn ) ⊂ ∆(Pp , Λ , Brn ) , by (1).
Therefore we can write:
M (2, ∆(f, Λ, Brn )) M (2, ∆(Pp , Λ , Brn ) ) M (, ∆(Pp , Λ , Brn )),
or equivalently:
q
r
M (, ∆(f, Λ, Brn )) C λ0 . . . λi ( )i ,
i=0
i
p i
λ0 . . . λi = (λj + ) = c λi1 . . . λi ( )i− .
r r
j=1 =0 i1 >...>i ∈{1,...,i}
i
λ0 . . . λi
c λ1 . . . λ ( )i− .
r
=0
We conclude that:
q
i
r r
c
M (, ∆(f, Λ, Brn )) C. λ0 λ1 . . . λ ( )−i ( )i
i=0 =0
q
r
c λ0 . . . λi ( )i .
i=0
q
r
M (, ∆(f, Λ, Brn )) c λ0 λ1 . . . λi ( )i , for ≥ Rk (f ),
i=0
q n−i
r i Rk (f ) k
M (, ∆(f, Λ, Br ))
n
c λ0 λ1 . . . λi ( ) , for Rk (f ),
i=0
112 9 Mappings of Finite Smoothness
where c and
c depend only on n, m and k.
Proof. For ≥ Rk (f ) the result follows from Proposition 9.1.
Let Rk (f ). We cover Brn by balls of radius r < r, where r is chosen
K
in such a way that Rk (f|B n ) . We can take for r , (r )k = , or:
r p!
k1
r = r. .
Rk (f )
The number of such balls we need to cover Brn is at most:
n
r n Rk (f ) k
N = C( ) = C. .
r
n
We apply Proposition 9.1 to the restriction of f to each of the N balls B(x ,
j ,r )
hence we obtain:
N
M (, ∆(f, Λ, Brn )) M (, ∆(f|B n n
, Λ, B(x ))
j ,r )
(xj ,r )
j=1
n
Rk (f ) k
q
r
M (, ∆(f, Λ, Brn )) C. .c λ0 . . . λi ( )i
i=0
nk q
Rk (f ) r r
=
c λ0 . . . λi ( )i ( )i .
i=0
r
Thus:
q n−i
r i Rk (f ) k
M (, ∆(f, Λ, Brn ))
c λ0 . . . λi ( ) .
i=0
Let us stress the fact that for ≥ Rk (f ) the expression of Theorem 9.2 is
the same as in the polynomial case. So in a resolution coarser than the Taylor
remainder term one cannot distinguish between the geometry of the critical
values of f and of its Taylor polynomial approximation. In finer resolution
a correction appears, expressed in terms of the remainder Rk (f ). In fact,
Theorem 9.2, as well as most of the results of Chapter 9, can be considered
as a “generalized Taylor formula” for the property in question: they contain
a “polynomial term” and a correction, expressed through Rk .
Let us denote now:
and
∆νf = f (Σfν ),
respectively the rank-ν set of critical points and of critical values. The closed
ball Brn being compact, there exist λ1 ≥ . . . . . . λq ≥ 0 such that for all
9 Mappings of Finite Smoothness 113
x ∈ Brn and for all i ∈ {1, . . . , q}, λi (Df(x) ) λi . Therefore we have ∆νf ⊂
∆(f, Λ, Brn )), for Λ = (λ1 , . . . , λν , 0, . . . , 0), and thus by Theorem 9.2:
n−i
ν
r i Rk (f ) k
M (, ∆νf ) M (, ∆νf ⊂ ∆(f, Λ, Brn )))
c λ0 . . . λi ,
i=0
so
n−i
ν
r i K.rk k
M (, ∆νf )
c λ0 . . . λi
i=0
p!
ν
K n−i 1 i+ n−i
=
c λ0 . . . λi k
.rn . k
.
i=0
p!
n−ν
Finally we have bounded M (, ∆νf ) by a polynomial of degree ν + in
k
1 n−ν
, hence by definition the entropy dimension of ∆νf is less than ν + ,
k
and so is the Hausdorff dimension of ∆f (by the inequality dimH dime of
ν
Comments.
• First of all, let us stress that the bound on the entropy dimension may be
much more restrictive than that for the Hausdorff dimension. Many examples
of this sort are given in Chapter 2. In particular, for ν = 0, i.e. for critical
values of rank 0, by Theorem 9.3 their entropy dimension is at most n/k.
This implies for instance that the sequences {1, 1/2β , . . . , 1/nβ , . . .} cannot be
k
contained among these critical values for β < + 1. Of course, the Hausdorff
n
dimension of any countable set is zero.
• Theorem 3.4.3 of [Fed 2] gives, for f a k-times continuously differen-
n−ν
tiable mapping (k ∈ N \ {0}), Hν+ k (∆νf ) = 0. In particular, it implies
n−ν
dimH (∆νf ) ν + . Theorem 9.3 shows that the assumption k-times con-
k
tinuously differentiable can be weakened to C k−1+1 , that is to say, f is (k−1)-
times differentable and Dk−1 f is Lipschitz.
• In terms of entropy (as well as in terms of Hausdorff) dimension, The-
orem 9.3 cannot be sharpened: we can construct a family of functions
2 ln 3
fβ : [0; 1]2 → [0; 1], with β ∈]2; ∞[, such that fβ is C ln β -smooth, ∆0fβ = I 13 ,
114 9 Mappings of Finite Smoothness
Of course, by [Com 1], equality (∗) would be the best quantitative result in
terms of Hausdorff measure.
• Equality (∗) has been first proved in two particular cases:
n−ν
(i) For k = (n−ν)/(m−ν), that is to say: if f : Rn → Rm is a C m−ν -smooth
mapping, then Hm (∆νf ) = 0 ([Bat 1]).
(ii) Forn ν = 0, that is to say: if f : Rn → Rm is a C k -smooth mapping, than
H k (∆0f ) = 0 ([Com 1]).
One can prove the classical Morse-Sard theorem (concerning the Hm -nullity
of ∆νf , see [Sar 1,2], [Bat 1], [Nor 1]) by induction on ν, starting from ν = 0,
and using Fubini’s theorem. Unfortunately Fubini’s theorem does not hold
anymore for non-integral Hausdorff measures (see [Fal]), thus one cannot
prove (∗) from (ii) as it is done in the classical case. One can only prove in
this way that for any point a ∈ ∆νf there exists an open subset U of Rm
n−ν
containing a, such that U ∩ ∆νf is Hν ⊗ H k -null, where Hν measures Rν
n−ν
and H k measures Rm−ν (in some suitable C k -coordinate system for U ),
n−ν n−ν
but in general, as mentioned above, Hν ⊗ H k
= Hν+ k (on this question
see [Fal], [Fed 2], [Nor 2] or for a counterexample see [Sar 3]).
• However if f : Rn → Rm is a C (p+α)+ -smooth mapping (it means that
there exists for every a ∈ Rn , a ball B(a,r)
n
, and a function a : R → R
such that: a (t) −→ 0 and ||Dp f(x) − Dp f(y) || a (||x − y||)||x − y||α , for all
t→0
n−ν
x, y ∈ B(a,r)
n
), the proof of theorem 3.4.3 of [Fed 2] gives (∗): Hν+ p+α (∆νf ) =
0. Finally, let us note that the proof of theorem 3.4.3 of [Fed 2] gives also:
n−ν
dimH (∆νf ) ν + , because if f : Rn → Rm is a C k -smooth mapping,
k
it is also a C + -smooth mapping, for any 1 < k, and thus one has by
n−ν
[Fed 2], Theorem 3.4.3: Hν+ (∆νf ) = 0, for every 1 < k, showing that
n−ν
dimH (∆νf ) ν + .
k
• Finally it has been proved in [Mor] by Carlos Gustavo T. de A. Moreira
that C k -regularity for f implies equality (∗). The difficulty mentioned above
concerning the use of Fubini’s theorem has been overcome by a “careful de-
composition of the critical set, combined with a parametrized strong version
of the so-called A. P. Morse lemma.
Let us give Moreira’s statement:
116 9 Mappings of Finite Smoothness
δ
M (, f (Brn ) ∩ Bδm ) = M (, Bδm ) = C.( )m
R1 (f ) n δ n Rk (f ) nk
min , ,
Thus: n
Rk (f ) k(m−n)
n R (f ) m
1
δ c.min , .
K k K
||f (x) − Pp (x)|| r = rk . = , for any x ∈ Brn .
p! p! Rk (f )
Hence:
f (Brn ) ∩ Bδm ⊂ Pp (Brn ) ∩ Bδn .
We thus get:
M (2, f (Brn ) ∩ Bδm ) M (2, Pp (Brn ) ∩ Bδn ) M (2, Pp (Rn ) ∩ Bδn )
118 9 Mappings of Finite Smoothness
+ δ n δ n
M (2, Pp (Rn ) ∩ Bδ+
n
) M (, Pp (Rn ) ∩ Bδ+
n
)
c. 2n
c. ,
since Pp (Rn ) is a semialgebraic set of fixed complexity and of dimension
n(< m) in B+δ m
, and since we assume here that δ. (see Corollary 5.7).
r n
Finally, the number of balls Brn one needs to cover Brn is C. that is
n r
to say C Rk(f ) k , completing the proof.
Remark. Notice that the latter quantity in Theorem 9.5 is smaller than the
δ k
first one only if ≥ Rk (f ) . Since the bounds for the -entropy of the
R1 (f )
subsets of Bδ are meaningful only if < δ, we see that, at least as δ −→ 0, for
m
δ k
any k > 1, we have the range of values of , namely δ ≥ ≥ Rk (f ) ,
R1 (f )
where our bounds for the -entropy of f (Brn ) ∩ Bδm are strictly better than
for C 1 -mappings.
The bound of Theorem 9.5 is “almost sharp”. Below we give correspond-
ing examples. Let us start with a mapping, given by an explicit analytic
formula (but providing roughly twice lower smoothness than in the optimal
k
construction.) Consider the following C 2 -smooth mapping g : [0; 1]n → R2n ,
where k > 2:
1 1
g(t1 , . . . , tn ) = (tk1 , . . . , tkn , tk1 sin( ), . . . , tkn sin( )).
t1 tn
We have:
√ 1 n
2( n) k −1 n δ n
n 1 k
V oln (g([0; 1] ) ∩
n
Bδ2n ) ≥ δ . + √ . (∗∗)
2k π(k − 1) δ n
1
Indeed, on one hand for (t1 , . . . , tn ) ∈ [0; δ k ]n , g(t1 , . . . , tn ) ⊂ [0; δ]2n ⊂
1
Bδ√n , thus V oln (g([0; 1]n )∩Bδ2n
2n √ ) ≥ V oln (g([0; δ k ]n ), and on the other hand
n
V oln (g([0; δ k ]n ) = V ol1 (γ) , because g([0; δ k ]n ) = γ n ⊂ (R2 )n , where
1 n 1
1 1
γ = {(tk , tk sin( )) ∈ R2 ; t ∈ [0; δ k ]}. Now let us show that V ol1 (γ) ≥
t
2 1 k1
δ. + δ n . We have:
2 π(k − 1) δ
k
k
1 1 1 1
V ol1 ({(tk , tk sin( )); t ∈ [ ; ]}) ≥ .
t π(p + 1) πp π(p + 12 )
1
where p0 is the smallest integer satisfying: p0 ≥ 1 . We obtain:
πδ k
+∞ k1 −1
1 dt 2 1
V ol1 (γ) ≥ +δ ≥ k +δ
πk 1
+1 (t + 12 )k 2 (k − 1)π δ
πδ 1/k
which proves (∗∗). Consequently we have the following lower bound, for → 0:
1
M (, g([0; 1]n ) ∩ Bδ2n ) ≥ c.( )n V oln (g([0; 1]n ) ∩ Bδ2n )
√ 1 −1 n
2( n) k δ n 1 nk δ n
≥ c. k . + c. √ ,
2 π(k − 1) δ n
where c depends only on n. More accurate analysis of the geometry of
g([0; 1]n ) ∩ Bδ2n allows one to show that this lower bound becomes valid
starting with ∼ δ 1+1/k and smaller. The upper bound given by Theorem
δ n 1 nk
9.5 being c . (g is a C k/2 -smooth mapping), we obtain:
2
√ 1 −1 n
2( n) k δ n 1 nk δ n
c. k . + c. √ M (, g([0; 1]n ) ∩ Bδ2n )
2 π(k − 1) δ n
δ n 1 nk
c . .
2
Hence the ratio of the lower and the upper bound is:
√ 1 −1 n 2
c 2( n) k nk c n
1 ≥ . k + √ n (2 ) k
c 2 π(k − 1) δ c ( n)
√ 1 −1 n
c 2( n) k 1 nk + 2n2
≥ . k k ,
c 2 π(k − 1) δ
1
for = δ 1+ k .
Additional results in this direction are given below.
Comparing the two expressions in Theorem 9.4, we see that the first one
(the one which gives the best bound for highly differentiable functions) is
better than the second (asymptotically, as −→ 0), if:
n n m
< , or k > .
k(m − n) m m−n
Once more, this bound is virtually sharp.
Consider the mapping g : [0; 1]n → R2n , built above: g(t1 , . . . , tn ) =
(t1 , . . . , tkn , tk1 sin( t11 ), . . . , tkn sin( t1n )). Its image is contained in the part of
k
of the considered points satisfy: |xn+i | |xi |, for all i ∈ {1, . . . , n}. So let a
point x = (x1 , . . . , xn , xn+1 , . . . , x2n ) ∈ R2n be given, with |xn+i | |xi |. We
1
claim that the distance d(x, g([0; 1]n )) is at most ||x||1+ k . Indeed, we can find
ξi , such that:
1 1 1 1 xn+i
1 1 + 2π and sin( ) = ∈ [−1; 1].
k
xi ξi x k ξi xi
i
We thus have:
1 1 1 1
0 xi −ξik xi − 1 k
xi (1+2πxik )k −1 xi (2kπxik (1+2πxik )k−1 )
( 1 + 2π)
xik
and finally:
1
1+ k
|xi − ξik | c.xi , (3)
where c depends only on k. Now, of course, we also have:
1 1 1 1+ 1
0 |xn+i − ξ k sin( )| = |xi sin( ) − ξ k sin( )| |xi − ξik | c.xi k . (4)
ξi ξi ξi
1
Inequalities (3) and (4) show that d(x, g([0; 1]n ) c.||x||1+ k , where
c depends
1
1+ k
only on k and n. Therefore, if we take = c.δ , the ball of radius δ centred
at the origin is contained in the -neighborhood of g([0; 1]n ). Thus the lower
1
bound for the maximal radius of a ball, contained in g([0; 1]n ) , is c .1− k+1 .
n
The upper bound, given by theorem 9.4 is c.1− (k/2)(2n−n) = c.1− k .
2
Examples in the same spirit can be given for any m and n, n < m, and
with an essentially maximal possible differentiability:
Let Ψ : B1n → Rn be a C ∞ -mapping with the following properties:
– Ψ|B1n \B1/2
n ≡ 0,
– B1 ⊂ Ψ (B1/4
n n
).
We assume that m > n and consider, for any s > 0 in Rm−n , the following
1
net Z s = {Zαs }: on each sphere S1/N
m−n−1
s , the points Zαs form an s+1 -net,
N
m−n−1
1/N s
N ∈ N \ {0}, and the number of Zαs on S1/N m−n−1
s is c. =
1/N s+1
c.N m−n−1 , where c only depends on n and m.
1 s
Now for each α and N , such that Zαs ∈ S1/Nm−n−1
s , let rα = ( ) k , where
N
1 1
k(= p + β) ∈ R. For any rα rα , the mapping: s Ψ ( x) is a C k -smooth
N rα
mapping, such that all its derivatives up to order p, as well as the Hölder
constant K of order β of Dp Ψ (see the very beginning of this chapter for the
definition of K) are bounded by a constant not depending on N , s, etc...
9 Mappings of Finite Smoothness 121
(but depending on the ratio rα /rα .), and such a mapping is zero off Brnα . It
follows that when the balls Brnα centred at certain points xα are contained in
B1n and disjoint, and all the ratios rα /rα are uniformly bounded with respect
1 1
to α, the mapping s
Ψ ( (x − xα )) is a C k -mapping.
α
N rα
Now let Φ : Brn → R be a C ∞ -function with the following properties:
– Φ|B1n \B1/2
n ≡ 0,
– Φ|B1/4
n ≡ 1.
If the sum rαn converges, we can find balls Brnα , contained in B1n ,
α
rα
disjoint, and with ≥ c, c not depending on α (see [Iva 1]). Let us take as
rα
n
xα the centers of Brα .
Thus if we define g : B1n → Rm as follows:
1 1 1
g(x) = s
Ψ ( (x − xα )), Zα .Φ( (x − xα )) ∈ Rn × Rm−n = Rm ,
s
α
N rα rα
s+1 1 1 s+1
where = δ s . Indeed, for δ = s , = s+1 = δ s . Now by construc-
N N
r
tion, the image under g of the ball of radius α , centred at xα , is the ball
4
(B n1s , Zαs ) ⊂ Rn × Rm−n = Rn . Thus Bδm is contained in the -neighborhood
N
1
of g(B1n ), for δ = 1− s+1 . It remains to determine the allowed values for s: the
∞ ∞
1 sn sn
condition rαn < ∞ means that N m−n−1 ( ) k = N m−n−1− k <
N
α N =1 N =1
k(m − n)
∞, ie s > .
n n
Finally the lower bound for δ = δ() in our example, is 1− n+k(m−n) , while
n
Theorem 9.4 gives an upper bound which is 1− k(m−n) . In particular, for
k −→ ∞, this bound is asymptotically sharp.
Estimation of the -entropy in this last example allows one to show es-
sential sharpness (as k → ∞) also of the bounds in Theorem 9.5.
Exactly the same arguments as in Theorem 9.5 can be applied in more
general situations. For instance:
Theorem 9.6. Let f : Brn → Rm be a C k -smooth mapping, q = min(n, m),
Λ = (λ1 , . . . , λq ), Bδm be some ball of radius δ in Rm , and δ. Then we
have:
M (, ∆(f, Λ, Brn ) ∩ Bδm )
122 9 Mappings of Finite Smoothness
nk
q
Rk (f ) r i δ
c. min λ0 . . . λi ( )i ( ) k , ( )i .
i=0
Rk (f )
Proof. The main argument in the proof of Theorem 9.5, consists in the ap-
k1
proximation of f by its Taylor polynomial on balls of radius r = r .
Rk (f )
n
As applied to ∆(f, Λ, Br ), it becomes the same as the argument in the
proof of Theorem 9.2, thus we obtain the same bound: M (, ∆(f, Λ, Brn ) ∩
n q
Rk (f ) k r i
Bδm ) c. λ0 . . . λi ( )i ( )k .
i=0
R k (f )
n q
Rk (f ) k δ i
Let us now prove that: M (, ∆(f, Λ, Br ) ∩ Bδ ) c .
n m
( ).
i=0
R (f ) n
As in the proof of Theorem 9.5, we cover the ball Brn by C (
k
) k balls of
k1
radius r = r . Then approximating f on each of these small balls,
Rk (f )
by Pp , its Taylor polynomial of degree p, we obtain (cf. the proof of Theorem
9.5):
m
Vi (Pp (Rn ) ∩ B m )
2δ
Now by Theorem 3.5: M (, Pp (Rn ) ∩ B2δ
m
) C(m) i
. But
i=0
Pp (Rn ) ∩ B2δ
m
is a semialgebraic set of B2δm
of dimension less than q, and of
fixed complexity, hence by Corollary 5.2, we obtain: Vi (Pp (Rn ) ∩ B2δ
m
) C.δ i ,
for i q, and Vi (Pp (R ) ∩ B2δ ) = 0, for i > q, completing the proof.
n m
We conclude that:
n
Rk (f ) k δ + ν
M (, [∆νf ] ∩ Bδm ) c .M (, ∆νf ∩ m
Bδ+ )
c. ( )
9 Mappings of Finite Smoothness 123
n
k
Rk (f )
C.
δ
( )ν .
Now if a ball Bδm is contained in the -neighborhood of ∆νf , we get:
n
Rk (f ) k δ ν
M (, [∆νf ] ∩ Bδm ) = M (, Bδm ) C. ( )
δ
Comparing this with M (, Bδm ) = C.( )m , we obtain the following corol-
lary:
Corollary 9.7. Let f : Brn → Rm be a C k -smooth mapping. The maximal
n
Rk (f ) k(m−ν)
ν
radius of a ball contained in [∆f ] is c. .
Examples of the type considered above show that this bound is virtually
sharp:
For instance, let us consider ϕ : R → [0; 1], a C ∞ -smooth function, such
1
that ϕ( ) = 1, ϕ(x) = 0, for all x ∈ R \ [0; 1] and |ϕ (x)| 2, for all x ∈ R.
2
If we denote by f the function defined as follows:
∞
1 (2p − 1)(2p + 1) 2p − 1
f (x) = .ϕ( x− ),
p=1
pk 2 2
k 1 1
we obtain a C 2 -smooth function with ∆0f = {0, 1, k
, . . . , k , . . .}. Now [∆0f ]
2 p
1 1 1
contains an interval of length δ() = + k , for = k − . It follows,
p p (p + 1)k
1 1 1
after an easy computation, that δ() = 1− k+1 ( k + k+1 ), where
1 k+1
(k + η( p ))
1 1
η( ) is a function tending to 0 as p tends to ∞. Finally, δ() ≥ c.1− k+1 ,
p
with c depending only on k.
2
Rk (f ) k
Under these asumptions, Corollary 9.7 gives δ() c. =
2
c 1− k .
Exercise. Find sequences αp → 0, βp → ∞, for which the function f (x) =
∞
1
k
ϕ(βp (x − αp )) is C γ , with γ arbitrarily close to k.
p=1
p
A more general example is the following: assume n m and represent Rm
as Rν ×Rm−ν . We build g : B1n → Rm exactly as the mapping g above, taking
Ψ : B1n → Rν to be the mapping covering the unit ball in Rν . Thus the rank
124 9 Mappings of Finite Smoothness
1
of Dg is ν on each ball of radius rα , centred at xα . Therefore the ∆νg form
s
4
an -net in Bδm , = δ s+1 and Bδm is contained in the -neighborhood of ∆νg ,
1 k(m − ν)
for δ = 1− s+1 and s > . Thus the lower bound for δ = δ() in this
n
n n
1− n+k(m−ν)
example is , while Corollary 9.7 gives: δ() c.1− k(m−ν) , showing
that this bound is asymptotically sharp, as k −→ ∞.
Our next result concerns quantitative transversality. As usual, we assume
that the domain and the image of our mappings are the Euclidean balls.
Moreover, we assume actually (although the statement of the Theorem 9.8
below is more general), that the submanifold in the image, to which our
mapping should be transversal, is the origin. (The general situation is reduced
to this special case by composing our mapping with the projection along the
submanifold to its normal plane).
So let f = (f1 , . . . , fm ) : Brn × Brm → Rm be a C k mapping. We assume
the following:
– For any (x, t) ∈ Brn × Brm , λm (Dt f(x,t) ) ≥ ρ > 0. By compactness of
Brn × Brm , this assumption is of course equivalent to the following:
Dt f(x,t) : Rm → Rm
is onto.
From this assumption, it follows that, for any x ∈ Brn , the C k -smooth mapping
f (x, .) : Brm → Rm is locally invertible. If furthermore we assume that:
– For any x ∈ Brn , f (x, .) : Brm → Rm is injective,
then f (x, .) : Brm → Rm is globally invertible and f −1 (x, .) satisfies the
Lipschitz condition with some constant L.
(This second assumption is technically convenient, but it can be easily
avoided.)
Now let, A1 ⊂ Brn , A2 ⊂ Rm , δ > 0 and Λ = (λ1 , . . . , λm ), with λ1 ≥
. . . ≥ λm ≥ 0 be given. We define Σ(f, Λ, A1 , A2 , δ) as the following set:
Σ(f, Λ, A1 , A2 , δ)
= {(x, t) ∈ A1 × Brm ; λi (Dx f(x,t) ) λi , 1 i m, f (x, t) ∈ [A2 ]δ }.
M (δ, A2 )
m
r δ δ
M (, ∆(f, Λ, A1 , A2 , δ)) c1 m
λ0 . . . λj ( )j (1+ +. . .+( )m−j ),
ρ j=0
if ≥ Rk (f ), and we have
c1 .c2 δ
M (r , A1 )M (δ, A2 )(1 + )m ×
ρm r
m
r j δ δ
λ0 . . . λj ( )j ( ) k (1 + + . . . + ( )m−j ),
j=0
R k (f )
if Rk (f ), where c2 = c2 max n m
(||f (x, t)||, ||Df(x,t) ||, 1), L and
(x,t)∈Br ×Br
1
r = r( )k .
Rk (f )
Proof. We can assume that A2 is a point (the origin in Rm ). Indeed, we
can cover [A2 ]δ by M (δ, A2 ) balls of radius 2δ, thus ∆(f, Λ, A1 , A2 , δ) is the
union of M (, A2 ) sets of type ∆(f, Λ, A1 , {0}, 2δ), and its entropy is at most
M (, A2 ) times the entropy of ∆(f, Λ, A1 , {0}, 2δ).
ρr
Let us consider first the case Rk (f ) . We denote by Pp the Taylor
2k
polynomial of degree p of f at the origin of Rn × Rm , where k = p + α,
α ∈]0; 1]. We have established in the proof of Theorem 9.1 the two following
inequalities:
||f (x, t) − Pp (x, t)|| Rk (f ) , for any (x, t) ∈ Brn × Brm (5)
pRk (f ) p
||Df(x,t) − DPp(x,t) || , for any (x, t) ∈ Brn × Brm (6)
r r
p ρ
Thus, by (6), for any (x, t) ∈ Brn × Brm , λm (Dt Pp (x,t) ) ≥ ρ − ≥ , since
r 2
ρr ρr
by assumptions, .
2k 2p
On the other hand, for any (x, t) ∈ Σ(f, Λ, A1 , {0}, δ), ||Pp (x, t)|| δ + ,
p
by (5), and λi (Dx Pp(x,t) ) λi = λi + , i ∈ {1, . . . , m},by (6).
r
Therefore, if Λ = (λ1 , . . . , λm ), then:
and
∆(f, Λ, A1 , {0}, δ) = π2 (Σ(f, Λ, A1 , {0}, δ))
⊂ π2 (Σ(Pp , Λ , A1 , {0}, δ + )) = ∆(Pp , Λ , A1 , {0}, δ + ).
126 9 Mappings of Finite Smoothness
Ĉ r
m
δ δ
m
λ0 . . . λj ( )j 1 + + . . . + ( )m−j .
ρ j=0
It follows that:
C1
m
k k r δ δ
(λ0 + ) . . . (λj + )( )j (1 + + . . . + ( )m−j )
ρm j=0 r r
C1
m
δ δ k j− r j
= (1 + + . . . + ( )m−j ) λj1 . . . λj ( ) ( )
ρm j=0 r
0 j1 ... j j
C2 δ m−j
m j
δ r
m
(1 + + . . . + ( ) ) λ0 . . . λ ( ) ( since λ1 ≥ . . . ≥ λm )
ρ j=0
=0
C2
m
r
m
δ δ
= λ0 . . . λ ( ) (1 + + . . . + ( )m−j )
ρm
=0 j=
c1
m
r δ δ
m
λ0 . . . λ ( ) (1 + + . . . + ( )m− ),
ρ
=0
||f (x1 , t1 ) − f (x1 , t0 )|| ||f (x1 , t1 ) − f (x0 , t0 )|| + ||f (x0 , t0 ) − f (x1 , t0 )||
2δ + M1 (f ).r
and by the Lipschitz condition on f −1 (x1 , .), we obtain:
L(2δ + M1 (f ).r ).
Finally, the total number of r -balls we need to cover Σ(f, Λ, A1 , {0}, δ)
is at most:
δ
M (r , A1 ).C3 (M1 (f ) · L)m .(1 + )m ,
r
and thus:
M (, ∆(f, Λ, A1 , {0}, δ)
δ m
m
c1 · c2 r j δ δ
m M (r , A1 )(1+ ) λ0 . . . λj ( )j ( ) k (1+ +. . .+( )m−j ).
ρ r j=0
R k (f )
n
In particular, for A1 = Br , we get:
Corollary 9.11. The set ∆(f, Λ, A1 , {0}, 0) of parameters t ∈ Brm , for which
there exists x ∈ Brn , such that ft at x is Λ-not-transversal to 0 ∈ Rm , satisfies:
m
r Rk (f ) n−j
M (, ∆(f, Λ, Brn , {0}, 0))
c. λ0 . . . λj ( )j ( ) k ,
j=0
c depends on M1 (f ), L and ρ.
where
That is to say: the set of parameters t ∈ Brm , for which there exists x in A1
such that ft (x) ∈ [A2 ] , has its -entropy bounded by c .M (, A1 )M (, A2 ),
where c only depends on M1 (f ), L, ρ.
Proof. It suffices to put λj = max ||Df(x,t) ||, = δ, and to apply
(x,t)∈Brn ×Brm
Theorem 9.8.
m
1 j
M δ, ∆(f, Λ, A1 , A2 , δ) c .M (r , A1 )M (δ, A2 ) λ0 . . . λj ( )j− k .
j=0
δ
Let us assume now that A1 = Brn and A2 = {0}. Then M (r , A1 ), for
1 n
= δ is of order ( ) k , and we get:
δ
9 Mappings of Finite Smoothness 129
m
1 n−j
M δ, ∆(f, Λ, Brn , {0}, δ) c λ0 . . . λj ( )j+ k .
j=0
δ
n−j n−m+1
Furthermore, we have m − j − ≥ 1− , for j m − 1 and
k k
k ≥ 1, thus for j m − 1 and k ≥ n − m + 1 we obtain:
n−j n−m+1
m−j− ≥1− ≥ 0.
k k
It follows from Corollary 9.14 and Proposition 2.6 that for any δ > 0:
Hm ∆(f, Λ, Brn , {0}, δ) S m ∆(f, Λ, Brn , {0}, δ)
δ m M δ, ∆(f, Λ, Brn , {0}, δ)
n−m+1 n−m
c .(δ 1− k + γδ − k ) = µ(δ, γ), (7)
where c depends only on M1 (f ), ρ, L, m,, assuming λ1 = . . . = λm−1 =
M1 (f ) and λm = γ.
In particular, for any k > n − m + 1, µ(δ, γ) −→ 0 as δ and γ tend to 0,
n−m
and γ δ k +ξ , ξ > 0.
Hence we have the following result:
Theorem 9.15. In any set E ⊂ Brm , with Hm (E) > η > 0, and for any δ, γ,
with µ(δ, γ) < η, there is a value t0 ∈ E of the parameter t, such that for any
xıBrn , if ||ft0 (x) δ||, then λm (Dft0 (x) ) ≥ γ.
Inequality (7) can be applied in various situations. We give here only the
following corollary:
Corollary 9.16. In any ball Bσm ⊂ Brm , there is t0 such that ft0 is δ-
transversal to 0 ∈ Rm (i.e. at each x where ||ft0 (x)|| δ, λj (Dft0 (x) ) ≥ δ, for
m m k−n−m+1
k
H (B1 ) mk
j ∈ {1, . . . , m}), for δ = σ k−n−m+1 .
2c
Proof. Substituting γ = δ in inequality (7), we obtain:
n−m+1 n−m n−m+1
µ(δ, δ) = c (δ 1− k + δ 1− k ) < 2c δ 1− k ,
1/(1− n−m+1 )
Hm (B1m )σ m k
hence for δ = , µ(δ, δ) < Hm (B1m )σ m = Hm (Bσm ).
2c
f (x, t) = g(x)+t. Then ∆(f, Λ, A, {0}, 0) is exactly the set of Λ-critical values
of g on the set A, and we have:
Theorem 9.17. Let f : Brn → Rm be a C k -smooth mapping, and ∆(f, Λ, A)
be the set of Λ-critical values of f on a subset A ⊂ Brn . Then for 0 <
Rk (f ):
1
m
r j
M (, ∆(f, Λ, A)) c.M ( k , A) λ0 . . . λj ( )j ( )k .
j=0
Rk (f )
As a special case (for A = Brn ) we obtain once more the result of Theo-
rem 9.2.
In order to stress the difference, consider only the consequence, concerning
the entropy dimension. Assuming that A ⊂ Σfν (the rank-ν set of critical
points of f , i.e. the set of x such that rank(Df (x)) ν)), we get:
1 1 1
M (, ∆νf ) c.M ( k , A)( )ν(1− k ) ,
1 1
dime (f (A)) dime (A) + ν(1 − ).
k k
In particular, for ν = 0:
1
dime (f (A)) dime (A).
k
Remark. It has been proved, in [Com 1], that, if A ⊂ Σf0 , Hs (A) = 0 im-
s 1
plies H k (f (A)) = 0, and thus that dimH (f (A)) dimH (A). Consequently,
k
Theorem 9.18 has an analogue for Hausdorff dimension.
Theorem 9.19. Let f : Brn → R be a C k -smooth function, if A ⊂ Brn is a
set of critical points, then:
1 1
dime (f (A)) dime (A) and dimH (f (A)) dimH (A).
k k
average bounds on volume of fibers etc. For M and N euclidean balls, ex-
plicit bounds are given in [Yom 3] for all the above quantities, in terms of
the bounds on the derivatives of f .
Also here, much more information can be presumably extracted by similar
methods: estimates for “bounded triangulations” of the fibers, more delicate
estimates of the geometry, including upper bounds for curvatures, estimates
for the spectrum of certain differential operators on these fibers (compare
Gromov’s paper [Gro 4]).
Till this point all the estimates in this section concerned a “typical” level
(fiber) of a mapping, or an “average” behavior of the fibers. There is another
important effect, related to all the level sets of “near-polynomial” differen-
tiable functions. It turns out that if a differentiable function f on the unit
ball B n is sufficiently close to a polynomial of degree k − 1 (i.e. if its partial
derivatives of the order k are sufficiently small with respect to its C 0 -norm),
then all its level sets resemble algebraic varieties of degree k − 1.
More accurately, it was shown in [Yom 2] that for f on B n k times
continuously differentiable, and with the C 0 -norm equal to 1, if the norm
of the k-th derivative of f in B n is bounded by 2−k−1 , then the set of zeroes
Y (f ) of f is contained in a countable union of smooth hypersurfaces. “Many”
straight lines intersect Y (f ) in at most k−1 points, and the n−1-dimensional
Hausdorff measure of Y (f ) is bounded by a constant C(n, k), depending only
on n and k.
This is in a strict contrast with the fact that any closed set in B n may be
the set of zeroes Y (f ) of an infinitely differentiable f , if we do not assume
restrictions on the derivatives.
We hope that many of the results of this book, concerning the -entropy
of semialgebraic sets and its behavior under polynomial mappings, can be
extended to “near-polynomial” functions. This approach presumably can be
applied in one important problem in Analysis, namely, description of the
“Nodal Sets” of the eigenfunctions of elliptic operators (see [CBar 1,2], [Donn-
Fef 1,2], [Han-Har-Lin], [Har-Sim], [Hof-Hof-Nad]). Recently, the results of
[Yom 2] have been applied in this context in [CBar 1].
A theorem of Kupka and Smale ([Kup 1], [Sma 1]) asserts that all the periodic
points of a generic diffeomorphism, or closed orbits of a generic flow, are
hyperbolic (i.e. no eigenvalue of the linearization of the mapping along the
orbit sits on the unit circle). The proof consists in a (rather delicate, as many
things in dynamics) application of a transversality theorem.
Quantitative transversality, obtained above, was used in [Yom 4] to get a
much more precise result: estimates of hyperbolicity, and not only for closed,
but for “almost closed” orbits.
10.1 Applications of Quantitative Sard and Transversality Theorems 135
The understanding of all the situation around the behavior of periodic and
almost periodic points and trajectories was recently changed dramatically
by striking results of Kaloshin ([Ka 1-3], [Ka-Hun]), who showed that any
prescribed growth rate can occure generically on one hand, and that the
exponential bound holds for “prevalent” diffeomorphisms, on the other hand!
incorporate the boundary data into the structure of the Taylor polynomials
at the grid points near the boundary.
Multi-Order Strategy
ones. At each order the maximal possible for this order accuracy is achieved,
and then the next order data is included as a “correction” to the previous
step. In many important situations this approach separates the scales and
strongly reduces the stiffness of the equations to be solved. In particular, it
excludes a necessity to divide the data by high degrees of . See [Wie-Yom]
for one specific implementation of the multi-order approach in solving elliptic
PDE’s and [Yom 27] for its more general mathematical treatment.
As the processing of the noisy empirical data is concerned (like digital
images) it is difficult to expect an overall smoothness. Here the multi-order
strategy is used also in order to discover a “hidden regularity” in the data
(which is closely related to the presence of singularities and to their Normal
Form representation). Consequently, in the implementations related to noisy
empirical data (like in Image Processing), the multi-order approach includes
an additional ingredient: the higher order data is used not everywhere, but
only in cases we the lower order analysis predicts a reliability of the higher
order information. See [Bri-Eli-Yom] and [Bri-Yom 6].
10.1.4.5 Flexible High Order Discretization: a General Structure.
In general, we combine the Kolmogorov representation of the smooth compo-
nents of the data with the Normal Form representation of the singularities.
The “geometric support” Z of the singularities (or the critical set) is explic-
itly memorized (Z can be called also the singular set of the data). Next, at
each point of a certain grid in a vicinity of Z the following data is stored: the
normal form of the local singularity and the coordinate transformations, (or
other allowed transformations) which bring this singularity to its normal form
(the “normalizing transformations”). Notice that in most cases the list of the
normal forms used is simple: there are several discrete types of normal forms,
and for each type, a small number of free parameters. Let us remind once
more that while the processing of singular data is usually ill-posed, bringing
singularity to its normal form and finding the normalizing transformations is
well-posed. Besides the improved stability and accuracy in producing our rep-
resentation from a noisy data, this fact has another important consequence:
our data size is usually strongly reduced in comparison with the original data
size.
10.1.4.6 Implementation Examples. The proposed approach has been
experimentally tested in several directions. In addition to the specific pub-
lications mentioned below, we plan to present a general description of the
underlying theory, of implemented algorithms as well as the experimental
results in [Yom-Bri 6], [Yom 27]. Here we just mention shortly the main
investigated problems.
1. We use at each grid point jets which satisfy the initial differential equation.
This reduces strongly the number of free parameters. For example, for two
independent variables, general jets of degree k contain (k+1)(k+2)
2 parameters.
Jets of degree k, satisfying a linear PDE with constant coefficients of order
2, have 2k + 1 free parameters.
(Notice that the requirement for local representing elements to satisfy
the initial differential equation is usually not compatible with the precise
adjustment of the values and the derivatives for the neighboring elements
(since the last requirement leads to elements with compact support)).
2. We find a relaxation scheme (i.e., equations relating a Taylor polynomial
at each grid point with its neighboring polynomials) which provides a highest
order discretization error. Since we spend no degrees of freedom to provide
boundary adjustment of neighboring polynomials, enough degrees of freedom
remain to “cancel” the Taylor coefficients in a discretization error up to a
high order. The following rough computation shows what order of accuracy
can be expected
For a second order linear PDE and jets of degree k we have 2k + 1 param-
eters at each grid point. Thus at four nearest neighboring points, we have
8k + 4 parameters. Assuming that the systems that arise are solvable, we can
reconstruct (at the central point) a jet of degree 4k satisfying the equation,
or we can cancel the Taylor terms in the discretization error up to the same
degree. In particular, for ∆u = 0, for a regular grid and k = 2, the corre-
sponding equations are solvable up to degree 10 (and not only 8) because of
a symmetry in the problem.
Notice that wider neighborhood stencils can be involved in the relaxation
procedure. However, geometrically compact schemes have important compu-
tational advantages.
Initial investigation of the parabolic equations has been started in [Bic-
Yom]. It confirmed the results obtained in the elliptic case. Investigation
of equations, developing singularities (like Burgers equation) has also been
started.
Inversion of Mappings
The inversion of the normal forms (the standard fold and cusp) is assumed
to be easily available. In practical realizations of the algorithm this inversion
is implemented in a special fast subroutine. Now to compute the inverse
function f −1 (y) for any given y, we just substitute y into the corresponding
model. For a regular point the k-jet of the inverse mapping at the point y is
computed. For the fold (cusp) the point y is substituted into the k-jet of the
the normalizing coordinate transformation in the target. Then the inverse of
the standard fold (cusp) is applied. Finally, the result is substituted into the
k-jet of the the inverse coordinate transformation in the source. In any case,
the overall complexity of computations is comparable with that of computing
a polynomial of degree k at one point.
The logical structure of the algorithm is as follows:
The algorithm analyses the input Taylor polynomials of the direct map-
ping. According to the Whitney description of the singularities of the plane
to plane mappings, a decision is taken, whether the considered point is classi-
fied as a regular point, near-fold or near-cusp. In the first case the inverse jet
is obtained by the standard jet-inversion formulae from the “Jet-calculus”.
If the point has been classified as a near-fold (near-cusp), the normalizing
coordinate transformations are computed via a “jet implementation” of the
Whitney procedure ([Whi 2]).
(Notice that the construction of a numerically stable jet implementation
for the procedure of bringing a singularity to its normal form is not trivial.
It requires, in particular, a careful study of the “balance of the smoothness”
for this specific singularity. Sometimes it may be necessary to introduce ad-
ditional parameters to the normal form, in order to increase stability of com-
putations (see [Eli-Yom 4], [Yom 27]). However, as a stable and efficient “jet
normalization” has been constructed, it becomes one of the standard formulae
of the Jet calculus library).
The main problem in the implementation of the algorithm is in the classi-
fication of the points. Indeed, if the point has been classified as a near-fold, a
real fold must be somewhere nearby, and the normalizing coordinate transfor-
mations to the normal form of this fold must be effectively bounded together
with their derivatives (otherwise their computing will not be reliable enough).
To guarantee this we must to know that this nearby fold (cusp) is sufficiently
non-degenerate.
So we put a number of thresholds, determining the decision at each
branching point, and the algorithm operates according to the basic pattern
described above and to the chosen thresholds. Of course, as the step-size of
the greed decreases, the non-degeneracy assumptions can be relaxed.
If the non-degeneracy assumptions are not satisfied in a neighborhood of
a certain point, we either subdivide the grid or restrict the processing to a
rough low-order approximation of the inverse mapping.
10.1 Applications of Quantitative Sard and Transversality Theorems 147
The most advanced implementation of our approach till now has been
achieved in Image Processing. A nonlinear high-order approximation, to-
gether with the appropriate “Normal forms” capturing image singularities,
have been used to achieve a very compact representation of images and “syn-
thetic video” sequences ([Bri-Eli-Yom], [Eli-Yom 5], [Y-E-B-S], [Bri-Yom 6]).
The results have been intensively tested against various known methods, and
have been used in working practical applications.
We do not discuss here the details of this implementation. Such a dis-
cussion would go too far away from the mainstream of this book. Let us
notice only that we indeed use in image analysis a third and fourth order lo-
cal polynomial approximation, which would be impossible without a proper
application of the multi-order strategy. We use three types of normal forms
for singularities (edges, ridges and “patches”) and a low order Kolmogorov
representation for smooth parts of the image. Since most of the edges and
ridges in a typical image of the real world appear as the visible contours of
the 3D objects, we do believe, that the methods of this book will ultimately
help us to optimize our representation.
“geometric complexity bounds” (in the spirit of the above sections of this
chapter) can be obtained for functions of bounded semialgebraic complexity.
where the infinum is taken over all the semialgebraic functions g, such that:
||f − g||C 1 .
over all polynomials p with C(p) dn . (Notice that above, C(p) = dn .) Writ-
ten as above, the definition shows that σp and Ep are constructed exactly as
σs and Es , only with a subclass of all semialgebraic functions. This proves
immediately that for any > 0 and d > 0,
Ep (f, d) ∼ q d , q < 1.
(We do not intend here to give accurate formulations of the results of
Approximation Theory. Consequently, some details, sometimes important,
are omitted).
Let us return to semialgebraic complexity. It is bounded by the polynomial
one, and one can show that for generic C k or analytic functions σs and σp are
equivalent. On the other hand, we can see immediately, that semialgebraic
complexity can be small for functions, not regular in the usual sense. Indeed,
let f (x) be defined as:
0, −1 x 0
f (x) =
x2 , 0 x 1 .
The result of this section is that the geometry of the set of critical values
of the C 1 -function is determined by its semialgebraic complexity (and not
by its regularity, as it appears in standard settings of Sard-like results). So
let f : Brn → R be a C 1 -function, Σ(f ) the set of its critical points, and
∆f = f (Σf ) the set of its critical values.
Theorem 10.8. For any > 0,
by Theorem 7.5 of Chapter 7 above. Now taking the infimum over all the
semialgebraic g with ||f − g||C 1 , we get
Applying this proposition to the partial sums of the infinite series, we get,
after simple computations:
∞
1
Theorem 10.12. For f = ( )i pi (x1 , . . . , xi ), deg pi = d, |pi | 1 on the
i=1
q
1
unit ball and q > 1, σQ (f, ) C(q, d)( )logq (2d) . In particular, for q > 2d,
f satisfies the Sard theorem (i.e., m(∆f ) = 0).
Returning to the Kupka example above, we see that it occurs exactly on
∞
1
the boundary: by Theorem 10.12, for any q > 6, functions f = pi (x, . . . ,
i=1
qi
xi ) with deg pi = 3, satisfy the Sard theorem. In the specific form of the
function f in Kupka’s example above it is enough to take q > 2. By formal
analogy we can say that the complexity of the function f : 2 → R, f =
∞
1
p (x , . . . , xi ), deg pi = d, is the same as the complexity of C k -functions
i i 1
i=1
q
g : B n → R, if k−1 n
= logq (2d) = β. In particular, the sequences of the form
1, 1/2 , 1/3 , . . . , 1/k s , . . ., may appear among the critical values of both f
s s
We’d like to briefly mention here some results concerning the singular be-
haviour of semialgebraic or tame mappings at infinity. To do this, let us
define the set K(f ) of generalized critical values of a semialgebraic or a tame
mapping f : Rn → Rm , n > m:
K(f ) = ∆f ∪ K∞ (f ),
∆f being the classical set of critical values of f (i.e. ∆f = f (Σf ), with Σf
the set of points x ∈ Rn such that Df(x) is not onto) and K∞ (f ) being the
set of critical values at infinity or asymptotic critical values, defined by:
K∞ (f ) = {y ∈ Rm , ∃xk ∈ Rn , |xk | → ∞, f (xk ) → y, |xk | · λm (Df(xk ) ) → 0}.
Of course, when xk is a singular point, λm (Df(xk ) ) = 0, thus K∞ (f ) is the
union of adh(∆f ) (the closure of ∆f ) and of the set:
156 10 Some Applications and Related Topics
We essentially refer here to [Pas] (see [Eva-Gar] for the basic definifions).
The main theorem of this paper is the following, generalizing the classical
Morse-Sard Theorem for Sobolev functions.
Theorem. [Pas] Let n > m be two integers and let, for p > n, f ∈
n−m+1,p
Woc (Rn ; Rm ). Then Hm (∆f ) = 0.
The continuous representative f of f is in C n−m,α , for some α = α(p) < 1
(see [Eva-Gar]), and the quantitative Morse-Sard Theorem for f shows that
n−m+1
the Hausdorff dimension of ∆f is at most min(m, m − 1 + ) = m.
n−m+α
In other words, the classical Morse-Sard Theorem gives no information on the
n−m+1,p
dimension of ∆f , when f ∈ Woc (Rn ; Rm ). As pointed out in [Pas], this
is the effect of the existence of another weak derivative summable enough.
The idea of the proof is the following: we know that there exists, for
every > 0, a set F ⊂ Rn and a mapping f ∈ C n−m+1 (Rn ; Rm ), such that
10.3 Additional Directions 157
f|F ≡ f|F and Hn (Rn \ F ) . Let us notice that Σf = (Σf ∩ F n1 ) ∪ N ,
n∈N
where Hn (N ) = 0, and that, by the classical Morse-Sard Theorem:
Hm (f (Σf ∩ F n1 )) = Hm (f n1 (Σf 1 ∩ F n1 )) = 0.
n
The main point of [Pas] is then to prove that Hm (N ) = 0, i.e. for several sets
A ⊂ Σf , such that Hn (A) = 0, we have Hm (f (A)) = 0.
Classical examples show that this result cannot be sharpened.
and
E = {(P̄ , , x); P̄ ∈ [Ḡn−i
n ]1 , ∈]0; 1], x ∈ A ∩ P̄ },
subanalytic case) morphism. One can stratify this morphism in the following
way: there exist a Whitney stratification Σ of G, compatible with E and
F , and a Whitney stratification Σ of [Ḡn−i n ]1 × [0; 1] (which is of course
finite, since [Ḡn−i
n ]1 × [0; 1] is compact), such that the fibers of p−1 ({(P̄ , )})
−1
and p ({(Q̄, η)}) are homeomorphic with respect to E and F . In particular
the sets A ∩ P̄ = p−1 ({(P̄ , )}) ∩ E and A ∩ Q̄ = p−1 ({(Q̄, η)}) ∩ E are
homeomorphic.
This shows that:
– the number of topological types of A ∩ P̄ is finite,
n ]1 , the topological type of A ∩ P̄ does not depend
– for a given P̄ ∈ [Ḡn−i
and
Λoc
i (A0 ) = lim c(n, i) χ(A ∩ P̄ ) dP̄
→0 P̄ ∈[Ḡn−i
n ]1
c(n, i)
= lim χ(A ∩ Bn ∩ P̄ ) dP̄ ,
→0 i P̄ ∈Ḡn−i
n
the local i-th variation of the germ A0 and the local Lipschitz-Killing curva-
ture of A0 (the global Lipschitz-Killing curvatures have been introduced in
[Wey], see also [Kla], [Had], [Sch], [Sch-McM]).
Of course we have a uniform bound on P̄ , in the semialgebraic case, for
these invariants, in terms of the diagram of A.
One can study the relation between the geometry of a multigerm (Ay )y∈Y
along a smooth set Y and the variation of the invariants Λoc i (Ay ) and
Vioc (Ay ) along Y . One can for instance prove that these invariants vary con-
tinuously along Y , when Y is a stratum of a Verdier stratification of adh(A)
(see [Com-Gra-Mer]).
Let DR denote the closed disk of radius R > 0, centered at the origin in C.
160 10 Some Applications and Related Topics
Definition 10.14. Let R > 0, and K > 0 be given and let f be holomorphic
in a neighborhood of DR . We say that f belongs to the Bernstein class BR,K
if the maximum of the absolute value of f over DR is at most K times the
maximum over D 12 R . The constant K is called the Bernstein constant of f .
This definition is motivated by one of the classical Bernstein inequalities:
let p(x) be a polynomial of degree d. Then
where ER is the ellipse in C with the focuses at −1, 1 and the semiaxes R
([Ber]).
A problem of computing Bernstein constants of algebraic functions has
recently appeared in several quite different situations.
In [Fef-Nar 1-3] this problem is investigated in relation with estimates of
a symbol of some pseudodifferential operators. In [B-L-M-T 1,2] and [Bru 1],
[Bar-Ple 1-3] (see also [Bos-Mi]) this problem is connected with some results
in Potential Theory and with a characterization of algebraic subsets.
In [Roy 1,2], [Roy-Yom], [Bri-Yom 5], [Fra-Yom 1,2], [Yak 1,2], [Yom 22]
and [Bru 2] Bernstein classes are used in counting zeroes in finite dimensional
families of analytic functions (this problem is closely related to the classical
problem of counting closed trajectories (limit cycles) of plane polynomial
vector-fields).
In [Yom 6,7], [Yom 21] and [Yom 23] various forms of Bernstein inequality
are used to prove results on a “C k -reparametrization” of semialgebraic sets,
which, in turn allow one to control the complexity growth in iterations of
smooth mappings (see Section 10.3.4 above). It was exactly the absence of
the Bernstein inequality for algebraic functions, which restricted the results
of [Yom 21] to one and two dimensional dynamics only.
By a structural Bernstein inequality for a certain class of functions, de-
fined by algebraic data (algebraic functions, solutions of algebraic differential
equations, etc.) we understand an inequality bounding the Bernstein constant
of the function on a couple of concentric disks in terms of the degree and the
relative position of these concentric disks in the maximal concentric disk of
regularity only.
As the example of rational functions shows, in a sense, this is the best
possible inequality one can expect for functions with singularities.
Let y(x) be an algebraic function, given by an equation:
4A(R + R1 ) 2m+2
Theorem 10.15. For any R1 < R, ỹ ∈ BR1 ,K , with K = [ ] .
R − R1
Here A is an absolute constant.
Theorem 10.15 provides a structural Bernstein inequality for algebraic
functions of one variable. It can be easily extended to algebraic functions of
several variables (see [Roy-Yom]).
One can restate this theorem (or, rather, its generalization to multivalued
functions) in a more geometric way: if an algebraic curve Y of degree d in C 2
is contained, over the disk DR in the x-axis, in a tube of the size K, and it
does not blow up to infinity over the disk D3R , then it is contained in a tube
of the size C(d)K over the disk D2R .
One can hope that this last result admits for a generalization to higher di-
mensions and more complicated (semi)algebraic sets. There is also an impor-
tant problem of obtaining structural Bernstein inequalities for other classes
of analytic functions, beyond algebraic ones (in particular, for solutions of
algebraic differential equations and, hopefully, for their Poincare mappings).
Some initial results in this direction are given in [Roy 1,2], [Roy-Yom], [Fra-
Yom 1,2], [Bri-Yom 5], [Yom 22], [Bru 2].
In the next section we discuss shortly polynomial control problems, stress-
ing situations, where semialgebraic geometry underlines the dynamics of the
trajectories.
A specific problem, related to the main topics of this book, and considered
in [Bri-Yom 1-4] is the validity of Sard’s theorem and its “quantitative” gen-
eralizations for an important class of nonlinear mappings, namely, input to
state mappings of nonlinear finite dimensional control problems of the type:
island Oi in the plane (x, u) the trajectory (x(t), u(t)) can pass either above
or below Oi . Now two trajectories, that pass on the same side of each of the
islands Oi , are “visible” one from another. Using properties of semialgebraic
sets, discussed in this book, one can join these trajectories inside S by paths
of controllable lengths, and estimate the difference of the derivatives of x(t).
As a result, we get a differential inequality, which, in turn, implies that the
endpoints of the two trajectories as above must be close to one another.
The following result is obtained in [Bri-Yom 2] by a detailed analysis on
the above lines:
Denote by WK the set of K-Lipschitz u on [0,1] with |u(t)| 1, and fix
the Lp -norm on the control space, p ≥ 1.
Theorem 10.17. Assume x0 = 0 in (1.1). Let f (x, u) be a polynomial of
degree d, satisfying |f (x, u)| 1 for |x| 1, |u| 1. Then for any γ ≥ 0 the
2
set of γ-critical values of Jf on WK can be covered by N (d) = dγ 23(d+1)
intervals of length δ = (qK)1/q γ q/q+1 . Here 1/p + 1/q = 1.
In particular, the measure of the γ-critical values of Jf does not exceed
N (d)δ.
Thus, a quantitave Sard theorem is valid for the control problems as above.
On can apply the approach of “Semialgebraic Complexity” described in this
chapter, and extend the result to right-hand sides more complicated than
polynomials. However, since the growth of the estimate of theorem 10.17
in d is very fast, f above can be replaced only by analytic functions of a
very restricted growth. Using the same considerations as above, but with an
infinite number of the “islands” Oi , one can easily construct control problems
of the above form with f infinitely smooth and with critical values of J
covering the whole interval (see [Bri-Yom 3,4]).
The approach outlined above can be applied also in higher-dimensional
control problems. In higher dimension the relation between the dynamics of
near-critical trajectories and semi-algebraic geometry of the right hand side
polynomials remains especially transparent for the trajectories of “rank zero”:
those for which the norm of the differntial of the input to state mapping J is
small. (For near-critical trajectories of higher rank their behavior is governed
by a combination of semi-algebraic restrictions with the Pontryagin maximum
principle, and all the considerations become more delicate).
However, also for near-critical trajectories of “rank zero” a new important
dynamical problem enters: consider, for example, the case of two-dimensional
x and one-dimensional u. Here as above critical trajectories of rank zero must
lie in a semi-algebraic set S in a three-dimensional phase space (x, u), defined
by the condition that fu (x, u) be small, where f is a (vector) right hand side
of the equation (1.1). We can consider the complement of S as a set of islands
Oi , but now these islands are three-dimensional bodies, and in addition to
a possibility of passing above or under the island, the trajectory can rotate
around it. See [Bri-Yom 3,4] for a detailed discussion of this direction.
164 10 Some Applications and Related Topics
The C k−1 -norm of the coordinate transformation from the original co-
ordinates to y1 , ..., yn (and of the inverse transformation) does not exceed
M (K, ).
166 10 Some Applications and Related Topics
b. For any i
= j, the distance between the critical values f (xi ) and f (xj ) is
not smaller than ψ2 > 0.
Then there is 0 > 0 (depending only on K, ψ1 , ψ2 ) such that for any
given , 0 > > 0, and for any f1 which is closer than to f in C k -norm,
f1 is equivalent to f via the diffeomeophisms G and H of the source and
the target, respectively. G and H differ (in C k−1 -norm) from the identical
diffeomeorphisms not more than by s(K, ψ1 , ψ2 , ). Here s(K, ψ1 , ψ2 , ) tends
to zero as tends to zero. For the proof see [Yom 28]. The next “quantitative”
result has no direct analogy in the classical Singularity Theory. It claims that
for a generic mapping each its “near-singular” point belongs to a controlled
neighborhood of one of exact singular points (its “organizing center”).
A more accurate statement of this result is as follows:
Statement 3. Let a C k function f be given, with all the derivatives up to
order k uniformly bounded by K. Then for any given > 0, we can find h
with hC k , such that for f = f0 + h the conditions i-iv of Statement 1
are satisfied, as well as the following additional condition.
v. There is η(K, ) > 0 such that for any point x if the norm of the gradf (x) is
smaller than η(K, ) then x belongs to one of the controlled neighborhoods
of the singular points xi of f .
[Bat 5] S. M. Bates, On the image size of singular maps, II. Duke Math. J. 68, n.
3, (1992), 463-476.
[Bat 6] S. M. Bates, A smooth surjective rank 1 endomorphism of a Hilbert space.
Int. Math. Res. Not. 6, (1991), 71-73.
[Bat 7] S. M. Bates, On smooth, nonlinear surjections of Banach spaces. Israel J.
Math., 100, (1997), 209-220.
[Bat-Mor] S. M. Bates, C. G. Moreira, De nouvelles perspectives sur le théorème
de Morse-Sard. C. R. Acad. Sci. Paris 332, (2001), Série I, 13-17.
[Bat-Nor] S. M. Bates; A. Norton, On sets of critical values in the real line. Duke
Math J. 83, No. 2, (1996), 399-413.
[Ben-Ris] R. Benedetti; J. J. Risler, Real algebraic and semi-algebraic sets. Actu-
alités Mathématiques, Hermann, Paris, (1990).
[Ber] S. Bernstein, Sur une propriété des polynomes. Proc. Kharkov Math. Society,
Serie 2, v.14 (1913), 1-6.
[Bic-Yom] E. Bichuch, Y. Yomdin, High order Taylor discretization for numerical
solution of parabolic PDE’s, preprint, 1995.
[Bie-Mi 1] E. Bierstone; P. Milman, Composite differentiable functions, Ann. of
Math (2) 116 (1982), no. 3, 541-558.
[Bie-Mi 2] E. Bierstone; P. Milman, Algebras of composite differentiable functions,
Singularities, Part 1 (Arcata, Calif., 1981), 127-136, Proc. Sympos. Pure Math.,
40, Amer. Math. Soc., Providence, RI 1983.
[Bie-Mi 3] E. Bierstone; P. Milman, Relations between analytic functions I, Ann.
Inst. Fourier (Grenoble) 37 (1987), no. 1, 187-239.
[Bie-Mi 4] E. Bierstone; P. Milman, Relations between analytic functions II, Ann.
Inst. Fourier (Grenoble) 37 (1987), no. 2, 49-77.
[Bie-Mi 5] E. Bierstone; P. Milman, Semianalytic and subanalytic sets. Inst. Hautes
Études Sci. Publ. Math. 67 (1988), 5-42.
[Bie-Mi 6] E. Bierstone; P. Milman, Canonical desingularization in characteristic
zero by blowing up the maximum strata of a local invariant, Invent. Math. 128
(1997), no. 2, 207-302.
[Bie-Mi 7] E. Bierstone; P. Milman, Geometric and differential properties of sub-
analytic sets, Ann. of Math (2) 147 (1998), no. 3, 731-785.
[Bie-Mi-Paw 1] E. Bierstone; P. Milman, W. Pawlucki, Composite differentiable
functions, Duke Math. J. 83 (1996), no. 3, 607-620.
[Bie-Mi-Paw 2] E. Bierstone; P. Milman, W. Pawlucki, Differentiable functions de-
fined in closed sets. A problem of Whitney, Invent. Math. 151 (2003), no. 2,
329-352.
[Bla] W. Blaschke, Vorlesungen über Integralgeometrie. 3te Aufl. Deutscher Verlag
der Wissenschaften, Berlin, 1955.
[Boc-Cos-Roy] J. Bochnak; M. Coste; M. F. Roy, Real algebraic geometry. Trans-
lated from the 1987 French original. Ergebnisse der Mathematik und ihrer
Grenzgebiete (3) 36, Springer-Verlag, Berlin, (1998).
[Boa] J. M. Boardman, Singularities of differential maps. Publ. Math. I.H.E.S. 33,
(1967), 21-57.
[B-L-M-T] L. Bos; N. Levenberg; P. Milman; B.A. Taylor, Tangential Markov in-
equalities characterize algebraic submanifolds of C n . Indiana Univ. Math. J.
44, (1995), 115-137.
[Bos-Mi] L. Bos; P. Milman, Sobolev-Gagliardo-Nirenberg and Markov type in-
equalities on subanalytic domains, Geom. Funct. Anal. 5 (1995), no. 6, 853-923.
References 175
[Pas] L. de Pascale, The Morse-Sard Theorem in Sobolev Spaces. Ind. Univ. Math.
J. 50, n. 3, (2001).
[Pet-Ole 1] I. G. Petrovskii; O. A. Oleinik, On the topology of real algebraic sur-
faces. (Russian) Izvestiya Akad. Nauk SSSR. Ser. Mat. 13, (1949), 389-402.
[Pet-Ole 2] I. G. Petrovskii; O. A. Oleinik, On the topology of real algebraic sur-
faces. Amer. Math. Soc. Translation 70 (1952), 20 p. Translated from Izvestiya
Akad. Nauk SSSR. Ser. Mat. 13, 389-402 (1949);
[Pey] J. Peyrieres, Introduction aux mesures et dimension de packing, in Fractals,
dimensions non entieres et applications, Scientific Editor G. Cherbit, Masson,
Paris 1987.
[Rab] P. J. Rabier, Ehresmann fibrations and Palais-Smale conditions for mor-
phisms of Finsler manifolds. Ann. of Math. (2), 146, n. 3, (1997), 647–691.
[Ren 1] J. Renegar, On the computational complexity and geometry of the first-
order theory of the reals. I. Introduction. Preliminaries. The geometry of semi-
algebraic sets. The decision problem for the existential theory of the reals. J.
Symbolic Comput. 13, no. 3, (1992), 255-299.
[Ren 2] J. Renegar, On the computational complexity and geometry of the first-
order theory of the reals. II. The general decision problem. Preliminaries for
quantifier elimination. J. Symbolic Comput. 13, no. 3 (1992), 301-327.
[Ren 3] J. Renegar, On the computational complexity and geometry of the first-
order theory of the reals. III. Quantifier elimination. J. Symbolic Comput. 13,
no. 3, (1992), 329-352.
[Rie 1] J. Rieger, Generic properties of edges and corners an smooth greyvalue
surfaces, Biol. Cybernet. 66 (1992), 497-502.
[Rie 2] J. Rieger, Generic evolution of edges on families of diffused greyvalue sur-
faces, J. Math. Imag. Vision 5 (1995), 207-217.
[Ris 1] J. J. Risler. Complexité et géométrie réelle (d’après A. Khovansky).
Séminaire Bourbaki, Vol. 1984/85. Astérisque 133-134, (1986), 89-100.
[Ris 2] J. J. Risler. Additive complexity and zeros of real polynomials. SIAM J.
Comput. 14, (1985), 178-183.
[Roc] R. T. Rockafeller, Favorable classes of Lipschitz continuous functions in sub-
gradient optimization. In Progress in Nondifferentiable Optimization, E. Nur-
minski,ed. Pergamon Press, New-York, (1981).
[Roh 1] A. Rohde, On Sard’s theorem for nonsmooth functions. Numer. Funct.
Anal. Optim. 18, (1997), n. 9-10, 1023-1039.
[Roh 2] A. Rohde, On the density of near-critical values of continuous selections
and maximum functions. Approximation and optimization in the Caribbean, II
(Havana, 1993), 569-581, Approx. Optim., 9, Lang, Frankfurt am Main, (1997).
[Roh 3] A. Rohde, On the - entropy of nearly critical values. J. Approx. Theory
76, (1994), n. 2, 166-176.
[Roy 1] N. Roytvarf, Generalized moments, composition of polynomials and Bern-
stein classes. In “Entire functions in modern analysis. B. Ya. Levin memorial
volume”, Isr. Math. Conf. Proc. 15, (2001), 339-355.
[Roy 2] N. Roytvarf, Markov type inequality for functional determinants. Qual.
Theory of Dyn. Systems 2, n. 1, (2001), 79-92.
[Roy-Yom] N. Roytvarf; Y. Yomdin, Bernstein Classes. Annales De l’Institut
Fourier, Grenoble 47 (1997), 3, 825-858.
[San] L. A. Santalo. Integral geometry and geometric probability. With a fore-
word by Mark Kac. Encyclopedia of Mathematics and its Applications Vol. 1.
Addison-Wesley Publishing Co., Reading, Mass.-London-Amsterdam, (1976).
References 183
[Sar 1] A. Sard. The measure of the critical values of differentiable maps. Bull.
Amer. Math. Soc. 48, (1942), 883-890.
[Sar 2] A. Sard, Images of critical sets. Ann. Math. 68, (1958), 247-259.
[Sar 3] A. Sard, Hausdorff measure of critical images on Banach manifolds. Amer.
J. Math. 87, (1965), 158-174.
[Sch] R. Schneider, Convex bodies: the Brunn-Minkowski theory. Encyclopedia of
Mathematics and its Applications 44. Cambridge University Press, Cambridge,
(1993), 490 p.
[Sch-McM] R. Schneider; P. McMullen, Valuations on convex bodies. Convexity
and its applications, Birkhäuser, Basel, (1983), 170-247.
[Sch-Sha] J. T. Schwartz; M. Sharir, On the “piano movers” problem II. General
techniques for computing topological properties of real algebraic manifolds.
Adv. Appl. Math. 4, (1983), 298-351.
[Shap-Yom] A. Shapiro; Y. Yomdin, On Functions representable as a difference of
two convex functions and necessary conditions in a constrained optimization.
Preprint, Beer-Sheva, (1981).
[Sham-Yom] T. Shamir; Y. Yomdin, Repeatability of redundant manipulators:
mathematical solution of the problem, IEEE Trans. on Automatic Control,
AC-33, No. 11 (1988), 1004–1009.
[Shi] M. Shiota. Geometry of subanalytic and semialgebraic sets. Progress in Math-
ematics 150, Birkhäuser Boston, (1997).
[Shu-Sma 1] M. Shub; S. Smale. Computational complexity. On the geometry of
polynomials and a theory of cost. I. Ann. Sci. éc. Norm. Sup. 4ème Série, t. 18
(1985), n. 1, 107–142.
[Shu-Sma 2] M. Shub; S. Smale, Computational complexity: on the geometry of
polynomials and a theory of cost. II. SIAM J. Comput. 15 (1986), n. 1, 145-
161.
[Shu-Sma 3] M. Shub; S. Smale, Complexity of Bezout’s Theorem. I: Geometric
aspects. J. Am. Math. Soc. 6, n. 2, (1993), 459-501.
[Shu-Sma 4] M. Shub; S. Smale, Complexity of Bezout’s Theorem. II: Volumes and
probabilities. Birkhauser Prog. Math. 109, (1993), 267-265.
[Shu-Sma 5] M. Shub; S. Smale, Complexity of Bezout’s Theorem. III: Condition
number and packing. J. Complexity 9, n. 1, (1993), 4-14.
[Shu-Sma 6] M. Shub; S. Smale, Complexity of Bezout’s Theorem. IV: Probability
of success, extensions. SIAM J. Number Anal. 33, n. 1, (1996), 128-148.
[Shu-Sma 7] M. Shub; S. Smale, Complexity of Bezout’s Theorem. V: Polynomial
time. Theor. Comput. Sci. 133, n. 1, (1994), 141-164.
[Sik] J.-C. Sikorav, Construction de sous-variétés symplectiques (d’après S. K. Don-
aldson et D. Auroux), Sḿinaire Bourbaki, Vol. 1997/98. Astérisque No. 252
(1998), Exp. No. 844, 4, 231-253.
[Sla 1] V. V. Slavskii, Integral-Geometric relation with orthogonal projection for
multidimensional surfaces, Dokl. Akad. Nauk SSSR 214 (1974), 48-51.
[Sla 2] V. V. Slavskii, Integral-Geometric relation with orthogonal projection for
hypersurfaces, Sibirsk. Mat. Z. 16 (1975), 103-123.
[Sla 3] V. V. Slavskii, Integral-Geometric relation with orthogonal projection for
surfaces, Sibirsk. Mat. Z. 16 (1975), 355-367.
[Sma 1] S. Smale, Stable manifolds for differential equations and diffeomorphisms.
Ann. Sc. Norm. Sup. Pisa 18, (1963), 97-116.
[Sma 2] S. Smale, An infinite dimensional version of Sard’s theorem. Amer. J.
Math. 87, (1965), 861-866.
184 References
[Yom 13] Y. Yomdin, Beta-spread of sets in metric spaces and critical values of
smooth functions. Preprint, MPI Bonn, (1983).
[Yom 14] Y. Yomdin, On represntability of convex functions as maxima of linear
families. Preprint, MPI Bonn, (1983).
[Yom 15] Y. Yomdin, Maxima of smooth families, III: Morse-Sard Theorem.
Preprint, MPI Bonn, (1984).
[Yom 16] Y. Yomdin, Approximational complexity of functions and entropy. Pro-
ceedings of the Israel Mathematical Union Conference (Tel Aviv, 1987), 1-9,
Tel Aviv Univ., Tel Aviv, (1987).
[Yom 17] Y. Yomdin, Approximational complexity of functions. Geometric aspects
of functional analysis (1986/87), 21-43, Lecture Notes in Math., 1317, Springer,
Berlin-New York, (1988).
[Yom 18] Y. Yomdin, Sard’s theorem and its improved versions in numerical analy-
sis. Proc. A.M.S. Conf., Colorado, 1988. Lectures in Applied Math. 29, (1990),
701-706.
[Yom 19] Y. Yomdin, Surjective mappings whose differential is nowhere surjective.
Proc. A.M.S., Vol. 111, n.1, (1991), 267-270.
[Yom 20] Y. Yomdin, Complexity of functions: some questions, conjectures and
results. J. of Complexity, 7, (1991), 70-96.
[Yom 21] Y. Yomdin, Local complexity growth for iterations of real analytic map-
pings and semi-continuity moduli of the entropy. Erg. Theory and Dynamical
Systems, 11, (1991), 583-602.
[Yom 22] Y. Yomdin, Global finiteness properties of analytic families and algebra
of their Taylor coefficients. Proceedings of the Arnoldfest, Toronto, 1997, Fields
Inst. Commun., 24, AMS, Providence, RI, (1999), 527-555.
[Yom 23] Y. Yomdin, Analytic reparametrization of semialgebraic sets and func-
tions. In preparation.
[Yom 24] Y. Yomdin, Semialgebraic complexity of functions. To appear in J. of
Complexity.
[Yom 25] Y. Yomdin, On the local structure of a generic central set, Compositio
Math. 43 (1981), no. 2, 225-238.
[Yom 26] Y. Yomdin, Quantitative Singularity theory A summary of a course given
at the Eur. Math. Union Summer School, Cluj, Romania, 1998, preprint.
[Yom 27] Y. Yomdin, Jet Representation and Processing of functions, in prepara-
tion.
[Yom 28] Y. Yomdin, Quantitative Morse Theorem, in preparation.
[Y-E-B-S] Y. Yomdin; Y. Elichai; S. Birman; E. Spiegel, Synthetic textures in
MPEG, to appear.
[Zel-Zh] I. Zelenko; M. Zhitomirskii, Rigid curves of generic 2-distributions on 3-
manifolds, Duke Math. Journal 79, no. 2 (1995), 281-307.
[Zer] M. Zerner, Approximation par des variétés algébriques dans les espaces
Hilbertiens. Studia Mathematica t. 38, (1970), 413-435.
[Zh] M. Zhitomirskii, Typical singularities of differential 1-forms and Pfaffian equa-
tions, Translations of Math. Monographs, Vol. 113, Amer. Math. Soc., Provi-
dence, 1992.