Semi-algebraic Geometry and generic

Hamiltonian stability
arXiv:2402.10707v1 [math.AG] 16 Feb 2024

S. Barbieri

Universitat de Barcelona
Gran Via de les Corts Catalanes, 585 - 08007 Barcelona, Spain
[email protected]

February 19, 2024

Mathematics subject classification: 37CXX - Smooth dynamical systems: gen-

eral theory, 14PXX - Real algebraic and real-analytic geometry.
Keywords: Hamiltonian dynamical systems, Nekhoroshev Theory, steepness, semi-
algebraic geometry.

Contents
1 Introduction 2

2 Main notations and definitions 13

3 Main results 15

4 The Thalweg and its properties 28

5 𝑠-vanishing polynomials 35

6 Proof of Theorem A 57

7 Proof of Theorem B and of its Corollaries 75

8 Partition of the set of 𝑠-vanishing polynomials 81

9 Proof of Theorems C1-C2-C3 94

1
A Semi-algebraic sets and semi-algebraic functions 106

B Analytic reparametrization of semi-algebraic sets 109

C Quantitative local inversion theorem 110

D Two elementary properties of Lie Groups 112

E Three auxiliary Lemmas 112

Abstract
The steepness property is a local geometric transversality condition on the gradient of
a 𝐶 2 -function which proves fundamental in order to ensure the stability of sufficiently-
regular nearly-integrable Hamiltonian systems over long timespans. Steep functions
were originally introduced by Nekhoroshev, who also proved their genericity. Namely,
given a pair of positive integers 𝑟, 𝑛, with 𝑟 high enough, and a point 𝐼0 ∈ ℝ𝑛 , the Taylor
polynomials of those 𝐶 2𝑟−1 functions which are not steep around 𝐼0 are contained in a
semi-algebraic set of positive codimension in the space of polynomials of 𝑛 variables
and degree bounded by 𝑟. The demonstration of this result was originally published in
1973 and has been hardly studied ever since, probably due to the fact that it involves
no arguments of dynamical systems: it makes use of quantitative reasonings of real-
algebraic geometry and complex analysis. The aim of the present work is two-fold. In
the first part, the original proof of the genericity of steepness is rewritten by making
use of modern tools of real-algebraic geometry: this allows to clarify the original rea-
sonings, that were obscure or sketchy in many parts. In particular, Yomdin’s Lemma
on the analytic reparametrization of semi-algebraic sets, together with non trivial esti-
mates on the codimension of certain algebraic varieties, turn out to be the fundamental
ingredients to prove the genericity of steepness. The second part of this work is entirely
new and is devoted to the formulation of explicit algebraic criteria to check steepness
of any given sufficiently regular function, which constitutes a very important result for
applications, as the original definition of steepness is not constructive. These criteria
involve both the derivatives of the studied function up to any given order and external
real parameters that, generically, belong to compact sets.

1 Introduction
1.1 Hamiltonian formalism and nearly-integrable systems
Hamiltonian formalism is the natural setting appearing in the study of many physical
systems. Namely, for any given positive integer 𝑛, we consider a symplectic manifold

2
 of dimension 2𝑛, endowed with a skew-symmetric two-form 𝜔, and with a func-
tion 𝐻 ∈ 𝐶 2 () classically called Hamiltonian. A Hamiltonian system on  is the
dynamical system governed by the vector field 𝑋𝐻 verifying

𝑖𝑋𝐻 𝜔 ∶= 𝜔(𝑋𝐻 , ⋅) = 𝑑𝐻 . (1.1)

In the simplest case, we consider the motion of a point on a 𝑛-dimensional Riemannian

manifold  endowed with the euclidean metric - called the configuration manifold -
under Newton’s second law
𝑞̈ = −∇𝑈 (𝑞) ,
where 𝑈 is a smooth potential function, and 𝑞 is a system of local coordinates for .
This system can be conjugated by duality due to Legendre’s transformation and reads

𝑝̇ = −𝜕𝑞 𝐻(𝑝, 𝑞) ; 𝑞̇ = 𝜕𝑝 𝐻(𝑝, 𝑞) (1.2)

where 𝐻(𝑝, 𝑞) is a real smooth function on the cotangent bundle 𝑇 ⋆ , and 𝑝 is the local
coordinate conjugated to 𝑞. In this example, if one takes  ≡ 𝑇 ⋆ , and if one chooses
∑
(𝑝, 𝑞) to be Darboux’s coordinates associated to the two-form 𝜔(𝑝, 𝑞) ≡ 𝑛𝑗=1 𝑑𝑝𝑖 ∧ 𝑑𝑞𝑖 ,
then system (1.2) is locally equivalent to (1.1).
Among Hamiltonian system, an important rôle is played by those which are inte-
grable by quadrature. Due to the classical Liouville-Arnol’d Theorem, under general
topological and algebraic assumptions, an integrable system depending on 2𝑛 variables
(𝑛 degrees of freedom) can be conjugated to a Hamiltonian system on the cotangent
bundle of the 𝑛-dimensional torus 𝕋 𝑛 , whose equations of motion take the form

𝐼̇ = −𝜕𝜗 ℎ(𝐼) = 0 , 𝜗̇ = 𝜕𝐼 ℎ(𝐼) ,

where (𝐼, 𝜗) ∈ ℝ𝑛 × 𝕋 𝑛 are called action-angle coordinates. Therefore, the phase space
of an integrable system is foliated by invariant tori carrying the linear motions of the
angular variables (called quasi-periodic motions).
Integrable systems are exceptional1 , but many important physical problems can be
described by Hamiltonian systems which are close to integrable. Namely, the dynamics
of a nearly-integrable Hamiltonian system is described by a Hamiltonian function whose
form in action-angle coordinates reads

𝐻(𝐼, 𝜗) ∶= ℎ(𝐼) + 𝜀𝑓 (𝐼, 𝜗) , (𝐼, 𝜗) ∈ ℝ𝑛 × 𝕋 𝑛

where 𝜀 is a small parameter that tunes the size of the perturbation 𝜀𝑓 w.r.t. the inte-
grable part ℎ.
The structure of the phase space of this kind of systems can be inferred with the help
of classical Kolmogorov-Arnol’d-Moser (KAM) theory. Namely, under the generic
1 Three examples of integrable systems are the classical Kepler’s problem, the harmonic oscillator, and

Lagrange’s top.

3
non-degeneracy condition that ∇ℎ is a local diffeomorphism, a Cantor-like set of pos-
itive Lebesgue measure of invariant tori carrying quasi-periodic motions for the inte-
grable flow persists under a suitably small perturbation (see e.g. ref. [3], [28]). As
this Cantor-like set is nowhere dense, it is extremely difficult to determine numerically
whether a given solution is quasi-periodic or not.
Moreover, for a Hamiltonian system depending on 𝑛 degrees of freedom (hence
a 2𝑛-dimensional system), the invariant tori provided by classical KAM theory are
𝑛-dimensional. Hence, if 𝑛 = 2, any pair of invariant tori disconnects the three-
dimensional energy level, so that the solutions of the perturbed system are global and
bounded over infinite times. However, an arbitrary large drift of the orbits is possible
in case 𝑛 ≥ 3. Actually, in ref. [2] Arnol’d proposed an example of a nearly-integrable
Hamiltonian system where an arbitrary large instability of the action variables occurs
for an arbitrarily small perturbation. This phenomenon is known under the name of
Arnol’d’s diffusion (see ref. [41] and references therein for the most recent develop-
ments in this field). Thus, results of stability for quasi integrable Hamiltonian systems
which are valid for an open set of initial condition can only be proved over finite times.

1.2 Long time stability of nearly-integrable systems

In the 1970s, Nekhoroshev2 proved that if we consider a real-analytic, integrable Hamil-
tonian whose gradient satisfies a suitable, quantitative transversality condition known
as steepness, then, for any sufficiently small perturbation the solutions of the perturbed
system are stable and have a very long time of existence3 .
The original definition of steepness given by Nekhoroshev is quite involved and
will be discussed at length in the sequel. In order to grasp an idea of what this property
means, it is worth mentioning that a real-analytic function is steep if and only if it has
no isolated critical points and if any of its restrictions to any affine proper subspace
admits only isolated critical points (see [40] and [51]). This is especially satisfied in
the important case of an integrable Hamiltonian which is strictly convex in the action
variables, since its critical points are non-degenerated.
Actually, the vast majority of the works on Nekhoroshev’s theory concerns small
perturbations of convex integrable Hamiltonian but Nekhoroshev also proved in [47]
that - unlike convexity - the steepness condition is generic, both in measure and topo-
logical sense. Namely, given a pair of positive integers 𝑟, 𝑛, with 𝑟 high enough, and
a point 𝐼0 ∈ ℝ𝑛 , the Taylor polynomials of those 𝐶 2𝑟−1 functions which are not steep
around 𝐼0 are contained in a semi-algebraic set of positive codimension in the space of
polynomials of 𝑛 variables and degree bounded by 𝑟. The proof and the refinement of
this property constitute the first part of the present work. However, before presenting
2 See [48], or [36] for a more modern presentation
3 The time of stability depends of the regularity of the considered system and is exponential (polynomial)
in the inverse of the size of the perturbation if the total Hamiltonian belongs to the Gevrey (Hölder) class.
See [45], [21], [11].

4
this and the other main results, we would like to highlight that - even though convex
systems are common 4 - non-convex integrable Hamiltonians occur in the investigation
of important problems of mechanics.
Namely, we consider a symplectic manifold (, Ω) of dimension 2𝑛, 𝑛 ∈ ℕ, where
Ω is an everywhere non-degenerate closed 2-form, a smooth symplectic vector field 𝑋
on  (meaning that the one-form 𝑖𝑋 Ω is closed) and an equilibrium point 𝑝∗ ∈ ,
that is 𝑋(𝑝∗ ) = 0.
We are interested in studying whether 𝑝∗ is stable or not.
Since we are in a conservative case, a first observation is that, if 𝑝∗ is stable, then
the spectrum of the linearized system around 𝑝∗ is {±𝑖𝛼1 , … , ±𝑖𝛼𝑛 } where 𝛼1 , … , 𝛼𝑛
are reals, so that 𝑝∗ is an elliptic equilibrium position.
The problem being local, we can assure without any loss of generality (this is spec-
ified in [22]) that we are working in (𝑀, Ω) = (ℝ2𝑛 , Ω0 ) where Ω0 ∶= 𝑑𝑥 ∧ 𝑑𝑦 is the
canonical symplectic structure of ℝ2𝑛 for the conjugated Darboux variables (𝑥, 𝑦) ∈
ℝ𝑛 × ℝ𝑛 . Moreover, under generic assumptions (see again [22]), we can assume that
the considered system derives from a Hamiltonian of the form:

∑
𝑛
𝐻(𝑥, 𝑦) = 𝛼𝑗 (𝑥2𝑗 + 𝑦2𝑗 )∕2 + 𝑂3 (𝑥, 𝑦). (1.3)
𝑗=1

Our standing assumption from now on is that 𝐻 is real-analytic. Such a system, under
a suitable rescaling, can be considered as nearly-integrable.
In this setting, there are two cases for which one knows that stability holds true for
the considered equilibrium.
The first case is when the quadratic part 𝐻2 is sign-definite, or, equivalently, when
the components of the vector 𝛼 ∈ ℝ𝑛 have the same sign. Indeed, the Hamiltonian
function has then a strict minimum (or maximum) at the origin, and as this function
is constant along the flow (it is in particular a Lyapounov function) one can construct,
using standard arguments, a basis of neighborhoods of the origin which are invariant,
and the latter property is obviously equivalent to stability.
The second case is when 𝑛 = 2 and when the so called Arnol’d’s iso-energetic non-
degeneracy condition is satisfied. Then, KAM stability occurs in every energy level
passing sufficiently close to the origin, implying Lyapounov stability, due to the fact
that the two-dimensional tori disconnect each three-dimensional energy level (see for
instance [1] and [46]). It is easy to see that the Arnold iso-energetic non-degeneracy
condition is generic in measure and topology as a function of the coefficients of the
𝑂4 (𝑥, 𝑦) part of the Taylor expansion of 𝐻 around the origin.
In the other cases, a large unstability due to Arnold diffusion can occur (see [32]),
but it has been proved in [22] that, generically, any solution starting sufficiently close
4 See e.g. [50], [12] in the study of the three-body problem, [5], [6] in the context of central force motions,

and [7], [4], [54], [31] in the framework of infinite-dimensional Hamiltonian systems.

5
to the equilibrium point remains close to it for an interval of time which is double-
exponentially large (exp ◦ exp) with respect to the inverse of the distance to the equi-
librium point. The latter result is obtained by making use of Nekhoroshev’s theory and
relies crucially on the genericity of steep functions, since one needs to build a suitable
steep integrable approximation of the complete system.
The same issue arises in order to apply Nekhoroshev’s theory to concrete exam-
ples. Especially, in Celestial Mechanics, we have important problems where an elliptic
equilibrium arises with a quadratic term in (1.3) which is not sign definite: this is the
case for the Lagrange’s equilibrium points L4, L5 in the restricted three body problem
(see [16]) and in the averaged ("secular") planetary three body problem (this is due to the
Herman’s resonance, see [33] and [44]). The latter system is a crucial approximation to
apply Hamiltonian perturbation theory (hence KAM or Nekhoroshev theory) in Celes-
tial Mechanics. Moreover, we cannot always build an integrable approximation of this
kind of systems which is convex in action variables, hence we have to consider steep non
convex Hamiltonians in order to infer stability results with the help of Nekhoroshev’s
theory. For the study of the Lagrange’s equilibrium points, it is possible in most cases
to recover steepness by considering higher order approximation (see [16]), actually this
corresponds to general considerations on functions with three variables which will be
specified in the sequel. For the secular planetary three-body problem, the associated
Hamiltonian is not convex w.r.t. the actions (see [53]) and much more variables are in-
volved than for the Lagrange’s points, hence we really need new criteria to ensure that
a given function is steep or not in this case. Up to now, generic explicit conditions for
steepness were known only for functions of three (the conditions given by Nekhoroshev
in [48]), four (see [57]) or five variables (see [10]). The second part of this work is de-
voted to proving explicit conditions for steepness which are generic for functions of an
arbitrary number of variables.
It can also be specified that if the steepness condition is dropped, large instablities
may occur over times of order 1∕𝜀, which is the shortest possible time of drift when
considering perturbations of magnitude 𝑂(𝜀) (see [51] and [23]).
In the context of KAM theory, Herman (see [38]) has shown that the lack of steep-
ness of the integrable Hamiltonian allows to build perturbations for which one can find
a 𝐺𝛿 −dense set of initial conditions leading to orbits whose action components are un-
bounded while the integrable Hamiltonian can also be Kolmogorov non-degenerate,
hence most of the orbits lie on invariant tori and we have simultanously, existence of
large zones of stability and instability.
Steepness also arises in the framework of Arnol’d’s diffusion (see [17]) for the op-
timality of the time of diffusion. Finally, in ref. [9] it is shown that Nekhoroshev’s
classical proof of stability for perturbations of steep integrable Hamiltonian systems
is also relevant in the study of PDE’s, considered as infinite-dimensional Hamiltonian
systems.

6
1.3 Genericity and explicit criteria for steepness

Now, we specify Nekhoroshev’s effective result of stability (see refs. [48], [49]), which
is valid for an open set of initial conditions provided that the total Hamiltonian is regular
enough and that its integrable part satisfies the following transversality property on its
gradient:

Definition 1.1 (Steepness). Fix 𝛿 > 0, 𝑅 > 0. A 𝐶 2 function ℎ ∶ 𝐵 𝑛 (0, 𝑅 + 2𝛿) → ℝ

is steep in 𝐵 𝑛 (0, 𝑅) with steepness indices α1 , … , α𝑛−1 ≥ 1 and steepness coefficients
𝐶1 , … , 𝐶𝑛−1 , 𝛿 if:

1. inf 𝐼∈𝐵𝑛 (0,𝑅) ||∇ℎ(𝐼)|| > 0;

2. for any 𝐼 ∈ 𝐵 𝑛 (0, 𝑅), for any integer 1 ≤ 𝑚 < 𝑛, and for any 𝑚-dimensional sub-
space Γ𝑚 orthogonal to ∇ℎ(𝐼) and endowed with the induced euclidean metric,
one has:

max min || 𝜋Γ𝑚 ∇ℎ(𝐼 + 𝑢) || > 𝐶𝑚 𝜉 α𝑚 , ∀𝜉 ∈ (0, 𝛿], (1.4)

0≤𝜂≤𝜉 𝑢∈Γ𝑚 , ||𝑢||2 =𝜂

where 𝜋Γ𝑚 stands for the orthogonal projection on Γ𝑚 .

Remark 1.1. Since in definition 1.1 the subspace Γ𝑚 ⊂ ℝ𝑛 is endowed with the induced
euclidean metric, for all 𝑢 ∈ Γ𝑚 one has ||𝜋Γ𝑚 ∇ℎ(𝐼 + 𝑢)|| = ||∇(ℎ|𝐼+Γ𝑚 )(𝐼 + 𝑢)||,
where ℎ|𝐼+Γ𝑚 indicates the restriction of ℎ to the affine subspace 𝐼 + Γ𝑚 .
As it is showed in [51], in the analytic case a function is steep if and only if, on any
affine hyperplane 𝐼 + Γ𝑚 , there exists no curve 𝛾 with one endpoint in 𝐼 such that the
restriction ∇(ℎ|𝐼+Γ𝑚 ) vanishes identically on 𝛾. From a heuristic point of view, for any
value 𝑚 ∈ {1, ..., 𝑛 − 1} the gradient ∇ℎ must "bend" towards 𝐼 + Γ𝑚 when "travelling"
along the curve 𝛾 ∈ 𝐼 + Γ𝑚 , so that critical points for the restriction of ℎ to 𝐼 + Γ𝑚 must
not accumulate.
Once the concept of steep function is introduced, Nekhoroshev’s effective result of
stability reads

Theorem 1.1 (Nekhoroshev, 1977). Consider a nearly-integrable Hamiltonian system

governed by

𝐻(𝐼, 𝜗) ∶= ℎ(𝐼) + 𝜀𝑓 (𝐼, 𝜗) , 𝐻 ∈ 𝐶 𝜔 (𝐵 𝑛 (0, 𝑟) × 𝕋 𝑛 ) ,

where 𝐵 𝑛 (0, 𝑟) is the open ball of radius 𝑟 in ℝ𝑛 , and ℎ is assumed to be steep. Then
there exist positive constants 𝑎, 𝑏, 𝜀0 , 𝐶1 , 𝐶2 , 𝐶3 such that, for any 𝜀 ∈ [0, 𝜀0 ) and for
any initial condition not too close to the boundary, one has |𝐼(𝑡) − 𝐼(0)| ≤ 𝐶2 𝜀𝑏 for
( )
any time 𝑡 satisfying |𝑡| ≤ 𝐶1 exp 𝐶3 ∕𝜀𝑎 .

Remark 1.2. The presence of a bound of the kind |𝐼(𝑡) − 𝐼(0)| ≤ 𝐶2 𝜀𝑏 on the variation
of the action variables is a consequence of the steepness property, whereas the time of

7
stability depends on the regularity of the function 𝐻 at hand. In the original formulation
by Nekhoroshev, 𝐻 is considered to be real-analytic, which yielded an exponentially-
long time in the inverse of the size of the pertubation (see also [36]). Exponentially-long
times of stability hold also in case 𝐻 is Gevrey (see [45]), whereas only polynomially-
long times of stability can be ensured for 𝐶 ∞ and Hölder functions (see refs. [8], [11]).
As it has already been anticipated in the previous paragraph, the steepness property
is generic - both in measure and in topological sense - in the space of Taylor polynomials
of sufficiently high order of smooth functions. Namely, let 𝑟, 𝑛 ≥ 2 be two positive
integers. We indicate by (𝑟, 𝑛) ⊂ ℝ[𝑥1 , … , 𝑥𝑛 ] the subspace of real polynomials in
𝑛 variables having degree bounded by 𝑟. For any point 𝐼0 ∈ ℝ𝑛 , and any function 𝑓
of class 𝐶 𝑟 near 𝐼0 , we call r-jet of 𝑓 at 𝐼0 the Taylor polynomial of 𝑓 up to order
𝑟 calculated at 𝐼0 . Moreover, we say that a subset  ⊂ ℝ𝑛 is semi-algebraic if it is
the finite union of subsets determined by a finite number of polynomial equalities or
inequalities (see Definition A.1). Nekhoroshev proved in [47]- [49] that

Theorem 1.2 (Nekhoroshev, 1973-1979). There exists a "bad" semi-algebraic subset

Ω(𝑟, 𝑛) of (𝑟, 𝑛) such that any function ℎ of class 𝐶 2𝑟−1 around a non-critical point
𝐼0 ∈ ℝ𝑛 , and whose 𝑟-jet at 𝐼0 lies outside of Ω(𝑟, 𝑛), is steep in a neighborhood of 𝐼0
with uniform indices. Moreover, the codimension of Ω(𝑟, 𝑛) in (𝑟, 𝑛) becomes positive
for 𝑟 ≳ [𝑛2 ∕4].

Although Nekhoroshev’s Theory has been a classic subject of study in the dynamical
systems community for more than forty years, the proof of Theorem 1.2 has remained
poorly understood. This is possibly due to the fact that such a demonstration does not
involve any arguments of dynamical systems, but combines quantitative reasonings of
real-algebraic geometry and complex analysis. Moreover, real-algebraic geometry was
at a more rudimentary level than nowadays at the time that Nekhoroshev’s was writ-
ing; for this reason, important properties of real-algebraic geometry are discovered5
in [47] at the same time that they are used to prove Theorem 1.2, which makes that
work obscure in many parts. In addition, the proofs of some lemmas in that work are
sketchy or presented in an old-fashioned way. For these reasons, the first part of this
work is devoted to proving and refining Theorem 1.2 by making use of modern results
of real-algebraic geometry. As we will discuss in detail in the sequel, Yomdin’s Lemma
about the analytic reparametrization of semi-algebraic sets (see [59]) turns out to be
the fundamental ingredient of real-algebraic geometry which is used in the proof of the
genericity of steepness.
Moreover, since the definition of steepness is not constructive, it is difficult to di-
rectly establish whether a given function is steep or not. Up to the author’s knowledge,
5 For example, it is remarkable that, up the author’s knowledge, a fundamental Bernstein’s inequality for

algebraic functions is proved for the first-time in Nekhoroshev’s work (see [13]). Such a result seems to have
passed unnoticed, until it has been widely rediscovered and generalized in the late nineties by Roytwarf and
Yomdin in [55], and subsequently developed by several authors.

8
there are only three articles on this topic (see [57], [29], [10]) which concern only poly-
nomials of low degree depending on a small number of variables. Actually, by develop-
ing the arguments used by Nekhoroshev to prove Theorem 1.2, it is possible to deduce
explicit sufficient algebraic conditions for steepness involving the derivatives up to an
arbitrary order of functions depending of an arbitrary number of variables. This proves
fundamental for applications of Nekhoroshev’s theory to physical models. The second
part of this work is dedicated to this topic. Namely, we will prove refined versions of
the Theorems below.

Theorem 1.3. The semi-algebraic set Ω(𝑟, 𝑛) in Theorem 1.2 verifies

( 𝑛−1 )
⋃
Ω(𝑟, 𝑛) = closure Proj(𝑟,𝑛) 𝑍(𝑟, 𝑚, 𝑛) (1.5)
𝑚=1

where, for any 𝑚 ∈ {1, … , 𝑛 − 1}, 𝑍(𝑟, 𝑚, 𝑛) is a semi-algebraic set of (𝑟, 𝑛) × ℝ𝐾 ×

ℝ𝑛 × 𝚄(𝑚 − 1, 𝑛), 𝐾 = 𝐾(𝑟, 𝑚) is a suitable positive integer, and 𝚄(𝑚 − 1, 𝑛) is the
compact 𝑚 − 1-dimensional Stiefel manifold in ℝ𝑛 (see section 2 for its definition).
Moreover, for any 𝑚 ∈ {1, … , 𝑛 − 1}, the form of 𝑍(𝑟, 𝑚, 𝑛) can be explicitly com-
puted.

Remark 1.3. Theorem 1.3 is a first example of an explicit criterion for steepness. In-
fact, as it is known, there exist general algorithms of real-algebraic geometry that allow
to compute the explicit form of the projection and of the closure of any given semi-
algebraic set (see e.g. [15]). Therefore, at least in principle, it is possible to com-
pute the r.h.s. of (1.5) - hence the form of Ω(𝑟, 𝑛) - as the expression of 𝑍(𝑟, 𝑚, 𝑛)
is known due to Theorem 1.3. However, the complexity of the classic algorithms grows
double-exponentially in the number of variables, so that they are of little use in practice
(see [37]).
Remark 1.4. Alternatively, one could use Theorem 1.3 in order to check if a func-
tion ℎ of class 𝐶 2𝑟−1 around a point 𝐼0 is steep in the following way. Indicating by
𝚃𝐼0 (ℎ, 𝑟, 𝑛) the 𝑟-jet of ℎ at 𝐼0 , by (1.5) one could check whether there exists 𝜏 > 0
such that, for any 𝑚 ∈ {1, … , 𝑛 − 1}, for any choice of parameters 𝛽 ∈ ℝ𝐾 × ℝ𝑛 ×
𝚄(𝑚 − 1, 𝑛), and for any polynomial 𝑆 ∈ (𝑟, 𝑛) which is 𝜏-close to 𝚃𝐼0 (ℎ, 𝑟, 𝑛), the
pair (𝑆, 𝛽) lies outside of 𝑍(𝑟, 𝑚, 𝑛). This would guarantee that 𝚃𝐼0 (ℎ, 𝑟, 𝑛) lies outside
(⋃ )
𝑛−1
of closure 𝑚=1 Proj (𝑟,𝑛) 𝑍(𝑟, 𝑚, 𝑛) , so that (1.5) and Theorem 1.2 would ensure
steepness. This is indeed one possibility, and we will make it more explicit in the next
section (see Theorem B). However, this criterion involves checking an explicit condi-
tion for a non-compact set of parameters (the first components of the vectors 𝛽 above
lie in ℝ𝐾 × ℝ𝑛 , whereas the remaining ones belong to the compact Stiefel manifold
𝚄(𝑚 − 1, 𝑛)). Nevertheless, as we show below, on "most subspaces" steepness can be
checked by making use of criteria involving only parameters belonging to a compact
set.

9
 Namely, let ℎ be a function of class 𝐶 2𝑟−1 around the origin, verifying ∇ℎ(0) ≠ 0.
Then,

Theorem 1.4. It is possible to find explicit algebraic criteria involving the derivatives
of ℎ up to order 𝑟 that ensure that ℎ is steep in a neighborhood of the origin on the
one-dimensional subspaces, with uniform index 𝜶 1 and uniform coefficients 𝐶1 , 𝛿.
Moreover, for any 𝑚 ∈ {2, … , 𝑛 − 1}, one has the following properties.

1. ℎ is steep at the origin with index 𝜶 𝑚 = 1 on the 𝑚-dimensional subspaces or-

thogonal to ∇ℎ(0) ≠ 0 on which the restriction of the hessian 𝐷2 ℎ(0) is non-
degenerate.

2. On the 𝑚-dimensional subspaces of ∇ℎ(0)⟂ on which the restriction of 𝐷2 ℎ(0)

has exactly one null eigenvalue, it is possible to construct explicit algebraic cri-
teria for steepness that involve the 𝑟-jet of ℎ at the origin and a finite number of
real parameters belonging to a compact subset. These criteria can be constructed
starting from the explicit form of subset 𝑍(𝑟, 𝑚, 𝑛) in Theorem 1.3 by the means
of algorithms involving only linear operations6 .

Therefore, explicit criteria for steepness at the origin involving only the 𝑟-jet of ℎ
exist in case 𝑚 = 1. In case 𝑚 ∈ {2, … , 𝑛 − 1}, instead, with the exception of the 𝑚-
dimensional subspaces of ∇ℎ⟂ (0) on which the restriction of 𝐷2 ℎ(0) has two or more
null eigenvalues, steepness can be checked by using a criterion which is simpler than
those stated in Remarks 1.3-1.4.
Moreover, for any value of 𝑚 ∈ {1, … , 𝑛 − 1}, the Hessian at the origin of a generic
function ℎ is non-degenerate on most subspaces of the 𝑚-dimensional Grassmannian
𝙶(𝑚, 𝑛), as the following result shows.

Theorem 1.5. Consider an integer 𝑚 ∈ {2, … , 𝑛 − 1}. For any bilinear, symmetric,
non-degenerate form 𝖡 ∶ ℝ𝑛 × ℝ𝑛 ⟶ ℝ, the 𝑚-dimensional subspaces on which the
restriction of 𝖡 is degenerate are contained in a submanifold of codimension one in the
Grassmannian 𝙶(𝑚, 𝑛).

1.4 Informal presentation of the proofs

Roughly speaking, the proof of Theorem 1.2 is split into two parts:

− General considerations of semi-algebraic geometry allow to prove that the com-

plicated condition (1.4) arising in the definition of steepness is an open property
in the space of polynomials (𝑟, 𝑛). Namely, if (1.4) holds for a given polyno-
mial 𝑄 ∈ (𝑟, 𝑛), then it holds also in a neighborhood of 𝑄 with uniform indices
𝜶 1 , … , 𝜶 𝑚 and uniform coefficients 𝐶1 , … , 𝐶𝑚 , 𝛿.
6 Hence, much simpler algorithms than the general algorithms of real-algebraic geometry.

10
 − Also, condition (1.4) is verified for given values of 𝜶 1 , … , 𝜶 𝑛−1 and for some
𝐶1 , … , 𝐶𝑛−1 , 𝛿 if the Taylor polynomial at 𝐼0 of the studied function ℎ lies outside
of the closure of a set of polynomials whose coefficients satisfy a certain number
of degeneracy conditions. A detailed analysis of these conditions shows that, for
sufficiently high 𝜶 1 , … , 𝜶 𝑛−1 , they only admit a non-generic set of solutions.
In the present work, we have results on the two sides of the proof.

1.4.1 Reparametrization of semi-algebraic sets and Bernstein’s inequality

We revisit Nekhoroshev’s reasonings of semi-algebraic geometry under the light of
more recent results in the field.
Due to (1.4), a central point to check steepness of a given function ℎ at a point 𝐼 ∈
𝑛
ℝ consists in ensuring a minimal growth of the projection of its gradient on any affine
subspace orthogonal to ∇ℎ(𝐼) ≠ 0. For a fixed affine subspace 𝐼 + Γ equipped with
local coordinates and with the induced euclidean metric, by Remark 1.1 the projection
of ∇ℎ(𝐼) on Γ corresponds to the gradient of the restriction ℎ|𝐼+Γ expressed in the
local coordinates. Hence, one is led to study the locus of minima of ||∇ℎ|𝐼+Γ ||. By
the above considerations, without entering into too many technicalities, a crucial step
in Nekhoroshev’s proof of the genericity of steepness consists in considering, for any
fixed polynomial 𝑃 ∈ ℝ[𝑋1 , … , 𝑋𝑚 ], the semi-algebraic set - called thalweg nowadays
(see [20]) - determined by :
{ }
ℝ𝑚 ⊃ 𝑃 ∶= 𝑢 ∈ ℝ𝑚 | ||∇𝑃 (𝑢)|| ≤ ||∇𝑃 (𝑣)|| ∀𝑣 ∈ ℝ𝑚 s.t. ||𝑢|| = ||𝑣|| . (1.6)
Nekhoroshev shows that 𝑃 contains the image of a semi-algebraic curve7 𝛾 which
admits a holomorphic extension with the exception of a finite set of singular complex
points whose cardinality depends only on the degree of 𝑃 and on the number of vari-
ables. In particular, one can ensure the existence of a uniform real interval of analyticity
and of a uniform complex analyticity width for the curve 𝛾, independently on the choice
of the polynomial 𝑃 ∈ (𝑟, 𝑚). More specifically, the graph of 𝛾 can be parametrized by
analytic-algebraic8 maps, and the existence of a Bernstein’s-like inequality controlling
uniformly the growth of this kind of functions in the complex plane ensures uniform
upper bounds on the derivatives of these charts (see [55], [59], [60], [13] and references
therein for a modern presentation).
Actually, this result about the thalweg in [47] is a particular case of a general the-
orem due to Yomdin [59] about analytic reparametrizations of semi-algebraic sets (the
finitely-differentiable case was firstly stated by Yomdin and Gromov in refs. [58], [35]
and then proved by Burguet in [27]). Generally speaking, the reparametrization of a
semi-algebraic set 𝐴 is a subdivision of 𝐴 into semi-algebraic pieces 𝐴𝑗 each of which
is the image of a semi-algebraic function9 of the unit cube. The uniform control on the
7 I.e.
a curve having semi-algebraic graph, see also Definition A.3.
8 I.e.
analytic maps whose graph solves a non-zero polynomial of two variables.
9 That is, a function whose graph is a semi-algebraic set, see also Definition A.3.

11
parametrization of the curve 𝛾 is unavoidable in [47], since it ensures that - for a smooth
function - steepness is an open property.
Moreover, it is proved that the coefficients of the Taylor expansions of non-steep
functions satisfy suitable algebraic conditions (one has a "finite-jet" determinacy of
steepness). In this way, the study of the genericity of steepness is reduced to a finite-
dimensional setting which involves polynomials of bounded order and this is another
crucial step in order to prove the genericity.
It is worth adding some remarks about the fact that Nekhoroshev proves a kind of
Bernstein’s inequality for algebraic functions (see [47], Lemma 5.1, p.446). Namely,
Nekhoroshev proves that an algebraic function which is real-analytic over a real inter-
val 𝐼 admits a bound on its growth on the complex plane which only depends on its
maximum over 𝐼 and on a constant depending on the degree of the polynomial solving
its graph and on the size of its complex domain of holomorphy. This result is proved
by exploiting the properties of algebraic curves of complex polynomials in two vari-
ables, and by making an intensive use of complex analysis. The original statements are
difficult to disentangle from the context of the genericity of steepness and the proofs
are very sketchy. This is different from the strategy used by Roytwarf and Yomdin
(see [55], [60] and references therein) which relies on arguments of analytic geometry.
Since we have not been able to find any reference that shows Nekhoroshev’s proof of
Bernstein’s inequality in detail except for the original paper [47], we have clarified and
extended Nekhoroshev’s reasonings in [13], and this allowed to obtain a simpler proof
on the existence of a Bernstein’s inequality for algebraic functions.
It is also worth mentioning that, in connection with arithmetic, the steepness con-
dition is introduced to prevent the abundance of rational vectors on certain sets and
it can be noticed that deep applications of the controlled analytic reparametrization of
semi-algebraic sets yield bounds on the number of integer points in semi-algebraic sets
(see [18] and [30]). In the future, these ideas may help to spread light on the connec-
tion between the stability of nearly-integrable Hamiltonian systems and the arithmetic
properties of semi-algebraic sets.
We also mention that, in the study of PDEs, the Yomdin-Gromov’s algebraic lemma
was used by Bourgain, Goldstein, and Schlag [25] to bound the number of integer points
in a two-dimensional semi-algebraic set.

1.4.2 Degeneracy condition

We describe heuristically the second part of the proof of Theorem 1.2. To make things
simple, we restrict this informal discussion to the case of a real-analytic function ℎ
around the origin. By formula (1.4), if ℎ is non-steep at the origin, then for some
𝑚 ∈ {1, … , 𝑛 − 1} there exists a 𝑚-dimensional subspace Γ𝑚 and a curve 𝛾 ⊂ Γ𝑚
starting at the origin on which the projection (𝜋Γ𝑚 ∇ℎ)|𝛾 is identically null. Assuming
that 𝛾 is sufficiently regular, this means that (𝜋Γ𝑚 ∇ℎ)|𝛾 has a zero of infinite order at

12
the origin. It can also be shown (see Theorem 4.1) that the the curve 𝛾 on which such
a condition is satisfied must possess a precise form. By these arguments, one can write
down explicitly the equations imposing to the derivatives of the function (𝜋Γ𝑚 ∇ℎ)|𝛾 to
be all identically null. Moreover, by performing complicated computations it is possi-
ble to check that these equations are all linearly independent. Then, estimates on the
codimension of a projected set show that, if one bounds the order of the derivatives that
are being considered in the equations by a positive integer 𝑟 ≥ 2, the Taylor polyno-
mials of non-steep functions belong to a semi-algebraic set of positive codimension in
the space of polynomials (𝑟, 𝑛). It is this kind of computations - which are expressed
explicitly for the first time in this work - that allow to prove Theorem 1.3.
Moreover, by construction, the equations that we are considering depend also on
the Taylor coefficients of the curve 𝛾 and on the vectors spanning the considered sub-
space Γ𝑚 . This explains the presence of the space of parameters ℝ𝐾 and of the Stiefel
manifold in the statement of Theorem 1.3. On the one hand, by suitably exploiting the
form of the equations, one can prove Theorem 1.4. On the other hand, Theorem 1.5 is
independent and its proof relies on the construction of a suitable system of coordinates
for the Grassmannian.

1.4.3 Structure of the work

This work is organized as follows: section 2 sets the main notations and definitions,
whereas the main results (refined versions of Theorems 1.2-1.3-1.4-1.5) are stated in
section 3. Section 4 construction of the thalweg and the reparametrization of semi-
algebraic sets, whereas section 5 is devoted to the study of the degeneracy condition
described in the paragraph above. Section 6 puts together the results of sections 4 and
5 in order to prove the genericity of steepness. Finally, sections 7-8-9 contain the proof
of the explicit criteria for steepness.

2 Main notations and definitions

Norms
For any 𝑛 ∈ ℕ⋆ , and for any 𝜈 ∈ ℕ⋆ ∪ {∞}, we denote by || ⋅ ||𝜈 the standard 𝓁 𝜈 -norm
in ℝ𝑛 .
For any integer 𝑞 ≥ 1, and for any open subset 𝐷 of ℝ𝑛 , the symbol 𝐶 𝑞 (𝐷) indicates
the set of 𝑞-times continuously differentiable maps 𝑓 ∶ 𝐷 → ℝ. Moreover, we indicate
by 𝐶𝑏𝑞 (𝐷) the subset of 𝐶 𝑞 (𝐷) containing those functions 𝑓 satisfying
‖𝑓 ‖𝐶 𝑞 (𝐷) ∶= sup sup |𝜕 𝛼 𝑓 (𝑥)| < +∞ . (2.1)
𝛼∈ℕ𝑛 𝑥∈𝐷
|𝛼|≤𝑞
( )
In particular, 𝐶𝑏𝑞 (𝐷), ‖⋅‖𝐶 𝑞 (𝐷) is a Banach space with multiplicative norm10 .
10 That is, satisfying an inequality of the form |𝑓 𝑔| ≤ 𝐾|𝑓 | |𝑔| for a suitable constant 𝐾.

13
Sets
In the sequel, we will make use of the following notations:

− For any 𝑑 > 0 and for any 𝑓 ∈ 𝐶𝑏𝑞 (𝐷), the symbol 𝕭𝑞 (𝑓 , 𝑑, 𝐷) indicates the
infinite-dimensional open ball of radius 𝑑 centered at 𝑓 for the norm (2.1);

− 𝜌 (𝑧0 ) indicates the open complex disc of radius 𝜌 > 0 centered at 𝑧0 ∈ ℂ;

− 𝐵 𝑛 (𝐼, 𝑅) indicates the real ball of radius 𝑅 > 0 centered at 𝐼 ∈ ℝ𝑛 .

− For any connected set  ⊂ ℂ, we denote the complex polydisk of width 𝑟 around
 by
{ }
′
()𝑟 ∶= 𝑧 ∈ ℂ s.t. ||𝑧 − ||2 ∶= inf′
||𝑧 − 𝑧 || < 𝑟 .
𝑧 ∈

Notations of real-algebraic geometry

For any pair (𝑟, 𝑛) of positive integers, and for any function ℎ of class 𝐶 𝑟 in a neighbor-
hood of some point 𝐼0 ∈ ℝ𝑛 , we denote by

− (𝑟, 𝑛) ⊂ ℝ[𝑋1 , … , 𝑋𝑛 ] the subspace of polynomials over the real field in 𝑛 real
variables with zero constant term and whose degree is bounded by 𝑟;

−  ⋆ (𝑟, 𝑛) ⊂ (𝑟, 𝑛) is the subset of those polynomials 𝑄 that verify ∇𝑄(0) ≠ 0;

− 𝚃𝐼0 (ℎ, 𝑟, 𝑛) ∈ (𝑟, 𝑛) the Taylor polynomial at order 𝑟 of the function ℎ(𝐼)−ℎ(𝐼0 )
centered at 𝐼0 .

Now, let 𝑘, 𝑛 be positive integers, with 𝑘 ≤ 𝑛.

− We indicate by 𝙶(𝑘, 𝑛) the 𝑘-dimensional Grassmanniann in ℝ𝑛 , i.e. the compact

manifold of 𝑘-dimensional linear subspaces in ℝ𝑛 .

− We also denote by 𝚄(𝑘, 𝑛) the compact 𝑘-dimensional Stiefel manifold in ℝ𝑛 , that

is the manifold of ordered orthonormal 𝑘-tuples of vectors in ℝ𝑛 .

For any set  ⊂ ℝ𝑛 × ℝ𝑚 , we indicate by Π𝑥  its projection onto the first 𝑛 coor-
dinates, that is the set

Π𝑥  ∶= {𝑥 ∈ ℝ𝑛 ∶ ∃𝑦 ∈ ℝ𝑚 |(𝑥, 𝑦) ∈ } .

Finally, let 𝑚, 𝑛 be positive integers satisfying 𝑚 ≤ 𝑛, and let {𝑢1 … , 𝑢𝑚 } be a set of lin-
early independent vectors in ℝ𝑛 . For each 𝑖 ∈ {1, … , 𝑛} and 𝑗 ∈ {1, … , 𝑚}, we denote
the 𝑖-th component of the vector 𝑢𝑗 by 𝑢𝑖𝑗 . For any multi-index 𝜇 = (𝜇1 , … , 𝜇𝑚 ) ∈ ℕ𝑚 ,

14
we set |𝜇| ∶= ||𝜇||1 . Given 𝐼0 ∈ ℝ𝑛 and a function ℎ of class 𝐶 |𝜇| in a neighborhood
of 𝐼0 , we also denote by
𝜇1 𝜇𝑚
|𝜇| [
⏞⏞⏞ ⏞⏞⏞ ]
ℎ𝐼 𝑢1 , … , 𝑢𝑚 ∶=
0

∑
𝑛
𝜕 |𝜇| ℎ(𝐼0 ) 𝑖 𝑖1𝜇 𝑖 𝑖𝑚𝜇 (2.2)
𝑢111 … 𝑢1 1 … 𝑢𝑚𝑚1 … 𝑢𝑚 𝑚
𝑖11 , … ,𝑖1𝜇 =1
𝜕𝐼𝑖11 … 𝜕𝐼𝑖1𝜇 … 𝜕𝐼𝑖𝑚1 … 𝜕𝐼𝑖𝑚𝜇
1 𝑚
… 1
𝑖𝑚1 , … ,𝑖𝑚𝜇𝑚 =1

the 𝜇-th order multilinear form associated to the 𝜇-th coefficient of the Taylor expansion
around 𝐼0 of the restriction of ℎ to Span (𝑢1 , … , 𝑢𝑚 ).

3 Main results
3.1 Genericity of steepness
As we discussed in Theorem 1.2 in the Introduction, the steepness property is generic,
both in measure and topological sense, in the space of jets of functions of sufficiently
high regularity. In this paragraph, we will give a more quantitative version of this result.
Namely, the statement below is a refined version of Nekhoroshev’s Theorem on the
genericity of steepness, which can be found in refs. [47]- [49].

Theorem (A). Let 𝑟, 𝑛 ≥ 2 be two integers, and let 𝚜 ∶= (𝑠1 , … , 𝑠𝑛−1 ) ∈ ℕ𝑛−1 be a
vector satisfying 1 ≤ 𝑠𝑚 ≤ 𝑟 − 1 for all 𝑚 = 1, … , 𝑛 − 1.
There exists a closed semi-algebraic subset 𝛀𝑟,𝚜
𝑛 of (𝑟, 𝑛) such that, for any 𝐼0 ∈
𝑛
ℝ𝑛 , for any real number 𝜚 > 0, for any open, bounded domain 𝒟 ⊂ 𝐶𝑏2𝑟−1 (𝐵 (𝐼0 , 𝜚)),
and for any function ℎ verifying

1. ℎ ∈ 𝒟 ,

2. ∇ℎ(𝐼0 ) ≠ 0 ,
|| ||
3. ||𝚃𝐼0 (ℎ, 𝑟, 𝑛) − 𝛀𝑟,𝚜 inf𝑟,𝚜 ||𝚃𝐼0 (ℎ, 𝑟, 𝑛) − 𝑄||∞ > 𝜏 > 0 ,
𝑛 ||||∞ ∶= 𝑄∈𝛀
|| 𝑛

one can introduce positive constants 𝜖 = 𝜖(𝑟, 𝚜, 𝜏, 𝑛), 𝑅 = 𝑅(𝑟, 𝚜, 𝜏, 𝑛, 𝜚), 𝐶𝑚 =

𝐶𝑚 (𝑟, 𝑠𝑚 , 𝜏, 𝑛) - where 𝑚 = 1, … , 𝑛 − 1 - and 𝛿 = 𝛿(𝑟, 𝚜, 𝜏, 𝑛, 𝜚, 𝒟 ) so that any function
𝑛
𝑓 ∈𝒟 , 𝑓 ∈ 𝔅2𝑟−2 (ℎ, 𝜀, 𝐵 (𝐼0 , 𝜚)) 𝜀 ∈ [0, 𝜖]

is steep in 𝐵 𝑛 (𝐼0 , 𝑅), with steepness coefficients 𝐶𝑚 , 𝛿 and with steepness indices bounded
by {
𝑠𝑚 , if 𝑚 = 1
𝜶 𝑚 (𝑠) ∶= (3.1)
2𝑠𝑚 − 1 , if 𝑚 ≥ 2 .

15
Moreover, in (𝑟, 𝑛) one has
{ }
codim 𝛀𝑟,𝚜
𝑛 ≥ max 0, min {𝑠𝑚 − 𝑚(𝑛 − 𝑚 − 1)} . (3.2)
𝑚∈{1,…,𝑛−1}

With the setting of Theorem (3.1), we give the following

𝑛
Definition 3.1. A function 𝑓 ∈ 𝐶𝑏2𝑟−1 (𝐵 (𝐼0 , 𝜚)) satisfying the hypotheses of Theorem
A is said to be steep at order 𝑟 at the point 𝐼0 for the vector 𝚜. The functions 𝑔 ∈ 𝒟
verifying
⋃
∇𝑔(𝐼0 ) ≠ 0 , 𝚃𝐼0 (𝑔, 𝑟, 𝑛) ∈ (𝑟, 𝑛)∖𝛀𝑟,𝚜
𝑛
𝚜∈ℕ𝑛−1
1≤𝑠𝑚 ≤𝑟−1
∀𝑚∈{1,…,𝑛−1}

are said to be steep at order 𝑟 at the point 𝐼0 .

With respect to the original result by Nekhoroshev, a few aspects are refined or
clarified in Theorem 3.1

1. The dependence of the steepness coefficients 𝐶𝑚 , 𝑚 ∈ {1, … , 𝑛 − 1}, and 𝛿 on

the distance 𝜏 to the bad set 𝛀𝑟,𝚜𝑛 , is made explicit. In particular, as it will be
shown in section 6, for fixed 𝑛, 𝑟, 𝚜, when 𝜏 → 0, then both 𝛿 → 0 and 𝐶𝑚 → 0
for all 𝑚 = 1, … , 𝑛 − 1, whereas the bounds 𝜶 𝑚 on the steepness indices are left
unchanged. Hence, when 𝜏 → 0, steepness may "break down" due to the steep-
ness coefficients tending to zero (whereas the steepness indices of those functions
whose 𝑟-jet lies outside of 𝛀𝑟,𝚜
𝑛 stay uniformly bounded).
It is important to stress that the above reasonings do not necessarily imply that a
𝑛
function 𝑔 ∈ 𝐶𝑏2𝑟−1 (𝐵 (𝐼0 , 𝜚)) whose 𝑟-jet satisfies 𝚃𝐼0 (𝑔, 𝑟, 𝑛) ∈ 𝛀𝑟,𝚜
𝑛 - for some
vector 𝚜 ∈ ℕ𝑛−1 as the one in Theorem A - is non-steep. For example, if for two
vectors 𝚜, 𝚜′ ∈ ℕ𝑛−1 , 𝚜′ ≠ 𝚜, having the same properties of the one in Theorem
′
A, one has ||𝚃𝐼0 (𝑔, 𝑟, 𝑛)−𝛀𝑟,𝚜 𝑟,𝚜
𝑛 ||∞ > 0, ||𝚃𝐼0 (𝑔, 𝑟, 𝑛)−𝛀𝑛 ||∞ = 0, then 𝑔 is steep
′
at order 𝑟 at 𝐼0 for the vector 𝚜 but not for the vector 𝚜. Hence, it admits different
bounds on the steepness indices than the functions whose jets lie outside of 𝛀𝑟,𝚜 𝑛 .
Therefore, the only result that one can infer from the relation 𝚃𝐼0 (𝑔, 𝑟, 𝑛) ∈ 𝛀𝑟,𝚜𝑛
is that, in case 𝑔 were steep, its steepness indices would admit a different upper
bound than the one in (3.1).
By Theorem A, definition 3.1, and by the above discussion, the set
⋂
𝛀𝑟,𝚜
𝑛 ⊂ (𝑟, 𝑛) (3.3)
𝚜∈ℕ𝑛−1
1≤𝑠𝑚 ≤𝑟−1
∀𝑚∈{1,…,𝑛−1}

contains the 𝑟-jets of all 𝐶𝑏2𝑟−1 functions around the non-critical point 𝐼0 which
are non-steep at order 𝑟 at 𝐼0 . In the same way, when 𝑟 → +∞, the Taylor

16
 coefficients of all non-steep analytic functions at the non-critical point 𝐼0 belong
to the set
⋂∞ ⋂ ⋃∞
𝜔𝑛 (𝐼0 ) ∶= 𝛀𝑟,𝚜
𝑛 ⊂ (𝑟, 𝑛) . (3.4)
𝑟=2 𝚜∈ℕ𝑛−1 𝑟=2
1≤𝑠𝑚 ≤𝑟−1
∀𝑚∈{1,…,𝑛−1}

Relation (3.4) is the explicit version of what Nekhoroshev meant when he wrote
"Hamiltonians which fail to be steep at a noncritical point are infinitely singular:
they satisfy an infinite number of independent conditions on the Taylor coeffi-
cients" (see [47], p. 426).
Actually, the strategy of proof given in the present work follows Nekhoroshev’s
reasonings by showing - for any given 𝑄 ∈ (𝑟, 𝑛)- the existence of an arc,
whose image is contained in the thalweg 𝑄 (see Definition 1.6), admitting a
fitted parametrization whose derivatives are controlled by constants depending
only on the number of variables 𝑛 and on the degree 𝑟, but not on 𝑄. This is a
particular occurrence of the fact that - with the exception of small neighborhoods
around singularities - a semi-algebraic set can be reparametrized by a collection
of images of holomorphic functions having a domain of analyticity and an upper
bound on their derivatives which depend only on the number of variables, and on
the number and on the degrees of the polynomials involved in the construction
(see [59] and [61] for a two-dimensional semi-algebraic set, and [18], [30] for
higher dimensional sets with more general properties than semi-algebraicness).
The considered analytic reparametrization is a partial extension of a theorem
(called algebraic lemma) due to Yomdin [58] and Gromov [35] which ensures
that, for any semi-algebraic set, there exists a collection of 𝐶 𝑘 −mappings which
parametrize entirely the considered set. Actually, in our proof (see Theorem 4.1),
we ask for the mappings to be holomorphic, and we must remove small neighbor-
hoods around their singularities; also, the use of 𝐶 𝑘 −reparametrizations would
not allow to obtain a characterization of non-steep functions as in 3.4, where a
control of all the derivative up to infinity is needed.

2. The vector 𝚜 does not appear in the original statement. Indeed, Nekhoroshev
implicitly sets
{ }
⎧max 1, 𝑟 − 1 + 𝑚(𝑛 − 𝑚 − 1) − 𝑛(𝑛 − 2) , for 𝑛 even
⎪ 4 }
𝑠𝑚 = 𝑠𝑚 ∶= ⎨ {
(𝑛 − 1)2
⎪max 1, 𝑟 − 1 + 𝑚(𝑛 − 𝑚 − 1) − , for 𝑛 odd .
⎩ 4
(3.5)
From a heuristic point of view, in ref. [49] this choice was probably conceived
in the following way: in estimate (3.2), one may want to get rid of the quantity
𝑚(𝑛 − 𝑚 − 1), which attains the maximal value 𝑛(𝑛 − 2)∕4 for 𝑚 = 𝑛∕2 when 𝑛

17
 is even, and (𝑛 − 1)2 ∕4 for 𝑚 = (𝑛 − 1)∕2 when 𝑛 is odd. Hence, (3.5) is the best
choice which allows to get rid of the term −𝑚(𝑛 − 𝑚 − 1) in (3.2) and which still
guarantees the essential condition 1 ≤ 𝑠𝑚 ≤ 𝑟 − 1.
Choice (3.5), in our case, yields (see (3.1)-(3.2))
{ }
⎧ 𝑛(𝑛 − 2)
⎪max 1, 𝑟 − 1 + 𝑚(𝑛 − 𝑚 − 1) − , for 𝑛 even, 𝑚 = 1
⎪ { 4 }
2
⎪max 1, 𝑟 − 1 + 𝑚(𝑛 − 𝑚 − 1) − (𝑛 − 1) , for 𝑛 odd, 𝑚 = 1
⎪ 4
𝜶 𝑚 (𝑠𝑚 ) ∶= ⎨ { }
𝑛(𝑛 − 2)
⎪max 1, 2𝑟 − 3 + 2𝑚(𝑛 − 𝑚 − 1) − , for 𝑛 even, 𝑚 ≥ 2
⎪ { 2 }
⎪ (𝑛 − 1)2
⎪max 1, 2𝑟 − 3 + 2𝑚(𝑛 − 𝑚 − 1) − , for 𝑛 odd, 𝑚 ≥ 2 ,
⎩ 2
(3.6)
and { }
⎧max 0, 𝑟 − 1 − 𝑛(𝑛 − 2) , if 𝑛 is even
⎪ 4 }
codim 𝛀𝑟,𝚜 ≥⎨ { (3.7)
𝑛 (𝑛 − 1)2
⎪max 0, 𝑟 − 1 − , if 𝑛 is odd .
⎩ 4
For 𝑚 = 1, the bound (3.6) on the steepness indices is half of the one found
in [49]. For 𝑚 ≥ 2, the estimates in [49] coincide with (3.6). Estimate (3.7) on
the codimension coincides with the one in [49].

3. Theorem 3.1 holds true even for functions in the class 𝐶𝑏𝑟+1 . In that case, one
considers 1 ≤ 𝑠𝑚 ≤ ⌊𝑟∕2⌋ for any 𝑚 = 1, … , 𝑛 − 1, which yields worse estimates
both on the indices of steepness and on the codimension. This is the case which
was considered in [47], whereas the regularity 𝐶𝑏2𝑟−1 was introduced in [49].

3.2 Explicit algebraic criteria for steepness

The kind of genericity stated in Theorem A implies the classic notions of genericity in
topological and in measure sense. However, Theorem A alone is not sufficient when
dealing with applications of Nekhoroshev’s Theory to physical models. Infact, in or-
der to infer long-time stability of a sufficiently regular integrable model Hamiltonian ℎ
under any sufficiently small and regular perturbation, one needs to have a criterion to
check at which points of its domain the given function ℎ is steep. As we shall show in
the sequel, establishing a criterion of this kind is a non-trivial developement of the proof
of the genericity of steepness. Namely, in the rest of this section we will present explicit
algebraic criteria for steepness which involve the Taylor coefficients at any order of the
studied function.
In order to give rigorous statements, we firstly need to introduce some notations.

18
3.2.1 Some additional notations
For any pair of positive integers 𝑛 ≥ 2 and 𝑘 ∈ {2, … , 𝑛}, we introduce the notation

𝒱 1 (𝑘, 𝑛) ∶= {(𝑣, 𝑢2 , … , 𝑢𝑘 ) ∈ ℝ𝑛 × 𝚄(𝑘 − 1, 𝑛)| rk(𝑣, 𝑢2 , … , 𝑢𝑘 ) = 𝑘} . (3.8)

For fixed integers 𝑚 ≥ 2 and 𝛼 ≥ 0, for any 𝛽 ∈ {0, ..., 𝛼}, and for any 𝑖 ∈ {1, ..., 𝑚},
we also define the family of multi-indices

{
(𝛽 + 1, 0, ..., 0) , for 𝑖 = 1
ℕ𝑚 ∋ 𝜈(𝑖, 𝛽) ∶= (3.9)
(𝛽, 0, ..., 0, 1, 0, ..., 0) , for 𝑖 = 2, ..., 𝑚

where the "1" fills the 𝑖-th slot for 𝑖 = 2, ..., 𝑚. When 𝛼 ≥ 1, we denote the multi-indices
𝜇 ∈ ℕ𝑚 of length 2 ≤ |𝜇| ∶= ||𝜇||1 ≤ 𝛼 + 1 not belonging to the family (3.9) with
⋃
𝑚 (𝛼) ∶= {𝜇 ∈ ℕ𝑚 , 2 ≤ |𝜇| ≤ 𝛼 + 1}∖ {𝜈(𝑖, 𝛽)} . (3.10)
𝑖=1,...,𝑚
𝛽=1,...,𝛼

Moreover, for given values of 𝛼 ∈ ℕ, 𝜇 = (𝜇1 , … , 𝜇𝑚 ) ∈ ℕ𝑚 and 𝓁 ∈ {1, ..., 𝑚},

we indicate by 𝜇̃(𝓁) = (̃
𝜇1 (𝓁), … , 𝜇̃𝑚 (𝓁)) the multi-index for which 𝜇̃𝑖 (𝓁) ∶= 𝜇𝑖 − 𝛿𝑖𝓁 ,
𝑖 = 1, … , 𝑚, and we introduce the sets
{
𝜇(𝓁), 𝛼) ∶= 𝑘 ∶= (𝑘22 , … , 𝑘2𝛼 , … , 𝑘𝑚2 , … , 𝑘𝑚𝛼 ) ∈ ℕ(𝑚−1)×(𝛼−1) ∶
𝑚 (̃
} (3.11)
∑
𝛼 ∑
𝑚 ∑
𝛼
𝑘𝑗𝑖 = 𝜇̃𝑗 (𝓁) , 𝜇̃1 (𝓁) + 𝑖 𝑘𝑗𝑖 = 𝛼 ,
𝑖=2 𝑗=2 𝑖=2

𝑚 (𝓁, 𝛼) ∶= {𝜇 ∈ ℕ𝑚 | 𝑚 (̃
𝜇 (𝓁), 𝛼) ≠ ∅} . (3.12)
For any 𝑘 ∈ 𝑚 (̃ 𝜇 (𝓁), 𝛼), we write 𝑘! ∶= 𝑘22 ! … 𝑘2𝛼 ! … 𝑘𝑚2 ! … 𝑘𝑚𝛼 ! and, simi-
larly, for any 𝜇 ∈ ℕ𝑚 , we use the notation 𝜇! ∶= 𝜇1 ! … 𝜇𝑚 ! .
Finally, we consider a quadruplet of positive integers 𝑟 ≥ 2, 𝑛 ≥ 3, 1 ≤ 𝑠 ≤ 𝑟 − 1,
2 ≤ 𝑚 ≤ 𝑛 − 1, a point 𝐼0 ∈ ℝ𝑛 , and a function ℎ of order 𝐶 𝑟 around 𝐼0 .
With this setting, for any given 𝛼 ∈ {1, … , 𝑠}, and 𝓁 ∈ {1, … , 𝑚}, we introduce
the functions
ℎ,𝐼0
𝑚,𝓁,𝛼 ∶ 𝒱 1 (𝑚, 𝑛) × ℝ(𝑚−1)𝑠 → ℝ

associating to any element (𝑣, 𝑢2 , … , 𝑢𝑚 ) ∈ 𝒱 1 (𝑚, 𝑛) and to any vector 𝑎(𝑚, 𝑠) =

(𝑎21 , … , 𝑎2𝑠 , … , 𝑎𝑚1 , … , 𝑎𝑚𝑠 ) ∈ ℝ(𝑚−1)×𝑠 the following quantities

19
 ℎ,𝐼
0
𝑚,1,1 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚, 𝑠)) ∶= ℎ2𝐼 [𝑣, 𝑣] , 𝛼=1, 𝓁=1
0
ℎ,𝐼0 2
[ ]
𝑚,𝓁,1 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚, 𝑠)) ∶= ℎ𝐼 𝑣, 𝑢𝓁 , 𝛼 = 1 , 𝓁 ∈ {2, … , 𝑚}
0
ℎ,𝐼0
𝑚,1,2 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚, 𝑠)) ∶= ℎ3𝐼 [𝑣, 𝑣, 𝑣] , if 𝑠 ≥ 2 , for 𝛼 = 2 , and 𝓁 = 1
0

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
if 𝑠 ≥ 3, for 𝛼 ∈ {3, … , 𝑠}, and 𝓁 = 1

ℎ,𝐼
0
𝑚,1,𝛼 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚, 𝑠)) ∶=
𝛽
1 𝛼+1 [ ∑
] 𝛼−1 1 [ ⏞⏞⏞ ∑ 𝑚
]
ℎ𝐼 𝑣, … , 𝑣 + ℎ𝛽+1
𝐼
𝑣 , 𝑎𝑖(𝛼−(𝛽−1)) 𝑢𝑖
𝛼! 0 𝛽=1
(𝛽 − 1)! 0 𝑖=2
𝜇1 −1 𝑘22 𝑘2𝛼 𝑘𝑚2 𝑘𝑚𝛼
|𝜇| [
⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ]
∑ ∑ ℎ𝐼
0
𝑣 , 𝑎22 𝑢2 , … , 𝑎2𝛼 𝑢2 , … , 𝑎𝑚2 𝑢𝑚 , … , 𝑎𝑚𝛼 𝑢𝑚 , 𝑣
+
𝜇∈𝑚 (1,𝛼) 𝑘∈𝑚 (̃
𝜇(1),𝛼)
(𝜇1 − 1)! 𝑘!
𝜇∈𝑚 (𝛼)
𝜇1 ≠0
(3.13)

if 𝑠 ≥ 2, for 𝛼 ∈ {2, … , 𝑠}, and 𝓁 ∈ {2, … , 𝑚}

𝛼
ℎ,𝐼0 1 [ ⏞⏞⏞ ]
𝑚,𝓁,𝛼 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚, 𝑠)) ∶= ℎ𝛼+1
𝐼
𝑣 , 𝑢𝓁
𝛼! 0
𝜇1 𝑘22 𝑘2𝛼 𝑘𝑚2 𝑘𝑚𝛼
|𝜇| [
⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ]
∑ ∑ ℎ𝐼
0
𝑣 , 𝑎22 𝑢2 , … , 𝑎2𝛼 𝑢2 , ..., 𝑎𝑚2 𝑢𝑚 , ..., 𝑎𝑚𝛼 𝑢𝑚 , 𝑢𝓁
+ .
𝜇∈𝑚 (𝓁,𝛼) 𝑘∈𝑚 (̃
𝜇(𝓁),𝛼)
𝜇1 ! 𝑘!
𝜇∈𝑚 (𝛼)
𝜇𝓁 ≠0
(3.14)

3.2.2 Theorem B and related corollaries

With the setting above, we can state the first explicit criterion for steepness. Its Corol-
lary B2 is a refined version of Theorem 1.3.

Theorem (B). Let 𝑟, 𝑛 ≥ 2 be two integers, and let 𝚜 ∶= (𝑠1 , … , 𝑠𝑛−1 ) ∈ ℕ𝑛−1 be a
vector verifying 1 ≤ 𝑠𝑖 ≤ 𝑟 − 1 for all 𝑖 = 1, … , 𝑛 − 1. Consider a point 𝐼0 ∈ ℝ𝑛 , a
real number 𝜚 > 0, and a function ℎ of class 𝐶𝑏2𝑟−1 (𝐵 𝑛 (𝐼0 , 𝜚)) verifying ∇ℎ(𝐼0 ) ≠ 0.

20
i) If the system
{
𝑤 ∈ 𝕊𝑛
𝑠 +1 (3.15)
ℎ1𝐼 [𝑤] = ℎ2𝐼 [𝑤, 𝑤] = ⋯ = ℎ𝐼1 [𝑤, … , 𝑤] = 0
0 0 0

has no solution11 ,
then ℎ is steep around the point 𝐼0 on the affine subspaces of dimen-
sion one, with steepness index bounded by 𝑠1 .
In the sequel, we set 𝑁 = 𝑁(𝑟, 𝑛) ∶= dim (𝑟, 𝑛).
ii) If, for some 𝑚 ∈ {2, … , 𝑛 − 1}, there exists 𝑅𝑚 > 0 such that for any polynomial
𝑆 ∈ 𝐵 𝑁 (𝚃𝐼0 (ℎ, 𝑟, 𝑛), 𝑅𝑚 ) the system

⎧(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛)
⎪
⎪𝑎(𝑚, 𝑠 ) ∶= (𝑎 , … , 𝑎 , … , 𝑎 , … , 𝑎 ) ∈ ℝ(𝑚−1)×𝑠𝑚
⎪ 𝑚 21 2𝑠𝑚 𝑚1 𝑚𝑠𝑚
⎪ ∑𝑚
⎨𝑣 = 𝑢1 + 𝑗=2 𝑎𝑗1 𝑢𝑗 (3.16)
⎪
⎪𝑆𝐼1 [𝑣] = 𝑆𝐼1 [𝑢2 ] = ⋯ = 𝑆𝐼1 [𝑢𝑚 ] = 0
⎪ 0 0 0
⎪∑𝑚 ∑𝑠𝑚 || 𝑆,𝐼0 (𝑣, 𝑢 , … , 𝑢 , 𝑎(𝑚, 𝑠 ))|| = 0
⎩ 𝓁=1 𝛼=1 | 𝑚,𝓁,𝛼 2 𝑚 𝑚 |

has no solution, then ℎ is steep in a neighborhood of 𝐼0 on the affine subspaces of

dimension 𝑚, with steepness index bounded by 𝜶 𝑚 ≤ 2𝑠𝑚 − 1 .
Though the quantities which are involved in Theorem B are quite cumbersome, the
idea behind the result is not difficult to grasp: condition (3.16) amounts to asking that
there exists a neighborhood of polynomials around the 𝑟-jet of ℎ(𝐼) − ℎ(𝐼0 ) which lies
outside of the closed semi-algebraic set 𝛀𝑟,𝚜𝑛 defined in Theorem A. This will be made
clearer in Corollary B2.
As it will be discussed in the technical sections of the present work (see e.g. sec-
tion 5), for any 𝑚 ∈ {2, … , 𝑛 − 1}, the real parameters 𝑎21 , … , 𝑎2𝑠𝑚 , … , 𝑎𝑚1 , … , 𝑎𝑚𝑠𝑚
appearing in (3.13)-(3.14) and in the statement of Theorem B represent the Taylor co-
efficients of analytic curves of the type
{ ∑
𝛾(𝑡) ∶= 𝑥1 (𝑡) ∶= 𝑡 , 𝑥𝑗 (𝑡) ∶= +∞𝑖=1 𝑎𝑗𝑖 𝑡
𝑖 𝑗 = 2, … , 𝑚 (3.17)

which, for any 𝑚-dimensional affine subspace 𝐼0 + Γ𝑚 , contain the locus of minima
of the projection ||𝜋Γ𝑚 ∇𝚃𝐼0 (ℎ, 𝑟, 𝑛)|| appearing in (1.4). For any given Γ𝑚 and for any
regular function ℎ around 𝐼0 , the existence of a minimal curve of the form (3.17) is
ensured by Theorem 4.1 in the sequel.
Theorem B comes together with important corollaries.
The following one is well-known: its statement can be found in [48], whereas its
proof can be found in [29]. As we shall see, in our context it is a simple consequence
of Theorem B.
11 In this case, ℎ is said to be 𝑠1 + 1-jet non-degenerate at the origin.

21
Corollary (B1). Consider an integer 𝑛 ≥ 2, a point 𝐼0 ∈ ℝ𝑛 , and a function ℎ of class
𝐶 5 around 𝐼0 , satisfying ∇ℎ(𝐼0 ) ≠ 0. If the system
{
𝑤 ∈ 𝕊𝑛
(3.18)
ℎ1𝐼 [𝑤] = ℎ2𝐼 [𝑤, 𝑤] = ℎ3𝐼 [𝑤, 𝑤, 𝑤] = 0 (3-jet non degeneracy)
0 0 0

has no solution, then ℎ is steep in a neighborhood of 𝐼0 , and its indices satisfy

𝛼1 = 2 , max {𝛼𝑚 } ≤ 3 .
𝑚=2,…,𝑛−1

Remark 3.1. Actually, as a more careful analysis of 3-jet non-degenerate functions

shows, the result is true for 𝐶 4 functions and one can take max𝑚=1,…,𝑛−1 {𝛼𝑚 } ≤ 2
(see [29]).
Thanks to Theorem B, moreover, we have a more explicit characterisation of the sets
𝛀𝑟,𝚜
𝑛 appearing in the statement of Theorem A. Namely, as it was the case in the hypothe-
ses of Theorem B, we consider two integers 𝑟, 𝑛 ≥ 2, and a vector 𝚜 ∶= (𝑠1 , … , 𝑠𝑛−1 ) ∈
ℕ𝑛−1 , with 1 ≤ 𝑠𝑖 ≤ 𝑟 − 1 for all 𝑖 = 1, … , 𝑛 − 1. Also, we take a point 𝐼0 ∈ ℝ𝑛 . This
time, differently to what we did in Theorem B, we do not consider a fixed function.
𝑟,𝑠1 ,1
Corollary (B2). For 𝑛 ≥ 2, and 𝑚 = 1, we indicate by 𝑛 the algebraic set of
 ⋆ (𝑟, 𝑛) × 𝕊𝑛 determined by
{
𝑤 ∈ 𝕊𝑛 , 𝑄 ∈  ⋆ (𝑟, 𝑛)
𝑠 +1 (3.19)
𝑄1𝐼 [𝑤] = 𝑄2𝐼 [𝑤, 𝑤] = ⋯ = 𝑄𝐼1 [𝑤, … , 𝑤] = 0 .
0 0 0

𝑟,𝑠𝑚 ,𝑚
For 𝑛 ≥ 3, and for any given 𝑚 ∈ {2, … , 𝑛 − 1}, we denote by 𝑛 the algebraic set
of  ⋆ (𝑟, 𝑛) × ℝ(𝑚−1)𝑠𝑚 × 𝒱 1 (𝑚, 𝑛) determined by

⎧(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛) , 𝑄 ∈  ⋆ (𝑟, 𝑛)

⎪
⎪𝑎(𝑚, 𝑠𝑚 ) ∶= (𝑎21 , … , 𝑎2𝑠𝑚 , … , 𝑎𝑚1 , … , 𝑎𝑚𝑠𝑚 ) ∈ ℝ(𝑚−1)×𝑠𝑚
⎨ ∑𝑚 (3.20)
⎪𝑣 = 𝑢1 + 𝑗=2 𝑎𝑗1 𝑢𝑗
⎪∑𝑚 ∑𝑠𝑚 || 𝑄,𝐼0 (𝑣, 𝑢 , … , 𝑢 , 𝑎(𝑚, 𝑠 ))|| = 0 .
⎩ 𝓁=1 𝛼=1 | 𝑚,𝓁,𝛼 2 𝑚 𝑚 |

With this setting, one has

⋃
𝑛−1
( 𝑟,𝑠 ,𝑚 )
closure Π ⋆ (𝑟,𝑛) 𝑛 𝑚 = 𝛀𝑟,𝚜
𝑛 , (3.21)
𝑚=1

where 𝛀𝑟,𝚜
𝑛 is the semi-algebraic set introduced in Theorem A.
𝑟,𝑠 ,𝑚
Remark 3.2. Since the sets 𝑛 𝑚 , 𝑚 = 1, … , 𝑛−1 in Corollary 3.2.2 are algebraic, the
Theorem of Tarski and Seidenberg (see Th. A.1) - together with expression (3.21) and
Proposition A.2 - ensures that 𝛀𝑟,𝚜
𝑛 is a semi-algebraic set of (𝑟, 𝑛), as we already knew

22
by Theorem A. Moreover, it is worth to notice that - at least in principle - one could
find the explicit expression for 𝛀𝑟,𝚜𝑛 . Infact, the Theorem of Tarski and Seidenberg
is somewhat "constructive", in the sense that there exist algorithms that allow to find
the explicit expression for the projection and the closure of any semi-algebraic set (see
e.g. [15]). However, these general algorithms are not very useful in applications, as
their complexity grows double-exponentially with the number of the involved variables
(see [37]). As we shall show in Theorems C1-C2 below, in "most cases" (in a sense
𝑟,𝑠 ,𝑚
that will be clarified in Theorem C3) the sets 𝑛 𝑚 , with 𝑚 = 2, … , 𝑛 − 1 can be
projected onto  ⋆ (𝑟, 𝑛) × 𝒱 1 (𝑚, 𝑛) with the help of a simple algorithm involving only
linear operations. Moreover, such a projection yields a closed semi-algebraic set of
 ⋆ (𝑟, 𝑛) × 𝒱 1 (𝑚, 𝑛). This implies a further criterion to check steepness of a given
function.
Finally, using Theorem B we can state a sufficient condition for non-steepness at a
given point, namely

Corollary (B3). Consider a point 𝐼0 ∈ ℝ𝑛 , and a function ℎ in the real-analytic class

around 𝐼0 verifying ∇ℎ(𝐼0 ) ≠ 0.
If at least one of the two following conditions is satisfied, then ℎ is non-steep at 𝐼0 .

1. There exists 𝑤 ∈ 𝕊𝑛 such that ℎ𝑟𝐼 [𝑤] = 0 , ∀𝑟 ∈ ℕ .

0

2. For some 𝑚 ∈ {2, … , 𝑛 − 1}, there exist

(a) 𝑚 − 1 real sequences {𝑎𝑗𝑖 }𝑖∈ℕ , 𝑗 = 2, … , 𝑚 and a number 𝚝 > 0 such that
∑
the expansions +∞ 𝑖
𝑖=1 𝑎𝑗𝑖 𝑡 admit a radius of convergence greater than 𝚝 for
all 𝑚 ;
(b) 𝑚 linearly-independent vectors 𝑣, 𝑢2 , … , 𝑢𝑚 ∈ 𝒱 1 (𝑚, 𝑛) ;

such that for all integer 𝑟 ≥ 2 the following system is satisfied:

⎧(𝑢 , … , 𝑢 ) ∈ 𝚄(𝑚, 𝑛)
⎪ 1 𝑚
⎪𝑎(𝑚, 𝑟 − 1) ∶= (𝑎21 , … , 𝑎2(𝑟−1) , … , 𝑎𝑚1 , … , 𝑎𝑚(𝑟−1) ) ∈ ℝ(𝑚−1)×(𝑟−1)
⎪ ∑𝑚
⎨𝑣 = 𝑢1 + 𝑗=2 𝑎𝑗1 𝑢𝑗
⎪ℎ1 [𝑣] = ℎ1 [𝑢 ] = ⋯ = ℎ1 [𝑢 ] = 0
⎪ 𝐼0 𝐼0 2 𝐼0 𝑚
⎪∑𝑚 ∑𝑟−1 || ℎ,𝐼0 (𝑣, 𝑢 , … , 𝑢 , 𝑎(𝑚, 𝑠 ))|| = 0 .
⎩ 𝓁=1 𝛼=1 | 𝑚,𝓁,𝛼 2 𝑚 𝑚 |
(3.22)

Remark 3.3. Since we consider any 𝑟 ≥ 2, we have an infinite system.

3.2.3 Theorems C1-C2-C3

As we have showed above, Theorem B constitutes an explicit criterion for steepness
which, however, for any given value of 𝑛 ≥ 3, 𝑚 ∈ {2, … , 𝑛−1} and 𝑠𝑚 ∈ {1, … , 𝑟−1}

23
depends on the additional parameters 𝑎21 , … , 𝑎2𝑠𝑚 , … , 𝑎𝑚1 , … , 𝑎𝑚𝑠𝑚 ∈ ℝ(𝑚−1)𝑠𝑚 and
on the vectors 𝑣, 𝑢2 , … , 𝑢𝑚 ∈ 𝒱 1 (𝑚, 𝑛). As we have already pointed out in Remark 3.2,
it is possible in principle to reduce these quantities from system (3.16), by the means
of classical algorithms of semi-algebraic geometry (see [15]). However, in general the
complexity of the latter grows double exponentially in the number of variables (see [37])
making them of little use in practice.
However, since the quantities in (3.13)-(3.14) are explicit, one may attempt to ex-
ploit their specific form in order to find an algorithm which is simpler than the classic
ones and that eliminates at least the parameters 𝑎22 , … , 𝑎2𝑠𝑚 , … , 𝑎𝑚2 , … , 𝑎𝑚𝑠𝑚 from
system (3.16). In this way, one would have an explicit criterion for steepness involving
only the multilinear forms of the tested function ℎ up to a given order, the parameters
𝑎21 , … , 𝑎𝑚1 , and the vectors 𝑣, 𝑢2 , … , 𝑢𝑚 . Moreover, as we shall show in the sequel,
without any loss of generality the numbers 𝑎21 , … , 𝑎𝑚1 and the vector 𝑣 can be assumed
to belong to a compact subset, whereas the vectors 𝑢2 , … , 𝑢𝑚 belong to 𝚄(𝑛, 𝑚 − 1),
which is compact by definition. Hence, having an explicit criterion for steepness in-
volving only the coefficients 𝑎21 , … , 𝑎𝑚1 and the vectors 𝑣, 𝑢2 , … , 𝑢𝑚 as additional free
parameters would be a qualititative improvement w.r.t. Theorem B in view of possible
applications, as one would only have to consider parameters belonging to a compactum.
Moreover, the presence of the vectors 𝑣, 𝑢2 , … , 𝑢𝑚 allows to keep track of the subspaces
one is working on; namely, it is possible to isolate the subspaces where the studied func-
tion is non-steep.
As we prove in sections 8-9, for a generic regular test function ℎ and for any 𝑚 ∈
{2, … , 𝑛}, on most of the 𝑚-dimensional subspaces of the Grassmannian 𝙶(𝑚, 𝑛) one is
able to apply an explicit criterion to check steepness that does not involve the parameters
𝑎22 , … , 𝑎2𝑠𝑚 , … , 𝑎𝑚2 , … , 𝑎𝑚𝑠𝑚 : this is the content of Theorems C1-C2-C3.
In order to state these results, we start by considering an integer 𝑛 ≥ 3, and a function
ℎ of class 𝐶 2 around the origin, verifying ∇ℎ(0) ≠ 0. Now, for any 𝑚 ∈ {2, … , 𝑛 − 1}
we need to consider some subsets of the Grassmannian manifold 𝙶(𝑚, 𝑛).
Definition 3.2. For any pair of integers 𝑚 ∈ {2, ..., 𝑛 − 1} and 𝑗 ∈ {0, … , 𝑚}, we
indicate by 𝑗 (ℎ, 𝑚, 𝑛) (resp. ≥𝑗 (ℎ, 𝑚, 𝑛)) the subset of 𝙶(𝑚, 𝑛) containing those 𝑚-
dimensional subspaces Γ𝑚 verifying
1. ∇ℎ(0) ⟂ Γ𝑚 ;

2. the Hessian matrix of the restriction of ℎ to Γ𝑚 , calculated at the origin, has

exactly 𝑗 null eigenvalues (resp. at least 𝑗 null eigenvalues).
With this definition, for any fixed 𝑚 ∈ {2, … , 𝑛 − 1}, we have the partition
⨆ ⨆
{Γ𝑚 ∈ 𝙶(𝑚, 𝑛)|Γ𝑚 ⟂ ∇ℎ(0)} = 0 (ℎ, 𝑚, 𝑛) 1 (ℎ, 𝑚, 𝑛) ≥2 (ℎ, 𝑚, 𝑛) . (3.23)

We are now ready to state Theorems C1-C2-C3.

Consider two positive integers 𝑟, 𝑛 ≥ 2, a vector 𝚜 ∶= (𝑠1 , … , 𝑠𝑛−1 ) ∈ ℕ𝑛−1 , with
1 ≤ 𝑠𝑖 ≤ 𝑟 − 1 for all 𝑖 = 1, … , 𝑛 − 1, and a function ℎ of class 𝐶𝑏2𝑟−1 around the

24
origin, satisfying ∇ℎ(0) ≠ 0. Then, for any given 𝑚 ∈ {2, ..., 𝑛 − 1}, one has the
following results (which, considered together, are refined versions of Theorems 1.4-1.5
in the introduction):
Theorem (C1). ℎ is steep at the origin, with index 𝛼𝑚 = 1, on the 𝑚-dimensional
subspaces belonging to 0 (ℎ, 𝑚, 𝑛).
Theorem (C2). If 𝑠𝑚 ≥ 2 there exist two semi-algebraic sets

𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚), 𝒜2 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) ⊂  ⋆ (𝑟, 𝑛) × ℝ𝑚−1 × 𝒱 1 (𝑚, 𝑛)

enjoying the following properties:

1. The form of 𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) can be explicitly computed starting from the expres-
𝑟,𝑠 ,𝑚
sion of set 𝑛 𝑚 in (3.20) by the means of an algorithm involving only linear
operations.

2. If system
{
(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛) , Span (𝑢1 , 𝑢2 , … , 𝑢𝑚 ) ∈ 1 (ℎ, 𝑚, 𝑛)
(3.24)
(𝚃0 (ℎ, 𝑟, 𝑛), 0, 𝑢1 , 𝑢2 , … , 𝑢𝑚 ) ∈ 𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚)

has no solution, then ℎ is steep at the origin with index 𝛼𝑚 ≤ 2𝑠𝑚 − 1 on any
subspace Γ𝑚 ∈ 1 (ℎ, 𝑚, 𝑛).

3. There exists a positive constant 𝒦 = 𝒦 (𝑟, 𝑛, 𝑚) such that if

⎧(𝑎 , … , 𝑎 ) ∈ 𝐵 𝑚−1 (𝒦 )
⎪ 21 𝑚1
∑
⎪(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛) , 𝑣 ∶= 𝑢1 + 𝑚 𝑖=2 𝑎𝑖1 𝑢𝑖
⎨ (3.25)
⎪ Span (𝑣, 𝑢2 , … , 𝑢𝑚 ) ∈ ≥2 (ℎ, 𝑚, 𝑛)
⎪(𝚃 (ℎ, 𝑟, 𝑛), 𝑎 , … , 𝑎 , 𝑢 , 𝑢 , … , 𝑢 ) ∈ 𝒜 (𝑟, 𝑠 , 𝑛, 𝑚)
⎩ 0 21 𝑚1 1 2 𝑚 2 𝑚

has no solution, then ℎ is steep at the origin with index 𝛼𝑚 ≤ 2𝑠𝑚 − 1 on any
subspace Γ𝑚 ∈ ≥2 (ℎ, 𝑚, 𝑛).

Remark 3.4. We observe that the statement above gives no information about the ex-
plicit expression of subset 𝒜2 (𝑟, 𝑠𝑚 , 𝑛, 𝑚). As it will be shown in sections 8-9, the linear
𝑟,𝑠 ,𝑚
algorithm used to deduce the form of 𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) starting from set 𝑛 𝑚 in (3.20)
fails in case 𝑣, 𝑢2 , … , 𝑢𝑚 span a subspace belonging to ≥2 (ℎ, 𝑚, 𝑛). Therefore, in or-
der to find the explicit expression for 𝒜2 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) one is obliged to apply the classi-
𝑟,𝑠 ,𝑚
cal, much slower algorithms of real-algebraic geometry to the set 𝑛 𝑚 determined
by system (3.20). Thus, checking steepness on the subspaces of ≥2 (ℎ, 𝑚, 𝑛) is more
complicated than on those belonging to 1 (ℎ, 𝑚, 𝑛), as one is obliged either to apply
slow algorithms to find the explicit expression of system (3.25), or to use the state-
ment of Theorem B, which nevertheless depends on the non-compact real coefficients
𝑎22 , … , 𝑎2𝑠𝑚 , … , 𝑎𝑚2 , … , 𝑎𝑚𝑠𝑚 .

25
 However, for a generic function ℎ the subsets 1 (ℎ, 𝑚, 𝑛) and ≥2 (ℎ, 𝑚, 𝑛) are "small"
inside the Grassmannian 𝙶(𝑚, 𝑛). Namely, in Theorem C3 below we prove that for any
𝑚 ∈ {2, … , 𝑛−1} and for any bilinear symmetric non-degenerate form 𝖡 ∶ ℝ𝑛 ×ℝ𝑛 ⟶
ℝ, the subspaces of dimension 𝑚 on which the restriction of 𝖡 has one or two null eigen-
values are rare in 𝙶(𝑚, 𝑛), both in measure and in topological sense.

Theorem (C3). Let 𝖡 ∶ ℝ𝑛 × ℝ𝑛 ⟶ ℝ be a bilinear, symmetric, nondegenerate form,

and let 𝑚 ∈ {2, … , 𝑛 − 1} be a positive integer.
For 𝑗 ∈ {1, 2}, denote by 𝙶≥𝑗 (𝖡, 𝑚, 𝑛) ⊂ 𝙶(𝑚, 𝑛) the subset of linear 𝑚-dimensional
subspaces on which the restriction of 𝖡 has at least 𝑗 null eigenvalues.
Then

1. 𝙶≥1 (𝖡, 𝑚, 𝑛) is contained in a submanifold of codimension one in 𝙶(𝑚, 𝑛);

2. 𝙶≥2 (𝖡, 𝑚, 𝑛) is obtained by the intersection of 𝙶≥1 (𝖡, 𝑚, 𝑛) with another subset
contained in a submanifold of codimension one in 𝙶(𝑚, 𝑛).

Finally, we state the following conjecture, which will hopefully be proved in a future
work.
Conjecture: for a generic bilinear form 𝖡, the subset 𝙶≥2 (𝖡, 𝑚, 𝑛) appearing in
Theorem C3 is contained in a submanifold of codimension two in 𝙶(𝑚, 𝑛).

3.2.4 Some observations on Theorems C1-C2

We stress that the criteria for steepness stated in Theorems C1-C2 are only valid for
a given function on specific subspaces. They do not guarantee steepness on the same
subspaces for an open neighborhood of functions around the studied one. Moreover,
Theorems C1-C2 yield a uniform bound on the indices of steepness on the subspaces
where their hypotheses are satisfied, but do not give any information on the coefficients
of steepness, which may tend to zero as a subspace where the conditions are not verified
is approached. This is why these theorems are somewhat "punctual", whereas Theorem
A is "local" (it ensures steepness for an open set of functions, in a neighborhood of a
considered point, with uniform estimates on the indices and coefficients).
To see an example showing why Theorems C1-C2 cannot be extended, in general,
to an open set of functions, and why they do not provide uniform steepness coefficients,
consider
(𝐼 + 𝐼2 )2
ℎ(𝐼1 , 𝐼2 , 𝐼3 ) = 1 + 𝐼3
2
and study on which subspaces of the origin it verifies the steepness condition. Even
though Theorems C1-C2 deal with steepness on 𝑚-dimensional subspaces, with 𝑚 ≥ 2,
for the sake of simplicity we will only consider one-dimensional subspaces in the sequel.
Infact, the phenomena that we will point out already suggest what kind of "problems"

26
may arise in higher dimensions when trying to extend the results of Theorems C1-
C2 either to an open set of functions, or when trying to get uniform estimates on the
steepness coefficients.
The Hessian matrix of ℎ at the origin reads

⎛1 1 0⎞
ℎ2(0,0) = ⎜1 1 0⎟ (3.26)
⎜ ⎟
⎝0 0 0⎠

and the vectors (𝑢1 , 𝑢2 , 𝑢3 ) ∈ ℝ3 which are both isotropic for ℎ2(0,0) and orthogonal to
∇ℎ(0) must verify

⎧(𝑢 , 𝑢 , 𝑢 ) ⋅ ∇ℎ(0) = 𝑢 = 0
⎪ 1 2 3 3
⎪ ⎛1 1 0⎞ ⎛𝑢1 ⎞
⎨ 2 ⎜ ⎟⎜ ⎟
⎪ℎ(0,0) [𝑢, 𝑢] = (𝑢1 , 𝑢2 , 𝑢3 ) ⎜1 1 0⎟ ⎜𝑢2 ⎟ = (𝑢1 + 𝑢2 )2 = 0
⎪ ⎜ ⎟⎜ ⎟
⎩ ⎝0 0 0⎠ ⎝𝑢3 ⎠

that is, they belong to the "bad" line

Γ1𝑏𝑎𝑑 ∶= Span{(𝑢1 , 𝑢2 , 𝑢3 ) | 𝑢3 = 0 , 𝑢1 + 𝑢2 = 0}

obtained by the intersection of the planes 𝑢3 = 0 and 𝑢1 + 𝑢2 = 0. Not only the Hessian
is completely degenerate on Γ1𝑏𝑎𝑑 : on a such a subspace the function ℎ is not even steep,
as the gradient of the restriction ∇(ℎ|Γ1 ) = 𝐼1 + 𝐼2 = 0 is identically null, so that
𝑏𝑎𝑑
condition (1.4) cannot be verified. However, for any parameter 𝜀 ≠ 0, the Hessian is
non-degenerate on the line

Γ1,𝜀
𝑔𝑜𝑜𝑑
∶= Span{(1 + 𝜀, −1, 0)}

as
ℎ2(0,0) [(1 + 𝜀, −1, 0), (1 + 𝜀, −1, 0)] = 𝜀2 > 0 .

This amounts to saying that ℎ is convex at the origin along Γ1,𝜀

𝑔𝑜𝑜𝑑
, hence that is it steep at
the origin on the subspace Γ1,𝜀
𝑔𝑜𝑜𝑑
with index equal to one. However, when 𝜀 is chosen to
be arbitrarily small, Γ1,𝜀
𝑔𝑜𝑜𝑑
approaches Γ1𝑏𝑎𝑑 and the steepness property breaks down as
the coefficients of steepness 𝐶1 , 𝛿 tend to zero. Hence, in the higher dimensional case,
if the conditions of Theorems C1-C2 are only matched on certain good subspaces, it
is reasonable to expect that one cannot have any uniform estimate on the steepness
coefficients, as these may go to zero when a bad subspace is approached.
Moreover, we introduce the one-parameter family of functions

(𝐼1 + (1 + 𝜆)𝐼2 )2
𝐻𝜆 (𝐼1 , 𝐼2 , 𝐼3 ) = + 𝐼3 , 𝜆∈ℝ
2

27
verifying 𝐻0 ≡ ℎ. For any given 𝜆0 ∈ ℝ, it is plain to check that the function 𝐻𝜆0
1,𝜆
0
is non-steep on the subspace Γ𝑔𝑜𝑜𝑑 , as the gradient of the restriction ∇(𝐻𝜆0 | 1,𝜆0 ) is
Γ𝑔𝑜𝑜𝑑
identically zero. However, as we have showed previously, the function 𝐻0 is steep on
Γ1,𝜆
𝑔𝑜𝑜𝑑
for any 𝜆 ≠ 0. Therefore, we have showed that for any given 𝜆 ≠ 0, there exists
a subspace Γ1,𝜆
𝑔𝑜𝑜𝑑
on which 𝐻0 is steep and a 𝜆-close function 𝐻𝜆 for the 𝐶 2 -norm
which is non-steep on Γ1,𝜆
𝑔𝑜𝑜𝑑
. Hence, in Theorems C1-C2 it is not reasonable to expect
that when ℎ is steep in a punctured neighborhood of the Grassmannian around a given
subspace, then the same holds true for all functions in a neighborhood of ℎ.

4 The Thalweg and its properties

It is clear from Definition 1.1 that studying the steepness property at the origin of a
given function ℎ ∈ 𝐶 2 (𝐵 𝑛 (0, 2𝛿), ℝ) verifying ∇ℎ(0) ≠ 0, amounts to studying the
projection of its gradient on any 𝑚-dimensional subspace Γ𝑚 perpendicular to ∇ℎ(0),
with 𝑚 ∈ {1, ..., 𝑛 − 1}. More precisely, given 𝛿 > 0, for any fixed 𝜂 ∈ (0, 𝛿] we are
interested in the quantity

𝜇ℎ (Γ𝑚 , 𝜂) ∶= min ||𝜋Γ𝑚 ∇ℎ(𝑢)||2 .

𝑢∈Γ𝑚 , ||𝑢||2 =𝜂

Since, for any given Γ𝑚 orthogonal to ∇ℎ(0) and for any 𝜂 ∈ (0, 𝛿], the value 𝜇ℎ (Γ𝑚 , 𝜂)
is attained at some point of the 𝑚-dimensional sphere

𝜂𝑚 ∶= {𝑢 ∈ Γ𝑚 | ||𝑢||2 = 𝜂} ,

it makes sense to give the following

Definition 4.1. We call Thalweg of ℎ on Γ𝑚 the set

 (ℎ, Γ𝑚 ) ∶= {𝐼 ⋆ ∈ Γ𝑚 ∶ ||𝜋Γ𝑚 ∇ℎ(𝐼 ⋆ )||2 = 𝜇ℎ (Γ𝑚 , 𝜂) for 0 ≤ 𝜂 ∶= ||𝐼 ⋆ ||2 ≤ 𝛿} .

In the sequel, we will be interested in studying the thalweg  (𝚃0 (ℎ, 𝑟, 𝑛), Γ𝑚 ) of the
Taylor polynomial 𝚃0 (ℎ, 𝑟, 𝑛). Namely, the goal of this section is to prove the following

Theorem 4.1. (Nekhoroshev, [47]) For any pair of integers 𝑟, 𝑛 ≥ 2, and for any real
𝛿 > 0, consider a function ℎ ∈ 𝐶 𝑟 (𝐵 𝑛 (0, 2𝛿)) verifying ∇ℎ(0) ≠ 0. Then, for any given
number 𝑚 ∈ {1, ..., 𝑛 − 1}, for any 𝑚-dimensional subspace Γ𝑚 orthogonal to ∇ℎ(0)
there exists a semi-algebraic curve 𝛾 with values in  (𝚃0 (ℎ, 𝑟, 𝑛), Γ𝑚 ) such that 𝛾(0) = 0
and

1. For any fixed 𝜂 ≥ 0, the intersection Im(𝛾) ∩ 𝜂𝑚 is a singleton;

2. There exists a positive integer 𝚍 = 𝚍(𝑟, 𝑛, 𝑚) that bounds the diagram (see Def.
A.2) of graph(𝛾);

28
 3. There exists 𝙺 = 𝙺(𝑟, 𝑛, 𝑚) > 1 such that, for any 𝜆 > 0, the curve 𝛾 is real-
analytic on some closed interval 𝙸𝜆 ⊂ [−𝜆, 𝜆] of length 𝜆∕𝙺, with complex ana-
lyticity width 𝜆∕𝙺;

4. Over 𝙸𝜆 , 𝛾 is an 𝑖-arc, i.e. it can be parametrized by

{
𝑥𝑖 (𝑡) = 𝑡 for some 𝑖 ∈ {1, ..., 𝑚}
𝛾(𝑡) ∶= 𝑡 ∈ 𝙸𝜆
𝑥𝑗 (𝑡) = 𝑓𝑗 (𝑡) for all 𝑗 ∈ {1, ..., 𝑚}, 𝑗 ≠ 𝑖

where the 𝑓𝑗 (𝑡) are Nash (i.e. analytic-algebraic) functions;

5. 𝛾 satisfies a Bernstein’s inequality on its Taylor coefficients over the interval 𝙸𝜆 .

Namely, indicating by

∑
+∞
𝑓𝑗 (𝑡) = 𝑎𝑗𝛽 (𝑢) 𝑡𝛽 , 𝑗 ∈ {1, ..., 𝑚} , 𝑗 ≠ 𝑖 ,
𝛽=0

the Taylor expansion of 𝑓𝑗 at some point 𝑢 ∈ 𝙸𝜆 , there exists a positive constants

𝐾2 (𝑟, 𝑛, 𝑚), and 𝙼 = 𝙼(𝑟, 𝑛, 𝑚, 𝛽) ∶= 𝛽! × (𝙺(𝑟, 𝑛, 𝑚))𝛽 × 𝐾2 (𝑟, 𝑛, 𝑚) for which the
following uniform estimate holds:
𝙼
max |𝑎𝑗𝛽 (𝑢)| ≤ . (4.1)
𝑢∈𝙸𝜆 𝜆𝛽−1

Remark 4.1. The Theorem above corresponds to reasonings holding in the polynomial
setting. Moreover, the constants 𝚍, 𝙺, 𝙼 depend only on the degree of the considered
polynomial. Consequently, this Theorem holds uniformly for any 𝑟-jet of any function
ℎ ∈ 𝐶 𝑟 (𝐵 𝑛 (0, 2𝛿), ℝ).
Remark 4.2. Nekhoroshev calls 𝛾 "minimal arc with uniform characteristics" (see [47],
section 4). In that work, the statement of Theorem 4.1 is not given in the form above but
is rather split in dispersed parts. Moreover, many of the modern tools of real-algebraic
geometry were lacking at that time, so that the redaction of his work appears quite
obscure in some parts. These two elements makes difficult for the reader to reconstruct
simply Theorem 4.1 from Nekhoroshev’s original paper.
Remark 4.3. The Bernstein’s inequality at point 5 of Theorem 4.1 is essential in order
to have stable12 lower estimates for the steepness coefficients of ℎ. For more details
about this result, which is interesting in itself and has applications in various fields of
mathematics, see refs. [55] and [13].
Some intermediate Lemmas are needed before demonstrating Theorem 4.1.

Lemma 4.1. Consider any triplet of integers 𝑟, 𝑛 ≥ 2 and 𝑚 ∈ {1, ..., 𝑛 − 1}. There
exists 𝖽 = 𝖽(𝑟, 𝑛, 𝑚) ≥ 0 such that, for any 𝑄 ∈ (𝑟, 𝑛) verifying ∇𝑄(0) ≠ 0, and for
12 In the sense given in Th. 3.1, that is valid for an open set of functions.

29
any subspace Γ𝑚 perpendicular to ∇𝑄(0), the thalweg  (𝑄, Γ𝑚 ) is a semi-algebraic
set satisfying diag( (𝑄, Γ𝑚 )) ≤ 𝖽 (see Def. A.2). Moreover, for any fixed 𝜂0 > 0, the
intersection of  (𝑄, Γ𝑚 ) with the sphere 𝜂𝑚 ⊂ Γ𝑚 is compact.
0

Proof. Γ𝑚 is obviously isomorphic toℝ𝑚 with the metric induced by the euclidean
ambient space ℝ𝑛 , and thus admits a global system of orthonormal coordinates 𝑥 =
(𝑥1 , ..., 𝑥𝑚 ). We denote by 𝑃 (𝑥) ∈ (𝑟, 𝑚) the restriction of 𝑄(𝐼) to Γ𝑚 ≃ ℝ𝑚 . Since
we endow Γ𝑚 with the induced euclidean metric, studying the norm of the projection of
∇𝐼 𝑄(𝐼) on Γ𝑚 amounts to studying the induced norm of ∇𝑥 𝑃 (𝑥) on Γ𝑚 ≃ ℝ𝑚 . Now,
consider the semi-algebraic set

 ∶= {(𝑥, 𝑦, 𝜂) ∈ ℝ2𝑚 × ℝ ∶ ||𝑥||22 = ||𝑦||22 = 𝜂 2 , 𝜂 > 0, ||∇𝑃 (𝑥)||2 > ||∇𝑃 (𝑦)||2 }
(4.2)
By the Theorem of Tarski and Seidenberg A.1 and Proposition A.1, we have that the
set ℝ𝑚 ∖𝜋𝑥  ∶= {𝑥 ∈ ℝ𝑚 × ℝ ∶ ∀(𝑦, 𝜂) ∈ ℝ𝑚 × ℝ, (𝑥, 𝑦, 𝜂) ∉ } is semi-algebraic.
We claim that it coincides with  (𝑄, Γ𝑚 ). Infact, by the definition of , it is clear that
for any given 𝜂 > 0 one has

𝑥 ∈ ℝ𝑚 ∖𝜋𝑥  ⟺ ||∇𝑃 (𝑥)||2 ≤ ||∇𝑃 (𝑦)||2 for all 𝑦 ∈ ℝ𝑚 s.t. ||𝑦||2 2 = ||𝑥||2 2 ,
(4.3)
so that 𝑥 ∈ ℝ ∖𝜋𝑥  is the locus of minima on any given sphere for ||∇𝑃 ||2 (that
𝑚

is for ||𝜋Γ𝑚 ∇𝑄||2 ), that is it coincides with the Thalweg  (𝑄, Γ𝑚 ). Moreover, since
deg 𝑃 ≤ 𝑟, the diagram of  is uniformly bounded w.r.t. any 𝑃 ∈ (𝑟, 𝑚) and, again
by the Theorem of Tarski and Seidenberg, the same is true for 𝜋𝑥  and for ℝ𝑚 ∖𝜋𝑥  ≡
 (𝑄, Γ𝑚 ).
It remains to prove that  (𝑄, Γ𝑚 )∩𝜂𝑚 is compact. By construction,  (𝑄, Γ𝑚 )∩𝜂𝑚
0 0
is the locus of minima of ||∇𝑃 ||2 on 𝜂𝑚 . Since the restriction of the function ||∇𝑃 ||2
0
to 𝜂𝑚 is continuous for the topology induced by Γ𝑚 , the inverse image of its minimal
0
value on 𝜂𝑚 is closed. Since 𝜂𝑚 is compact, the thesis follows.
0 0

The next Lemma shows how an analytic curve with uniform characteristics can be
extracted from the Thalweg.

Lemma 4.2. Fix a triplet of integers 𝑟, 𝑛 ≥ 2 and 𝑚 ∈ {1, .., 𝑛 − 1}. There exist
positive constants 𝐷(𝑟, 𝑛, 𝑚) ∈ ℕ, and 𝐾𝑖 = 𝐾𝑖 (𝑟, 𝑛, 𝑚) ∈ ℝ, 𝑖 = 1, 2, such that,
for any 𝜉 > 0, for any polynomial 𝑄 ∈ (𝑟, 𝑛) satisfying ∇𝑄(0) ≠ 0, and for any
𝑚-dimensional subspace Γ𝑚 orthogonal to ∇𝑄(0), there exists a semi-algebraic curve
𝜙 = (𝜙1 (𝜂), ..., 𝜙𝑚 (𝜂)) ∶ [0, 𝜉] ⟶  (𝑄, Γ𝑚 ) having the following properties

1. For any fixed 𝜂 ∈ [0, 𝜉], the intersection Im(𝜙) ∩ 𝜂𝑚 is a singleton;

2. The diagram of graph(𝜙) (see Definition A.2) is bounded by 𝐷;

3. There exists a closed interval 𝜉 ⊂ [0, 𝜉] of length 𝜉∕𝐾1 over which 𝜙 is real-
analytic, with complex analyticity width 𝜉∕𝐾1 ;

30
 4. On the closed complex polydisk (𝜉 )𝜉∕𝐾1 of width 𝜉∕𝐾1 around 𝜉 , one has the
uniform Bernstein’s inequality

max |𝜙𝑗 (𝜂)| ≤ 𝐾2 𝜉 for any 𝑗 ∈ {1, ..., 𝑚}.

𝜂∈(𝜉 )𝜉∕𝐾
1

Proof. As in Lemma 4.1, we consider the isomorphism Γ𝑚 ≃ ℝ𝑚 , and we endow Γ𝑚

with a global system of orthonormal coordinates 𝑥 = (𝑥1 , ..., 𝑥𝑚 ) for the scalar product
induced by the ambient space ℝ𝑛 . We proceed by steps. At the first step, we build
a semi-algebraic function 𝜙1 associating to a sphere 𝜂𝑚 of given radius 𝜂 > 0 the
minimal value attained by the coordinate 𝑥1 on 𝜂𝑚 . At Step 2 we apply Yomdin’s
reparametrization Lemma (see appendix B) to one of the algebraic components of 𝜙1
and we get a function with the suitable properties. Finally, at Step 3, we repeat the same
construction for the other coordinates.
Step 1. For any 𝜉 > 0, by Lemma 4.1 the set

𝜉 (𝑄, Γ𝑚 ) ∶=  (𝑄, Γ𝑚 ) ∩ {𝑥 ∈ ℝ𝑚 ∶ ||𝑥||22 ≤ 𝜉 2 }

is semi-algebraic and its diagram is bounded by a positive constant 𝖽 = 𝖽(𝑟, 𝑛, 𝑚). By

the Theorem of Tarski and Seidenberg A.1, the continuous function

𝑓1 ∶= 𝜉 (𝑄, Γ𝑚 ) ⟶ ℝ , 𝑥 ⟼ 𝑥1

is semi-algebraic and its diagram is bounded by a quantity depending only on 𝑟, 𝑛, 𝑚.

Infact,

graph(𝑓1 ) ∶= Πℝ𝑚 ×ℝ {(𝑢, 𝑣) ∈ ℝ𝑚 × ℝ𝑚 | 𝑢 ∈ 𝜉 (𝑄, Γ𝑚 ), 𝑢 = 𝑣} .

Moreover, Lemma 4.1 ensures that for any 0 < 𝜂0 < 𝜉 the set 𝜉 (𝑄, Γ𝑚 )∩𝜂𝑚 is compact,
0
so that 𝑓1 has a minimum on it. On the other hand, the function 𝑔1 ∶ 𝜉 (𝑄, Γ𝑚 ) ⟶ ℝ,
𝑥 ⟼ ||𝑥||2 is also semi-algebraic and its diagram is bounded by a quantity depending
only on 𝑟, 𝑛, 𝑚, because

graph(𝑔1 ) ∶= {(𝑥, 𝑦) ∈ ℝ𝑚 × ℝ | 𝑥 ∈ 𝜉 (𝑄, Γ𝑚 ) , ||𝑥||22 − 𝑦2 = 0} .

Then, by applying Proposition A.11, we have that the function

𝜙1 ∶ [0, 𝜉] ⟶ ℝ 𝜂⟼ inf 𝑓1 (𝑥) = min {𝑥1 } (4.4)

𝑥∈𝑔1−1 (𝜂) 𝑥∈𝜉 (𝑄,Γ𝑚 )∩𝜂𝑚

is semi-algebraic and we indicate by 𝑑1 = 𝑑1 (𝑟, 𝑛, 𝑚) its diagram.

Step 2. Corollary A.1 ensures the existence of a number 𝑁1 = 𝑁1 (𝑑1 ) and of an
open interval  1 ⊂ [0, 𝜉] of length 𝜉∕𝑁1 over which the restriction 𝜙1 | 1 is algebraic.
By Proposition A.6, 𝜙1 | 1 is 𝑑1 -valent and has no more than 𝑑1 zeros on its domain so
that there exists an interval  1 ⊂  1 of length 𝜉∕(𝑁1 (𝑑1 ) × (𝑑1 + 1)) over which the

31
restriction 𝜙1 | 1 has definite sign. Without loss of generality we can assume 𝜙1 (𝜂) ≥ 0
for all 𝜂 ∈  1 (one considers −𝜙1 otherwise).
We denote by 𝜉1 and 𝜉2 the extremal points of the interval  1 and we rescale the
domain by setting
( )
𝑢+1 𝑢−1
𝜑1 ∶ [−1, 1] ⟶ ℝ , 𝜑1 (𝑢) ∶= 𝜙1 𝜉2 − 𝜉1 . (4.5)
2 2
We also define the function
𝜑1 (𝑢)
̃ 1 (𝑢) ∶=
𝜑 , 0≤𝜑
̃ 1 (𝑢) ≤ 1 , (4.6)
𝜉
which satisfies the hypotheses of Theorem B.1. With the notations of Theorem B.1,
1
we choose the value 𝛿 ∶= so that, once at most 𝚈1 neighborhoods of length
8𝚈1 (𝑑1 )
2𝛿 around the singularities of 𝜑̃ 1 are eliminated from [−1, 1], the remaining set has a
∑𝚈1
measure which is no less than 2 − 2 × 1∕(8𝚈1 ) = 7∕4 . Moreover, the number of the
𝑖=1 ( )
partition intervals is bounded by the uniform quantity 𝚈2 log2 8𝚈1 , so that there exists
an interval Δ1 ⊂ [−1, 1] verifying
7
|Δ1 | = ( )
4 𝚈2 log2 8𝚈1

on which 𝜑 ̃ 1 is real-analytic with uniform analyticity width 32 |Δ1 |. Infact, by Propo-

sition B.1, the complex singularities of 𝜑 ̃ 1 are at distance no less than 3|Δ1 | from the
center 𝑐1 of Δ1 . By Theorem B.1, Δ1 can be affinely reparametrized by a function
𝜓1 ∶ [−1, 1] ⟶ Δ1 , which maps the closed complex disc 3 (0) into the closed com-
plex disc 𝜚 (𝑐1 ) of radius 𝜚 ∶= 32 |Δ1 |. Hence, we can write

max |̃
𝜑1 (𝑢)| ≤ max |̃ ̃ 1 (0)| + |̃
𝜑1 (𝑢) − 𝜑 𝜑1 (0)|
𝑢∈𝜚 (𝑐1 ) 𝑢∈𝜚 (𝑐1 )
(4.7)
= max |̃ ̃ 1 ◦𝜓1 (0)| + |̃
𝜑1 ◦𝜓1 (𝑧) − 𝜑 𝜑1 (0)|
𝑧∈3 (0)

so that, by Definition B.1 and Theorem B.1 and by the fact that |̃
𝜑1 (𝑢)| ≤ 1 for any
𝑢 ∈ [−1, 1], we obtain

max |̃
𝜑1 (𝑢)| ≤ 2 . (4.8)
𝑢∈𝜚 (𝑐1 )

Scaling back to the original variables, by (4.5) the interval Δ1 is mapped into an interval
𝜉 −𝜉
Δ𝜉1 of length |Δ𝜉1 | = |Δ1 | 2 2 1 = |Δ1 | 2𝑁 (𝑑 )×(𝑑
𝜉
+1)
and center 𝑐̂1 and, in the same way,
1 1 1
𝜉
the radius rescales as 𝜚 ⇝ 𝜚 2𝑁 (𝑑 )×(𝑑 . Therefore, taking into account (4.6) and
1 1 1 +1)
(4.8), there exists a uniform constant 𝙼1 = 𝙼1 (𝑑1 ) such that the following Bernstein’s
inequality is satisfied
max |𝜙1 (𝑧)| ≤ 2𝜉 . (4.9)
𝑧∈𝜉∕𝙼 (̂
𝑐1 )
1

32
 Step 3. Since Δ𝜉1 is compact, 𝜙1 (Δ𝜉1 ) is also compact and the inverse image
𝚄1𝜉 (𝑄, Γ𝑚 ) ∶= 𝑔1−1 (𝜙−1
1
(𝜙1 (Δ𝜉1 ))) is closed. Moreover, since the diagrams of 𝜙1 and 𝑔1
depend only on 𝑟, 𝑛, 𝑚, then by Propositions A.7-A.8-A.9 the diagram of 𝚄1𝜉 (𝑄, Γ𝑚 ) also
depends only on 𝑟, 𝑛, 𝑚. Hence, for any fixed 𝜂 ∈ Δ𝜉1 we have that the set 𝜂𝑚 ∩𝚄1𝜉 (𝑄, Γ𝑚 ),
which contains the points of the Thalweg that have minimal coordinate 𝑥1 on the sphere
of radius 𝜂, is compact and semialgebraic with a bound on its diagram depending only
on 𝑟, 𝑛, 𝑚. Hence, the coordinate 𝑥2 admits a minimum on this set and we can repeat
the same argument of Step 2 on the function
𝜙2 ∶ Δ𝜉1 ⟶ ℝ 𝜂 ⟼ min 𝑓2 (𝑥) = min {𝑥2 } (4.10)
𝑥∈𝑔2−1 (𝜂) 𝑥∈𝜂𝑚 ∩𝚄1𝜉 (𝑄,Γ𝑚 )

where we have set 𝑓2 ∶ 𝜂𝑚 ∩ 𝚄1𝜉 (𝑄, Γ𝑚 ) ⟶ ℝ , 𝑥 ⟼ 𝑥2 and 𝑔2 ∶ 𝜂𝑚 ∩

0 0
𝚄1𝜉 (𝑄, Γ𝑚 )⟶ ℝ, 𝑥 ⟼ ||𝑥||2 . The curve 𝜙 ∶= (𝜙1 , 𝜙2 , ..., 𝜙𝑚 ) is constructed by
iterating this procedure 𝑚 times.
Points 1, 2, and 3 of the thesis follows easily from this construction. Point 4 is a
consequence of estimate (4.9) applied to the complex polydisk of uniform width 𝜉∕𝐾1
around the common uniform real interval of analyticity 𝜉 of the functions 𝜙1 , ..., 𝜙𝑚 .

We are now ready to state the proof of Theorem 4.1.

Proof. (Theorem 4.1) We assume the setting of Lemma 4.2 with 𝑄 equal to the Taylor
expansion 𝚃0 (ℎ, 𝑟, 𝑛), and we proceed by steps. At Step 1, we show that there exists
a component 𝜙𝑖 , 𝑖 ∈ {1, … , 𝑚}, of the curve 𝜙 introduced in Lemma 4.2 whose first
derivative admits a lower bound on a domain of uniform length. Then, at the sec-
ond step, we use this fact to apply a quantitative inverse function Theorem and we
reparametrize 𝜙 by the 𝑖-th coordinate. Steps 3 and 4 contain, respectively, the proofs
of points 1-4 and of point 5 in the statement.
Step 1. We cut the uniform interval of analyticity 𝜉 into three equal intervals and
we denote by ̂𝜉 the central one, whose length is |̂𝜉 | = |𝜉 |∕3. We indicate by 𝜉̂1 and
𝜉̂2 the extreme points of ̂𝜉 . Since for any given 𝜂 ∈ ̂𝜉 by Lemma 4.2 we have 𝜂 2 =
𝜙21 (𝜂) + ... + 𝜙2𝑚 (𝜂), there must be some component 𝜙𝑖 of the curve, with 𝑖 ∈ {1, ..., 𝑚},
verifying
|̂𝜉 | 𝜉
|𝜙𝑖 (𝜉̂2 ) − 𝜙𝑖 (𝜉̂1 )| ≥ = . (4.11)
𝑚 3𝑚𝐾1
At the same time, for some point 𝜉̂3 ∈ ̂𝜉 we have
𝜉
|𝜙𝑖 (𝜉̂2 ) − 𝜙𝑖 (𝜉̂1 )| = |𝜙′𝑖 (𝜉̂3 )||̂𝜉 | = |𝜙′𝑖 (𝜉̂3 )| . (4.12)
3𝐾1
On the one hand, relations (4.11) and (4.12) together imply
1
|𝜙′𝑖 (𝜉̂3 )| ≥ . (4.13)
𝑚

33
On the other hand, for any 𝜂 ∈ [𝜉̂3 − |̂𝜉 |, 𝜉̂3 + |̂𝜉 |] ⊂ 𝜉 one has the estimate

|𝜙′𝑖 (𝜂) − 𝜙′𝑖 (𝜉̂3 )| ≤ max |𝜙′′

𝑖 ||𝜂 − 𝜉3 |
̂ (4.14)
𝜉

which, thanks to the classic Cauchy estimate and to the Bernstein inequality of Lemma
4.2, implies

2𝐾12 2𝐾12 𝐾2
|𝜙′𝑖 (𝜂) − 𝜙′𝑖 (𝜉̂3 )| ≤ max |𝜙𝑖 ||𝜂 − 𝜉̂3 | ≤ |𝜂 − 𝜉̂3 | . (4.15)
𝜉2 (𝜉 )𝜉∕𝐾
1
𝜉
[ ]
𝜉 𝜉
Hence, for any 𝜂 in the interval 𝐽𝜉 ∶= 𝜉̂3 − , 𝜉̂3 + ⊂ 𝜉 we have
4𝑚𝐾12 𝐾2 4𝑚𝐾12 𝐾2
by (4.13) and (4.15) that
2
| | 1 2𝐾1 𝐾2 𝜉 1
|𝜙′𝑖 (𝜂)| ≥ ||𝜙′𝑖 (𝜉̂3 )| − |𝜙′𝑖 (𝜂) − 𝜙′𝑖 (𝜉̂3 )|| ≥ − = . (4.16)
| | 𝑚 𝜉 4𝑚𝐾 2 𝐾2 2𝑚
1

Step 2. By Lemma 4.2 and by the construction at Step 1 we can apply the quanti-
tative local inversion Theorem C.2 for 𝜙𝑖 at any point 𝜂 ∈ 𝐽𝜉 ⊂ 𝜉 . By making use of
the notations in Theorem C.2, we can set the uniform parameters

𝜉 1 𝐾12 𝐾2
𝑅 ∶= , |𝜙′𝑖 (𝜂)| ≥ , max |𝜙′′
𝑗|≤2 . (4.17)
𝐾1 2𝑚 (𝐽𝜉 )𝜉∕𝐾
1
𝜉

Hence, 𝜙𝑖 is invertible in the complex closed polydisk (𝐽𝜉 )𝑅′ ∕16 around the real interval
𝐽𝜉 , where
{ }
1 min𝐽𝜉 |𝜙′𝑖 | 𝜉
′
𝑅 ∶= × min 𝑅, , = .
2 max(𝐽𝜉 )𝜉∕(2𝐾 ) |𝜙′′ | 8𝑚𝐾 2𝐾
𝑖 1 1 2

Since, by construction, 𝜙𝑖 is real-analytic in 𝐽𝜉 , the continuity of the derivative ensures

that 𝜙𝑖 (𝐽𝜉 ) is an interval of ℝ and, by the definition of the function 𝜙𝑖 in (4.4), one has
𝜙𝑖 (𝐽𝜉 ) ⊂ [−𝜉, 𝜉]. The inverse function is analytic in the complex polydisc of uniform
width
𝑅′ 𝑅′ 𝜉
𝑅′′ ∶= min |𝜙′𝑖 | ≥ = (4.18)
𝐽𝜉 8 16𝑚 128 𝑚2 𝐾 2 𝐾2
1
around 𝜙𝑖 (𝐽𝜉 ). Moreover, using (4.16), one has that

1 𝜉
|𝜙𝑖 (𝐽𝜉 )| ≥ min |𝜙′𝑖 | × |𝐽𝜉 | ≥ × . (4.19)
𝐽𝜉 2𝑚 2𝑚𝐾 2 𝐾2
1

Step 3. Point 1 of Theorem 4.1 follows by Point 1 of Lemma 4.2 and by the local inver-
sion Theorem applied at Step 2. Points 2, 3, and 4 of Theorem 4.1 are also immediate
consequences of the local inversion Theorem at Step 2.

34
 Namely, by keeping in mind the notations at Point 4 of Theorem 4.1, the curve
𝛾 ∶= 𝜙◦𝜙−1 𝑖 can be defined as
{
𝑥𝑖 = 𝑡
𝛾(𝑡) ∶=
𝑥𝑗 (𝑡) = 𝑓𝑗 (𝑥𝑖 ) ∶= 𝜙𝑗 (𝜙−1
𝑖 (𝑥𝑖 )) for all 𝑗 ∈ {1, ..., 𝑚} , 𝑗 ≠ 𝑖 .
(4.20)
The existence of an interval of analyticity with uniform length and complex width
for 𝛾 is a consequence of (4.18) and (4.19) and the constant 𝙺 in the statement can be
taken equal to
𝙺 ∶= 128 𝑚2 𝐾12 𝐾2 . (4.21)
Indeed, for any 0 < 𝜆 ≤ 𝜉, 𝙸𝜆 can be chosen to be any interval of length 𝜆∕𝙺 contained
in the interval 𝜙𝑖 (𝐽𝜆 ) ⊂ [−𝜆, 𝜆] (see (4.19)). For later convenience, we also observe
that the above discussion implies that

(𝜆 )𝜆∕𝐾1 ⊃ (𝐽𝜆 )𝜆∕𝐾1 ⊃ 𝜙−1

𝑖 (𝙸𝜆 )𝜆∕𝙺 , ∀𝜆 ∈ (0, 𝜉] . (4.22)

The fact that the diagram of graph(𝛾) depends only on 𝑟, 𝑛, 𝑚 is an immediate con-
sequence of (4.20), together with point 2 of Lemma 4.2 and with Propositions A.8-A.9.
Step 4. It remains to prove the Bernstein’s inequality at Point 5 of the statement. By
Lemma 4.2, for any 𝜆 > 0 we have

max |𝜙𝑗 (𝜂)| ≤ 𝐾2 𝜆 (4.23)

𝜂∈(𝜆 )𝜆∕𝐾
1

for some uniform constant 𝐾2 = 𝐾2 (𝑟, 𝑛, 𝑚) and for any 𝑗 ∈ {1, ..., 𝑚}. By construction
in (4.20), 𝑓𝑗 (𝑥1 ) = 𝜙𝑗 ◦𝜙−1
𝑖 (𝑥𝑖 ), and for any 𝛽 ∈ ℕ ∪ {0} the classic Cauchy estimate
implies

max𝑧∈(𝙸𝜆 )𝜆∕𝙺 |𝑓𝑗 (𝑧)| max𝑧∈(𝙸𝜆 )𝜆∕𝙺 |𝜙𝑗 ◦𝜙−1

𝑖 (𝑧)|
max |𝑓𝑗(𝛽) (𝑡)| ≤ 𝛽! 𝙺 𝛽
= 𝛽! 𝙺𝛽
. (4.24)
𝑡∈𝙸𝜆 𝜆𝛽 𝜆𝛽

For any 0 < 𝜆 ≤ 𝜉, by (4.22) one has 𝜙−1

𝑖 ((𝙸𝜆 )𝜆∕𝙺 ) ⊂ (𝜆 )𝜆∕𝐾1 . Taking this into
account, (4.24) and (4.23) yield
max𝜂∈(𝜆 )𝜆∕𝐾 |𝜙𝑗 (𝜂)| 𝐾2 𝜆 𝐾
max |𝑓𝑗(𝛽) (𝑡)| ≤ 𝛽! 𝙺𝛽 1
≤ 𝛽! 𝙺𝛽 = 𝛽! 𝙺𝛽 2 . (4.25)
𝑡∈𝙸𝜆 𝜆𝛽 𝜆𝛽 𝜆𝛽−1
The thesis at Point 5 in the statement follows by setting 𝙼 = 𝛽! × 𝐾2 × 𝙺𝛽 .

5 𝑠-vanishing polynomials
We take into account the results and the notations of the previous section, in particular
Theorem 4.1.

35
5.1 Heuristics and Definitions
The goal of the first part of this paragraph is to provide the reader with a heuristic justi-
fication for introducing the special class of 𝑠-vanishing polynomials in the study of the
genericity of steepness. A rigorous description of the rôle played by these polynomials
will be given in the next paragraphs and sections.
For any fixed integer 𝑛 ≥ 2, we consider the euclidean space ℝ𝑛 and we endow
any of its linear subspaces with the induced metric. For any pair of positive integers
1 ≤ 𝑚 ≤ 𝑛 − 1 and 𝑟 ≥ 2, for any given function ℎ of class 𝐶 𝑟 near the origin verifying
∇ℎ(0) ≠ 0, and for any 𝑚-dimensional subspace Γ𝑚 orthogonal to ∇ℎ(0), by Def. 4.1,
the set  (𝚃0 (ℎ, 𝑟, 𝑛), Γ𝑚 ) is the locus of minima of ||𝜋Γ𝑚 ∇𝚃0 (ℎ, 𝑟, 𝑛)||2 on the spheres
𝜂𝑚 (0) ⊂ Γ𝑚 , with 𝜂 > 0.
In Theorem 4.1 we have proved the existence of a minimal semi-algebraic arc 𝛾
(see (4.20)) of diagram 𝚍(𝑟, 𝑛, 𝑚) parametrized by one coordinate and whose image is
contained in the thalweg  (𝚃0 (ℎ, 𝑟, 𝑛), Γ𝑚 ). Due to Proposition A.4 - 𝛾 is piecewise
algebraic, with a maximal number of algebraic components depending only on its di-
agram 𝚍(𝑟, 𝑛, 𝑚). With the exception of a finite set of complex points, any algebraic
function admits a local holomorphic extension, and the number of its singularities is
bounded by a quantity depending only on its diagram (see appendix B, or [13] for more
details). Therefore, 𝛾(𝑡) is real-analytic with the exception of a finite number of points
whose cardinality is bounded uniformly by a quantity depending solely on 𝚍(𝑟, 𝑛, 𝑚).
In particular, for any 𝜆 > 0, this ensures the existence of an interval 𝙸𝜆 ⊂ [−𝜆, 𝜆] of
uniform length 𝜆∕𝙺(𝑟, 𝑚, 𝑛), where 𝙺 = 𝙺(𝑟, 𝑚, 𝑛) is a suitable constant, over which 𝛾(𝑡)
is real-analytic with complex analyticity width 𝜆∕𝙺.
By the above reasonings, for sufficiently small 𝜆 > 0 the interval (−3𝜆, 3𝜆) contains
no singularities of 𝛾(𝑡). In particular, 𝛾(𝑡) is real analytic in 𝙻𝜆 ∶= (𝜆, 2𝜆), with com-
plex analyticity width 𝜆, and the same holds also for || 𝜋Γ𝑚 ∇𝚃0 (ℎ, 𝑟, 𝑛)||𝛾(𝑡) ||2 in that
2

interval. Hence, if the function || 𝜋Γ𝑚 ∇𝚃0 (ℎ, 𝑟, 𝑛)||𝛾(𝑡) ||2 has a zero of infinite order at
2

some point 𝑡⋆ ∈ 𝙻𝜆 , then it is identically null in 𝙻𝜆 . Then, by Definition 1.1, and by the
minimality of 𝛾, this implies that the polynomial 𝚃0 (ℎ, 𝑟, 𝑛) cannot satisfy the steepness
property at the origin on the subspace Γ𝑚 .
We claim that a kind of converse result - involving 𝙸𝜆 instead of 𝙻𝜆 - is also true: if
𝚃0 (ℎ, 𝑟, 𝑛) is non-steep at the origin on the subspace Γ𝑚 , then || 𝜋Γ𝑚 ∇𝚃0 (ℎ, 𝑟, 𝑛)||𝛾(𝑡) ||2
2

must have a zero of infinite order in 𝙸𝜆 . This observation is fundamental in order to

prove Theorem A. Actually, the necessity of a zero of infinite order has been proved for
a real-analytic function in [51] via the curve-selection Lemma; however, in the polyno-
mial setting considered here, we have a much stronger quantitative result.
Motivated by this heuristic argument, we are interested in studying the properties
of those real polynomials of 𝑚 ≥ 1 variables whose gradient has a zero of sufficiently
high order on some curve 𝛾 parametrized by one coordinate. In a first moment, we
do not consider the fact that these polynomials are the restrictions to a 𝑚-dimensional
subspace Γ𝑚 of polynomials defined in ℝ𝑛 , with 𝑛 > 𝑚. This will be taken into account

36
in section 6. Therefore, we give the following definitions:
Definition 5.1. We indicate by Θ𝑚 the set of curves 𝛾(𝑡) with values in ℝ𝑚 such that
𝛾(𝑡) is real-analytic around the origin, and 𝛾(0) = 𝛾(0)
̇ = 0.
Remark 5.1. By the inverse function theorem, for any element 𝛾 ∈ Θ𝑚 there exists
some neighborhood 𝑈𝛾 and some 𝑘 ∈ {1, ..., 𝑚} such that for all 𝑡 ∈ 𝑈𝛾 one has the
parametrization
{
𝑥𝑘 (𝑡) = 𝑡
𝛾(𝑡) ∶= (5.1)
𝑥𝑗 (𝑡) = 𝑓𝑗 (𝑡) ∀𝑗 ∈ {1, ..., 𝑚} , 𝑗 ≠ 𝑘 .

From now on, when considering an element of Θ𝑚 , we will always assume implicitly
that it is parametrized as in (5.1).
For fixed 𝑖 ∈ {1, … , 𝑚}, we denote by Θ𝑖𝑚 the subset of curves in Θ𝑚 that can be
parametrized by the 𝑖-th coordinate. Clearly, one has the decomposition
⋃
𝑚
Θ𝑚 = Θ𝑖𝑚 .
𝑖=1

Remark 5.2. We are asking the arc 𝛾 to be analytic at the origin, but the minimal arc
obtained in Theorem 4.1 did not necessarily have this property (the origin was not in-
cluded, in general, in the uniform interval of analyticity 𝙸𝜆 ⊂ [−𝜆, 𝜆]). As it has already
been discussed in the introduction, this is an issue that comes from the use of analytic
reparametrizations of semi-algebraic sets. We will deal with this apparent difficulty in
section 6.
Remark 5.3. For the moment, we do not make any assumption on the sizes of the neigh-
borhoods of analyticity 𝑈𝛾 of the arcs in Θ𝑚 . Hence, the results of this section do not
require any uniform lower bound on |𝑈𝛾 | as in Theorem 4.1. Nevertheless, the exis-
tence of a uniform lower bound will prove to be necessary in order to to demonstrate
the results of section 6.
Definition 5.2. For any pair of integers 𝑠 ≥ 1 and 𝑖 ∈ {1, … , 𝑚}, we indicate by
𝜗𝑚 (𝑠) (resp. 𝜗𝑖𝑚 (𝑠)) the subset of ( ⋆ (𝑠, 1))𝑚 =  ⋆ (𝑠, 1) × ⋯ ×  ⋆ (𝑠, 1) containing the
truncations at order 𝑠 of the Taylor expansions at the origin of all curves in Θ𝑚 (resp.
in Θ𝑖𝑚 ). The elements of 𝜗𝑚 (𝑠) will henceforth be referred to as 𝑠-truncations.
Clearly, one has the following decomposition:
⋃
𝑚
𝜗𝑚 (𝑠) = 𝜗𝑖𝑚 (𝑠) . (5.2)
𝑖=1

Remark 5.4. Clearly, 𝜗𝑚 (𝑠) (resp. 𝜗𝑖𝑚 (𝑠)) is isomorphic to the set of 𝑠-jets of curves
in Θ𝑚 (resp. Θ𝑖𝑚 ). Moreover, for each 𝑖 ∈ {1, … , 𝑚}, the set 𝜗𝑖𝑚 (𝑠) is isomorphic to
ℝ(𝑚−1)×𝑠 , since for any curve 𝛾 ∈ Θ𝑖𝑚 its 𝑠-truncation 𝑠,𝛾 ∈ 𝜗𝑖𝑚 (𝑠) is determined by the
first 𝑠 Taylor coefficients at the origin of the functions 𝑓𝑗 , with 𝑗 ∈ {1, ..., 𝑚}, 𝑗 ≠ 𝑖.

37
Definition 5.3. Fix three integers 𝑟 ≥ 2, 𝑚 ≥ 1, and 1 ≤ 𝑠 ≤ 𝑟 − 1. A polynomial
𝑃 ∈ (𝑟, 𝑚) is said to be 𝑠-vanishing if there exists an arc 𝛾 ∈ Θ𝑚 such that on its
𝑠-truncation 𝑠,𝛾 ∈ 𝜗𝑚 (𝑠) the gradient of 𝑃 has a zero of order 𝑠 at the origin, namely
( )
𝑑𝛼 𝜕𝑃 ||
= 0 , ∀ 𝓁 ∈ {1, ..., 𝑚} , ∀ 𝛼 ∈ {0, ..., 𝑠} . (5.3)
𝑑𝑡𝛼 𝜕𝑥𝓁 ||𝑠,𝛾 (𝑡)
𝑡=0
The set of 𝑠-vanishing polynomials in (𝑟, 𝑚) is denoted by 𝜎(𝑟, 𝑠, 𝑚).
In paragraph 5.2, we shall investigate the properties of the set 𝜎(𝑟, 𝑠, 𝑚) ⊂ (𝑟, 𝑚)
of 𝑠-vanishing polynomials: in particular, we shall prove that
1. for any given values of 𝑟 ≥ 2, 𝑚 ≥ 1, 1 ≤ 𝑠 ≤ 𝑟 − 1, it is the semi-algebraic
projection onto (𝑟, 𝑚) of an algebraic set 𝑍(𝑟, 𝑠, 𝑚) of (𝑟, 𝑚) × 𝜗𝑚 (𝑠) whose
ideal can be explicitly computed ;
2. it has positive codimension.
Secondly, in paragraph 5.3, we shall show that any polynomial 𝑃 belonging to the com-
plementary of the closure of 𝜎(𝑟, 𝑠, 𝑚) in (𝑟, 𝑚) satisfies a "stable" lower estimate on
its gradient. As we shall see, "stable" means that the estimate holds uniformly true for
any polynomial belonging to a neighborhood of 𝑃 .
Finally, in section 6, we shall prove that a polynomial 𝑄 ∈ (𝑟, 𝑛) satisfying
∇𝑄(0) ≠ 0 is steep around the origin iff there exists 1 ≤ 𝑠 ≤ 𝑟 − 1 such that, for
all 𝑚 ∈ {1, ..., 𝑛 − 1}, the restriction of 𝑄 to any 𝑚-dimensional linear subspace Γ𝑚 per-
pendicular to ∇𝑄(0) is contained in the complementary of closure(𝜎(𝑟, 𝑠, 𝑚)) in (𝑟, 𝑚).

5.2 Algebraic properties

We assume the notations of the previous paragraph, and we consider a triplet of integers
𝑟 ≥ 2, 𝑚 ≥ 1, 1 ≤ 𝑠 ≤ 𝑟 − 1. We work in the euclidean space ℝ𝑚 equipped with co-
ordinates (𝑥1 , ..., 𝑥𝑚 ), and we consider a polynomial 𝑃 = 𝑃 (𝑥) ∈ (𝑟, 𝑚) verifying the
𝑠-vanishing condition on the 𝑠-truncation 𝑠,𝛾 ∈ 𝜗𝑚 (𝑠) of some curve 𝛾 ∈ Θ𝑚 . Unless
explicitly specified, we will henceforth work in the case in which 𝛾 is parametrized by
the first coordinate, as the generalization to other cases is immediate. Hence, from now
on we set 𝛾(𝑡) ∶= (𝑡, 𝑥2 (𝑡), … , 𝑥𝑚 (𝑡)) ∈ Θ1𝑚 .

5.2.1 Case 𝑚 = 1
We observe that, for 𝑚 = 1, we have the following simple result:
∑ 𝜇
Lemma 5.1. For 𝑚 = 1, a polynomial 𝑃 (𝑥) = 𝜇∈ℕ 𝑝𝜇 𝑥 of one real variable
1≤𝜇≤𝑟
belongs to the set 𝜎(𝑟, 𝑠, 1) ⊂ (𝑟, 1) if and only if
𝑝𝜇 = 0 ∀𝜇 ∈ ℕ such that 1 ≤ 𝜇 ≤ 𝑠 + 1 . (5.4)
Moreover, 𝜎(𝑟, 𝑠, 1) is closed and its codimension in (𝑟, 1) is equal to 𝑠 + 1.

38
Proof. For 𝑚 = 1, the set Θ1 is the singleton containing the line 𝛾(𝑡) ∶= 𝑥(𝑡) = 𝑡.
By Definition 5.3, it is clear that a polynomial verifying condition (5.4) in the state-
ment satisfies also the 𝑠-vanishing condition. Conversely, again by Definition 5.3, it is
plain to check that the 𝑠-vanishing condition for 𝑚 = 1 imposes that the coefficients
of the studied polynomial must be null up to order 𝑠 + 1. The closure of 𝜎(𝑟, 𝑠, 1) is
due to continuity, whereas codim 𝜎(𝑟, 𝑠, 1) = 𝑠 + 1 as such a set is determined by 𝑠 + 1
independent equations in (𝑟, 1).

5.2.2 Notations (case 𝑛 ≥ 3, 2 ≤ 𝑚 ≤ 𝑛 − 1)

Up to the end of this paragraph, we will restrict to the case 𝑛 ≥ 3, 2 ≤ 𝑚 ≤ 𝑛 − 1.
The goal is to introduce useful notations in order to study the properties of 𝑠-vanishing
polynomials in (𝑟, 𝑚).
Using standard notations, we set 𝜇 ∶= (𝜇1 , ..., 𝜇𝑚 ) ∈ ℕ𝑚 , |𝜇| ∶= ||𝜇||1 , and for
any 𝑃 ∈ (𝑟, 𝑚) we write
∑ 𝜇 𝜇 ∑
𝑃 (𝑥) ∶= 𝑝(𝜇1 ,...,𝜇𝑚 ) 𝑥1 1 ... 𝑥𝑚𝑚 =∶ 𝑝𝜇 𝑥 𝜇 , (5.5)
𝜇∈ℕ𝑚 𝜇∈ℕ𝑚
1≤|𝜇|≤𝑟 1≤|𝜇|≤𝑟

where we have taken into account the fact that 𝑃 has no constant term by the definition
of (𝑟, 𝑚).
We also consider a curve 𝛾 ∈ Θ1𝑚 and - for 𝑗 = 2, ..., 𝑚 - we develop its components
𝑓𝑗 at the origin, and we write
∑
+∞
𝑥𝑗 (𝑡) = 𝑓𝑗 (𝑡) =∶ 𝑎𝑗𝑖 𝑡𝑖 , (5.6)
𝑖=1

where we have taken into account the fact that 𝛾(0) = 0 by Definition 5.1. Thus, the
𝑠-truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) of the curve 𝛾 is identified by the (𝑚 − 1)𝑠 real coefficients
(𝑎21 , ..., 𝑎2𝑠 , ..., 𝑎𝑚1 , ..., 𝑎𝑚𝑠 ) of the truncated expansion, namely
( 𝑠 )
∑ ∑
𝑠
𝑠,𝛾 = 𝑠,𝛾 (𝑡) = 𝑡, 𝑎2𝑖 𝑡𝑖 , … , 𝑎𝑚𝑖 𝑡𝑖 . (5.7)
𝑖=1 𝑖=1

In the rest of this paragraph, we will try to find an explicit expression for the 𝑠-
vanishing condition in terms of the coefficients of 𝑃 and 𝑠,𝛾 . We first observe that the
𝑠-vanishing condition (5.3) for 𝛼 = 0 implies
𝑝𝜇 = 0 for all 𝜇 ∈ ℕ𝑚 such that |𝜇| = 1 . (5.8)
Thus, without any loss of generality, in (5.5) we can only consider the multi-indices
𝜇 ∈ ℕ𝑚 that satisfy 2 ≤ |𝜇| ≤ 𝑟. Moreover, for 𝓁 = 1, ..., 𝑚, the 𝓁-th component of the
gradient of 𝑃 reads
𝜕𝑃 ∑
∶= 𝜇𝓁 𝑝𝜇 𝑥𝜇̃(𝓁) , 𝜇̃𝑗 (𝓁) ∶= 𝜇𝑗 −𝛿𝑗𝓁 , 𝑗 = 1, ..., 𝑚 , |𝜇| = |̃
𝜇 (𝓁)|+1 , (5.9)
𝜕𝑥𝓁 𝜇∈ℕ𝑚
2≤|𝜇|≤𝑟

39
where 𝛿𝑗𝓁 is the Kronecker symbol. At this point, we indicate by
Φ1 ∶ (𝑟, 𝑚) × 𝜗1𝑚 (𝑠) ⟶ ℝ𝑀 × ℝ(𝑚−1)𝑠 , 𝑀 ∶= dim (𝑟, 𝑚) (5.10)
the trivial chart associating (𝑃 , 𝑠,𝛾 ) ⟼ (𝑝𝜇 , 𝑎21 , ..., 𝑎2𝑠 , ..., 𝑎𝑚1 , ..., 𝑎𝑚𝑠 ) and we define
the functions 𝑞𝓁𝛼1 ∶ ℝ𝑀 × ℝ(𝑚−1)𝑠 ⟶ ℝ , 𝓁 ∈ {1, ..., 𝑚} , 𝛼 ∈ {0, ..., 𝑠} in the

following way:

( )
1 𝜕𝑃 ||
For 𝛼 = 0 𝑞𝓁0 ◦Φ1 (𝑃 , 𝑠,𝛾 ) ∶= = 𝑝(0,...,0,1,0,...,0)
𝜕𝑥𝓁 ||𝑠,𝛾 (𝑡)
( 𝑡=0 )
1 1 𝑑 𝛼 𝜕𝑃 ||
For 𝛼 ∈ {1, ..., 𝑠} 𝑞𝓁𝛼 ◦Φ (𝑃 , 𝑠,𝛾 ) ∶= 𝛼
𝑑𝑡 𝜕𝑥𝓁 ||𝑠,𝛾 (𝑡) (5.11)
𝑡=0
⎡ ( 𝑠 )𝜇̃2 (𝓁) ( 𝑠 )𝜇̃𝑚 (𝓁) ⎤
𝑑𝛼 ⎢ ∑ ∑ ∑ ⎥
= 𝛼⎢ 𝜇 𝑝 𝑡𝜇̃1 (𝓁) 𝑎2𝑘 𝑡𝑘 ... 𝑎𝑚𝑗 𝑡𝑗 ⎥
𝑑𝑡 ⎢ 𝜇∈ℕ𝑚 𝓁 𝜇 𝑘=1 𝑗=1 ⎥
⎣2≤|𝜇|≤𝑟 ⎦𝑡=0
where the "1" fills the 𝓁-th slot in the multi-index at the rightest member of the first line
and where the last line is obtained by injecting (5.7) into expression (5.9).
Remark 5.5. In a similar way, when 𝛾 is parametrized by the 𝑖-th coordinate, with 𝑖 ≠ 1,
one can denote by Φ𝑖 ∶ (𝑟, 𝑚) × 𝜗𝑖𝑚 (𝑠) ⟶ ℝ𝑀 × ℝ(𝑚−1)𝑠 the chart associating
(𝑃 , 𝑠,𝛾 ) ↦ (𝑝𝜇 , 𝑎11 , ..., 𝑎1𝑠 , ..., 𝑎(𝑖−1)1 , ..., 𝑎(𝑖−1)𝑠 , 𝑎(𝑖+1)1 , ..., 𝑎(𝑖+1)𝑠 , ..., 𝑎𝑚1 , ..., 𝑎𝑚𝑠 ), and
introduce the maps 𝑞𝓁𝛼 𝑖 ◦Φ𝑖 (𝑃 ,  ), 𝓁 ∈ {1, … , 𝑚}, 𝛼 ∈ {0, … , 𝑠} exchanging the
𝑠,𝛾
rôle of the first coordinate with that of the 𝑖-th coordinate in (5.11).
Remark 5.6. Comparing expressions (5.11) with Definition 5.3 and expression (5.8),
and taking Remark 5.5 into account, it is easy to see that any given polynomial 𝑃 ∈
(𝑟, 𝑚) satisfies the 𝑠-vanishing condition on some truncation 𝑠,𝛾 ∈ 𝜗𝑚 (𝑠) if and only
if
𝑖
𝑞𝓁𝛼 ◦Φ𝑖 (𝑃 , 𝑠,𝛾 ) = 0 , for some 𝑖 ∈ {2, … , 𝑚} , for all 𝓁 ∈ {1, ..., 𝑚} , 𝛼 ∈ {0, ..., 𝑠} .
(5.12)
By the above discussion, we see that the set of 𝑠-vanishing polynomials in (𝑟, 𝑚)
is given by
⋃𝑚
𝜎(𝑟, 𝑠, 𝑚) = 𝜎 𝑖 (𝑟, 𝑠, 𝑚) , (5.13)
𝑖=1
where we have introduced the sets
𝜎 𝑖 (𝑟, 𝑠, 𝑚) ∶= Π(𝑟,𝑚) 𝑍 𝑖 (𝑟, 𝑠, 𝑚) (5.14)
and
𝑍 𝑖 (𝑟, 𝑠, 𝑚) ∶= {(𝑃 , 𝑠,𝛾 ) ∈ (𝑟, 𝑚) × 𝜗𝑖𝑚 (𝑠)}|
𝑖
(5.15)
𝑞𝓁𝛼 ◦Φ𝑖 (𝑃 , 𝑠,𝛾 ) = 0 for all 𝓁 ∈ {1, ..., 𝑚}, 𝛼 ∈ {0, ..., 𝑠}} .

40
We also set
⋃
𝑚
𝑍(𝑟, 𝑠, 𝑚) ∶= 𝑍 𝑖 (𝑟, 𝑠, 𝑚) . (5.16)
𝑖=1
It turns out that the ideal of 𝑍(𝑟, 𝑠, 𝑚) can be explicitly computed for any given value
of the integers 𝑟 ≥ 2, 1 ≤ 𝑠 ≤ 𝑟 − 1, and 𝑚 ≥ 2 13 , i.e., one can find explicit expressions
𝑖 ◦Φ𝑖 (𝑃 ,  ), for any value of 𝓁 ∈ {1, ..., 𝑚}, 𝛼 ∈ {0, ..., 𝑠}, and
for the quantities 𝑞𝓁𝛼 𝑠,𝛾
𝑖 ∈ {1, ..., 𝑚}.
Before stating this result, for any given value of 𝑖 ∈ {1, … , 𝑚} we will introduce
new global charts for (𝑟, 𝑚) × 𝜗𝑖𝑚 (𝑠) which - though unessential for the validity of our
𝑖 ◦Φ𝑖 (𝑃 ,  ) = 0 than the standard
results - yield nicer expressions for the equations 𝑞𝓁𝛼 𝑠,𝛾
𝑖
chart Φ . As it will be shown in the next paragraph, the variables 𝑎𝑗1 , 𝑗 ≠ 𝑖, associated
to the linear terms of the 𝑠-truncation 𝑠,𝛾 can be incorporated in the coordinates of the
polynomial 𝑃 . This simplifies the calculations and yields more readable formulas.

5.2.3 A useful chart for (𝑟, 𝑚) × 𝜗𝑚 (𝑠) (case 𝑛 ≥ 3, 2 ≤ 𝑚 ≤ 𝑛 − 1)

Here too, we restrict to the case 𝑛 ≥ 3, 2 ≤ 𝑚 ≤ 𝑛 − 1.
Once again, we only consider the case in which 𝛾 is parametrized by the coordinate
𝑖 = 1, the other cases being trivial generalizations. Some of the quantities introduced in
the sequel should be labeled with an index 1, as their definition depends in an obvious
way from the choice of the parametrizing coordinate. However, in order not to burden
notations, we drop, when possible, the reference to the fact that we are considering the
case 𝑖 = 1.
In order to define a new chart for (𝑟, 𝑚) × 𝜗1𝑚 (𝑠), we start by observing that, if we
denote by 𝐴1 , ..., 𝐴𝑚 the canonical basis associated to the coordinates 𝑥1 , ..., 𝑥𝑚 in ℝ𝑚 ,
for any fixed vector 𝚋 ∶= (𝑏21 , ..., 𝑏𝑚1 ) ∈ ℝ𝑚−1 , we can define the new parametric basis

𝑣𝚋 ∶= 𝐴1 + 𝑏21 𝐴2 + 𝑏31 𝐴3 + ... + 𝑏𝑚1 𝐴𝑚 , 𝑢2 ∶= 𝐴2 , ... , 𝑢𝑚 ∶= 𝐴𝑚 (5.17)

associated to the parametric change of variables

𝚋 ∶ ℝ𝑚 ⟶ ℝ𝑚 , (𝑥1 , ..., 𝑥𝑚 ) ⟼ 𝚢(𝚋) ∶= (𝚢1 , 𝚢2 (𝚋), ..., 𝚢𝑚 (𝚋)), (5.18)

where

𝚢1 ∶= 𝑥1 , 𝚢2 = 𝚢2 (𝚋) ∶= 𝑥2 − 𝑏21 𝑥1 …
𝚢𝑚 = 𝚢𝑚 (𝚋) ∶= 𝑥𝑚 − 𝑏𝑚1 𝑥1 .
(5.19)
Obviously, for any fixed 𝚋 ∈ ℝ𝑚−1 , the change of coordinates (5.19) in ℝ𝑚 induces a
change of coordinates also in (𝑟, 𝑚). Infact, the pull-back of the polynomial 𝑃 (𝑥) is
indicated by
∑ ∑
𝙿𝚋 (𝚢(𝚋)) ∶= 𝑃 ◦−1
𝚋 (𝚢(𝚋)) = 𝑝𝜇 (𝑥(𝚢(𝚋)))𝜇 =∶ 𝚙𝜇 (𝑝𝜇 , 𝚋)𝚢𝜇 (𝚋) , (5.20)
𝜇∈ℕ𝑚 𝜇∈ℕ𝑚
2≤|𝜇|≤𝑟 2≤|𝜇|≤𝑟
13 The case 𝑚 = 1 is easier, see Lemma 5.1.

41
where the new coefficients 𝚙𝜇 = 𝚙𝜇 (𝑝𝜇 , 𝑏21 , ...., 𝑏𝑚1 ) are polynomial functions of coef-
ficients 𝑝𝜇 and on the parameters 𝚋 ∈ ℝ𝑚−1 . For any given 𝚋 ∈ ℝ𝑚−1 , there is a 1 − 1
correspondence between the quantities 𝑝𝜇 and 𝚙𝜇 , as they represent the coordinates of
the same polynomial written in different bases. Moreover, by (5.6), in the new vari-
ables (5.19) the components of the push-forward 𝚋 ◦𝑠,𝛾 = 𝚋 ◦𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) of any
𝑠-truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) read

∑
𝑠
𝚢1 (𝑡) = 𝑡 , 𝚢𝑗 (𝑡) ∶= (𝑎𝑗1 − 𝑏𝑗1 )𝑡 + 𝑎𝑗𝑖 𝑡𝑖 , 𝑗 ∈ {2, ..., 𝑚} . (5.21)
𝑖=2

By looking at expression (5.21), we see that, for any 𝑠 ≥ 2 and for any given 𝑠-
truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠), it is possible to find a parametric change of coordinates 𝚋 in
ℝ𝑚 such that the image 𝚋 ◦𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) has no linear terms except for the parametriz-
ing component: it suffices to choose 𝚋 = 𝚊 ∶= (𝑎21 , ..., 𝑎𝑚1 ), with (𝑎21 , ..., 𝑎𝑚1 ) the
coefficients of the linear terms of 𝑠,𝛾 . Taking (5.20) into account, this also entails
that, for any given truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠), there exists an associated set of coordinates
𝚙𝜇 = 𝚙𝜇 (𝑝𝜇 , 𝚊), with 𝚊 ∶= (𝑎21 , ..., 𝑎𝑚1 ), in the space of polynomials (𝑟, 𝑚). Hence,
we can parametrize the coordinates of any polynomial 𝑃 through the linear coefficients
of any truncation 𝑠,𝛾 .
Namely, we firstly observe that 𝜗1𝑚 (1) is isomorphic to

𝜗1𝑚 (1) ≃ {𝚊 ∶= (𝑎21 , … , 𝑎𝑚1 ) ∈ ℝ𝑚−1 } . (5.22)

Also, indicating by 𝜗1𝑚 (𝑠, 2) the subset of 𝑠-truncations having null linear terms, we have
that
𝜗1𝑚 (𝑠, 2) ≃ {(𝑎22 , … , 𝑎2𝑠 , … , 𝑎𝑚2 , … , 𝑎𝑚𝑠 ) ∈ ℝ(𝑚−1)(𝑠−1) } . (5.23)
Clearly, one has
𝜗1𝑚 (𝑠) = 𝜗1𝑚 (1) × 𝜗1𝑚 (𝑠, 2) . (5.24)
At this point, with the notation in (5.20), we define the invertible transformation

ℱ 1 ∶ (𝑟, 𝑚) × ℝ𝑚−1 ⟶ ℝ𝑀 × ℝ𝑚−1 , 𝑀 ∶= dim (𝑟, 𝑚) (5.25)

associating
⎛ ⎞ ⎛ ( ) ⎞
⎜ ∑ ⎟ ⎜ ⎟
⎜𝑃 = 𝑝𝜇 𝑥𝜇 , 𝚊⎟ ⟼ ⎜𝚙𝜇′ (𝑝𝜇 ) 𝜇∈ℕ𝑚 , 𝚊 , 𝚊⎟ (5.26)
⎜ 𝜇∈ℕ𝑚 ⎟ ⎜ 1≤|𝜇|≤𝑟 ′
𝜇 ∈ℕ 𝑚
⎟
⎝ 1≤|𝜇|≤𝑟 ⎠ ⎝ 1≤|𝜇 ′ |≤𝑟 ⎠
and we indicate its image by

𝑊 1 (𝑟, 𝑚) ∶= ℱ 1 ((𝑟, 𝑚) × ℝ𝑚−1 ) . (5.27)

In other words, 𝑊 1 (𝑟, 𝑚) is constructed by attaching to any point 𝚊 ∈ ℝ𝑚−1 ≃ 𝜗1𝑚 (1)
the fiber of all polynomials in (𝑟, 𝑚) expressed in the variables (5.19) associated to
the value 𝚊.

42
 Furthermore, setting
𝙹𝑠,𝛾,𝚊 ∶= 𝚊 ◦𝑠,𝛾 (5.28)
we have 𝙹𝑠,𝛾,𝚊 ∈ 𝜗1𝑚 (𝑠, 2) by construction because in the adapted variables - as we had
shown in (5.21) by setting 𝚋 = 𝚊 - with the exception of the parametrizing component,
any truncation starts at order two. Taking the notation in (5.20) into account, we can
also define the chart
Υ1 ∶ (𝑟, 𝑚) × 𝜗1𝑚 (𝑠) ⟶ 𝑊 1 (𝑟, 𝑚) × ℝ(𝑚−1)(𝑠−1) (5.29)
associating
⎛ ( ) ⎞
⎜ ⎟
(𝑃 , 𝑠,𝛾 ) ⟼ ⎜𝚙𝜇′ (𝑝𝜇 ) 𝜇∈ℕ𝑚 , 𝚊 , 𝚊, 𝑎22 , 𝑎23 , ..., 𝑎𝑚𝑠 ⎟ . (5.30)
⎜ 1≤|𝜇|≤𝑟 𝜇′ ∈ℕ𝑚 ⎟
⎝ 1≤|𝜇′ |≤𝑟 ⎠
Remark 5.7. For further convenience, we also denote by
𝔘1 ∶ (𝑟, 𝑚) × 𝜗1𝑚 (1) ⟶ 𝑊 1 (𝑟, 𝑚) (5.31)
the restriction of Υ1 to (𝑟, 𝑚) × 𝜗1𝑚 (1). It is plain to check by formulas (5.20), (5.26),
and (5.30) that 𝔘1 is polynomial, invertible and that its inverse is a polynomial map.
Remark 5.8. The generalization of the arguments above to the case in which the curve
𝛾 ∈ Θ𝑚 is parametrized by the 𝑖-th coordinate, with 𝑖 ∈ {2, … , 𝑚}, is immediate. In
particular, one can define functions ℱ 𝑖 , Υ𝑖 , 𝔘𝑖 , together with sets 𝑊 𝑖 (𝑟, 𝑚).
Remark 5.9. With slight abuse of notation, in the rest of this work we will often write
(𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) and (𝙿𝚊 , 𝚊) to indicate a point of Υ1 (𝑃 , 𝑠,𝛾 ) and 𝑊 1 (𝑟, 𝑚) respectively.
For further convenience, we also observe that
Lemma 5.2. Any polynomial 𝑃 ∈ (𝑟, 𝑚) satisfies the 𝑠-vanishing condition
( )
𝑑𝛼 𝜕𝑃 (𝑥) ||
= 0 ∀𝛼 ∈ {0, ..., 𝑠} , ∀𝓁 ∈ {1, ..., 𝑚} (5.32)
𝑑𝑡𝛼 𝜕𝑥𝓁 ||𝑠,𝛾 (𝑡)
𝑡=0

on the 𝑠-truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) of some curve 𝛾 ∈ Θ1𝑚 , if and only if it satisfies
( )
𝑑𝛼 𝜕𝙿𝚊 (𝚢) ||
= 0 ∀𝛼 ∈ {0, ..., 𝑠} , ∀𝓁 ∈ {1, ..., 𝑚}
𝑑𝑡𝛼 𝜕𝚢𝓁 ||𝙹𝑠,𝛾,𝚊 (𝑡)
𝑡=0
in the adapted coordinates associated to the linear terms of 𝑠,𝛾 .
Proof. By (5.19) one has
⎧𝑥 = 𝚢 ⎛ 1 0 0 ... 0⎞
⎪ 1 1
⎜𝑎 1 0 ... 0⎟
⎪𝑥2 = 𝑎21 𝚢1 + 𝚢2 ⎜ 21 ⎟
−1
𝚊 (𝚢) ∶= ⎨
−1
, 𝚊 (𝚢) = ⎜ 𝑎31 0 1 ... 0⎟ . (5.33)
⎪... ⎜ ... ... ... ... ...⎟
⎪𝑥 = 𝑎 𝚢 + 𝚢 ⎜ ⎟
⎩ 𝑚 𝑚1 1 𝑚 ⎝𝑎𝑚1 0 0 ... 1⎠

43
Indicating with (−1
𝚊 (𝚢))𝑘,𝓁 the (𝑘, 𝓁)-th entry of the Jacobian of the inverse transfor-
mation −1
𝚊 in (5.33), by the Leibniz formula one has
( )
⎛ ⎞
𝑑 𝛼 ⎜ 𝜕(𝑃 ◦𝚊 (𝚢)) ||
−1
𝑑𝛼 𝜕𝙿𝚊 (𝚢) || ⎟
∶= |
𝑑𝑡𝛼 𝜕𝚢𝓁 ||𝙹𝑠,𝛾,𝚊 (𝑡) 𝑑𝑡𝛼 ⎜ 𝜕𝚢𝓁 |
| ⎟
𝑡=0 ⎝ 𝙹 𝑠,𝛾,𝚊 ⎠𝑡=0
(𝑡)
[𝑚 ]
𝑑 𝛼 ∑ 𝜕𝑃 −1
|
| −1 | (5.34)
= 𝛼 ◦ (𝚢)| × (𝚊 (𝚢))𝑘,𝓁 |
𝑑𝑡 𝑘=1 𝜕𝑥𝑘 𝚊 |𝚊 ◦𝑠,𝛾 (𝑡) |𝙹𝑠,𝛾,𝚊 (𝑡)
𝑡=0
[𝑚 ]
𝑑 𝛼 ∑ 𝜕𝑃 (𝑥) || |
= 𝛼 × (−1
𝚊 (𝚢))𝑘,𝓁 ||𝙹 .
𝑑𝑡 𝑘=1 𝜕𝑥𝑘 ||𝑠,𝛾 (𝑡) 𝑠,𝛾,𝚊 (𝑡)
𝑡=0

Since the entries of the matrix −1

𝚊 in (5.33) are constant, one has
|
(−1
𝚊 (𝚢))𝑘,𝓁 ||𝙹 = (−1
𝚊 (𝚢))𝑘,𝓁 ,
𝑠,𝛾,𝚊 (𝑡)

so that, finally, (5.34) reads

( ) [ ]
𝑑𝛼 𝜕𝙿𝚊 (𝚢) || ∑𝑚
𝑑 𝛼 𝜕𝑃 (𝑥) ||
= × (−1
𝚊 (𝚢))𝑘,𝓁 . (5.35)
𝑑𝑡𝛼 𝜕𝚢𝓁 ||𝙹𝑠,𝛾,𝚊 (𝑡) 𝑘=1
𝑑𝑡𝛼 𝜕𝑥𝑘
|
|𝑠,𝛾 (𝑡)
𝑡=0 𝑡=0
By the expression above, it is immediate to check that if 𝑃 satisfies the 𝑠-vanishing
condition on 𝑠,𝛾 then 𝙿𝚊 does the same on 𝙹𝑠,𝛾,𝚊 . The proof of the converse is immediate
by applying the same arguments to 𝑃 (𝑥) ≡ 𝙿𝚊 ◦𝚊 (𝑥).

By the discussion above, one has the choice to write the equations determining
the set 𝑍 1 (𝑟, 𝑠, 𝑚) in (5.15) either in the original coordinates, where they assume the
1 ◦Φ(𝑃 ,  ) = 0 for all 𝓁 ∈ {1..., 𝑚} and 𝛼 ∈ {0, ..., 𝑠}, or in the new set
form 𝑞𝓁𝛼 𝑠,𝛾
of coordinates associated to the change of variables 𝚊 in ℝ𝑚 , defined in (5.19). In
particular, by performing the same computations that led to expression (5.11) in the
new variables, and by taking into account the fact that the expansion of 𝙹𝑠,𝛾,𝚊 (𝑡) starts
at order two in 𝑡, one can introduce the functions
𝚀𝓁𝛼 ∶ 𝑊 1 (𝑟, 𝑚) × ℝ(𝑚−1)(𝑠−1) ⟶ ℝ , 𝓁 ∈ {1, … , 𝑚} 𝛼 ∈ {0, … , 𝑠} (5.36)
in the following way:
For 𝛼 = 0, 𝚀𝓁0 ◦Υ1 (𝑃 , 𝑠,𝛾 ) = 𝚀𝓁0 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ∶= 𝚙(0,…,0,1,0,…,0)
( )
1 𝑑𝛼 𝜕𝙿𝚊 (𝚢) ||
For 𝛼 ∈ {1, ..., 𝑠}, 𝚀𝓁𝛼 ◦Υ (𝑃 , 𝑠,𝛾 ) = 𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ∶= 𝛼
𝑑𝑡 𝜕𝚢𝓁 ||𝙹𝑠,𝛾,𝚊 (𝑡)
𝑡=0
⎡ ( 𝑠 )𝜇̃2 (𝓁) ( 𝑠 )𝜇̃𝑚 (𝓁) ⎤
𝑑𝛼 ⎢ ∑ ∑ ∑ ⎥
= 𝛼⎢ 𝜇 𝚙 𝑡𝜇̃1 (𝓁) 𝑎2𝑘 𝑡𝑘 … 𝑎𝑚𝑗 𝑡𝑗 ⎥ .
𝑑𝑡 ⎢ 𝜇∈ℕ𝑚 𝓁 𝜇 𝑘=2 𝑗=2 ⎥
⎣2≤|𝜇|≤𝑟 ⎦𝑡=0
(5.37)

44
Expressions (5.11) and (5.37), considered together with Lemma 5.2, imply that condi-
1 ◦Φ1 (𝑃 ,  ) = 0 holds if and only if
tion 𝑞𝓁𝛼 𝑠,𝛾

𝚀𝓁𝛼 ◦Υ1 (𝑃 , 𝑠,𝛾 ) = 𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 , ∀𝓁 ∈ {1, … , 𝑚}, ∀𝛼 ∈ {0, … , 𝑠} , (5.38)

so that the ideal of the set 𝑍 1 (𝑟, 𝑠, 𝑚) in the new variables is given by the equations
𝚀𝓁𝛼 ◦Υ1 (𝑃 , 𝑠,𝛾 ) = 𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 for all 𝓁 ∈ {1..., 𝑚} and 𝛼 ∈ {0, ..., 𝑠}. In the
sequel, we will work in these new coordinates, since the involved expressions are nicer.

5.2.4 Computations and estimate on the codimension of 𝜎(𝑟, 𝑠, 𝑚) (case 𝑛 ≥ 3,

2 ≤ 𝑚 ≤ 𝑛 − 1)
As in the previous paragraphs, we set 𝑛 ≥ 3, 2 ≤ 𝑚 ≤ 𝑛 − 1.
Once again, we will only consider the case in which 𝛾 is parametrized by the first
coordinate.
Before stating the main results of this paragraph, we still need to introduce a few no-
tations. For any fixed 𝛼 ∈ {0, … , 𝑠}, for any 𝛽 ∈ {0, … , 𝛼}, and for any 𝑖 ∈ {1, … , 𝑚},
we set
{
(𝛽 + 1, 0, … , 0) , for 𝑖 = 1
𝜈(𝑖, 𝛽) ∶= (5.39)
(𝛽, 0, … , 0, 1, 0, … , 0) , for 𝑖 = 2, … , 𝑚

where the "1" fills the 𝑖-th slot for 𝑖 = 2, ..., 𝑚. For 𝛼 ∈ {1, ..., 𝑠}, we also denote the
multi-indices 𝜇 ∈ ℕ𝑚 of length 2 ≤ |𝜇| ≤ 𝛼 + 1 not belonging to this family with
⋃
𝑚 (𝛼) ∶= {𝜇 ∈ ℕ𝑚 , 2 ≤ |𝜇| ≤ 𝛼 + 1}∖ {𝜈(𝑖, 𝛽)} . (5.40)
𝑖=1,...,𝑚
𝛽=1,...,𝛼

Moreover, for any given 𝛼 ∈ {1, .., 𝑠}, 𝜇 ∈ ℕ𝑚 and 𝓁 ∈ {1, ..., 𝑚} we introduce
{
𝜇 (𝓁), 𝛼) ∶= (𝑘𝑗2 , ..., 𝑘𝑗𝛼 ) ∈ ℕ(𝑚−1)×(𝛼−1) , 𝑗 ∈ {2, ..., 𝑚} ∶
𝑚 (̃
} (5.41)
∑
𝛼 ∑
𝑚 ∑
𝛼
𝑘𝑗𝑖 = 𝜇̃𝑗 (𝓁) , 𝜇̃1 (𝓁) + 𝑖 𝑘𝑗𝑖 = 𝛼
𝑖=2 𝑗=2 𝑖=2

and we set
𝑚 (𝓁, 𝛼) ∶= {𝜇 ∈ ℕ𝑚 | 𝑚 (̃
𝜇 (𝓁), 𝛼) ≠ ∅} . (5.42)
Finally, for any 𝓁 ∈ {1, … , 𝑚} and for any 𝜇 ∈ ℕ𝑚 , we remind that (see (5.9))

𝜇̃𝑗 (𝓁) ∶= 𝜇𝑗 − 𝛿𝑗𝓁 𝑗 ∈ {1, … , 𝑚} .

With these notations, we can now state the following

45
Theorem 5.1. For any choice of integers 𝑚 ≥ 1, 𝑟 ≥ 2, 1 ≤ 𝑠 ≤ 𝑟 − 1, the set
𝑍(𝑟, 𝑠, 𝑚) in (5.16) is an algebraic set of (𝑟, 𝑚) × 𝜗𝑚 (𝑠), whose ideal can be explicitly
computed. In particular, with the notations in (5.37), (5.39), (5.40) and (5.41), the set
𝑍 1 (𝑟, 𝑠, 𝑚) in (5.15) is the image through the inverse of the transformation Υ1 in (5.30)
of the algebraic set determined by the following equations:
𝚀10 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 𝚙𝜈(1,0) = 0 ,
𝚀11 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 2 𝚙𝜈(1,1) = 0 ,
𝚀1𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ∑ ∑
𝑚 𝛼−1
= (𝛼 + 1)𝚙𝜈(1,𝛼) + 𝛽 𝚙𝜈(𝑖,𝛽) 𝑎𝑖(𝛼−(𝛽−1))
𝛼! 𝑖=2 𝛽=1
( ( ) )
∑ ∑ ∏
𝑚
𝜇̃𝑗 (1) 𝑘𝑗2 𝑘𝑗𝛼
+ 𝜇1 𝚙𝜇 𝑎𝑗2 ...𝑎𝑗𝛼 =0,
𝜇∈𝑚 (1,𝛼) 𝑘∈𝑚 (̃
𝜇(1),𝛼) 𝑗=2 𝑘𝑗2 ... 𝑘𝑗𝛼
𝜇∈𝑚 (𝛼)
𝜇1 ≠0
(5.43)
for 𝓁 = 1, 𝛼 = 2, ..., 𝑠, and
𝚀𝓁0 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 𝚙𝜈(𝓁,0) = 0 ,
𝚀𝓁1 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 𝚙𝜈(𝓁,1) = 0 ,
𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 )
= 𝚙𝜈(𝓁,𝛼)
𝛼! ( )
( )
∑ ∑ ∏
𝑚
𝜇̃𝑗 (𝓁) 𝑘𝑗2 𝑘𝑗𝛼
+ 𝜇𝓁 𝚙𝜇 𝑎𝑗2 ...𝑎𝑗𝛼 =0,
𝜇∈𝑚 (𝓁,𝛼) 𝑘∈𝑚 (̃
𝜇(𝓁),𝛼) 𝑗=2 𝑘𝑗2 ... 𝑘𝑗𝛼
𝜇∈𝑚 (𝛼)
𝜇𝓁 ≠0

for 𝓁 = 2, ..., 𝑚 , 𝛼 = 2, ..., 𝑠 .

(5.44)
Remark 5.10. It is plain to check that the coefficients of the vector 𝚊 ∈ 𝜗1𝑚 (1) containing
the linear terms of the truncation 𝑠,𝛾 do not appear explicitly in expressions (5.43)-
(5.44). However, they are "hidden" in the terms 𝚙𝜇 = 𝚙𝜇 (𝑝𝜇 , 𝚊) (see (5.20)).
As an almost immediate consequence of Theorem 5.1, we have that 𝑠-vanishing
polynomials are rare in (𝑟, 𝑚), namely
Corollary 5.1. 𝑍(𝑟, 𝑠, 𝑚) has codimension 𝑚(𝑠 + 1) in (𝑟, 𝑚) × 𝜗𝑚 (𝑠) and 𝜎(𝑟, 𝑠, 𝑚)
is a semi-algebraic set of codimension 𝑠 + 𝑚 in (𝑟, 𝑚).
Proof. (Theorem 5.1) For fixed 𝑠 ∈ {1, ..., 𝑟−1}, we consider a polynomial 𝑃 ∈ (𝑟, 𝑚)
verifying the 𝑠-vanishing condition on some truncation 𝑠,𝛾 (𝑡) ∈ 𝜗1𝑚 (𝑠).
Step 1. By Lemma 5.2, in the adapted coordinates (5.19) one must have
( )
𝑑𝛼 𝜕𝙿𝚊 (𝚢) ||
=0 ∀𝛼 ∈ {0, ..., 𝑠} , ∀𝓁 ∈ {1, ..., 𝑚} . (5.45)
𝑑𝑡𝛼 𝜕𝚢𝓁 ||𝙹𝑠,𝛾,𝚊 (𝑡)
𝑡=0

46
For 𝛼 = 0, it is plain to check that the terms of order zero in 𝑡 in (5.45) are the linear
terms of 𝙿𝚊 , for which |𝜇| = 1. Expressions (5.45) and (5.37) yield the thesis for this
value of 𝛼.
Then, as we did in (5.9), we drop the linear terms in 𝙿𝚊 and we write the quantity
𝜕𝙿𝚊 (𝚢)∕𝜕𝚢𝓁 explicitly

𝜕𝙿𝚊 (𝚢) ∑
∶= 𝜇𝓁 𝚙𝜇 𝚢𝜇̃(𝓁) , 𝜇̃𝑗 (𝓁) ∶= 𝜇𝑗 − 𝛿𝑗𝓁 , 𝑗 = 1, ..., 𝑚 , |𝜇| = |̃
𝜇(𝓁)| + 1 .
𝜕𝚢𝓁 𝜇∈ℕ𝑚
2≤|𝜇|≤𝑟
(5.46)
Taking (5.21) into account (with 𝚋 ∶= (𝑏21 , … , 𝑏𝑚1 ) = 𝚊 ∶= (𝑎21 , … , 𝑎𝑚1 )), we inject
in (5.46) the components of the 𝑠-truncation 𝙹𝑠,𝛾,𝚊 (𝑡), namely

∑
𝑠
𝚢1 (𝑡) = 𝑡 , 𝚢𝑗 (𝑡) = 𝑎𝑗𝑖 𝑡𝑖 , 𝑗 ∈ {2, ..., 𝑚} , (5.47)
𝑖=2

and we obtain
( )𝜇̃2 (𝓁) ( )𝜇̃𝑚 (𝓁)
𝜕𝙿𝚊 (𝚢) || ∑ ∑
𝑠 ∑
𝑠

| = 𝜇𝓁 𝚙𝜇 𝑡𝜇̃1 (𝓁) 𝑎2𝑖 𝑡𝑖 ... 𝑎𝑚𝑢 𝑡𝑢 . (5.48)

𝜕𝚢𝓁 |𝙹𝑠,𝛾,𝚊 (𝑡) 𝜇∈ℕ 𝑚 𝑖=2 𝑢=2
2≤|𝜇|≤𝑟

Step 2. For 𝛼 = 1, we must look for the coefficient of the linear term (in 𝑡) in
expression (5.48). Hence, for fixed 𝓁 = 1, ..., 𝑚, since the sums in (5.48) start at order
two in 𝑡, only the multi-index for which 𝜇̃𝑗 (𝓁) = 0 for all 𝑗 ∈ {2, ..., 𝑚} and 𝜇̃1 (𝓁) = 1
must be retained in the sum in expression (5.48). The first condition implies 𝜇𝑗 (𝓁) =
𝛿𝑗𝓁 for all 𝑗 ∈ {2, ..., 𝑚}, whereas the second yields 𝜇1 (𝓁) = 1 + 𝛿1𝓁 . Therefore,
by definition (5.39), for fixed 𝓁 ∈ {1, ..., 𝑚} only the multi-index 𝜈(𝓁, 1) appears in
expression (5.48) for 𝛼 = 1. Again by (5.39), one has 𝜈1 (1, 1) = 2 and 𝜈𝓁 (𝓁, 1) = 1 for
𝓁 ∈ {2, ..., 𝑚} so that the thesis in the case 𝛼 = 1 follows.
Step 3. For any given 𝛼 ∈ {2, ..., 𝑠}, we are interested in the coefficients of the terms
of order 𝑡𝛼 in (5.48). Hence, we can truncate the internal sums in (5.48) at order 𝛼. For
the same reason, for any 𝓁 ∈ {1, ..., 𝑚}, we can neglect from the leftmost sum in (5.48)
the monomials 𝜇 satisfying |̃ 𝜇 (𝓁)| > 𝛼 (hence |𝜇| > 𝛼 + 1), as their contribution is of
order at least 𝑡𝛼+1 . Thus, for fixed 𝛼 ∈ {2, ..., 𝑠} and 𝓁 ∈ {1, ..., 𝑚}, we have

𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 )

⎛ ( 𝛼 )𝜇̃2 (𝓁) ( 𝛼 )𝜇̃𝑚 (𝓁) ⎞
𝑑𝛼 ⎜ ∑ 𝜇̃1 (𝓁)
∑
𝑖
∑
𝑢 ⎟ (5.49)
= 𝛼⎜ 𝜇𝓁 𝚙𝜇 𝑡 𝑎2𝑖 𝑡 ... 𝑎𝑚𝑢 𝑡 ⎟
𝑑𝑡 ⎜ 𝜇∈ℕ𝑚 𝑖=2 𝑢=2 ⎟
⎝2≤|𝜇|≤𝛼+1 ⎠𝑡=0

due to formula (5.37). Now, for any 𝑗 = 2, … , 𝑚 and 𝓁 ∈ {1, … , 𝑚}, the multinomial

47
expansion yields:
( )𝜇̃𝑗 (𝓁) ( )
∑
𝛼 ∑ 𝜇̃𝑗 (𝓁) 𝑘𝑗2 𝑘𝑗𝛼
𝑎𝑗𝑖 𝑡 𝑖
= 𝑎𝑗2 ... 𝑎𝑗𝛼 𝑡2𝑘𝑗2 +....+𝛼𝑘𝑗𝛼 ,
𝑖=2 𝑘𝑗2 ,...,𝑘𝑗𝛼 ∈ℕ 𝑘𝑗2 ... 𝑘𝑗𝛼
𝑘𝑗2 +...+𝑘𝑗𝛼 =̃ 𝜇𝑗 (𝓁)
(5.50)
( )
𝜇̃𝑗 (𝓁) 𝜇̃𝑗 (𝓁)!
where we have used the notation . ∶=
𝑘𝑗2 ... 𝑘𝑗𝛼 𝑘𝑗2 ! ... 𝑘𝑗𝛼 !
Replacing each truncated Taylor development in (5.49) by its multinomial expan-
sion (5.50), expression (5.49) reads
⎡ ⎛ ⎞ ⎤
⎢ ( ) ⎥
𝑑𝛼 ⎢ ∑ ∏⎜ 𝑚
∑ 𝜇
̃ (𝓁) ⎟ ⎥
𝜇𝓁 𝚙𝜇 𝑡𝜇̃1 (𝓁) ⎜ 𝑗 𝑘 𝑗2 𝑘 𝑗𝛼
𝑡2𝑘𝑗2 +....+𝛼𝑘𝑗𝛼 ⎟
𝑑𝑡𝛼 ⎢⎢ ⎥ =
𝑎𝑗2 ...𝑎𝑗𝛼
𝜇∈ℕ𝑚 𝑗=2
⎜ 𝑘 ,...,𝑘 ∈ℕ 𝑘𝑗2 ... 𝑘𝑗𝛼 ⎟ ⎥
⎢ 2≤|𝜇|≤𝛼+1 ⎜𝑘 +...+𝑘
𝑗2 𝑗𝛼
⎟ ⎥
⎣ ⎝ 𝑗2 𝑗𝛼 =̃
𝜇𝑗 (𝓁) ⎠ ⎦𝑡=0
⎡ ⎤
⎢ ( ) ⎥
𝑑 ⎢
𝛼 ∑ ∑ ∏
𝑚
𝜇̃𝑗 (𝓁) 𝑘𝑗2 𝑘𝑗𝛼 ⎥
𝜇𝓁 𝚙𝜇 𝑡𝜇̃1 (𝓁) 𝑡2𝑘𝑗2 +....+𝛼𝑘𝑗𝛼
𝑑𝑡𝛼 ⎢⎢ ⎥ .
𝑎𝑗2 ...𝑎𝑗𝛼
𝜇∈ℕ𝑚 𝑘∈ℕ(𝑚−1)×(𝛼−1) 𝑗=2 𝑘𝑗2 ...𝑘𝑗𝛼 ⎥
⎢ 2≤|𝜇|≤𝛼+1 𝑘=(𝑘22 ,…,𝑘2𝛼 ,…,𝑘𝑚2 ,…,𝑘𝑚𝛼 ) ⎥
⎣ ∀𝑖∈{2,…,𝑚} ⎦𝑡=0
𝑘𝑖2 +...+𝑘𝑖𝛼 =̃
𝜇𝑖 (𝓁)
(5.51)

Moreover, taking (5.41) into account, the class of multi-indices 𝑚 (̃ 𝜇(𝓁), 𝛼) selects
those terms whose contribution inside the brackets of (5.51) is of order 𝑡𝛼 . Hence,
by the above discussion, by (5.37) and by (5.42), for any fixed 𝛼 ∈ {2, ..., 𝑠}, and
𝓁 ∈ {1, ..., 𝑚}, we can write
( 𝑚 ( ) )
∑ ∑ ∏ 𝜇̃𝑗 (𝓁) 𝑘𝑗2 𝑘𝑗𝛼
𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) =𝛼! 𝜇𝓁 𝚙𝜇 𝑎𝑗2 ...𝑎𝑗𝛼 .
𝜇∈ (𝓁,𝛼) 𝜇(𝓁),𝛼) 𝑗=2 𝑘𝑗2 ... 𝑘𝑗𝛼
𝑘∈ (̃
𝑚 𝑚
𝜇𝓁 ≠0
(5.52)

Now, we split the leftmost sum in (5.52) into the partial sums with respect to the
families of indices defined in (5.39) and (5.40), namely for any fixed 𝛼 ∈ {2, ..., 𝑠} and

48
𝓁 ∈ {1, ..., 𝑚}, we write
𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 )
𝛼! ( 𝑚 ( )
)
𝑚 ∑
∑ 𝛼 ∑ ∏ 𝜈̃𝑗 (𝑖, 𝛽)(𝓁) 𝑘𝑗2 𝑘𝑗𝛼
= 𝜈𝓁 (𝑖, 𝛽) 𝚙𝜈(𝑖,𝛽) 𝑎𝑗2 … 𝑎𝑗𝛼
𝑖=1 𝛽=1 𝑘∈𝑚 (̃
𝜇(𝓁),𝛼) 𝑗=2 𝑘𝑗2 … 𝑘𝑗𝛼
𝜈𝓁 (𝑖,𝛽)∈𝑚 (𝓁,𝛼)
𝜈𝓁 (𝑖,𝛽)≠0
( 𝑚 ( ) )
∑ ∑ ∏ 𝜇̃𝑗 (𝓁) 𝑘𝑗2 𝑘𝑗𝛼
+ 𝜇𝓁 𝚙𝜇 𝑎𝑗2 … 𝑎𝑗𝛼 .
𝜇∈𝑚 (𝓁,𝛼) 𝑘∈𝑚 (̃
𝜇(𝓁),𝛼) 𝑗=2 𝑘𝑗2 … 𝑘𝑗𝛼
𝜇∈𝑚 (𝛼)
𝜇𝓁 ≠0
(5.53)
Step 4. We first study the case in which 𝓁 ≠ 1. For fixed 𝓁 ∈ {2, ..., 𝑚}, for any
𝑖 ∈ {1, ..., 𝑚}, 𝑖 ≠ 𝓁, and for any 𝛽 ∈ {1, ..., 𝛼}, the monomials corresponding to the
indices 𝜈(𝑖, 𝛽) do not contribute to the leftmost sum at the r.h.s. of (5.53). Infact, by
(5.39), the 𝓁-th element 𝜈𝓁 (𝑖, 𝛽) of multi-index 𝜈(𝑖, 𝛽) is equal to zero for 𝓁 ∈ {2, ..., 𝑚}
and 𝑖 ∈ {1, ..., 𝑚}, 𝑖 ≠ 𝓁.
Moreover, still for fixed 𝓁 ∈ {2, ..., 𝑚}, the indices 𝜈(𝓁, 𝛽), with 𝛽 ∈ {1, ..., 𝛼},
satisfy 𝜈̃1 (𝓁, 𝛽)(𝓁) = 𝛽 and 𝜈̃𝑗 (𝓁, 𝛽)(𝓁) = 0 for all 𝑗 ∈ {2, ..., 𝑚}, so that by (5.41) we
have {
∅, if 𝛽 = 1, ..., 𝛼 − 1
𝑚 (̃
𝜈 (𝓁, 𝛽)(𝓁), 𝛼) = .
{0} , if 𝛽 = 𝛼
Consequently, the only monomial that contributes to the leftmost sum at the r.h.s. of
(5.53) is the one associated to the multi-index 𝜈(𝓁, 𝛼), and one has 𝑘𝑗2 = 0, … , 𝑘𝑗𝛼 =
0 when 𝜇 = 𝜈(𝓁, 𝛼). Moreover, by hypothesis we have 𝜈𝓁 (𝓁, 𝛼) = 1 for any 𝓁 ∈
{2, ..., 𝑚}. Due to these arguments, for any fixed 𝓁 ∈ {2, ..., 𝑚}, we can rewrite (5.53)
in the form
𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 )
𝛼! ( )
( )
∑ ∑ ∏
𝑚
𝜇̃𝑗 (𝓁) 𝑘𝑗2 𝑘𝑗𝛼
= 𝚙𝜈(𝓁,𝛼) + 𝜇𝓁 𝚙𝜇 𝑎𝑗2 ...𝑎𝑗𝛼 . (5.54)
𝜇∈𝑚 (𝓁,𝛼) 𝑘∈𝑚 (̃
𝜇(𝓁),𝛼) 𝑗=2 𝑘𝑗2 ... 𝑘𝑗𝛼
𝜇∈𝑚 (𝛼)
𝜇𝓁 ≠0

This proves the Lemma for 𝓁 = 2, ..., 𝑚, 𝛼 = 2, ..., 𝑠.

Step 5. We now consider the case 𝓁 = 1. For all 𝑗 ∈ {2, ..., 𝑚}, the sub-family of
indices 𝜈(1, 𝛽), with 𝛽 ∈ {1, ..., 𝛼}, satisfies 𝜈̃1 (1, 𝛽)(1) = 𝛽 and 𝜈̃𝑗 (1, 𝛽)(1) = 0. Hence,
thanks to (5.41), we find
{
∅, if 𝛽 = 1, ..., 𝛼 − 1
𝑚 (̃
𝜈 (1, 𝛽)(1), 𝛼) = . (5.55)
{0} , if 𝛽 = 𝛼

49
Moreover, we have 𝜈1 (1, 𝛼) = 𝛼 + 1 by construction.
On the other hand, for 𝓁 = 1, 𝛽 ∈ {1, ..., 𝛼} and 𝑖, 𝑗 ∈ {2, ..., 𝑚}, the multi-indices
𝜈(𝑖, 𝛽) satisfy 𝜈̃1 (𝑖, 𝛽)(1) = 𝛽 − 1 and 𝜈̃𝑗 (𝑖, 𝛽)(1) = 𝛿𝑗𝑖 . Hence, by (5.41) one can write

∑
𝛼
𝜈̃𝑗 (𝑖, 𝛽)(1) = 𝛿𝑗𝑖 = 𝑘𝑗𝑢 ⟺ 𝑘𝑗𝑢 = 𝛿𝑗𝑖 𝛿𝑢𝑣 for some 𝑣 ∈ {2, ..., 𝛼} (5.56)
𝑢=2

and
𝑚 ∑
∑ 𝛼 𝑚 ∑
∑ 𝛼
𝜈̃1 (𝑖, 𝛽)(1) + 𝑢 𝑘𝑗𝑢 = 𝛽 − 1 + 𝑢 𝑘𝑗𝑢 = 𝛼 . (5.57)
𝑗=2 𝑢=2 𝑗=2 𝑢=2

By (5.56), we see that condition (5.57) can be satisfied by some vector of multi-integers
(𝑘22 , … , 𝑘2𝛼 , … , 𝑘𝑚2 , … , 𝑘𝑚𝛼 ) if 𝛽 ∈ {1, ..., 𝛼 − 1}, but cannot be fulfilled for 𝛽 = 𝛼.
Injecting (5.56) into (5.57) one has

∑
𝑚 ∑
𝛼
𝑢 𝛿𝑗𝑖 𝛿𝑢𝑣 = 𝛼 − (𝛽 − 1) for all 𝛽 ∈ {1, ..., 𝛼 − 1} , 𝑖 ∈ {2, ..., 𝑚} (5.58)
𝑗=2 𝑢=2

which implies

𝑘𝑗𝑢 = 𝛿𝑗𝑖 𝛿𝑢𝑣 𝛿𝑣,𝛼−(𝛽−1) for all 𝛽 ∈ {1, ..., 𝛼 − 1} , 𝑖 ∈ {2, ..., 𝑚} . (5.59)

Hence, for all 𝛽 ∈ {1, ..., 𝛼 − 1}, and 𝑖 ∈ {2, ..., 𝑚}, we can finally write

𝑚 (̃
𝜈 (𝑖, 𝛽)(1), 𝛼) = {(𝑘𝑗1 , ..., 𝑘𝑗𝛼 ), 𝑗 ∈ {2, ..., 𝑚}, 𝑘𝑗𝑢 = 𝛿𝑗𝑖 𝛿𝑢,𝛼−(𝛽−1) } (5.60)

and
𝑚 (̃
𝜈 (𝑖, 𝛼)(1), 𝛼) = ∅ . (5.61)
Moreover, for 𝑖 = 2, ..., 𝑚, by (5.39) we have 𝜈1 (𝑖, 𝛽) = 𝛽.
By taking (5.55), (5.60), (5.61) into account, expression (5.53) with 𝓁 = 1 yields

𝚀1𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ∑ ∑

𝑚 𝛼−1
= (𝛼 + 1)𝚙𝜈(1,𝛼) + 𝛽 𝚙𝜈(𝑖,𝛽) 𝑎𝑖(𝛼−(𝛽−1))
𝛼! 𝑖=2 𝛽=1
( 𝑚 ( ) )
∑ ∑ ∏ 𝜇̃𝑗 (1) 𝑘𝑗2 𝑘𝑗𝛼 (5.62)
+ 𝜇1 𝚙𝜇 𝑎𝑗2 ...𝑎𝑗𝛼 .
𝜇∈𝑚 (1,𝛼) 𝑘∈𝑚 (̃
𝜇(1),𝛼) 𝑗=2 𝑘𝑗2 ... 𝑘𝑗𝛼
𝜇∈𝑚 (𝛼)
𝜇1 ≠0

This concludes the proof for the case in which 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠).
The proof of the case in which 𝑠,𝛾 ∈ 𝜗𝑖𝑚 (𝑠), with 𝑖 = 2, … , 𝑚, is the same: one
just has to take into account that the rôle of the special index is played by 𝑖 instead of 1.
Hence, the ideals of the sets 𝑍 𝑖 (𝑟, 𝑠, 𝑚) can be explicitly computed and, by expression
(5.16), the proof is concluded.

50
 We are now able to prove that 𝑠-vanishing polynomials are rare in (𝑟, 𝑚).

Proof. (Corollary 5.1) We want to show that, for a given pair (𝑃 , 𝑠,𝛾 ) ∈ (𝑟, 𝑚) ×
𝜗1𝑚 (𝑠), the 𝑚𝑠 + 𝑚 equations in (5.43) and (5.44) are all linearly independent.
For all 𝓁 = 1, ..., 𝑚 and 𝛼 = 0, ..., 𝑠, we collect in table (5.64) the derivatives of the
functions 𝚀𝓁𝛼 ∕𝛼! defined in (5.36)-(5.37) - and whose action is made explicit in (5.43)
- (5.44) - with respect to the coefficients of 𝙿𝚊 , 𝚊, and to the Taylor coefficients of 𝙹𝑠,𝛾,𝚊 .
We have indicated

1. with the symbol 𝔻, the 𝑠 × 𝑠 diagonal matrix whose entries are the numbers 𝛼 + 1,
for 𝛼 = 1, ..., 𝑠;

2. with the symbol 𝕀𝑠 , the 𝑠 × 𝑠 identity matrix;

3. with the symbol 𝔹𝑖 , 𝑖 ∈ {2, ..., 𝑚}, an 𝑠 × 𝑠 matrix whose entry at position 𝛼, 𝛽,
with 𝛼 ∈ {1, ..., 𝑠} and 𝛽 ∈ {1, ..., 𝑠}, reads

⎧0 , if 𝛼 = 1 ,
⎪
(𝔹𝑖 )𝛼,𝛽 ∶= ⎨𝛽𝑎𝑖(𝛼−(𝛽−1)) , if 2 ≤ 𝛼 ≤ 𝑠 , 1 ≤ 𝛽 ≤ 𝛼 − 1 , (5.63)
⎪
⎩0 , if 2 ≤ 𝛼 ≤ 𝑠 , 𝛼 ≤ 𝛽 ≤ 𝑠 .

𝓁 𝛼 𝜕𝑎𝑗𝑖 𝜕𝚙𝜇 𝜕𝚙𝜈(𝑖,0) 𝜕𝚙𝜈(1,𝛽) 𝜕𝚙𝜈(2,𝛽) … 𝜕𝚙𝜈(𝑚,𝛽)

𝜇 ∈ (𝑠) 𝑖 = 1, … , 𝑚 𝛽 = 1, … , 𝑠 𝛽 = 1, … , 𝑠 … 𝛽 = 1, … , 𝑠

1, … , 𝑚 0 0 0 𝕀𝑚 0 0 0 0

1 1, … , 𝑠 … … 0 𝔻 𝔹2 … 𝔹𝑚

2 1, … , 𝑠 … … 0 0 𝕀𝑠 0 0

… … … … 0 0 0 … 0

𝑚 1, … , 𝑠 … … 0 0 0 0 𝕀𝑠
(5.64)

Table 5.64: Jacobian of 𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 with 𝓁 ∈ {1, ..., 𝑚} and 𝛼 = {0, .., 𝑠}.
The first and the second column contain, respectively, all the possible values for the
parameters 𝓁 and 𝛼. The third column corresponds to the derivatives of 𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 )
with respect to the variables of the vector 𝚊 ∈ 𝜗1𝑚 (1), and to the Taylor coefficients
of the 𝑠-jet 𝙹𝑠,𝛾,𝚊 . The remaining columns contain the derivatives with respect to the
coefficients 𝚙𝜇 of 𝙿𝚊 associated with the families of multi-indices (5.39) and (5.40), in
suitable order.
It is plain to check that matrix (5.64) contains a submatrix of maximal rank 𝑚𝑠 + 𝑚
- corresponding to the derivatives w.r.t. those coefficients associated to the family of

51
multi-indices (5.39) - independently of 𝚊 and of the 𝑠-truncation 𝙹𝑠,𝛾,𝚊 on which the
𝑠-vanishing condition is realized. Hence, since the transformation Υ in (5.29) is invert-
ible, by Theorem 5.1 the set 𝑍 1 (𝑟, 𝑠, 𝑚) is determined by 𝑚𝑠 + 𝑚 linearly independent
algebraic equations and has codimension 𝑚𝑠 + 𝑚 in (𝑟, 𝑚) × 𝜗1𝑚 (𝑠).
As it was the case in the proof of Theorem 5.1, the same strategy of proof applies
for 𝑍 𝑗 (𝑟, 𝑠, 𝑚), 𝑗 = 2, … , 𝑚, one just has to switch the rôle of the indices 1 and 𝑗.
Since 𝜎 1 (𝑟, 𝑠, 𝑚) ∶= Π(𝑟,𝑚) 𝑍 1 (𝑟, 𝑠, 𝑚) (see (5.14)) and 𝑍 1 (𝑟, 𝑠, 𝑚) is algebraic, by
the Theorem of Tarski and Seidenberg (see Th. A.1) 𝜎 1 (𝑟, 𝑠, 𝑚) is a semi-algebraic set
of (𝑟, 𝑚). Moreover, as Jacobian (5.64) has rank 𝑚𝑠 + 𝑚 w.r.t. the 𝑚𝑠 + 𝑚 polynomial
coefficients associated to the multi-indices of the family (5.39), for 𝛼 ∈ {2, ..., 𝑠} and
𝑖 ∈ {1, ..., 𝑚}, by Theorem 5.1 and by the implicit function theorem, the conditions
𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 imply

𝚙𝜈(𝑖,0) = 𝚙𝜈(𝑖,1) = 0
(5.65)
𝚙𝜈(𝑖,𝛼) = 𝑔𝑖𝛼 (𝚙𝜇 , 𝚊, 𝑎22 , ..., 𝑎2𝑠 , ..., 𝑎𝑚2 , ..., 𝑎𝑚𝑠 ) , 𝜇 ∈ (𝛼)

for some implicit functions 𝑔𝑖𝛼 . That is, one can express the polynomial coefficients
𝚙𝜈(𝑖,0) , 𝚙𝜈(𝑖,1) , 𝚙𝜈(𝑖,𝛼) as implicit functions of the remaining coefficients - associated to
the multi-indices in the family (𝛼) defined in (5.39) - and of the (𝑚 − 1)𝑠 parameters
of 𝚊 and 𝙹𝑠,𝛾,𝚊 . Moreover, since the functions 𝚀𝓁𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) are polynomial for all
𝓁 ∈ {1, ..., 𝑚} and 𝛼 ∈ {0, .., 𝑠}, the implicit functions 𝑔𝑖𝛼 are all analytic. Therefore,
one has an analytic parametrization of 𝜎 1 (𝑟, 𝑠, 𝑚) given by the 𝑚(𝑠 + 1) independent
equations (5.65), for 𝑖 ∈ {1, ..., 𝑚}, 𝛼 ∈ {0, ..., 𝑠}. This, in turn, yields that

dim 𝜎 1 (𝑟, 𝑠, 𝑚) = dim 𝑊 1 (𝑟, 𝑚) − 𝑚(𝑠 + 1) = dim (𝑟, 𝑚) + dim 𝜗1𝑚 (𝑠) − 𝑚(𝑠 + 1)
= dim (𝑟, 𝑚) + (𝑚 − 1)𝑠 − 𝑚(𝑠 + 1) ,
(5.66)

which implies that the codimension of 𝜎 1 (𝑟, 𝑠, 𝑚) in (𝑟, 𝑚) is

codim 𝜎 1 (𝑟, 𝑠, 𝑚) =𝑚(𝑠 + 1) − (𝑚 − 1)𝑠 = 𝑚 + 𝑠 . (5.67)

Once again, it is plain to check that the same result holds true also in the case in which
the parametrizing coordinate of the curve gamma is the 𝑗-th, with 𝑗 = 2, … , 𝑚. Hence,
one finds codim 𝜎 𝑗 (𝑟, 𝑠, 𝑚) = 𝑚 + 𝑠 for 𝑗 = 2, … , 𝑚, which, together with expression
(5.13), proves the statement.

5.3 Geometric properties

For fixed integers 𝑟 ≥ 2, 𝑚 ≥ 2, 1 ≤ 𝑠 ≤ 𝑟 − 1, and for any 𝑖 ∈ {1, … , 𝑚}, we indicate
respectively by Σ(𝑟, 𝑠, 𝑚) ∶= 𝜎(𝑟, 𝑠, 𝑚) and Σ𝑖 (𝑟, 𝑠, 𝑚) ∶= 𝜎 𝑖 (𝑟, 𝑠, 𝑚) the closures in

52
(𝑟, 𝑚) of the sets 𝜎(𝑟, 𝑠, 𝑚) and 𝜎 𝑖 (𝑟, 𝑠, 𝑚) introduced in the previous section. Taking
(5.13) into account, one has
⋃
𝑚
Σ(𝑟, 𝑠, 𝑚) = Σ𝑖 (𝑟, 𝑠, 𝑚) . (5.68)
𝑖=1

For 𝑚 ≥ 2, Corollary 5.1 and Proposition A.3 ensure that for any 𝑖 ∈ {1, … , 𝑚}

codim Σ𝑖 (𝑟, 𝑠, 𝑚) = codim 𝜎 𝑖 (𝑟, 𝑠, 𝑚) = 𝑠 + 𝑚 > 0

in (𝑟, 𝑚), so that (𝑟, 𝑚)∖Σ𝑖 (𝑟, 𝑠, 𝑚) is an open set of full Lebesgue measure. There-
fore, by (5.68), the same holds true also for (𝑟, 𝑚)∖Σ(𝑟, 𝑠, 𝑚).
As we did previously, when 𝑚 ≥ 2 we only consider the case in which the index of
the parametrizing coordinate is 𝑖 = 1, as the other cases are immediate generalizations.
In case 𝑚 = 1, instead, Lemma 5.1 ensures that

𝜎(𝑟, 𝑠, 1) = 𝜎(𝑟, 𝑠, 1) =∶ Σ(𝑟, 𝑠, 1) . (5.69)

and
codim 𝜎(𝑟, 𝑠, 1) = codim Σ(𝑟, 𝑠, 1) = 𝑠 + 1 . (5.70)
Still for 𝑚 = 1, in order to make use of uniform notations w.r.t. the case 𝑚 = 2 and
to simplify the exposition in the sequel, we write

𝜎 1 (𝑟, 𝑠, 𝑚) ≡ Σ1 (𝑟, 𝑠, 1) ∶= Σ(𝑟, 𝑠, 1) (5.71)

and we extend the notations of subsection 5.2.3 by setting

𝚊 ∶= id , 𝙿𝚊 (𝚢) ∶= 𝑃 (𝚢) , for 𝑚 = 1 . (5.72)

The rest of this section will be devoted to the proof of the following
Lemma 5.3. Let 𝑚 be a positive integer. For any open set 𝙳 ⊂ (𝑟, 𝑚)∖Σ1 (𝑟, 𝑠, 𝑚)
verifying 𝙳 ∩ Σ1 (𝑟, 𝑠, 𝑚) = ∅, there exist positive constants 𝐶1 (𝙳), 𝐶2 (𝑠, 𝑚) such that for
any polynomial 𝑃 (𝑥) ∈ 𝙳 and for any arc 𝛾 ∈ Θ1𝑚 one has the following lower estimates
| 𝛼 ( ) |
|𝑑 𝜕𝙿 (𝚢) | |
max || 𝛼 𝚊 |
|
| > 𝐶1 (𝙳)
|
𝓁=1,...,𝑚 | 𝑑𝑡 𝜕𝚢𝓁 |𝚊 ◦𝛾(𝑡) |
𝛼=1,...,𝑠 | 𝑡=0 |

in case 𝑠 = 1 or 𝑚 = 1,
| 𝛼 ( ) | (5.73)
|𝑑 𝜕𝙿 (𝚢) | | 𝐶1 (𝙳)
max || 𝚊 | |>
𝓁=1,...,𝑚 | 𝑑𝑡𝛼 𝜕𝚢𝓁 ||𝚊 ◦𝛾(𝑡) | 1 + 𝐶 (𝑠, 𝑚) × max
| 2 𝓁=2,...,𝑚 |𝑎𝓁𝛼 |
𝛼=1,...,𝑠 | 𝑡=0 | 𝛼=2,...,𝑠
in case 2 ≤ 𝑠 ≤ 𝑟 − 1 and 𝑚 ≥ 2 ,
where - for 𝑚 ≥ 2 - 𝙿𝚊 is the polynomial 𝑃 written in the adapted system of coordinates
for 𝛾 introduced in paragraph 5.2.3 and 𝚊 is the associated transformation defined in
(5.33), whereas for 𝑚 = 1 the symbols 𝙿𝚊 and 𝚊 are to be intended as in (5.72).

53
Remark 5.11. As we shall see in the next section, for any 𝜆 > 0, when 𝛾 is the minimal
arc of Theorem 4.1, one can give a positive upper bound - that only depends on 𝑟, 𝑠, 𝑚, 𝜆 -
to the quantity max𝓁=1,...,𝑚 |𝑎𝓁𝛼 | at the denominator of (5.73). This is due to the fact that
𝛼=2,...,𝑠
all minimal arcs satisfy a uniform Bernstein-like inequality on their Taylor coefficients
(see formula (4.1) in Theorem 4.1).
Before proving Lemma 5.3, we need an intermediate result and a few additional
notations.
In case 𝑚 ≥ 2, for any given arc 𝛾 ∈ Θ1𝑚 with associated change of coordinates 𝚊
(see paragraph 5.2.3, in particular formulas (5.19)-(5.33)), we define the direct sum

(𝑟, 𝑚) = 𝚊♯ (𝑟, 𝑚) ⊕ 𝚊♭ (𝑟, 𝑚)

in the following way: for any polynomial 𝑅(𝑥) ∈ (𝑟, 𝑚), we consider its expression
𝚁𝚊 (𝚢) ∶= 𝑅◦−1 𝚊 (𝚢) ∈ (𝑟, 𝑚) in the adapted coordinates (5.19) for 𝛾; 𝚁𝚊 (𝚢) can be
decomposed uniquely into the partial sums with respect to the families of multi-indices
defined in (5.39) and (5.40), namely:
∑ ∑
𝑚 ∑
𝑠 ∑
𝚁𝚊 (𝚢) = 𝚛𝜇 𝚢𝜇 = 𝚛𝜈(𝑖,𝛽) 𝚢𝜈(𝑖,𝛽) + 𝚛𝜇 𝚢𝜇 =∶ 𝚁♯𝚊 (𝚢) + 𝚁♭𝚊 (𝚢) , (5.74)
𝜇∈ℕ𝑚 𝑖=1 𝛽=0 𝜇∈ℕ𝑚
1≤|𝜇|≤𝑟 𝜇∈(𝑠)

and we set 𝑅♯ (𝑥) ∶= 𝚁♯𝚊 ◦𝚊 (𝑥) ∈ 𝚊♯ (𝑟, 𝑚) and 𝑅♭ (𝑥) ∶= 𝚁♭𝚊 ◦𝚊 (𝑥) ∈ 𝚊♭ (𝑟, 𝑚).
Clearly, the decomposition 𝑅(𝑥) = 𝑅♯ (𝑥) + 𝑅♭ (𝑥) is unique, as the function associ-
ating 𝑅 ⟼ 𝚁𝚊 ∶= 𝑅◦−1 𝚊 , with 𝑅 ∈ (𝑟, 𝑚), is invertible.
Finally, we set
𝚀◦Υ1 ∶ (𝑟, 𝑚) × 𝜗1𝑚 (𝑠) ⟶ ℝ𝑚(𝑠+1)
(5.75)
(𝑅, 𝑠,𝛾 ) ⟼ 𝚀𝓁𝛼 ◦Υ1 (𝑅, 𝑠,𝛾 ) ≡ 𝚀𝓁𝛼 (𝚁𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ,
where 𝓁 = 1, ..., 𝑚, 𝛼 = 0, ..., 𝑠, 𝑠,𝛾 is the 𝑠-truncation of the curve 𝛾, and the explicit
form of 𝚀𝓁𝛼 ◦Υ1 is given in Theorem 5.1. We also indicate by 𝒩 (⋅) the zero sets of the
transformations which will appear henceforth.
With this setting, one has the following intermediate result:
Lemma 5.4. In case 𝑚 ≥ 2, for any given 𝛾 ∈ Θ1𝑚 with associated 𝑠-truncation 𝑠,𝛾 ∈
𝜗1𝑚 (𝑠), and for any given polynomial 𝑅(𝑥) ∈ (𝑟, 𝑚)∖Σ1 (𝑟, 𝑠, 𝑚), there exists a unique
polynomial 𝑆(𝑥) ∈ 𝜎 1 (𝑟, 𝑠, 𝑚) such that

(𝑆, 𝑠,𝛾 ) ∈ 𝒩 (𝚀◦Υ1 ) , 𝑆 ♭ = 𝑅♭ .

In particular, 𝑆 satisfies the 𝑠-vanishing condition on the truncation 𝑠,𝛾 .

Proof. Given 𝛾 ∈ Θ1𝑚 with its associated 𝑠-truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) and a polynomial
𝑅(𝑥) ∈ (𝑟, 𝑚)∖Σ1 (𝑟, 𝑠, 𝑚), we denote by

𝙰♯ ♭ ∶ Υ1 (𝚊♯ ⊕ {𝑅♭ } × {𝑠,𝛾 }) → ℝ𝑚(𝑠+1) (5.76)

𝑅 ,𝑠,𝛾

54
the restriction of 𝚀 to the set Υ1 (𝚊♯ ⊕ {𝑅♭ } × {𝑠,𝛾 }) .
As it was shown in the proof of Corollary 5.1 (see Table (5.64)), for all 𝛼 ∈ {0, ..., 𝑠}
and 𝓁 ∈ {1, ..., 𝑚}, the Jacobian matrix of 𝙰♯ ♭ reads
𝑅 ,𝑠,𝛾

⎛ 𝕀𝑚 0 0 0 ... 0⎞
⎜ ⎟
⎜ ⎟
⎜0 𝔻 𝔹2 𝔹3 ... 𝔹𝑚 ⎟
⎜ ⎟
⎜ ⎟
⎜0 0 𝕀𝑠 0 ... 0⎟
 ∶= ⎜ ⎟ , (5.77)
⎜ ⎟
⎜0 0 0 𝕀𝑠 ... 0⎟
⎜ ⎟
⎜ ⎟
⎜0 0 0 ... 0 0⎟
⎜ ⎟
⎜ ⎟
⎝0 0 0 0 0 𝕀𝑠 ⎠

where the blocks 𝔻 and 𝔹𝑖 , 𝑖 = 2, ..., 𝑚, were defined in (5.63). Also, by Theorem 5.1,
when 𝑅♭ and 𝑠,𝛾 are fixed, the restriction of the function 𝚀 to the set Υ1 (𝚊♯ ⊕ {𝑅♭ } ×
{𝑠,𝛾 }) is affine. Hence, 𝙰♯ ♭ is represented by matrix (5.77) acting on Υ1 (𝚊♯ ⊕
𝑅 ,𝑠,𝛾
{𝑅♭ } × {𝑠,𝛾 }) plus a constant term depending only on 𝚁♭𝚊 ≡ 𝑅♭ ◦−1
𝚊 and 𝙹𝑠,𝛾,𝚊 ≡
𝚊 ◦𝑠,𝛾 ; thus, it is globally invertible in Υ1 (𝚊♯ ⊕ {𝑅♭ } × {𝑠,𝛾 }). We set
( )−1 ( ( ♯ ))
𝑆 ♯ ∶= Υ1 𝒩 𝚀𝑅, ∈ 𝚊♯ ,
𝑠,𝛾

which is equivalent to saying that

(𝑆 ♯ + 𝑅♭ , 𝑠,𝛾 ) ∈ 𝒩 (𝚀◦Υ1 ) ,

i.e., by (5.75) and (5.37), 𝑆 ♯ + 𝑅♭ satisfies the 𝑠-vanishing condition on 𝑠,𝛾 .

We are now able to prove Lemma 5.3.

Proof. (Lemma 5.3)

We consider a polynomial 𝑃 ∈ 𝙳 ⊂ (𝑟, 𝑚)∖Σ1 (𝑟, 𝑠, 𝑚), with 𝙳 ∩ Σ1 (𝑟, 𝑠, 𝑚) = ∅.
Case 𝑚 = 1. In case 𝑚 = 1, by Lemma 5.1 there exists a constant 𝐶1 (𝙳) such that
the truncation at order 𝑠 + 1 of 𝑃 - indicated by 𝑃𝑠+1 - satisfies

||𝑃𝑠+1 ||∞ > 𝐶1 (𝙳) .

The thesis follows easily by the expression above and by Definition 5.3.
Case 𝑚 ≥ 2. For any fixed arc 𝛾 ∈ Θ1𝑚 , we shift to its associated adapted coor-
dinates, and we consider the 𝑠-truncation 𝙹𝑠,𝛾,𝚊 ∈ 𝜗1𝑚 (𝑠), together with the pull-back
𝙿𝚊 ∶= 𝑃 ◦−1𝚊 of the polynomial 𝑃 w.r.t. the change of coordinates 𝚊 introduced in

55
paragraph 5.2.3. Due to the hypothesis, to Lemma 5.2 (see especially formulas (5.33)
and (5.35)), and to the fact that the linear terms of the curve 𝛾 are uniformly bounded
(see the Bernstein’s estimate (4.1)), there exists a constant 𝐶1 (𝙳) > 0 such that

||𝙿𝚊 − Σ1 (𝑟, 𝑠, 𝑚)||∞ ∶= inf ||𝙿𝚊 − ̂

𝙿𝚊 ||∞ > 𝐶1 (𝙳) > 0 . (5.78)
̂
𝙿𝚊 ∈Σ1 (𝑟,𝑠,𝑚)

It suffices to prove the statement for the quantity

| 𝛼 ( ) |
|𝑑 𝜕𝙿𝚊 || |
max | 𝛼 | |
𝓁=1,...,𝑚 | 𝑑𝑡
|
𝜕𝚢𝓁 |𝙹𝑠,𝛾,𝚊 (𝑡) |
|
𝛼=1,...,𝑠 | 𝑡=0 |

instead of
| 𝛼 ( ) |
|𝑑 𝜕𝙿 | |
max || 𝛼 𝚊 |
|
|
| (5.79)
𝓁=1,...,𝑚 | 𝑑𝑡 𝜕𝚢𝓁 |𝚊 ◦𝛾(𝑡) |
𝛼=1,...,𝑠 | 𝑡=0 |

because - as we had already pointed out in paragraph 5.2.4 - the terms of order higher
than 𝑠 in the Taylor developement of 𝚊 ◦𝛾 yield a null contribution to (5.79). With
the decomposition in (5.74), by Lemmata 5.2 and 5.4 there exists a unique polynomial
𝚂𝚊 = 𝚂♯𝚊 + 𝚂♭𝚊 fulfilling the 𝑠-vanishing condition on 𝙹𝑠,𝛾,𝚊 and satisfying 𝚂♭𝚊 = 𝙿♭𝚊 .
Hence, (5.78) yields

||𝙿𝚊 − 𝚂𝚊 ||∞ = ||𝙿♯𝚊 − 𝚂♯𝚊 ||∞ > 𝐶1 (𝙳) > 0 . (5.80)

By the proof of Lemma 5.4, we also know that - since 𝚂♭𝚊 = 𝙿♭𝚊 and 𝙹𝑠,𝛾,𝚊 are kept fixed
- the function 𝙰♯ ♭ in (5.76) is affine and invertible in Υ1 (𝚊♯ ⊕ {𝑃 ♭ } × {𝑠,𝛾 }). In
𝑃 ,𝑠,𝛾
particular, it is represented by matrix  in (5.77) plus a constant term depending only
on 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 . Taking into account the fact that 𝙰♯ ♭ is the restriction of 𝚀 to the set
𝑃 ,𝑠,𝛾
Υ1 (𝚊♯ ⊕ {𝑃 ♭ } × {𝑠,𝛾 }), one can write

||𝙿♯𝚊 − 𝚂♯𝚊 ||∞ ≤ ||−1 ||∞ ||𝚀(𝙿♯𝚊 + 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 ) − 𝚀(𝚂♯𝚊 + 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 )||∞ , (5.81)

where ||−1 ||∞ indicates the matrix norm of the inverse. Expressions (5.80) and (5.81)
together yield

𝐶1 (𝙳)
||𝚀(𝙿♯𝚊 + 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 ) − 𝚀(𝚂♯𝚊 + 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 )||∞ > . (5.82)
||−1 ||∞

Moreover, by construction one has 𝚂♯𝚊 ∈ 𝒩 (𝙰♯ ♭ ), that is (𝚂♯𝚊 + 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 ) ∈ 𝒩 (𝚀),
𝑃 ,𝑠,𝛾
so that (5.82) implies

𝐶1 (𝙳)
||𝚀(𝙿♯𝚊 + 𝙿♭𝚊 , 𝙹𝑠,𝛾,𝚊 )||∞ > . (5.83)
||−1 ||∞

56
From the explicit form (5.77) of the (𝑚𝑠 + 𝑚) × (𝑚𝑠 + 𝑚) matrix , one can easily infer
the form of −1 , namely

⎛𝕀𝑚 0 0 0 ... 0 ⎞
⎜ ⎟
⎜ ⎟
⎜0 𝔻−1 −𝔻−1 𝔹2 −𝔻−1 𝔹3 ... −𝔻−1 𝔹𝑚 ⎟
⎜ ⎟
⎜ ⎟
⎜0 0 𝕀𝑠 0 ... 0 ⎟
−1 ∶= ⎜ ⎟ . (5.84)
⎜ ⎟
⎜0 0 0 𝕀𝑠 ... 0 ⎟
⎜ ⎟
⎜ ⎟
⎜0 0 0 0 ... 0 ⎟
⎜ ⎟
⎜ ⎟
⎝0 0 0 0 0 𝕀𝑠 ⎠
∑
The induced matrix norm is, by construction, ||−1 ||∞ ∶= sup𝑖∈{1,...,𝑛} 𝑚𝑠+𝑚 𝑗=1 |𝑖𝑗 | .
−1

By the definition of 𝔻 given above table (5.64), and by the definition of the blocks 𝔹𝑖 ,
with 𝑖 ∈ {2, ..., 𝑚}, in (5.63) one has ||𝔻−1 ||∞ = 1 and for 𝑚 ∈ {2, ..., 𝑚} one can write

⎧
⎪0 if 𝑠 = 1
sup ||𝔹𝓁 || ≤ ⎨
𝑠 × (𝑠 − 1) max𝓁=2,...,𝑚 |𝑎𝓁𝛼 | if 2 ≤ 𝑠 ≤ 𝑟 − 1 .
𝓁∈{2,...,𝑚} ⎪
⎩ 𝛼=2,...,𝑠

Hence, one finally has

⎧
−1 ⎪1 if 𝑠 = 1
|| ||∞ ≤⎨ (5.85)
1 + (𝑚 − 1) 𝑠 (𝑠 − 1) × max𝓁=2,...,𝑚 |𝑎𝓁𝛼 | if 2 ≤ 𝑠 ≤ 𝑟 − 1 .
⎪
⎩ 𝛼=2,...,𝑠

Estimate (5.85), together with formulas (5.83), (5.75), and (5.37) implies the thesis,
with 𝐶2 (𝑠, 𝑚) = (𝑚 − 1) 𝑠 (𝑠 − 1)

6 Proof of Theorem A
In order to prove Theorem A, we need to combine the results of the previous sections
with several intermediate Lemmata.

6.1 Codimension estimates

For any pair of integers 𝑛 ≥ 2 and 1 ≤ 𝑚 ≤ 𝑛 − 1, we indicate by 𝑂(𝑛, 𝑚) the space of
𝑛 × 𝑚 real matrices whose columns are orthonormal vectors of ℝ𝑛 . Clearly, a matrix
𝐴 ∈ 𝑂(𝑛, 𝑚) induces a map from ℝ𝑚 to ℝ𝑛 associating ℝ𝑚 ∋ 𝑥 ⟼ 𝐼 = 𝐴𝑥 ∈ ℝ𝑛 .

57
From a geometric point of view, for any integer 𝑟 ≥ 2, the restriction of any polynomial
𝑄(𝐼) ∈ (𝑟, 𝑛) to any 𝑚-dimensional subspace Γ𝑚 ⊂ ℝ𝑛 endowed with the Euclidean
metric can be identified through 𝑄|Γ𝑚 (𝑥) ∶= 𝑄(𝐴𝑥) =∶ 𝑃 (𝑥) ∈ (𝑟, 𝑚), where the
columns of 𝐴 ∈ 𝑂(𝑛, 𝑚) span Γ𝑚 .
We also indicate by 𝑂(𝑚) the 𝑚×𝑚 orthogonal group and by 𝙶(𝑚, 𝑛) the 𝑚-dimensional
Grassmannian, which is locally isomorphic to 𝑂(𝑛, 𝑚)∕𝑂(𝑚).
With this setting, for any integer 1 ≤ 𝑠 ≤ 𝑟 − 1, we define

 =  (𝑟, 𝑠, 𝑚, 𝑛) ∶= {(𝑄, 𝐴, 𝑃 ) ∈  ⋆ (𝑟, 𝑛) × 𝑂(𝑛, 𝑚) × (𝑟, 𝑚)|

(6.1)
𝑃 (𝑥) = 𝑄(𝐴𝑥) , 𝑃 (𝑥) ∈ Σ(𝑟, 𝑠, 𝑚)} ,

and we indicate by

 = (𝑟, 𝑠, 𝑚, 𝑛) ∶= Π ⋆ (𝑟,𝑛)  (𝑟, 𝑠, 𝑚, 𝑛) (6.2)

its projection onto the first component, i.e. the set of those polynomials 𝑄 ∈  ⋆ (𝑟, 𝑛)
for which the origin is non-critical, and such that, for some 𝑚-dimensional subspace Γ𝑚
orthogonal to ∇𝑄(0), the restriction 𝑄|Γ𝑚 ∈ (𝑟, 𝑚) belongs to the closure Σ(𝑟, 𝑠, 𝑚)
of the set of 𝑠-vanishing polynomials introduced in section 5.
Remark 6.1. We observe that it is implicit in Definition (6.1) that Γ𝑚 must be orthogonal
to ∇𝑄(0). Infact, any 𝑃 ∈ 𝜎(𝑟, 𝑠, 𝑚) must satisfy ∇𝑃 (0) = 0 (see equation (5.8)).
Hence, the limit 𝑃̂ of any converging sequence {𝑃𝑛 ∈ 𝜎(𝑟, 𝑠, 𝑚)}𝑛∈ℕ must have the
same property. Since Σ(𝑟, 𝑠, 𝑚) = 𝜎(𝑟, 𝑠, 𝑚), one has ∇𝑃̂(0) = 0 for any 𝑃̂ ∈ Σ(𝑟, 𝑠, 𝑚).
As in our case we are considering 𝑃̂(𝑥) = 𝑄(𝐴𝑥)̂ for some 𝑄 ̂ ∈  ⋆ (𝑟, 𝑛), we have
̂ † ̂
∇𝑃 (0) = 𝐴 ∇𝑄(0) = 0, which is equivalent to saying that all the columns of 𝐴 are
̂
orthogonal to ∇𝑄(0).
The first result that we prove in this section is the following

Lemma 6.1. (𝑟, 𝑠, 𝑚, 𝑛) is a closed set in  ⋆ (𝑟, 𝑛) for the topology induced by (𝑟, 𝑛).

Proof. Consider a sequence {𝑄𝑗 }𝑗∈ℕ in (𝑟, 𝑠, 𝑚, 𝑛), converging to some polynomial
𝑄 ∈  ⋆ (𝑟, 𝑛) for the topology induced by (𝑟, 𝑛) on  ⋆ (𝑟, 𝑛). Then, for any fixed 𝑗 ∈
ℕ, by (6.1)-(6.2) there exists 𝐴𝑗 ∈ 𝑂(𝑛, 𝑚) such that 𝑃𝑗 (𝑥) ∶= 𝑄𝑗 (𝐴𝑗 𝑥) ∈ Σ(𝑟, 𝑠, 𝑚).
Since 𝑂(𝑛, 𝑚) is compact, there exists 𝐴 ∈ 𝑂(𝑛, 𝑚) and a subsequence {𝐴𝑗𝑘 }𝑘∈ℕ ⟶ 𝐴.
Hence, there exists a polynomial 𝑃 ∈ (𝑟, 𝑚) such that the subsequence {𝑃𝑗𝑘 (𝑥) ∶=
𝑄𝑗𝑘 (𝐴𝑗𝑘 𝑥)}𝑘∈ℕ converges to 𝑃 (𝑥) ∶= 𝑄(𝐴𝑥). Since 𝑃𝑗𝑘 (𝑥) ∈ Σ(𝑟, 𝑠, 𝑚) for all 𝑘 ∈ ℕ
by construction, and Σ(𝑟, 𝑠, 𝑚) is closed, 𝑃 (𝐴𝑥) ∈ Σ(𝑟, 𝑠, 𝑚), whence the thesis.

Moreover, for given values of 𝑚, 𝑛, when 𝑟 and 1 ≤ 𝑠 ≤ 𝑟 − 1 are sufficiently high,

the set (𝑟, 𝑠, 𝑚, 𝑛) becomes generic, namely

Lemma 6.2. (𝑟, 𝑠, 𝑚, 𝑛) is a semi-algebraic set of  ⋆ (𝑟, 𝑛) satisfying

codim (𝑟, 𝑠, 𝑚, 𝑛) ≥ max{0, 𝑠 − 𝑚(𝑛 − 𝑚 − 1)} . (6.3)

58
Proof. By hypothesis, Σ(𝑟, 𝑠, 𝑚) ∶= 𝜎(𝑟, 𝑠, 𝑚) and 𝜎(𝑟, 𝑠, 𝑚) is a semi-algebraic set (see
Corollary 5.1). Hence, Proposition A.2 assures that Σ(𝑟, 𝑠, 𝑚) is also semi-algebraic.
Therefore, set  in (6.1) is clearly semi-algebraic, since it is determined by a finite
number of semi-algebraic relations. Finally, the Theorem of Tarski and Seidenberg
A.1 implies that  is semi-algebraic since it is obtained by projecting  onto its first
component.
As for the codimension of , we start by estimating the dimension of  . We remark
that, for a fixed choice of 𝐴 ∈ 𝑂(𝑛, 𝑚) and 𝑃 ∈ Σ(𝑟, 𝑠, 𝑚), one has

dim( ∩ {(𝑄, 𝐴, 𝑃 ) ∈  ⋆ (𝑟, 𝑛) × 𝑂(𝑛, 𝑚) × (𝑟, 𝑚) ∶ 𝐴 = 𝐴 , 𝑃 = 𝑃 })

(6.4)
= dim  ⋆ (𝑟, 𝑛) − dim (𝑟, 𝑚) .

Infact, since the matrix 𝐴 ∶= (𝐴1 | … |𝐴𝑚 ), 𝐴1 , … , 𝐴𝑚 ∈ ℝ𝑛 , has been fixed, one can
construct an orthonormal basis of ℝ𝑛 by completing 𝐴1 , … , 𝐴𝑚 , with 𝑛 − 𝑚 vectors
𝐴𝑚+1 , … , 𝐴𝑛 . Since in the set above the restriction of any polynomial 𝑄 ∈ (𝑟, 𝑛)
to the space generated by 𝐴1 , … , 𝐴𝑚 is fixed, all the monomials of 𝑄 corresponding
to the coordinates associated to 𝐴1 , … , 𝐴𝑚 are uniquely determined. The number of
these monomials is dim (𝑟, 𝑚), and the total number of monomials in 𝑄 is dim (𝑟, 𝑛),
whence equality (6.4). In order to compute dim  , one must add to the r.h.s. of (6.4)
the dimension of the spaces corresponding to the fixed variables, namely

dim  = dim( ∩ {(𝑄, 𝐴, 𝑃 ) ∈  ⋆ (𝑟, 𝑛) × 𝑂(𝑛, 𝑚) × (𝑟, 𝑚) ∶ 𝐴 = 𝐴 , 𝑃 = 𝑃 })

+ dim 𝑂(𝑛, 𝑚) + dim Σ(𝑟, 𝑠, 𝑚)
= dim  ⋆ (𝑟, 𝑛) − dim (𝑟, 𝑚) + dim 𝑂(𝑛, 𝑚) + dim Σ(𝑟, 𝑠, 𝑚) .
(6.5)

We observe that, by Definition 6.1, if (𝑄(𝐼), 𝐴, 𝑄(𝐴𝑥)) ∈  , for any orthogonal matrix
𝑀 ∈ 𝑂(𝑚) also (𝑄(𝐼), 𝐴𝑀, 𝑄(𝐴𝑀𝑥)) ∈  , since one has the freedom to choose the
orthonormal basis spanning the 𝑚-dimensional subspace Γ𝑚 ∈ 𝙶(𝑚, 𝑛). More precisely,
if we define the action of 𝑂(𝑚) on any element (𝑄(𝐼), 𝐴, 𝑄(𝐴𝑥)) ∈  as

(𝑄(𝐼), 𝐴, 𝑄(𝐴𝑥)) ⟼ (𝑄(𝐼), 𝐴𝑀, 𝑄(𝐴𝑀𝑥)) , (6.6)

we can define an orbit of 𝑂(𝑚) starting at a given point (𝑄(𝐼), 𝐴, 𝑄(𝐴𝑥)) ∈  as

{(𝑄(𝐼), 𝐴𝑀, 𝑄(𝐴𝑀𝑥)) ∈  , 𝑀 ∈ 𝑂(𝑚)} . (6.7)

Since the first component in (6.7) is invariant, by (6.2) we see that the set  can be
found as the projection of the set of orbits  ∕𝑂(𝑚) onto  ⋆ (𝑟, 𝑛), namely

 ∶= Π ⋆ (𝑟,𝑛)  = Π ⋆ (𝑟,𝑛) ( ∕𝑂(𝑚)) , (6.8)

so that one can write

dim  = dim Π ⋆ (𝑟,𝑛) ( ∕𝑂(𝑚)) . (6.9)

59
Obviously, the action of 𝑂(𝑚) on  is free and smooth, hence by D.1 it is also proper.
Therefore, Theorem D.1 ensures that

dim( ∕𝑂(𝑚)) = dim  − dim 𝑂(𝑚) . (6.10)

By (6.9), we have

codim  ≥ codim ( ∕𝑂(𝑚)) − dim 𝑂(𝑛, 𝑚) − dim (𝑟, 𝑚) (6.11)

and equations (6.5) and (6.10) imply

codim ( ∕𝑂(𝑚))
= dim  ⋆ (𝑟, 𝑛) + dim 𝑂(𝑛, 𝑚) + dim (𝑟, 𝑚) − dim  + dim 𝑂(𝑚) . (6.12)
≥ 2 dim (𝑟, 𝑚) − dim Σ(𝑟, 𝑠, 𝑚) + dim 𝑂(𝑚)

Expressions (6.11) and (6.12) together yield

codim  ≥ dim (𝑟, 𝑚) − dim Σ(𝑟, 𝑠, 𝑚) + dim 𝑂(𝑚) − dim 𝑂(𝑛, 𝑚)

(6.13)
= codim Σ(𝑟, 𝑠, 𝑚) + dim 𝑂(𝑚) − dim 𝑂(𝑛, 𝑚) .

By Proposition A.3, codim Σ(𝑟, 𝑠, 𝑚) = codim 𝜎(𝑟, 𝑠, 𝑚), so by Corollary 5.1 one has
codim Σ(𝑟, 𝑠, 𝑚) = 𝑠 + 𝑚. Moreover, since dim 𝑂(𝑚) = 𝑚(𝑚 − 1)∕2 and dim 𝑂(𝑛, 𝑚) =
𝑚𝑛 − 𝑚(𝑚 − 1)∕2 − 𝑚, (6.13) reads

codim  ≥ 𝑠 − 𝑚(𝑛 − 𝑚 − 1) . (6.14)

Since the codimension is a nonnegative number, the thesis follows.

6.2 Stable lower estimates

The set (𝑟, 𝑠, 𝑚, 𝑛) introduced in the previous section is important because, for any
given polynomial 𝑄 ∈  ⋆ (𝑟′ , 𝑛) - with 𝑟′ ≥ 𝑟 - whose truncation at order 𝑟 lies out-
side of (𝑟, 𝑠, 𝑚, 𝑛), all polynomials belonging to a small open neighborhood of 𝑄 in
 ⋆ (𝑟′ , 𝑛) are steep around the origin on the subspaces of dimension 𝑚, with uniform
indices and coefficients. More precisely, one has

Theorem 6.1. Take five integers 𝑟′ ≥ 𝑟 ≥ 2, 1 ≤ 𝑠 ≤ 𝑟 − 1, 𝑛 ≥ 2, 𝑚 ∈ {1, ..., 𝑛 − 1}.

Consider a polynomial 𝑄 ∈  ⋆ (𝑟′ , 𝑛), and suppose that for some 𝜏 > 0 its truncation
at order 𝑟, indicated by 𝑄𝑟 , satisfies

||𝑄𝑟 −  (𝑟, 𝑠, 𝑚, 𝑛)|| ∶= inf ||𝑄𝑟 − 𝑅||∞ > 𝜏 . (6.15)

|| ||∞ 𝑅∈(𝑟,𝑠,𝑚,𝑛)

There exist constants 𝜀0 = 𝜀0 (𝑟, 𝑠, 𝑚, 𝜏, 𝑛), 𝙲𝑚 = 𝙲𝑚 (𝑟′ , 𝑟, 𝑠, 𝜏, 𝑛), 𝜆0 = 𝜆0 (𝑟, 𝑠, 𝑚, 𝜏),

and 𝛿̂ = 𝛿(𝑟,
̂ 𝑠, 𝑚, 𝜏, 𝑛) such that any polynomial 𝑆 ∈  ⋆ (𝑟′ , 𝑛) contained in a ball of

60
radius 𝜀 ∈ [0, 𝜀0 ] around 𝑄 in  ⋆ (𝑟′ , 𝑛) is steep on the subspaces of dimension 𝑚 at
̂ with uniform steepness coefficients 𝙲𝑚 , 𝜆0 and with steepness
any point 𝐼 ∈ 𝐵 𝑛 (0, 𝛿),
indices bounded by
{
𝑠 , if 𝑚 = 1
𝛼 𝑚 (𝑠) ∶= (6.16)
2𝑠 − 1 , if 𝑚 ≥ 2 .

Remark 6.2. By Lemma 6.2, since 1 ≤ 𝑠 ≤ 𝑟 − 1, if 𝑟 > 𝑚(𝑛 − 𝑚 − 1) + 1 and

𝑠 ≥ 𝑚(𝑛 − 𝑚 − 1) + 1 one has codim (𝑟, 𝑠, 𝑚, 𝑛) ≥ 1, so that hypothesis (6.15) is
generic in  ⋆ (𝑟, 𝑛).
In order to prove Theorem 6.1, we need a Lemma used by Pyartli in the study Dio-
phantine approximation. This result is crucial in KAM Theory, as it was shown by
Herman, Rüssmann, Sevryuk, and others (see [26] and the references therein). We give
its statement in a version provided by Rüssmann [56].

Lemma 6.3. Let 𝑓 ∶ [𝑎, 𝑏] ⟶ ℝ, with 𝑎 < 𝑏, be a 𝑞-times continuously differentiable

function satisfying
|𝑓 (𝑞) (𝑡)| > 𝛽
for all 𝑡 ∈ [𝑎, 𝑏], for some 𝑞 ∈ ℕ and 𝛽 > 0.
Then one has the estimate
( )1∕𝑞
𝜌
meas{𝑡 ∈ [𝑎, 𝑏] | |𝑓 (𝑡)| ≤ 𝜌} ≤ 4 𝑞!
2𝛽
for all 𝜌 > 0.

We also need the following auxiliary

Lemma 6.4. With the hypotheses of Theorem 6.1, there exist positive constants 𝜀⋆ =
𝜀⋆ (𝑟, 𝑠, 𝑚, 𝜏, 𝑛), 𝜒 = 𝜒(𝑟, 𝑠, 𝑚, 𝜏, 𝑛), and 𝜁 = 𝜁 (𝑟, 𝑠, 𝑚, 𝜏), such that, for any 𝜀 ∈ [0, 𝜀⋆ ],
the truncation 𝑆𝑟 ∈  ⋆ (𝑟, 𝑛) of any polynomial 𝑆 ∈  ⋆ (𝑟′ , 𝑛) contained in a ball of
radius 𝜀 around 𝑄 in  ⋆ (𝑟′ , 𝑛) verifies

||𝑆𝑟 − (𝑟, 𝑠, 𝑚, 𝑛)||∞ > 𝜒 , (6.17)

and for any 𝑚-dimensional subspace Γ𝑚 orthogonal to ∇𝑆(0), one has

||𝑆𝑟 |Γ𝑚 − Σ(𝑟, 𝑠, 𝑚)||∞ > 𝜁 . (6.18)

Proof. (Lemma 6.4)

We split the proof into three steps. In the first one, we introduce appropriate sets and
notations that are helpful in the proof. In second step, suitable continuous functions are
defined by exploiting the existence of local continuous sections for the Grassmannian.
In the third and last step, the statement is proved by exploiting the first two steps and
the compactness of the Grassmannian.

61
 We also observe that, due to Remark 6.2, estimate (6.17) is trivial for 𝑟 > 𝑚(𝑛 −
𝑚 − 1) + 1 and 𝑠 ≥ 𝑚(𝑛 − 𝑚 − 1) + 1.
Step 1. For any given polynomial 𝑉 ∈  ⋆ (𝑟, 𝑛), we denote by 𝙶𝑉 (𝑚, 𝑛) ⊂ 𝙶(𝑚, 𝑛)
the compact subset of 𝑚-dimensional subspaces orthogonal to ∇𝑉 (0) ≠ 0. We also
define the set
Λ𝑚 ∶= {(𝑉 , Γ𝑚 ) | 𝑉 ∈  ⋆ (𝑟, 𝑛), Γ𝑚 ∈ 𝙶𝑉 (𝑚, 𝑛)} . (6.19)
Now, setting 𝑁 ∶= dim  ⋆ (𝑟′ , 𝑛), for sufficiently small 𝜀 one has ∇𝑆(0) ≠ 0 for any
𝑆 ∈ 𝐵 𝑁 (𝑄, 𝜀). Moreover, the map 𝚏 ∶  ⋆ (𝑟′ , 𝑛) ⟶ ℝ𝑛 associating 𝑆 ⟼ ∇𝑆(0)
is obviously continuous and surjective, and the same holds true for the function 𝚑 ∶
ℝ𝑛 ⟶ 𝙶(𝑛 − 1, 𝑛) which to a vector 𝜔 associates 𝜔⟂ . Hence, 𝚑◦𝚏 is also continuous
and surjective. Therefore there exists an open set of 𝑛 − 1 dimensional hyperplanes -
indicated by  𝑛−1 (𝑄, 𝜀) ⊂ 𝙶(𝑛 − 1, 𝑛) - whose inverse image 𝚏−1 (𝚑−1 ( 𝑛−1 (𝑄, 𝜀)))
contains 𝐵 𝑁 (𝑄, 𝜀). Hence, for 𝑚 ∈ {1, … , 𝑛 − 1}, we can define the open set
 𝑚 (𝑄, 𝜀) ∶= {Γ𝑚 ∈ 𝙶(𝑚, 𝑛) | Γ𝑚 ⊆ Γ𝑛−1 for some Γ𝑛−1 ∈  𝑛−1 (𝑄, 𝜀)} .
The above construction implies that, for any 𝑚 ∈ {1, ..., 𝑛 − 1} and for sufficiently small
𝜀, the choice of an open ball 𝐵 𝑁 (𝑄, 𝜀) determines a set
Ξ𝑚 (𝑄, 𝜀) ∶= {(𝑆, Γ𝑚 ) | 𝑆 ∈ 𝐵 𝑁 (𝑄, 𝜀), Γ𝑚 ∈  𝑚 (𝑄, 𝜀), Γ𝑚 ∈ 𝙶𝑆 (𝑚, 𝑛)} ⊂ Λ𝑚 .
(6.20)
Remark 6.3. To carry out the construction at this step, one only needs to perturb the
linear terms of 𝑄. The bound on 𝜀 that must be considered at this step, therefore, does
not depend on the degree of the polynomial 𝑄.
Step 2. Now, we take into account the fact that it is always possible to define a local
continuous section for the Grassmannian 𝙶(𝑚, 𝑛). Namely, for any element Γ ∈ 𝙶(𝑚, 𝑛)
there exists an open neighborhood Γ ⊂ 𝙶(𝑚, 𝑛) of Γ and a continuous map 𝜉 ∶ Γ ⟶
𝑂(𝑛, 𝑚) such that, if 𝜋 ∶ 𝑂(𝑛, 𝑚) ⟶ 𝙶(𝑚, 𝑛) is the canonical projection, then 𝜋◦𝜉 is
the identity. Hence, for any 𝑚-dimensional subspace Γ𝑚 ∈ 𝙶𝑄 (𝑚, 𝑛), we consider its
associated open neighborhood Γ𝑚 , and a compact neighborhood 𝖵Γ𝑚 ⊂ Γ𝑚 centered
at Γ𝑚 . Since 𝙶𝑄 (𝑚, 𝑛) is compact, it can be covered by a finite number 𝐿 > 0 of compact
neighborhoods 𝖵𝑖 and open neighborhoods 𝑖 , with 𝑖 = 1, ..., 𝐿, of this kind. Hence, if
𝜀 is sufficiently small, then one has
⋃
𝐿 ⋃
𝐿
𝑚
 (𝑄, 𝜀) ⊂ 𝖵𝑖 ⊂ 𝑖 . (6.21)
𝑖=1 𝑖=1

Moreover, if we indicate by 𝜉𝑖 , 𝑖 = 1, ..., 𝐿 the continuous section associated to each

neighborhood 𝑖 , it makes sense to define the sets
⋃
𝐿
Ξ𝑚
𝑖 (𝑄, 𝜀) ∶= {(𝑆, Γ ) | (𝑆, Γ ) ∈ Ξ (𝑄, 𝜀), Γ ∈ 𝑖 } ,
𝑚 𝑚 𝑚 𝑚 𝑚
Ξ (𝑄, 𝜀) = Ξ𝑚
𝑖 (𝑄, 𝜀)
𝑖=1
(6.22)

62
and the continuous functions
′
𝐹𝑖 ∶ Ξ 𝑚
𝑖 (𝑄, 𝜀) ⟶ (𝑟 , 𝑚)
(6.23)
(𝑆(𝐼), Γ𝑚 ) ⟼ 𝑇 (𝑥) ∶= 𝑆(𝐴𝑥) , 𝐴 ∶= 𝜉𝑖 (Γ𝑚 ) ∈ 𝑂(𝑛, 𝑚) .

Step 3. Fix 𝑖 = 1, ..., 𝐿. By hypothesis ||𝑄𝑟 − ||∞ > 𝜏 and ∇𝑄(0) = ∇𝑄𝑟 (0) ≠ 0,
so that by the definition of  and  in (6.1)-(6.2), by Remark 6.1, and by the com-
pactness of 𝙶𝑄 (𝑚, 𝑛) = 𝙶𝑄𝑟 (𝑚, 𝑛), there exists 𝜁𝑖 = 𝜁𝑖 (𝑟, 𝑠, 𝑚, 𝜏) > 0 such that - on any
subspace Γ𝑚 ∈ 𝙶𝑄 (𝑚, 𝑛), Γ𝑚 ∈ 𝑖 - one has

||𝑃𝑟 − Σ(𝑟, 𝑠, 𝑚)||∞ > 2 𝜁𝑖 , (6.24)

where 𝑃𝑟 (𝑥) ∶= 𝑄𝑟 (𝐴𝑥) - with 𝐴 = 𝜉𝑖 (Γ𝑚 ) - is the restriction to the subspace Γ𝑚 of the
truncation 𝑄𝑟 .
Now, fix Γ𝑚 ∈ (𝙶𝑄 (𝑚, 𝑛) ∩ 𝖵𝑖 ) ⊂ 𝑖 . By the continuity of 𝐹𝑖 , there exists 𝜀⋆ 𝑖,Γ𝑚
=
𝜀𝑖,Γ𝑚 (𝑟, 𝑠, 𝑚, 𝜏, 𝑛) > 0 such that, for any 𝜀 ∈]0, 𝜀⋆
⋆
𝑖,Γ𝑚
], the open ball 𝜐 𝑚 (𝜀) ⊂ Ξ𝑚 (𝑄, 𝜀)
𝑖,Γ𝑚 𝑖
centered at (𝑄, Γ𝑚 ) verifies the following property: for any (𝑆, Γ ̂𝑚 ) ∈ 𝜐𝑚 𝑚 (𝜀), the
𝑖,Γ
restricted truncated polynomial 𝑇̂𝑟 (𝑥) ∶= 𝑆𝑟 (𝐴𝑥),
̂ with 𝐴̂ = 𝜉𝑖 (Γ
̂𝑚 ), is contained in an
open ball of radius 𝜁𝑖 around 𝑃𝑟 (𝑥) ∶= 𝑄𝑟 (𝐴𝑥), with 𝐴 = 𝜉𝑖 (Γ𝑚 ). Hence, on the one
hand by (6.24) one infers

||𝑇̂𝑟 − 𝑃𝑟 ||∞ < 𝜁𝑖 ⟹ ||𝑇̂𝑟 − Σ(𝑟, 𝑠, 𝑚)||∞ > 𝜁𝑖 . (6.25)

On the other hand, by construction one has

( ) ⋃
𝑄, 𝙶𝑄 (𝑚, 𝑛) ∩ 𝖵𝑖 ⊂ 𝜐𝑚 ⋆
𝑖,Γ𝑚 (𝜀𝑖,Γ𝑚 ) (6.26)
Γ𝑚 ∈ 𝙶𝑄 (𝑚,𝑛)∩𝖵𝑖

and - due to the compactness of the fiber (𝑄, 𝙶𝑄 (𝑚, 𝑛) ∩ 𝖵𝑖 ) - it is possible to extract
a finite number 𝐽𝑖 of subspaces Γ𝑚 ∈ 𝙶𝑄 (𝑚, 𝑛) ∩ 𝖵𝑖 from (6.26) and write (with slight
abuse of notation)

( ) 𝐽𝑖
⋃ ] ]
𝑄, 𝙶𝑄 (𝑚, 𝑛) ∩ 𝖵𝑖 ⊂ 𝜐𝑚 𝑚
𝑖,𝑗 (𝜀) ⊂ Ξ𝑖 (𝑄, 𝜀) , 𝜀 ∈ 0, 𝜀⋆
𝑖 , (6.27)
𝑗=1

where we have set

𝜀⋆ ⋆
𝑖 = 𝜀𝑖 (𝑟, 𝑠, 𝑚, 𝜏, 𝑛) ∶= min {𝜀⋆
𝑖,𝑗 } . (6.28)
𝑗∈{1,…,𝐽𝑖 }

𝐽
Inclusion (6.27), together with (6.22), yields that the finite union 𝜐𝑚 (𝜀) ∶= ∪𝐿 ∪ 𝑖
𝑖=1 𝑗=1
𝑚 𝑚
𝜐𝑖,𝑗 (𝜀) is an open neighborhood of Ξ (𝑄, 𝜀) containing the fiber (𝑄, 𝙶𝑄 (𝑚, 𝑛)). There-
fore, by setting

𝜁 = 𝜁 (𝑟, 𝑠, 𝑚, 𝜏) ∶= min {𝜁𝑖 } > 0 , 𝜀⋆ = 𝜀⋆ (𝑟, 𝑠, 𝑚, 𝜏, 𝑛) ∶= min {𝜀⋆

𝑖 }>0,
𝑖∈{1,…,𝐿} 𝑖=1,…,𝐿

63
and by taking (6.25) into account, one has that for 0 < 𝜀 ≤ 𝜀⋆ and for any (𝑆, Γ̂𝑚 ) ∈
𝜐𝑚 (𝜀) ⊂ Ξ𝑚 (𝑄, 𝜀), the restricted truncated polynomial 𝑇̂𝑟 (𝑥) ∶= 𝑆𝑟 (𝐴𝑥) - with 𝐴 =
̂𝑚 ) for some 𝑗 = 1, ..., 𝐿 - verifies
𝜉𝑗 ( Γ

||𝑇̂𝑟 − Σ(𝑟, 𝑠, 𝑚)||∞ > 𝜁𝑗 ≥ 𝜁 . (6.29)

Therefore, we have proved that, for any 0 ≤ 𝜀 ≤ 𝜀⋆ there exists 𝜁 > 0 such that

||𝑆𝑟 |Γ𝑚 − Σ(𝑟, 𝑠, 𝑚)||∞ > 𝜁 , (6.30)

for any 𝑆 ∈ 𝐵 𝑁 (𝑄, 𝜀) and for any Γ𝑚 orthogonal to ∇𝑆(0) = ∇𝑆𝑟 (0) ≠ 0. Hence,
the Definition of set (𝑟, 𝑠, 𝑚, 𝑛) in (6.2) ensures that for any 0 ≤ 𝜀 ≤ 𝜀⋆ there exists
𝜒 = 𝜒(𝑟, 𝑠, 𝑚, 𝜏, 𝑛) > 0 such that for any 𝑆 ∈ 𝐵 𝑁 (𝑄, 𝜀) one has

||𝑆𝑟 − (𝑟, 𝑠, 𝑚, 𝑛)||∞ > 𝜒 . (6.31)

This concludes the proof.

We need another intermediate result in order to demonstrate Theorem 6.1. Before

giving its statement, for any polynomial 𝑆 ∈  ⋆ (𝑟′ , 𝑛), we firstly consider its associated
minimal arc 𝛾 constructed in Theorem 4.1. Also, for any 𝜆 > 0 we indicate by 𝙸′𝜆 ⊂
[−𝜆, 𝜆] the interval obtained by cutting the interval 𝙸𝜆 at point 3) of Theorem 4.1 into
three equal pieces and by taking the central one. In particular, we have

|𝙸𝜆 | 𝜆
|𝙸′𝜆 | = = ,
3 3𝙺
where 𝙺 = 𝙺(𝑟′ , 𝑚, 𝑛) is a suitable constant.
We also assume the setting of Lemmata 6.1 and 6.4, and we consider a ball 𝐵 𝑁 (𝑄, 𝜀)
of radius 𝜀 < 𝜀⋆ ∕2 around 𝑄 ∈  ⋆ (𝑟′ , 𝑛). Within this framework, one has

Lemma 6.5. There exist two constants  ′ =  ′ (𝑟′ , 𝑟, 𝑠, 𝑚, 𝜏) and 𝜆0 = 𝜆0 (𝑟, 𝑠, 𝑚, 𝜏)

such that - for any number 0 < 𝜆 ≤ 𝜆0 , for any polynomial 𝑆 ∈  ⋆ (𝑟′ , 𝑛) contained in
𝐵 𝑁 (𝑄, 𝜀), and for any 𝑚-dimensional subspace Γ𝑚 orthogonal to ∇𝑆(0), the restriction
𝑇 ∶= 𝑆|Γ𝑚 satisfies the following estimates
| 𝛼 ( ) |
|𝑑 𝜕 𝑇 (𝑥) | |
max || 𝛼 |
|
| > ′ ,
| for 𝑚 = 1,
𝛼=1,...,𝑠 | 𝑑𝑡 𝜕 𝑥1 |𝛾(𝑡) |
| 𝑡=𝑡⋆ |
| 𝛼 ( ) | (6.32)
|𝑑 𝜕 𝑇 (𝑥) || |
max || 𝛼 |
| >  ′ 𝜆𝑠−1
| , for 𝑚 ≥ 2
𝓁=1,...,𝑚 | 𝑑𝑡 𝜕 𝑥𝓁 |𝛾(𝑡) |
𝛼=1,...,𝑠 | 𝑡=𝑡⋆ |

at any point 𝑡⋆ ∈ 𝙸′𝜆 .

64
Proof. (Lemma 6.5) We proceed by steps.
Step 1. We consider the value 𝜀⋆ of Lemma 6.4 and we fix a polynomial 𝑆 in the
ball 𝐵 𝑁 (𝑄, 𝜀), where 𝜀 ∈ [0, 𝜀⋆ ∕2]. By Lemma 6.4, there exists a parameter 𝜁 =
𝜁 (𝑟, 𝑠, 𝑚, 𝜏) > 0 such that on any 𝑚-dimensional subspace Γ𝑚 orthogonal to ∇𝑆(0) ≠ 0
- the truncation 𝑆𝑟 varifies

||𝑆𝑟 |Γ𝑚 − Σ(𝑟, 𝑠, 𝑚)||∞ > 𝜁 . (6.33)

Now, for a given subspace Γ𝑚 ∈ 𝙶𝑆 (𝑚, 𝑛), one can choose a matrix 𝐴 ∈ 𝑂(𝑛, 𝑚)
whose columns span Γ𝑚 and set 𝑇 (𝑥) ∶= 𝑆(𝐴𝑥). Then, for any 𝜆 > 0, by Theorem 4.1
there exists a minimal real-analytic arc
{
𝑥1 (𝑡) = 𝑡
𝛾(𝑡) ∶= 𝑡 ∈ 𝙸𝜆 ⊂ [−𝜆, 𝜆] , |𝙸𝜆 | = 𝜆∕𝙺(𝑟′ , 𝑛, 𝑚) ,
𝑥𝑗 (𝑡) = 𝑓𝑗 (𝑡), 𝑗 ∈ {2, ..., 𝑚}
(6.34)
whose image is contained in the thalweg  (𝑆, Γ𝑚 ). We observe that, up to a change
in the order of the vectors spanning Γ𝑚 , in Theorem 4.1, we can always suppose that
the coordinate parametrizing 𝛾 is the first one. We divide the interval 𝙸𝜆 into three
equal parts of length 𝜆∕(3𝙺) and we denote by 𝙸′𝜆 ∶= [𝜆𝑚𝑖𝑛 , 𝜆𝑚𝑎𝑥 ] the central one.
Then, for any given 𝑡⋆ ∈ 𝙸′𝜆 associated to the point 𝛾(𝑡⋆ ) = 𝑥⋆ , we consider the affine
reparametrization
{
⋆ 𝑥1 (𝑢) = 𝑢 + 𝑡⋆
𝛾 (𝑢) ∶= 𝑢 ∈ [𝜆𝑚𝑖𝑛 − 𝑡⋆ , 𝜆𝑚𝑎𝑥 − 𝑡⋆ ] .
𝑥𝑗 (𝑢) = 𝑓𝑗 (𝑢 + 𝑡⋆ ) , 𝑗 ∈ {2, ..., 𝑚}

It is clear that 𝛾 and 𝛾 ⋆ share the same image, have the same speed everywhere, and
that 𝛾 ⋆ (0) = 𝛾(𝑡⋆ ) = 𝑥⋆ , so that for all 𝛼 ∈ {0, ..., 𝑠} and 𝓁 ∈ {1, ..., 𝑚}, one has
( ) ( )
𝑑𝛼 𝜕𝑇 (𝑥) || 𝑑𝛼 𝜕𝑇 (𝑥) ||
= 𝛼 . (6.35)
𝑑𝑡𝛼 𝜕𝑥𝓁 ||𝛾(𝑡) 𝑑𝑢 𝜕𝑥𝓁 ||𝛾 ⋆ (𝑢)
𝑡=𝑡⋆ 𝑢=0

Now, indicating by 𝐿⋆ ∶ ℝ𝑚 ⟶ ℝ𝑚 , 𝑥 ⟼ 𝑥 − 𝑥⋆ =∶ 𝑥 ̃ the translation w.r.t. 𝑥⋆ , the

curve 𝛾 ⋆ is mapped into ̃
𝛾 ⋆ ∶= 𝐿⋆ ◦𝛾 ⋆ , which reads
{
⋆ ̃1 (𝑢) = 𝑥1 (𝑢) − 𝑥⋆
𝑥 1
= 𝑢 + 𝑡⋆ − 𝑡⋆ = 𝑢
𝛾 (𝑢) ∶=
̃ , (6.36)
̃𝑗 (𝑢) ∶= 𝑥𝑗 (𝑢) − 𝑥𝑗 (𝑡⋆ ) = 𝑓𝑗 (𝑢 + 𝑡⋆ ) − 𝑓𝑗 (𝑡⋆ ) =∶ 𝑓̃𝑗 (𝑢)
𝑥

with 𝑢 ∈ [𝜆𝑚𝑖𝑛 − 𝑡⋆ , 𝜆𝑚𝑎𝑥 − 𝑡⋆ ]. The polynomial 𝑇 written in the new coordinates reads

𝑇 (𝑥) = 𝑇 ◦𝐿−1
⋆ (̃ 𝑥 + 𝑥⋆ ) =∶ 𝑈 ⋆ (̃
𝑥) = 𝑇 (̃ 𝑥) . (6.37)

Since 𝑥⋆ is fixed, one has

𝜕𝑇 (𝑥) 𝜕𝑈 ⋆ (̃
𝑥)
= , ∀𝓁 ∈ {1, ..., 𝑚} . (6.38)
𝜕𝑥𝓁 𝜕̃𝑥𝓁

65
Moreover, if one takes into account the fact that ̃ 𝛾 ⋆ (0) = 0, and that the origin for the
̃ corresponds to the point 𝑥 = 𝑥⋆ in the old coordinates, equality (6.35),
coordinates 𝑥
together with (6.38), yields, for all 𝛼 ∈ {0, ..., 𝑠} and for all 𝓁 ∈ {1, ..., 𝑚},
( ) ( ) ( )
𝑑𝛼 𝜕𝑇 (𝑥) || 𝑑𝛼 𝜕𝑇 (𝑥) || 𝑑𝛼 𝑥) ||
𝜕𝑈 ⋆ (̃
= 𝛼 = 𝛼 | .
𝑑𝑡𝛼 𝜕𝑥𝓁 ||𝛾(𝑡) ⋆
𝑑𝑢 𝜕𝑥𝓁 ||𝛾 ⋆ (𝑢) 𝑑𝑢 𝑥𝓁 ||̃𝛾⋆ (𝑢)
𝜕̃
𝑡=𝑡 𝑢=0 𝑢=0
(6.39)
Step 2. Bernstein’s estimate (4.1) applied to the components of 𝛾 in (6.34) reads

max′ |𝑓𝑗 (𝑡)| ≤ 𝙼0 𝜆 , 𝑗 = 2, … , 𝑚 (6.40)

𝑡∈𝙸𝜆

for some uniform constant 𝙼0 = 𝙼(𝑟′ , 𝑛, 𝑚, 0), so that - for any 𝑡⋆ ∈ 𝙸′𝜆 - one has

||𝑥⋆ ||∞ ∶= ||𝛾(𝑡⋆ )||∞ ≤ 𝙼0 𝜆 . (6.41)

For any given 𝜀 ∈ [0, 𝜀⋆ ], with the help of the arguments in the proof of Lemma 6.4
- in particular taking (6.20), (6.22), and (6.23) into account - the set Ξ𝑚 (𝑄, 𝜀) admits
the covering Ξ𝑚 (𝑄, 𝜀) = ∪𝐿 Ξ𝑚 (𝑄, 𝜀), and for any index 𝑖 ∈ {1, ..., 𝐿} there exists a
𝑖=1 𝑖
continuous function
′
𝐹𝑖 ∶ Ξ𝑚 𝑖 (𝑄, 𝜀) ⟶ (𝑟 , 𝑚) (6.42)
that maps
(𝑅, Γ𝑚 ) ⟼ 𝑅(𝐴𝑥) , 𝐴 ∶= 𝜉𝑖 (Γ𝑚 ) ∈ 𝑂(𝑛, 𝑚) , (6.43)
where 𝜉𝑖 is a local continuous section for the Grassmannian 𝙶(𝑚, 𝑛).
Hence, taking Lemma 6.4 and formula (6.37) into account, the function
′
𝙵⋆ 𝑚 𝑚
𝑖 ∶ Ξ𝑖 (𝑄, 𝜀) × ℝ ⟶ (𝑟 , 𝑚)
(6.44)
(𝑅, Γ𝑚 , 𝑥⋆ ) ⟼ 𝑅(𝐴(̃
𝑥 + 𝑥⋆ )) , 𝐴 ∶= 𝜉𝑖 (Γ𝑚 ) ∈ 𝑂(𝑛, 𝑚)

is continuous. Moreover, (6.41) and the Theorem of Heine-Cantor ensure that 𝙵⋆

𝑖 is
𝑚 𝑛
uniformly continuous over the restricted compact domain Ξ𝑖 (𝑄, 𝜀⋆ ∕2) × 𝐵 ∞ (0, 𝙼0 𝜆).
Choosing the value 𝜁 (𝑟, 𝑠, 𝑚, 𝜏) > 0 in (6.33), there exists a uniform positive real 0 <
𝜆𝑖 = 𝜆𝑖 (𝜁 ) ≤ 1 such that, for any 0 < 𝜆 ≤ 𝜆𝑖 and for any (𝑆, Γ𝑚 ) ∈ Ξ𝑚 ⋆
𝑖 (𝑄, 𝜀 ∕2), the
𝑛
image of the set (𝑆, Γ𝑚 , 𝐵 ∞ (0, 𝙼0 𝜆)) through 𝙵⋆
𝑖 verifies
( ) ( )
𝑛 𝜁
⋆ 𝑚 𝑀 ⋆
𝙵𝑖 𝑆, Γ , 𝐵 ∞ (0, 𝙼0 𝜆) ⊂ 𝐵∞ 𝙵𝑖 (𝑆, Γ , 0), 𝑚
, 𝑀 ∶= dim (𝑟′ , 𝑚) . (6.45)
2

The above reasonings imply for any 𝑡⋆ ∈ 𝙸′𝜆 ⊂ [−𝜆, 𝜆], with 0 < 𝜆 ≤ 𝜆𝑖 , one has

𝜁
||(𝑇 ◦𝐿−1
⋆ )𝑟 − 𝑇𝑟 ||∞ =∶ ||𝑈𝑟 − 𝑇𝑟 ||∞ ≤ ||𝑈 − 𝑇 ||∞ <
⋆ ⋆
, (6.46)
2
where 𝑈𝑟⋆ , 𝑇𝑟 are the truncations at order 𝑟 of polynomial 𝑈 ⋆ introduced in (6.37) and
of polynomial 𝑇 ∶= 𝑆|Γ𝑚 , respectively.

66
 Repeating the same argument for any index 𝑗 ∈ {1, ..., 𝐿}, relations (6.33), (6.45),
and (6.46) imply that, if

0 < 𝜆 ≤ 𝜆0 ∶= min {𝜆𝑖 = 𝜆𝑖 (𝜉(𝑟, 𝑠, 𝑚, 𝜏))} ≤ 1 , (6.47)

𝑖∈{1,...,𝐿}

then, for any (𝑆𝑟 , Γ𝑚 ) ∈ Ξ𝑚 (𝑄𝑟 , 𝜀∕2) and for any 𝑡⋆ ∈ 𝙸′𝜆 one has

𝜁
||𝑈𝑟⋆ − Σ(𝑟, 𝑚, 𝑛)||∞ = ||𝑇𝑟 ◦𝐿−1
⋆ − Σ(𝑟, 𝑚, 𝑛)||∞ > . (6.48)
2
Step 3. Without any loss of generality, we suppose that the minimal curve 𝛾 ⊂
 (𝑆, Γ𝑚 ) is parametrized by the first coordinate. Hence, for 𝑛 ≥ 3 and 2 ≤ 𝑚 ≤ 𝑛 − 1
we indicate by 𝚊 = (𝑎12 , ..., 𝑎1𝑚 ) the linear coefficients of the Taylor expansion of the
translated curve ̃𝛾 ⋆ in (6.36). One can make use of the set of adapted coordinates ̃ 𝚢 ∶=
𝚊 (̃
𝑥) for the curve ̃ ⋆
𝛾 , as defined in paragraph 5.2.3. We remind that 𝚊 ∶= id in case
𝑚 = 1 (see also (5.72)).
In particular, for all 𝑚 ∈ {1, … , 𝑛−1} we write 𝚄⋆ ⋆ −1 𝛾𝚊⋆ ∶= 𝚊 ◦̃
𝛾 ⋆.
𝑟,𝚊 ∶= 𝑈𝑟 ◦𝚊 and ̃

By construction, the curve ̃ 𝛾𝚊⋆ ∈ Θ1𝑚 (see Definition (5.1) and Remark 5.1) is
analytic in [𝜆𝑚𝑖𝑛 − 𝑡 , 𝜆𝑚𝑎𝑥 − 𝑡⋆ ] with complex analyticity width 𝜆∕𝙺, and |𝙸′𝜆 | =
⋆

|𝜆𝑚𝑎𝑥 − 𝜆𝑚𝑖𝑛 | = 𝜆∕(3𝙺), as 𝙸′𝜆 was obtained by cutting 𝙸𝜆 into three equal pieces and by
taking the central one. By (6.48) and Lemma 5.3, there exist constants14 𝐶1 = 𝐶1 (𝜁 )
and 𝐶2 = 𝐶2 (𝑠, 𝑚) such that one has the lower estimate
| 𝛼 ( 𝜕 𝚄⋆ (𝚢) | ) |
|𝑑 | |
max | 𝛼 | 𝑟,𝚊
| | > 𝐶1
𝛼=1,...,𝑠 | 𝑑𝑢 | |
𝜕 𝚢1 |̃𝛾⋆ (𝑢) |
| 𝚊 𝑢=0 |
in case 𝑠 = 1 or 𝑚 = 1,
| 𝛼 ( 𝜕 𝚄⋆ (𝚢) | ) | (6.49)
|𝑑 | | 𝐶1
max || |>
𝑟,𝚊
| | 1 + 𝐶 × max
𝓁=1,...,𝑚 | 𝑑𝑢𝛼 𝜕 𝚢𝓁 ||̃𝛾⋆ (𝑢) | 2 𝓁=2,...,𝑚 |𝑎𝓁𝛼 |
𝛼=1,...,𝑠 | 𝚊 𝑢=0 | 𝛼=2,...,𝑠
in case 2 ≤ 𝑠 ≤ 𝑟 − 1 and 𝑚 ≥ 2 ,

where the 𝑎𝓁𝛼 ’s, with 𝓁 ∈ {2, ..., 𝑚} and 𝛼 ∈ {2, ..., 𝑠}, are the Taylor coefficients of
𝛾𝚊⋆ (𝑢) at the origin. Definition (6.36) assures that the Taylor coefficients of order equal
̃
or higher than one of the curve ̃ 𝛾 ⋆ (𝑢) at 𝑢 = 0, and those of the curve 𝛾 calculated at
𝑡 = 𝑡⋆ coincide. Moreover, by construction (see paragraph (5.2.3)) ̃ 𝛾𝚊⋆ (𝑢) and ̃
𝛾 ⋆ (𝑢)
share the same Taylor coefficients of order greater or equal than two calculated at the
origin. Hence, the Bernstein estimate in (4.1) applied to the second relation in (6.49)
and the fact that 0 < 𝜆 ≤ 𝜆0 ≤ 1 (see (6.47)) yield that there exists a uniform constant
14 InLemma 5.3, 𝐶1 depends on the open set 𝙳 ∈ (𝑟, 𝑚)∖Σ(𝑟, 𝑠, 𝑚). In our case, by (6.48), 𝙳 is the
open ball of radius 12 𝜁 around 𝑇 , which is at distance at least 12 𝜁 from Σ(𝑟, 𝑠, 𝑚); hence, with slight abuse of
notation, we can write 𝐶1 = 𝐶1 (𝜁).

67
𝙼 = 𝙼(𝑟′ , 𝑛, 𝑚, 𝑠) ≥ 1 such that estimate
| 𝛼 ( 𝜕 𝚄⋆ (𝚢) | ) |
|𝑑 | | 𝐶1
max | 𝛼 | 𝑟,𝚊
| |> 𝜆𝑠−1
𝓁=1,...,𝑚 | 𝑑𝑢 | | (6.50)
𝜕 𝚢𝓁 |̃𝛾⋆ (𝑢) | (1 + 𝐶2 ) 𝙼
𝛼=1,...,𝑠 | 𝚊 𝑢=0 |

holds in case 2 ≤ 𝑠 ≤ 𝑟 − 1 and 𝑚 ≥ 2.

Now, expression (5.33) together with estimate (4.1) yields

||−1
𝚊 ||∞ ≤ 1 + 𝙼 ,

for the matrix norm of −1 𝚊 . Therefore, by (5.35), by the first estimate in (6.49) and by
(6.50), we infer
| 𝛼 ( ) |
|𝑑 𝜕 𝑈𝑟⋆ (𝑥) || |
max || | | > 𝐶1
𝛼=1,...,𝑠 | 𝑑𝑢𝛼 𝜕 𝑥1 ||̃𝛾⋆ (𝑢) | 1+𝙼
|
| 𝑢=0 |
for 𝑠 = 1 or 𝑚 = 1,
(6.51)
| 𝛼 ( ) |
|𝑑 𝜕 𝑈𝑟⋆ (𝑥) || | 𝐶1
max | 𝛼 | | |> 𝜆𝑠−1
𝓁=1,...,𝑚 | 𝑑𝑢 | |
𝜕 𝑥𝓁 |̃𝛾⋆ (𝑢) | (1 + 𝐶2 )𝙼(1 + 𝙼)
𝛼=1,...,𝑠 | 𝑢=0 |

for 2 ≤ 𝑠 ≤ 𝑟 − 1, 𝑚 ≥ 2 .

Since in expression (6.51) one is considering only the derivatives up to order 𝑠 ∈

𝛾 ⋆ (𝑢) contains no constant terms, the same estimate
{1 … , 𝑟−1} at the origin 𝑢 = 0 and ̃
holds true for the polynomial 𝑈 ⋆ .
𝐶1
The thesis follows from (6.51) and from (6.39) by setting  ′ = .
(1 + 𝐶2 )𝙼(1 + 𝙼)

With the help of Lemmata 6.4 and 6.5, we are now able to prove Theorem 6.1.

Proof. (Theorem 6.1)

Introduction. We assume the setting and the notations of Lemmata 6.4-6.5. In
particular, for 0 < 𝜀 ≤ 𝜀⋆ ∕2 we consider a polynomial 𝑆 ∈  ⋆ (𝑟, 𝑛) in the ball
𝐵 𝑁 (𝑄, 𝜀), and a given 𝑚-dimensional subspace Γ𝑚 orthogonal to ∇𝑆(0) ≠ 0. We
denote by 𝛾 the minimal arc of Theorem 4.1 - whose image is contained in the thalweg
 (𝑆, Γ𝑚 ) - and for any 0 < 𝜆 ≤ 𝜆0 = 𝜆0 (𝑟, 𝑠, 𝑚, 𝜏) we indicate by 𝙸𝜆 its interval of
analyticity of length 𝜆∕𝙺, where 𝙺 = 𝙺(𝑟′ , 𝑛, 𝑚) is a uniform constant. We also indicate
by 𝙸′𝜆 the interval which is obtained by dividing 𝙸𝜆 into three equal parts and by taking
the central one.
Finally, we set 𝑇 (𝑥) ∶= 𝑆|Γ𝑚 (𝑥) ∶= 𝑆(𝐴𝑥), with 𝐴 ∈ 𝑂(𝑛, 𝑚) a matrix belonging
to the image of the continuous section 𝜉𝑖 ∶ 𝑖 (Γ𝑚 ) ⟶ 𝑂(𝑛, 𝑚), with 𝑖 ∈ {1, ..., 𝐿},
and whose columns span Γ𝑚 (see the proof of Lemma 6.4).

68
 We proceed by steps.
Step 1. For 𝓁 = 1, ..., 𝑚 and 𝛼 = {1, ..., 𝑠}, we consider the functions
( )
(𝛼) ⋆ 𝑑𝛼 𝜕𝑇 (𝑥) ||
𝑔𝓁 (𝑡 ) ∶= 𝛼 , 𝑡⋆ ∈ 𝙸′𝜆 (6.52)
𝑑𝑡 𝜕𝑥𝓁 ||𝛾(𝑡)
𝑡=𝑡⋆

and the constant functions

𝑔𝑚+1 (𝑡⋆ ) ∶=  ′ , 𝑔𝑚+2 (𝑡⋆ ) ∶= − ′ , 𝑔𝑚+3 (𝑡⋆ ) ∶=  ′ 𝜆𝑠−1 , 𝑔𝑚+4 (𝑡⋆ ) ∶= − ′ 𝜆𝑠−1 .

The degree of 𝑇 is bounded by 𝑟′ and - on the interval 𝙸𝜆 - 𝛾 is an analytic-algebraic

function whose diagram is bounded by a positive integer 𝚍 = 𝚍(𝑟′ , 𝑛, 𝑚) (see Point
2 of the thesis in Theorem 4.1). Hence, for any given choice of 𝛼 ∈ {1, ..., 𝑠} and
𝓁 ∈ {1, ..., 𝑚}, the function 𝑔𝓁(𝛼) (𝑡⋆ ) is Nash (i.e. semi-algebraic of class 𝐶 ∞ ) due to
Propositions A.8 and A.10, and its diagram is bounded by a quantity depending only on
𝑟′ , 𝑚, 𝑛. In addition, Proposition A.5 ensures that 𝑔𝓁(𝛼) (𝑡⋆ ) is actually analytic-algebraic
in 𝙸′𝜆 . The same is obviously true also for 𝑔𝑚+𝑗 (𝑡⋆ ) for 𝑗 ∈ {1, 2, 3, 4}. Therefore, we
̂ ′ , 𝑚, 𝑛) ∶= max𝑖∈{1,...,𝑚+4} max𝛼∈{1,...,𝑠} {diag (𝑔 (𝛼) )}.
set 𝑑̂ = 𝑑(𝑟 𝑖
Now, for any choice of 𝛼 ∈ {1, ..., 𝑠} and 𝑖 ∈ {1, ..., 𝑚 + 4}, the graph of 𝑔𝑖(𝛼) over
𝙸′𝜆 belongs to the algebraic curve of some non-constant polynomial 𝑉𝑖(𝛼) ∈ ℝ[𝑥, 𝑦] of
two variables, whose degree depends on 𝑑(𝑟 ̂ ′ , 𝑚, 𝑛). If we indicate by
( )𝑝𝑘
𝑉𝑖(𝛼) (𝑥, 𝑦) = Π𝐾(𝑖,𝛼)
𝑘=1
𝑉 (𝛼)
𝑖,𝑘
(𝑥, 𝑦) , with 𝐾(𝑖, 𝛼), 𝑝𝑘 ∈ ℕ⋆ (6.53)

the decomposition of 𝑉𝑖(𝛼) (𝑥, 𝑦) into its irreducible factors, by Bézout’s Theorem (see
{ }
(𝛼)
Th. E.2) the irreducible components (𝑥, 𝑦) ∈ ℝ2 |𝑉𝑖,𝑘 (𝑥, 𝑦) = 0 of the algebraic
{ }
curve (𝑥, 𝑦) ∈ ℝ2 |𝑉𝑖(𝛼) (𝑥, 𝑦) = 0 intersect at most at a finite number of points which
( )2
is bounded by deg 𝑉𝑖(𝛼) , which in turn is a quantity depending only on 𝑑(𝑟 ̂ ′ , 𝑚, 𝑛).

This fact, together with the regularity of 𝑔𝑖(𝛼) in 𝙸′𝜆 , implies that there exist two positive
integers 𝑘(𝑖, 𝛼) ∈ {1, … , 𝐾(𝑖, 𝛼)}, 𝓌 = 𝓌(𝑟′ , 𝑚, 𝑛), and a subinterval of 𝙸⋆
𝜆,𝑖,𝛼
⊂ 𝙸′𝜆 of
length |𝙸𝜆 |∕𝓌 verifying the two following conditions:
′

( ) { }
|
graph 𝑔𝑖(𝛼) | ⋆ ⊂ 2
(𝑥, 𝑦) ∈ ℝ |𝑉 (𝛼)
(𝑥, 𝑦) = 0
|𝙸𝜆,𝑖,𝛼 𝑖,𝑘(𝑖,𝛼)
( ) { } (6.54)
|
graph 𝑔𝑖(𝛼) | ⋆ (𝛼)
∩ (𝑥, 𝑦) ∈ ℝ2 |𝑉𝑖,𝑘 (𝑥, 𝑦) = 0 = ∅
|𝙸𝜆,𝑖,𝛼

for all 𝑘 ∈ {1, … , 𝐾(𝑖, 𝛼)}∖{𝑘(𝑖, 𝛼)}.

The above reasoning can be repeated for all other pairs of integers belonging to
{1, … , 𝑠}×{1, … , 𝑚+4} and which are different from (𝛼, 𝑖). Hence, finally there exists

69
an interval 𝙸⋆
𝜆
⊂ 𝙸′𝜆 of length |𝙸′𝜆 |∕(𝓌(𝑚+4) 𝑠 ) on which for any (𝑖′ , 𝛼 ′ ) ∈ {1, … , 𝑠} ×
{1, … , 𝑚 + 4} the relations
( ) { }
′ | ′)
graph 𝑔𝑖(𝛼′ ) | ⋆ ⊂ (𝑥, 𝑦) ∈ ℝ2 |𝑉 (𝛼 (𝑥, 𝑦) = 0
|𝙸𝜆 𝑖′ ,𝑘(𝑖′ ,𝛼 ′ )
( ) { } (6.55)
(𝛼 ′ ) | 2 (𝛼 ′ )
graph 𝑔𝑖′ | ⋆ ∩ (𝑥, 𝑦) ∈ ℝ |𝑉𝑖′ ,𝑘 (𝑥, 𝑦) = 0 = ∅
|𝙸𝜆

are verified for some integer 𝑘(𝑖′ , 𝛼 ′ ) ∈ {1, … , 𝐾(𝑖′ , 𝛼 ′ )} and for any integer 𝑘 ∈
{1, … , 𝐾(𝑖′ , 𝛼 ′ )}∖{𝑘(𝑖′ , 𝛼 ′ )}.
Then, by (6.55) and again by Bézout’s Theorem, there exists a positive integer
𝖭 = 𝖭(𝑟′ , 𝑚, 𝑛) such that for any distinct pairs of integers (𝛼, 𝑖) and (𝛽, 𝑗) belonging
to {1, … , 𝑠} × {1, … , 𝑚 + 4}, the algebraic curves {(𝑥, 𝑦) ∈ (𝙸⋆ 𝜆
, ℝ)|𝑉 (𝛼) (𝑥, 𝑦) = 0}
𝑖,𝑘(𝑖,𝛼)
and {(𝑥, 𝑦) ∈ (𝙸⋆
𝜆
, ℝ)|𝑉 (𝛽) (𝑥, 𝑦) = 0} either coincide or intersect at most at 𝖭 =
𝑗,𝑘(𝑗,𝛽)
𝖭(𝑟′ , 𝑚, 𝑛) points.
By repeating this reasoning for all possible distinct pairs and by taking (6.55) into
account, one finally has that there exists a positive constant 𝖬 = 𝖬(𝑟′ , 𝑠, 𝑚, 𝑛) and
an interval 𝙹⋆ 𝜆
of uniform length |𝙹⋆
𝜆
| = |𝙸⋆𝜆
|∕𝖬 over which the graphs of any pair
of functions among 𝑔1(1) , ..., 𝑔𝑚(1) , ..., 𝑔1(𝑠) , ..., 𝑔𝑚(𝑠) , 𝑔𝑚+1 , ..., 𝑔𝑚+4 either do not intersect or
coincide.
These reasonings - together with the fact that expression (6.32) in Lemma 6.5 holds
for all 𝑡⋆ ∈ 𝙹⋆
𝜆
⊂ 𝙸′𝜆 - yield that there must exist 𝛼 ∈ {1, ..., 𝑠} and 𝓁 ∈ {1, ..., 𝑚}
verifying
| |
min⋆ ||𝑔 (𝛼) (𝑡⋆ )|| >  ′ for 𝑚 = 1
𝑡 ∈𝙹𝜆 |
⋆ 𝓁 |
(6.56)
| (𝛼) ⋆ |
min⋆ ||𝑔 (𝑡 )|| >  ′ 𝜆𝑠−1 for 𝑚 ≥ 2 .
𝑡⋆ ∈𝙹𝜆 | 𝓁 |
Step 2. We apply Lemma 6.3 to 𝑔𝓁 , with 𝑞 ≡ 𝛼, [𝑎, 𝑏] ≡ 𝙹⋆
𝜆
and with 𝛽 equal to the
r.h.s. of (6.56). If we ask for
( )1∕𝛼
𝜌 |𝙹⋆ | 𝜆
4 𝛼! ≤ 𝜆 = (6.57)
2𝛽 2 6 𝙺 𝖬 𝓌(𝑚+4) 𝑠
and we take into account the fact that 𝛼 ∈ {1, ..., 𝑠}, we can choose
{
2 𝜆𝑠  ′ for 𝑚 = 1
𝜌= [ ]𝑠 × (6.58)
𝑠 ! × 24 𝙺 𝖬 𝓌(𝑚+4) 𝑠  ′ 𝜆𝑠−1 for 𝑚 ≥ 2 .

|𝙹⋆
𝜆
| 𝜆
Hence, in an open set 𝙰𝜆 ⊂ 𝙹⋆
𝜆
of measure = , one has
2 6 𝙺 𝖬 𝓌(𝑚+4) 𝑠
{
𝙲 1 𝜆𝑠 , for 𝑚 = 1
|𝑔𝓁 (𝑡)| > 𝜌 = ∀𝑡 ∈ 𝙰𝜆 (6.59)
𝙲𝑚 𝜆2𝑠−1 , for 𝑚 ≥ 2

70
for some 𝓁 ∈ {1, ..., 𝑚}, and for a constant
2 ′ (𝑟′ , 𝑟, 𝑠, 𝑚, 𝜏)
𝙲𝑚 = 𝙲𝑚 (𝑟′ , 𝑟, 𝑠, 𝜏, 𝑛) = [ ]𝑠 , (6.60)
𝑠 ! × 24 𝙺(𝑟′ , 𝑚, 𝑛) 𝖬(𝑟′ , 𝑠, 𝑚, 𝑛) 𝓌(𝑟′ , 𝑚, 𝑛)(𝑚+4) 𝑠
where 𝑚 ∈ {1, … , 𝑛 − 1}.
Step 3. Taking the definition of 𝙰𝜆 into account, by construction (see (6.52)) we
have
| | | |
| | | 𝜕𝑇 (𝑥) || | | 𝜕𝑆|Γ𝑚 (𝑥) || |
max |𝑔𝓁 (𝑡)| ∶= max || | || ∶= max || | || . (6.61)
𝑡∈𝙰𝜆 | | 𝑡∈𝙰𝜆 | 𝜕𝑥 | | 𝑡∈𝙰𝜆 | 𝜕𝑥𝓁 ||𝛾(𝑡) |
| 𝓁 |𝛾(𝑡) | | |
Due to Theorem 4.1, the image of 𝛾 is contained in the thalweg  (𝑆, Γ𝑚 ), that is in the
locus of minima of 𝑇 ∶= 𝑆|Γ𝑚 on the spheres 𝜂𝑚 ⊂ Γ𝑚 of radius 𝜂 > 0 centered at
the origin. Moreover, the curve 𝛾 was constructed by a uniform local inversion theorem
applied to the curve 𝜙 of Lemma 4.2 that was parametrized by the radius 𝜂 > 0 of the
spheres 𝜂𝑚 ⊂ Γ𝑚 and shared the same image with 𝛾. So, to any value of 𝑡 ∈ 𝙰𝜆 there
𝑚 ⊂ Γ𝑚 .
corresponds a unique radius 𝜂(𝑡) associated to a sphere 𝜂(𝑡)
Hence, taken any pair of real numbers 𝜆, 𝜉 satisfying 0 < 𝜆 ≤ 𝜉 ≤ 𝜆0 - where
𝜆0 is the quantity defined in Lemma 6.5 - by the discussion at step 3 of the proof of
Theorem 4.1 (in particular, the inclusions in (4.22)), one has that the inverse image of
𝙰𝜆 is contained in the interval 𝜉 ⊂ [0, 𝜉] defined in Lemma 4.2. This argument and
(6.61) imply that for some 𝓁 ∈ {1, ..., 𝑚} one has
| | | |
| 𝜕𝑆|Γ𝑚 (𝑥) || | | 𝜕𝑆|Γ𝑚 (𝑥) || | | 𝜕𝑆| 𝑚 (𝑥) |
| |
max || | || ≤ max || | || = max min |
Γ
| (6.62)
𝑡∈𝙰𝜆 | 𝜕𝑥𝓁 ||𝛾(𝑡) | 𝜂∈[0,𝜉] | 𝜕𝑥𝓁 ||𝜙(𝜂) | 𝜂∈[0,𝜉] ||𝑥||2 =𝜂 || 𝜕𝑥𝓁 ||
| | | |
which in turn, as 0 < 𝜆 ≤ 𝜉 ∈ (0, 𝜆0 ], by taking (6.59)-(6.61) and the equivalence of
norms into account, implies that
{
|| || 𝙲1 𝜆𝑠 , for 𝑚 = 1
max min ||∇𝑆|Γ𝑚 (𝑥)||2 > , ∀ 0 < 𝜆 ≤ 𝜉 ∈ (0, 𝜆0 ] .
𝜂∈[0,𝜉] ||𝑥||2 =𝜂 𝙲𝑚 𝜆 2𝑠−1 , for 𝑚 ≥ 2
(6.63)
Since the coordinates 𝑥 are associated to an orthonormal basis spanning Γ , for any 𝑚

point 𝐼 ∈ ℝ𝑛 contained in the subspace Γ𝑚 one has 𝜋Γ𝑚 (∇𝐼 𝑆(𝐼)) ≡ ∇𝑥 (𝑆|Γ𝑚 )(𝑥), and
by choosing 𝜆 = 𝜉 in (6.63), we have proved that any polynomial 𝑆 ∈  ⋆ (𝑟′ , 𝑛) in the
ball 𝐵 𝑁 (𝑄, 𝜀), with 𝜀 ≤ 𝜀⋆ ∕2, is steep at the origin on the subspaces of dimension 𝑚
with index bounded as in (6.16) and with coefficients 𝙲𝑚 , 𝜆0 . It remains to prove that
this holds true also in a neighborhood of the origin.
Step 4. For any polynomial 𝑆 ∈  ⋆ (𝑟′ , 𝑛), we consider the translation
𝙷⋆ ∶  ⋆ (𝑟′ , 𝑛) × ℝ𝑛 ⟶  ⋆ (𝑟′ , 𝑛) , (𝑆(𝐼), 𝐼 ⋆ ) ⟼ 𝑆(𝐼 + 𝐼 ⋆ ) . (6.64)
𝑁 𝑛
𝙷⋆ is uniformly continuous over the compact 𝐵 (𝑄, 𝜀⋆ ∕4) × 𝐵 (0, 1). Hence, there
̂ ⋆ ) > 0 such that for any 𝑆 ∈ 𝐵 𝑁 (𝑄, 𝜀⋆ ∕4), one has
exists a number 𝛿̂ = 𝛿(𝜀
𝑛
𝙷⋆ ({𝑆} × 𝐵 (0, 𝛿))
̂ ⊂ 𝐵 𝑁 (𝑆, 𝜀⋆ ∕4) . (6.65)

71
Hence, for any given point 𝐼 ⋆ satisfying ||𝐼 ⋆ ||2 < 𝛿̂ and for any polynomial 𝑆 ∈
𝑁
𝐵 (𝑄, 𝜀⋆ ∕4), the polynomial 𝑆(𝐼 + 𝐼 ⋆ ) belongs to 𝐵 𝑁 (𝑄, 𝜀⋆ ∕2).
𝑁
Now, we consider a polynomial 𝑆 ∈ 𝐵 (𝑄, 𝜀⋆ ∕4). By the above reasonings,
for any 𝐼 ⋆ verifying ||𝐼 ⋆ ||2 < 𝛿,̂ one has that its translation 𝑆(𝐼 + 𝐼 ⋆ ) belongs to
𝐵 𝑁 (𝑄, 𝜀⋆ ∕2). We have proved at Step 3 that any polynomial in  ⋆ (𝑟′ , 𝑛) belonging to
𝐵 𝑁 (𝑄, 𝜀) - with 𝜀 ∈ [0, 𝜀⋆ ∕2] - is steep at the origin on the subspaces of dimension
𝑚, with index as in (6.16), and with uniform coefficients 𝙲𝑚 , 𝜆0 . Consequently, for any
given 𝐼 ⋆ satisfying ||𝐼 ⋆ ||2 < 𝛿,
̂ the polynomial 𝑆(𝐼 + 𝐼 ⋆ ) is steep at the origin on the
𝑚-dimensional subspaces, with uniform index and coefficients. This is equivalent to
stating the same property for polynomial 𝑆 at any point 𝐼 ⋆ in a ball of radius 𝛿̂ around
the origin.
The thesis follows by setting 𝜀0 = 𝜀0 (𝑟, 𝑠, 𝑚, 𝜏, 𝑛) ∶= 𝜀⋆ ∕4.

6.3 Proof of the genericity of steepness

With the help of Theorem 6.1, we are finally able to prove Theorem A.

Proof. (Theorem A)
It is sufficient to study the case in which 𝐼0 = 0, else one considers the translated
function ℎ0 (𝐼) ∶= ℎ(𝐼 + 𝐼0 ). We proceed by steps.
Step 1. For any choice of integers 𝑟, 𝑛 ≥ 2, and for any given 𝚜 = (𝑠1 , … , 𝑠𝑛−1 ) ∈
𝑛−1
ℕ , where 𝑠𝑚 ∈ {1, … , 𝑟 − 1} for all 𝑚 ∈ {1, … , 𝑛 − 1}, by taking (6.2) into account,
we define
⋃
𝑛−1
𝛀𝑟,𝚜
𝑛 ∶= (𝑟, 𝑠𝑚 , 𝑚, 𝑛) ⊂  ⋆ (𝑟, 𝑛) . (6.66)
𝑚=1
The above set is closed due to Lemma 6.1.
𝑛
For any given pair 𝜚, 𝜏 > 0, we consider a function ℎ ∈ 𝒟 ⊂ 𝐶𝑏2𝑟−1 (𝐵 (0, 𝜚))
satisfying
∇ℎ(0) ≠ 0 , ||𝚃0 (ℎ, 𝑟, 𝑛) − 𝛀𝑟,𝚜 || > 𝜏 . (6.67)
|| 𝑛 ||∞

Now, for 𝑚 = 1, … , 𝑛 − 1, taking the definition of 𝜀0 (𝑟, 𝑠𝑚 , 𝑚, 𝜏, 𝑛) in Theorem 6.1

into account, we set
1
𝜖 = 𝜖(𝑟, 𝚜, 𝜏, 𝑛) ∶= × min {𝜀 (𝑟, 𝑠𝑚 , 𝑚, 𝜏, 𝑛) > 0} . (6.68)
2 𝑚∈{1,...,𝑛−1} 0
𝑛
Then, for 𝜀 ∈ [0, 𝜖], we consider a function 𝑓 ∈ 𝒟 ⊂ 𝐶𝑏2𝑟−1 (𝐵 (0, 𝜚)) satifying
𝑛
𝑓 ∈ 𝕭2𝑟−2 (ℎ, 𝜀, 𝐵 (0, 𝜚)) . (6.69)

Due to (6.69), 𝚃0 (𝑓 , 2𝑟−2, 𝑛) is contained in a ball of radius 𝜀 around 𝚃0 (ℎ, 2𝑟−2, 𝑛) in

(𝑟, 𝑛). Hence, as by construction we have set 𝜀 ≤ 𝜖, where 𝜖 was defined in (6.68), the

72
definition of set 𝛀𝑟,𝚜
𝑛 in (6.66), together with condition (6.67) yields that we can apply
Theorem 6.1 with 𝑟′ = 2𝑟 − 2. In turn, this ensures the existence of positive constants
𝙲𝑚 = 𝙲𝑚 (𝑟′ = 2𝑟 − 2, 𝑟, 𝑠𝑚 , 𝜏, 𝑛),

𝑑̂ = 𝑑(𝑟,
̂ 𝚜, 𝜏, 𝑛) ∶= min ̂ 𝑠𝑚 , 𝑚, 𝜏, 𝑛) ,
𝛿(𝑟,
𝑚∈{1,…,𝑛−1}

and
𝜆 = 𝜆(𝑟, 𝚜, 𝜏) ∶= min 𝜆0 (𝑟, 𝑠𝑚 , 𝑚, 𝜏)
𝑚∈{1,...,𝑛−1}

such that 𝚃0 (𝑓 , 2𝑟 − 2, 𝑛) is steep in an open ball of radius 𝑑̂ around the origin with
̄ 𝙲𝑚 , 𝑚 = 1, … , 𝑛 − 1, and with indices
coefficients 𝜆,
{
𝑠1 , if 𝑚 = 1
𝛼 𝑚 (𝑠𝑚 ) ∶= (6.70)
2𝑠𝑚 − 1 , if 𝑚 ≥ 2 .

Step 2. For any 𝐼 ∈ 𝐵 𝑛 (0, 𝑅) - with

{ }
̂ 𝚜, 𝜏, 𝑛)
𝜚 𝑑(𝑟,
𝑅 = 𝑅(𝑟, 𝚜, 𝜏, 𝑛, 𝜚) ∶= min , , (6.71)
3 2

for any 𝑚 ∈ {1, … , 𝑛 − 1}, and for any 𝑚-dimensional affine subspace Γ𝑚 = Γ𝑚 (𝐼)
passing through 𝐼 and orthogonal to ∇𝑓 (𝐼) ≠ 0, we indicate by 𝑓 |Γ𝑚 the restriction
of 𝑓 to Γ𝑚 . We assume that any given Γ𝑚 (𝐼) is endowed with the induced euclidean
metric, and we indicate by 𝑥 a suitable system of coordinates on Γ𝑚 (𝐼) whose origin
𝑥 = 0 corresponds to point 𝐼. Moreover, for any 𝛽 ∈ {1, … , 𝑛}, we set 𝜕𝛽 ∶= 𝜕𝑥𝜕 and
𝛽
{ 𝜚}
𝜅 ∶= min 𝜆, ,
3
Now, we fix both 𝐼 ∈ 𝐵 𝑛 (0, 𝑅) and Γ𝑚 (𝐼). By standard calculus, at any point 𝑥
verifying ||𝑥|| ≤ 𝜅 (hence, sufficiently close to 𝐼), one can write

|𝜕𝛽 (𝑓 |Γ𝑚 )(𝑥) − 𝚃0 (𝜕𝛽 (𝑓 |Γ𝑚 ), 2𝑟 − 3, 𝑚)(𝑥)|

||𝑥||22𝑟−2 (6.72)
≤ 𝐾(𝑟, 𝑚) max𝑛 max |𝐷𝛼 𝑓 (𝐼 ′ )|
𝑛
α∈ℕ 𝐼 ′ ∈𝐵 (0,𝜚) (2𝑟 − 2)!
|α|=2𝑟−1

for some constant 𝐾(𝑟, 𝑚) > 0.

Since 𝚃0 (𝜕𝛽 (𝑓 |Γ𝑚 ), 2𝑟 − 3, 𝑚)(𝑥) = 𝜕𝛽 [𝚃0 (𝑓 |Γ𝑚 , 2𝑟 − 2, 𝑚)](𝑥), taking (6.72) into
account, we have
|𝜕𝛽 (𝑓 |Γ𝑚 )(𝑥)|
| |
≥ | | 𝜕𝛽 [ 𝚃0 (𝑓 |Γ𝑚 , 2𝑟 − 2, 𝑚) ] (𝑥) | − |𝜕𝛽 (𝑓 |Γ𝑚 )(𝑥) − 𝚃0 (𝜕𝛽 (𝑓 |Γ𝑚 ), 2𝑟 − 3, 𝑚) (𝑥)| |
| |
≥ | 𝜕𝛽 [ 𝚃0 (𝑓 |Γ𝑚 , 2𝑟 − 2, 𝑚) ] (𝑥) | − 𝑐(𝑟, 𝑛, 𝑚, 𝜚, 𝒟 ) ||𝑥||2 2𝑟−2 ,
(6.73)

73
where we have indicated
𝐾(𝑟, 𝑚)
𝑐 = 𝑐(𝑟, 𝑛, 𝑚, 𝜚, 𝒟 ) ∶= max ||𝑔||𝐶 2𝑟−1 (𝐵 𝑛 (0,𝜚)) .
(2𝑟 − 2)! 𝑔∈𝒟 𝑏

Estimate (6.73) implies that for any 𝑥 ∈ Γ𝑚 (𝐼) verifying ||𝑥|| ≤ 𝜅 we can write

||∇(𝑓 |Γ𝑚 )(𝑥)||1 ≥ ||||∇𝚃0 (𝑓 |Γ𝑚 , 2𝑟 − 2, 𝑚)(𝑥)||||1 − 𝑐 𝑛 ||𝑥||2 2𝑟−2 ,

and, by the equivalence of norms,

1 ||
||∇(𝑓 |Γ𝑚 )(𝑥)||2 ≥ ∇𝚃 (𝑓 |Γ𝑚 , 2𝑟 − 2, 𝑚)(𝑥)||||2 − 𝑐 ||𝑥||2 2𝑟−2 . (6.74)
𝑛 || 0
Step 3. By the discussion at Step 1, 𝚃0 (𝑓 , 2𝑟 − 2, 𝑛) is steep in an open ball of radius
𝑑̂ around the origin 𝐼 = 0, with coefficients 𝜆, ̄ 𝙲𝑚 , 𝑚 = 1, … , 𝑛 − 1, and with indices
as in (6.70). This property, together with expression (6.74) and with the fact that

− the origin 𝑥 = 0 on Γ𝑚 (𝐼) corresponds to point 𝐼 ∈ 𝐵 𝑛 (0, 𝑅) by construction;

̂ by (6.71);
− 𝑅 ≤ 𝑑∕2

yields
𝙲1 𝑠
max min ||∇(𝑓 |Γ1 )(𝑥)||2 > 𝜉 1 − 𝑐 𝜉 2𝑟−2 ∀𝜉 ∈ [0, 𝜅] (𝑚 = 1) (6.75)
𝜂∈[0,𝜉] ||𝑥||2 ∈Γ1 (𝐼) 𝑛
||𝑥||2 =𝜂

𝙲𝑚 2𝑠 −1
max min𝑚 ||∇(𝑓 |Γ𝑚 )(𝑥)||2 > 𝜉 𝑚 −𝑐 𝜉 2𝑟−2 ∀𝜉 ∈ [0, 𝜅] (2 ≤ 𝑚 ≤ 𝑛−1) .
𝜂∈[0,𝜉] ||𝑥||2 ∈Γ (𝐼) 𝑛
||𝑥||2 =𝜂
(6.76)
If we impose

⎧ 2𝑟−2 𝙲1 𝑠
⎪𝑐 𝜉 ≤ 𝜉1 , if 𝑚 = 1
⎪ 2𝑛
⎨ (6.77)
⎪ 2𝑟−2 𝙲𝑚 2𝑠 −1
⎪𝑐 𝜉 ≤ 𝜉 𝑚 , if 𝑚 = 2, … , 𝑛 − 1
⎩ 2𝑛
by (6.75)-(6.76), and by the fact that 𝑠𝑚 ≤ 𝑟 − 1 for all 𝑚 = 1, … , 𝑛 − 1, we have that 𝑓
is steep in a ball of radius 𝑅 around the origin with coefficients (we have set 𝑟′ = 2𝑟 − 2)
𝛿 = 𝛿(𝑟, 𝚜, 𝜏, 𝑛, 𝜚, 𝒟 )
{ ( ) 1 {( ) 1 }}
𝙲1 (𝑟′ , 𝑟, 𝑠1 , 𝜏, 𝑛) 2𝑟−2−𝑠1 𝙲𝑚 (𝑟′ , 𝑟, 𝑠𝑚 , 𝜏, 𝑛) 2(𝑟−𝑠𝑚 )−1
∶= min 𝜅, , min ,
2 𝑛 𝑐(𝑟, 𝑛, 𝑚, 𝜚, 𝒟 ) 𝑚∈{2,…,𝑛−1} 2 𝑛 𝑐(𝑟, 𝑛, 𝑚, 𝜚, 𝒟 )
(6.78)

𝙲𝑚 (𝑟′ , 𝑟, 𝑠𝑚 , 𝑚, 𝜏, 𝑛)
𝐶𝑚 (𝑟′ , 𝑟, 𝑠𝑚 , 𝜏, 𝑛) ∶= ,
2𝑛

74
and with indices bounded as in (6.70).
It remains to prove the estimate on the codimension of 𝛀𝑟,𝚜
𝑛 . By (6.66), Lemma 6.2
and Proposition A.3 we have
{ }
𝑟,𝚜
codim 𝛀𝑛 ≥ max 0, min {𝑠𝑚 − 𝑚(𝑛 − 𝑚 − 1)} .
𝑚∈{1,…,𝑛−1}

This concludes the proof.

7 Proof of Theorem B and of its Corollaries

Hereafter, we assume the notations and the results of the previous sections.

7.1 Proof of Theorem B

It suffices to prove the statement for 𝐼0 = 0, otherwise one considers ℎ0 (𝐼) ∶= ℎ(𝐼 +𝐼0 ).

7.1.1 Case 𝑚 = 1.
Let Γ1 be a 1-dimensional subspace (a line) orthogonal to ∇ℎ(0) ≠ 0, and let 𝑤 ∈ 𝕊𝑛
be its generating vector. By standard results of calculus, the restriction of the Taylor
polynomial 𝚃0 (ℎ, 𝑟, 𝑛) to Γ1 , indicated by 𝚃0 (ℎ|Γ𝑚 , 𝑟, 1), reads
∑𝑟
1 𝑖
𝚃0 (ℎ|Γ1 , 𝑟, 1)(𝑥) = ℎ(0) + ℎ [𝑤, … , 𝑤]𝑥𝑖 , (7.1)
𝑖=1
𝑖!
where the multi-linear notation in (2.2) has been taken into account, and where 𝑥 is the
coordinate associated to the vector 𝑤.
By (7.1) and Lemma 5.1, condition (3.15) amounts to asking that, for any sub-
space Γ1 , the polynomial 𝚃0 (ℎ|Γ1 , 𝑟, 1) belongs to the complementary of the set of 𝑠1 -
vanishing polynomials 𝜎(𝑟, 𝑠1 , 1) in (𝑟, 1). Moreover, again by Lemma 5.1, one has
𝜎(𝑟, 𝑠1 , 1) = 𝜎(𝑟, 𝑠1 , 1) =∶ Σ(𝑟, 𝑠1 , 1). Hence, by definitions (6.1)-(6.2), by Theorem A
and by (6.66), ℎ is steep on the subspaces of dimension one in a neighborhood of the
origin, with steepness index bounded by 𝑠1 .

7.1.2 Case 𝑛 ≥ 3, 2 ≤ 𝑚 ≤ 𝑛 − 1.
It is sufficient to prove that, for fixed 𝑚 ∈ {1, … , 𝑛 − 1}, under the assumptions at
point 𝑖𝑖) of Theorem B, one has 𝚃0 (ℎ, 𝑟, 𝑛) ∈ (𝑟, 𝑛)∖(𝑟, 𝑠𝑚 , 𝑚, 𝑛), where the set
(𝑟, 𝑠𝑚 , 𝑚, 𝑛) was defined in (6.2). The thesis then follows by Theorem A and by ex-
pression (6.66).
By absurd, suppose that the claim is false. Then, by (6.1)-(6.2), there exists some
subspace Γ𝑚 orthogonal to ∇ℎ(0) ≠ 0 such that 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, 𝑠𝑚 , 𝑚).
Hence, since Σ(𝑟, 𝑠𝑚 , 𝑚) ∶= 𝜎(𝑟, 𝑠𝑚 , 𝑚) by construction, there are two possibilities:

75
 1. either 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝜎(𝑟, 𝑠𝑚 , 𝑚) ;

2. or 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, 𝑠𝑚 , 𝑚)∖𝜎(𝑟, 𝑠𝑚 , 𝑚) .

We consider the two cases separately and we prove that in both cases we end up being
in contradiction with the hypotheses.
Case 1. If 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝜎(𝑟, 𝑠𝑚 , 𝑚), then by construction 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) verifies
the 𝑠𝑚 vanishing condition at the origin on some curve 𝛾 ∈ Θ𝑚 , whose image is con-
tained in Γ𝑚 . Since one is free to choose the orthonormal basis {𝑢1 , … , 𝑢𝑚 } ∈ 𝚄(𝑚, 𝑛)
spanning Γ𝑚 , up to a changement in the order of the vectors we can suppose without
any loss of generality that the coordinate which parametrizes the curve 𝛾 is the first one,
that is 𝛾 ∈ Θ1𝑚 , and 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝜎 1 (𝑟, 𝑠𝑚 , 𝑚). Moreover, following section 5.2.3,
we can make use of the adapted coordinates for the curve 𝛾, which are associated to the
basis (see expression (5.17))
{ }
∑𝑚
𝑣 ∶= 𝑢1 + 𝑎𝑖1 𝑢𝑖 , 𝑢2 , … , 𝑢𝑚 ∈ 𝒱 1 (𝑚, 𝑛) , (7.2)
𝑖=2

where, as we did in 5.2.3, we indicate by 𝚊 ∶= (𝑎21 , … , 𝑎1𝑚 ) ∈ ℝ𝑚−1 the vector contain-
ing the linear coefficients of the Taylor expansion of 𝛾 at the origin. Hence, following
the notations of section 5.2.3 (especially, formula (5.20)), we write

𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚)(𝚢) ∶= 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚)◦−1

𝚊 (𝚢) .

Then, by standard results of calculus, taking (2.2) into account, one can write
𝜇1 𝜇2 𝜇𝑚
∑ [ ⏞⏞⏞ ] 𝜇
1 |𝜇| ⏞⏞⏞ ⏞⏞⏞ 𝜇
𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚)(𝚢) = ℎ0 𝑣 , 𝑢2 , … , 𝑢𝑚 𝚢1 1 … 𝚢𝑚𝑚 , (7.3)
𝜇∈ℕ𝑚
𝜇!
1≤|𝜇|≤𝑟

where 𝜇! ∶= 𝜇1 ! … , 𝜇𝑚 !.
Since 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝜎 1 (𝑟, 𝑠𝑚 , 𝑚), by (5.14) and Theorem 5.1, one has

𝚀𝑖𝛼 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹𝑠𝑚 ,𝛾,𝚊 ) = 0 ∀ 𝑖 ∈ {1, … , 𝑚} , ∀𝛼 ∈ {0, … , 𝑠} , (7.4)

where 𝙹𝑠𝑚 ,𝛾,𝚊 ∶= 𝚊 ◦𝑠𝑚 ,𝛾 denotes the 𝑠𝑚 -truncation at the origin of the curve 𝛾 ex-
pressed in the adapted variables.
We now try to simplify the expressions in (7.4), by taking formulas (5.43)-(5.44) in
Theorem 5.1 into account and by exploiting the form of the polynomial 𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚)
in (7.3). Namely - due to (7.3) and (3.9) - the coefficients 𝚙𝜈(1,𝛼) and 𝚙𝜈(𝓁,𝛼) , with 𝓁 ∈
{2, … , 𝑚}, 𝛼 ∈ {0, … , 𝑠𝑚 }, appearing in (5.43)-(5.44) read
𝛼
[ ⏞⏞⏞ ]
1 1 (7.5)
𝚙𝜈(1,𝛼) = ℎ𝛼+1 [𝑣, … , 𝑣] , 𝚙𝜈(𝓁,𝛼) = ℎ𝛼+1 𝑣 , 𝑢𝓁 .
(𝛼 + 1)! 0 𝛼! 0

76
Moreover, if 𝑠𝑚 ≥ 2, for 𝛼 ∈ {2, … , 𝑠𝑚 }, exploiting (7.3) and the linearity, in our case
the second addend at the right hand side at the third line of (5.43) reads
𝛽
∑ ∑
𝑚 𝛼−1 ∑ ∑ 𝛽 𝛽+1 [ ⏞⏞⏞ ]
𝑚 𝛼−1
𝛽 𝚙𝜈(𝑖,𝛽) 𝑎𝑖(𝛼−(𝛽−1)) = ℎ 𝑣 , 𝑢𝑖 𝑎𝑖(𝛼−(𝛽−1))
𝑖=2 𝛽=1 𝑖=2 𝛽=1
𝛽! 0
(7.6)
𝛽
∑
𝛼−1 [ ⏞⏞⏞ ∑𝑚 ]
1
= ℎ𝛽+1
0
𝑣 , 𝑎𝑖(𝛼−(𝛽−1)) 𝑖 .
𝑢
𝛽=1
(𝛽 − 1)! 𝑖=2

Henceforth, in order to simplify our formulas, we set

{
𝑣 , if 𝑖 = 1
𝐕𝑖 ∶= .
𝑢𝑖 , if 𝑖 ∈ {2, … , 𝑚}

Considering again the case 𝑠𝑚 ≥ 2, for 𝛼 ∈ {2, … , 𝑠𝑚 }, and for any 𝑖 = 1, … , 𝑚, by

(7.3) and by (3.11)-(3.12), the last addend at the right hand side of the third relation in
(5.43)-(5.44) reads
( 𝑚 ( ) )
∑ ∑ ∏ 𝜇̃𝑗 (𝑖) 𝑘𝑗2 𝑘𝑗𝛼
𝜇𝑖 𝚙𝜇 𝑎𝑗2 ...𝑎𝑗𝛼
𝜇∈ (𝑖,𝛼)
𝑚 𝜇 (𝑖),𝛼) 𝑗=2 𝑘𝑗2 ... 𝑘𝑗𝛼
𝑘∈ (̃ 𝑚
𝜇∈𝑚 (𝛼)
𝜇𝑖 ≠0
𝜇̃1 (𝑖) 𝜇̃2 (𝑖) 𝜇̃𝑚 (𝑖)
∑ 𝜇𝑖 |𝜇| [ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ]
= ℎ0 𝑣 , 𝑢2 , … , 𝑢𝑚 , 𝐕𝑖
𝜇∈𝑚 (𝑖,𝛼)
𝜇!
𝜇∈𝑚 (𝛼)
𝜇𝑖 ≠0
( ( ) )
∑ ∏
𝑚
𝜇̃𝑗 (𝑖) 𝑘𝑗2 𝑘𝑗𝛼
× 𝑎𝑗2 ...𝑎𝑗𝛼
𝑘∈𝑚 (̃
𝜇(𝑖),𝛼) 𝑗=2 𝑘𝑗2 ...𝑘𝑗𝛼
𝜇̃1 (𝑖) 𝑘22 𝑘2𝛼 𝑘𝑚2 𝑘𝑚𝛼
[ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ⏞⏞⏞ ]
|𝜇|
∑ ∑ ℎ0 𝑣 , 𝑎22 𝑢2 , … , 𝑎2𝛼 𝑢2 , 𝑎𝑚2 𝑢𝑚 , … , 𝑎𝑚𝛼 𝑢𝑚 , 𝐕𝑖
= ,
𝜇∈𝑚 (𝑖,𝛼) 𝑘∈𝑚 (̃
𝜇(𝑖),𝛼)
𝜇̃1 (𝑖)! 𝑘!
𝜇∈𝑚 (𝛼)
𝜇𝑖 ≠0
(7.7)

where the last passage is a consequence of the multi-linearity and of the fact that, for all
∑
𝑖 ∈ {1, … , 𝑚}, 𝑗 ∈ {2, … , 𝑚}, we have 𝛼𝑢=2 𝑘𝑗𝑢 = 𝜇̃𝑗 (𝑖) by construction (see (3.11)).
Now, by (5.43)-(5.44), it is trivial to observe that for 𝛼 = 0 and for all 𝑖 ∈ {1, … , 𝑚},
one has

𝚀𝑖,0 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹𝑠𝑚 ,𝛾,𝚊 ) = 0 ⟺ ℎ10 [𝑣] = ℎ10 [𝑢2 ] = … ℎ10 [𝑢𝑚 ] = 0 (7.8)

77
which simply means that the basis vectors {𝑣, 𝑢2 , … , 𝑢𝓁 } are orthogonal to ∇ℎ(0) ≠ 0.
We observe that, in case 𝑖 = 1, … , 𝑚 and 𝛼 = 1, taking (7.5) and (5.43)-(5.44) into
account, condition (7.4) amounts to requiring that for all 𝑖 ∈ {1, … , 𝑚}

𝚀𝑖,1 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹𝑠𝑚 ,𝛾,𝚊 ) = 0 ⟺ ℎ20 [𝑣, 𝑣] = ℎ20 [𝑣, 𝑢2 ] = ⋯ = ℎ20 [𝑣, 𝑢𝑚 ] = 0 .
(7.9)
For 𝑖 = 1 and 𝛼 = 2, instead, we have
1 3
𝚙𝜈(1,𝛼) = ℎ [𝑣, 𝑣, 𝑣] (7.10)
3!
and for 𝛽 = 1 the term in (7.6) is null due to (7.9).
Moreover, still for 𝑖 = 1 and 𝛼 = 2 we observe that also the term in (7.7) does not
yield any contribution to condition (7.4). In order to see this, we start by remarking
that the multi-indices to be taken into account in (7.7) for 𝑖 = 2 and 𝛼 = 2 must verify
𝜇 ∈ 𝑚 (1, 2), that is, by (3.11)-(3.12)

∑
𝑚
𝑘𝑗2 = 𝜇̃𝑗 (1) ∀ 𝑗 ∈ {2, … , 𝑚} , 𝜇̃1 (1) + 2𝑘𝑗2 = 2 . (7.11)
𝑗=2

Conditions (7.11) are only possible if

1. 𝜇̃1 (1) = 2 and 𝜇̃𝑗 (1) = 0 for all 𝑗 ∈ {2, … , 𝑚}, which implies 𝜇 = (3, 0, … , 0).

2. 𝑘𝑗2 = 𝜇̃𝑗 (1) = 𝛿𝑗𝑝 for some 𝑝 ∈ {2, … , 𝑚} and 𝜇̃1 (1) = 0. This implies that
𝜇1 = 1 and 𝜇𝑗 = 𝛿𝑗𝑝 for all 𝑗 ∈ {2, … , 𝑚}, so that finally |𝜇| = 2.

The first case is incompatible with the condition 𝜇 ∈ 𝑚 (2) required in (7.7) (see
(3.10)), as (3, 0, … , 0) ≡ 𝜈(1, 2). The second case does not yield any contribution to
(7.7) because of (7.9).
Hence one has

𝚀1,2 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹𝑠𝑚 ,𝛾,𝚊 ) = 0 ⟺ ℎ3 [𝑣, 𝑣, 𝑣] = 0 . (7.12)

Finally, for 𝑖 = 1 and 𝛼 ≥ 3, and for 𝑖 ∈ {2, … , 𝑚} and 𝛼 ≥ 2, comparing expres-

sions (7.5)-(7.6)-(7.7) with the quantities (5.43)-(5.44) in Theorem 5.1, and taking the
ℎ,0
definition of the quantities 𝑚,𝓁,𝛼 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚)) in (3.13)-(3.14) into account, one
has that
ℎ,0
𝚀𝑖,𝛼 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹𝑠𝑚 ,𝛾,𝚊 ) = 0 ⟺𝑚,𝓁,𝛼 (𝑣, 𝑢2 , … , 𝑢𝑚 , 𝑎(𝑚, 𝑠𝑚 )) = 0 .
(7.13)
Putting together (7.8)-(7.9)-(7.12)-(7.13) with (7.2), we see that the polynomial 𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚)

78
belongs to 𝜎 1 (𝑟, 𝑠𝑚 , 𝑚) if and only if the system

⎧(𝑢 , … , 𝑢 ) ∈ 𝚄(𝑚, 𝑛)
⎪ 1 𝑚
⎪𝑎(𝑚) ∶= (𝑎 , … , 𝑎 , … , 𝑎 , … , 𝑎 ) ∈ ℝ(𝑚−1)×𝑠𝑚
⎪ 21 2𝑠𝑚 𝑚1 𝑚𝑠𝑚
⎪ ∑𝑚
⎨𝑣 = 𝑢1 + 𝑗=2 𝑎𝑗1 𝑢𝑗 (7.14)
⎪
⎪ℎ1 [𝑣] = ℎ1 [𝑢2 ] = ⋯ = ℎ1 [𝑢𝑚 ] = 0
⎪ 0 0 0
⎪ ℎ,0 (𝑣, 𝑢 , … , 𝑢 , 𝑎(𝑚, 𝑠 )) = 0 𝑖 = 1, … , 𝑚 , 𝛼 = 1, … , 𝑠
⎩ 𝑚,𝓁,𝛼 2 𝑚 𝑚

admits a solution. However, this is in contradiction with hypothesis (3.16) in the state-
ment, therefore 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∉ 𝜎 1 (𝑟, 𝑠𝑚 , 𝑚).
Case 2. We now assume that 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, 𝑠𝑚 , 𝑚)∖𝜎(𝑟, 𝑠𝑚 , 𝑚). Up to chang-
ing the order of the vectors spanning Γ𝑚 , by (5.68) without any loss of generality we can
suppose 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ1 (𝑟, 𝑠𝑚 , 𝑚)∖𝜎(𝑟, 𝑠𝑚 , 𝑚). Then, there must exist a sequence of
polynomials {𝑃𝑘 ∈ 𝜎 1 (𝑟, 𝑠𝑚 , 𝑚)}𝑘∈ℕ approaching 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚). To conclude the proof
of Case 2, we need the following

Lemma 7.1. There exists a sequence {𝑆𝑘 ∈ (𝑟, 𝑛)}𝑘∈ℕ converging to 𝚃0 (ℎ, 𝑟, 𝑛) in
(𝑟, 𝑛) and verifying 𝑆𝑘 |Γ𝑚 = 𝑃𝑘 for any given 𝑘 ∈ ℕ.

Proof. We indicate by 𝐴1 , … , 𝐴𝑚 ∈ 𝚄(𝑚, 𝑛) an orthonormal basis of Γ𝑚 , and we choose

𝑛 − 𝑚 orthonormal supplementary vectors 𝐴𝑚+1 , … , 𝐴𝑛 to form a orthonormal basis
of ℝ𝑛 . In the coordinates (𝑥1 , … , 𝑥𝑛 ) associated to 𝐴1 , … , 𝐴𝑛 , the restriction of any
polynomial 𝑄 ∈ (𝑟, 𝑛) to the subspace Γ𝑚 is obtained by simply setting 𝑥𝑚+1 = ⋯ =
𝑥𝑛 = 0 in the expression of 𝑄. Conversely, any polynomial 𝑃 (𝑥1 , … , 𝑥𝑚 ) ∈ (𝑟, 𝑚)
depending only on the first 𝑚 variables is the projection on Γ𝑚 of any polynomial 𝑄 ∈
(𝑟, 𝑛) of the form 𝑄(𝑥1 , … , 𝑥𝑛 ) = 𝑃 (𝑥1 , … , 𝑥𝑚 ) + 𝑞(𝑥1 , … , 𝑥𝑛 ), where 𝑞 ∈ (𝑟, 𝑛)
verifies 𝑞(𝑥1 , … , 𝑥𝑚 , 0) = 0.
By the above discussion and with slight abuse of notation, one can define the poly-
nomial

𝑞ℎ = 𝑞ℎ (𝑥1 , … , 𝑥𝑛 ) ∶= 𝚃0 (ℎ, 𝑟, 𝑛)(𝑥1 , … , 𝑥𝑛 ) − 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚)(𝑥1 , … , 𝑥𝑚 ) (7.15)

which verifies 𝑞ℎ (𝑥1 , … , 𝑥𝑚 , 0) = 0 by construction.

Then, for 𝑘 ∈ ℕ, we consider the polynomials

𝑆𝑘 = 𝑆𝑘 (𝑥1 , … , 𝑥𝑛 ) ∶= 𝑃𝑘 (𝑥1 , … , 𝑥𝑚 ) + 𝑞ℎ (𝑥1 , … , 𝑥𝑛 ) , (7.16)

where {𝑃𝑘 }𝑘∈ℕ is the sequence approaching 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) introduced above. The se-
quence {𝑆𝑘 }𝑘∈ℕ has the properties we seek. Infact, as 𝑞ℎ (𝑥1 , … , 𝑥𝑚 , 0) = 0, on the one
hand 𝑆𝑘 verifies

𝑆𝑘 |Γ𝑚 = 𝑆𝑘 (𝑥1 , … , 𝑥𝑚 , 0) = 𝑃𝑘 ∀𝑘∈ℕ; (7.17)

79
on the other hand, as 𝑃𝑘 ⟶ 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) by hypothesis, by taking (7.15)-(7.16) into
account one has
𝑆𝑘 ⟶ 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) + 𝑞ℎ = 𝚃0 (ℎ, 𝑟, 𝑛) . (7.18)

Since by Lemma 7.1 one has 𝑆𝑘 |Γ𝑚 = 𝑃𝑘 , and since 𝑃𝑘 ∈ 𝜎 1 (𝑟, 𝑠𝑚 , 𝑚) by construc-
tion, the same arguments developed at Case 1 yield that for any 𝑘 ∈ ℕ the system

⎧(𝑢 , … , 𝑢 ) ∈ 𝚄(𝑚, 𝑛)
⎪ 1 𝑚
⎪𝑎(𝑚) ∶= (𝑎 , … , 𝑎 , … , 𝑎 , … , 𝑎 ) ∈ ℝ(𝑚−1)×𝑠𝑚
⎪ 21 2𝑠𝑚 𝑚1 𝑚𝑠𝑚
⎪ ∑𝑚
⎨𝑣 = 𝑢1 + 𝑗=2 𝑎𝑗1 𝑢𝑗 (7.19)
⎪
⎪(𝑆𝑘 )1 [𝑣] = (𝑆𝑘 )1 [𝑢2 ] = ⋯ = (𝑆𝑘 )1 [𝑢𝑚 ] = 0
0 0 0
⎪
⎪ 𝑆𝑘 ,0 (𝑣, 𝑢 , … , 𝑢 , 𝑎(𝑚, 𝑠 )) = 0 𝑖 = 1, … , 𝑚 , 𝛼 = 1, … , 𝑠
⎩ 𝑚,𝓁,𝛼 2 𝑚 𝑚

must be verified. However, this fact and the fact that, by Lemma 7.1, one also has
𝑆𝑘 ⟶ 𝚃0 (ℎ, 𝑟, 𝑛), contradicts the hypotheses of Theorem B. Hence, we must have
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∉ Σ(𝑟, 𝑠𝑚 , 𝑚)∖𝜎(𝑟, 𝑠𝑚 , 𝑚).
By the discussion at Cases 1-2 above, the assumptions of Theorem B imply that -
for any 𝑚-dimensional subspace Γ𝑚 , with 𝑚 ∈ {2, … , 𝑛 − 1} - the Taylor polynomial
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) lies outside of Σ(𝑟, 𝑠𝑚 , 𝑚). Hence 𝚃0 (ℎ, 𝑟, 𝑛) ∈ (𝑟, 𝑛)∖(𝑟, 𝑠𝑚 , 𝑚, 𝑛).
This, together with (6.66) and Theorem A, concludes the proof.

7.2 Proof of the Corollaries

7.2.1 Proof of Corollary B1
We start by studying the one-dimensional affine subspaces orthogonal to ∇ℎ(𝐼0 ) ≠ 0.
Hypothesis (3.18) is equivalent to hypothesis (3.15) in Theorem B with 𝑟 = 3, 𝑠1 = 2,
whence the thesis.
On the other hand, for any fixed 𝑚 ∈ {2, … , 𝑛 − 1}, since 𝙶(𝑚, 𝑛) and 𝕊𝑛 are both
compact, by hypothesis (3.18) there exists 𝜏𝑚 > 0 such that for any 𝑚-dimensional
affine subspace 𝐼0 + Γ𝑚 orthogonal to ∇ℎ(𝐼0 ), and for any vector 𝑤 ∈ 𝕊𝑛 ∩ Γ𝑚 we have

ℎ1𝐼 [𝑤] = 0 , |ℎ2𝐼 [𝑤, 𝑤]| + |ℎ3𝐼 [𝑤, 𝑤, 𝑤]| ≥ 𝜏𝑚 > 0 . (7.20)
0 0 0

By continuity, up to dividing the constant 𝜏𝑚 by a factor two, relation (7.20) holds

uniformly also for all polynomials in a small ball around 𝚃0 (ℎ, 3, 𝑛). Hence, ℎ matches
the hypotheses at point 𝑖𝑖) of Theorem B for 𝑟 = 3 and 𝑠𝑚 = 2, whence the thesis for
affine subspaces of dimension higher or equal than two.

80
7.2.2 Proof of Corollary B2
⋃𝑛−1
In the proof of Theorem A we have set 𝛀𝑟,𝚜
𝑛 ∶= 𝑚=1 (𝑟, 𝑠𝑚 , 𝑚, 𝑛) (see (6.66)). Fur-
thermore, formula (6.2) ensures that for any given function ℎ of class 𝐶 2𝑟−1 around
𝐼0
by definition ∃Γ𝑚 ∈𝙶(𝑚,𝑛), Γ𝑚 ⟂∇ℎ(𝐼0 ) s.t.
𝚃𝐼0 (ℎ, 𝑟, 𝑛) ∈ (𝑟, 𝑠𝑚 , 𝑚, 𝑛) ⟺ 𝚃𝐼 (ℎ|Γ𝑚 ,𝑟,𝑚)∈Σ(𝑟,𝑠𝑚 ,𝑚)∶=𝜎(𝑟,𝑠𝑚 ,𝑚) . (7.21)
0

Since one is free to choose the order of the orthonormal vectors spanning Γ𝑚 , without
any loss of generality we can also write
∃Γ𝑚 ∈𝙶(𝑚,𝑛), Γ𝑚 ⟂∇ℎ(𝐼0 ) s.t.
𝚃𝐼0 (ℎ, 𝑟, 𝑛) ∈ (𝑟, 𝑠𝑚 , 𝑚, 𝑛) ⟺ 𝚃𝐼 (ℎ|Γ𝑚 ,𝑟,𝑚)∈Σ1 (𝑟,𝑠𝑚 ,𝑚)∶=𝜎 1 (𝑟,𝑠𝑚 ,𝑚) . (7.22)
0

In the proof of Theorem B, we have also seen that for any 𝑚 ∈ {1, … , 𝑛 − 1}, and
for any given subspace Γ𝑚 ∈ 𝙶(𝑚, 𝑛) orthogonal to ∇ℎ(𝐼0 ), condition 𝚃𝐼0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈
𝜎 1 (𝑟, 𝑠𝑚 , 𝑚) holds if and only if system (3.19) (if 𝑚 = 1) or (3.20) (if 𝑚 ≥ 2) admits a
solution when 𝑄 is set to be the Taylor polynomial at order 𝑟 of function ℎ.
Hence we have that, for any fixed 𝑚 ∈ {1, … , 𝑛 − 1}, condition 𝚃𝐼0 (ℎ, 𝑟, 𝑛) ∈
(𝑟, 𝑠𝑚 , 𝑚, 𝑛) is equivalent to asking that 𝚃𝐼0 (ℎ, 𝑟, 𝑛) belongs to the closure in  ⋆ (𝑟, 𝑛)
of the set of polynomials solving system (3.19) (if 𝑚 = 1) or system (3.20) (for 𝑚 ∈
{2, … , 𝑛 − 1}).
The thesis follows by the arguments above and by (6.66).

7.2.3 Proof of Corollary B3

By the proof of Theorem B, the hypotheses in 𝑖) and 𝑖𝑖) amount to asking for the exis-
tence of a real-analytic curve
{
𝑤𝑡 for 𝑚 = 1
𝛾(𝑡) ∶= ( ∑+∞ ∑ +∞ ) (7.23)
𝑡, 𝑖=1 𝑎2𝑖 𝑡𝑖 , … , 𝑖=1 𝑎𝑚𝑖 𝑡𝑖 for 𝑚 = {2, … , 𝑛 − 1}

whose image is contained in some 𝑚-dimensional subspace orthogonal to ∇ℎ(𝐼0 ) ≠ 0,

and such that (𝜋Γ𝑚 ∇ℎ)◦𝛾 has a zero of infinite order at 𝛾(0) = 𝐼0 . By analyticity, then,
(𝜋Γ𝑚 ∇ℎ)◦𝛾 is identically zero. This implies by Definition 1.1 that ℎ is not steep.

8 Partition of the set of 𝑠-vanishing polynomials

The proof of Theorems C1-C2-C3 is quite long and requires intermediate results which
will be presented in this section. Before stating them, in the following two paragraphs
we introduce some definitions and notations.

81
8.1 Initial setting
Consider three integers 𝑟, 𝑚 ≥ 2, and 𝑠 ∈ {2, … , 𝑟 − 1}. In sections 5-6, we have indi-
cated by Σ(𝑟, 𝑠, 𝑚) ⊂ (𝑟, 𝑚) the closure of the set 𝜎(𝑟, 𝑠, 𝑚) of 𝑠-vanishing polynomials.
In particular, by (5.13)-(5.14) one has

⋃
𝑚
𝜎(𝑟, 𝑠, 𝑚) = 𝜎 𝑖 (𝑟, 𝑠, 𝑚) , 𝜎 𝑖 (𝑟, 𝑠, 𝑚) ∶= Π(𝑟,𝑚) 𝑍 𝑖 (𝑟, 𝑠, 𝑚) ,
𝑖=1
(8.1)
⋃
𝑛 ⋃
𝑛
𝑖
⋃
𝑛
𝑖
Σ(𝑟, 𝑠, 𝑚) = Σ (𝑟, 𝑠, 𝑚) ∶= 𝜎 (𝑟, 𝑠, 𝑚) ∶= Π(𝑟,𝑚) 𝑍 𝑖 (𝑟, 𝑠, 𝑚) ,
𝑖=1 𝑖=1 𝑖=1

where the sets 𝑍 𝑖 (𝑟, 𝑠, 𝑚) ⊂ (𝑟, 𝑚) × 𝜗𝑖𝑚 (𝑠), with 𝑖 ∈ {1, … , 𝑚}, are defined in (5.15),
and one has decomposition (5.16), namely

⋃
𝑚
(𝑟, 𝑚) × 𝜗𝑚 (𝑠) ⊃ 𝑍(𝑟, 𝑠, 𝑚) =∶ 𝑍 𝑖 (𝑟, 𝑠, 𝑚) .
𝑖=1

The expression of the sets 𝑍 𝑖 (𝑟, 𝑠, 𝑚), 𝑖 ∈ {1, … , 𝑚}, is given explicitly 15 in Theorem
5.1.
In the previous sections, we have seen that, in order to check if a given polynomial
𝑄 ∈  ⋆ (𝑟, 𝑛) is steep at the origin16 on a fixed subspace Γ𝑚 orthogonal to ∇𝑄(0) ≠ 0,
one must check whether the restriction 𝑃 ∶= 𝑄|Γ𝑚 ∈ (𝑟, 𝑚) belongs to the comple-
mentary of Σ(𝑟, 𝑠, 𝑚) ∶= 𝜎(𝑟, 𝑠, 𝑚). We now claim that it is not strictly necessary to
consider the closure of the whole set 𝜎(𝑟, 𝑠, 𝑚). Indeed, in practice, the curves on which
the 𝑠-vanishing condition must be tested are minimal arcs with uniform characteristics,
like the one defined in Theorem 4.1.
Namely, by the arguments in the proof of Lemma 6.5 - for any given 𝑄 ∈  ⋆ (𝑟, 𝑛)
and for any fixed Γ𝑚 perpendicular to ∇𝑄(0) ≠ 0, it is sufficient to check if there exists
a threshold 𝜆0 > 0 such that, for any 0 < 𝜆 ≤ 𝜆0 , there exists an interval 𝐼𝜆 ⊂ [−𝜆, 𝜆]
of length 𝜆∕𝙺 17 , on which - for any curve 𝛾 ∈ Θ𝑚 verifying the Bernstein’s inequality
(4.1) - the composition (∇𝑃 )◦𝛾 ∶= (∇𝑄|Γ𝑚 )◦𝛾 has no zeros of order greater or equal
than 𝑠. In particular, we are interested in testing the 𝑠-vanishing conditions on those
15 Actually, in Theorem 5.1, only the expression of 𝑍 1 (𝑟, 𝑠, 𝑚) is explicit. However, as it was already
pointed out in section 5, the cases 𝑖 = 2, … , 𝑚 are trivial generalizations of the case 𝑖 = 1: in order to find
the expression of 𝑍 𝑖 (𝑟, 𝑠, 𝑚) for 𝑖 ≠ 1, one simply has to follow the same steps needed to find the expression
for 𝑍 1 (𝑟, 𝑠, 𝑚), and to exchange the rôle of the first coordinate with the 𝑖-th one.
16 It is clear that the arguments devoped in the sequel hold also if the considered point is not the origin.
17 More details about the threshold 𝜆 are given in Lemma 6.5, whereas the constant 𝙺 is the one introduced
0
in Theorem 4.1.

82
analytic curves 𝛾 over 𝐼𝜆 that, for some 𝑖 ∈ {1, … , 𝑚}, verify
∑
⎛ 𝑥1 (𝑡) = +∞ 𝑘=1 𝑎1𝑘 𝑡
𝑘 ⎞
⎜ … ⎟
⎜ ∑+∞ 𝑘
⎟
⎜𝑥𝑖−1 (𝑡) = 𝑘=1 𝑎(𝑖−1)𝑘 𝑡 ⎟ 𝙼(𝑟, 𝑛, 𝑚, 𝑘)
𝛾(𝑡) = ⎜ 𝑥𝑖 (𝑡) = 𝑡 ⎟ , max max |𝑎𝑗𝑘 (𝑢)| ≤ .
⎜ ∑+∞ 𝑘⎟ 𝑢∈𝐼𝜆 𝑗∈{1,…,𝑚} 𝜆𝑘−1
⎜ 𝑖+1
𝑥 (𝑡) = 𝑘=1 (𝑖+1)𝑘 ⎟
𝑎 𝑡 𝑗≠𝑖
⎜ … ⎟
⎜ ∑+∞ ⎟
⎝ 𝑥𝑚 (𝑡) = 𝑘=1 𝑎𝑚𝑘 𝑡𝑘 ⎠
(8.2)
By Theorem 4.1, the constant 𝙼 = 𝙼(𝑟, 𝑛, 𝑚, 𝑘) in (8.2) can be explicitly computed
and it is uniform for all curves 𝛾 ∈ Θ𝑚 .
As we have shown in the proof of Lemma 6.5, for any given 𝑃 ∈ (𝑟, 𝑚), the
threshold 𝜆0 - if it exists - goes to zero with the distance of 𝑃 to the "bad" set 𝜎(𝑟, 𝑠, 𝑚).
Therefore, by formula (8.2), the Taylor coefficients of the curves 𝛾 on which the 𝑠-
vanishing condition must be tested may take any value, except for those of order one
which, independently from the choice of 𝜆 > 0, are always uniformly bounded by
𝙼(𝑟, 𝑛, 𝑚, 1).
Inspired by the above reasonings, with the notations in (8.2), we give the following
Definition 8.1. For 𝑖 = 1, … , 𝑚, we introduce the sets
⎧ ⎫
̂ ⎪ ⎪
Θ𝑚 ∶= ⎨𝛾 ∈ Θ𝑚 | max {|𝑎𝑗1 (0)|} ≤ 𝙼(𝑟, 𝑛, 𝑚, 1)⎬
𝑖 𝑖
(8.3)
𝑗∈{1,…,𝑚}
⎪ ⎪
⎩ 𝑗≠𝑖 ⎭
and
⋃
𝑚
̂ 𝑚 ∶=
Θ ̂𝑖 ,
Θ (8.4)
𝑚
𝑖=1

and we denote respectively by 𝜗̂𝑖𝑚 (𝑠) ⊂ 𝜗𝑖𝑚 (𝑠) and by 𝜗̂𝑚 (𝑠) ⊂ 𝜗𝑚 (𝑠) their associated
subsets of 𝑠-truncations.
Remark 8.1. By formula (8.3), if one introduces the decomposition

𝜗̂𝑖𝑚 (𝑠) = 𝜗̂𝑖𝑚 (1) × 𝜗̂𝑖𝑚 (𝑠, 2) (8.5)

as in (5.24), the space 𝜗̂𝑖𝑚 (1) is compact and 𝜗̂𝑖𝑚 (𝑠, 2) ≡ 𝜗𝑖𝑚 (𝑠, 2).
Definition 8.2. In section 5.2.3 (see formula (5.27)), we defined the set 𝑊 1 (𝑟, 𝑚) as

𝑊 1 (𝑟, 𝑚) ∶= ℱ 1 ((𝑟, 𝑚) × ℝ𝑚−1 ) , (8.6)

where the function ℱ 1 was introduced in (5.25)-(5.26). Similarly, for any 𝑗 ∈ {2, … , 𝑚},
taking Remark 5.8 into account, we had set

𝑊 𝑗 (𝑟, 𝑚) ∶= ℱ 𝑗 ((𝑟, 𝑚) × ℝ𝑚−1 ) . (8.7)

83
 Now, by (8.3) and by the above discussion, for any 𝑖 ∈ {1, … , 𝑚} it makes sense to
define also

̂ 𝑖 (𝑟, 𝑚) ∶= ℱ 𝑖 ((𝑟, 𝑚) × 𝐵 𝑚−1 (0, 𝙼(𝑟, 𝑛, 𝑚, 1))) .

𝑊 (8.8)

We remind that, due to Definition 8.1 and to Remark 5.7, for any given 𝑖 ∈ {1, … , 𝑚}
there exists a polynomial bijection 𝔘𝑖 between (𝑟, 𝑚) × 𝜗̂𝑖𝑚 (1) and 𝑊 ̂ 𝑖 (𝑟, 𝑚): one is
free to work either in the standard coordinates (𝑝𝜇 , 𝚊) ∈ ℝ𝑀 × ℝ(𝑚−1)𝑠 , with 𝑀 ∶=
dim (𝑟, 𝑚), or in the adapted coordinates (𝚙𝜇 , 𝚊) ∈ 𝑊 ̂ 𝑖 (𝑟, 𝑚).
By the arguments above, without any loss of generality, for any fixed 𝑖 ∈ {1, … , 𝑚}
it is sufficient to consider the set of those polynomials 𝑃 ∈ (𝑟, 𝑚) verifying the 𝑠-
vanishing condition on the 𝑠-jet 𝑠,𝛾 ∈ 𝜗̂𝑖𝑚 (𝑠) of some curve 𝛾 ∈ Θ̂ 𝑖 . Namely, following
𝑚
(5.15) and (8.1), for 𝑖 = 1, … , 𝑚 we introduce the semi-algebraic sets

̂ 𝑖 (𝑟, 𝑠, 𝑚) ∶={(𝑃 , 𝑠,𝛾 ) ∈ (𝑟, 𝑚) × 𝜗̂𝑖 (𝑠)}|(𝑃 , 𝑠,𝛾 ) satisfies

𝑍 𝑚
𝑖
𝑞𝓁𝛼 ◦Φ𝑖 (𝑃 , 𝑠,𝛾 ) = 0 for all 𝓁 ∈ {1, ..., 𝑚}, 𝛼 ∈ {0, ..., 𝑠}}
⋃
𝑚
̂ 𝑠, 𝑚) ∶=
𝑍(𝑟, ̂ 𝑖 (𝑟, 𝑠, 𝑚)
𝑍
𝑖=1
(8.9)
⋃
𝑚
𝑖 ̂ 𝑖 (𝑟, 𝑠, 𝑚) , 𝑖
̂ (𝑟, 𝑠, 𝑚) ∶=Π(𝑟,𝑚) 𝑍
𝜎 ̂(𝑟, 𝑠, 𝑚) =
𝜎 ̂ (𝑟, 𝑠, 𝑚)
𝜎
𝑖=1
⋃
𝑛 ⋃
𝑛
( 𝑖 )
̂ 𝑠, 𝑚) ∶=
Σ(𝑟, ̂𝑖 (𝑟, 𝑠, 𝑚) ∶=
Σ ̂ (𝑟, 𝑠, 𝑚)
closure 𝜎
𝑖=1 𝑖=1

and, of course, we have

̂ 𝑖 (𝑟, 𝑠, 𝑚) ⊂ 𝑍 𝑖 (𝑟, 𝑠, 𝑚) ,
𝑍 ̂𝑖 (𝑟, 𝑠, 𝑚) ⊂ 𝜎 𝑖 (𝑟, 𝑠, 𝑚)
𝜎 ∀ 𝑖 ∈ {1, … , 𝑚}. (8.10)

For further convenience, we also state the following simple

̂(𝑟, 𝑠, 𝑚) has codimension 𝑠 + 𝑚 = codim 𝜎(𝑟, 𝑠, 𝑚) in (𝑟, 𝑚).

Lemma 8.1. 𝜎

Proof. By the third line of (8.9) and by the second inclusion in (8.10), it suffices to
prove the statement for 𝜎 ̂1 (𝑟, 𝑠, 𝑚). By the first line of (8.9) and by Definition 8.1, one
has
⋂{ }
̂ 1 (𝑟, 𝑠, 𝑚) = 𝑍 1 (𝑟, 𝑠, 𝑚)
𝑍 (𝑃 , 𝑠,𝛾 ) ∈ (𝑟, 𝑚) × 𝜗̂1𝑚 (𝑠) . (8.11)

As it was shown in Corollary 5.1, the Jacobian associated to the equalities deter-
mining 𝑍 1 (𝑟, 𝑠, 𝑚) has full rank 𝑚𝑠 + 𝑚. Namely, by the discussion below expression
(5.64), in the adapted coordinates of section 5.2.3, such a Jacobian has non-zero pivots
corresponding to the derivatives w.r.t. the coefficients of the polynomial 𝙿𝚊 associated
to multi-indices in the family (5.39). As we had shown in the proof of Corollary 5.1,

84
this fact and the Implicit Function Theorem imply that for any pair (𝑃 , 𝑠,𝛾 ) belong-
ing to 𝑍 1 (𝑟, 𝑠, 𝑚), one can express 𝑚𝑠 + 𝑚 coefficients of 𝑃 as implicit functions of
the other coefficients of 𝑃 and of the (𝑚 − 1)𝑠 parameters of 𝑠,𝛾 . This was the ar-
gument that led to estimate codim 𝜎 1 (𝑟, 𝑠, 𝑚) = 𝑠 + 𝑚 in Corollary 5.1. The thesis
follows by putting this argument together with formulas (8.9)-(8.11) and with the fact
that dim 𝜗1𝑚 (𝑠) = dim 𝜗̂1𝑚 (𝑠).

8.2 ̂ 1 (𝑟, 𝑚)
Partition of (𝑟, 𝑚) and 𝑊
Let 𝑟, 𝑚 ≥ 2 be two integers. In this paragraph, we introduce a partition of the spaces
(𝑟, 𝑚) and 𝑊̂ 1 (𝑟, 𝑚) which will turn out to be useful in the sequel. In order to do this,
we first need to introduce a family of multi-indices.

8.2.1 A family of multi-indices

For 𝑏, 𝑐 ∈ {2, … , 𝑚}, 𝑏 ≤ 𝑐 18 , we set

𝜛(𝑏, 𝑐) ∶= 𝜇 = (𝜇1 , … , 𝜇𝑚 ) ∈ ℕ𝑚 | 𝜇1 = 0 , 𝜇𝑗 = 𝛿𝑗𝑏 + 𝛿𝑗𝑐 ∀𝑗 ∈ {2, … , 𝑚} .

(8.12)
Comparing (8.12) with the sub-family 𝜈(𝚒, 1) defined in (5.39), it is plain to check that
one has the disjoint union

⎛ 𝑚 ⎞ (𝑚 )
⎜⋃ ⎟⨆ ⋃
⎜ {𝜛(𝑏, 𝑐)}⎟ {𝜈(𝚒, 1)} = {𝜇 ∈ ℕ𝑚 | |𝜇| = 2} . (8.13)
⎜𝑏,𝑐=2 ⎟ 𝚒=1
⎝ 𝑏≤𝑐 ⎠

Moreover, we have the following

Lemma 8.2. For any polynomial 𝑃 ∈ (𝑟, 𝑚), the coefficients 𝑝𝜇 associated to the
multi-indices 𝜇 belonging to the family (8.12) are invariant under the transformations
of paragraph 5.2.3. Namely, using the notations in (5.20), for any given 𝚊 ∈ 𝜗̂1𝑚 (1) one
has
𝚙𝜛(𝑏,𝑐) = 𝑝𝜛(𝑏,𝑐) for all 𝑏, 𝑐 ∈ {2, … , 𝑚} , 𝑏≤𝑐.

Proof. We indicate by 𝐴1 , … , 𝐴𝑚 the orthonormal basis of ℝ𝑚 associated to the coor-

dinates 𝑥1 , … , 𝑥𝑚 on which polynomials in (𝑟, 𝑚) depend.
For any 𝚊 ∶= (𝑎21 , … , 𝑎𝑚1 ) ∈ 𝜗̂1𝑚 (1), we also denote by

𝑣𝚊 ∶= 𝐴1 + 𝑎21 𝐴2 + 𝑎31 𝐴3 + ... + 𝑎𝑚1 𝐴𝑚 , 𝑢2 ∶= 𝐴2 , ... , 𝑢𝑚 ∶= 𝐴𝑚 , (8.14)

18 We have set 𝑏 ≤ 𝑐 in (8.12) only for convenience, in order not to have two indices 𝑏, 𝑐 corresponding to

the same multi-index 𝜇 ∈ ℕ𝑚 . Infact, it is clear that if we eliminate this constraint we have 𝜛(𝑏, 𝑐) = 𝜛(𝑐, 𝑏)
for all 𝑏, 𝑐 ∈ {2, … , 𝑚}.

85
the basis associated to the adapted variables defined in Section 5.2.3 (see (5.17)), namely

𝚢1 ∶= 𝑥1 , 𝚢2 = 𝚢2 (𝚊) ∶= 𝑥2 − 𝑎21 𝑥1 …
𝚢𝑚 = 𝚢𝑚 (𝚊) ∶= 𝑥𝑚 − 𝑎𝑚1 𝑥1 .
(8.15)
By (5.20), (5.39) and (8.13), the quadratic terms of the transformed polynomial 𝙿𝚊 read

⎧𝚙𝜈(1,1) 𝚢2 = 𝚙𝜈(1,1) 𝑥2
1 1
⎪
⎨𝚙𝜈(𝓁,1) 𝚢1 𝚢𝓁 = 𝚙𝜈(𝓁,1) 𝑥1 (𝑥𝓁 − 𝑎𝓁1 𝑥1 ) 𝓁 = 2, … , 𝑚
⎪
⎩𝚙𝜛(𝑗,𝓁) 𝚢𝑗 𝚢𝓁 = 𝚙𝜛(𝑗,𝓁) (𝑥𝑗 − 𝑎𝑗1 𝑥1 )(𝑥𝓁 − 𝑎𝓁1 𝑥1 ) 𝑗, 𝓁 = 2, … , 𝑚 , 𝑗 ≤ 𝓁 .
(8.16)

By expression (8.16), we infer that - in the original variables 𝑥1 , … , 𝑥𝑚 - for any

𝑗, 𝓁 ∈ {2, … , 𝑚}, 𝑗 ≤ 𝓁, the coefficient associated to the monomial 𝑥𝑗 𝑥𝓁 is 𝚙𝜛(𝑗,𝓁) ,
that is 𝑝𝜛(𝑗,𝓁) = 𝚙𝜛(𝑗,𝓁) .

8.2.2 Partition
For any 𝑃 ∈ (𝑟, 𝑚), we set

⎛2𝑝𝜛(2,2) 𝑝𝜛(2,3) … 𝑝𝜛(2,𝑚) ⎞

⎜ ⎟
𝑝 2𝑝𝜛(3,3) … 𝑝𝜛(3,𝑚) ⎟
ℍ1 (𝑃 ) ∶= ⎜ 𝜛(2,3) (8.17)
⎜ … … ⋱ … ⎟
⎜𝑝 𝑝𝜛(3,𝑚) … 2𝑝𝜛(𝑚,𝑚) ⎟⎠
⎝ 𝜛(2,𝑚)

and
𝖲11 (𝑟, 𝑚) ∶= {𝑃 ∈ (𝑟, 𝑚)| det ℍ1 (𝑃 ) ≠ 0} . (8.18)
Remark 8.2. Matrix ℍ1 (𝑃 ) is invariant under the transformations of section 5.2.3. Namely,
by Lemma 8.2, we have

⎛2𝚙𝜛(2,2) 𝚙𝜛(2,3) … 𝚙𝜛(2,𝑚) ⎞

⎜ ⎟
𝚙 2𝚙𝜛(3,3) … 𝚙𝜛(3,𝑚) ⎟
ℍ1 (𝙿𝚊 ) ∶= ⎜ 𝜛(2,3) = ℍ1 (𝑃 ) . (8.19)
⎜ … … ⋱ … ⎟
⎜𝚙 … 2𝚙𝜛(𝑚,𝑚) ⎟⎠
⎝ 𝜛(2,𝑚) 𝚙𝜛(3,𝑚)

We also define

𝖲12 (𝑟, 𝑚) ∶= (𝑟, 𝑚)∖𝖲11 (𝑟, 𝑚) = {𝑃 ∈ (𝑟, 𝑚)| det ℍ1 (𝑃 ) = 0} , (8.20)

so that we have the disjoint union

⨆
(𝑟, 𝑚) = 𝖲11 (𝑟, 𝑚) 𝖲12 (𝑟, 𝑚) . (8.21)

86
 We now consider the images of 𝖲11 (𝑟, 𝑚) and 𝖲12 (𝑟, 𝑚) through the transformation 𝔘1
defined in Remark 5.7, namely
( ) { }
𝒮11 (𝑟, 𝑚) ∶= 𝔘1 𝖲11 (𝑟, 𝑚) × 𝜗̂1𝑚 (1) = (𝙿𝚊 , 𝚊) ∈ 𝑊̂ 1 (𝑟, 𝑚)| det ℍ1 (𝙿𝚊 ) ≠ 0
( ) { }
𝒮21 (𝑟, 𝑚) ∶= 𝔘1 𝖲12 (𝑟, 𝑚) × 𝜗̂1𝑚 (1) = (𝙿𝚊 , 𝚊) ∈ 𝑊̂ 1 (𝑟, 𝑚) | det ℍ1 (𝙿𝚊 ) = 0 .
(8.22)

By (8.22), we have the partition

⨆
̂ 1 (𝑟, 𝑚) = 𝒮 1 (𝑟, 𝑚)
𝑊 𝒮21 (𝑟, 𝑚) . (8.23)
1

Remark 8.3. It is clear that the above partition can be implemented also in case one
considers adapted variables (𝙿𝚊 , 𝚊) ∈ 𝑊 ̂ 𝑖 (𝑟, 𝑚), with 𝑖 ∈ {2, … , 𝑚}. By suitably modi-
fying the family of indices in (8.12), as well as by introducing an adapted matrix ℍ𝑖 (𝑃 ),
it is possible to define sets 𝖲𝑖1 , 𝖲𝑖2 whose disjoint union yields (𝑟, 𝑚) and sets 𝒮1𝑖 , 𝒮2𝑖
whose disjoint union yields 𝑊 ̂ 𝑖 (𝑟, 𝑚). However, the underlying reasonings are not con-
ceptually different from the ones we did above, therefore we omit them in order not to
burden the exposition.

8.3 Two important results

Consider three integers 𝑟, 𝑚 ≥ 2 and 𝑠 ∈ {1, … , 𝑟 − 1}. The two results below are the
cornerstones of the proof of Theorems C1-C2-C3.

Theorem 8.1. In case 𝑟 ≥ 2, 𝑠 = 1, for any 𝑖 ∈ {1, … , 𝑚} the semi-algebraic sets

̂𝑖 (𝑟, 1, 𝑚), and 𝜎
𝜎 ̂(𝑟, 1, 𝑚) are closed in (𝑟, 𝑚), that is, by formulas (8.9),

̂𝑖 (𝑟, 1, 𝑚) = Σ
𝜎 ̂𝑖 (𝑟, 1, 𝑚) ∀ 𝑖 ∈ {1, … , 𝑚} , ̂ 1, 𝑚) .
̂(𝑟, 1, 𝑚) = Σ(𝑟,
𝜎

Moreover, for any 𝑖 ∈ {1, … , 𝑚}, taking the definition of transformation Υ𝑖 into account
(see (5.29)), the set Π𝑊 𝑖 ̂𝑖 ̂𝑖
̂ 𝑖 (𝑟,𝑚) Υ (𝑍 (𝑟, 1, 𝑚)) is closed in 𝑊 (𝑟, 𝑚), and its form can be
explicitly computed.

Theorem 8.2. For any given values of 𝑟 ≥ 3, 𝑠 ≥ 2, and 𝑖 ∈ {1, … , 𝑚} there exist two
semi-algebraic subsets of (𝑟, 𝑚)

𝑋1𝑖 (𝑟, 𝑠, 𝑚) ⊂ 𝖲𝑖1 (𝑟, 𝑚) , 𝑋2𝑖 (𝑟, 𝑠, 𝑚) ⊂ 𝖲𝑖2 (𝑟, 𝑚) , (8.24)

̂ 𝑖 (𝑟, 𝑚)
and two semi-algebraic subsets of 𝑊

𝑌1𝑖 (𝑟, 𝑠, 𝑚) ⊂ 𝒮1𝑖 (𝑟, 𝑚) , 𝑌2𝑖 (𝑟, 𝑠, 𝑚) ⊂ 𝒮2𝑖 (𝑟, 𝑚) , (8.25)

verifying the following properties:

⨆
̂𝑖 (𝑟, 𝑠, 𝑚) = 𝑋1𝑖 (𝑟, 𝑠, 𝑚)
1. one has the partition 𝜎 𝑋2𝑖 (𝑟, 𝑠, 𝑚) ;

87
 2. for 𝑗 ∈ {1, 2}, one has
( )
𝑋𝑗𝑖 (𝑟, 𝑠, 𝑚) = Π(𝑟,𝑚) (𝔘𝑖 )−1 (𝑌𝑗𝑖 (𝑟, 𝑠, 𝑚)) ,

where 𝔘𝑖 was defined in Remarks 5.7-5.8;

3. 𝑌1𝑖 (𝑟, 𝑠, 𝑚) is closed in 𝒮1𝑖 (𝑟, 𝑚) for the topology induced by (𝑟, 𝑚);

4. 𝑋1𝑖 (𝑟, 𝑠, 𝑚) is closed in 𝖲𝑖1 (𝑟, 𝑚) for the topology induced by (𝑟, 𝑚);

5. the form of 𝑌1𝑖 (𝑟, 𝑠, 𝑚) can be explicitly computed starting from the expression of
𝑍̂ 𝑖 (𝑟, 𝑠, 𝑚) by the means of an algorithm involving only linear operations.

The rest of this section is devoted to the proof of the above results.
We will only prove Theorems 8.1-8.2 in the case 𝑖 = 1, as the other cases are simple
generalizations.

8.3.1 Proof of Theorem 8.1 and strategy of proof of Theorem 8.2.

We have seen in section 5 that the equations determining 𝑍 1 (𝑟, 𝑠, 𝑚), can be written in
the adapted coordinates introduced in section 5.2.3. Namely, by recalling the functions

𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ∶ 𝑊 1 (𝑟, 𝑚)×ℝ(𝑚−1)(𝑠−1) ⟶ ℝ , 𝚒 ∈ {1, … , 𝑚} 𝛼 ∈ {0, … , 𝑠} ,

(8.26)
presented in (5.36)-(5.37), Theorem 5.1 ensures that 𝑍 1 (𝑟, 𝑠, 𝑚) is the image through
the inverse of the transformation Υ1 in (5.30) of the zero set
⋂
{(𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) ∈ 𝑊 1 (𝑟, 𝑚) × ℝ(𝑚−1)(𝑠−1) |𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0} , (8.27)
𝚒∈{1,…,𝑚}
𝛼∈{0,…,𝑠}

and the explicit form of the quantities 𝚀𝑖𝛼 is given in (5.43)-(5.44).

Then, if we indicate by ̂ ̂ 1 (𝑟, 𝑚)×𝜗1 (𝑠, 2) ⊂
𝚀𝚒𝛼 the restriction of 𝚀𝚒𝛼 to the subset 𝑊
( ) 𝑚
1 1 ̂
𝑊 (𝑟, 𝑚) × 𝜗𝑚 (𝑠, 2), and if we denote by 𝒩 {𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} the zero set of the non-
𝛼∈{0,…,𝑠}
linear maps {̂
𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} , it is clear by the discussion at paragraph 8.1, in particular
𝛼∈{0,…,𝑠}
by (8.11), that
( ( ))
̂ 1 1 −1 ̂
𝑍 (𝑟, 𝑠, 𝑚) ∶= (Υ ) 𝒩 {𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} (8.28)
𝛼∈{0,…,𝑠}

that is
( )
1 ̂ 1 (𝑟, 𝑠, 𝑚)) = 𝒩
Υ (𝑍 {̂
𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} . (8.29)
𝛼∈{0,…,𝑠}

88
Proof. (Theorem 8.1) When 𝑠 = 1, then in (5.43)-(5.44) one must consider 𝛼 ∈ {0, 1}.
We observe that
∀ 𝛼 ∈ {0, 1} , ∀𝚒 = 1, … , 𝑚 ̂
𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 ⟺ 𝚙𝜈(𝚒,𝛼) = 0 , (8.30)
where the family of multi-indices 𝜈(𝚒, 𝛼) was defined in (5.39). As we see, no parame-
ters belonging to the space 𝜗1𝑚 (𝑠, 2) appear in equation (8.30), which determines a closed
subset of (𝑊̂ 1 (𝑟, 𝑚). In other words, by (8.29) with 𝑠 = 1, we have that the projection
)
Π𝑊 1 ̂1 ̂1
̂ 1 (𝑟,𝑚) Υ (𝑍 (𝑟, 1, 𝑚)) is closed in 𝑊 (𝑟, 𝑚) and its form is given in (8.30). More-
over, as the function 𝔘 1
( introduced( in Remark 5.7 ))is continuous by construction, the
inverse image (𝔘 )1 −1 Π𝑊 1 ̂ 1
̂ 1 (𝑟,𝑚) Υ (𝑍 (𝑟, 1, 𝑚)) is closed in (𝑟, 𝑛) × 𝜗̂1𝑚 (1). Fi-
nally, as all the elements of 𝜗̂1𝑚 (1) live in a compact ball (see (8.3) and Remark (8.1)),
using Lemma E.1 with 𝐸 ≡ (𝑟, 𝑚) and 𝐾 ≡ 𝜗̂1𝑚 (1) and taking formulas (8.9)-(8.29)
into account one has that
[ ( ( ))]
𝜎̂1 (𝑟, 1, 𝑚) = Π(𝑟,𝑚) (𝔘1 )−1 Π𝑊 ̂ (𝑟,𝑚)
1 Υ 1 ̂1
(𝑍 (𝑟, 1, 𝑚)) (8.31)
is closed. This concludes the proof.

Now, taking (8.29), the third line of (8.9) and the proof of Theorem 8.1 into account,
the key idea behind the proof of Theorem 8.2 consists in understanding under which
conditions the set Υ1 (𝑍 ̂ 1 (𝑟, 𝑠, 𝑚)) admits a closed projection onto 𝑊 ̂ 1 (𝑟, 𝑚).
In particular, we will see that if we only consider polynomials in the set 𝖲11 (𝑟, 𝑚) ⊂
(𝑟, 𝑚), it is possible to extrapolate linearly from the 𝑚𝑠 + 𝑚 equations in (8.29) the
(𝑚 − 1)𝑠 parameters of 𝜗1𝑚 (𝑠, 2) as explicit algebraic functions of the parameters of
𝒮11 (𝑟, 𝑚) ∶= 𝔘1 (𝖲11 (𝑟, 𝑚) × 𝜗̂1𝑚 (1)), as the discussion at the following paragraph shows.
We start by setting
( )
⋂
𝖠(𝑟, 𝑠, 𝑚) ∶=𝒩 {̂ 𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} Υ1 (𝖲11 (𝑟, 𝑠, 𝑚) × 𝜗̂1𝑚 (𝑠)) , (8.32)
𝛼∈{0,…,𝑠}
and we remind that, by Remark 5.7,
( )
Υ1 (𝖲11 (𝑟, 𝑠, 𝑚) × 𝜗̂1𝑚 (𝑠)) = 𝔘1 𝖲11 (𝑟, 𝑠, 𝑚) × 𝜗̂1𝑚 (1) × 𝜗1𝑚 (𝑠, 2) = 𝒮11 (𝑟, 𝑚) × 𝜗1𝑚 (𝑠, 2) .
(8.33)
We now claim that 𝖠(𝑟, 𝑠, 𝑚) is a graph of the form

𝖠(𝑟, 𝑠, 𝑚) = 𝑌11 (𝑟, 𝑠, 𝑚) × 𝜃(𝑟, 𝑠, 𝑚) , (8.34)

where
𝑌11 (𝑟, 𝑠, 𝑚) ∶= Π𝑊
̂ 1 (𝑟,𝑚) 𝖠(𝑟, 𝑠, 𝑚)

is a closed subset of 𝒮11 (𝑟, 𝑚) for the induced topology determined by algebraic equa-
tions involving the coordinates of elements in 𝒮11 (𝑟, 𝑚), and the points of 𝜃(𝑟, 𝑠, 𝑚) ⊂
𝜗1𝑚 (𝑠, 2) are parametrized by 𝑌11 (𝑟, 𝑠, 𝑚). In the following subparagraphs, we will prove
this claim and show how it can be used in order to demonstrate Theorem 8.2.

89
8.3.2 Explicit form of the equations (case 𝑟 ≥ 3, 𝑠 ≥ 2)
( )
The goal of this paragraph is to give a more explicit expression of 𝒩 {̂
𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} ,
𝛼∈{0,1}
in case 𝑟 ≥ 3, 𝑠 ≥ 2.
Remark 8.4. We remind that equations (5.43)-(5.44) are recursive w.r.t. the parame-
ters of the curve 𝛾. Namely, for any given integer 𝛽 ∈ {2, … , 𝑠}, the coefficients of
order 𝛽 belonging to the space 𝜗1𝑚 (𝑠, 2) - that is 𝑎22 , … , 𝑎2𝛽 , … , 𝑎𝑚2 , … , 𝑎𝑚𝛽 - appear
in equations (5.43)-(5.44) only for 𝛼 ≥ 𝛽.
For any polynomial 𝑃 ∈ (𝑟, 𝑚) and any curve 𝛾 ∈ Θ ̂ 1 , we indicate by 𝙿(>2)
𝚊 the
𝑚
associated polynomial 𝙿𝚊 written in the adapted coordinates for 𝛾 (see paragraph 5.2.3)
deprived of its monomials of degree less or equal than two 19 . Also, if 𝑠,𝛾 ∈ 𝜗̂1𝑚 (1) is
the 𝑠-truncation of 𝛾, for any given 𝛼 ∈ {2, … , 𝑠} we denote by

⎛ 0 ⎞
⎜ ∑𝑠 𝑎 𝑡 𝑖 ⎟
𝙹𝑠,𝛾,𝚊 (<𝛼) (𝑡) ∶= 𝙹𝑠,𝛾,𝚊 − ⎜ 𝑖=𝛼 2𝑖 ⎟
⎜ … ⎟
⎜∑𝑠 𝑎 𝑡𝑗 ⎟
⎝ 𝑗=𝛼 𝑚𝑗 ⎠

its truncation at order 𝛼 − 1 written in the adapted coordinates for 𝛾.

Remark 8.5. We observe that, for 𝛼 = 2, 𝙹𝑠,𝛾,𝚊 (<2) (𝑡) reduces to the line (𝑡, 0, … , 0),
since with the exception of the parametrizing coordinate, the components of 𝙹𝑠,𝛾,𝚊 start
at order two in 𝑡 (see paragraph 5.2.3).
With this setting, we have

Lemma 8.3. For any polynomial 𝑃 ∈ (𝑟, 𝑚), there exists a linear change of coor-
dinates 𝔇 = 𝔇(𝑃 ) ∶ ℝ𝑚 ⟶ ℝ𝑚 such that for any given 𝛼 ∈ {2, … , 𝑠}, and for
𝚒 = 1, … , 𝑚, the algebraic equations ̂
𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 in (5.43)-(5.44) take the
form
( )
⎛0 ⎞ ⎛̂𝚀1𝛼 𝙿(>2)
𝚊 , 𝙹𝑠,𝛾,𝚊 (<𝛼) ⎞
0 0 … 0 ⎛ 0 ⎞ ⎜ ( )⎟
⎜ ′ ⎟⎜ ′ ⎟ ⎜ ̂ (>2)
, (<𝛼) ⎟
⎜0 2𝚙 0 … 0 ⎟ ⎜ 𝑎2𝛼 ⎟ 𝚀 2𝛼 𝙿 𝙹𝑠,𝛾,𝚊
𝜛(2,2) ( 𝚊 )⎟
⎜0 0 2𝚙′𝜛(3,3) … 0 ⎟ ⎜ 𝑎′ ⎟ + 𝔇 ⎜⎜ ̂ (>2) (<𝛼) ⎟ = 0 .
⎜ ⎟ 3𝛼 𝚀 3𝛼 𝙿𝚊 , 𝙹𝑠,𝛾,𝚊
⎜0 … … … … ⎟⎜ … ⎟ ⎜ ⎟
⎜0 ⎟ ⎜ ′ ⎟ ⎜ ( … ) ⎟
⎝ 0 0 … 2𝚙𝜛(𝑚,𝑚) ⎠ ⎝𝑎𝑚𝛼 ⎠
′
⎜̂ (>2) (<𝛼) ⎟
⎝ 𝑚𝛼 𝚊
𝚀 𝙿 , 𝙹 𝑠,𝛾,𝚊 ⎠
(8.35)
19 One has 𝙿(>2)
𝚊 ≠ 0 since we are considering the case of polynomials having degree 𝑟 ≥ 3.

90
Proof. Step 1. We firstly claim that equations (5.43)-(5.44) can be but into the form
( )
⎛̂𝚀1𝛼 𝙿𝚊(>2) , 𝙹𝑠,𝛾,𝚊 (<𝛼) ⎞
⎛0 0 0 … 0 ⎞⎛ 0 ⎞ ⎜ ( )⎟
⎜0 2𝚙 𝚙 … 𝚙 ⎟ ⎜𝑎 ⎟ ⎜̂ 𝚀2𝛼 𝙿(>2)𝚊 , 𝙹𝑠,𝛾,𝚊 (<𝛼) ⎟
⎜ 𝜛(2,2) 𝜛(2,3) 𝜛(2,𝑚) ⎟ ⎜ 2𝛼 ⎟ ⎜ ( )⎟
⎜0 𝚙𝜛(2,3) 2𝚙𝜛(3,3) … 𝚙𝜛(3,𝑚) ⎟ ⎜ 𝑎3𝛼 ⎟ + ⎜ ̂ 𝚀3𝛼 𝙿(>2)𝚊 , 𝙹 𝑠,𝛾,𝚊
(<𝛼) ⎟ = 0 .
⎜0 … … … … ⎟⎜ … ⎟ ⎜ ⎟
⎜ ⎟⎜ ⎟ ⎜ ( … ) ⎟
⎝0 𝚙𝜛(2,𝑚) 𝚙𝜛(3,𝑚) … 2𝚙𝜛(𝑚,𝑚) ⎠ ⎝𝑎𝑚𝛼 ⎠ ⎜̂ (>2) (<𝛼) ⎟
⎝𝚀𝑚𝛼 𝙿𝚊 , 𝙹𝑠,𝛾,𝚊 ⎠
(8.36)
By (8.30), for any 𝛼 ∈ {2, … , 𝑠}, and for any 𝚒, 𝚓 ∈ {1, … , 𝑚}, the monomials
𝚙𝜈(𝚓,0) , 𝚙𝜈(𝚓,1) do not yield any contribution to equations 𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0. There-
fore, taking (8.13) into account, the only monomials of order two which may appear in
equation 𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 are those associated to the multi-indices 𝜛(𝑏, 𝑐), with
𝑏, 𝑐 ∈ {2, … , 𝑚}, 𝑏 ≤ 𝑐.
Moreover, fixing the values of 𝚒 ∈ {1, … , 𝑚} and 𝛼 ∈ {2, … , 𝑠}, by (5.43)-(5.44),
the multi-indices 𝜇 ∈ ℕ𝑚 contributing to equation ̂ 𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 are those for
which the set 𝑚 (̃ 𝜇 (𝚒), 𝛼) in (5.41) is non-empty. This amounts to requiring that the
components of the vector (𝑘22 , … , 𝑘2𝛼 , … , 𝑘𝑚2 , … , 𝑘𝑚𝛼 ) ∈ ℕ(𝑚−1)(𝛼−1) appearing in
formulas (5.43)-(5.44) satisfy
∑
𝛼 ∑
𝑚 ∑
𝛼
𝑘𝑗𝑖 = 𝜇̃𝑗 (𝚒) ∀𝑗 ∈ {2, … , 𝑚} , 𝜇̃1 (𝚒) + 𝑖 𝑘𝑗𝑖 = 𝛼 . (8.37)
𝑖=2 𝑗=2 𝑖=2

In particular, for fixed 𝚒 ∈ {1, … , 𝑚}, 𝛼 ∈ {2, … , 𝑠} and 𝓁 ∈ {2, … , 𝑚}, if we look at
the monomials containing the coefficient 𝑎𝓁𝛼 in equation ̂ 𝚀𝚒𝛼 (𝙿𝚊 , 𝙹𝑠,𝛾,𝚊 ) = 0 - that is, at
the form of the terms for which 𝑘𝓁𝛼 ≠ 0 in (5.43)-(5.44) - by (8.37) we must have

𝜇̃1 (𝚒) = 0 , 𝑘𝑗𝑖 = 𝛿𝑖𝛼 𝛿𝑗𝓁 , 𝜇̃𝑗 (𝚒) = 𝛿𝑗𝓁 , 𝑗 ∈ {2, … , 𝑚} . (8.38)

Firstly, expression (8.38) implies |𝜇| = 2. Secondly, as we said above, the only multi-
indices of length |𝜇| = 2 which may yield a contribution to 𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 are
those belonging to the family {𝜛(𝑏, 𝑐)}𝑏,𝑐∈{2,…,𝑚}𝑏≤𝑐 in (8.12). Therefore, we have two
cases.
Case 𝚒 = 1. If 𝜇 ∈ {𝜛(𝑏, 𝓁)}𝑏,𝓁∈{2,…,𝑚}𝑏≤𝓁 then 𝜇1 = 0, and all terms in formula
(5.43) which are associated to these indices are multiplied by zero. Hence, the coeffi-
cients 𝑎𝓁𝛼 do not appear in ̂𝚀1𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0, nor do any of the monomials of order
two in 𝙿𝚊 . This, together with Remark 8.4 proves the claim for 𝚒 = 1 (the first line of
(8.36)).
Case 𝚒 ∈ {2, … , 𝑚}. Taking (8.38) into account, for any given 𝓁 ∈ {2, … , 𝑚} one
has that
1. if 𝚒 ≤ 𝓁, the only term to which the coefficient 𝑎𝓁𝛼 is associated in equation
̂
𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 is the one corresponding to the multi-index 𝜇 ∈ 𝜛(𝚒, 𝓁),
that is, by (5.44), the monomial (1 + 𝛿𝓁𝚒 )𝚙𝜛(𝚒,𝓁) 𝑎𝓁𝛼 ;

91
 2. if 𝚒 > 𝓁, the term containing 𝑎𝓁𝛼 is the one associated to the multi-index 𝜇 ∈
𝜛(𝓁, 𝚒), that is, by (5.44), (1 + 𝛿𝓁𝚒 )𝚙𝜛(𝓁,𝚒) 𝑎𝓁𝛼 .

Conversely, if a monomial associated to an index 𝜛(𝚒, 𝓁), with 𝚒, 𝓁 ∈ {2, … , 𝑚},

𝚒 ≤ 𝓁, appears in equations ̂ 𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0, then, by (8.12) and by (8.37), one
must necessarily have
{ ∑ ∑𝛼
𝜇̃1 (𝚒) + 𝑚 𝑗=2 𝑖=2 𝑖 𝑘𝑗𝑖 = 𝛼 , 𝜇̃1 (𝚒) = 0
∑𝛼 (8.39)
𝑖=2 𝑘𝑗𝑖 = 𝜇
̃𝑗 (𝚒) ∶= 𝛿𝑗𝓁 + 𝛿𝑗𝚒 − 𝛿𝑗𝚒 = 𝛿𝑗𝓁 ∀𝑗 ∈ {2, … , 𝑚} ,

which is true if and only if for some 𝑣 ∈ {2, … , 𝛼} one has

{
𝑘𝑗𝑖 = 𝛿𝑗𝓁 𝛿𝑖𝑣
∑𝑚 ∑𝛼 (8.40)
𝑗=2 𝑖=2 𝑖 𝛿𝑗𝓁 𝛿𝑖𝑣 = 𝛼

that is if and only if 𝑘𝑗𝑖 = 𝛿𝑗𝓁 𝛿𝑖𝛼 . One can check by formula (5.44) that this ensures
that such a term must be of the form (1 + 𝛿𝓁𝚒 )𝚙𝜛(𝚒,𝓁) 𝑎𝓁𝛼 . This reasoning, together
with Remark 8.4, and with the fact that - as we showed at the beginning of the proof
- no monomials of order two appear in equation ̂ 𝚀𝚒𝛼 (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 ) = 0 other than those
associated to the family (8.12), proves the claim for 𝚒 ∈ {2, … , 𝑚} (rows 2, … , 𝑚 of
(8.36)).
Step 2. For any 𝑃 ∈ (𝑟, 𝑚), taking Lemma 8.2 into account we indicate by 𝔾1 (𝑃 )
the matrix
⎛0 0 … 0 ⎞ ⎛0 0 … 0 ⎞
⎜ ⎟ ⎜ ⎟ ( )
⎜0 2𝑝𝜛(2,2) … 𝑝𝜛(2,𝑚) ⎟ ⎜0 2𝚙𝜛(2,2)
=
… 𝚙𝜛(2,𝑚) ⎟
=
0 0
⎜0 … ⋱ … ⎟ ⎜0 … ⋱ … ⎟ 0 ℍ1 (𝑃 )
⎜0 𝑝𝜛(2,𝑚) … ⎟ ⎜
2𝑝𝜛(𝑚,𝑚) ⎠ ⎝0 𝚙𝜛(2,𝑚) … ⎟
2𝚙𝜛(𝑚,𝑚) ⎠
⎝
(8.41)
appearing in (8.36). 𝔾1 (𝑃 ) is symmetric, hence diagonalizable. Hence, for any 𝑃 ∈
(𝑟, 𝑚), there exist a basis of eigenvectors, indicated by

𝑣𝚊 , 𝑢′2 = 𝚞′2 (𝑃 ), … , 𝑢′𝑚 = 𝑢′𝑚 (𝑃 ) , (8.42)

and a real 𝑚 × 𝑚 invertible matrix 𝔇 = 𝔇(𝑃 ), such that equation (8.35) takes the form
( )
(>2)
⎛0 ⎞⎛ 0 ⎞ ⎛ 𝚀1𝛼 𝙿𝚊 , 𝙹 𝑠,𝛾,𝚊
(<𝛼) ⎞
0 0 … 0 ⎜ ( )⎟
⎜ ′ ⎟⎜ ′ ⎟ ⎜ 𝚀2𝛼 𝙿𝚊(>2) , 𝙹𝑠,𝛾,𝚊 (<𝛼) ⎟
⎜0 2𝚙 0 … 0 ⎟ ⎜ 𝑎2𝛼 ⎟
𝜛(2,2) ( )⎟
⎜0 0 2𝚙′𝜛(3,3) … 0 ⎟ ⎜ 𝑎′ ⎟ + 𝔇 ⎜⎜ (>2) (<𝛼) ⎟ = 0 ,
⎜ ⎟ 3𝛼 𝚀3𝛼 𝙿𝚊 , 𝙹 𝑠,𝛾,𝚊
⎜0 … … … … ⎟⎜ … ⎟ ⎜ ⎟
⎜0 ⎟ ⎜ ′ ⎟ ⎜ ( … ) ⎟
⎝ 0 0 … 2𝚙𝜛(𝑚,𝑚) ⎠ ⎝𝑎𝑚𝛼 ⎠
′
⎜ (>2) (<𝛼) ⎟
⎝𝚀𝑚𝛼 𝙿𝚊 , 𝙹𝑠,𝛾,𝚊 ⎠
(8.43)
where the primed quantities indicate that we are working in the new basis (8.42).

92
Remark 8.6. Comparing (8.14) with (8.42), we observe that the vector 𝑣𝚊 was left un-
changed. This is due to the fact that, by (8.41), 𝑣𝚊 is already an eigenvector of 𝔾1 (𝑃 )
(associated to a null eigenvalue).

8.3.3 Proof of Theorem 8.2

For 𝑟 ≥ 3, 𝑠 ≥ 2, we define
⋂
𝑋11 (𝑟, 𝑠, 𝑚) ∶= 𝜎
̂1 (𝑟, 𝑠, 𝑚) 𝖲11 (𝑟, 𝑚) . (8.44)

Step 1. Firstly, we show that, if 𝑃 ∈ 𝖲11 (𝑟, 𝑚), then the parameters of the space
𝜗1𝑚 (𝑠, 2)can be reduced iteratively from equation (8.35) for 𝛼 ∈ {2, … , 𝑠}.
When 𝛼 = 2 the second term at the l.h.s. of formula (8.35) in Lemma 8.3.2
does not depend on the parameters of 𝜗1𝑚 (𝑠, 2) (see Remark 8.5), so that the coeffi-
cients 𝑎22 , … , 𝑎𝑚2 can be reduced, as matrix ℍ1 (𝑃 ) is invertible by construction when
𝑃 ∈ 𝖲11 (𝑟, 𝑚).
If, for 𝛼 ∈ {3, … , 𝑠}, we assume that the parameters 𝑎𝑗𝛽 , with 𝑗 ∈ {2, … , 𝑚} and
𝛽 ∈ {2, … , 𝛼 − 1}, have been reduced, then the first equation in (8.35) does not contain
any new parameter, whereas the terms 𝑎2𝛼 , … , 𝑎𝑚𝛼 can be found by inverting ℍ1 (𝑃 )
once again.
The above considerations and (8.44) imply that if 𝑃 ∈ 𝑋11 (𝑟, 𝑠, 𝑚) then

1. the parameters of the space 𝜗1𝑚 (𝑠, 2) can be reduced from the equations in (8.35)
by the means of a recursive algorithm which only involves linear computations
and the inversion of ℍ1 (𝑃 );

2. there exists a truncation 𝑠,𝛾 ∈ 𝜗1𝑚 (𝑠) such that 𝔘1 (𝑃 × 𝑠,𝛾 ) = (𝙿𝚊 , 𝚊, 𝙹𝑠,𝛾,𝚊 )
solves (8.35).

Taking (8.32) and the above arguments into account, with the notations of section
8.3.1 we have that 𝖠(𝑟, 𝑠, 𝑚) is determined by a system of 𝑚𝑠 + 𝑚 − (𝑚 − 1)(𝑠 − 1) =
𝑠 + 2𝑚 − 1 explicit algebraic equations involving only the coordinates of 𝒮11 (𝑟, 𝑚), and
of (𝑚−1)(𝑠−1) explicit algebraic equations that parametrize the coefficients of 𝜗1𝑚 (𝑠, 2)
as functions of the points in 𝒮11 (𝑟, 𝑚). In other words, 𝖠(𝑟, 𝑠, 𝑚) has the form of a graph
of the type indicated in (8.34), and 𝑌11 (𝑟, 𝑠, 𝑚) is a closed algebraic subset of 𝒮11 (𝑟, 𝑚)
whose form can be explicitly computed - starting from the expression of Υ1 (𝑍 ̂ 1 (𝑟, 𝑠, 𝑚))
- by the means of an algorithm involving only linear operations.
Moreover, by Remark 5.7 the function 𝔘1 is polynomial, and 𝖲11 (𝑟, 𝑠, 𝑚) is obviously
semi-algebraic in (𝑟, 𝑚) by construction. Therefore, by (8.32)-(8.33) the set 𝖠(𝑟, 𝑠, 𝑚)
̂ 1 (𝑟, 𝑚) × 𝜗1 (𝑠, 2), and 𝑌 1 (𝑟, 𝑠, 𝑚) is semi-algebraic in 𝑊
is semi-algebraic in 𝑊 ̂ 1 (𝑟, 𝑚)
𝑚 1
by the Theorem of Tarski and Seidenberg A.1.

93
 Step 2. Since the invertible transformation 𝔘1 defined in Remark 5.7 is polynomial,
due to Step 1 (see (8.32)-(8.33)-(8.34)) and to continuity we have that the inverse image

(𝔘1 )−1 (𝑌11 (𝑟, 𝑠, 𝑚)) ⊂ 𝖲11 (𝑟, 𝑚) × 𝜗̂1𝑚 (1) (8.45)

is closed in 𝖲11 (𝑟, 𝑚) × 𝜗̂1𝑚 (1) for the induced topology. Finally by taking (8.32)-(8.34)-
(8.44) into account, and by considering the fact that, as we have already pointed out in
Remark 8.1, 𝜗̂1𝑚 (1) is compact - Lemma E.1 ensures that the projection
( ) ⋂
Π(𝑟,𝑚) (𝔘1 )−1 (𝑌11 (𝑟, 𝑠, 𝑚)) = 𝑋11 (𝑟, 𝑠, 𝑚) = 𝜎
̂1 (𝑟, 𝑠, 𝑚) 𝖲11 (𝑟, 𝑚) (8.46)

is closed in 𝖲11 (𝑟, 𝑠, 𝑚) for the topology induced by (𝑟, 𝑚).

The semi-algebraicness of the projection in (8.46) is a consequence of the semi-
algebraicness of 𝑌11 (𝑟, 𝑠, 𝑚) demonstrated at Step 1, of the fact that 𝔘1 is polynomial,
and of the Theorem of Tarski and Seidenberg.
The above arguments prove Theorem 8.2 once one sets

𝑋21 (𝑟, 𝑠, 𝑚) ∶= 𝜎
̂1 (𝑟, 𝑠, 𝑚)∖𝑋11 (𝑟, 𝑠, 𝑚)
( ( ))
1
(8.47)
𝑌2 (𝑟, 𝑠, 𝑚) ∶= Π𝑊 ̂ 1 (𝑟,𝑚) 𝒩
̂
{𝚀𝚒𝛼 }𝚒∈{1,…,𝑚} ∖𝑌11 (𝑟, 𝑠, 𝑚) .
𝛼∈{0,…,𝑠}

9 Proof of Theorems C1-C2-C3

We assume the notations of section 8. In this section, we consider two positive integers
𝑟 ≥ 2, 𝑛 ≥ 3, a vector 𝚜 ∶= (𝑠1 , … , 𝑠𝑛−1 ) ∈ ℕ𝑛−1 , with 1 ≤ 𝑠𝑖 ≤ 𝑟 − 1 for all
𝑖 = 1, … , 𝑛 − 1, and a function ℎ of class 𝐶𝑏2𝑟−1 around the origin, verifying ∇ℎ(0) ≠ 0.
Also, for any 𝑚 ∈ {2, … , 𝑛 − 1}, we set 𝒦 (𝑟, 𝑛, 𝑚) ∶= 𝙼(𝑟, 𝑛, 𝑚, 1), where the
constant 𝙼(𝑟, 𝑛, 𝑚, 1) is the one appearing in the statement of Theorem 4.1 (see formula
(4.1)).

9.1 Proof of Theorem C1

Fix 𝑚 ∈ {2, … , 𝑛 − 1}. Let Γ𝑚 be a given 𝑚-dimensional subspace belonging to the
subset 0 (ℎ, 𝑚, 𝑛) ⊂ 𝙶(𝑚, 𝑛) introduced in Definition 3.2.
Taking Definition 8.3 into account, we consider a curve 𝛾 ∈ Θ ̂ 𝑚 whose image is
𝑚
contained in Γ . Without any loss of generality, up to changing the order of the vectors
spanning Γ𝑚 , we can suppose that 𝛾 is parametrized by the first coordinate, hence that
𝛾∈Θ ̂ 1 . Following (8.2), we indicate by 𝚊 = (𝑎21 , … , 𝑎𝑚1 ) ∈ 𝐵 𝑚−1 (0, 𝒦 (𝑟, 𝑛, 𝑚)) the
𝑚
linear coefficients coefficients of the Taylor expansion of 𝛾 at the origin, and by 𝑠,𝛾 its
𝑠-truncation (with 1 ≤ 𝑠 ≤ 𝑟 − 1).
We also indicate by 𝑢1 , … , 𝑢𝑚 ∈ 𝚄(𝑚, 𝑛) a orthonormal basis spanning Γ𝑚 , and by
∑
𝑣 ∶= 𝑢1 + 𝑚 𝑖=2 𝑎𝑖1 𝑢𝑖 , 𝑢2 … , 𝑢𝑚 the basis associated to the adapted coordinates for 𝛾

94
introduced in section 5.2.3. As we have already shown in (7.3), the Taylor polynomial
𝚃0 (ℎ, 𝑟, 𝑛) restricted to Γ𝑚 written in the adapted coordinates reads
𝜇1 𝜇2 𝜇𝑚
∑ [ ⏞⏞⏞ ] 𝜇
1 |𝜇| ⏞⏞⏞ ⏞⏞⏞ 𝜇
𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚)(𝚢) = ℎ0 𝑣 , 𝑢2 , … , 𝑢𝑚 𝚢1 1 … 𝚢𝑚𝑚 , (9.1)
𝜇∈ℕ𝑚
𝜇!
1≤|𝜇|≤𝑟

where we have used the notation introduced in formula (2.2). Moreover, as customary,
𝑠,𝛾 reads 𝙹𝑠,𝛾,𝚊 in the new coordinates.
By the arguments at paragraph 8.1, and by taking (5.29) and (8.9) into account, if
we manage to prove that condition
(𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹1,𝛾,𝚊 ) ∈ Υ1 (𝑍
̂ 1 (𝑟, 1, 𝑚))

is never satisfied for any choice of the subspace Γ𝑚 ∈ 0 (ℎ, 𝑚, 𝑛) and of the curve 𝛾,
which is equivalent - due to Theorem 8.1 for 𝑠 = 1 - to condition
̂1 (𝑟, 1, 𝑚) = Σ
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∉ 𝜎 ̂1 (𝑟, 1, 𝑚) , ∀ Γ𝑚 ∈ 0 (ℎ, 𝑚, 𝑛) , (9.2)
then by the definitions in (6.1)-(6.2) we have
𝚃0 (ℎ, 𝑟, 𝑛) ∈
 ⋆ (𝑟, 𝑛)∖Π ⋆ (𝑟,𝑛) {(𝑄, 𝐴, 𝑃 ) ∈  ⋆ (𝑟, 𝑛) × 𝑂(𝑛, 𝑚) × (𝑟, 𝑚)| (9.3)
Span (𝐴) ∈ 0 , 𝑃 (𝑥) = 𝑄(𝐴𝑥) , 𝑃 (𝑥) ∈ Σ(𝑟, 1, 𝑚)} .
Hence, if claim (9.2) is true, taking (9.3) into account one retraces exactly the same
steps used to demonstrate Theorem A, (see section 6), and the thesis follows. As it has
already been pointed out in subparagraph 3.2.4, one just needs to be careful about the
fact that the steepness coefficients, in this case, do not admit a uniform lower bound,
as they may tend to zero when a subspace not belonging to 0 (ℎ, 𝑚, 𝑛) and on which
ℎ is not steep is approached. Moreover, one cannot expect to have steepness on the
subspaces of 0 (ℎ, 𝑚, 𝑛) for an open set of functions around ℎ.
For any choice of Γ𝑚 ∈ 0 (ℎ, 𝑚, 𝑛), by Theorem 5.1 and by (5.37), condition
(𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹1,𝛾,𝚊 ) ∉ Υ1 (𝑍
̂ 1 (𝑟, 1, 𝑚))

is equivalent to asking that system

𝚀𝑖𝛼 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹1,𝛾,𝚊 ) = 0 ∀ 𝑖 ∈ {1, … , 𝑚} , ∀𝛼 ∈ {0, 1} , (9.4)
has no solution. Then, by expressions (7.8)-(7.9), the system in (9.4) is not satisfied if
and only if
⎧(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛) , Span (𝑢1 , … , 𝑢𝑚 ) = Γ𝑚
⎪ 𝑚−1
⎪(𝑎21 , … , 𝑎𝑚1 ) ∈ 𝐵 (0, 𝒦 (𝑟, 𝑛, 𝑚)) , 𝑣 ∶= 𝑢1 + 𝑎21 𝑢2 + ⋯ + 𝑎𝑚1 𝑢𝑚
⎨ 1 1 1
⎪ℎ0 [𝑣] = ℎ0 [𝑢2 ] = ⋯ = ℎ0 [𝑢𝑚 ]
⎪ℎ2 [𝑣, 𝑣] = ℎ2 [𝑣, 𝑢 ] = ⋯ = ℎ2 [𝑣, 𝑢 ] = 0
⎩ 0 0 2 0 𝑚
(9.5)

95
has no solution.
By construction, the Hessian of the restriction ℎ|Γ𝑚 has no null eigenvalues, so that
system (9.5) admits no solution, and both claim (9.2) and expression (9.3) hold true, as
wished. This concludes the proof.

9.2 Proof of Theorem C2

9.2.1 Case of a subspace belonging to 1 (ℎ, 𝑚, 𝑛)
With the usual setting, let 𝑚 ∈ {2, … , 𝑛 − 1} be an integer, and Γ𝑚 ∈ 1 (ℎ, 𝑚, 𝑛) ⊂
𝙶(𝑚, 𝑛) be a 𝑚-dimensional subspace spanned by vectors 𝑢1 , … , 𝑢𝑚 ∈ 𝚄(𝑚, 𝑛).
As the Hessian matrix of the restriction ℎ|Γ𝑚 has at most one null eigenvalue, with-
out any loss of generality one can choose 𝑢1 to be the eigenvector associated to the
unique null eigenvalue, that is

⎧ℎ1 [𝑢1 ] = ℎ1 [𝑢2 ] = ⋯ = ℎ1 [𝑢𝑚 ] = 0

⎪ 20 0 0
⎨ℎ0 [𝑢1 , 𝑢1 ] = ℎ20 [𝑢1 , 𝑢2 ] = ℎ20 [𝑢1 , 𝑢𝑚 ] = 0 (9.6)
⎪ 𝑚
⎩Span(𝑢1 , 𝑢2 , … , 𝑢𝑚 ) = Γ
so that at the same time one must have
⎛ ℎ2 [𝑢2 , 𝑢2 ] ℎ2 [𝑢2 , 𝑢3 ] … ℎ20 [𝑢2 , 𝑢𝑚 ] ⎞
⎜ ℎ20 [𝑢 , 𝑢 ] ℎ20 [𝑢 , 𝑢 ] … ℎ20 [𝑢3 , 𝑢𝑚 ] ⎟⎟
det ⎜ 0 3 2 0 3 3 ≠0. (9.7)
⎜ … ⎟
⎜ℎ2 [𝑢 , 𝑢 ] ℎ2 [𝑢 , 𝑢 ] … ℎ20 [𝑢𝑚 , 𝑢𝑚 ]⎟⎠
⎝ 0 𝑚 2 0 𝑚 3

The expression of 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) w.r.t. the coordinates 𝑥1 , … , 𝑥𝑚 associated to the

vectors 𝑢1 , 𝑢2 , … , 𝑢𝑚 reads
𝜇1 𝜇2 𝜇𝑚
∑ [ ⏞⏞⏞ ] 𝜇
1 |𝜇| ⏞⏞⏞ ⏞⏞⏞ 𝜇
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚)(𝑥) = ℎ0 𝑢1 , 𝑢2 , … , 𝑢𝑚 𝑥1 1 … 𝑥𝑚𝑚 . (9.8)
𝜇∈ℕ𝑚
𝜇!
1≤|𝜇|≤𝑟

We now claim that

Lemma 9.1. If ℎ is non-steep at the origin on Γ𝑚 at some given order 𝑠𝑚 ≥ 2, then
⋂
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝑋11 (𝑟, 𝑠𝑚 , 𝑚) ∶= 𝜎
̂1 (𝑟, 𝑠𝑚 , 𝑚) 𝖲11 (𝑟, 𝑚) (9.9)

̂1
and 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) satisfies the 𝑠𝑚 -vanishing condition on Γ𝑚 on some curve 𝛾 ∈ Θ 𝑚
whose Taylor expansion at the origin has null linear terms.
Proof. Looking at (9.8), it is easy to check that the coefficients of 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) asso-
ciated to the family of indices {𝜛(𝑏, 𝑐)}𝑏,𝑐∈{2,…,𝑚},𝑏≤𝑐 introduced in (8.12) read
( ) 1
𝑝𝜛(𝑏,𝑐) = 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) 𝜛(𝑏,𝑐) = ℎ2 [𝑢 , 𝑢 ] , (9.10)
1 + 𝛿𝑏𝑐 0 𝑏 𝑐

96
where 𝛿𝑏𝑐 is the Kronecker delta.
By putting together expressions (9.7) - (9.10) with the definition of set 𝖲11 (𝑟, 𝑚) in
(8.18), one has that
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝖲11 (𝑟, 𝑚) . (9.11)
Since we have assumed ℎ is non-steep at the origin on Γ𝑚 at some order 𝑠𝑚 ≥ 2
then, by the discussions at section 6 and at paragraph 8.1, one has

̂ 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, (9.12)

which, together with (9.11), yields

⋂
̂ 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, 𝖲11 (𝑟, 𝑚) . (9.13)

We now claim that

Lemma 9.2.
̂𝑖 (𝑟, 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∉ Σ ∀ 𝑖 ∈ {2, … , 𝑚} . (9.14)

Proof. Suppose, by absurd, that 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ ̂𝑖 (𝑟, 𝑠𝑚 , 𝑚) for some 𝑖 ∈ {2, … , 𝑚}.
It is clear from Theorem 5.1 and (8.9) that, as 𝑠𝑚 ≥ 2, one has 𝜎 ̂𝑖 (𝑟, 𝑠𝑚 , 𝑚) ⊂ 𝜎
̂𝑖 (𝑟, 1, 𝑚),
and therefore Σ ̂𝑖 (𝑟, 𝑠𝑚 , 𝑚) ⊂ Σ̂𝑖 (𝑟, 1, 𝑚). This fact and the initial hypothesis imply
̂ 𝑖
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ (𝑟, 1, 𝑚), so that by Theorem 8.1, one must have

̂𝑖 (𝑟, 1, 𝑚) .
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝜎 (9.15)

Relation (9.15) implies that there must exist a curve 𝛾 ∈ Θ ̂ 𝑖 with values in Γ𝑚 such that
𝑚
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) satisfies the 1-vanishing condition on 𝛾. As it was shown in the proof of
Theorem C1 (see the discussion around formula (9.4)) this is equivalent to asking that
system
{
ℎ10 [𝑢1 ] = ℎ10 [𝑢2 ] = ⋯ = ℎ10 [𝑢𝑖 ] = ⋯ = ℎ10 [𝑢𝑚 ] = 0
𝑖≠1
ℎ20 [𝑢𝑖 , 𝑢1 ] = ℎ10 [𝑢𝑖 , 𝑢2 ] = ⋯ = ℎ10 [𝑢𝑖 , 𝑢𝑖 ] = ⋯ = ℎ10 [𝑢𝑖 , 𝑢𝑚 ] = 0
(9.16)
admits a solution, which contradicts (9.7).

Due to (9.13) and to Lemma 9.2, we then have that

⋂
̂1 (𝑟, 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ 𝖲11 (𝑟, 𝑚) . (9.17)
⋂
Moreover, by Theorem 8.2, the set 𝑋11 (𝑟, 𝑠, 𝑚) ∶= 𝜎 ̂1 (𝑟, 𝑠, 𝑚) 𝖲11 (𝑟, 𝑚) defined in
(8.44) is closed in 𝖲11 (𝑟, 𝑚) for the topology induced by (𝑟, 𝑚), whence one deduces
that actually
( 1 )⋂ 1 ⋂
𝑋11 (𝑟, 𝑠𝑚 , 𝑚) = closure 𝜎̂ (𝑟, 𝑠𝑚 , 𝑚) ̂1 (𝑟, 𝑠𝑚 , 𝑚)
𝖲1 (𝑟, 𝑚) = Σ 𝖲11 (𝑟, 𝑚) . (9.18)

97
Relations (9.17) and (9.18) together imply

𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝑋11 (𝑟, 𝑠𝑚 , 𝑚) . (9.19)

Therefore, by (9.19) and by the definition of 𝑋11 (𝑟, 𝑠𝑚 , 𝑚) in (8.44), there exists a
̂ 1 , with image in Γ𝑚 , on which the Taylor polynomial of the restriction
curve 𝛾 ∈ Θ 𝑚
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) satisfies the 𝑠𝑚 -vanishing condition. Namely, if 𝚊 = (𝑎21 , … , 𝑎𝑚1 ) ∈
𝑚−1
𝐵 (0, 𝒦 (𝑟, 𝑛, 𝑚)) indicates the linear coefficients of 𝛾 and 𝙹̂𝑠𝑚 ,𝛾,𝚊 its 𝑠𝑚 -truncation
written in the adapted coordinates, by (8.9) one must have

(𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹̂𝑠𝑚 ,𝛾,𝚊 ) ∈ Υ1 (𝑍

̂ 1 (𝑟, 𝑠𝑚 , 𝑚)) , (9.20)

that is, by (5.37),

̂
𝚀𝓁𝛼 (𝚃0,𝚊 (ℎ|Γ𝑚 , 𝑟, 𝑚), 𝚊, 𝙹̂𝑠𝑚 ,𝛾,𝚊 ) = 0 , 𝓁 = 1, … , 𝑚 , 𝛼 = 0, … , 𝑠𝑚 . (9.21)

In particular, as 𝑠𝑚 ≥ 2, due to Theorem 5.1 and to (7.5), the equations in (9.21) for
𝛼 = 0, 1 read
⎧ℎ1 [𝑣] = ℎ1 [𝑢1 ] = ⋯ = ℎ1 [𝑢𝑚 ] = 0
⎪ 20 0 0
⎨ℎ0 [𝑣, 𝑣] = ℎ20 [𝑣, 𝑢2 ] = ℎ20 [𝑣, 𝑢𝑚 ] = 0 (9.22)
⎪ 𝑚
⎩Span(𝑣, 𝑢2 , … , 𝑢𝑚 ) = Γ
where
∑
𝑚
𝑣 = 𝑢1 + 𝑎𝑖1 𝑢𝑖 (9.23)
𝑖=2
is the anisotropic vector associated to the curve 𝛾. Comparing (9.6) to (9.22), as the
Hessian of ℎ|Γ𝑚 has only one null eigenvalue we see that the vectors 𝑢1 and 𝑣 must be
parallel, but then (9.23) yields

𝑣 = 𝑢1 , 𝑎21 = ⋯ = 𝑎𝑚1 = 0 , (9.24)

so that by the arguments of subsection (5.2.3) the coordinates adapted to the curve 𝛾
coincide with the original coordinates.

We now recall that, due to Theorem 8.2, the form of the set
( ( ) )
⋂
1 ̂ 1 1 ̂1
̂ 1 (𝑟,𝑚) 𝒩
𝑌1 (𝑟, 𝑠𝑚 , 𝑚) ∶= Π𝑊 {𝚀𝚒𝛼 } 𝚒∈{1,…,𝑚} Υ (𝖲1 (𝑟, 𝑚) × 𝜗𝑚 (𝑠)) (9.25)
𝛼∈{0,…,𝑠𝑚 }

satisfying20
( )
𝑋11 (𝑟, 𝑠𝑚 , 𝑚) = Π(𝑟,𝑚) (𝔘1 )−1 𝑌11 (𝑟, 𝑠𝑚 , 𝑚) (9.26)
20 The transformation 𝔘1 was introduced in Remark 5.7.

98
can be explicitly computed - starting from the relations determining Υ1 (𝑍 ̂ 1 (𝑟, 𝑠𝑚 , 𝑚)) -
by the means of an algorithm involving only linear operations. By this fact, the form of
the set ⋂
𝒴11 (𝑟, 𝑠𝑚 , 𝑚) ∶= 𝑌11 (𝑟, 𝑠𝑚 , 𝑚) {(𝚙𝜇 , 𝚊) ∈ 𝑊
̂ 1 (𝑟, 𝑚)|𝚊 = 0} (9.27)
̂ 1 (𝑟, 𝑠𝑚 , 𝑚)). Then, due to Lemma 9.1
can also be deduced explicitly starting from Υ1 (𝑍
and to (9.26), one has that

ℎ non-steep at 0
at order 𝑠𝑚 on Γ𝑚 ⟹ 𝔘1 (𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚), 0) = (𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚), 0) ∈ 𝒴11 (𝑟, 𝑠𝑚 , 𝑚) .
(9.28)
Moreover, we observe the following facts:
𝑟,𝑠 ,𝑚
1. in section 7, the explicit expression of set 𝑛 𝑚 ⊂  ⋆ (𝑟, 𝑛)×ℝ(𝑚−1)𝑠𝑚 ×𝒱 1 (𝑚, 𝑛)
introduced in Corollary B2 is obtained by injecting into the explicit expression
for
{ }
Υ1 (𝑍
̂ 1 (𝑟, 𝑠𝑚 , 𝑚)) ∶= 𝒩 𝚀𝑖,𝛼 (𝚃0,𝚊 (𝙿𝚊 , 𝑟, 𝑚), 𝚊, 𝙹𝑠 ,𝛾,𝚊 )
𝑚 𝑖=1,…,𝑚 (9.29)
𝛼=0,…,𝑠𝑚

given in Theorem 5.1 the explicit form of the coefficients of 𝚃0,𝚊 (𝑄|Γ𝑚 , 𝑟, 𝑚) in
(7.1)-(7.3) (which we have written again in(9.1)), with 𝑄 any polynomial be-
longing to  ⋆ (𝑟, 𝑛), and by considering the vectors 𝑣, 𝑢2 , … , 𝑢𝑚 in (9.1) as free
parameters of 𝒱 1 (𝑚, 𝑛).

2. In the same way, for any 𝑄 ∈  ⋆ (𝑟, 𝑛), one can inject in the expressions deter-
mining 𝒴11 (𝑟, 𝑠𝑚 , 𝑚) the explicit form of the coefficients of 𝚃0,𝚊 (𝑄|Γ𝑚 , 𝑟, 𝑚), given
in (9.1), with the vectors 𝑣, 𝑢2 , … , 𝑢𝑚 considered as free parameters of 𝒱 1 (𝑚, 𝑛).
The resulting subset is indicated by

𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) ⊂  ⋆ (𝑟, 𝑛) × ℝ𝑚−1 × 𝒱 1 (𝑚, 𝑛) .

By the arguments above, and by the fact that the form of 𝒴11 (𝑟, 𝑠𝑚 , 𝑚) is obtained
̂ 1 (𝑟, 𝑠𝑚 , 𝑚)) through linear algorithms, we have
starting from the expression of Υ1 (𝑍
that the explicit expression of 𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) can be found starting from the expressions
𝑟,𝑠 ,𝑚
determining 𝑛 𝑚 by only performing linear operations.
Furthermore, the above arguments together with (9.28) and with formula (9.24)
yield that if system
{
(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛) , Span (𝑢1 , 𝑢2 , … , 𝑢𝑚 ) = Γ𝑚 ∈ 1 (ℎ, 𝑚, 𝑛)
(9.30)
(𝚃0 (ℎ, 𝑟, 𝑛), 0, 𝑢1 , 𝑢2 , … , 𝑢𝑚 ) ∈ 𝒜1 (𝑟, 𝑠𝑚 , 𝑛, 𝑚)

has no solution, then ℎ is steep at the origin with index 𝛼𝑚 ≤ 2𝑠𝑚 − 1 on any subspace
Γ𝑚 ∈ 1 (ℎ, 𝑚, 𝑛).

99
9.2.2 Case of a subspace belonging to 2 (ℎ, 𝑚, 𝑛)
As we did in paragraph 9.2.1, we choose a basis 𝑢1 , … , 𝑢𝑚 ∈ 𝚄(𝑚, 𝑛) spanning Γ𝑚 such
that 𝑢1 coincides with a normalized eigenvector of the Hessian associated to one of the
null eigenvalues. Then, we have

Lemma 9.3. If ℎ is non-steep at the origin on Γ𝑚 ∈ 2 (ℎ, 𝑚, 𝑛) at a given order 𝑠𝑚 ≥ 2,

up to suitably changing the order of the vectors 𝑢1 , … , 𝑢𝑚 one has
⋂
̂1 (𝑟, 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ 𝖲12 (𝑟, 𝑚) . (9.31)

Proof. By (9.10), and by the fact that the hessian of ℎ|Γ𝑚 has two or more null eigen-
values, we have that

⎛ ℎ2 [𝑢2 , 𝑢2 ] ℎ2 [𝑢2 , 𝑢3 ] … ℎ20 [𝑢2 , 𝑢𝑚 ] ⎞

⎜ ℎ02 [𝑢 , 𝑢 ] ℎ02 [𝑢 , 𝑢 ] … ℎ20 [𝑢3 , 𝑢𝑚 ] ⎟⎟
det ⎜ 0 3 2 0 3 3 =0, (9.32)
⎜ … ⎟
⎜ℎ2 [𝑢 , 𝑢 ] ℎ2 [𝑢 , 𝑢 ] … ℎ20 [𝑢𝑚 , 𝑢𝑚 ]⎟⎠
⎝ 0 𝑚 2 0 𝑚 3

hence 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ 𝖲12 (𝑟, 𝑚) following definition (8.20).

Moreover, ℎ is non-steep at the origin on Γ𝑚 at a given order 𝑠𝑚 ≥ 2 so that, by the
discussions at section 6 and at paragraph 8.1, one has

̂ 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, (9.33)

so that by the previous considerations we have

⋂
̂ 𝑠𝑚 , 𝑚)
𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ(𝑟, 𝖲12 (𝑟, 𝑚) . (9.34)

By the above expression and by (8.9), we have that there must exist 𝑖 ∈ {1, … , 𝑚} such
̂𝑖 (𝑟, 𝑠𝑚 , 𝑚) ⋂ 𝖲1 (𝑟, 𝑚). As it was already discussed in the proof of
that 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ
̂𝑖 (𝑟, 𝑠𝑚 , 𝑚) ⋂ 𝖲1 (𝑟, 𝑚) then the vector 𝑢𝑖 must satisfy
2
Theorem C1, if 𝚃0 (ℎ|Γ𝑚 , 𝑟, 𝑚) ∈ Σ 2
(9.16). If 𝑖 = 1, there is nothing else to prove. If 𝑖 ≠ 1, it suffices to interchange the
vector 𝑢1 with the vector 𝑢𝑖 .
⋆ 21 determining the semi-
Now, if for any 𝑄 ((∈  (𝑟, 𝑛) we ⋂
inject in )
the expressions
)
algebraic subset 𝔘 1 ̂ (𝑟, 𝑠𝑚 , 𝑚) 𝖲 (𝑟, 𝑚) × 𝜗̂ (1) ⊂ 𝑊
Σ 1 1 1 ̂ 1 (𝑟, 𝑚) the explicit ex-
2 𝑚
pression of the coefficients of the polynomial 𝚃0,𝚊 (𝑄|Γ𝑚 , 𝑟, 𝑚), and we let the vectors
𝑣, 𝑢2 , … , 𝑢𝑚 appearing in (9.1) vary in 𝒱 1 (𝑚, 𝑛), we obtain a set

𝒜2 (𝑟, 𝑠𝑚 , 𝑛, 𝑚) ⊂ (𝑟, 𝑛) × ℝ𝑚−1 × 𝒱 1 (𝑚, 𝑛) .

21 Contrary to the case studied in the previous paragraph, here these expressions cannot be found easily, in

general.

100
 Moreover, by Lemma 9.3, we have that if system

⎧(𝑎 , … , 𝑎 ) ∈ 𝐵 𝑚−1 (𝒦 )
⎪ 21 𝑚1
∑
⎪(𝑢1 , … , 𝑢𝑚 ) ∈ 𝚄(𝑚, 𝑛) , 𝑣 ∶= 𝑢1 + 𝑚 𝑖=2 𝑎𝑖1 𝑢𝑖
⎨ 𝑚
(9.35)
⎪ Span (𝑣, 𝑢2 , … , 𝑢𝑚 ) = Γ ∈ 2 (ℎ, 𝑚, 𝑛)
⎪(𝚃 (ℎ, 𝑟, 𝑛), 𝑎 , … , 𝑎 , 𝑣, 𝑢 , … , 𝑢 ) ∈ 𝒜 (𝑟, 𝑠 , 𝑛, 𝑚)
⎩ 0 21 𝑚1 2 𝑚 2 𝑚

has no solution, then ℎ is steep at the origin with index 𝛼𝑚 ≤ 2𝑠𝑚 − 1 on any subspace
Γ𝑚 ∈ 2 (ℎ, 𝑚, 𝑛).

9.3 Proof of Theorem C3

9.3.1 Construction of an atlas of analytic maps for the Grassmannian
It is well known that for any pair of positive integers 𝑘, 𝑛, with 𝑘 < 𝑛, the Grassmannian
𝙶(𝑘, 𝑛) has the structure of a projective algebraic variety (see e.g. [19], [52]). In this
subparagraph, for any positive integer 𝑛 ≥ 3 and for any 𝑚 ∈ {2, … , 𝑛 − 1} we will
construct a suitable atlas of analytic maps for 𝙶(𝑚, 𝑛) by exploiting classic arguments
of real-algebraic geometry.
Namely, we fix two integers 𝑛 ≥ 3, and 𝑚 ∈ {2, … , 𝑛 − 1} and we consider a
subset 𝐼 ∶= (𝑖1 , … , 𝑖𝑚 ) ⊂ {1, … , 𝑛} of cardinality 𝑚, as well as its complementary
𝐽 ∶= (𝑗1 , … , 𝑗𝑛−𝑚 ) of cardinality 𝑛 − 𝑚 in {1, … , 𝑛}.
We work in the euclidean space ℝ𝑛 , and we fix a bilinear, symmetric, non-degenerate
form 𝖡 ∶ ℝ𝑛 × ℝ𝑛 ⟶ ℝ. The Spectral Theorem ensures the existence of an orthonor-
mal basis - indicated by 𝑒1 , … , 𝑒𝑛 - that diagonalizes 𝖡. Namely, in the basis 𝑒1 , … , 𝑒𝑛
the form 𝖡 is represented by matrix

⎛ 𝛼1 0 0 … 0⎞
⎜ ⎟
0 𝛼2 0 … 0⎟
𝖡(𝑒1 ,…,𝑒𝑛 ) ∶= ⎜ , (9.36)
⎜… … … … …⎟
⎜0 0 0 … 𝛼𝑛 ⎟⎠
⎝
where 𝛼1 × ⋯ × 𝛼𝑛 ≠ 0.
We indicate by 𝐸𝐼 (resp. 𝐸𝐽 ) the 𝑚-dimensional subspace spanned by the vectors
(𝑒𝑖1 , … , 𝑒𝑖𝑚 ) (resp. the 𝑛 − 𝑚-dimensional subspace spanned by 𝑒𝑗1 , … , 𝑒𝑗𝑛−𝑚 ). One
clearly has ℝ𝑛 = 𝐸𝐼 ⊕ 𝐸𝐽 . We also denote by 𝑈𝐽 the subset of 𝙶(𝑚, 𝑛) containing the
𝑚-dimensional subspaces which are supplementary for 𝐸𝐽 .
𝑚 times
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
With this setting, we consider the cartesian product 𝐸𝐽𝑚 ∶=𝐸𝐽 × ⋯ × 𝐸𝐽 and we
have that
Lemma 9.4. The map

ℱ𝐽 ∶ 𝐸𝐽𝑚 ⟶ 𝑈𝐽 (𝑤1 , … , 𝑤𝑚 ) ⟼ Span(𝑒𝑖1 + 𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) (9.37)

101
is bijective.

Proof. We proceed by steps. In the first two steps, we check that definition (9.37) is
well-posed. In Steps 3 and 4 we prove injectivity and surjectivity.
Step 1. We claim that for any choice of (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 , the vectors (𝑒𝑖1 +
𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) are linearly independent. Infact, if by absurd there exist 𝑚 vectors
(𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 such that (𝑒𝑖1 + 𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) are not linearly independent, then
∑
there must be a vector 𝜆 = (𝜆1 , … , 𝜆𝑚 ) ∈ ℝ𝑚 ∖{0} satisfying 𝑚 𝓁=1 𝜆𝓁 (𝑒𝑖𝓁 + 𝑤𝓁 ) = 0,
∑𝑚 ∑𝑚
that is 𝓁=1 𝜆𝓁 𝑒𝑖𝓁 = − 𝓁=1 𝜆𝓁 𝑤𝑖𝓁 .
∑ ∑𝑚
As 𝑚 𝓁=1 𝜆𝓁 𝑒𝑖𝓁 ∈ 𝐸𝐼 , and − 𝓁=1 𝜆𝓁 𝑤𝑖𝓁 ∈ 𝐸 , and as ℝ𝑛 = 𝐸𝐼 ⊕ 𝐸𝐽 by con-
∑𝑚 𝐽 ∑
struction, by the assumptions one must have − 𝓁=1 𝜆𝓁 𝑤𝑖𝓁 = 𝑚 𝓁=1 𝜆𝓁 𝑒𝑖𝓁 = 0. The
previous relation - together with the fact that the vectors 𝑒𝑖1 , … , 𝑒𝑖𝑚 are linearly inde-
pendent by hypothesis, yields 𝜆 = 0, in contradiction with the initial assumption 𝜆 ≠ 0.
Step 2. We claim that for any choice of (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 one has Span (𝑒𝑖1 +
𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) ∈ 𝑈𝐽 .
By absurd, we suppose that for some (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 there exist two non-
∑
zero vectors 𝑢 ∈ 𝐸𝐽 , 𝜎 = (𝜎1 , … , 𝜎𝑚 ) ∈ ℝ𝑚 , verifying 𝑢 = 𝑚 𝓁=1 𝜎𝓁 (𝑒𝑖𝓁 + 𝑤𝓁 ),
∑𝑚 ∑𝑚 ∑
that is 𝓁=1 𝜎𝓁 𝑒𝑖𝓁 = 𝑢 − 𝓁=1 𝜎𝓁 𝑤𝓁 . By construction, one has 𝑚 𝓁=1 𝜎𝓁 𝑒𝑖𝓁 ∈ 𝐸𝐼 ,
∑
and 𝑢 − 𝑚 𝜎 𝑤 ∈ 𝐸 . Hence, as ℝ𝑛 = 𝐸 ⊕ 𝐸 , the previous formula yields
∑ 𝓁=1 𝓁 𝓁 ∑ 𝐽 𝐼 𝐽
𝑢− 𝑚 𝓁=1 𝜎𝓁 𝑤𝓁 =
𝑚
𝓁=1 𝜎𝓁 𝑒𝑖𝓁 = 0, which in turn implies 𝜎 = 0, as the vectors
(𝑒𝑖1 , … , 𝑒𝑖𝑚 ) are linearly independent.
Hence, one has 𝐸𝐽 ∩ Span (𝑒𝑖1 + 𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) = {0}. Therefore, since
( )
dim 𝐸𝐽 = 𝑛 − 𝑚 and dim Span (𝑒𝑖1 + 𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) = 𝑚 (the vectors 𝑒𝑖1 +
𝑤1 , … , 𝑒𝑖𝑚 +𝑤𝑚 are linearly independent by Step 1), the subspace given by Span (𝑒𝑖1 +
𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 ) is supplementary to 𝐸𝐽 and thus belongs to 𝑈𝐽 .
Step 3. We prove that ℱ𝐽 is injective. By absurd, we suppose that there exists a
subspace in 𝑈𝐽 which has two different pre-images. Namely, we suppose that there exist
vectors (𝑢1 , … , 𝑢𝑚 ) ∈ 𝐸𝐽𝑚 , and (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 , with (𝑢1 , … , 𝑢𝑚 ) ≠ (𝑤1 , … , 𝑤𝑚 ),
satisfying ℱ𝐽 (𝑢1 , … , 𝑢𝑚 ) = ℱ𝐽 (𝑤1 , … , 𝑤𝑚 ). Hence, as 𝑒𝑖1 + 𝑢1 , … , 𝑒𝑖𝑚 + 𝑢𝑚 and
𝑒𝑖1 + 𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 generate the same subspace, for any 𝓁 ∈ {1, … , 𝑚} there must
exist 𝛽 𝓁 = (𝛽1𝓁 , … , 𝛽𝑚𝓁 ) ∈ ℝ𝑚 ∖{0} such that

∑
𝑚
𝑒𝑖𝓁 + 𝑤𝓁 = 𝛽𝑘𝓁 (𝑒𝑖𝑘 + 𝑢𝑘 ) , (9.38)
𝑘=1
that is
∑
𝑚 ∑
𝑚
𝑒𝑖𝓁 − 𝛽𝑘𝓁 𝑒𝑖𝑘 = 𝛽𝑖𝓁 𝑢𝑖 − 𝑤𝓁 . (9.39)
𝑘=1 𝑖=1
∑ ∑𝑚 𝓁
By construction one has 𝑒𝑖𝓁 − 𝑚 𝓁
𝑘=1 𝛽𝑘 𝑒𝑖𝑘 ∈ 𝐸𝐼 and 𝑖=1 𝛽𝑖 𝑢𝑖 − 𝑤𝓁 ∈ 𝐸𝐽 , so due
∑
to (9.39) and to the direct sum ℝ = 𝐸𝐼 ⊕ 𝐸𝐽 one infers 𝑚
𝑛 𝓁
𝑖=1 𝛽𝑖 𝑢𝑖 − 𝑤𝓁 = 𝑒𝑖𝓁 −
∑𝑚 𝓁
𝑘=1 𝛽𝑘 𝑒𝑖𝑘 = 0. Since the vectors 𝑒𝑖1 , … , 𝑒𝑖𝑚 are linearly independent, we finally

102
obtain
𝛽𝑘𝓁 = 𝛿𝓁𝑘 . (9.40)
where 𝛿𝓁𝑘 is the Kronecker symbol. Formulas (9.38) and (9.40) together imply that

𝑒𝑖𝓁 + 𝑤𝓁 = 𝑒𝑖𝓁 + 𝑢𝓁 ⟺ 𝑤𝓁 = 𝑢𝓁 ∀𝓁 ∈ {1, … , 𝑚} , (9.41)

in contradiction with the hypothesis (𝑢1 , … , 𝑢𝑚 ) ≠ (𝑤1 , … , 𝑤𝑚 ).

Step 4. We prove that ℱ𝐽 is surjective. Consider a subspace 𝑉 ∈ 𝑈𝐽 . Since 𝑉 is
supplementary of 𝐸𝐽 , one has the direct sum ℝ𝑛 = 𝑉 ⊕ 𝐸𝐽 , and for any 𝓁 ∈ {1, … , 𝑚}
there exist unique vectors (𝑣𝓁 , 𝑤𝓁 ) ∈ 𝑉 × 𝐸𝐽 such that 𝑒𝑖𝓁 = 𝑣𝓁 − 𝑤𝓁 . Hence, for
any 𝓁 ∈ {1, … , 𝑚} there exists a unique choice of vectors (𝑤1 , … , 𝑤𝑚 ) satisfying
𝑣𝓁 = 𝑒𝑖𝓁 + 𝑤𝓁 . The vectors 𝑣1 , … , 𝑣𝑚 belong to 𝑉 by construction, and are linearly
independent by Step 1. Therefore, since dim 𝑉 = 𝑚 by hypothesis (𝑉 ∈ 𝑈𝐽 ), one has
Span(𝑣1 , … , 𝑣𝑚 ) = 𝑉 .

We indicate by 𝒥 𝑛−𝑚 the subsets of cardinality 𝑛 − 𝑚 in {1, … , 𝑛}. One has the
following covering of the 𝑚-dimensional Grassmannian:
⋃
𝙶(𝑚, 𝑛) = 𝑈𝐽 . (9.42)
𝐽 ∈𝒥 𝑛−𝑚

By construction, any vector 𝑤𝓁 ∈ 𝐸𝐽 can be expressed uniquely as

∑
𝑛−𝑚
𝑤𝓁 = 𝑀𝓁𝑘 𝑒𝑗𝑘 , (9.43)
𝑘=1

where (𝑀𝓁𝑘 ) 𝓁=1,…,𝑚 is a real 𝑚 × (𝑛 − 𝑚) matrix. By (9.42), and by Lemma 9.4, there
𝑘=1,…,𝑛−𝑚
exists an atlas sending 𝙶(𝑚, 𝑛) to the open union
⋃ ⋃
ℱ𝐽−1 (𝑈𝐽 ) = 𝐸𝐽𝑚 ⊂ ℝ𝑚×(𝑛−𝑚) . (9.44)
𝐽 ∈𝒥 𝑛−𝑚 𝐽 ∈𝒥 𝑛−𝑚

9.3.2 Proof of Theorem C3

Taking (9.42) into account, we fix 𝐽 ∈ 𝒥 𝑛−𝑚 together with its associated sets 𝐸𝐽 , 𝑈𝐽 .
Let 𝑉 be a 𝑚-dimensional subspace belonging to 𝑈𝐽 . By Lemma 9.4, one has

𝑉 = Span {𝑒𝑖1 + 𝑤1 , … , 𝑒𝑖𝑚 + 𝑤𝑚 } (9.45)

for a unique (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 .

Now, as 𝙶≥1 (𝑚, 𝑛) contains those subsets of 𝙶(𝑚, 𝑛) on which the restriction of the
bilinear form 𝖡 has at least one null eigenvalue, 𝑉 ∈ 𝙶≥1 (𝑚, 𝑛) if and only if 𝖡 is

103
degenerated on 𝑉 . Namely, taking (9.45) into account, 𝑉 ∈ 𝙶≥1 (𝑚, 𝑛) iff there exists a
∑
vector 𝑣 = 𝑚 ′
𝓁=1 𝑣𝓁 (𝑒𝑖𝓁 + 𝑤𝓁 ) ∈ 𝑉 such that for all 𝓁 ∈ {1, … , 𝑚} one has

∑
𝑚
𝖡(𝑣, 𝑒𝑖𝓁′ + 𝑤𝓁′ ) = 𝑣𝓁 𝖡(𝑒𝑖𝓁 + 𝑤𝓁 , 𝑒𝑖𝓁′ + 𝑤𝓁′ )
𝓁=1
∑𝑚 ( )
= 𝑣𝓁 𝖡(𝑒𝑖𝓁 , 𝑒𝑖𝓁′ ) + 𝖡(𝑤𝓁 , 𝑤𝓁′ ) (9.46)
𝓁=1
∑
𝑚
= 𝑣𝓁 ′ 𝛼𝑖𝓁′ + 𝑣𝓁 𝖡(𝑤𝓁 , 𝑤𝓁′ ) .
𝓁=1

To pass from the first to the second line in the above expression, we have taken into
account the fact that (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 , that 𝐸𝐽 = Span(𝑒𝑗1 , … , 𝑒𝑗𝑛−𝑚 ), and that the
form 𝖡 is diagonal in the basis 𝑒1 , … , 𝑒𝑚 ; in the last passage, we have considered (9.36).
Setting

⎛ 1 1 ⎞
⎜ 𝛼 𝖡(𝑤1 , 𝑤1 ) … 𝛼𝑖1
𝖡(𝑤1 , 𝑤𝑚 ) ⎟
⎜ 𝑖1 ⎟
ℳ𝖡 = ℳ𝖡 (𝑤1 , … , 𝑤𝑚 ) ∶= ⎜ … … … ⎟ (9.47)
⎜ 1 1 ⎟
⎜ 𝛼 𝖡(𝑤𝑚 , 𝑤1 ) … 𝛼𝑖𝑚
𝖡(𝑤𝑚 , 𝑤𝑚 )⎟
⎝ 𝑖𝑚 ⎠

it is plain to check that (9.46) can be rewritten in the form

ℳ𝖡 𝑣 = −𝑣 (9.48)

that is, we are asking for −1 to be an eigenvalue of ℳ𝖡 , hence (9.48) is equivalent to

det(ℳ𝖡 + 𝕀𝑚 ) = 0 , (9.49)

where 𝕀𝑚 is the 𝑚 × 𝑚 identity matrix.

Since ℳ𝖡 depends on (𝑤1 , … , 𝑤𝑚 ) ∈ 𝐸𝐽𝑚 and since (𝑤1 , … , 𝑤𝑚 ) are in bijec-
tion with ℝ𝑚(𝑛−𝑚) by (9.43), the quantity det(ℳ𝖡 + 𝕀𝑚 ) determines a polynomial map
ℝ𝑚(𝑛−𝑚) ⟶ ℝ.
If, by absurd, det(ℳ𝖡 + 𝕀𝑚 ) is the null polynomial, then relation (9.49) holds on the
whole inverse image ℱ𝐽−1 (𝑈𝐽 ). In particular, we observe that (𝑤1 = 0, … , 𝑤𝑚 = 0) ∈
ℱ𝐽−1 (𝑈𝐽 ) because ℱ𝐽 (0, … , 0) = Span{𝑒𝑖1 , … , 𝑒𝑖𝑚 } = 𝐸𝐼 and 𝐸𝐼 is supplementary of
𝐸𝐽 by construction. Therefore, by the above reasonings one must have

det(ℳ𝖡 (0, … , 0) + 𝕀𝑚 ) = det 𝕀𝑚 = 0

which is clearly false. Consequently, the polynomial function det(ℳ𝖡 (𝑤1 , … , 𝑤𝑚 )+𝕀𝑚 )
is not identically null over ℝ𝑚(𝑛−𝑚) and, due to Lemma E.3, its zero set is contained in
a submanifold of codimension one in ℝ𝑚(𝑛−𝑚) . Hence, also the subset of degenerated

104
subspaces of 𝑈𝐽 is contained in a submanifold of codimension one in 𝙶(𝑚, 𝑛). The
reasoning can be repeated for all 𝐽 ∈ 𝒥 𝑛−𝑚 . As, by its definition and by (9.42), 𝙶1 (𝑚, 𝑛)
is the finite union over 𝐽 ∈ 𝒥 𝑛−𝑚 of the degenerated subspaces of 𝑈𝐽 , we have that
𝙶1 (𝑚, 𝑛) is contained in a submanifold of codimension one in 𝙶(𝑚, 𝑛). This proves point
1 of the statement.
With the setting above, for any fixed 𝐽 ∈ 𝒥 𝑛−𝑚 we observe that a 𝑚-dimensional
subspace 𝑉 ′ ∈ 𝑈𝐽 belongs to 𝙶≥2 (𝑚, 𝑛) if and only if there exist at least two linearly
independent vectors 𝑣, 𝑢 ∈ 𝑉 ′ satisfying (9.48). In particular, the subset 𝙶≥2 (𝑚, 𝑛) ∩ 𝑈𝐽
of "doubly-degenerated" subspaces of 𝑈𝐽 is contained in the intersection of 𝙶≥1 (𝑚, 𝑛) ∩
𝑈𝐽 with the set
𝑇𝐽 ∶= {𝑊 ∈ 𝑈𝐽 |Δ(𝑃ℳ𝖡 ) = 0} (9.50)
where 𝑃ℳ𝖡 is the characteristic polynomial of matrix ℳ𝖡 , and Δ(𝑃ℳ𝖡 ) is its discrim-
inant. By the above arguments and by (9.42), point 2 of the statement follows if we
manage to prove that 𝑇𝐽 is contained in a submanifold of codimension one in 𝙶(𝑚, 𝑛).
The rest of the proof will be devoted to demonstrating this property.
Clearly, by the same arguments used in the proof of point 1 of the statement, Δ(𝑃ℳ𝖡 )
is a polynomial function over ℝ𝑚(𝑛−𝑚) . If, by absurd, Δ(𝑃ℳ𝖡 ) is identically zero in
ℝ𝑚(𝑛−𝑚) , then in particular it must be zero over the open set ℱ𝐽−1 (𝑈𝐽 ).
Now, choose 𝑚 numbers 𝑗1 , … , 𝑗𝑚 ∈ 𝐽 , and consider the vectors
√ √ √
| 𝛼𝑖 | | 𝛼𝑖 | | 𝛼𝑖 |
| | | | | |
𝑤′1 ∶= | |𝑒𝑗1 , 𝑤′2 ∶= 2 | |𝑒𝑗2 , … , 𝑤′𝑚 ∶= 𝑚 | 𝑚 |𝑒𝑗𝑚 , (9.51)
1 2
| 𝛼𝑗 | | 𝛼𝑗 | | 𝛼𝑗 |
| 1| | 2| | 𝑚|
which are well defined by the fact that 𝖡 is non-degenerate (see (9.36)). It is clear that
(𝑤′1 , … , 𝑤′𝑚 ) ∈ 𝐸𝐽𝑚 , so that by Lemma (9.4) one has ℱ𝐽 (𝑤′1 , … , 𝑤′𝑚 ) ∈ 𝑈𝐽 . As 𝖡 is
diagonal for the basis 𝑒1 , … , 𝑒𝑛 by hypothesis, matrix ℳ𝖡 (𝑤′1 , … , 𝑤′𝑚 ) in (9.36) reads

⎛sgn(𝛼𝑖 ) sgn(𝛼𝑗 ) 0 0 … 0 ⎞
⎜ 1
0
1
2 sgn(𝛼𝑖2 ) sgn(𝛼𝑗2 ) 0 … 0 ⎟
⎜ ⎟ , (9.52)
⎜ … … … … … ⎟
⎜ 0 0 0 … 𝑚 sgn(𝛼𝑖𝑚 ) sgn(𝛼𝑗𝑚 )⎟⎠
⎝

and it is clear that the discriminant of the characteristic polynomial of this matrix cannot
be zero, in contradiction with the fact that Δ(𝑃ℳ𝖡 ) = 0 on the whole set ℱ𝐽−1 (𝑈𝐽 ).
Therefore, the polynomial function Δ(𝑃ℳ𝖡 ) is not identically null over ℝ𝑚(𝑛−𝑚) and its
zero set is contained in a submanifold of codimension one in ℝ𝑚(𝑛−𝑚) , by Lemma E.3.
This proves that 𝑇𝐽 is contained in a submanifold of codimension one in 𝙶(𝑚, 𝑛), which
concludes the proof.

105
Tools of real-algebraic geometry
The goal of this appendix is to provide the reader with an overview of some standard
results of real-algebraic geometry that are used throughout the present work. The inter-
ested reader can find a complete exposition in [19] and [15].

A Semi-algebraic sets and semi-algebraic functions

Definition A.1. A set 𝐴 ⊂ ℝ𝑛 is said to be semi-algebraic if it can be written in the
form
⋃𝑠
{𝑅𝑖 (𝑥) = 0, 𝑄𝑖1 (𝑥) < 0, … , 𝑄𝑖 𝑟𝑖 (𝑥) < 0} , (A.1)
𝑖=1
where 𝑅𝑖 , 𝑄𝑖1 , ..., 𝑄𝑖 𝑟𝑖 ∈ ℝ[𝑥].

Remark A.1. If only equalities are present in (A.1), 𝐴 is said to be algebraic.

It is clear that the polynomials generating a given semi-algebraic set 𝐴 are not
uniquely determined, nor is their number. However, one can introduce a unique quantity
associated to a semi-algebraic set, namely

Definition A.2 (Diagram). For any semi-algebraic set 𝐴 ⊂ ℝ𝑛 , we denote by

− 𝑘1 (𝐴) the minimal number of polynomials 𝑅𝑖 (𝑋), 𝑄𝑖𝑗 (𝑋) ∈ ℝ[𝑋] that are nec-
essary in order to determine 𝐴 as in (A.1).
∑ ̂𝑖 ) + ∑𝑠 ∑𝑟𝑖 deg(𝑄
− 𝑘2 (𝐴) the minimal value that the sum 𝑠𝑖=1 deg(𝑅 𝑖=1 𝑗=1
̂𝑖𝑗 ) can
̂𝑖 (𝑋), 𝑄
attain, with 𝑅 ̂𝑖𝑗 (𝑋) ∈ ℝ[𝑋] determining 𝐴 as in (A.1).

Following [24] (Def. 9.1), we call diagram of 𝐴 the quantity

diag(𝐴) ∶= 𝑘1 (𝐴) + 𝑘2 (𝐴) . (A.2)

Semi-algebraic sets are stable under projection, namely

Theorem A.1. (Tarski and Seidenberg, quantitative version)

Take 𝑛, 𝑚 > 0 and let 𝐴 ⊂ ℝ𝑛 × ℝ𝑚 be a semi-algebraic set. We indicate by
Π𝑛 ∶ ℝ𝑛 × ℝ𝑚 ⟶ ℝ𝑚 , (𝑥, 𝑦) ⟼ 𝑥 the projector onto the first 𝑛 coordinates. Then,
the set Π𝑛 (𝐴) is semi-algebraic and its diagram depends only on diag(𝐴), 𝑚 and 𝑛.

The classic versions of the Theorem of Tarski and Seidenberg do not usually make
any reference to the diagram of the projected set. The statement given here can be found
in [24] (Proposition 9.2) and its proof is contained in [14].
The Theorem of Tarski and Seidenberg is fundamental in order to demonstrate the
following results (see [19] for proofs)

106
Proposition A.1. The complementary of a semi-algebraic set 𝐴 ⊂ ℝ𝑛 is semi-algebraic,
and its diagram depends only on the diagram of 𝐴.
Proposition A.2. The closure, the interior and the boundary of a semi-algebraic set
𝐴 ⊂ ℝ𝑛 are semi-algebraic and their diagrams depend only on the diagram of 𝐴.
Proposition A.3. Let 𝐴 be a semi-algebraic set of ℝ𝑛 . Then, indicating with 𝐴 the
closure of 𝐴, one has that diag(𝐴) depends only on diag(𝐴), and

dim 𝐴 = dim(𝐴) .

The notion of semi-algebraicness can be easily extended to functions by making

reference to their graphs, namely
Definition A.3. Let 𝐴 ⊂ ℝ𝑛 and 𝐵 ∈ ℝ𝑚 be semi-algebraic sets. A map 𝜑 ∶ 𝐴 ⟶ 𝐵
is said to be semi-algebraic if graph(𝜑) is a semi-algebraic set of ℝ𝑛 × ℝ𝑚 .
Semi-algebraic functions are piecewise algebraic, namely one has
Proposition A.4. Let 𝐴 be a semialgebraic subset 𝐴 ⊂ ℝ𝑛 and 𝜑 ∶ 𝐴 ⟶ ℝ be a semi-
algebraic function of diagram 𝑑 > 0. There exist a positive integer 𝑀(𝑑), a partition
of 𝐴 into a finite number of semi-algebraic sets 𝐴𝑖 , 𝑖 = 1, … , 𝑚, with 𝑚 ≤ 𝑀(𝑑), and
for every value of 𝑖 there exists a polynomial 𝑆𝑖 (𝑋, 𝑌 ) in 𝑛 + 1 variables such that, for
every 𝑥 in 𝐴𝑖 , 𝑆𝑖 (𝑥, 𝑌 ) is not identically zero and solves 𝑆𝑖 (𝑥, 𝜑(𝑥)) = 0.
Proof. Except for the existence of the bound 𝑀 = 𝑀(𝑑), the proof can be found in
ref. [19] (Lemma 2.6.3). Assume, by absurd, that the bound 𝑀(𝑑) does not exist; then,
one can find a sequence {𝜑𝑗 }𝑗∈ℕ of semi-algebraic functions of 𝐴 with diagram 𝑑 such
that, for each fixed 𝑗 ∈ ℕ, the Proposition holds with the minimal number of pieces
in the partition of 𝐴 being equal to 𝑚𝑗 , and 𝑚𝑗 ⟶ +∞. In particular, one can write
𝑚𝑗
𝐴 = ⊔𝑖=1 𝐴𝑖,𝑗 for any 𝑗 ∈ ℕ, and there exist polynomials 𝑆𝑖,𝑗 (𝑋1 , … , 𝑋𝑛 , 𝑌 ) with the
required properties. On the one hand, for any given 𝑗 ∈ ℕ and 𝑖 ∈ {1, … , 𝑚𝑗 }, by
decomposition (A.1), one has that graph(𝜑𝑗 |𝐴𝑖,𝑗 ) is the finite union of sets of the kind
{ }
(𝑥, 𝑦) ∈ 𝐴𝑖,𝑗 × ℝ | 𝑆𝑖,𝑗 (𝑥, 𝑦) = 0, 𝑄𝑖1,𝑗 (𝑥, 𝑦) < 0, … , 𝑄𝑖𝑟𝑖 ,𝑗 (𝑥, 𝑦) < 0 (A.3)

for some polynomials 𝑄𝑖1,𝑗 , … , 𝑄𝑖𝑟𝑖 ,𝑗 ∈ ℝ[𝑋1 , … , 𝑋𝑛 , 𝑌 ]. On the other hand, one has
the disjoint union
𝑚
⨆𝑗
graph(𝜑𝑗 ) = graph(𝜑𝑗 |𝐴𝑖 ,𝑗 ) , (A.4)
𝑖=1
𝑚𝑗
which is a consequence of the fact that 𝐴 = ⊔𝑖=1 𝐴𝑖,𝑗 is a partition.
By Def. A.2, formulas (A.3)-(A.4), and the fact that 𝑚𝑗 is the minimal number of
pieces in the partition of 𝐴 for 𝜑𝑗 , one has diag(𝜑𝑗 ) ≥ 𝑚𝑗 and, for sufficiently high 𝑗, one
has also 𝑚𝑗 > 𝑑 since 𝑚𝑗 ⟶ +∞, in contradiction with the hypothesis diag(𝜑𝑗 ) = 𝑑
for any 𝑗 ∈ ℕ. This concludes the proof.

107
 An immediate consequence of Proposition A.4 is the following

Corollary A.1. Let 𝐴 ⊂ ℝ be an interval (finite or infinite) and 𝜑 ∶ 𝐴 ⟶ ℝ be a

semi-algebraic function of diagram 𝑑 > 0. There exist a positive integer 𝑀(𝑑) and
an interval  of length |𝐴|∕𝑀(𝑑) over which the function 𝜑 is algebraic, namely there
exists a polynomial 𝑆(𝑋, 𝑌 ) in 𝑛 + 1 variables such that, for every 𝑥 in , 𝑆(𝑥, 𝑌 ) is
not identically zero and solves 𝑆(𝑥, 𝜑(𝑥)) = 0.

Among semi-algebraic functions, an important class is that of Nash functions:

Definition A.4. Let 𝐴 be an open semi-algebraic subset of ℝ𝑛 . A semi-algebraic func-

tion 𝜑 ∶ 𝐴 ⟶ ℝ belonging to the 𝐶 ∞ class is said to be a Nash function. The set of
Nash functions on 𝐴 is a ring under the usual operations of sum and function multipli-
cation.

Moreover, if we define analytic-algebraic functions as those real-analytic functions

𝑓 defined on an open semi-algebraic set 𝐴 ⊂ ℝ𝑛 and satisfying 𝑃 (𝑥, 𝑓 (𝑥)) = 0 for
some polynomial 𝑃 of 𝑛 + 1 variables and for all 𝑥 ∈ 𝐴, it turns out that

Proposition A.5 (ref. [19], Prop. 8.1.8). A function 𝜑 ∶ 𝐴 ⟶ ℝ is Nash on 𝐴 if and

only if it is analytic-algebraic.

Another important property of more general complex analytic-algebraic functions

is stated in the following

Proposition A.6. Let 𝐷 ⊂ ℂ be an open, bounded domain. An analytic-algebraic

function 𝑓 ∶ 𝐷 ⟶ ℂ, whose graph solves a polynomial 𝑆 ∈ ℂ[𝑧, 𝑤] of degree 𝑘 ∈ ℕ,
is 𝑘-valent: that is, if 𝑓 is not constant then each value of Im(𝑓 ) is the image of at most
𝑘 points in 𝐷.

Proof. Assume, by absurd, that 𝑓 is non-constant and that there exists 𝑤0 ∈ Im(𝑓 )
which is the image of at least 𝑝 > 𝑘 points in 𝐷. The polynomial 𝑆 𝑤0 (𝑧) ∶= 𝑆(𝑧, 𝑤0 )
would admit 𝑝 > 𝑘 roots while deg(𝑆 𝑤0 ) ≤ 𝑘 by hypothesis. The Fundamental The-
orem of Algebra ensures that 𝑆 𝑤0 must be identically zero and one has the factoriza-
tion 𝑆(𝑧, 𝑤) = (𝑤 − 𝑤0 )𝛼 𝑆(𝑧,̂ 𝑤), where 𝛼 ∈ {1, ..., 𝑘}, while 𝑆̂ cannot be divided by
(𝑤−𝑤0 ) in ℂ[𝑧, 𝑤]. Since 𝑓 is analytic and not constant, then the preimage 𝑓 −1 ({𝑤0 })
̂ 𝑓 (𝑧)) = 0 out of 𝑓 −1 ({𝑤0 }). By conti-
is a finite set and the graph of 𝑓 must fulfill 𝑆(𝑧,
̂ 𝑓 (𝑧)) = 0 on the whole domain of definition of 𝑓 since 𝑓 −1 ({𝑤0 }) is
nuity, one has 𝑆(𝑧,
finite. But deg 𝑆̂𝑤0 ≤ 𝑘, with 𝑆̂𝑤0 (𝑧) ∶= 𝑆(𝑧,
̂ 𝑤0 ), and 𝑆̂𝑤0 admits more than 𝑘 roots,
hence the previous argument ensures that 𝑆̂ can be divided by (𝑤−𝑤0 ), in contradiction
with the construction.

Semi-algebraicness is preserved by composition and inversion of semi-algebraic

maps. In the following propositions, 𝐴 ⊂ ℝ𝑛 and 𝐵 ⊂ ℝ𝑚 are supposed to be semi-
algebraic sets.

108
Proposition A.7. Let 𝜑 ∶ 𝐴 ⟶ 𝐵 a semi-algebraic map. If 𝑆 ⊂ 𝐴 and 𝑇 ⊂ 𝜑(𝐴) are
semi-algebraic, so are 𝜑(𝑆) and the inverse image 𝜑−1 (𝑇 ). Moreover, their diagrams
depend only on the diagram of graph(𝜑).
Proposition A.8. Let 𝑓 ∶ 𝐴 ⟶ 𝐵 and 𝑔 ∶ 𝐵 ⟶ 𝐶 be two semi-algebraic functions.
Then 𝑓 ◦𝑔 is semi-algebraic and the diagram of its graph depends only on the diagram
of graph(𝑓 ) and on the diagram of graph(𝑔).
Proposition A.9. Let 𝑓 ∶ 𝐴 ⟶ 𝐵 be an injective semi-algebraic function. Then, its
inverse 𝑓 −1 ∶ 𝑓 (𝐴) ⟶ 𝐴 is semi-algebraic and the diagram of its graph depends only
on the diagram of graph(𝑓 ).
Proposition A.10. Let 𝐼 ⊂ ℝ be an open interval and 𝑓 ∶ 𝐼 ⟶ ℝ be a semi-algebraic
function differentiable in 𝐼. Then its derivative 𝑓 ′ is a semi-algebraic function and its
diagram only depends on the diagram of graph(𝑓 ).
We refer to [19] for the proofs of these statements. The dependence of the diagrams
on the diagram of the initial function is, once again, a consequence, of the quantitative
version A.1 of the Theorem of Tarski and Seidenberg.
Finally, we give the following statement, which will prove to be helpful in our work
Proposition A.11. (see e.g. ref. [39], pag. 23-24) Let 𝑓 ∶ 𝐴 ⟶ ℝ and 𝑔 ∶ 𝐴 ⟶ ℝ𝑚
be semi-algebraic functions and suppose that 𝑓 is bounded from below. Then

𝜑 ∶ 𝑔(𝐴) ⟶ ℝ 𝑦⟼ inf 𝑓 (𝑥)

𝑥∈𝑔 −1 (𝑦)

is semi-algebraic and the diagram of its graph depends only on the diagrams of graph(𝑓 )
and graph(𝑔).

B Analytic reparametrization of semi-algebraic sets

Generally speaking, the reparametrization of a semi-algebraic set 𝐴 is a subdivision of
𝐴 into semi-algebraic pieces 𝐴𝑗 each of which is the image of a semi-algebraic function
of the unit cube. On the one hand, it is possible to cover the whole of 𝐴 if one asks for the
covering functions to be of finite regularity, with a uniform control on their derivatives
(see [35]). On the other hand, if one requires analyticity of the covering functions
together with a uniform control on their derivatives, it is only possible to cover 𝐴 up to
a "small" subset.
Hereafter, we state this result only in the case we need, that is for reparametrizations
of graphs of algebraic functions, referring to [59] for the general theory.
It is known that algebraic functions can only have two type of complex singularities:
ramification points and poles (where the function may also ramify). If we denote by 𝑑
the diagram of an algebraic function, the number of its complex singularities is bounded
by a quantity depending only on 𝑑 (see e.g. [13]). It is exactly the neighborhoods of
these singularities that cannot be analytically covered.

109
Definition B.1. Let 𝛿 > 0 and 𝑔 ∶ 𝐼 ∶= [−1, 1] ⟶ ℝ be an algebraic function. An
analytic 𝛿-reparametrization of 𝑔 consists of

− A finite number 𝑁 of open subintervals 𝑈𝑖 of 𝐼, with length(𝑈𝑖 ) ≤ 2𝛿 for any

𝑖 = 1, ..., 𝑁.

− A partition of 𝐼∖ ∪𝑁 𝑈 into a finite number of subsegments Δ𝑗 , 𝑗 = 1, ..., 𝑀,

𝑖=1 𝑖
together with a collection of real-analytic maps 𝜓𝑗 ∶ 𝐼 ⟶ Δ𝑗 such that for any
𝑗 ∈ {1, ..., 𝑁}

1. 𝜓𝑗 is an affine reparametrization of the segment Δ𝑗 .

2. 𝜓𝑗 and 𝑔◦𝜓𝑗 are both holomorphic in 3 ∶= {𝑧 ∈ ℂ ∶ |𝑧| ≤ 3} .
3. 𝜓𝑗 and 𝑔◦𝜓𝑗 both satisfy a Bernstein inequality, that is

max |𝜓𝑗 (𝑧) − 𝜓𝑗 (0)| ≤ 1 max |𝑔◦𝜓𝑗 (𝑧) − 𝑔◦𝜓𝑗 (0)| ≤ 1 . (B.1)
𝑧∈3 𝑧∈3

Theorem B.1. (Yomdin, [59] Th. 3.2) Let 𝑑 be a positive integer. There exist constants
𝚈1 = 𝚈1 (𝑑) and 𝚈2 = 𝚈2 (𝑑) such that for each algebraic function 𝑔(𝑥) of diagram
𝑑 defined on 𝐼 satisfying 0 ≤ 𝑔(𝑥) ≤ 1 and for each 𝛿 > 0 there is an analytic 𝛿-
reparametrization of 𝑔 with the number 𝑁 of the removed intervals bounded by 𝚈1 and
the number 𝑀 of the covering maps bounded by 𝚈2 log2 (1∕𝛿). Each of the removed
intervals is centered at the real part of a complex singularity of 𝑔(𝑥).

Moreover, one has the following auxiliary result concerning the distance of each
interval Δ𝑗 to the singularities of the complex extension of the function 𝑔

Proposition B.1. (Yomdin, [59], Lemma 3.6) The center of any of the segments Δ𝑗 in
the partition given by Theorem B.1 is at distance no less than 3×|Δ𝑗 | from any complex
singularity of 𝑔.

Other auxiliary Results

C Quantitative local inversion theorem

We start by stating a Lipschitz inverse function Theorem. Its proof can be found, for
example, in [34] (Th. 14.6.6).

Theorem C.1. Let 𝑈 be an open subset of a Banach space 𝐸 and that 𝑘 ∶ 𝑈 ⟶ 𝐸 is

a Lipschitz mapping with constant 𝐾 < 1. Set ℎ(𝑥) = 𝑥 + 𝑘(𝑥). If the closed ball 𝙱𝜀 (𝑥)
of radius 𝜀 around 𝑥 is contained in 𝑈 then 𝙱(1−𝐾)𝜀 (ℎ(𝑥)) ⊂ ℎ(𝙱𝜀 (𝑥)) ⊂ 𝙱(1+𝐾)𝜀 (ℎ(𝑥)).
The mapping ℎ is a homeomorphism of 𝑈 onto ℎ(𝑈 ), ℎ−1 is a Lipschitz mapping with
constant 1∕(1 − 𝐾), and ℎ(𝑈 ) is an open subset of 𝐸.

This result is crucial in order to prove an analytic inverse function theorem, namely

110
Theorem C.2. Take a function 𝑓 ∈ 𝐶 𝜔 (𝑅 (0)) and a point 𝑧∗ ∈ 𝑅∕2 (0) satisfying
𝑓 ′ (𝑧∗ ) ≠ 0. Then, 𝑓 is invertible in the closed disk 𝑅′ ∕16 (𝑧∗ ) and its inverse 𝑓 −1 is
analytic in |𝑓 ′ (𝑧∗ )|𝑅′ ∕8 (𝑓 (𝑧∗ )), where
{ }
′ 1 |𝑓 ′ (𝑧∗ )|
𝑅 ∶= × min 𝑅, , .
2 max (0) |𝑓 ′′ |
𝑅

𝑓 (𝑧) − 𝑓 (𝑧∗ )
Proof. We define ℎ(𝑧) ∶= + 𝑧∗ , 𝑘(𝑧) ∶= ℎ(𝑧) − 𝑧; both these functions
𝑓 ′ (𝑧∗ )
𝑓 ′ (𝑧)
are obviously holomorphic in 𝑅∕2 (𝑧∗ ). Since 𝑘′ (𝑧) = ℎ′ (𝑧) − 1 = −1 =
𝑓 ′ (𝑧∗ )
′ ′
𝑓 (𝑧) − 𝑓 (𝑧 ) ∗ |𝑓 (𝑧) − 𝑓 (𝑧 )|
′ ′ ∗ |𝑧 − 𝑧 | ∗
, one has |𝑘′ (𝑧)| = ≤ max |𝑓 ′′ | . If we
𝑓 ′ (𝑧∗ ) |𝑓 ′ (𝑧∗ )| 𝑅 (0) |𝑓 ′ (𝑧∗ )|

1
choose to consider only the 𝑧 ∈ 𝑅′ (𝑧∗ ), we obtain that 𝑘 is −Lipschitz on this
2
set.
At this point, we exploit Theorem C.1 and we have that the function ℎ(𝑧) = 𝑘(𝑧) + 𝑧
is a homeomorphism of 𝑅′ (𝑧∗ ) onto its image. Moreover, one has 𝑅′ ∕8 (ℎ(𝑧∗ )) ⊂
ℎ(𝑅′ ∕4 (𝑧∗ )) ⊂ 3𝑅′ ∕8 (ℎ(𝑧∗ )) and, since ℎ(𝑧∗ ) = 𝑧∗ , this yields

𝑅′ ∕8 (𝑧∗ ) ⊂ ℎ(𝑅′ ∕4 (𝑧∗ )) ⊂ 3𝑅′ ∕8 (𝑧∗ ) . (C.1)

We can define 𝑓 −1 by exploiting ℎ and its inverse, namely
( )
𝑓 (𝑧) − 𝑓 (𝑧∗ )
𝑧 = ℎ−1 (ℎ(𝑧)) = ℎ−1 𝑧∗ + =∶ 𝑓 −1 (𝑓 (𝑧)) (C.2)
𝑓 ′ (𝑧∗ )
Indeed, by expressions (C.1) and (C.2), we see that, if we choose
| ∗ 𝑓 (𝑧) − 𝑓 (𝑧∗ ) | 𝑅′
|𝑧 + − 𝑧∗ || ≤ ,
| ′ ∗
| 𝑓 (𝑧 ) | 8
𝑅′
that is |𝑓 (𝑧) − 𝑓 (𝑧∗ )| ≤ |𝑓 ′ (𝑧∗ )| , we have defined the inverse over the closed disc
8
|𝑓 ′ (𝑧∗ )|𝑅′ ∕8 (𝑓 (𝑧∗ )). Finally, we prove that

𝑓 (𝑅′ ∕16 (𝑧∗ )) ⊂ |𝑓 ′ (𝑧∗ )|𝑅′ ∕8 (𝑓 (𝑧∗ )) .

In order to see this, for 𝑧 ∈ 𝑅′ ∕16 (𝑧∗ ) we consider the identity
1
𝑓 (𝑧) − 𝑓 (𝑧∗ ) = 𝑓 ′ (𝑧∗ )(𝑧 − 𝑧∗ ) + 𝑓 ′′ (𝑡𝑧 + (1 − 𝑡)𝑧∗ ) (𝑧 − 𝑧∗ )2 (1 − 𝑡)𝑑𝑡
∫0
that yields the estimate
|𝑓 (𝑧) − 𝑓 (𝑧∗ )| ≤|𝑓 ′ (𝑧∗ )||𝑧 − 𝑧∗ | + max |𝑓 ′′ | |𝑧 − 𝑧∗ |2
𝑅 (0)
(C.3)
𝑅′ 𝑅′2 𝑅′
≤|𝑓 ′ (𝑧∗ )| + max |𝑓 ′′ | ≤ |𝑓 ′ (𝑧∗ )| ,
16 𝑅 (0) 256 8

111
where the last estimate is a consequence of the definition of 𝑅′ .
The fact that 𝑓 −1 inherits the same regularity of 𝑓 is a standard consequence of the
classic local inversion theorem.

D Two elementary properties of Lie Groups

We refer to [43] (Cor. 21.6, Th. 21.10) for proofs.

Proposition D.1. Every continuous action by a compact Lie group on a manifold is

proper.

Theorem D.1 (Quotient manifold). Suppose 𝐺 is a Lie group acting smoothly, freely,
and properly on a smooth manifold . Then the orbit space ∕𝐺 is a topological
manifold of dimension equal to dim  − dim 𝐺 and has a unique smooth structure with
the property that the quotient map 𝜋 ∶  ⟶ ∕𝐺 is a smooth submersion.

E Three auxiliary Lemmas

The following Lemma is an elementary criterion to establish when the projection of a
closed set is still closed.

Lemma E.1. Let 𝐸 be a metric space, 𝐾 a compact subset of some metric space and
Δ a closed subset of 𝐸 × 𝐾. Then, the projection of Δ on 𝐸, indicated by Π𝐸 (Δ), is
closed.

Proof. Let {𝑝𝑛 }𝑛∈ℕ be a sequence in Π𝐸 (Δ) converging to a point 𝑝̄ and {𝑘𝑛 }𝑛∈ℕ a
sequence in 𝐾 satisying (𝑝𝑛 , 𝑘𝑛 ) ∈ Δ. Since 𝐾 is a compact subset of some metric
space, one can extract a subsequence {𝑘𝑛𝑙 }𝑙∈ℕ converging to a point 𝑘̄ ∈ 𝐾. Hence, the
sequence {(𝑝𝑛𝑙 , 𝑘𝑛𝑙 )}𝑙∈ℕ in Δ converges to (𝑝, ̄ ∈ Δ, since Δ is closed. This implies
̄ 𝑘)
that 𝑝̄ belongs to Π𝐸 (Δ), which is therefore closed.

The following statement is a known Theorem due to Bézout (see [42], Th. 3.4a).

Lemma E.2. For any couple of positive integers 𝑘1 , 𝑘2 consider two non-zero irre-
ducible, non-proportional polynomials 𝑄1 ∈ (𝑘1 ) and 𝑄2 ∈ (𝑘2 ). Then the system
𝑄1 (𝑧, 𝑤) = 𝑄2 (𝑧, 𝑤) = 0 has at most 𝑘1 × 𝑘2 solutions.

For the sake of completeness, we also state the following simple result on the codi-
mension of the zero set of a non-null polynomial.

Lemma E.3. Let 𝑛 be a positive integer. Consider a non-null real polynomial 𝑃 ∈ ℝ[𝑥]
of the variables 𝑥 = (𝑥1 , … , 𝑥𝑛 ) ∈ ℝ𝑛 . The zero set 𝑍𝑃 ∶= {𝑥 ∈ ℝ𝑛 |𝑃 (𝑥) = 0} is
contained in a submanifold of codimension one in ℝ𝑛 .

112
Proof. If 𝑃 is a non-zero constant, then there is nothing to prove.
If 𝑃 is non-constant, the proof is by induction on the degree of 𝑃 .
If deg 𝑃 = 1, then 𝑍𝑃 is a hyperplane, which is obviously a submanifold of codi-
mension one.
Suppose, now, that the statement is true for polynomials of degree 𝑘 − 1 ≥ 1.
Consider a polynomial of degree 𝑘, together with its associated open set of non-critical
points 𝑆𝑃 ∶= {𝑥 ∈ ℝ𝑛 | ∇𝑃 (𝑥) ≠ 0}. On the one hand, locally around any point of
𝑍𝑃 ∩ 𝑆𝑃 one can apply the implicit function theorem, so that 𝑍𝑃 ∩ 𝑆𝑃 is indeed a
submanifold of codimension one in ℝ𝑛 . On the other hand, ℝ𝑛 ∖𝑆𝑃 is the common zero
set of the 𝑛 polynomials 𝜕𝑃 ∕𝜕𝑥1 , … , 𝜕𝑃 ∕𝜕𝑥𝑛 ; moreover, since deg 𝑃 = 𝑘 ≥ 2, at least
one among 𝜕𝑃 ∕𝜕𝑥1 , … , 𝜕𝑃 ∕𝜕𝑥𝑛 has degree 𝑘 − 1 ≥ 1. Hence, by hypothesis, ℝ𝑛 ∖𝑆𝑃
is contained in a submanifold of codimension one in ℝ𝑛 .
This proves that the set 𝑍𝑃 ∩ (ℝ𝑛 ∖𝑆𝑃 ) ⊂ (ℝ𝑛 ∖𝑆𝑃 ) is contained in a submanifold
of codimension one in ℝ𝑛 . Obviously, the thesis follows by the fact that

𝑍𝑃 = (𝑍𝑃 ∩ (ℝ𝑛 ∖𝑆𝑃 )) ∪ (𝑍𝑃 ∩ 𝑆𝑃 ) .

Acknowledgements
I am extremely grateful to my PhD advisors, Laurent Niederman and Luca Biasco, for
putting me on this subject, for their useful suggestions and comments, and for reading
and correcting the present manuscript. I am also grateful to J.P. Marco for very useful
discussions. In the months preceeding the end of the redaction of this work, I have been
funded by the ERC project 757802 Haminstab; therefore, I wish to acknowledge both
the ERC and the PI of the project (prof. M. Guardia) for their support.

References
[1] V. Arnol’d. The stability of the equilibrium position of a Hamiltonian system of
ordinary differential equations in the general elliptic case. Soviet. Math. Dokl.,
2:247–249, 1961.

[2] V. Arnol’d. Instability of dynamical systems with several degrees of freedom.

Soviet. Math. Dokl., 156(1):9–12, 1964.

[3] V. Arnol’d, V. Kozlov, A.I. Neishtadt, and E. Khukhro.

Mathematical Aspects of Classical and Celestial Mechanics. Springer, 2010.

[4] D. Bambusi. Nekhoroshev theorem for small amplitude solutions in nonlinear

Schrödinger equations . Math Z., 230:345–387, 1999.

113
 [5] D. Bambusi and A. Fusè. Nekhoroshev Theorem for perturbations of the central
motion. Regular and Chaotic Dynamics, 22:18–26, 2017.

[6] D. Bambusi, A. Fusè, and M. Sansottera. Exponential stability in the perturbed

central force problem. Regular and Chaotic Dynamics, 23:821–841, 2018.

[7] D. Bambusi and A. Giorgilli. Exponential stability of states close to resonance in

infinite dimensional hamiltonian systems. Journal of Statistical Physics, 71:569–
606, 1993.

[8] D. Bambusi and B. Langella. A 𝑐 ∞ nekhoroshev theorem. Mathematics in

Engineering, 3(2):1–17, 2020.

[9] D. Bambusi and B. Langella. Growth of Sobolev Norms in quasi-integrable quan-

tum systems. https://arxiv.org/abs/2202.04505, 2022.

[10] S. Barbieri. On the algebraic properties of exponentially stable integrable Hamil-

tonian systems. Annales de la Faculté de Sciences de Toulouse, 31(5):1365–1390,
2022.

[11] S. Barbieri, J.P. Marco, and J.E. Massetti. Analytic smoothing and Nekhoroshev
estimates for Hölder steep Hamiltonians . Commun. Math. Phys., 396(1), 2022.

[12] S. Barbieri and L. Niederman. Sharp Nekhoroshev estimates for the three-body
problem around periodic orbits. Journal of Differential Equations, 268(7):3749–
3780, 2020.

[13] S. Barbieri and L. Niederman. Bernstein-Remez inequality for algebraic func-

tions: a topological approach. Nonlinear Anaysis, (237), 2023.

[14] S. Basu, R. Pollack, and M.-F. Roy. On the combinatorial and algebraic complex-
ity of quantifier elimination. J. ACM, (45):1002–1045, 1996.

[15] S. Basu, R. Pollack, and M.-F. Roy. Algorithms in real-algebraic geometry.

Springer, 2006.

[16] G. Benettin, F. Fassò, and M. Guzzo. Nekhoroshev stability of L4 and L5 in the

spatial restricted three body problem. Regular Chaotic Dyn., (3):56–72, 1998.

[17] M. Berti, L. Biasco, and P. Bolle. Drift in phase space: a new variational mecha-
nism with optimal diffusion time. Journal de mathématiques pures et appliquées,
82(6):613–664, 2003.

[18] G. Binyamini and D. Novikov. Complex cellular structures. Ann. of Math.,

190:145–248, 2019.

[19] J. Bochnak, M. Coste, and M.-F. Roy. Real algebraic geometry. Springer, 1998.

114
[20] J. Bolte, A. Daniilidis, O. Ley, and L. Mazet. Characterizations of Łojasiewicz
inequalities: subgradient flows, talweg, convexity. Trans. Amer. Math. Soc.,
362:3319–3363, 2010.

[21] A. Bounemoura. Nekhoroshev estimates for finitely differentiable, quasi-convex

Hamiltonian systems. Journal of Differential Equations, 249(11):2905–2920,
2010.

[22] A. Bounemoura, B. Fayad, and L. Niederman. Super-exponential stability for

generic real-analytic elliptic equilibrium points. Advances in Mathematics,
366(3), 2020.

[23] A. Bounemoura and V. Kaloshin. Generic fast diffusion for a class of non-convex
Hamiltonians with two degrees of freedom. Mosc. Math. J., 14(2):181–203, 426,
2014.

[24] J. Bourgain. Green’s functions estimates for lattice Schrödinger operators and
applications. Princeton University Press, 2005.

[25] J. Bourgain, M. Goldstein, and W. Schlag. Anderson localization for Schrödinger

operators on ℤ2 with quasi-periodic potential. Acta Math., (188):41–86, 2002.

[26] H.W. Broer, G.B. Huitema, and M. Sevryuk.

Quasi-periodic motions in families of dynamical systems - Order amidst Chaos.
Springer, 1996.

[27] D. Burguet. A proof of Yomdin-Gromov’s algebraic lemma . Israel Journal of

Mathematics, 168:291–316, 2008.

[28] L. Chierchia. Kolmogorov’s 1954 paper on nearly-integrable Hamiltonian sys-

tems. Regular and Chaotic Dynamics, 13(2):130–139, 2008.

[29] L. Chierchia, M. A. Faraggiana, and M. Guzzo. On steepness of 3-jet non-

degenerate functions. Ann. Mat. Pura Appl. (4), (6):2151–2165, 2019.

[30] R. Cluckers, J. Pila, and A. Wilkie. Uniform parameterization of subanalytic sets

and diophantine applications. Ann. Sci. Éc. Norm. Super., 53:1–42, 2020.

[31] E. Faou and B. Grébert. A Nekhoroshev-type theorem for the nonlinear

Schrödinger equation on the torus . Analysis and Partial Differential Equations,
6(6):1243–1262, 2013.

[32] B. Fayad. Lyapunov unstable elliptic equilibria. J. Amer. Math. Soc., 36(1):81–
106, 2023.

[33] J. Féjoz. Démonstration du Théorème d’Arnol’d sur la stabilité du système plané-

taire. Ergodic Theory Dynam. Systems, 24(5):1521–1582, 2004.

115
[34] D.J.H. Garling. A course in mathematical analysis. Cambridge University Press,
2014.

[35] M. Gromov. Entropy, homology and semi-algebraic geometry. Astérisque -

Séminaire Bourbaki 663, pages 225–240, 1987.

[36] M. Guzzo, L. Chierchia, and G. Benettin. The Steep Nekhoroshev’s Theorem.

Commun. Math. Phys., 342:569–601, 2016.

[37] J. Heintz, M. Roy, and P. Solernó. Sur la complexité du principe de Tarski-

Seidenberg. Bulletin de la S. M. F., 118(1):101–126, 1990.

[38] M.-R. Herman. Dynamics connected with indefinite normal torsion. In Twist
mappings and their applications, volume 44 of IMA Vol. Math. Appl., pages 153–
182. Springer, New York, 1992.

[39] H.-V. Hà and T. Pham. Genericity in Polynomial Optimization. World Scientific,

2017.

[40] Y. Ilyashenko. A criterion of steepness for analytic functions. Uspekhi Mat. Nauk.,
41:193–194, 1986.

[41] V. Kaloshin and K. Zhang. Arnold diffusion for smooth systems of two and a half
degrees of freedom, volume 208 of Annals of Mathematics Studies. Princeton
University Press, Princeton, NJ, 2020.

[42] K. Kendig. A guide to plane algebraic curves. Mathematical Association of

America, 2012.

[43] J.M. Lee. Introduction to smooth manifolds. Springer, 2013.

[44] F. Malige, P. Robutel, and J. Laskar. Partial reduction in the 𝑛-body planetary prob-
lem using the angular momentum integral. Celestial Mech. Dynam. Astronom.,
84(3):283–316, 2002.

[45] J-P. Marco and D. Sauzin. Stability and instability for gevrey quasi-convex near-
integrable hamiltonian systems. Publ. Math. Inst. Hautes Études Sci., 96:199–275,
2003.

[46] J. Moser. On invariant curves of area-preserving mappings of an annulus. Nachr.

Akad. Wiss. Göttingen, II:1–20, 1962.

[47] N. N. Nekhoroshev. Stable lower estimates for smooth mappings and for gradients
of smooth functions. Math USSR Sb., 19(3):425–467, 1973.

[48] N. N. Nekhoroshev. An exponential estimate of the time of stability of nearly-

integrable Hamiltonian systems. I. Russian Mathematical Surveys, 32(6):1–65,
1977.

116
[49] N. N. Nekhoroshev. An exponential estimate of the time of stability of nearly-
integrable Hamiltonian systems. II. Topics in Modern Mathematics, Petrovskii
Seminar, (5):5–50, 1979.

[50] L. Niederman. Stability over exponentially long times in the planetary problem.
Nonlinearity, 9:1703–1751, 1996.

[51] L. Niederman. Hamiltonian stability and subanalytic geometry . Annales de

l’Institut Fourier, 56(3):795–813, 2006.

[52] D. Perrin. Algebraic Geometry. An Introduction. Springer, 2008.

[53] G. Pinzari. Aspects of the planetary Birkhoff normal form. Regular and Chaotic
Dynamics, 18(6):860–906, 2013.

[54] J. Pöschel. On Nekhoroshev Estimates for a Nonlinear Schrödinger Equation and

a Theorem by Bambusi . Nonlinearity, 12:1587–1600, 1999.

[55] N. Roytwarf and Y. Yomdin. Bernstein Classes. Annales de l’Institut Fourier,

47(3):825–858, 1998.

[56] H. Rüssmann. Invariant tori in non-degenerate nearly integrable Hamiltonian sys-

tems. Regular and Chaotic Dynamics, 6(2):119–204, 2001.

[57] G. Schirinzi and M. Guzzo. On the formulation of new explicit conditions for
steepness from a former result of N.N. Nekhoroshev. Journal of Mathematical
Physics, 54(7):1–23, 2013.

[58] Y. Yomdin. 𝐶 𝑘 -resolution of semialgebraic mappings. Addendum to: Volume

growth and entropy. Israel J. Math., 57:301–317, 1987.

[59] Y. Yomdin. Analytic reparametrization of semi-algebraic sets. Journal of

Complexity, (24):54–76, 2008.

[60] Y. Yomdin. Remez-type inequality for discrete sets. Israel J. Math., 186:45–60,
2011.

[61] Y. Yomdin. Smooth parametrizations in dynamics, analysis, diophantine and com-

putational geometry. Japan Journal Ind. Appl. Math., 32:411–435, 2015.

117