Treball final de grau

GRAU DE
MATEMÀTIQUES

Facultat de Matemàtiques
Universitat de Barcelona

TOOLS OF DIFFERENTIAL CALCULUS TO

SOLVE FUNCTIONAL EQUATIONS

Gabriel Garcia Arenas

Director: Ernest Fontich Julià

Realitzat a: Departament de Matemàtiques i Informàtica.
Barcelona, 26 de juny de 2016
Abstract
In this work we first prove the converse of Taylor’s Theorem. This allows us to
prove next the Omega-lemma and the differentiability of the Evaluation map for
certain Banach spaces of analytic functions. These two theorems together with the
Implicit Function Theorem are applied to certain functional equations in order to
prove Poincaré’s Linearization Theorem and the Analytic Stable Manifold Theorem.

Resum
En aquest treball provem en primer lloc el recı́proc del Teorema de Taylor. Aquest
resultat ens permet provar l’Omega-lemma i la diferenciabilitat de l’aplicació Aval-
uació per a certs espais de Banach de funcions analı́tiques. Aquests dos teoremes
juntament amb el Teorema de la Funció Implı́cita els apliquem a certes equacions
funcionals per tal de provar el Teorema de Linealització de Poincaré aixı́ com el
Teorema de la Varietat Estable Analı́tica.

1
Acknowledgements
I am very grateful to my advisor Professor Ernest Fontich for selecting a very
interesting subject which has thrilled me a lot, and also for his good advice and
guidance in the realization of this work.
I also wish to thank the personal support and encouragement received from my
friends and family.

2
Contents
1 Introduction 4

2 Normed spaces 7

3 Continuous linear and multilinear maps 9

4 Differentiability in Banach spaces 15

4.1 Basic definitions and Taylor’s formula . . . . . . . . . . . . . . . . . 15
4.2 The converse of Taylor’s Theorem . . . . . . . . . . . . . . . . . . . 16

5 The Inverse and the Implicit Function Theorem 19

6 Banach spaces of analytic functions 21

6.1 The spaces Aδ (E, F ) . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.2 The Omega-lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.3 The Evaluation map . . . . . . . . . . . . . . . . . . . . . . . . . . 31

7 Poincaré’s Linearization Theorem 34

8 The Analytic Stable Manifold Theorem 42

9 Conclusions 45

References 46

3
1 Introduction
Project
This work has two main objectives:
On the one hand, to become acquainted with some facts and techniques of math-
ematical analysis not widely known. First, the converse of Taylor’s Theorem and
then, a couple of theorems from global analysis, namely the so called Omega-lemma
and the differentiability of the Evaluation map, both in the setting of certain Ba-
nach spaces of analytic maps. And on the other, to have a close look at the power of
these tools of functional analysis when applied to certain functional equations in or-
der to supply clear proofs of deep results such as Poincaré’s Linearization Theorem
and the analytic version of the Stable Manifold Theorem.
Structure
This work mainly consists of three blocks, the first of which contains the pre-
requisites needed in the subsequent parts. These prerequisites aim at covering the
main topics of differential calculus in Banach spaces. This block consists of Sections
2 to 5 and in them one of the most relevant theorems is the converse of Taylor’s
Theorem. The second block is Section 6, where the background and properties of
certain Banach spaces is built including the Omega-lemma and the differentiability
of the Evaluation map. All these results are then applied in the third block which
consists of Sections 7 and 8. To be more precise we next give a brief description of
the different sections of the work.
In Section 2 we describe some basic facts on normed spaces, either real or com-
plex, which are needed in the subsequent sections.
In Section 3, relying mainly on [2] and [4], we focus on those linear and mul-
tilinear maps between normed spaces which are also continuous. This leads us to
the concept of operator norms but it turns out that on some occasions there are
other interesting norms which will be most useful in our later considerations and to
which we have paid special attention. A central point in dealing with the various
continuous multilinear maps between normed spaces is what is called a consistent
family of norms. This section also contains the fundamentals of polynomials in the
general setting of Banach spaces because of their essential role in Taylor’s formula.
Section 4 starts with the definition of differential or derivative of a map between
Banach spaces at a point, and this needs the consideration of the Banach space of
continuous linear maps. In this section we have given the basic definitions and facts
concerning differentiability in dealing with maps between Banach spaces, including
examples because, especially when dealing with higher derivatives, the framework
required is more difficult than in the finite-dimensional situation and requires deal-
ing with spaces of continuous multilinear maps. One of the important results in this
section is Taylor’s formula; we have presented the version which uses Landau’s o-
notation. But the most interesting result in this section is the converse to Taylor’s
Theorem because it represents the key point in proving most subsequent results
which are essential in this work. We have been inspired by the proof occurring in

4
[4] but we have simplified and modified it. Its statement has been given in terms of
Landau’s o-notation.
Section 5 quickly revisits the Inverse and the Implicit Function Theorems because
the latter will be applied in the next sections. Fortunately, the statements of these
key theorems are quite close to those of the well-known case of Rn . This section also
includes comments on partial derivatives in view of their applications in proving the
C ∞ -differentiability character of certain maps.
After these preliminary considerations, we begin, in Section 6, with the main
machinery in this work. Here and following the sketchy notes of Meyer [5] we have
introduced the spaces Aδ (E, F ) and we have provided an elementary proof that they
are Banach. These spaces consist of generalized power series, in the sense that we are
working over a Banach space rather than just working with power series involving a
finite number of variables, an important classical case of course covered by Aδ (E, F ).
The role of δ can be thought of as a generalization of radius of multiconvergence
for these series. It turns out that the functions of Aδ (E, F ) (from a centered ball
of E into F ) are easily seen to be continuous in a neighbourhood of the origin of
E, but the real aim of the beginning of this section is to prove that these functions
are actually C ∞ -differentiable. We have essentially followed the proof of [5] and it
turns out that the key ingredient in the proof is the converse to Taylor’s Theorem.
At the same time some bounds on certain norms of the derivatives for functions
in Aδ (E, F ) are proved, since they are needed later. The second part of Section
6 is devoted to the so called Omega-lemma, which proves C ∞ -differentiability of
composition of functions lying in spaces of type Aδ (E, F ). Here we have given an
original proof because of the essential gaps found in trying to follow Meyer’s proof.
It has not been easy to fill these gaps. In the course of the proof we have needed
not only the converse of Taylor’s Theorem again but also the bounds in norm found
in the previous subsection. We have also included a third subsection which deals
with the C ∞ -differentiability of the so called Evaluation map, because, apart from
its intrinsic importance, we have made use of it in the applications given in the
following sections. We have supplied a couple of original proofs, one of which has
been inspired by [4].
Section 7 is devoted to our main application, which is Poincaré’s linearization
Theorem on the local behaviour near a fixed point of an analytic map from Rn or Cn
into itself when all eigenvalues of its linear part have modulus less than one and are
different from zero. We begin by establishing and proving a lemma of our own whose
role is to make everything clear when it comes to applying the Implicit Function
Theorem in later proofs. Following Meyer’s paper [5] but trying to make clearer
some obscure points in it we have applied the Implicit Function Theorem in order
to reduce everything to formal algebraic computations. But in order to satisfy the
assumptions required by the Implicit Function Theorem, we have needed to make
use of the Omega-lemma in trying to solve a certain functional equation. The
interesting fact here is that by using the properties of the Banach spaces Aδ (E, F )
(in the case of E = F = Rn or Cn ) all problems concerning convergence of power
series have been solved. We have treated both the real and the complex case, as
well as both the so called resonant and non-resonant cases.

5
 Finally in Section 8 we deal with the analytic case of the Stable Manifold The-
orem concerning an analytic map with hyperbolic linear part and whose proof in
this work is similar to that of Poincaré’s Theorem, in the sense that a functional
equation is established and then the Omega-lemma is used in order to guarantee
that the Implicit Function Theorem can be applied to prove the solubility of this
functional equation.

6
2 Normed spaces
In this section, we introduce a few facts from differential calculus in Banach spaces,
either real or complex, since they are the foundation for properly understanding the
results of the paper [5]. We first recall some basic facts on normed spaces.

Definition 2.1. A norm on a real or complex vector space E is a map k k : E → R

satisfying the following properties:

(i) For all x in E, kxk ≥ 0 and equality holds if and only if x = 0,

(ii) kλ xk = |λ| · kxk, for all λ ∈ R or C and all x ∈ E,

(iii) triangle inequality: kx + yk ≤ kxk + kyk, for all x, y ∈ E.

Definition 2.2. A normed space is a pair (E, k k) consisting of a (real or complex)

vector space E and a norm k k defined on it.

When there is no danger of confusion, we just speak of a normed space E, with

k k tacitly understood.
A normed space E may always be considered as a metric space with distance
defined by d(x, y) := kx − yk, and moreover, as a topological space through the
topology induced by the distance. This topological space is always Hausdorff, by
virtue of (i) in Definition 2.1. Actually if we do not assume that kxk = 0 implies
x = 0, we speak of a seminorm, a pseudodistance, and the Hausdorff condition fails
to hold.

Definition 2.3. Two norms | | and k k on a vector space E are said to be equivalent
if they induce the same topology on E.

The following is a well-known fact (cf. [2] I, Prop. 1.6.1, [4] Prop. 2.1.9).

Proposition 2.4. Two norms | | and k k on a vector space E are equivalent if and
only if there exist strictly positive real numbers λ and µ such that, for all x in E,

λ |x| ≤ kxk ≤ µ |x|.

Definition 2.5. A Banach space is a complete normed space, i.e. a normed space
in which every Cauchy sequence is convergent.

Proposition 2.6. If E is a finite dimensional vector space, then there exists at least
a norm on E. Furthermore, any two norms in E are equivalent and E is complete
with respect to any norm.

See [4] Prop. 2.1.10.

7
Remarks 2.7.

(i) Any Cauchy sequence with respect to a norm in a normed space is also Cauchy
with respect to any other equivalent norm. The same result obviously holds
for convergent sequences.

(ii) If E is a normed space, the sum (x, y) 7→ x + y and product by scalars

(λ, x) 7→ λ x from E × E and R × E or C × E into E are both continuous, so
that E may be considered as a topological vector space. In fact, the addition
is uniformly continuous on E × E (see [3] 5.1.5).

Examples 2.8.

(i) In Rn or Cn there are very interesting norms, namely

1
x = (x1 , . . . , xn ) 7→ kxkp = (|x1 |p + · · · + |xn |p ) p , for p ≥ 1

(which in case p = 2 gives the euclidean norm), and the norm

x = (x1 , . . . , xn ) 7→ max (|x1 |, . . . , |xn |)

which corresponds to the case p → ∞. As stated before, all these, as well as

any other norm, are equivalent.
P
∞
(ii) The sequences (xn ) of real or complex numbers such that |xn |p < ∞, for
n=0
p ≥ 1, constitute a vector space under componentwise addition and
λ · (xn ) = (yn ), with yn = λxn for all n.
P∞ 1
Defining k(xn )kp := ( |xn |p ) p , we actually get a Banach space, usually
n=0
denoted by `p . (When p = ∞, `∞ is the space of all bounded sequences with
k(xn )k = sup |xn |).
n≥0

(iii) If E and F are normed spaces then (x, y) 7→ kxkE + kykF defines a norm on
the vector space E × F . The same occurs with (x, y) 7→ max (kxkE , kykF ).

8
3 Continuous linear and multilinear maps
We first recall that if E, F, G, . . . are normed vector spaces, the continuous linear
maps from E into F form a vector subspace of the vector space of all linear maps
from E into F . In fact, the space L(E; F ) of continuous linear maps from E to F
can be normed by defining for f ∈ L(E; F ) its operator norm kf k as the infimum
of those real numbers K such that

kf (x)kF ≤ KkxkE , ∀x ∈ E.

Actually continuity needs to be checked only at 0 thanks to linearity, and we have

(cf. [2] I, Thm. 1.4.1 or [4] Def. 2.2.3)
kf (x)kF
kf k = sup kf (x)kF = sup kf (x)kF = sup kxkE
,
kxkE =1 kxkE ≤1 x6=0

which turns out to be finite if and only if f is continuous and this tells us that f is
bounded on the unit ball of E centered at 0, this being the reason why sometimes
one speaks of bounded linear maps as in [4] or [5].
Similar considerations and results hold for the case of continuous bilinear maps
from E × F into G: for instance, if f belongs to the vector space L(E, F ; G) of
continuous bilinear maps from E × F into G, its operator norm kf k is defined as
the infimum of those K such that:

kf (x, y)kG ≤ K kxkE kykF , ∀(x, y) ∈ E × F.

And we also have (cf. [4] Def. 2.2.8)

kf (x, y)kG
kf k = sup kf (x, y)kG = sup kf (x, y)kG = sup .
kxkE =1,kykF =1 kxkE ≤1,kykF ≤1 x6=0,y6=0 kxkE kykF

This is directly generalized to the case of multilinear maps and we get for in-
stance the normed space of continuous k-linear maps L(E1 , · · · , Ek ; F ). Concerning
notation, if E1 = E2 = · · · = Ek = E, this space will simply be denoted by Lk (E; F ).
It is remarkable that the map f 7→ fe from L(E, F ; G) into L(E; L(F ; G)) defined
by fe(x) : y 7→ f (x, y), where x ∈ E and y ∈ F , turns out to be an isomorphism
of vector spaces which preserves norms (see [2] I, Section 1.9), i.e. kf k = kfek, or
said in other words, it is an isometry. This directly generalizes to the case (see
[4] Prop. 2.2.9)

L(E1 , . . . , Eh , . . . , Eh+k ; F ) ' L(E1 , . . . , Eh ; L(Eh+1 , . . . , Eh+k ; F )).

And in particular we have Lh+k (E; F ) ' Lh (E; Lk (E; F )).

Another essential fact is that if F is assumed to be a Banach space, then
L(E; F ), L(E1 , E2 ; F ), . . . are Banach spaces too (the proofs are straightforward,
cf. [3] 5.7.3 and [4] Prop. 2.2.4).
The vector subspace Lks (E; F ) of Lk (E; F ) consisting of the symmetric k-linear

9
maps f from E k into F , i. e. those continuous k-linear maps f satisfying f σ = f ,
where
f σ (x1 , . . . , xk ) := f (xσ(1) , . . . , xσ(k) ),
i.e.
f (xσ(1) , . . . , xσ(k) ) = f (x1 , . . . , xk ),
for any permutation σ of {1, 2, . . . , k}, is easily seen to be closed in Lk (E; F ) (see
[4] Section 2.2), and consequently is a Banach space in case F is Banach.
So far we have considered the so-called operator norms in the various spaces

L(E1 , . . . , Ek ; F )

but it turns out that in the case E = Rn or Cn there are other norms which are key
in our project and which we now describe:
Let us consider the norm |x| = max(|x1 |, . . . , |xn |) for x = (x1 , . . . , xn ) in E (this
norm will be important when dealing with the expansions of analytic functions).
Then if f : E k −→ F is k-linear, it is well-known that f is determined by the images
f (ei1 , . . . , eik ), where 1 ≤ i1 , i2 , . . . , ik ≤ n, of all the k-tuples obtained from any
Pn
basis, say {e1 , . . . , en } of E: in fact, if xi = xij ej for 1 ≤ i ≤ k, by multilinearity
j=1
we have

n
X n
X
f (x1 , . . . , xk ) = ... x1i1 · · · xkik f (ei1 , . . . , eik ). (1)
i1 =1 ik =1

Then, we associate to f the following norm

X
|f |k := kf (ei1 , . . . , eik )kF .
i1 ,...,ik

It is easy to see that f 7→ |f |k is actually a norm. This norm can be Pseen as the
n n
corresponding one to the 1-norm of R or C , i. e. k(x1 , . . . , xn )k1 = |xi |.
i

Proposition 3.1. Let us consider the norm |(x1 , . . . , xn )|E = max (|x1 |, . . . , |xn |)
in E (= Rn or Cn ). Then the norm |f |k just defined above is equivalent to the
operator norm kf k.

The case when F is finite-dimensional is automatic: all norms in a finite-dimensional

normed space are equivalent as stated in Proposition 2.6, so the assertion becomes
interesting when dim F = ∞.

10
Proof. By (1) we have

° °
°X °
° °
kf (x1 , . . . , xk )kF = ° f (ei1 , . . . , eik )x1i1 · · · xkik °
° °
i1 ,...,ik F
X
≤ kf (ei1 , . . . , eik )kF |x1i1 | · · · |xkik |
i1 ,...,ik
à !
X
≤ kf (ei1 , . . . , eik )kF |x1 |E · · · |xk |E
i1 ,...,ik
= |f |k |x1 |E · · · |xk |E ,

which proves kf k ≤ |f |k . And from

X X X
|f |k = kf (ei1 , . . . , eik )kF ≤ kf k·|ei1 |E · · · |eik |E = kf k = nk ·kf k ,
i1 ,...,ik i1 ,...,ik i1 ,...,ik

we finally obtain the equivalence of both norms.

Our next considerations will turn out to be important in dealing with the ex-
pansions of the analytic maps we will introduce in Section 6. We first begin with
the following

Definition 3.2. (see [5]). A family of norms | |k on the vector spaces Lk (E, F ),
for k = 1, 2, . . . , is said to be consistent (or simply, the norms | |k are consistent),
if F is a Banach space and the following four properties hold:

(i) {Lk (E, F ), | |k } is a Banach space,

(ii) kf (x1 , . . . , xk )kF ≤ |f |k · kx1 kE · · · kxk kE , for all xi ∈ E,

(iii) The isomorphism of Lh+k (E; F ) ' Lh (E; Lk (E; F )) is norm-preserving, and

(iv) |f σ |k = |f |k , for any permutation σ of {1, 2, . . . , k}.

It is well-known that the family of the usual operator norms is consistent. On

the other hand, not every norm in Lk (E, F ) satisfies (ii): for instance f 7→ 12 kf k
where kf k stands for the operator norm fails to satisfy (ii). Let us see now that the
norms | |k just introduced are consistent.

Proposition 3.3. The norms | |k , for k = 1, 2, . . ., are consistent on the spaces

Lk (E; F ), where E = Rn or Cn and F is a Banach space.

11
Proof. Obviously (i) holds by Proposition 3.1, since Lk (E; F ) is Banach under the
operator norm.
Concerning (ii), by multilinearity it suffices to assume kxi kE = 1, for all xi , and
with the notations of (1), we have

° °
°X °
° °
kf (x1 , . . . , xk )kF = ° f (ei1 , . . . , eik ) x1i1 · · · xkik °
° °
i1 ,...,ik F
X
≤ kf (ei1 , . . . , eik )kF = |f |k .
i1 ,...,ik

For (iii), if f ∈ Lh+k (E; F ) and fe is its corresponding element in Lh (E; Lk (E; F )),
for any (x1 , . . . , xh ) ∈ E h , as

fe(x1 , . . . , xh ) = f (x1 , . . . , xh , ∗, . . . , ∗) ∈ Lk (E; F ),

we have
X
|fe|h = kfe(ei1 , . . . , eih )kLk (E;F )
i1 ,...,ih
à !
X X
= k(fe(ei1 , . . . , eih ))(ej1 , . . . , ejk )kF
i1 ,...,ih j1 ,...,jk
à !
X X
= kf (ei1 , . . . , eih , ej1 , . . . , ejk kF
i1 ,...,ih j1 ,...,jk
X
= kf (ei1 , . . . , eih , ej1 , . . . , ejk kF = |f |h+k
i1 ,...,ih ,j1 ,...,jk

which proves (iii), and (iv) is obvious.

Next, we recall a few basic facts related to polynomials. First we assume E, F

to be vector spaces over any field K of characteristic 0, but later on we will be
interested in the case of normed spaces over R or C.

Definition 3.4. A homogeneous polynomial map of degree k, or just k-homogeneous

polynomial from E into F is a map ϕ : E → F induced by a nonzero k-linear map
f : E k → F in the following way:

ϕ(x) = f (x, . . . , x), ∀x ∈ E.

12
Remarks 3.5.

(i) f can be taken to be symmetric in the preceding definition: in any case,

1 X
g(x1 , . . . , xk ) := f (xσ(1) , . . . , xσ(k) )
k! σ∈S
k

is k-linear, symmetric and

P ϕ(x) = g(x, . . . , x). Usually we write g = Symk (f ),
so that Symk (f ) = k!1 f σ , where f σ (x1 , . . . , xk ) = f (xσ(1) , . . . , xσ(k) ).
σ∈Sk

(ii) Furthermore, g is unique, i.e., there exists just one symmetric k-linear function
g : E k → F satisfying g(x, . . . , x) = ϕ(x). One possible proof (based on [4] Prop.
2.2.11 (iii)) may be the following: for any k-tuple (v1 , . . . , vk ) of vectors of E,
expanding ϕ(x), with x = t1 v1 + · · · + tk vk and t1 , . . . , tk being indeterminates,
by multilinearity and symmetry of g, we have

ϕ(x) = g(x, . . . , x) = g(t1 v1 + · · · + tk vk , . . . , t1 v1 + · · · + tk vk )

X
= tα1 1 · · · tαk k g(v1 , . . . , v1 , . . . , vk , . . . , vk )
α +···+α =k
| {z } | {z }
1 k α1 αk
αi ≥0
X
= tα1 1 · · · tαk k gα1 ,...,αk ,
α1 +···+αk =k
αi ≥0

a usual homogeneous polynomial (with coefficients in F ) of degree k in the in-

determinates t1 , . . . , tk , where we have set gα1 ,...,αk = g(v1 , . . . , v1 , . . . , vk , . . . , vk ).
| {z } | {z }
α1 αk

By formally deriving with respect to t1 we get

∂ X
ϕ(t1 v1 + · · · + tk vk ) = α1 t1α1 −1 tα2 2 · · · tαk k gα1 ,...,αk
∂t1 α +···+α =k 1 k
αi ≥0

and going on deriving with respect to t2 , . . . , tk , we eventually get

∂k X
ϕ(t1 v1 + · · · + tk vk ) = α1 · · · αk t1α1 −1 · · · tαk k −1 gα1 ,...,αk .
∂t1 · · · ∂tk α +···+α =k 1 k
αi ≥0

As (α1 − 1) + · · · + (αk − 1) = α1 + · · · + αk − k = 0, the last sum reduces

to just one term, namely g1,...,1 = g(v1 , . . . , vk ), which actually proves that g
is uniquely determined by ϕ. Another proof of this uniqueness may be found
in [2] I Cor. 6.3.3.
(iii) When E is finite-dimensional, taking a basis e1 , . . . , en in E and having a look
at (1) with x = x1 = · · · = xk , we see that a k-homogeneous polynomial is
just a usual homogeneous polynomial of degree k in the n coordinates of x,
with coefficients in F .

13
Definition 3.6. A polynomial (function) from E into F is a function ϕ : E → F
defined by a finite sum of homogeneous polynomials, in which case, we can write
ϕ = ϕ0 + ϕ1 + · · · + ϕk , for some k, where each ϕj is j-homogeneous. If ϕk 6= 0 we
say ϕ has degree k, and if ϕ = 0 (and here we allow ϕ0 , . . . , ϕk to be zero), we set
degree (ϕ) = −∞.

Remarks 3.7.

(i) The expression of ϕ as a sum of j-homogeneous polynomials for distinct j’s

is unique, i.e., if ϕ = ϕ0 + ϕ1 + · · · + ϕk = ψ0 + · · · + ψk , where ϕj and
ψj are assumed j-homogeneous, then ϕj = ψj for all j. In other words, the
summands ϕj are uniquely determined by ϕ.
A proof of this fact can be given by induction on k: the details can be found
in [2] I, Cor. 6.3.2.

(ii) It also arises the question of continuity: the fact is that global continuity comes
from continuity at the origin (this is obvious since this is the case in the context
of multilinear maps and polynomials are obtained from them). Moreover, if
dimE < ∞, as all multilinear maps from E into F are continuous (this follows
easily from (1)), we gather that all polynomial functions are continuous. If
dim E = ∞, then a polynomial ϕ = ϕ0 + ϕ1 + · · · + ϕk is continuous if and
only if each ϕj is continuous and this is so if and only if, for each j, the unique
j-linear symmetric map defining ϕj is continuous (see [2] I, Thm. 6.4.1). As we
will deal with continuous multilinear maps, all polynomials will, in turn, be
continuous.

14
4 Differentiability in Banach spaces
Here we introduce some basic facts concerning differentiability of functions defined
on open sets of a Banach space and taking values in another Banach space. Special
attention will be devoted to Taylor’s formula and specially to the converse of Taylor’s
Theorem.

4.1 Basic definitions and Taylor’s formula

Definition 4.1. A function f : U → F , where E, F are Banach spaces and U is
open in E, is differentiable at x ∈ U if and only if there exists a continuous linear
map, necessarily unique, Df (x) : E → F such that, using the Landau notation, for
x + h ∈ U,
kf (x + h) − f (x) − Df (x)(h)k = o(khk),
in which case we call Df (x) the (first) derivative of f at x. If Df (x) exists for all
x ∈ U , f is said to be differentiable on U and x 7→ Df (x) is a map from U into the
Banach space L(E; F ). In case this latter map is continuous, we say f is of class
C 1 on U and write f ∈ C 1 (U ).

If Df : U → L(E; F ) turns out to be also differentiable we get the second

derivative of f , denoted by D2 f which is none other than the derivative of Df :

D2 f : U → L(E; L(E; F )).

But we have a canonical norm preserving isomorphism L(E; L(E; F )) ' L2 (E; F )
into the Banach space of continuous bilinear maps from E into F , so that in what
follows, we will consider D2 f as a map from U into L2 (E; F ).
Now a remarkable fact occurs: If a ∈ U and D2 f (a) exists, then D2 f (a) is
symmetric i.e., D2 f (a)(u, v) = D2 f (a)(v, u), for all (u, v) ∈ E 2 (proof in [2]
I, Thm. 5.1.1). We do not need here D2 f to be continuous at a. If D2 f : U →
L2s (E; F ) turns out to be continuous, f is said to be of class C 2 . Similarly, the n-th
derivative of f , if it exists is a map Dn f : U → Lns (E; F ), etc.
Examples 4.2.

(i) By definition, if f : U → F is the restriction of a continuous linear map

fe : E → F , then Df : U → L(E; F ) is the constant map Df (x) = fe, for all
x ∈ U , and as the derivative of a constant map is zero, we see that D2 f = 0.
(ii) Any continuous bilinear map f : E ×F → G is differentiable and its derivative
at (a, b) ∈ E × F is the map Df (a, b) ∈ L(E × F ; G) defined by

Df (a, b)(h, k) = f (a, k) + f (h, b)

(see [2] I, Thm. 2.4.3, or [4] Ch.2 Ex 2.3-1). As the maps h 7→ f (h, b) and
k 7→ f (a, k) are continuous and linear, we get D3 f = 0.

15
(iii) If f : E1 × · · · × En → F is continuous and n-linear, then f is differentiable
and we have
Df (a1 , . . . , an )(h1 , . . . , hn )

= f (h1 , a2 , . . . , an ) + f (a1 , h2 , a3 , . . . , an ) + · · · + f (a1 , . . . , hn ).

In particular, for the map ϕ : x 7→ f (x, . . . , x) from E into F , where f is

assumed continuous and n-linear, we have Dϕ(x) : E → L(E; F ) given by

Dϕ(x)(h) = f (h, x, . . . , x) + f (x, h, x, . . . , x) + · · · + f (x, x, . . . , x, h)

and in case f is symmetric the above sum is just n f (x, x, · · · , x, h).

Let us turn now to a simple version of Taylor’s formula which will turn out to
be quite useful in our later applications.
Theorem 4.3. Let E, F be Banach spaces and U open in E. Assume f : U → F
is n − 1 times differentiable on U and that Dn−1 f is differentiable at a ∈ U . Then
whenever a + h ∈ U , we have, using the Landau notation,
Df (a) D2 f (a) 2 Dn f (a) n
f (a + h) = f (a) + ·h+ · h + ··· + · h + o(khkn ).
1! 2! n!

(Here, for instance, Dn f (a) · hn is the value of Dn f (a) : E n → F on (h, . . . , h),

and Dn f (a) · (hn−1 , k) = Dn f (a)(h, . . . , h, k), etc).
Observe that this formula approximates in F the value f (a + h) by a polynomial
of degree at most n.
The proof of this theorem is by induction on n (details in [2] Part I Thm. 5.6.3).

4.2 The converse of Taylor’s Theorem

Using the preceding notations, we state and prove a converse to Theorem 4.3,
namely the so-called converse to Taylor’s Theorem (cf. [4]).
Theorem 4.4. Given f : U → F , assume there exist continuous maps

ϕj : U ⊂ E → Ljs (E; F ), j = 1, . . . , n

such that, for any a ∈ U and any h with a + h ∈ U , we have

ϕ1 (a) ϕ2 (a) 2 ϕn (a) n
f (a + h) = f (a) + ·h+ · h + ··· + · h + o(khkn ).
1! 2! n!

Then f is C n in U and ϕj (a) = Dj f (a), for all j.

Remark 4.5. The expression on the right hand side before the o-term is a poly-
nomial function of degree ≤ n, and it is continuous by the assumptions on the
ϕj ’s.

16
Proof. We follow the lines of ([4] Supplement 2.4.B and [6]) .
We proceed by induction on n, the case n = 1 being obvious by the definition of
derivative of a map at a point. So assume the theorem holds for j = 1, . . . , n − 1
and let us prove that it also holds for j = n. This entails that we have

ϕ1 (a) = Df (a), . . . , ϕn−1 (a) = Dn−1 f (a)

and we contend that ϕn (a) = Dn f (a), for every a in U . We will be considering,

for any a ∈ U , elements h, k in U such that a + h and a + h + k lie in U and such
that rkkk ≤ khk ≤ skkk, for some 0 < r < s. This condition will ensure that
o(khkn ) = o(kkkn ) = o(kh + kkn ), for all n ≥ 1. Now we write the formula in the
theorem in two different ways, namely
1
f ((a + h) + k) = f (a + h) + Df (a + h)k + · · · + Dn−1 f (a + h)k n−1
(n − 1)!
1
+ ϕn (a + h)k n + o(kkkn ).
n!

1
f (a + (h + k)) = f (a) + Df (a)(h + k) + · · · + Dn−1 f (a)(h + k)n−1
(n − 1)!
1
+ ϕn (a)(h + k)n + o(kh + kkn ).
n!

Subtracting these two expansions and collecting terms homogeneous in k j we get,

using symmetry,

D2 f (a) 2 Dn−1 f (a) n−1 ϕn (a) n

0 = f (a + h) − f (a) − Df (a)h − h − ··· − h − h
2! (n − 1)! n!
D2 f (a) Dn−1 f (a)
+ Df (a + h)k − Df (a)k − 2(h, k) − · · · − (n − 1)(hn−2 , k)
2! (n − 1)!
n−2 n−2
ϕn (a) D f (a + h) n−2 D f (a) n−2
− n(hn−1 , k) + · · · + k − k
n! (n − 2)! (n − 2)!
Dn−1 f (a) ϕn (a) ³ n ´ 2 n−2 Dn−1 f (a + h) n−1
− (n − 1)(h, k n−2 ) − · (h , k ) + k
(n − 1)! n! 2 (n − 1)!
Dn−1 f (a) n−1 ϕn (a) ϕn (a + h) n ϕn (a) n
− k − n(h, k n−1 ) + k − k + o(kkkn ).
(n − 1)! n! n! n!

Calling g0 (h) the 1st line above, g1 (h)k the 2nd, g2 (h)k 2 the 3rd,. . ., and finally
gn (h)k n the (n + 1)th without the o-term, we see that

g0 (h) + g1 (h)k + · · · + gn−1 (h)k n−1 + gn (h)k n = o(kkkn ).

As gj (0) = 0 and the gj are continuous for j = 0, 1, . . . , n, so that, in particular,

17
 gn (h)k n is o(kkkn ), then we get

g0 (h) + g1 (h)k + · · · + gn−1 (h)k n−1 = o(kkkn ).

We want to prove that each term in the preceding sum is actually of order
o(kkkn ). This can be achieved by taking distinct numbers λ1 , . . . , λn and replace k
by λj k in the above expression. So we are led to the system of n linear equations
with unknowns the vectors g0 (h), . . . , gn−1 (h)k n−1 on F :
 n−1
 g0 (h) + g1 (h)λ1 k + · · · + gn−1 (h)(λ1 k)
 = o(kkkn )
·································


g0 (h) + g1 (h)λn k + · · · + gn−1 (h)(λn k)n−1 = o(kkkn )

or written otherwise
    
1 λ1 · · · λn−1
1 g0 (h) o(kkkn )
1 λ2 · · · λn−1   g1 (h)k  o(kkkn )
 2  = 
· ··· ··· ···  ···   ··· .
1 λn · · · λn−1
n gn−1 (h)k n−1 o(kkkn )

As the matrix of coefficients

Q of this system of equations is a Vandermonde matrix
with nonzero determinant (λi − λj ), we can multiply on the left by its inverse
i>j
and get that all vectors g0 (h), g1 (h)k, . . . , gn−1 (h)k n−1 are of order o(kkkn ), since
they can be expressed as linear combinations of the independent terms.
We will fix our attention to the last vector, but first observe that the restriction
rkkk ≤ khk ≤ skkk does not affect the following results because in dealing with
multilinear maps (such as Dn−1 f (a) and ϕn (a)), scalars may be placed either inside
or outside them without altering values.
Now gn−1 (h)k n−1 = o(kkkn ) = o(khkn ) says that
° n−1 °
° D f (a + h) n−1 Dn−1 f (a) n−1 ϕ (a) °
° k − k −
n
(h, k n−1
)° = o(kkkn ),
° (n − 1)! (n − 1)! (n − 1)! °

or equivalently, that

kDn−1 f (a + h) − Dn−1 f (a) − ϕn (a)(h, ∗)k = o(khk),

so that ϕn (a) = D(Dn−1 f )(a) = Dn f (a).

18
5 The Inverse and the Implicit Function Theorem

In this section we recall two of the central theorems of differential calculus in the
context of Banach spaces, which are the Inverse Map Theorem and the Implicit
Function Theorem, but for the latter we also need to introduce partial derivatives
because of their appearance in the Implicit Function Theorem. Partial derivatives
also play a relevant role concerning C r -differentiability.
Definition 5.1. A map f : U → V , where U and V are open sets in the respective
Banach spaces E and F , is a C r diffeomorphism (r ≥ 1) if f is C r -differentiable, f
is bijective and f −1 is also of class C r .

Now we are going to establish the Inverse Map Theorem:

Theorem 5.2. With the preceding notations, if f : U → V is of class C r , where
r ≥ 1, a ∈ U and Df (a) : E → F is a linear isomorphism, then f is a C r -
diffeomorphism of some neighbourhood of a onto some neighbourhood of f (a). Fur-
thermore,
Df −1 (y) = [Df (f −1 (y))]−1 ,
for y in this neighbourhood of f (a).

Proofs may be found in [2], or [4].

Definition 5.3. Let U be open in E1 × E2 and f : U → F . If (a, b) ∈ U and the
derivatives of the maps x 7→ f (x, b), y 7→ f (a, y) exist at a and b respectively, for
x ∈ E1 , y ∈ E2 , they are called the partial derivatives of f at (a, b) ∈ U and will be
denoted by D1 f (a, b) ∈ L(E1 ; F ) and D2 f (a, b) ∈ L(E2 ; F ).

Obviously it may happen that one partial derivative exists but not the other,
or that none of them exist and, of course, the preceding definition may be directly
generalized to the case of any finite direct product E1 × · · · × En of Banach spaces
instead of E1 × E2 .
In this context we have the following
Proposition 5.4. With the above notations, if f : U → F is differentiable at
(a, b) ∈ U , then both partial derivatives exist and
D1 f (a, b)(v) = Df (a, b)(v, 0),
D2 f (a, b)(w) = Df (a, b)(0, w),
Df (a, b)(v, w) = D1 f (a, b)(v) + D2 f (a, b)(w).
Moreover f is of class C r on U (r ≥ 1) if and only if both D1 f and D2 f are of
class C r−1 on U .

Proofs may be found in [2], [3], or [4].

We end this section with one of the most fundamental theorems in Analysis, the
Implicit Function Theorem:

19
Theorem 5.5. Let U ⊂ E, V ⊂ F be open in the Banach spaces E and F , and
let f : U × V → G be a C r map (r ≥ 1) into the Banach space G. Assume
that, for (a, b) ∈ U × V, D2 f (a, b) : F → G is an isomorphism. Then there exist
neighbourhoods U0 of a and W0 of f (a, b) and a unique C r map g : U0 × W0 → V ,
such that, for all (x, w) ∈ U0 × W0 ,

f (x, g(x, w)) = w.

Proof in [4].

Remarks 5.6.

(i) By a Theorem of Banach (see [4] 2.2.16 or [2] I Section 1.6) if A : F → G is

an algebraic isomorphism which is continuous, then A−1 is also continuous,
i.e., A is automatically a homeomorphism. As, by definition, D2 f (a, b) is a
continuous map from F into G, D2 f (a, b) is actually a homeomorphism.

(ii) The Implicit Function Theorem 5.5 is very often used when w is fixed, where we
may even assume w = 0. In this case we say that in the equation f (x, y) = 0,
for (x, y) near (a, b), y can be locally solved in the sense that y = g(x), with
g as regular as f is.

20
6 Banach spaces of analytic functions

In this section we first introduce the Banach spaces of analytic functions we will be
interested in and then we will focus on the Omega-lemma and the Evaluation map.

6.1 The spaces Aδ (E, F )

Assume that E, F are Banach spaces, both real or both complex, and consider a
family of norms | |n on the Banach spaces Ln (E; F ) which is consistent in the sense
of Definition 3.2.
We now introduce the basic Banach spaces we will be dealing with in what
follows, and it will easily be observed that the norms we will define are best adapted
for the case of analytic functions in terms of the coefficients occurring in their
expansions. For simplicity in what follows we will omit subscripts in the norms.
Thus, take δ > 0 and consider the set Aδ (E, F ) of all formal power series f =
P
∞ P
∞
ak , with ak ∈ Lks (E; F ), such that |ak |δ k < ∞, and define kf kδ to be the
k=0 k=0
P
∞
finite value of the sum of the preceding series: kf kδ = |ak |δ k . It turns out that
k=0
Aδ (E, F ), with addition and product by scalars defined in the usual way, is a vector
space, k kδ is a norm in it, and actually Aδ (E, F ) is a Banach space under k kδ , since
this norm is essentially the `1 norm. Obviously it contains all polynomial functions
(see Definition 3.6).
Observe that if 0 < ρ < δ, then

Aδ (E, F ) ⊆ Aρ (E, F ), and kf kδ ≥ kf kρ .

Let us prove, for the sake of completeness, that Aδ (E, F ) is actually a Banach
space.

Proposition 6.1. Aδ (E, F ) with the norm k · kδ is a Banach space.

Proof. We have just mentioned that Aδ (E, F ) is a normed vector space (over R or
C) and now we have to prove that it is complete, and for this, the proof that `1 is
Banach may easily be adapted here. It runs as follows:
P
∞
(n)
Let fn := ak , n = 1, 2, . . . , be a Cauchy sequence of elements in Aδ (E, F ).
k=0
P
∞
(n)
This means that, for each n, |ak |δ k < ∞ and that for any given ε > 0 there
k=0
exists an n0 , depending on ε, such that
∞
X (p) (q)
kfp − fq kδ = |ak − ak |δ k < ε, whenever p, q ≥ n0 .
k=0

21
 (p) (q)
But then,
³ ´ for any fixed k, |ak − ak |δ k < ε, which means that the sequence
(n)
ak is Cauchy in Lks (E, F ) and therefore convergent, say to ak ∈ Lks (E, F ),
n∈N
since Lks (E, F ) is Banach (being closed in Lk (E, F ) which is Banach because F is
P
∞
so). Now there remains to see that f = ak lies in Aδ (E, F ) and is in fact the
k=0
limit of fn .
Observe that, for each t ∈ N, if p, q ≥ n0 ,
t ¯
X ¯ ∞ ¯
X ¯
¯ (p) (q) ¯ k ¯ (p) (q) ¯ k
¯ak − ak ¯ δ ≤ ¯ak − ak ¯ δ < ε
k=0 k=0

so that
t ¯
X ¯ t ¯
X ¯
¯ (p) (q) ¯ ¯ (q) ¯
lim ¯ak − ak ¯ δ k = ¯ak − ak ¯ δ k ≤ ε.
p→∞
k=0 k=0

As this holds, for any t, letting t → ∞, we get

∞ ¯
X ¯
¯ (q) ¯ k
¯ak − ak ¯ δ ≤ ε.
k=0

This tells us that f − fq ∈ Aδ (E, F ), and as fq ∈ Aδ (E, F ), we conclude that

f = (f − fq ) + fq ∈ Aδ (E, F ).

Furthermore, the last inequality shows that lim fq = f.

q→∞

Now to any f in Aδ (E, F ), with our preceding notation, we associate the map

fe : B(0, δ) ⊂ E → F

from the closed ball B(0, δ) = {x ∈ E : |x|E ≤ δ} of E into F , defined by

∞
X
fe(x) = ak (xk ),
k=0

where xk = (x, . . . , x), for each k, and x ∈ B(0, δ), and we will identify in the sequel
fe with its power series representation f .
P
This is due to the fact that f 7→ fe is injective, i.e. that if ak (xk ) = 0 whenever
k
|x| ≤ δ, then all ak = 0. Otherwise take the first nonzero ak and an x ∈ E such
that |x| ≤ δ and ak (xk ) 6= 0. By homogeneity we have for |λ| < 1

0 = ak ((λx)k ) + ak+1 ((λx)k+1 ) + ak+2 ((λx)k+2 ) + · · ·

= λk ak (xk ) + λk+1 (ak+1 (xk+1 ) + λ(ak+2 (xk+2 ) + · · · ).

22
 But the series in the last term is absolutely convergent for |λ| < 1, and dividing
through by λk and letting λ → 0 we get ak (xk ) = 0, a contradiction. (Actually this
generalizes Remark 3.7 (i)).
By virtue of the Weierstrass M -test, f is absolutely and uniformly convergent
for |x| ≤ δ, so that f is continuous on B(0, δ). Furthermore, as the series defining
kf kδ consists of nonnegative terms, we see that each term satisfies |ak |δ k ≤ kf kδ or,
equivalently, |ak | ≤ kfδkkδ (a Cauchy-type inequality), and also, obviously, |f (x)|F ≤
kf kδ whenever |x|E ≤ δ.
Now, one of our major goals here will be to show that f is not only continuous but
also C ∞ on the open ball B(0, δ) = {x ∈ E : |x|E < δ}, as is the case of the usual
analytic functions of one or several variables. As is well-known these latter functions
are always C ∞ , but not conversely, and their derivatives are easily recognized in the
coefficients of their expansions. But first we make a couple of observations, the first
of which pays attention to Aδ (E, F ) in case E is finite-dimensional, say E = Rn or
Cn :
We take a basis e1 , . . . , en in E with |ei |E = 1, for i = 1, . . . , n. Then any
ak ∈ Lks (E, F ) may be described as follows:
P
n
If x = xj ej , then as seen in (1) we have
j=1
n
X
k
ak (x ) = ak (ei1 , . . . , eik ) xi1 · · · xik
i1 ,...,ik =1

But using the fact that ak is symmetric we can gather those terms in which e1 occurs
α1 times, e2 occurs α2 times,. . ., and en , αn times, (with α1 + α2 + · · · + αn = k)
and get
X k!
ak (xk ) = ak (e1 , . . . , e1 , . . . , en , . . . , en ) xα1 1 · · · xαnn .
α1 +α2 +···+αn =k
α1 ! · · · αn ! | {z } | {z }
αi ≥0 α1 αn

Here we recover the usual homogeneous polynomial expansion for ak in the coor-
dinates x1 , . . . , xn of x with coefficients in F and see that the norm |ak | appearing
in Proposition 3.1 is just the sum of the norms (in F ) of the coefficients associated
with ak . When F = Cn , for instance, the coefficients are indeed vectors in Cn ,
so that in this case f ∈ Aδ (Cn , Cn ) is given by a usual convergent power series in
x1 , . . . , xn with coefficients in Cn . A similar result obviously holds for Aδ (Rn , Rn )
when F = Rn .
The second observation is contained in the following proposition and shows that
the familiar analytic functions are in the spaces just introduced when E = Rn or
Cn .

Proposition 6.2. If f : {x ∈ Cn : |xi | < r} → F is analytic and bounded in norm

by M , then for each η ∈ (0, r), f lies in Aη (Cn , F ) and kf kη ≤ M (1 − ηr )−n .

23
Proof. Consider the power series expansion for f when |xi | < r:
X
f (x1 , . . . , xn ) = ak1 ···kn xk11 · · · xknn
ki ≥0

By Cauchy’s inequalities, |ak1 ···kn | ≤ M r−(k1 +···+kn ) and consequently the series
X
kf kη = |ak1 ···kn | η k1 +···+kn
ki >0

is majorized term by term by the series

à ! à !
X ³ η ´k1 +···+kn X ³ η ´ k1 X ³ η ´kn
M =M ··· =
k ≥0
r k ≥0
r k ≥0
r
i 1 n

³ η ´−1 ³ η ´−1 ³ η ´−n

=M 1− ··· 1 − =M 1− .
r r r

Remark 6.3. This proof works when the spaces Lks (Cn ; F ) are normed as in Propo-
sition 3.1, but as the operator norms are dominated by them the assertion also holds
in the usual situation.

Next, we give a technical lemma:

P
∞
Lemma 6.4. Let ak ≥ 0, k = 0, 1, 2, . . . , and δ > 0 be such that ak δ k = M < ∞.
k=0
Then for any positive integer k, and ρ such that 0 < ρ < δ, we have

∞
X j! k!M
aj ρj−k ≤ .
j=k
(j − k)! (δ − ρ)k

P
∞
Proof. The complex-valued function g(z) = ak z k is analytic in the disc |z| < δ
k=0
and bounded by M .
P
∞
Consider its expansion g(z) = bk (z − ρ)k around ρ. This power series has
k=0
(k)
radius of convergence ≥ δ − ρ and bk = g k!(ρ) . Since M obviously bounds |g(z)| on
M
the ball centered at ρ of radius δ − ρ, the Cauchy’s inequalities entail |bk | ≤ (δ−ρ) k,

i.e.,
k!M
|g (k) (ρ)| ≤ .
(δ − ρ)k
P
∞
But from g(z) = ak z k we see that
k=0
∞
X
(k) j!
g (z) = aj z j−k .
j=k
(j − k)!

24
Setting z = ρ, we get the result.

We now come to a very important fact concerning differentiability:

Theorem 6.5. If f ∈ Aδ (E, F ) then, for all i ≥ 1, there exists Di f . Moreover, for
0 < ρ < δ it belongs to Aρ (E, Lis (E; F )), and
i!kf kδ
kDi f kρ ≤ .
(δ − ρ)i
In particular f is C ∞ in the disc B(0, δ) = {x ∈ E : |x|E < δ}.

P
∞
Proof. Let f (x) = ak (xk ), as in our previous notations. We have
k=0

∞
X ∞
X X∞ X k µ ¶
k k k
∞ > kf kδ = |ak |δ = |ak |(ρ + (δ − ρ)) = |ak |ρk−i (δ − ρ)i
k=0 k=0 k=0 i=0
i
∞
Ã∞ µ ¶ !
X X k
= |ak |ρk−i (δ − ρ)i , (2)
i=0 k=i
i

where we have rearranged terms, since they are all positive. Let |x| < ρ and
|y| < 12 (δ − ρ). Then

∞ ∞
à k µ ¶ !
X X X k
f (x + y) = ak ((x + y)k ) = ak (xk−i , y i )
k=0 k=0 i=0
i
∞
à ∞ µ ¶ !
X X k
= ak (xk−i , ∗) (y i ) (3)
i=0 k=i
i

the last equality holding because the norms are assumed to be consistent and the
last two series in (2) majorize (term by term) the last two series in (3), which implies
absolute convergence.
Formula (3) suggests that the ith derivative of f at x should be

X∞ µ ¶
i k
D f (x) = i! ak (xk−i , ∗) (4)
k=i
i

and the next aim is to prove this. Let us denote the right-hand side of (4) by ϕi (x).
We then have by Lemma 6.4

X∞ µ ¶
k kf kδ
kϕi kρ = i! |ak |ρk−i ≤ i! (5)
k=i
i (δ − ρ)i

so that ϕi ∈ Aρ (E, Lis (E; F )).

25
 Next, we go on with (3) using the notation of (4):

∞
X n
X X∞
ϕi (x) i ϕi (x) i ϕi (x) i
f (x + y) = (y ) = (y ) + (y ) (6)
i=0
i! i=0
i! i=n+1
i!

δ−ρ
Now we estimate the last term in (6) assuming |y| < 2
and bearing in mind (5)
we have:

¯ ∞ ¯ ¯Ã ∞ !¯
¯ X ϕ (x) ¯ ¯ X ϕ (x) ¯
¯ i i ¯ ¯ i i−n−1 n+1 ¯
¯ (y )¯ = ¯ (y , ∗))(y ) ¯
¯ i! ¯ ¯ i! ¯
i=n+1 i=n+1
à ∞ !
X kf kδ µ δ − ρ ¶i−n−1
≤ i
· kykn+1
i=n+1
(δ − ρ) 2
X∞
kf kδ 1 2kf kδ
= n+1 t
kykn+1 = n+1
· kykn+1 = o(kykn ).
(δ − ρ) t=0
2 (δ − ρ)

We can now apply the converse of Taylor’s Theorem 4.4 and conclude that Di f = ϕi ,
which lies in Aρ (E, Lis (E; F )) and the bound for kDi f kρ is given in (5).

6.2 The Omega-lemma

Our next goal is to prove the so-called Omega-lemma in the spaces Aδ (E, F ) which
deals with composition of functions. The techniques involved follow the same pat-
tern as that of substitution of convergent power series into convergent power se-
ries (cf. [3] Ch. IX.2 and 5.5.3). So assume f ∈ Aδ (E, F ) and g ∈ Aη (D, E) with
kgkη ≤ δ. Then for each x ∈ D such that |x| ≤ η, we know that |g(x)| ≤ kgkη ≤ δ
and thus f (g(x)) makes sense, and even we have |f (g(x))| ≤ kf kδ . This suggests
that kf ◦ gkη ≤ kf kδ . But in order to establish this last inequality we need to
expand f ◦ g as a power series of continuous symmetric multilinear maps. All this
requires some explanations.
Assume ak ∈ Lk (E; F ) and bj ∈ Lj (D; E). Then

ak (bj1 (∗), . . . , (bjk (∗)) ∈ Lj1 +···+jk (D; F ).

Indeed, if ` = j1 + · · · + jk , the map

(x1 , . . . , x` ) 7→ ak (bj1 (x1 , . . . , xj1 ), . . . , bjk (x`−jk +1 , . . . , x` ))

from D` into F is obviously continuous and multilinear because ak and the bj ’s are
so. Observe that even in case ak and the bj ’s are symmetric, it is not clear whether
the preceding map is symmetric.

26
 P
∞ P
∞
Now assume f = ak ∈ Aδ (E, F ) and g = bj ∈ Aη (D, E) (with kgkη =
k=0 j=0
P
|bj |η j ≤ δ). We have, for x ∈ D, |x| ≤ η,
j
à !
P
∞ P
∞ P
∞
(f ◦ g)(x) = ak bj1 (xj1 ), . . . , bjk (xjk )
k=0 j1 =0 jk =0

P
∞ P
∞ P
∞
= ··· ak (bj1 (xj1 ), . . . , bjk (xjk )) ,
k=0 j1 =0 jk =0

the last equality due to continuity and multilinearity. But by absolute convergence
this can be rewritten grouping first those terms sharing a common sum j1 + · · · + jk ,
say equal to `. But before proceeding further and as remarked above observe that
ak (bj1 (∗), . . . , bjk (∗)) generally fails to be symmetric. For instance, we can easily
check this taking k = 2, ` = 3, paying attention to

a2 (b1 (x1 ), b2 (x2 , x3 )) + a2 (b2 (x1 , x2 ) + b1 (x3 ))

and comparing this to the corresponding expression with x1 and x2 interchanged.
If we momentarily denote ak (bj1 (∗), . . . , bjk (∗)) by f = f (x1 , . . . , x` ), and
f σ stands for the map
(x1 , . . . , x` ) 7→ f (xσ(1) , . . . , xσ(`) ),
for any permutation σ of {1, 2, . . . , `}, then the symmetrized map
1X σ
Sym` f = f
`! σ

obtained from f (cf. Remark 3.5 (i)) satisfies

¯ ¯ ¯ ¯
¯1 X ¯ 1 ¯X ¯ 1X σ
¯ ¯ ¯ ¯
|Sym` f | = ¯ f σ¯ = ¯ f σ¯ ≤ |f |,
¯ `! σ ¯ `! ¯ σ ¯ `! σ

and recalling Definition 3.2 (iv), we get

1X
|Sym` f | ≤ |f | = |f |.
`! σ

Of course when x1 = · · · = x` = x, we have (Sym` f )(x, . . . , x) = f (x, . . . , x) so that

the difference between f and Sym` f becomes clear when x1 , . . . , x` are unequal.
Coming back to our previous discussion we realize that we can write
∞
X
f ◦g = γ` ,
`=0

where ∞
X X
γ` = Sym` ak (bj1 (∗), . . . , bjk (∗))
k=0 j1 +···+jk =`

27
 P
is symmetric `-linear and continuous. Note that when ` > 0 the sum is
j1 +···+jk =`
void for k = 0.
As

¡ ¢ ¡ ¢
|Sym` ak bj1 (xj1 ), . . . , bjk (xjk | = |ak bj1 (xj1 ), . . . , bjk (xjk ) |
≤ |ak | · |bj1 (xj1 )| · · · |bjk (xjk )|
≤ |ak | · |bj1 | · |x|j1 · · · |bjk | · |x|jk
= |ak | · |bj1 | · · · |bjk | · |x|`

we see that

∞
X ∞ X
X ∞ X
kf ◦ gkη = |γ` |η ` ≤ |ak ||bj1 | · · · |bjk |η `
`=0 `=0 k=0 j1 +···+jk =`
ji ≥0
 
∞
X ∞
X X
 
=  |ak | |bj1 | · · · |bjk | η `
`=0 k=0 j1 +···+jk =`
ji ≥0
 
∞
X ∞
X X 
= |ak |  |bj1 | · · · |bjk |η ` 
k=0 `=0 j1 +···+jk =`
ji ≥0

∞
Ã∞ !k ∞ ∞
X X X X
j
= |ak | |bj |η = |ak | · kgkkη ≤ |ak | δ k = kf kδ .
k=0 j=0 k=0 k=0

What we have just proved may be summarized in the following

Lemma 6.6. If g ∈ Aη (D, E) with kgkη ≤ δ and f ∈ Aδ (E, F ), then

f ◦ g ∈ Aη (D, F ), and kf ◦ gkη ≤ kf kδ .

Next we prove another lemma which will be needed in the subsequent proposi-
tions.

Lemma 6.7. Let f ∈ Aδ (E, F ) and g, h ∈ Aη (D, E) such that kgkη = α < δ and
khkη ≤ β := 31 (δ − α). Then
° k ° ° k °
° D f (g(∗)) ° ° D f (∗) °
° k ° °
(h(∗) )° ≤ ° ° · khkkη .
° k! k! °
η α

28
 P
∞ P
∞ P
∞
Proof. Write f = ak , g = b` and h = h` , with ak ∈ Lks (E; F ), and
k=0 `=0 `=0
b` , h` ∈ L`s (D; E). Then, as seen in the proof of Theorem 6.5,
µ ¶
Dk f (g(x)) X j
= aj (g(x)j−k , ∗)
k! j≥k
k

and consequently,
X µ ¶
Dk f (g(x)) j
(h(x)k ) = aj (g(x)j−k , h(x)k ).
k! j≥k
k

Now we proceed to expand this last sum as in the arguments proving Lemma 6.6
k (g(x))
and have that D fk! (h(x)k ) is equal to
Xµj ¶ X
aj (b`1 (∗), . . . , b`j−k (∗), h`j−k+1 (∗), . . . , h`j (∗))x`1 +···+`j
j≥k
k ` ,...,` ≥0
1 j

from which, if ι = `1 + · · · + `j−k + `j−k+1 + · · · + `j , we see that

° k °
° D f (g(∗)) ° Xµj ¶ X
° (h(∗)k °
) ≤ |aj | · |b`1 | · · · |b`j−k | · |h`j−k+1 | · · · |h`j |η ι
° k! ° k
η j≥k `1 ,...,`j ≥0
à !j−k à !k
Xµj ¶ X X
= |aj | · |b` |η ` · |h` |η `
j≥k
k `≥0 `≥0
Xµj ¶
= |aj | · kgkηj−k · khkkη
j≥k
k
X j¶µ ° k °
°D f °
= |aj |α j−k
· khkη = °
k ° k
° k! ° · khkη
j≥k
k α

the last equality by virtue of formula (5) of Theorem 6.5.

Our final aim in this subsection, as mentioned earlier, is to prove the Ω-lemma,
i.e. differentiability of composition. To begin with, let us consider continuity.

Proposition 6.8. The map Ω : (f, g) 7→ f ◦ g from Aδ (E, F ) × U , where U = {g ∈

Aη (D, E) : kgkη < δ}, into Aη (D, F ) is continuous.

Proof. We have by Lemma 6.6

k(f + f1 ) ◦ g − f ◦ gkη = kf1 ◦ gkη ≤ kf1 kδ ,

29
and this entails Ω is uniformly continuous in the first variable and independently of
the second argument g. Now we examine what happens with the second variable.
Assume as in Lemma 6.7 that kgkη = α < δ and set β := 13 (δ − α), so that
α + β = δ − 2β < δ. By Theorem 6.5 we know that, for f ∈ Aδ (E, F ), its kth
derivative Dk f ∈ Aδ−2β (E, Lks (E; F )) and

kDk f kδ−2β ≤ k! · kf kδ · (2β)−k .

Take now h ∈ U such that khkη < β, which entails

kg + hkη ≤ kgkη + khkη < α + β < δ.

From
X∞
Dk f (g(x))
f ◦ (g + h)(x) − f ◦ g(x) = f (g(x) + h(x)) − f (g(x)) = (h(x)k ),
k=1
k!

we have,
°∞ ° ∞ ° k °
°X Dk f (g(∗)) ° X °D f °
° k ° ° ° k
° (h(∗) )° ≤ ° k! ° · khkη
° k! ° α
k=1 η k=1
X∞ ° °
X∞
° Dk f ° kf kδ
≤ ° ° k
· khkη ≤ · khkkη
° k! ° (2β)k
k=1 α+β k=1
X∞ ∞
kf kδ k−1 kf kδ X 1
≤ k
· β · khk η = · k
· khkη
k=1
(2β) β k=1
2
kf kδ
= · khkη ,
β

where the first inequality is due to Lemma 6.7, the second holds because α < α+β =
δ − 2β < δ, the third by Theorem 6.5, and the fourth is obvious. This implies
continuity with respect to the second argument. Now continuity with respect to
both variables follows from

|Ω(f 0 , g 0 ) − Ω(f, g)| = |Ω(f 0 , g 0 ) − Ω(f, g 0 ) + Ω(f, g 0 ) − Ω(f, g)|

≤ |Ω(f 0 , g 0 ) − Ω(f, g 0 )| + |Ω(f, g 0 ) − Ω(f, g)|.

Theorem 6.9. The map Ω of the previous proposition is C ∞ .

Proof. Keeping our previous notations, we first show that Ω has continuous partial
derivatives with respect to its second argument. In particular, we assume that khkη
is small enough. We have

30
 Xn
Dk f (g(x))
f ◦ (g + h)(x) = f (g(x) + h(x)) = f (g(x)) + (h(x)k )
k=1
k!
X∞
Dk f (g(x))
+ (h(x)k )
k=n+1
k!

and the k kη -norm of the last term, following the lines of the second part of the
proof of Proposition 6.8, is bounded by

kf kδ
· khkn+1
η = o(khknη ).
2n β n+1

As (Dk f ) ◦ g is continuous by Lemma 6.6, we can apply the converse of Taylor’s

Theorem 4.4 and thus D2k Ω, for 0 ≤ k ≤ n, does exist and equals Dk f ◦ g. Now,
since Dk f ∈ Aρ (E; Lks (E; F )), for all ρ < δ, by Theorem 6.5, and composition is
continuous by the previous proposition, D2k Ω is continuous in both arguments.
With respect to the first argument, we observe that (f, g) 7→ Ω(f, g) = f ◦ g is
linear (and continuous) in f , so that

D1 Ω(f, g) = Ω(∗, g)

(cf. Example 4.2 (i)), which is continuous by Proposition 6.8. But then

D12 Ω = D13 Ω = · · · = 0.

Let us check this for the most interesting case, namely that of D12 : As

D1 Ω(f + h, g) − D1 Ω(f, g) = Ω(∗, g) − Ω(∗, g) = 0,

the equality holding because D1 Ω(f, g) = Ω(∗, g) as just seen above, we get that
D12 Ω = 0 (cf. Example 4.2 (i) again).
Concerning mixed partial derivatives, as, for k ≥ 1, D2k Ω(f, g) = Dk f ◦ g, we
see that (f, g) 7→ D2k Ω(f, g) is continuous and linear in f , as in the case of Ω, and
therefore the same reasoning as before can be applied, so that

D1 D2k Ω(f, g) = D2k Ω(∗, g), and D1j D2k Ω(f, g) = 0, for j ≥ 2.

Thus Ω has continuous partial derivatives of all orders, i.e. Ω is C ∞ (cf. Proposition 5.4).

6.3 The Evaluation map

Another important theorem of global analysis is related to the Evaluation map. It
will be useful in considering some closed subspaces of the Banach spaces Aδ (E, F ),
as we will see in the next two sections.

31
Theorem 6.10. The Evaluation map

Ev : Aδ (E, F ) × {x ∈ E : |x| < δ} → F

defined by Ev(f, x) = f (x) is C ∞ , and we have

D2k Ev(f, x)(y k ) = Dk f (x)(y k ).

Proof. It follows the lines of the preceding proof since obviously Ev is linear in its
first variable f and is continuous. In fact, continuity is proved as in Proposition 6.8:
Ev is uniformly continuous with respect to f and independently of x because

|Ev(f + f1 , x) − Ev(f, x)|F = |f1 (x)|F ≤ kf1 kδ

and Ev is continuous with respect to x because, as said at the beginning of this

section, any f in Aδ (E, F ) yields a continuous function on {x ∈ E : |x| ≤ δ}. From
this, we see, as in the preceding proof, that

D1 Ev(f, x) = Ev(∗, x), and D12 Ev = D13 Ev = · · · = 0.

If we now fix the first variable f and let vary the second, i.e. x, we are just considering
the function x 7→ f (x) which according to Theorem 6.5 is C ∞ . This gives the result

D2k Ev(f, x) = Dk f (x),

as stated in the theorem. The mixed partial derivatives are dealt with in the
same way as in the preceding proof, since D2k Ev(f, x) is obviously linear in f , and
continuous because of Theorem 6.5, from which follows the C ∞ -differentiability of
Ev.

As a matter of fact and using the converse of Taylor’s Theorem 4.4 again we end
this section with an explicit expression of the kth derivative of the Evaluation map.
We have

Theorem 6.11. The kth derivative Dk Ev of the Evaluation map of Theorem 6.10
at (f, x) is the continuous k-linear map from (Aδ (E, F ) × E)k into F given by

Dk Ev(f, x)((g1 , y1 ), . . . , (gk , yk )) = Dk f (x)(y1 , . . . , yk )

Xk
+ Dk−1 gi (x)(y1 , . . . , ybi , . . . , yk ).
i=1

Proof. Let us denote by ϕk (f, x) the map from (Aδ (E, F ) × E)k into F given by
the right hand side of the equality in the statement of the Theorem 6.11. Obviously
ϕk (f, x) is symmetric in (g1 , y1 ), . . . , (gk , yk ), and ϕk (f, x) is k-linear and continuous
because Dk f (x) and Dk−1 gi (x) are both multilinear and continuous. Moreover, the
map (f, x) 7→ ϕk (f, x) is continuous: as f is continuous in {x ∈ E : |x| < δ}, it

32
turns out that ϕk is continuous in the second variable, and with respect to the first,
it is uniformly continuous by virtue of the bounds given in Theorem 6.5. Hence,
ϕk is continuous as it is indicated in the proof of Proposition 6.8. Now we are in a
situation where the converse of Taylor’s Theorem can be applied. In fact, for (g, y)
small,
Ev((f, x) + (g, y)) − Ev(f, x) = Ev(f + g, x + y) − Ev(f, x)
= (f + g)(x + y) − f (x) = f (x + y) − f (x) + g(x + y)
Pn
D k f (x) k P
∞
Dk f (x) k P Dk g(x) k
n−1 P
∞
D k g(x) k
= k!
(y ) + k!
(y ) + k!
(y ) + k!
(y )
k=1 k=n+1 k=0 k=n
n ³ k
P ´ P
∞ P
∞
D f (x) Dk−1 g(x) Dk f (x) D k g(x) k
= k!
(y k ) + (k−1)!
k−1
(y ) + k!
(y k ) + k!
(y ).
k=1 k=n+1 k=n

The last two series can be treated as in the last part of the proof of Theorem 6.5:
the first, involving f , is of order o(kykn ), and the second may be bounded in norm
by
2kgkδ
n+1
· kykn , for a suitable ρ < δ,
(δ − ρ)

which tells us that it is of order o(k(g, y)kn ), so that

Ev((f, x) + (g, y)) − Ev(f, x)

Xn
1 k
= (D f (x)(y k ) + k Dk−1 g(x)(y k−1 )) + o(k(g, y)kn )
k=1
k!
Xn
1
= ϕk (f, x)(y k ) + o(k(g, y)kn ),
k=1
k!

and the theorem follows by virtue of the converse of Taylor’s Theorem 4.4.

33
7 Poincaré’s Linearization Theorem

In this section we want to prove Poincaré’s linearization Theorem for an analytic

map from Rn or Cn into itself near a fixed point. It states that under certain
conditions it is conjugate, via an analytic isomorphism, to its linear part (see [5],
cf. [1] Ch.5).
We begin with the complex case i.e., that of Cn , but first let us prove a technical
lemma.

Lemma 7.1. The subspace

V = {u ∈ Aδ (Cn , Cn ) : u(0) = Du(0) = 0}

of Aδ (Cn , Cn ) is closed in Aδ (Cn , Cn ) and, as a result, V is a Banach space under

the induced norm k kδ .

Proof. We give two proofs.

First proof. It is easy to check that the complementary set of V is open in
Aδ (Cn , Cn ): in the expansion of any u not in V some vector coefficient corresponding
to zeroth or first order has to be nonzero and by definition of the norm k kδ , an
ε > 0 can be found so that any u e in the ball centered at u with radius ε has a
nonvanishing coefficient of order < 2, i.e., the ball considered has no elements of V .
More specifically, in case u(0) = u0 is nonzero, it suffices to take 0 < ε < |u0 |, since
if ∞
X
kv − ukδ = |vk − uk | δ k < ε,
k=0

then, in particular, |v0 − u0 | < ε, which is not possible if v0 = 0, i.e., v ∈

/ V . And
similarly, if u1 = Du(0) 6= 0, then it suffices to take 0 < ε < |u1 | δ, in which case,
from kv − ukδ < ε, we would have |v1 − u1 | < ε, which is not satisfied if v1 = 0.
Second proof. The maps u 7→ u(0) and u 7→ Du(0) defined on Aδ (Cn , Cn ) are
continuous, because the Evaluation map

Ev : Aδ (Cn , Cn ) × {x ∈ Cn : |x| < δ} → Cn

is C ∞ -differentiable and D2 Ev(f, 0) = Df (0), as stated in Theorem 6.10. Thus the

preimages of zero of the former two maps are closed, but V is just the intersection
of these two closed preimages and is therefore closed.

Now we state Poincaré’s Theorem:

34
Theorem 7.2. Let Φ ∈ Aδ (Cn , Cn ) be such that Φ(0) = 0, and A := DΦ(0) is an
n × n- matrix that satisfies the following conditions:

(i) A diagonalizes.

(ii) All eigenvalues λ1 , . . . , λn of A are nonzero and less than one in modulus.

(iii) For all nonnegative integers k1 , . . . , kn such that k1 + · · · + kn ≥ 2 and for all
j = 1, . . . , n, one has λj 6= λk11 · · · λknn (non-resonance condition).

Then there exist η > 0 and Ψ ∈ Aη (Cn , Cn ) such that Ψ(0) = 0, DΨ(0) = identity
and Ψ ◦ Φ ◦ Ψ−1 = A.
(So that Ψ is an analytic change of variables near the origin in Cn which by
conjugation linearizes Φ).

Proof. As Φ ∈ Aδ (Cn , Cn ) with Φ(0) = 0, and DΦ(0) = A, we can write

Φ(x) = A(x) + g(x)

with g ∈ Aδ (Cn , Cn ) of order ≥ 2, i.e., with g(0) = Dg(0) = 0, and we contend that
there exists u ∈ Aδ (Cn , Cn ) of order ≥ 2 such that if we take as new coordinates
Ψ(x) = w = x+u(x), Φ is described in these coordinates by the transformation w 7→
Aw. This can be more clearly viewed in terms of the existence of a commutative
diagram (actually Φ is only defined for x ∈ Cn with |x| ≤ δ)

Φ / Cn
Cn
Ψ Ψ
² A
²
Cn / Cn

and taking elements

Φ / Φ(x)
x
Ψ Ψ
² A
²
w / Aw

for a suitable Ψ such that Ψ(x) = w = x + u(x) with u ∈ Aδ (Cn , Cn ) of order ≥ 2.

Following the diagram we get in one direction

(A ◦ Ψ)(x) = A(Ψ(x)) = A(x + u(x)) = Ax + Au(x)

and in the other

(Ψ ◦ Φ)(x) = Ψ(Φ(x)) = Ψ(Ax + g(x)) = Ax + g(x) + u(Ax + g(x))

35
so that the condition of commutativity reads

Ax + Au(x) = Ax + g(x) + u(A(x) + g(x))

and cancelling the Ax terms, we are led to the following functional equation for u:

Au(x) − g(x) − u(Ax + g(x)) = 0

As all eigenvalues of A are less than one in modulus, there exists a norm on A
such that |A| < α < 1. This allows us to write

|Ax + g(x)| ≤ |Ax| + |g(x)| ≤ αδ + kgkδ ,

if |x| ≤ δ and leads us to introduce the function F : U × U0 → U , where U0 is the

centered ball of radius (1 − α)δ in the Banach space

U = {ϕ ∈ Aδ (Cn , Cn ) : ϕ(0) = Dϕ(0) = 0}

(see Lemma 7.1), defined by

F (u, g)(x) = Au(x) − g(x) − u(Ax + g(x))

Observe that F is well defined and C ∞ -differentiable by virtue of Lemma 6.6,

Theorem 6.9 and the chain rule.
Now, it is clear that F (0, 0) = 0. Our aim is to show that given g (in our case
g is given by Φ), there exists u (as a function, actually of class C ∞ , of g) satisfying
F (u, g) = 0 and for this we will need to apply the Implicit Function Theorem 5.5,
so we have to deal with the first partial derivative of F with respect to the first
argument at (0, 0).
As F (u, 0)(x) = Au(x) − u(Ax), we immediately see that u 7→ F (u, 0) is contin-
uous and linear in u, and this being the case, its derivative is the same continuous
linear function (defined on U ), for any u, and in particular when u = 0. Thus

D1 F (0, 0)(v) = Av − v ◦ A.

Let us denote D1 F (0, 0) by L. Now, to satisfy the requirements of the Implicit

Function Theorem 5.5 we need to prove that L is invertible. Let us seek a formal
complex solution for v to Lv = w where v and w are formal power series of order
≥ 2 say in the indeterminates x1 , . . . , xn (with coefficients in Cn ). Recall that
Lv = Av − v ◦ A. As we have assumed that A diagonalizes, there exists an invertible
complex n×n-matrix P such that P −1 AP = D =diag(λ1 , . . . , λn ). Introducing new
indeterminates y1 , . . . , yn such that x = P y, where x = (x1 , . . . , xn )T , and similarly
for y, and setting

ν(y) := P −1 v(P y), ω(y) := P −1 w(P y),

we see that the equation Lv = w may be translated into the form Λν = ω,

36
where
Λν(y) = Dν(y) − ν(Dy),
because

P (Dν − ν ◦ D) ◦ P −1 = P DP −1 P ν ◦ P −1 − P ν ◦ (P −1 P DP −1 )
= Av − v ◦ A = Lv = w = P ω ◦ P −1 .

We proceed now, after diagonalization, to isolate ν in Λν = ω, but this requires

working with explicit expansions for ν and ω. So let us consider the set K of
integer vectors k = (k1 , . . . , kn ) with ki ≥ 0 and abbreviate y1k1 · · · ynkn to y k and set
|k| := k1 + · · · + kn . With these notations, if the expansions of ν and ω are
X X
ν(y) = νk y k , ω(y) = ωk y k ,
|k|≥2 |k|≥2

(t) (t)
and ωk and νk , for t = 1, 2, . . . , n, are the respective components of the vectors
ωk and νk , we have for each t, that the t-component of Λν(y) is given by
X (t) X (t)
λt νk y k − νk (λy)k ,
|k|≥2 |k|≥2

where (λy)k means (λ1 y1 )k1 · · · (λn yn )kn = λk11 · · · λknn y1k1 · · · ynkn = λk y k .
But this is just X (t)
(λt − λk )νk y k (t = 1, . . . , n)
|k|≥2

or in vector notation, Mk standing for the matrix diag (λ1 − λk , · · · , λn − λk ) =

D − λk I,
X X
Λν(y) = (D − λk I) νk y k = Mk νk y k ,
|k|≥2 |k|≥2

We have that Mk is invertible for all k, |k| ≥ 2, in view of assumption (iii).

Consequently, X X
Λν(y) = Mk νk y k = ω(y) = ωk y k
|k|≥2 |k|≥2

and we get νk = Mk−1 ωk , which yields a formal solution to Λν = ω or equivalently

to Lv = w.
But, by condition (ii), we have λk = λk11 . . . λknn → 0 when the kj ’s are large
and this implies that both Mk and Mk−1 are bounded, for any k ∈ K, |k| ≥ 2. In
particular, there exists a real number H such that |Mk−1 | < H, for all these k, and
this implies |νk | = |Mk−1 ωk | < H |ωk |, from which we see that ν ∈ Aδ (Cn , Cn ),
whenever ω ∈ Aδ (Cn , Cn ), and thus D1 F (0, 0) is invertible.
Now that we have just proved that D1 F (0, 0) is invertible, the Implicit Function
Theorem 5.5 assures us the existence of an ε > 0 such that u is uniquely determined

37
by g as far as kgkδ < ε and kukδ < ε. In order to get rid of this restriction
P on g we
−1
can rescale as follows: We set ge(x) := α g(αx), for α ∈ (0, 1]. If g = gk , with
k≥2
gk ∈ Lks (Cn , Cn ),
then by definition of k kδ , we have
X X X
g kδ = α−1
ke |gk | αk δ k = α |gk | αk−2 δ k ≤ α |gk | δ k = αkgkδ
k≥2 k≥2 k≥2

the last inequality holding because 0 < α ≤ 1. This allows us to choose α such
that ke
g kδ < ε, and by the Implicit Function Theorem 5.5, for this ge there exists a
unique ue ∈ Aδ (Cn , Cn ) such that F (e u(α−1 x) for
u, ge) = 0. Next, define u(x) := αe
the α chosen. Now, F (e u, ge) = 0 means

Ae
u(x) − u
e(Ax + ge(x)) − ge(x) = 0,

which, in terms of u and g, is

Aα−1 u(αx) − α−1 u(αAx + g(αx)) − α−1 g(αx) = 0.

Cancelling α−1 , we have under the change of variables αx = y

Au(y) − u(Ay + g(y)) − g(y) = 0,

P
i.e. F (u, g)(y) = 0. And setting η = αδ, if u
e= u
ek , we see that
k≥2

X X ³ η ´k X
kukη = α uk α−k |η k = α
|e |e
uk | =α uk |δ k = αke
|e ukδ < ∞,
k≥2 k≥2
α k≥2

so that u ∈ Aαδ (Cn , Cn ).

Corollary 7.3. Poincaré’s Theorem 7.2 holds true if Cn is replaced by Rn every-

where, i.e., if Φ ∈ Aδ (Rn , Rn ), so that in particular A = DΦ(0) has real entries,
then Ψ can be found in Aη (Rn , Rn ) satisfying Ψ ◦ Φ ◦ Ψ−1 = A.

Remark 7.4. The corollary establishes that all transformations involved are real
independently of the fact that the eigenvalues λ1 , . . . , λn of A may not be real.

Proof. (Keeping the notations occurring both in Poincaré’s Theorem 7.2 and in its
proof)
If A is real and λ is a non-real eigenvalue of A then its conjugate λ is also
another eigenvalue of A and obviously if v is a complex eigenvector corresponding
to λ, then v is an eigenvector corresponding to λ (since from λv = Av we get
λv = λv = Av = Av because A = A). These facts allow us to consider a complex
basis of eigenvectors

{e1 , . . . , e` , e`+1 = e1 , . . . , e2` = e` , e2`+1 , . . . , en }

38
corresponding to the eigenvalues
{λ1 , . . . , λ` , λ`+1 = λ1 , . . . , λ2` = λj , λ2`+1 , . . . , λn },
the 2` first of which are non-real and the last n − 2` are real. Let P be the complex
n×n-matrix whose columns are the (complex) coordinates of the ei ’s in the standard
basis of Cn , i.e., the matrix such that P −1 AP = diag(λ1 , . . . , λn ) = D, and let Q
stand for the matrix which interchanges the complex conjugate vectors of the basis
{e1 , . . . , en } i.e. the matrix  
0 I` 0
 I` 0 0 .
0 0 In−2`

From this, it is obvious that QD = DQ and P = P Q. Now, as in the proof of

the theorem, set
x = P y, ν(y) = P −1 v(P y), ω(y) = P −1 w(P y)
and observe that v and w are real if and only if v(x) = v(x) and w(x) = w(x)
(since x real means x = x). Let us express this in terms of ν and ω: v(x) = v(x)
in terms of ν = P −1 v ◦ P is expressed as P ν ◦ P −1 = P ν ◦ P −1 , but from P = P Q,
we see that P ν ◦ P −1 = P Qν ◦ (Q−1 P −1 ), i.e. ν = Qν ◦ Q−1 or Qν = ν ◦ Q (as
expected since Q interchanges conjugate basis vectors), and similarly Qω = ω ◦ Q
or ω = Q−1 ω ◦ Q, which means that ωk = Qωq for q = kQ, where q = (q1 , . . . , qn )
and k = (k1 , . . . , kn ).
Coming back to the proof of the theorem, we have
−1 k
νk = Mk ωk = (D − λ I)−1 Qωq = Q(D − λk I)−1 ωq = Qνq
i.e., Qν = ν ◦ Q so that v is real when w is real.

Remark 7.5. Observe that the proof of this corollary is vacuous if all eigenvalues
of A are real.

Now, we will deal with the resonant case

Theorem 7.6. Let us just drop assumption (iii) in Theorem 7.2 and keep the same
notations. Then the conjugate Ψ ◦ Φ ◦ Ψ−1 sends w to Aw + h(w) where h lies in
the kernel of L, i.e. Ah(w) − h(Aw) = 0, and h is a polynomial.

Proof. We follow the lines of the proof of Poincaré’s Theorem 7.2 and consider the
set (of resonant terms):
© ª
S = (j, k) : j ∈ {1, . . . , n}, k ∈ K, |k| ≥ 2, λj = λk .

By condition (ii) we see that λk = λk11 · · · λknn → 0 as the kj ’s become large. This
implies that the set S is finite. Then ν ∈ Ker Λ if and only if ν is of the form
X
ν(y) = α(j,k) y k ej ,
(j,k)∈S

39
and as S is finite, Ker Λ is finite-dimensional. Here ej stands for the jth element of
the canonical basis of Cn .
Let now π be the projection of Aδ (Cn , Cn ) defined by
X (j)
π ν(y) = νk y k ej ,
(j,k)∈S

i.e., π sends ν(y) to the finite sum of its resonant terms i.e., those where λj − λk
vanishes.
Obviously π is linear and kπ νkδ ≤ kνkδ , which implies π is continuous.
Under these circumstances we can assure (see [3] 5.4.2) that Aδ (Cn , Cn ) is the
topological direct sum of the Banach spaces π Aδ (Cn , Cn ) and (I − π) Aδ (Cn , Cn ).
It is clear that Λ π = π Λ = 0 and that Λ sends (I − π) Aδ (Cn , Cn ) into itself.
Furthermore, Λ has a (continuous) inverse on the Banach space

U = {` ∈ (I − π) Aδ (Cn , Cn ) : `(0) = D`(0) = 0} .

Let U0 be the open (centered) ball of radius δ in the preceding Banach space,

V = {h ∈ π A2δ (Cn , Cn ) : h(0) = Dh(0) = 0} ,

© ª
W = g ∈ A(1−α)δ (Cn , Cn ) : g(0) = Dg(0) = 0 ,
and
Z = {m ∈ Aδ (Cn , Cn ) : m(0) = Dm(0) = 0} .

Recall that A2δ (Cn , Cn ) ⊂ Aδ (Cn , Cn ). Now, proceeding as at the beginning of the
proof of Poincaré’s Theorem 7.2, we are led to consider the functional equation
which both u and h must satisfy:

F (u, h, g) = A u(x) − u (Ax + g(x)) − g(x) + h(x + u(x)) = 0,

for F : U0 × V × W → Z.
As before, and bearing in mind that by restricting C ∞ -differentiable functions to
subspaces we get C ∞ -differentiable functions again, we have that F is well-defined
and C ∞ -differentiable, and as

F (0, 0, 0) = 0, F (u, 0, 0) = A u − u ◦ A, and F (0, h, 0) = h,

we see that
e and D2 F (0, 0, 0) = I,
D1 F (0, 0, 0) = L| (I − π) Aδ (Cn , Cn ) =: L,

and these partial derivatives clearly define a continuous homomorphism from U × V

into Z, namely (u, h) 7→ L e u + h. If we show this is invertible we can apply the
Implicit Function Theorem 5.5 and rescale as before to conclude the proof. But
invertibility is immediate: In order to uniquely solve
ev +h = g
L

40
for v ∈ (I − π) Aδ (Cn , Cn ) and h ∈ π A2δ (Cn , Cn ), for any given g ∈ Aδ (Cn , Cn ),
with all three functions of order at least two, we see that we have to take precisely
h = π g (which is a polynomial and therefore lies in π Aη (Cn , Cn ), for any η > 0)
and v = L e−1 (I − π)g.

Corollary 7.7. The statement of Corollary 7.3 is also true in the case of Theo-
rem 7.6, i.e., in the real case.

Proof. It is the same as that of Corollary 7.3 except for the case of the resonant terms
(j, k) ∈ S. Recalling that v is real if and only if v(x) = v(x), i.e., Q ν(y) = ν(Q y),
and this obviously holds when ν is replaced by (I − π) ν, there remains the case of
π ν, i.e. we are asking whether Q π ν(y) = π ν(Q y). But this is true because if A is
k
a real matrix and (j, k) ∈ S, i.e., λj = λk , then λj = λ is also a resonant term for
A (recall that the set of eigenvalues of A is invariant by conjugation).

We end this section covering the case when A does not diagonalize. It turns out
that the same conclusions as those of the preceding theorems hold.
Theorem 7.8. If we drop condition (i ) in Poincaré’s Theorem 7.2, the same con-
clusion holds. And this also occurs in Theorem 7.6, Corollary 7.3 and Corollary 7.7.

Proof. If A does not diagonalize, by the theory of the Jordan blocks there exists an
invertible matrix P such that P −1 A P = D + N with D diagonal and N nilpotent,
and we can further assume N is small (this can easily be achieved by substituting the
vectors of a Jordan basis by suitable multiples of them). Then with our previous
changes, namely v = P ν ◦ P −1 , w = P ω ◦ P −1 , the equation L v = w, with
L v = A v − v ◦ A, is transformed into

Av − v ◦ A = P (D + N )P −1 P ν ◦ P −1 − P ν ◦ [P −1 P (D + N )P −1 ]
= P (D + N ) ν ◦ P −1 − P ν ◦ [(D + N )P −1 ]
= P [(D + N ) ν − ν ◦ (D + N )] ◦ P −1
= P [D ν + N ν − ν ◦ (D + N )] ◦ P −1
= P [D ν − ν ◦ D + N ν − ν ◦ (D + N ) + ν ◦ D] ◦ P −1 = P ω ◦ P −1

i.e., Λ ν = ω, where Λ = Λ1 + Λ2 , with

Λ1 ν(y) = D ν(y) − ν (Dy), Λ2 ν(y) = N ν(y) − (ν(D y + N y) − ν (Dy)).

These expressions allow us to conclude that Λ1 is, as seen before, invertible (in the
resonant case, leaving aside the resonant terms), and Λ2 is small. As the invertible
elements in the Banach space of continuous linear maps between two Banach spaces
form an open set (see [2] I, Thm. 1.7.3(a)) we see that Λ = Λ1 + Λ2 can be made
invertible, which is all we need.

41
8 The Analytic Stable Manifold Theorem
This section deals with the analytic Stable Manifold Theorem:

Theorem 8.1. Let C be an n × n real nonsingular matrix with k eigenvalues less

than one in modulus and the other n − k eigenvalues greater than one in modulus
and Φ ∈ Aδ (Rn , Rn ) be such that Φ(0) = 0 and D Φ(0) = C. Then there exists a
neighbourhood N of 0 in Rn such that the subset of those p ∈ N satisfying Φn (p) ∈
N , for all n ∈ N, is a k-dimensional real analytic submanifold W s of N . In fact,
for any p ∈ W s , we have lim Φn (p) = 0.
n→∞
s
(We say that W is the stable submanifold of Φ).

Proof. The assumption on the eigenvalues of C allows a linear change of variables

leading to the case when µ ¶
A 0
C=
0 B

with A a real k × k-matrix with norm |A| = α < 1 and B a real (n − k) × (n − k)-
matrix such that the norm |B −1 | = β < 1. We will write Rn = Rk × Rn−k and
(x, y) ∈ Rk × Rn−k , so that Φ may be described as

Φ : (x, y) 7→ (x∗ , y ∗ ) = (Ax + f (x, y), By + g(x, y))

where f ∈ Aδ (Rn , Rk ), g ∈ Aδ (Rn , Rn−k ), with both f and g of order ≥ 2. Let us

seek a change of variables of the form ξ = x, η = y −h(x) for some h ∈ Aδ (Rk , Rn−k )
of order ≥ 2 such that the ξ-axis is Φ-invariant. Then we will show that the ξ-axis,
i.e. the graph of h, is W s .
In the new variables {ξ, η}, Φ may be described as

(ξ, η) 7→ (ξ ∗ , η ∗ ) = (A ξ + f ∗ (ξ, η), B η + g ∗ (ξ, η))

with both f ∗ and g ∗ of order ≥ 2.

The relationship between f, g and f ∗ , g ∗ comes from the fact that if (x, y) corre-
sponds to (ξ, η) then (x∗ , y ∗ ) must correspond to (ξ ∗ , η ∗ ), i.e.

ξ ∗ = x∗ , and η ∗ = y ∗ − h(x∗ ).

Paying attention only to the last equation, note it can be rewritten as

B η + g ∗ (ξ, η) = B y + g(x, y) − h(A x + f (x, y))

and after substituting x by ξ and y by η + h(x) = η + h(ξ), we arrive at

B η + g ∗ (ξ, η) = B(η + h(ξ)) + g(ξ, η + h(ξ)) − h(A ξ + f (ξ, η + h(ξ))),

42
or, cancelling B η in both members,
g ∗ (ξ, η) = B(h(ξ)) + g(ξ, η + h(ξ)) − h(A ξ + f (ξ, η + h(ξ))).
Observe that the ξ-axis η = 0 is invariant under Φ if and only if η = 0 implies
η ∗ = 0. As η ∗ = B η + g ∗ (ξ, η) we realize that this is exactly the same as requiring
g ∗ (ξ, 0) = 0, or using the above expression for g ∗ (ξ, η), that h has to satisfy the
following functional equation
B(h(ξ)) + g(ξ, h(ξ)) − h(A ξ + f (ξ, h(ξ))) = 0.
Let F (h, f, g)(ξ) stand for the left hand side of this functional equation, so that F
can be seen as a map from U0 × V0 × W into U , where U0 is the centered ball of
radius δ in the Banach space
U = {h ∈ Aδ (Rk , Rn−k ) : h(0) = Dh(0) = 0},
V0 is the centered open ball of radius (1 − α)δ of the Banach space
V = {f ∈ Aδ (Rn , Rk ) : f (0) = Df (0) = 0}
and W is the Banach space
W = {g ∈ Aδ (Rn , Rn−k ) : g(0) = Dg(0) = 0}.

By Lemma 6.6, Theorem 6.9 and the chain rule, F : U0 × V0 × W → U is

well-defined and C ∞ -differentiable.
Obviously F (0, 0, 0) = 0 and as F (h, 0, 0) = B h − h ◦ A is linear in h, we have
that the first partial derivative D1 F (0, 0, 0) of F at (0, 0, 0) is the continuous linear
map Λ : U → U defined by Λ(`) = B ` − ` ◦ A. The map Λ is invertible: in fact,
∞
X
Λ−1 (`)(ξ) = B −1 `(ξ) + B −2 `(A ξ) + B −3 `(A2 ξ) + · · · = B −s−1 `(As ξ).
s=0

Formally
Λ−1 (Λ`) = B −1 (B` − ` ◦ A) + B −2 (B` − ` ◦ A) ◦ A + · · ·
= (` − B −1 ` ◦ A) + (B −1 ` ◦ A − B −2 ` ◦ A2 ) + · · · = `,

and
Λ(Λ−1 `) = B(B −1 ` + B −2 ` ◦ A + B −3 ` ◦ A2 + · · · )
−(B −1 ` + B −2 ` ◦ A + B −3 ` ◦ A2 + · · · ) ◦ A
= ` + B −1 ` ◦ A + B −2 ` ◦ A2 + · · · − (B −1 ` ◦ A + B −2 ` ◦ A2 + · · · ) = `.
P
∞
But the series B −s−1 ` ◦ As obviously converges in U , since by Lemma 6.6,
s=0

∞
X ∞
X
−s−1 s β
kB ` ◦ A kδ ≤ β s+1 k`kδ = k`kδ .
s=0 s=0
1−β

43
 β
This also shows that kΛ−1 k ≤ 1−β . Now we can apply the Implicit Function
Theorem 5.5 and conclude that there exists an ε > 0 and a unique h ∈ U0 which
is a C ∞ -function of (f, g), for f ∈ V0 , kf kδ < ε and g ∈ W, kgkδ < ε, satisfying
F (h, f, g) = 0. But we would like to solve F (h, f, g) = 0 without the assumptions
that f and g have to be small in norm. This can be done by rescaling as we have
done in the proof of the Poincaré’s Theorem 7.2, and we eventually see that there
exists an α > 0 and an h ∈ Aαδ (Rk , Rn−k ) satisfying F (h, f, g) = 0. So we have
shown there is a change of variables

x = ξ, y = η − h(ξ)

such that the ξ-axis is Φ-invariant. Considering now Φ−1 instead of Φ, we see there
is a change of variables leaving the η-axis invariant. After making both changes of
variables we consider Φ again but in the new variables: As before,

Φ(ξ, η) = (ξ ∗ , η ∗ ) = (Aξ + f ∗ (ξ, η), Bη + g ∗ (ξ, η)),

with f ∗ (0, 0) = Df ∗ (0, 0) = 0 and g ∗ (0, 0) = Dg ∗ (0, 0) = 0, but now, by the

Φ-invariance of the axes, we also have f ∗ (0, η) = 0, g ∗ (ξ, 0) = 0.
By the Mean Value Theorem(see [4] 2.4.8) there is a neighbourhood N of the
origin in Rn and a ϑ ∈ (0, 1) such that

|Aξ + f ∗ (ξ, η)| < ϑ |ξ|, and |Bη + g ∗ (ξ, η)| > ϑ−1 |η|, for all (ξ, η) ∈ N.

Let us justify the former inequality. So consider the map

(ξ, η) 7→ ξ ∗ = Aξ + f ∗ (ξ, η).

As f ∗ (0, η) = 0, we have by the Mean Value Theorem

|A ξ +f ∗ (ξ, η)| ≤ |A ξ|+|f ∗ (ξ, η)−f ∗ (0, η)| ≤ |A|·|ξ|+ sup |Df ∗ (t ξ, η)|·|ξ| < ϑ1 |ξ|,
0≤t≤1

for some ϑ1 ∈ (0, 1), since |A| = α < 1 and Df ∗ (t ξ, η) is near 0 if (ξ, η) is near the
origin, because Df ∗ (0, 0) = 0. The same reasoning applies to the map (ξ, η) 7→ η ∗ =
Bη + g ∗ (ξ, η) to get near the origin |B η + g ∗ (ξ, η)| > ϑ−1
2 |η|, for some ϑ2 ∈ (0, 1)
and, of course, by replacing ϑ1 and ϑ2 by max(ϑ1 , ϑ2 ), we can assume ϑ1 = ϑ2 .
Let (ξ0 , η0 ) be in N and set (ξn , ηn ) = Φn (ξ0 , η0 ) where Φn stands for the nth
iterate of Φ. Then the preceding inequalities given by the Mean Value Theorem
imply that |ξn | < ϑn (ξ0 ) and |ηn | > ϑ−n |η0 | as long as (ξn , ηn ) ∈ N . So, if η0 = 0,
and ξ0 is small, i.e., (ξ0 , 0) ∈ N , we obtain

lim (ξn , ηn ) = lim (ξn , 0) = 0.

n→∞ n→∞

However, if η0 6= 0, then given any compact set K of N , an n can be found such

that (ξn , ηn ) ∈
/ K, and this concludes the proof of the theorem.

44
9 Conclusions
In this work we have dealt with some essential topics in Mathematical Analysis
which do not seem to be widely known, namely the converse of Taylor’s Theorem,
the Omega-lemma and the Evaluation map. In this work we have introduced and
elaborated in detail the proofs of these important theorems of Global Analysis
in the case of the Banach spaces Aδ (E, F ) since these are those we have been
interested in, because with them we have supplied nice proofs of the deep theorems
of Poincaré concerning linearization of certain analytic maps and of the Analytic
Stable Manifold Theorem. Guided by the sketchy paper of Meyer [5], we have relied
on good standard books such as Arnold [1], Cartan [2], Dieudonné [3], and Marsden
[4]. In the end we realize that functional analysis, in our case concerning the key
spaces Aδ (E, F ), turns out to be useful not only in itself but also in important
applications as those considered in this work.

45
References
[1] Arnold, V.: Chapitres Supplémentaires de la Théorie des Équations
Différentielles Ordinaires, Éditions MIR, Moscou, 1980.

[2] Cartan, H.: Cours de Calcul Différentiel, Collection Méthodes, Hermann,

Paris, 1977.

[3] Dieudonné, J.: Foundations of Modern Analysis, Academic Press, New York,
1960.

[4] Marsden, J. E.; Ratiu, T. (with collaboration of Abraham, R.): Manifolds, Ten-
sor Analysis, and Applications, Applied Mathematical Sciences 75. Springer,
3rd edition, New York, 2001.

[5] Meyer, K. R.: The implicit function theorem and analytic differential equa-
tions, Warwick Dynamical Systems 1974, Lecture Notes in Mathematics 468,
Springer-Verlag, New York-Berlin, 1975.

[6] Nelson, E.: Topics in Dynamics I: Flows, Princeton University Press, New
York, 1969.

[7] Palis, J. Jr.; de Melo, W.: Geometric theory of dynamical systems. An intro-
duction. Springer-Verlag, New York-Berlin, 1982.

46