INSTRUCTIONS

To produce the compuscript, Latex twice the file
 (so as to get correct cross references).


 The text between the lines BODY and ENDBODY is made  of


  1864 lines and    68955 bytes (not counting <line feed> or <carriage return>)



 In the following table this count is broken down by ASCII  code;
 immediately following the code is the corresponding  character.

     41142  lowercase letters
      1954  uppercase letters
      2090  digits

      8349  ASCII characters  32  
        17  ASCII characters  33 !
        43  ASCII characters  34 "
        52  ASCII characters  35 #
      1720  ASCII characters  36 $
       734  ASCII characters  37 %
        46  ASCII characters  38 &
       169  ASCII characters  39 '
       715  ASCII characters  40 (
       764  ASCII characters  41 )
       144  ASCII characters  42 *
        92  ASCII characters  43 +
       830  ASCII characters  44 ,
       293  ASCII characters  45 -
       634  ASCII characters  46 .
         3  ASCII characters  47 /
       127  ASCII characters  58 :
        75  ASCII characters  59 ;
         5  ASCII characters  60 <
       309  ASCII characters  61 =
        19  ASCII characters  62 >
         2  ASCII characters  64 @
        56  ASCII characters  91 [
      3574  ASCII characters  92 \
        56  ASCII characters  93 ]
       395  ASCII characters  94 ^
       906  ASCII characters  95 _
        84  ASCII characters  96 `
      1734  ASCII characters 123 {
        53  ASCII characters 124 |
      1734  ASCII characters 125 }
        35  ASCII characters 126 ~

 BODY
%%% This is a Latex file producing the compuscript entitled
%%% "A Direct Proof of a Theorem by Kolmogorov
%%%  in Hamiltonian Systems" 
%%% by L.Chierchia and C. Falcolini
%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% FORMATO GRANDE
%
\documentstyle[amssymbols,12pt]{article}
\setlength{\hoffset}{-1.cm}
\setlength{\voffset}{-1.8cm}
\setlength{\textwidth}{ 15.8cm}
\setlength{\textheight}{22cm}
\setlength{\parindent}{8mm}
\setlength{\footskip}{2.truecm}
\frenchspacing
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%
%%%        FOR SHORT
%%%
%%%
\newcommand\beq[1]{ \begin{equation}\label{#1} }
\newcommand{\eeq}{ \end{equation} }
\newcommand{\beqno}{ \[ }
\newcommand{\eeqno}{ \] }
\newcommand\beqa[1]{ \begin{eqnarray} \label{#1}}
\newcommand{\eeqa}{ \end{eqnarray} }
\newcommand{\beqano}{ \begin{eqnarray*} }
\newcommand{\eeqano}{ \end{eqnarray*} }
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%
%%%        THEOREMS AND ALIKE
%%%
%%%
\renewcommand{\theequation}{\arabic{section}.\arabic{equation}}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{definition}{Definition}[section]
\newtheorem{proposition}{Proposition}[section]
\newtheorem{lemma}{Lemma}[section]
\newtheorem{remark}{Remark}[section]
\newtheorem{sublemma}{Sublemma}[section]
\newtheorem{corollary}{Corollary}[section]
\newcommand\dfn[1]{ \begin{definition}\label{#1} \rm}
\newcommand\dfntwo[2]{ \begin{definition}[#1]\label{#2} \rm}
\newcommand\edfn{ \end{definition} }
\newcommand\rem[1]{ \begin{remark}\label{#1} \rm}
\newcommand\erem{ \end{remark} }
\newcommand\thm[1]{ \begin{theorem}\label{#1}}
\newcommand\thmtwo[2]{ \begin{theorem}[#1]\label{#2}}
\newcommand\ethm{ \end{theorem} }
\newcommand\pro[1]{ \begin{proposition}\label{#1}}
\newcommand\protwo[2]{ \begin{proposition}[#1]\label{#2}}
\newcommand\epro{ \end{proposition} }
\newcommand\lem[1]{ \begin{lemma}\label{#1}}
\newcommand\lemtwo[2]{ \begin{lemma}[#1]\label{#2}}
\newcommand\elem{ \end{lemma} }
\newcommand\sublem[1]{ \begin{sublemma}\label{#1}}
\newcommand\sublemtwo[2]{ \begin{sublemma}[#1]\label{#2}}
\newcommand\esublem{ \end{sublemma} }
\newcommand\cor[1]{ \begin{corollary}\label{#1}}
\newcommand\cortwo[2]{ \begin{corollary}[#1]\label{#2}}
\newcommand\ecor{ \end{corollary} }
%%%
%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%
%%%       REFERRING TO
%%%
%
\newcommand\equ[1]{{\rm (\ref{#1})}}
%
%
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%
%%%       VARIOUS
%%%
\newcommand{\ie}{{\it i.e. }}
\newcommand{\eg}{{\it e.g. }}
\newcommand{\etc}{{\it etc}}
\newcommand{\HB}{\hfill\break}
\newcommand{\acapo}{\linebreak}
\newcommand{\HALF}{{\textstyle{1\over 2}}}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%                 A CAPO.....
%%%
%%%
%%%
\newcommand{\giu}{{\medskip\noindent}}
\newcommand{\Giu}{{\bigskip\noindent}}
\newcommand{\nl}{{\smallskip\noindent}}
\newcommand{\nin}{{\noindent}}
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
\newcommand{\qed}{\hskip.5truecm
\vrule width 1.7truemm height 3.5truemm depth 0.truemm}
\newcommand{\qedd}{\vrule width 1.7truemm height 1.7truemm depth 0.truemm}
\newcommand{\QED}{\hfill\smallskip
         \line{\hfill\vrule height 1.8ex width 2ex depth +.2ex
               \ \ \ \ \ \ }
         \bigskip}
\newcommand{\PROOF}{\medskip\noindent{\bf Proof\ } \allowbreak}
\newcommand{\REMARK}{\medskip\noindent{\bf Remark\ }}
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%
%%%                         MATH MODE DEFINITIONS:
%%%
%
%
\newcommand{\io}{ {\infty}}
\newcommand{\ci}{ {C^\infty}   }
\newcommand{\ig}{ {\int}   }
\newcommand{\dpr}{ {\partial}   }
%
\newcommand{\torus}{ {\Bbb T}   }
\renewcommand{\natural}{ {\Bbb N}   }
\newcommand{\real}{ {\Bbb R}   }
\newcommand{\integer}{ {\Bbb Z}   }
\newcommand{\complex}{ {\Bbb C}   }
\newcommand{\tN}{ {\torus^N }   }
\newcommand{\rN}{ {\real^N}   }
\newcommand{\cN}{ {\complex^N }   }
\newcommand{\zN}{ {\integer^N }   }
\newcommand{\tM}{ {\torus^M }   }
\newcommand{\rM}{ {\real^M}   }
\newcommand{\cM}{ {\complex^M }   }
\newcommand{\zM}{ {\integer^M }   }
%
\renewcommand{\a }{ {\alpha}   }
\renewcommand{\b}{ {\beta}   }
\newcommand{\g}{ {\gamma}   }
\newcommand{\G}{ {\Gamma}   }
\renewcommand{\d}{ {\delta}   }
\newcommand{\D}{ {\Delta}   }
\newcommand{\e }{ {\varepsilon}   }
\newcommand{\et}{ {\eta}   }
\newcommand{\th }{ {\theta}   }
\newcommand{\vth }{ {\vartheta}   }
\newcommand{\k}{ {\kappa}   }
\renewcommand{\l}{ {\lambda}   }
\renewcommand{\L}{ {\Lambda}   }
\newcommand{\m}{ {\mu}   }
\newcommand{\n}{ {\nu}   }
\newcommand{\x }{ {\csi}   }
\newcommand{\X }{ {\Csi}   }
\newcommand{\p}{ {\pi}   }
\renewcommand{\P}{ {\Pi}   }
\newcommand{\r}{ {\rho}   }
\newcommand{\s}{ {\sigma}   }
\newcommand{\Si}{ {\Sigma}   }
\renewcommand{\t}{ {\tau}   }
\newcommand{\f}{ {\varphi}   }
\newcommand{\ph}{ {\phi}   }
\newcommand{\F}{ {\Phi}   }
\renewcommand{\c}{ {\chi}   }
\newcommand{\ps}{ {\psi}   }
\renewcommand{\o}{ {\omega}   }
\renewcommand{\O}{ {\Omega}   }
%
\renewcommand{\Im}{{\rm Im}}
\renewcommand{\Re}{{\rm Re}}
\renewcommand{\=}{ {\ \equiv\ }   }
%
\newcommand{\implies}{\ \Longrightarrow\ }
\newcommand{\sse}{\ \Longleftrightarrow\ }
%%%
\newcommand{\ppk}{ { (k)}}
\newcommand{\ppo}{ { (0)}}
\newcommand{\Xk}{ X^{(k)} }
\newcommand{\Xj}{ X^{(j)} }
\newcommand{\Xh}{ X^{(h)} }
\newcommand{\Tr}{ {T_r} }
%
\newcommand{\cA}{ {\cal A} }
\newcommand{\cT}{ {\cal T} }
\newcommand{\cR}{ {\cal R} }
\newcommand{\cH}{ {\cal H} }
\newcommand{\cC}{ {\cal C} }
\newcommand{\cF}{ {\cal F} }
\newcommand{\calN}{ {\cal N} }
%
\newcommand{\noin}{ {\not\in} }
\renewcommand{\lg}{ {\langle} }
\newcommand{\rg}{ {\rangle} }
\newcommand{\getr}{{{\ge}_{{\phantom{.}}_{\!\Tr}}}}
\newcommand{\baj}{ {\bar\jmath}}
\newcommand{\bai}{ {\bar\imath}}
\newcommand\bks{\backslash}
\newcommand{\zNn}{ {\zN\backslash \{0\} } }
%
\newcommand\bcR{ {\overline {\cal R}} }
\newcommand\bR{ {\overline  R} }
\newcommand\bT{{ \overline  T}}
\newcommand\bu{{ \bar u}}
\newcommand\bv{{ \bar v}}
\newcommand\bw{{ \bar w}}
\newcommand\bk{{ \bar \k}}
\newcommand\bl{{ \bar \l}}
\newcommand\bs{{ \bar \s}}
\newcommand\tT{{ \widetilde T}}
\newcommand\tcT{{ \widetilde \cT}}
\newcommand\hT{{ \widehat T}}
\newcommand\hcT{{ \widehat \cT}}
\newcommand\agl[1]{ |\lg{#1}\rg|  }
\newcommand\gl[1]{ \lg{#1}\rg  }
%********************************************
\newcommand\sumsud[2]{\sum_{ \scriptstyle {#1} \atop \scriptstyle  {#2} }}
\newcommand\sumsut[3]{\sum_{ {\scriptstyle {#1} \atop \scriptstyle  {#2}}
\atop \scriptstyle {#3}  }}
\newcommand\sumsuq[4]{\sum_{ 
{ {\scriptstyle {#1} \atop \scriptstyle  {#2}}
\atop \scriptstyle {#3} } 
\atop \scriptstyle {#4}
 }}
%
\newcommand\littlesum{ {\scriptscriptstyle \sum }}
\newcommand\wt{\widetilde}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
\setcounter{page}{1}
\renewcommand{\thepage}{\roman{page}}
\title{Compensations in small divisor problems}

\author{ L. Chierchia and C. Falcolini\\ 
{\footnotesize Dipartimento di Matematica} \\
{\footnotesize Universit\`a di Roma  ``Tor Vergata"}\\
{\footnotesize via della Ricerca Scientifica,
00133 Roma (Italy)} \\ 
{\footnotesize (Internet: chierchia@mat.utovrm.it and 
falcolini@mat.utovrm.it)}
\thanks{The authors gratefully acknowledge helpful discussions
with C.~Liverani.}
}

\maketitle

\date{August 1994}

\begin{abstract}

\nin
{\footnotesize
Several small divisor problems arising in the perturbative theory
of Hamiltonian and Lagrangian systems are considered. A general
method that allows to prove compensations among the elementary 
contributions of the formal power series expansions associated to
invariant surfaces is presented.
}
\end{abstract}

{\footnotesize  \tableofcontents}
%********************************************

\newpage
\setcounter{page}{1}
\renewcommand{\thepage}{\arabic{page}}

\section{Introduction}
\label{sec:par1}
\setcounter{equation}{0}
%
C.L.~Siegel \cite{S} was the first to solve a
small divisor problem (linearization of germs of analytic
functions). His solution was based on a ``direct method" consisting in
computing the $k^{\rm th}$ coefficient of the formal power series
solution and estimating such coefficient by a constant to the 
$k^{\rm th}$ power. The solution of small divisor problems arising
in perturbative series
in the theory of Hamiltonian systems has been 
one of the major contributions to dynamical 
systems of this century: 
the techniques employed for such a task, started by A.N.~Kolmogorov in the
early fifties and developed in the early sixties by V.I.~Arnold and 
J.K.~Moser, are known as
``Kolmogorov--Arnold--Moser (KAM) theory" (see \eg \cite{A1}
and references therein). Such methods show {\sl
indirectly} the convergence of the formal expansions of the perturbative 
series 
(longly before considered by
astronomers and especially by H.~Poincar\`e \cite{Po}). 

In 1988 H.~Eliasson
proposed a direct proof of the convergence of formal expansions
in the Hamiltonian context (\cite{E1}, \cite{E2}, \cite{E3}).
More recently direct proofs (in the Hamiltonian context)
have been reconsidered in several
papers: \cite{CF1}, \cite{Ga1}, \cite{Ga2}, \cite{Ga3} 
\cite{GG}, \cite{Ge1}, \cite{Ge2}; a small divisor problem
arising in elliptic systems
of PDE's has been solved by similar methods in \cite{CF2}. 

\giu
In this paper 
we consider several small divisor problems and discuss a general
method that allows to prove compensations among 
``elementary" contributions to the
$k^{\rm th}$ order of the associated formal expansion.
Such compensations,
modulo technical estimates (which have been thoroughly discussed in
\cite{CF1}), yield new proofs of the 
{\sl convergence} of the formal power series.

\giu
The starting point in this subject is a formal power series whose
coefficients are {\sl bona fide} functions determined by linear recursive
equations. ``Recursive" means that the $k^{\rm th}$ coefficient can
be expressed (typically by Taylor's expansions and by inversion of
suitable differential operator $D$)
in terms of the derivatives of a given function 
(``the Hamiltonian" or ``the Lagrangian") 
and in terms of the $h^{\rm th}$ coefficients with $h<k$.
If one keeps on expressing each coefficient in terms of the preceding ones,
one will end up with a finite (but huge) decomposition 
of the $k^{\rm th}$ coefficient involving only
the derivatives of the given function or, more precisely, involving
simple operations, coming from inverting $D$, on such derivatives;
the inversion of $D$ brings in the small divisors. We shall 
call the terms obtained through this procedure 
``elementary contributions". In writing out the elementary contributions
in the most explicit way, it appears readily clear that a graphical
representation is of big help. On a more formal level this is done 
using graph theory (more specifically the theory of labeled,
rooted trees).
Estimating the $k^{\rm th}$ coefficient by inserting absolute values
within the sum of the elementary contributions leads, 
in general (and inevitably), to divergences
(\ie to estimates of the $k^{\rm th}$ coefficients by a power of 
the factorial $k!$). 
But if the formal series were convergent (and in many examples one {\sl 
knows} by KAM that this is the case) it would mean that it is possible to
group together the elementary contributions in such a way 
that  huge (size $k!$) terms {\sl  compensate}
among themselves leaving out well behaving (exponential in $k$) quantities.
Graph theory provides a very useful tool in order to recognizes  the
families of elementary contributions that compensate. Most of the
known small divisor problems and some new ones\footnote{e.g. the elliptic 
systems considered in \cite{CF2}.} can be solved by this approach.

\giu
The small divisor problems we shall consider in this paper
(always in the real--analytic
category) are: P1) maximal 
quasi--periodic solutions (or ``invariant tori") for nearly--integrable
non--degenerate Hamiltonian systems; 
P2) maximal quasi--periodic solutions for nearly--inte\-gra\-ble
non--degenerate Lagrangians; P3) lower dimensional hyperbolic 
(whi\-ske\-r\-ed)
tori for ``second order" nearly--integrable Hamiltonian systems.
We shall also consider a small divisor problem, P4),
(lower dimensional ``indefinite" tori
for second order Hamiltonian systems) which is ``non compensable"
(in the precise sense described below),
strongly supporting that the associated formal power series are divergent;
in fact, it would be nice to give a direct proof of the divergence
of such series.

\giu
Our approach to P1) (integrated by the estimates of \cite{CF1})
yields a  direct proof of the classical ``KAM theorem" alternative to
that of Eliasson \cite{E1}; the direct proofs in the context of P2) 
and P3) are new\footnote{
The KAM theory in the context of P2) is relatively recent and is
sometimes referred to as ``KAM in configuration space" (see 
\cite{SZ}, \cite{CC}). For the KAM theory in the context of P3) see 
\cite{Gr}. The small divisor problem P3) is particularly
relevant in connection with the so--called Arnold diffusion
(see \cite{A2}, \cite{CG}, \cite{X}).}.

\giu
In \S~\ref{sec:par2} we give a precise description of the models 
P1) $\div$ P4) and of 
the results obtained. In \S~\ref{sec:par3} a new family of trees 
(simply related to 
the standard labeled, rooted trees) is introduced. In 
\S~\ref{sec:par4}, \S~\ref{sec:par5}, \S~\ref{sec:par6} the
compensation mechanism for the models P1), P2), P3) is discussed.
Finally, in \S~\ref{sec:par7} we exhibit an example where no 
compensations (of the type described in this paper) take place.
\section{Notations and Results}
\label{sec:par2}
\setcounter{equation}{0}
%
Let us describe more formally
what we mean by compensations in small divisor problems.
Given a ``nearly--integrable" dynamical system (say, as in P1) $\div$ P3)),
one is interested in finding invariant surfaces run by a linear flow
or, equivalently, in finding quasi--periodic solutions.
Quasi--periodic solutions are described in terms of functions on 
the standard $N$--dimensional torus  $\torus^N\=\real^N/(2\p \integer^N)$
substituting the phase variable $\th\in\tN$ with $\o t=(\o_1 t, ...,\o_N t)$
where $\o\in \rN$ is a given (rationally independent) vector
and $t$ denotes time\footnote{In the PDE case $t$ is a multidimensional
independent variable and $\o$ is a matrix; see \cite{CF2}.}.
The system being ``nearly--integrable" means that there is a perturbative
parameter, say $\e$, such that for $\e=0$ the system is completely
``solvable". It is therefore natural to try 
to establish the existence of formal solutions: this constitutes
the {\sl first step}.
Formal quasi--periodic solutions have been considered already in the last 
century; for a modern discussion of the formal solutions of the models
discussed here we refer the reader to Appendix B of \cite{CF1}\footnote{
In \cite{CF1} the formal expansion for P1) is proved; formal solutions for 
P2) $\div$ P4) can be proved in a completely similar way.}.

\noindent 
The formal solutions we will be dealing with have the form
\beq{1.1}
Z\=(Z_1,...,Z_d)\sim \sum_{k\ge 0} Z^k(\th) \e^k\ ,\qquad \th\in 
\torus^N
\eeq
where,
for each $k\ge 1$, the vector--valued function
$Z^k$ is a
real--analytic function over $\tN$  having an exponentially 
fast converging Fourier expansion of the form\footnote{Note that
we are using the suffix $k$ with different meaning as $\e^k$ denotes
the $k^{\rm th}$ power of the number $\e$ while $Z^k$ is a vector valued
function and $k$ is used as an index. 
The $h^{\rm th}$ power of the
$j^{\rm th}$ component of $Z^k$ will be denoted  $(Z^k_j)^h$;
the $j^{\rm th}$ 
component of the Fourier coefficients of $Z^k$ will be denoted
$Z^k_{nj}$.}
\beq{1.2}
Z^k= \sum_{n\in \zN} Z^k_n e^{i n\cdot \theta}\ .
\eeq
while for $k = 0$ $Z^0$ is the solution of the unperturbed problem.

\noindent
The {\sl second step} is to give $Z$ a representation
in terms of {\sl trees} \ie a representation of the form:
\beq{1.tr}
Z^k_{n}=\frac{1}{k!}\sum_{T\in \cT^k_*} 
\sum_{ {\a:V\to \zN}\atop{\Sigma \a_v=n}}
\sum_{\b:V\to B} \L(T,\a,\b) \ \ \prod_{v\in V} \g_v(T,\a,\b)\ ,
\eeq
where, intuitively speaking, $\cT^k_*$ is a suitable family of trees of 
``order $k$" taking care of the combinatorics coming out of Taylor's 
formula and its repeated applications (see \S~\ref{sec:par3}); 
the sum over the indices 
$\a$ attached to each vertex comes from expanding everything in terms
of Fourier series (hence the constrain of the total sum of such indices
to be $n$ since on the left hand side we have the $n$--Fourier coefficient
of the solution); the final sum  is finite and 
encodes all possible (case--by--case depending) indices which may help in 
writing out ``in the most explicit way" the solution\footnote{\eg
$\b_v$ will typically be related to the degree of the small divisors, see 
\equ{1.ga} below.}; 
the addenda of these sums are split into two terms:
$\L(T,\a,\b)$, which 
are complex vectors related to the derivatives of the given function
(the Hamiltonian or the Lagrangian) ruling the evolution of the dynamical 
system and the products of $\g_v$'s which are the small divisors.
More formally:
$\cT^k_*$ is a suitable family of labeled rooted trees\footnote{See
any introductory book on graph theory, such as \cite{Bo},
for the standard terminology or Appendix A in \cite{CF1}.}, $V\=V(T)$ 
denotes the 
set of vertices of $T$ and the suffix 
(``order") $k$ refers to the following estimates on cardinalities
\beq{1.ft}
\#\ \cT^k_* \le k!\ c^k_1 \ ,\qquad \# \ V(T) \le c_2\  k\ , \qquad
(\forall\ 
T\in\cT^k_*)\ ,
\eeq
(for a suitable constant $c_i>0$);
the second sum runs over all possible 
functions assigning to each vertex 
$v\in V$ of a rooted tree $T$ an integer vector $\a_v\in\zN$ 
with the constrain
$\sum_{v\in V} \a_v=n$; 
the third sum runs over a suitable set of
(possibly) vector--valued indices taking a finite number of
values ($\# B<\io$);
$\L$ is a complex
vector depending on $T$, $\{\a_v\}_{v\in V}$,  
$\{\b_v\}_{v\in V}$ and on the Hamiltonian (or Lagrangian);
finally $\g_v\in\real$ are {\sl divisors} that are
described as follows. Rooted trees can be naturally equipped with 
a partial order: 
we say that $v'\le v$ if the path joining the root $r$ of $T$
with $v'$ contains $v$; (obviously $v'< v$ means $v'\le v$ and $v'\neq v$
and $r\ge v$ $\forall$ $v\in V$ \ie the root $r$ is the first vertex of 
the rooted tree $T$).
Given a function $\a$, we define 
\beq{1.dl}
\d_v\=\d_v(T,\a)\=\sum_{v'\in V:\ v'\le v} \a_v\ .
\eeq
The divisors $\g_v$, which may assume arbitrarily small values,
are defined in terms of a real--valued function
$\lg \cdot \rg$ which satisfies the {\sl Diophantine condition}
\beq{1.dp}
|\lg n \rg|^{-1} \le \g |n|^\t\ ,\qquad \forall\ n\in \zN\backslash\{0\}
\eeq
with suitable positive constants $\g,\t$.\footnote{ 
Typically, in dynamical systems, 
$\lg n \rg = \o \cdot n$ (where the dot denotes the standard 
inner product in $\rN$) but in other situations (\eg \cite{CF2}) 
the function $\lg \cdot \rg$ might be more complicate 
(\eg non linear); in the case $\lg n \rg = \o \cdot n$
it is well known that, if $\t>N-1$ , up to a set of Lebesgue measure zero,
all $\o\in\rN$ satisfy \equ{1.dp} for some $\g$.}
Then
\beq{1.ga}
\g_v\=\g_v(T,\a,\b)\=\cases{ \lg \d_v\rg^{-\s_v} & if $\d_v\neq 0$\cr
1 & if $\d_v=0$.\cr}
\eeq
where $\s_v$ is, say, the first component of the index $\b_v$ and
takes value $0$, $1$ or $2$.
In \equ{1.tr} only the second sum runs over an infinite set of indices:
therefore we assume that there exist positive numbers $\xi,\xi',a>0$
such that, if we set
\beq{1.co}
a^k_{n}\= 
\frac{1}{k!}\sum_{T\in \cT^k_*} 
\sum_{ {\a:V\to \zN}\atop{\Sigma \a_v=n}}
\sum_{\b:V\to B} |\L(T,\a,\b)| \prod_{v\in V} e^{\xi'|\a_v|} \ ,
\eeq
then, for all $k$ and $n$ one has
\beq{1.co'}
a^k_{n}\le a^k e^{-\xi|n|}\ .
\eeq
In the models considered here, such an assumption is an immediate 
consequence of the 
well--posedness of the (formal) problem and of the analyticity assumptions
on the Hamiltonian (Lagrangian).
>From \equ{1.co} it follows at once that if one could bound the product of the 
divisors as\footnote{From now on we will adhere to the
the common abuse of notation
$v\in T$ in place of the more proper $v\in V(T)$ and also if $vv'=v'v$
denotes an edge of $T$, we shall denote $vv'\in T$ rather than the 
more proper $vv'\in E(T)$.}
\beq{1.si}
\prod_{v\in T} |\g_v|\le c^k_3 \ \prod_{T} (1+|\a_v|^b)
\eeq
for some $c_3 > 1$ and $b > 0$,
then from \equ{1.ft}, \equ{1.co}
and \equ{1.si} it would follow
\beq{1.5}
|Z^k_{n}| \le \frac{1}{k!}\sum_{T\in \cT^k_*} 
\sum_{ {\a:V\to \zN}\atop{\Sigma \a_v=n}}
\sum_{\b:V\to B} |\L(T,\a,\b)| \ c^k_4 \ \prod_{v\in V} (1+|\a_v|^b)
\le c^k_5 e^{-\xi |n|} ,
\eeq
(for suitable $c_1 > 0$) leading to ``absolute" convergence
(\ie convergence without compensations) of the formal expansion $Z$.
Indeed Siegel's original proof is based on a similar argument,
even though the set up is slightly different (and simpler). Technically
Siegel's problem corresponds to $\th$ varying in a small (complex) ball
so that Fourier series is replaced by Taylor series and 
$\a_v$ ranges over $\integer^N_+$: in such a case
$\d_v\neq \d_{v'}$ whenever $v>v'$ and Siegel's method \cite{S}
yields the estimates \equ{1.si}\footnote{For a detailed
discussion, in the present language, of Siegel's methods
see Appendix C of \cite{CF1}; for a different approach 
see \cite{Br}.}.
The problem with $\a_v\in \zN$ is that
one can have ``resonances" \ie $\d_v=\d_{v'}$ for $v>v'$ which may lead
to obstinate  repetitions of particularly small divisors. It is well
known (see \eg \cite{CF1}, Appendix B) that, in general, one has,
for arbitrarily large $k$,
subfamilies $\cF_{\rm div}\subset \cT^k_*$ and
a choice of $\bar \a$ and $\bar \b$ (depending only on the subfamily)
such that, for suitable $\bar a,\bar b>0$, 
\beq{1.di}
\frac{1}{k!}\sum_{T\in \cF_{\rm div}} 
\L(T,\bar \a,\bar \b) \ \ \prod_{v\in V} \g_v 
\ge \bar a^k k!^{\bar b}\ .
\eeq
Such families are obtained by taking 
{\sl chains of resonances} which are defined as follows.
Given  $T\in \cT^k_*$ and  $\a$ (\ie $\{\a_v\}_{v\in T}$),
a {\sl resonance} is 
%
a subtree\footnote{When referred to trees the notation $T'\subset 
T$ will  always mean ``$T'$ (unrooted) subtree  of $T$". Other special
conventions we are adopting are the following: a) for rooted trees the root
(usually denoted $r$) may be identified by adding an extra edge $\eta r$
where $\eta$ is a symbol (not a vertex of the tree) sometimes called
the ``earth"; with these positions one has, for trees, $\# E= \#V-1$
and, for rooted trees, $\# E= \#V$; b) the degree of a vertex $v$ is the 
number of edges $vv'$ incident with $v$ and if $v=r$ is the root, the edge
$\eta r$ is included in the count;
c) the degree of a subtree 
$T'\subset T$ is the number of edges connecting $T'$ with $T\backslash 
T'$;  if $T$ is rooted and the root $r$ belongs to $T'$ the edge $\eta r$
must be included in the count.}
%
$R\subset T$ such that: i) $R$ is of degree two 
(\ie $R$ is connected to $T\backslash R$
by two edges); ii)  if $u$ is the first vertex\footnote{Recall that 
$T$ is a rooted tree and hence partially ordered
(the order being such that 
the first vertex of $T$ is always its root).}
in $R$ and $z$ is the first vertex following $R$, then 
$\d_u=\d_z\neq 0$; iii)
$R$ cannot be disconnected, by removal of one edge, into two subtrees 
of degree two satisfying i) and ii). 
It will be important to consider different choices of the index $\b$
(\ie $\{\b_v\}_{v\in T}$): in particular given a resonance $R$ 
and given $\b$ we call {\sl order of the resonance} the number (see 
\equ{1.ga})
$\s\=\s_u$ ($u$ being the first vertex of $R$).
A {\sl chain of resonances} is a maximal series of resonances 
$R_1,...,R_h$ with $R_i$ 
adjacent\footnote{\ie  connected by one edge.}
to $R_{i+1}$; given a choice of $\b$, 
the {\sl order of the chain} is defined
to be $\bar \s\=\s_1+\cdots +\s_h$ where $\s_i\=\s_{u_i}$, $u_i$ being the
first vertex of $R_i$.
>From these positions it follows that
if $C$ is a chain of
order $\bar \s$, if $n=\d_z$ where $z$ is the first 
vertex following the chain (\ie following the last resonance, which by 
convention will be $R_h$), then
\beq{1.ch}
\prod_{v\in C} \g_v = \lg n \rg^{-\bar \s} \prod_{ {v\in C}\atop {v\neq u_i}}
\g_v
\eeq
(where $v\in C$ means $v\in \bigcup_i V(R_i)$).
The examples for which \equ{1.di} holds are based on chains 
with $\bar \s\sim k$ and $|\lg n \rg|\sim k^{-1}$. 
This phenomenon may be cured through compensations. 
To be more precise we introduce the notion of ``compensable chain".
Consider a chain $C=(R_1,...,R_h)$, ($h\ge 1$) and fix the indices $\b_v$.
Let, as above,
$u_i$ be the first vertex in $R_i$,
let $R_i$ be connected to $R_{i+1}$ by the edge $w_i u_{i+1}$ 
with $w_i\in R_i$ and $u_1\ge w_1>u_2\ge ...\ge w_h$; let 
$z$ be the first vertex following the chain (\ie following $w_h$)
and $n\=\d_z$; and, finally,
let $P_i$ be the path joining $u_i$ with $w_i$. 
Consider the following function of $x\in \complex$
\beq{1.pi}
\pi_C(x;T,\a,\b)\=\prod_{i=1}^h \pi_{R_i}(x;T,\a,\b)\ ,
\qquad
\pi_{R_i}(x)
\=\prod_{v\in  (R_i\backslash P_i)} \g_v\ 
\prod_{{v\in P_i}\atop{v\neq u_i}} \bar \g_v(x)\ ,
\eeq
where, if $u_i=w_i$ (\ie $P_i=\{u_i\}$), the product over $P_i$ is absent,
while if $v\in P_i\neq \{u_i\}$ we set
\beq{1.pi'}
\bar \g_v(x)\= \lg \sum_{ {v'\in R_i}\atop {v'\le v}} \a_{v'}
+ x n \rg^{-\s_v} \ .
\eeq
Thus, $\g_v=\bar \g_v (1)$ and
\beq{1.10}
\prod_{v\in C} \g_v= \lg n \rg^{-\bar \s} 
\pi_C(1)\ .
\eeq
We say that {\sl a chain $C\subset T\in \cT^k_*$ 
{\rm (\ie $C=(R_1,...,R_h)$ with $R_i\subset T$)}
is compensable} 
if there exists a family of trees $\cF_C\subset \cT^k_*$ whose elements
$T'$ have $C$ as common chain of resonance, and, for each $T'$,
there exists a choice of indices $\b'\=\b'(T')$, 
such that the function\footnote{
In other words, the elements $T'$ of $\cF_C$ are obtained from $T$ and $C$
by (possibly) changing the edges connecting $T\backslash C$ with $R_1$,
$R_1$ with $R_2$,...,$R_h$ with $T\backslash C$, and by (possibly) 
changing the values of the indices $\b$ on $C$.}
\beq{1.11}
\bar \pi_C(x)\=\sum_{T'\in \cF_C} \L(T',\a,\b')\  \pi_C(x;T',\a,\b')
\eeq
has a zero in $x$ of order at least $\bar \s-1$.
>From the detailed estimates in \cite{CF1}, it follows easily
that {\sl if in a small divisor problem all chains of
resonances are compensable then
the associated formal series \equ{1.1} is in fact convergent}.
In the rest of the paper we shall prove the following statements.

\giu
{\bf P1)} Let $H=H_0(y)+\e H_1(x,y)$ be a real--analytic Hamiltonian 
with $y\in B({y_0})\subset \rN$, $x\in \tN$
and $(x,y)$ standard symplectic coordinates\footnote{This simply means 
that the Hamiltonian (evolution) equations are given by
$\dot x= \dpr_y H$, $\dot y=- \dpr_x H$.}; here $B({y_0})$
denotes some  $N$--sphere centered at $y_0$. 
Assume the following standard 
``non degeneracy conditions" on the ``integrable part" $H_0$: 
i) the Hessian matrix $\dpr^2_y H_0(y_0)$ is invertible;
ii) $\o\=\dpr_y H_0\in\rN$
is a Diophantine vector \ie there exists $\g,\t>0$
such that\footnote{Compare with \equ{1.dp}.}
\beq{1.om}
|\o \cdot n|^{-1} \le \g |n|^\t\ ,
\qquad \forall\  n\in \zN\backslash \{0\}\ .
\eeq
Then, quasi--periodic solutions with frequencies $\o$, \ie solutions
of the form $(x(t),y(t))=Z(\o t)$ with $Z^0(\th)=(\th,y_0)$ and
$Z^k:\th\in \tN \to Z^k(\th)\in \real^{2N}$, satisfy the equations
\beq{1.h1}
D Z = J \dpr H\Big( Z(\th) \Big)\ ,
\eeq
where $D\=\o\cdot \dpr_\th$, $J$ is the standard symplectic matrix
$\pmatrix{0 & I\cr -I & 0\cr}$ and $\dpr$ is the gradient 
$\dpr_{(x,y)}$ with respect to the variables $(x,y)$.
Then it is well known\footnote{See  \cite{Po} or Appendix B of
\cite{CF1}.} that there exists a unique formal solutions 
$Z\sim \sum_{k\ge 0} \e^k Z^k$ of
\equ{1.h1} with the normalization condition 
\beq{1.av}
\ig_{\tN} \p_1\circ Z^k d\th= 0\ \quad (k\ge 1) ,
\eeq
where $\p_1$ is the projection onto the first coordinates:
$\p_1(x,y)=x$.
We can then prove
\thm{thm1} There exists a tree expansion \equ{1.tr} for $Z$ such
that all chains of resonances are compensable.
\ethm
%
{\bf P2)} The formulation for the Lagrangian case is very similar. We
let $L\=L_0(y)+\e L_1(x,y)$ be a real--analytic Lagrangian\footnote{The
Lagrangian (evolution equations) are $\frac{d}{dt} \dpr_y L(x,\dot x)=
\dpr_x L(x,\dot x)$.}
with $L_i$
satisfying the same hypotheses assumed above for $H_i$ and $\o$, except 
that now $\o \= y_0$. 
Quasi--periodic solutions $x(t)=Z(\o t)=\sum_{k\ge 0} \e^k Z^k(\o t)$, 
where now $Z^0(\th)=\th$ and
$Z^k:\th\in \tN\to Z^k(\th)\in \rN$,  satisfy the equations
\beq{1.h2}
D \ \dpr_y L\big(Z(\th), DZ(\th)\big)=
\dpr_x L\big(Z(\th), DZ(\th)\big)
\ ,\qquad 
(D\=\o\cdot\dpr_\th)\ .
\eeq
Mimicking the proof of Appendix B in \cite{CF1} one can show that
there exists a unique formal solution 
$Z\sim \sum_{k\ge 0} \e^k Z^k$ of \equ{1.h2} with the normalization
condition as in \equ{1.av} but without the projection $\p_1$.
Then, {\sl Theorem~\ref{thm1} holds also in this case}.

\giu
{\bf P3)} 
Maximal invariant tori correspond to analytic continuation (in $\e$)
of unperturbed tori having ``all frequencies excited" \ie tori run by 
linear flow $t\to \o t$ with $\o\=\dpr_y H_0(y_0)$ rationally independent
over $\zN$ and $N=\#$ of degrees of freedom. More delicate is what happens 
to unperturbed 
``resonant" tori\footnote{Here, the word ``resonant" which is related
to the ``resonances" of celestial mechanics, is used in a technically 
different meaning from the one used throughout this paper.}
\ie invariant unperturbed tori for which there exist $n\in \zNn$
such that $\o\cdot n=0$. We shall argue (see next item P4))
that {\sl in general, 
such tori are not analytically continuable for $\e\neq 0$}. Nevertheless
we will show, under suitable conditions,
how to construct, with the (intrinsically analytic)
methods outlined above, lower dimensional invariant (unstable) tori 
when $\e\neq 0$.
For simplicity, 
we shall discuss only a special case 
of ``second order" nearly integrable Hamiltonian systems 
with Hamiltonian function
\beq{1.12}
H\= \frac{1}{2} y^2 + \frac{1}{2} p^2 + \e f(x,q)
\eeq
where $(x,y)$ are symplectic variables as above (\ie with $x\in \tN$)
and so are $(q,p)$ with $q\in \tM$; hence the number of degrees of freedom 
is $N+M$ and we shall study $N$--dimensional invariant tori.
The specification ``second order"
is due to the particular form in which can be cast
the Hamilton equations for $H$:
\beq{1.13}
\ddot x= -\e \dpr_x f\ ,\qquad \ddot q = -\e \dpr_q f\ .
\eeq
For $\e=0$, $N$--dimensional invariant tori
(up to a trivial linear and
symplectic change of coordinates\footnote{A transformation of 
(standard) symplectic coordinates $(q,p)$ (of a $2d$--dimensional phase 
space)  is called {\sl symplectic} if it preserves the (standard) two form 
$\sum_{i=1}^d dq_i\wedge dp_i$.})
are spanned by solutions of the form
$(y,p)=(\o,0)$, $(x,q)=(x_0+\o t, q_0)$.
%
>From classical transformation theory it follows that  (if 
$\o$ satisfies \equ{1.om}) the  
evolution equations of $H$ in \equ{1.12} are equivalent
to the evolution equations for a Hamiltonian of the form
$\frac{y^2}{2} +\frac{p^2}{2} +\e f_0(q) +o(\e)$ where $f_0$ is the average
(w.r.t. $x$) over $\tN$ of $f$. Motivated by this observation, we consider
the Hamiltonian
\beq{1.12'}
\frac{1}{2} y^2 + \frac{1}{2} p^2 + \e f_0(q) + 
\e^{2} \tilde f(x,q) 
\= H_0(q,y,p;\e) + H_1(x,q;\e) \ ,\qquad H_1\=\e^2 \tilde f\ .
\eeq
Even though $H_0$ is not, in general, integrable,
{\sl if $q_0$ is a critical point for $f_0$}, $H_0$ still admits
the invariant $N$--torus $\cT_0$ spanned by $(y,p)=(\o,0)$, $q=q_0$ and
$x= x_0+ \o t$ and we want to study the persistence of such torus for the 
full Hamiltonian. To attack the problem perturbatively, 
we introduce a new analyticity parameter $\m$
with respect to which we shall make formal (and eventually convergent
with radius of convergence greater than one)
power series expansion and  consider the Hamiltonian
$H_0+\m H_1$ (so that  for $\m=1$ we recover the Hamiltonian \equ{1.12'}).
We also
make the following {\sl hyperbolicity assumption}:
we assume that $q_0$ is such that the matrix
$\e \dpr^2_q f_0(q_0)$ is {\sl positive definite}. Under these hypotheses
it is easy to prove
that there exists a (unique) formal expansion
\beq{1.15}
Z\=(Z_1,...,Z_d)\sim \sum_{k\ge 0} Z^k(\th;\e) \m^k\ ,\qquad \th\in  
\torus^N\ ,\qquad d=2(N+M)
\eeq
such that $t\to Z(\o t)$ is a formal 
quasi--periodic solution of the Hamiltonian equations
governed by $H_0+\m H_1$ and such that the set 
$\{ Z^{0}(\th):\th\in \tN\}$ coincides with the unperturbed
torus $\cT_0$ introduced above. Uniqueness is 
achieved by requiring \equ{1.av} where $\p_1$ denotes
the projection onto the $x$ variable.
Then, {\sl Theorem~\ref{thm1} holds also in this case}.
>From this result, as already remarked, it follows that the formal
power series is actually convergent but, what is more interesting in this
case, one can show that for $\e\neq 0$ 
small enough, {\sl the radius of convergence
(in  $\m$) is greater than one} so that the set 
$\{ Z(\th):\th\in \tN, \m=1\}$ is an invariant $N$--torus for the 
Hamiltonian \equ{1.12'}. We finally mention that from \cite{Gr}
it follows easily that these tori are {\sl whiskered} in the sense of
\cite{A2}.

\giu
{\bf P4)} 
Consider again the Hamiltonian \equ{1.12}. Indeed, one can consider
formal power series in $\e$ and one can show
that {\sl if $q_0$ is 
a non degenerate critical point of the $x$--average of $f$
then there exists a (unique) formal power series 
\equ{1.1} (with $d=2(N+M)$) such that $t\to Z(\o t)$ 
($\o$ satisfying \equ{1.om}) is a formal quasi--periodic solution
for \equ{1.12} and the set $\{Z^{0}(\th):\th\in\zN\}$ coincide
with the torus spanned by $y=\o$, $p=0$, $q=q_0$, $x= x_0 +\o t$.}
Uniqueness, again, is enforced by the requirement \equ{1.av}.
Also for such a formal series  one can write down a tree expansion
completely analogous to those referred to in P1)$\div$P3) above.
However, we shall prove that 
there exist chains $C=(R_1,...,R_h)$ such that if
$\cF_0$ denotes the family of {\sl all} trees with chain $C$ 
then 
\beq{1.16}
\sum_{T'\in\cF_0} \L (T')\ \p_C(0)\neq 0\ .
\eeq
In view of this fact it seems natural 
to {\sl conjecture that in the present case the formal power series
$Z$ is divergent}. 
\section{Weighted trees}
\label{sec:par3}
\setcounter{equation}{0}
%
Here we describe the tree family $\cT^k_*$, which appears  in the
basic formula \equ{1.tr}.

\giu
For the models  P3) and P4) introduced in \S~\ref{sec:par2},
$\cT^k_*$ is simply the family of all labeled,
rooted trees with $k$ vertices\footnote{The basic terminology can
be found in any introductory book on graphs or in Appendix~A of 
\cite{CF1}.}, which we will denote by $\cT^k$. In this case,
as is well known (see \eg \cite{Bo}),
$\# \cT^k = k^{k-1}$  and \equ{1.ft} is clearly satisfied.

\giu
To treat the cases P1) and P2) one has to distinguish, in the Taylor's
expansion, the contributions coming from $H_0$ and $L_0$ from those coming
from $H_1$ and $L_1$. To do this 
we introduce the following family of ``weighted trees".
Given a rooted (unlabeled) tree $T$ we call a function of the vertices of 
$T$
\beq{3.1}
\chi : v\in V(T) \to \chi_v\in\{0,1\}
\eeq
a {\sl weight function}. A {\sl weighted rooted tree} is a couple 
$(T,\chi)$ with $T$ a rooted tree and $\chi$ a weight function. We now
denote $\widetilde \cT_k$ the set of weighted rooted trees satisfying
\beq{3.2}
(i) \quad \deg v\le 2 \ \implies \ \chi_v=1\ ,\qquad
(ii) \quad \sum_{v\in T} \chi_v = k\ .
\eeq
Notice that, in particular, all final vertices (\ie vertices
of degree 1) have weight 1
and that, for any $T\in \widetilde  \cT_k$,
$\# V(T) \le 2k-1$ as is easy to verify\footnote{
Let $V_i=\{v\in V:\chi_v=i\}$, and let $k_i=\#V_i$, so that 
$k=k_1$. It is well known (see, \eg, \cite{Bo} and recall our
convention on degree of the root) that 
$\sum\limits_{V_0} \deg v + \sum\limits_{V_1} \deg v= 2(k_0+k_1)-1$.
If $v\in V_0$ then $\deg v\ge 3$, thus 
the sum over $V_0$ can be bounded from below by $3 k_0$
while the sum over $V_1$ can be bounded from below by $k_1$. This leads
to $k_0\le k_1-1$ which is the claim. Such inequality is
optimal.}.
We now define the class $\cT_k$ of
{\sl labeled, weighted rooted trees} obtained from $\widetilde  \cT_k$
by labelling with $k$ different labels the $k$ vertices with weight 1.

\giu
In cases P1) and P2) we let $\cT^k_*=\cT_k$; it 
is easy to check that \equ{1.ft} holds also in this case\footnote{
Since (\cite{Bo}) 
$\# \widetilde \cT^h\le 4^h$ and $\# V(T)\le 2k-1$ for any
$T\in \widetilde \cT_k$ one sees that $\# \widetilde \cT_k\le 4^{2k}$.
Since the labels are attached to $k$ vertices, it
is $\# \cT_k\le k! 4^{2k}$.}.

\giu
The relation between Taylor's formula and trees may be based
on the following operation $*$ (that we now briefly discuss for the
case of $\cT_k$; for the case $\cT^k$ see 
Appendix~B of \cite{CF1}). 
If $T\in\cT_k$ we denote by $\widetilde  T$ the tree in $\widetilde  \cT_k$
obtained by removing the labels from $T$; we also denote
$T^0$ (or $\widetilde  T^0$) 
the unrooted tree obtained from $T$ (or $\widetilde  T$)
by removing the edge $\eta r$ \ie by not distinguishing any more the root
from the other vertices; finally if $T$ (or $\widetilde  T$) is an unrooted
labeled (or unlabeled) tree and $r$ is one of its vertices,
we denote by $T_r$ (or $\widetilde  T_r$) the rooted tree obtained by adding
the edge $\eta r$ (\ie by decreeing that $r$ is the root). 
Let $\chi\in \{0,1\}$ and $\ell\ge 1$,
let $h_i$ be $\ell$ 
positive integers such that $h_1+\cdots+h_\ell=k-\chi$ and 
pick trees $T_i\in\widetilde \cT_{h_i}$. 
We can form a tree $T\in \widetilde \cT_k$ with
root $r$ ($r$ being a vertex different from the vertices of $T_i$, 
$\forall$ $i$) of weight $\chi_r\=\chi$ by setting
\beq{3.4}
T\=T_{h_1}*\cdots *T_{h_\ell}\=
\Big(T_{h_1}^0\cup\cdots\cup T^0_{h_\ell} \cup \{r\}+
\sum_{i=1}^\ell rr_i\Big)_r
\eeq
where $r_i$ is the root of $T_i$ (and summing an edge $e$ to a tree $S$ 
means, obviously, to add $e$ to $E(S)$). 
Then, one has the following
\pro{pro:1}
Let $F$ be a complex valued function defined on trees in 
$\widetilde  \cT_k$
(for any $k$). Then
\beq{3.5}
\frac{1}{k!} \sum_{T\in \cT_k} F(\widetilde  T) =
\sum_{\chi=0,1} \sum_{\ell=2-\chi}^{k-\chi} \frac{1}{\ell!} 
\ \sum_{h_1+\cdots+h_\ell=k-\chi}\prod_{i=1}^\ell \frac{1}{h_i!}
\ \sum_{ T_i\in \cT_{h_i}} F(\widetilde  T_{h_1}*\cdots *
\widetilde  T_{h_\ell})\ .
\eeq
\epro
For the proof we refer to \cite{CF1} 
(Corollary~B.1 of Appendix~B)\footnote{In \cite{CF1}
it is treated the case of $\cT^k$
(the $*$ operation in $\widetilde  \cT^k$ 
is defined as above, replacing $\chi$ systematically by 1); 
adapting the proof to $\cT_k$ is a trivial exercise.} .

%
\begin{figure}[hbt]
\begin{center}
\begin{picture}(350,80)
\thicklines

\put(5,20){\begin{picture}(30,60)
\put(0,40){\circle{8}}
\multiput(0,40)(15,0){3}{\circle*{3}}
\multiput(-2,48)(15,0){3}{{\scriptsize $1$}}
\put(0,40){\line(1,0){30}}
\end{picture}}

\put(60,20){\begin{picture}(30,60)
\put(15,40){\circle{8}}
\multiput(0,40)(15,0){3}{\circle*{3}}
\multiput(-2,48)(15,0){3}{{\scriptsize $1$}}
\put(0,40){\line(1,0){30}}
\end{picture}}

\put(115,20){\begin{picture}(30,60)
\put(0,40){\circle{8}}
\multiput(0,40)(15,0){2}{\circle*{3}}
\put(0,40){\line(1,0){15}}
\put(-2,48){{\scriptsize $1$}}
\put(13,48){{\scriptsize $0$}}
\put(35,50){{\scriptsize $1$}}
\put(35,28){{\scriptsize $1$}}
\put(30,50){\circle*{3}}
\put(15,40){\line(3,2){15}}
\put(30,30){\circle*{3}}
\put(15,40){\line(3,-2){15}}
\end{picture}}

\put(170,20){\begin{picture}(30,60)
\put(15,40){\circle{8}}
\multiput(0,40)(15,0){2}{\circle*{3}}
\put(0,40){\line(1,0){15}}
\put(-2,48){{\scriptsize $1$}}
\put(13,48){{\scriptsize $0$}}
\put(35,50){{\scriptsize $1$}}
\put(35,28){{\scriptsize $1$}}
\put(30,50){\circle*{3}}
\put(15,40){\line(3,2){15}}
\put(30,30){\circle*{3}}
\put(15,40){\line(3,-2){15}}
\end{picture}}

\put(225,20){\begin{picture}(45,60)
\put(15,40){\circle{8}}
\multiput(0,40)(15,0){4}{\circle*{3}}
\put(-2,48){{\scriptsize $1$}}
\put(13,48){{\scriptsize $0$}}
\put(28,48){{\scriptsize $1$}}
\put(43,48){{\scriptsize $1$}}
\put(0,40){\line(1,0){45}}
\end{picture}}

\put(295,20){\begin{picture}(45,60)
\put(15,40){\circle{8}}
\multiput(0,40)(15,0){3}{\circle*{3}}
\put(0,40){\line(1,0){30}}
\put(-2,48){{\scriptsize $1$}}
\put(13,48){{\scriptsize $0$}}
\put(29,48){{\scriptsize $0$}}
\put(45,50){\circle*{3}}
\put(30,40){\line(3,2){15}}
\put(45,30){\circle*{3}}
\put(30,40){\line(3,-2){15}}
\put(50,50){{\scriptsize $1$}}
\put(50,28){{\scriptsize $1$}}
\end{picture}}

\parbox{12cm}{\caption{\label{fig:wt1}The elements of 
$\widetilde  \cT_3$. (The
numbers $0$, $1$ are the values of the index $\chi$ for the corresponding 
vertices and the encircled vertex is the root of the tree).}}
\end{picture}
\end{center}
\end{figure}
%
%
\begin{figure}[hbt]
\begin{center}
\begin{picture}(350,80)
\thicklines
\put(115,20){\begin{picture}(30,60)
\put(0,40){\circle{8}}
\multiput(0,40)(20,0){2}{\circle*{3}}
\multiput(-2,48)(20,0){2}{{\scriptsize $1$}}
\put(-4,25){$u_2$}
\put(16,25){$u_1$}
\put(0,40){\line(1,0){20}}
\end{picture}}

\put(170,20){\begin{picture}(30,60)
\put(0,40){\circle{8}}
\multiput(0,40)(20,0){2}{\circle*{3}}
\multiput(-2,48)(20,0){2}{{\scriptsize $1$}}
\put(-4,25){$u_1$}
\put(16,25){$u_2$}
\put(0,40){\line(1,0){20}}
\end{picture}}

\put(225,20){\begin{picture}(30,60)
\put(0,40){\circle*{3}}
\put(0,40){\circle{8}}
\put(-2,48){{\scriptsize $0$}}
\put(13,55){{\scriptsize $1$}}
\put(13,35){{\scriptsize $1$}}
\put(20,50){$u_1$}
\put(20,30){$u_2$}
\put(15,50){\circle*{3}}
\put(0,40){\line(3,2){15}}
\put(15,30){\circle*{3}}
\put(0,40){\line(3,-2){15}}
\end{picture}}

\parbox{12cm}{\caption{\label{fig:wt2}The elements of $\cT_2$
($u_1$ and $u_2$ are the labels of $\cT_2$).}}

\end{picture}
\end{center}
\end{figure}
%
\section{Compensations I (Maximal Hamiltonian tori)}
\label{sec:par4}
\setcounter{equation}{0}
\subsection{Tree expansion}
\label{subsec:4.1}
%
Consider the model introduced in P1) of \S~\ref{sec:par2}
and recall that there exists a unique formal solutions
$Z\sim\sum_{k\ge 0} \e^k Z^k$ satisfying 
\equ{1.h1} and \equ{1.av}. Denote by $Z^{(1)k}$ the $x$--component
(\ie the first $N$ components) of the vector $Z^k$ and by 
$Z^{(2)k}$ the $y$--component;
consistently, let $\dpr^{(1)}\=\dpr_x$
and $\dpr^{(2)}\=\dpr_y$. We also let  
\beqno
[\ \cdot\ ]_k\= \frac{1}{k!} \frac{d^k}{d\e^k}\ (\cdot)|_{\e=0}
\eeqno
denote the $k^{\rm th}$ order operator which to a (possibly 
formal) power
series $a\sim \sum a_k \e^k$ associates its $k^{\rm th}$ order
coefficient: $[ a ]_k\= a_k$. 
Finally let
\beq{4.0}
A \= \dpr^2_y H_0(y_0)\ .
\eeq
With these definitions, we can rewrite
\equ{1.h1} as 
\beq{4.1}
D Z^{(1)k} = A Z^{{(2)} k} + [\dpr^{(2)} H_0]_k^{(k-1)} + 
[\dpr^{(2)} H_1]_{k-1}^{(k-1)}\ ,
\qquad
D Z^{{(2)} k} = - [\dpr^{(1)} H_1]_{k-1}^{(k-1)}\ ,
\eeq
where the suffix $\phantom{[\cdot]}^{(k-1)}$ 
means that the argument of the
function within square brackets is, for $k\ge 2$, the polynomial
in $\e$ of degree $(k-1)$ given by
\beq{4.pol}
x=\th+\sum_{h=1}^{k-1} \e^h Z^{{(1)} k}\ ,
\qquad y= y_0+ \sum_{h=1}^{k-1} \e^h Z^{{(2)} k}\ ,
\eeq
and, for $k=1$, is $(x,y)=(\th,y_0)$.
Since the average of $[\dpr^{(1)} H_1]_{k-1}^{(k-1)}$ vanishes
(as is clear from the second of \equ{4.1})
we can apply to it the operator $D^{-1}$ obtained by inverting
the constant coefficient operator\footnote{If $f$ is a (smooth) function
on $\tN$ with vanishing mean value we denote by $D^{-1} f$
the unique solution with vanishing mean value of the equation for $g$:
$D g = f$. Expanding in Fourier series one has $g(\th)=D^{-1}f(\th)=
-i \sum\limits_{n\in\zNn} \frac{f_n}{\o\cdot n} \exp({i n\cdot \th})$, 
where
$i=\sqrt{-1}$: the inversion of $D$ introduces the small divisors.} 
$D$ and rewrite \equ{4.1} in a more compact way as
\beq{4.2}
D Z^{(\r)k}= (2-\r) A Z^{(2)k}+
\sum_{\chi=0,1} (-1)^{3-\r} [\dpr^{(3-\r)} H_\chi]^{(k-1)}_{k-\chi}\ ,
\qquad (\r=1,2)\ .
\eeq
Notice that while, by \equ{1.av}, the average of $Z^{(1)k}$ vanishes,
the average of $Z^{(2)k}$ can be read (for $\r=1$) from \equ{4.2}
by integrating over $\tN$:
\beq{4.3}
\ig_{\tN} Z^{(2)k}=- A^{-1} \sum_{\chi=0,1} \ig_{\tN} [\dpr^{(2)} 
H_\chi]^{(k-1)}_{k-\chi}\ .
\eeq
Taking the $n$--Fourier coefficient of \equ{4.2}, \equ{4.3}
one gets
\beq{4.4}
Z^{(\r)k}_n= \sumsud{\s\in \{0,1,2\}^*}{\chi\in\{0,1\} }
\lg n\rg^{-\s}\ 
\Big\{ [D^{(\s,\r)} H_\chi]_{k-\chi}^{(k-1)}\Big\}_n\ ,\qquad
\big(\lg n\rg\= \o\cdot n\big)\ ,
\eeq
where $D^{(\s,\r)}$ is the vector--valued 
operator\footnote{
For example the $j^{\rm th}$ component of $D^{(2,1)}$ is given by
$D_{j}^{(2,1)}=\sum\limits_{s=1}^N
\frac{\dpr^2 H_0}{\dpr y_j\dpr y_s}(y_0) \frac{\dpr}{\dpr x_s}$.}
\beq{4.5}
D^{(\s,\r)}\=(-1)^{1-\s\r} \ i^{-\s}\   A^{\s-1}\ 
\dpr^{(4-\s-\r)}\ ,
\eeq
and the $*$ attached to the range of $\s$ 
means that the
following {\sl constraints} have to be satisfied: 
$\s+\r\in \{2,3\}$ and $\s=0$ $\iff$ $n=0$
in which case we adopt the convention that $\lg 
n\rg^\s=0^0\=1$.
Notice that $D^{(\s,\r)} H_\chi=0$ if $\chi=0$ and $\s+\r=3$
(as in such a case $\dpr^{(4-\s-\r)}=\dpr_x$ and $H_0$ is independent of 
$x$).

\giu
We shall now use Taylor's formula in the following form.
If $f:x\in \real^m \to f(x)\in \real$ is a $C^\io$ function and
if  $a(\e)\sim\sum_{s\ge 1} \e^s a^{(s)}$ is a $\real^m$--valued
(possibly formal) power series, then
\beq{4.6}
f\big(a(\e)\big) \sim f(0) + \sum_{h\ge 1} \e^h
\sum_{\ell=1}^h \frac{1}{\ell!}
\sum_{{h_1+\cdots+h_\ell=h}\atop{1\le h_i\le \ell}}
\sum_{{j_1,...,j_\ell}\atop{1\le j_i\le m}} 
\frac{\dpr^\ell f}{\dpr x_{j_1}\cdots\dpr x_{j_\ell}}(0)\ 
a^{(h_1)}_{j_1}\cdots a^{(h_\ell)}_{j_\ell}\ .
\eeq
Thus, if $z\=(x,y)$ and $z^{(1)}\=x$, $z^{(2)}\=y$, by \equ{4.4}
and \equ{4.6} we get for the $j^{th}$ component of the vector 
$Z^{(\r)k}_n$, and for $k\ge 2$,
\beqa{4.7}
&& Z^{(\r)k}_{nj}= \\
&& \sum_{ {\s\in \{0,1,2\}^*}\atop{\chi\in \{0,1\}}}
\sum_{\ell=2-\chi}^{k-\chi} \frac{1}{\ell!} 
\sumsut{1\le h_i\le k-1}{\Sigma h_i=k-\chi}{(1\le i\le \ell)}
\sumsut{n_i\in\zN}{\Sigma n_i=n}{(0\le i\le \ell)}
\sum_{{\r_1,...,\r_\ell}\atop{\r_i\in \{1,2\}}}
\sum_{ {j_1,...,j_\ell}\atop{1\le j_i\le N}}
\lg n\rg^{-\s}
\Big\{ \frac{ \dpr^\ell\  D_{j}^{(\s,\r)} H_\chi}{\dpr z^{(\r_1)}_{j_1}
\cdots \dpr z^{(\r_\ell)}_{j_\ell}} \Big\}_{n_0}
\prod_{i=1}^\ell Z^{(\r_i)h_i}_{n_i j_i}\nonumber\ ,
\eeqa
where the derivatives of $H_\chi$ are evaluated at
$(\th,y_0)$ and then one takes the $n_0$--Fourier coefficient 
(with respect to $\th$); for $k=1$, since $[D^{(\s,\r)} H_0]_1^{(0)}=0$,
one has the simple formula
\beq{4.7'}
Z^{(\r)1}_{nj}=
\sum_{\s\in \{0,1,2\}^*} \lg n\rg^{-\s} \{D^{(\s,\r)}_j H_1\}_n\ .
\eeq
We are ready to prove the following {\sl tree expansion formula}
(recall \equ{1.ga}, \equ{1.dl})
\beq{4.8}
Z^{(\r_0)k}_{nj_0} = 
\frac{1}{k!} \sum_{T_r\in\cT_k} \quad 
\sum_{\a:V\to \a_v\in \zN \atop \Sigma_v \a_v = n}\quad
\sumsud{\b:V\to\b_v\in B}{\r_r=\r_0}
\quad 
\sumsud{j:V\to j_v\in\{1,...,N\}}{j_r=j_0}\ 
\prod_{v\in T_r} \big\{ \L_v(T_r,\b) H_{\chi_v} \big\}_{\a_v}\ 
\prod_{v\in T_r} \g_v \ ,
\eeq
where: the index set $B$, which
depends on the function $\d_v$,
is defined as 
\beqa{4.B}
& B & \=\Big\{\b=(\s,\r):\ \s\in\{0,1,2\};\ \r\in\{1,2\};\ 
{\rm s.t.}\nonumber\\ 
&& \qquad 
\s+\r\in\{2,3\}\ ,\s=0\ \iff \  \d_v= 0\Big\}\ ;
\eeqa
the scalar operator $\L_v(T_r,\b)$ is given by
\beq{4.9}
\L_v(T_r,\b)\=
D^{(\s_v,\r_v)}_{j_v} 
\prod_{v'\in \calN_v} \dpr_{j_{v'}}^{(\r_{v'})}\ ,
\qquad
\calN_v\=\{v'<v; v'\ {\rm adjacent\  to\ } v\}\ ,
\eeq
(if $\calN_v=\emptyset$, \ie $\deg v=1$, the product is omitted).
To recover \equ{1.tr} from \equ{4.8}, one simply defines
the  component $j_0$ of $\L(T_r,\a,\b)$ as
\beq{4.11'}
\L(T_r,\a,\b)\= 
\sumsud{j:V\to \{1,...,N\}}{j_r=j_0}\quad
\prod_{v\in T_r} \{\L_v(\b) H_{\chi_v}\}_{\a_v}\ .
\eeq
{\sl The proof of \equ{4.8}} 
is by induction. For $k=1$ \equ{4.8} is an immediate 
consequence of \equ{4.7'}. Assume that \equ{4.8} holds with $k$
replaced  by $1,...,k-1$. 
Given $h\ge 1$, $\r_0$, $j_0$, $n$, consider the function of (unlabeled)
trees $T_r\in\wt\cT_h$,  given by
\beq{4.10}
F^{(\r_0)h}_{nj_0}(T_r)
\= \sumsut{ \a:V\to\zN:\Sigma \a_v=n}{\b\in B}{j:V\to \{1,...,N\}:j_r=j_0}
\prod_{v\in T_r} \g_v\ 
\prod_{v\in T_r} \{\L_v(\b) H_{\chi_v}\}_{\a_v}
\eeq
and observe that if $T_r = T_1 *\cdots *T_\ell$ with 
$T_i\in \wt \cT_{h_i}$ and $h_1+\cdots+h_\ell=h$, then 
(fixing $\chi$, $n$, $\r_0$, $j_0$)
\beqa{4.11}
&&F^{(\r_0)h}_{nj_0}(T_1 *\cdots *T_\ell)\\
&&\qquad =
\sum_{\s\in \{0,1,2\}^*} \ 
\sum_{n_0+\cdots+n_\ell=n}\ 
\sumsud{\r_1,...,\r_\ell}{j_1,...,j_\ell}
\lg n_0\rg^{-\s}
\Big\{ \frac{ \dpr^\ell\  D_{j}^{(\s,\r)} H_\chi}{\dpr z^{(\r_1)}_{j_1}
\cdots \dpr z^{(\r_\ell)}_{j_\ell}} (\cdot,y_0)\Big\}_{n_0}
\prod_{i=1}^\ell F^{(\r_i)h_i}_{n_i j_i}(T_i)
\nonumber
\eeqa
Equation \equ{4.8} follows now from \equ{4.7}, \equ{4.11} and 
Proposition~\ref{pro:1}. 
%%%%%%%%%%%%%%
%
\subsection{Compensations}
\label{subsec:4.2}
Here, we show how to choose families of trees $\cF$ and corresponding
indices so that all chains of resonances are compensable 
(recall the definitions given in \S~\ref{sec:par2}).

\giu
Given $T=T_r$, $\a$ and $\b$, consider a resonance $R$ and let
$u$ denote its first vertex and $w'<u$ the first vertex following $R$.
By definition of resonance it is $\d_u=\d_{w'}\neq 0$
(\ie $\sum_{v\in R} \a_v=0$) and  $\d_v\neq \d_{w'}$
if $u>v>w'$.
%
We shall classify resonances by assigning to them an integer 
$s\=s_R\in\{0,1,2\}$, which we shall call {\sl index of the
resonance $R$}. Then,  to each resonance $R\subset T$ we 
shall associate a family $\cF_R$ of trees $T'$ obtained by
(possibly) changing the edges connecting $R$ with $T\bks R$
and choosing a suitable set of indices .
The family $\cF_R$ and the index $\b'\=\b'(T')$ will be chosen so that
(recall \equ{1.pi})
\beq{4.12}
\sum_{T'\in\cF_R} \L(T',\a,\b')\ 
\pi_R(x;T',\a,\b')=O(x^s)
\eeq
and the family $\cF_C$ will simply be given by
\beq{4.13}
\cF_C=\bigcup_{i=1}^h \cF_{R_i} \ .
\eeq
Let $R$ be a resonance and let $u$ be its first vertex
and  $w'$ the first vertex following $R$. We define the {\sl index
of $R$} as
\beq{4.in}
s_R\=\s_u+\r_u-\r_{w'}\ .
\eeq
Thus if $C=(R_1,...,R_h)$ is a chain of resonances and if $\bar \s$
denotes its order (see \S~\ref{sec:par2}), then
\beq{4.14}
\sum_{i=1}^h s_{R_i}= \bar \s +\r_{u_1} - \r_{z}
\eeq
where $z$ is the first vertex following $R_h$ (which, by convention,
is the last resonance in the chain $C$). Hence, if \equ{4.12} holds, 
from \equ{4.13}, \equ{4.14}, \equ{1.pi} and the definition of $\L$
\equ{4.11'} it follows  easily that
\beq{4.15}
\sum_{T'\in \cF_C} \L(T',\a,\b')\ \pi_{C}(x;T',\a,\b')=
O(x^{\bar \s +\r_{u_1} - \r_{z}})
\eeq
which implies that the chain $C$ is compensable. 

\giu
Let $R\subset T$ be a resonance and let $\hat u u$ and $w {w'}$
be the edges connecting $R$ with $T\bks R$, with $\hat u>u\ge w>{w'}$
(hence $u,w\in R$).
We proceed by constructing the family $\cF_R$.
If $\r_{w'}\neq 1$ we set $\cF_R'=\{T\}$; if 
$\r_{w'}=1$ we let $\cF_R'$ be the family of all trees
$T'$ obtained from $T$ by replacing the edge $w {w'}$ with the
edge $\bar w {w'}$, as $\bar w$ varies in $R$. Hence, 
$T\in\cF_R'$ and if $\r_{w'}=1$, $\# \cF_R'=\# R$.
If $\s_u+\r_u\neq 3$ we set $\cF_R''=\{T\}$; if $\s_u+\r_u=3$
(\ie $(\s_u,\r_u)=(1,2)$ or $(\s_u,\r_u)=(2,1)$) 
we let 
$\cF_R''$ be the family of all trees
$T'$ obtained from $T$ by replacing the edge $\hat u u$ with the
edge $\hat u\bar u$, as $\bar u$ varies in $R$.
As above,  
$T\in\cF_R''$ and if $\s_u+\r_u=3$, $\# \cF_R'=\# R$.
In the first case (\ie for $T'\in\cF_R'$) 
we do not modify the values $\b_v$
(\ie $\b'(T')\=\b$). In the second case (\ie for $T'\in \cF_R''$)
we define $\b'\=\b'(T')$ as follows. 
If $v\notin R$ then $\b'_v=\b_v$.
Recall that, by definition, 
the first vertex of $R$ considered as subtree of $T'$ is
$\bar u$ while the first vertex of $R$ considered as subtree of $T$ is $u$.
Let $\bar u\neq u$ (otherwise, obviously, we set $\b'\=\b$) and
consider the path $P(u,\bar u)$ connecting $u$ with $\bar u$.
The path $P$ will be formed by $p\ge 2$
ordered vertices that we denote $v_i$: the order is such that 
$v_1=u$, $v_p=\bar u$ and the edges of the path are $v_1v_2$,...,$v_{p-1}v_p$.
We then set 
\beqa{4.be}
& \b_{v_p}'\= \b_{v_1}\ , &\nonumber\\
& \b_{v_i}'\=\big( \s_{v_{i+1}}, 
4-\s_{v_{i+1}}-\r_{v_{i+1}}\big)\ ,  &{\rm for}\ 1\le i\le p-1\nonumber\\
& \b_v'\=\b_v\ , &\forall\ v\notin P(u,v)\ .
\eeqa
It is easy to see that this definition is well posed
(see also (ii) of the following Remark).
Finally, we define 
\beq{4.fa}
\cF_R\=\cF_R' \cup \cF_R''\ .
\eeq
\rem{rem:4.1} (i) If $s_R=0$, it is $\cF_R=\{T\}$ and $\b'\=\b$. 

\noindent
(ii) 
The map $\b\to \b'$ is involutive: more precisely,  
if we denote by $\b'(\b;u,\bar u)$ the map defined in \equ{4.be}
(definition which depends on  the ordered path $P(u,\bar u)$)
then  $\b'(\b'(\b;u,\bar u);\bar u,u)=\b$.

\noindent 
(iii) One might say that the families $\cF_R'$ and $\cF_R''$ are
constructed ``going around" the resonance $R$ with a ``discrete
curve" obtained by moving, respectively, the edge connecting 
$R$ with $w'$ and the edge connecting $\hat u$ with $R$.
This interpretation might explain the name ``index" given to the quantity 
$s_R$: $\cF_R$ is obtained by ``going around" $R$ exactly $s_R$ times.

\noindent
(iv) (On the definition of $\cF_R'$) 
In practice, moving around the edge $\bar w z$ produces a factor 
proportional to $\a_{\bar w}$ in \equ{4.11'}  coming from the 
$\dpr^{(\r_z)}=\dpr^{(1)}$ appearing in \equ{4.9} (as
$z\in \calN_{\bar w}$). It is easy to see that, for $x=0$,
summing $\L$ $\pi_R(x)$ over the family $\cF_R'$ produces
a {\sl common} factor $\sum_{\bar w\in R}\a_{\bar w}$
which vanishes by definition of resonance.

\noindent
(v) (On the definition of $\cF_R''$) 
The idea is similar: one wants to produce a factor proportional
to $\a_{\bar u}$ when moving around the ``first" edge 
$\hat u \bar u$ connecting $R$ with $\hat u$. But now the
situation is more delicate as changing the connection $\hat u\bar u$
{\sl changes the order} in $R$
and, consequently, change the small divisors and also
the structure of the derivatives (since both $\g_v$ and $\L_v$
depend on the order).
Since  one wants eventually
to set $x=0$  and collect the factor $\sum_{\bar u\in R}\a_{\bar u}$
one sees the necessity of changing the values of the indices $\b$.
In fact $\b'$ is defined in such a way that, when $x=0$, one can
factor out the {\sl product} of divisors and the {\sl products}
of the operators (derivatives) $\L_v$.
\erem
%
With the above remarks in mind, it is not difficult to verify that 
\equ{4.12} holds, proving Theorem~\ref{thm1} in case P1). 
\section{Compensations II (Maximal Lagrangian tori)}
\label{sec:par5}
\setcounter{equation}{0}
%
Consider the model introduced in P2) of \S~\ref{sec:par2} for a
real--analytic Lagrangian $L\= L_0(y)+\e L_1(x,y)$. 
Formal quasi--periodic solutions of the Lagrangian equations have the form
$x(t)=Z(\o t)$ for an $\o$ satisfying \equ{1.om} and a function $Z(\th)$
which is the (unique) formal
solution $Z\sim\sum_{k\ge 0} \e^k Z^k$, with $Z\in\real^N$, of \equ{1.h2}.
The vector valued function $Z^k$
can be put in a form which is amazingly similar to the Hamiltonian case P1).

\nin
Let us denote by $Z^{(1)k}$ the full vector $Z^k \in \real^N$ and by
$Z^{(2)k}$ the vector $DZ^k=DZ^{(1)k}$ (where $D\= \o\cdot\partial_\th$) 
and, as before,
let $\dpr^{(1)}\=\dpr_x$, $\dpr^{(2)}\=\dpr_y$ and
$A_L \= \dpr^2_y L_0(y_0)$. 
Equation \equ{1.h2} can then be written as
\beq{5.1}
A_L D Z^{{(2)} k} = - D[\dpr^{(2)} L_0]_k^{(k-1)} - 
D[\dpr^{(2)} L_1]_{k-1}^{(k-1)} + [\dpr^{(1)} L_1]_{k-1}^{(k-1)}\ .
\eeq
adopting the same convention used in \equ{4.1}: in particular the arguments
of the functions within square brackets are as in \equ{4.pol}.
As for the previous case (see \equ{4.4}), the $n$--Fourier coefficient of
\equ{5.1} is 
\beq{5.2}
Z^{(\r)k}_n= \sumsud{\s\in \{0,1,2\}^*}{\chi\in\{0,1\} }
\lg n\rg^{-\s}\ 
\Big\{ [D_L^{(\s,\r)} L_\chi]_{k-\chi}^{(k-1)}\Big\}_n
\eeq
where $D_L^{(\s,\r)}$ is now the vector--valued 
operator
\beq{5.3}
D_L^{(\s,\r)}\=(-1)^{1-(\s+\r)} \ i^{-\s}\   A_L^{-1}\ 
\dpr^{(4-\s-\r)}\ ,
\eeq
and $\{0,1,2\}^*$ is the same set of equation \equ{4.4}.

\nin
The Lagrangian problem is now in a form which is identical to 
the Hamiltonian case, except for the definition of $D_L^{(\s,\r)}$. Therefore
the tree expansion formula for the  component $j_0$ of the $n$--Fourier 
coefficient of $Z^k$, with
fixed values of $\r_0$ and $j_0$ at the root $r$, is given again by
\equ{4.8} with the only proviso of replacing $D^{(\s_v,\r_v)}_{j_v}$ in 
\equ{4.9} with $D^{(\s_v,\r_v)}_{L,j_v}$.

\noindent
Also, the families $\cF$ of trees for which there are compensations are found
exactly as in the Hamiltonian case and we refer to \S~\ref{subsec:4.2}
for details.
\section{Compensations III (Lower dimensional tori)}
\label{sec:par6}
\setcounter{equation}{0}
%
Recall the notations of \S~\ref{sec:par2}, P3). Because of the particular 
form of the Hamiltonian, the formal solution\footnote{Recall that here $\e$
is a fixed real number different from zero, while the (complex) 
perturbation parameter appearing in the formal power series is $\m$.}
$Z(\th,\m)$ is of the form
$Z\=(X,Q,DX,DQ)$ where, as usual, 
$D\=\o\cdot \dpr_\th$, $\th\in\tN$ and $X\in \rN$, $Q\in \rM$.
Denote
\beq{6.2}
Z^{(1)}\=X\ ,\quad
Z^{(2)}\=Q\ ,\quad
\dpr^{(1)}\=\dpr_x\ ,\quad
\dpr^{(2)}\=\dpr_q\ ,\quad
A\=\e \dpr^2_q f(q_0)\  (>0)\ .
\eeq
One checks immediately that the recursive equations for $Z^{(\r)k}$
($\r=1,2$) are
\beq{6.eq}
\Big( -D^2+(\r-1) A\Big) Z^{(\r)k}= \sum_{\chi=0,1}\big[\dpr^{(\r)} H_\chi
\big]^{(k-1)}_{k-\chi}\ ,\qquad \r=1,2
\eeq
where the argument of the derivatives of 
$H_\chi$ is $x=\th$, $q=q_0$. From \equ{6.eq} one can see that 
the average (over $\tN$) of $Z^{(1)k}$ vanishes, while the average of
$Z^{(2)k}$ is as in \equ{4.3} but with the minus sign replaced by  a plus.
Taking Fourier coefficients of \equ{6.eq} we get the analogous of \equ{4.4}, 
namely
\beq{6.fo}
Z^{(\r)k}_n= \sumsud{\s\in \{0,2\}^*}{\chi\in\{0,1\} }
\lg n\rg^{-\s}\ 
\Big\{ [D^{(\r)}_n H_\chi]_{k-\chi}^{(k-1)}\Big\}_n\ ,\qquad
\big(\lg n\rg\= \o\cdot n\big)\ ,
\eeq
which differs from \equ{4.4} for the range of $\s$ and relative constraints:
\beq{6.co}
\s \in \{0,2\}^*\ \iff \ \s+\r\in\{2,3\}\ ,\qquad
n=0 \ \implies \ \s=0\ ,
\eeq
and for the definition of the vector valued operator $D^{(\r)}_n$:
\beq{6.dr}
D^{(\r)}_n\= \Big( \lg n\rg^2+ A\Big)^{1-\r} \ \dpr^{(\r)}\ .
\eeq
Notice that the components $D^{(\r)}_{nj}$ are $N$ if $\r=1$ and $M$
if $\r=2$; we therefore let
\beq{6.N}
N_1\=N\ ,\qquad N_2\= M \ \implies j\in\{1,...,N_\r\}\ .
\eeq
Recall that $A=\e \dpr^2 f_0(q_0)$ is
positive definite, and so is $a+A$ for any $a\ge 0$; thus 
\beq{6.aa}
\|(a+A)^{-1}\| \le \frac{b}{|\e|}
\eeq
for a suitable constant $b$ depending only on $\dpr^2 f_0(q_0)$.
We also remark that, by \equ{6.co}, $\r=2$ implies $\s=0$ (hence no
divisors)
while if $\r=1$ then $\s=2$ and $n\neq 0$, in which case the 
divisors are $\lg n\rg^2$.
In view of these remarks, we see that \equ{4.8} holds also in the present
case provided we change the following items:
$N$ in the fourth sum is replaced (see \equ{6.N})
by $N_{\r_v}$; the index set $B$ is defined as
\beqa{6.B}
& B & \=\Big\{\b=(\s,\r):\ \s\in\{0,2\};\ \r\in\{1,2\};\ 
{\rm s.t.}\nonumber\\ 
&& \qquad 
\s+\r\in\{2,3\}\ ,\d_v=0\ \implies \  \s=0 \Big\}\ ;
\eeqa
finally the operator $\L_v$ now depends also on $\a$:
$\L_v(T_r,\b)$ in \equ{4.8} is now replaced by
\beq{6.L}
\L_v(T_r,\a,\b)\=
D^{(\r_v)}_{\d_v j_v} 
\prod_{v'\in \calN_v} \dpr_{j_{v'}}^{(\r_{v'})}\ .
\eeq
Also \equ{4.11'} is readily adapted replacing $N$ by $N_{\r_v}$
and $\L_v$ with \equ{6.L}.
We can proceed to define the families of trees $\cF$ and relative
indices $\b'$ which exhibit compensations. Given a resonance
$R$ (and a choice of $\a$ and $\b$), we define  the index of $R$ and 
the family $\cF_R''$ and relative indices $\b'$
exactly in the same way we did in
\S~\ref{subsec:4.2}. Also the family $\cF_R'$ 
is defined in the same way but the relative indices $\b'$ are now
defined in a slightly different way
(due to the different definition of $D^{(\r)}_n$): we let 
(same notations as in \S~\ref{subsec:4.2})
\beqa{6.be}
& \b_{v_p}'\= \b_{v_1}\ , &\nonumber\\
& \b_{v_i}'\=\b_{v_{i+1}}\ ,
&{\rm for}\ 1\le i\le p-1\nonumber\\
& \b_v'\=\b_v\ , &\forall\ v\notin P(u,v)\ .
\eeqa
With these definitions it is easy to check that 
\equ{4.12} holds and hence that
Theorem~\ref{thm1} is valid in case P3) too. 

\giu
We close by a remark on the $\m$--radius of convergence 
of $\sum_{k\ge 1} \m^k Z^k$. It is an easy exercise to
adapt the estimates in \cite{CF1} to the present case
and to check how the radius of convergence depends on $\e$.
In fact, observing that, by \equ{6.aa}, one has 
\beq{6.es}
|\lg n\rg^{-\s}|\ |\big(\lg n\rg^2+A\big)^{1-\r}|
\le \max\{\g^2, \frac{b}{|\e|}\}\ \ 
\max\{|n|^{2\t}\ ,\ 1\}\ ,
\eeq
leading to an estimate on the radius of convergence $\m_0$
of the type
\beq{6.mu}
\m_0\ge {\rm const}\ \min\{\e\ ,\ \g^{-2}\}\ .
\eeq
Thus $\m=\e^2$ (or $\m=\e^c$ with any $c>1$) is within the
domain of analyticity provided $\e$ is small enough.
\section{An example with no compensations}
\label{sec:par7}
\setcounter{equation}{0}
%
Consider the Hamiltonian \equ{1.12} with 
\beq{p.1}
f(x,q) =\sum_{s\geq 1} f_s 
e^{i (n^{(s)}\cdot x + m^{(s)}\cdot q)}
\eeq
where $n^{(s)} \in \integer^N$, $m^{(s)} \in \integer^M$
are given integer vectors (with $|n^{(s)}|+|m^{(s)}|>0$) and the Fourier 
coefficients $f_s$ decay exponentially fast with $|n^{(s)}|+|m^{(s)}|$.
Let $q_0$ be a non degenerate critical point of the $x$--average of $f$
(i.e. of $f_0=\sum_{s\geq 1} f_s e^{i m^{(s)}\cdot q}$); then,
as for the previous cases, there exists a (unique) formal power series
\beq{p.2}
Z \= (Z_1,\dots , Z_d)\  \sim\  \sum_{k\geq 0} Z^k(\th)\e^k\ ,
\ \th\in\torus^N
\eeq
with $d=2(N+M)$,
such that $t\to Z(\o t)$ is a formal quasi--periodic solution for \equ{1.12}
and the set $\{Z^0(\th):\th\in\torus^N\}$ coincide with the torus spanned by
$y=\o$, $p=0$, $q=q_0$, $x=x_0+\o t$, where $\o\in\real^N$ satisfies condition
\equ{1.om}.

\nin
Expanding $Z^k$ in Fourier series also in the variable $q$, besides the variable
$x$ (see \equ{1.2}), it is still possible to write $Z_n^k$ as in 
\equ{1.tr} provided one makes the following changes.
$\cT^k_* = \cT^k$; $B$ is the trivial set $B=\{\b =\s = 2\}$;
$\L(T,\a,\b)$ is replaced by
\beq{p.3}
\L(T,\hat \a,\b) = \sum_{\a':V\to \a_v'\in \integer^M}\ \prod_{v\in V} 
f_{\hat\a_v} \prod_{vv'\in E} \hat\a_v \cdot \hat\a_{v'}
\eeq
where $\hat\a_v\= (\a_v,\a_v')\in \integer^{N+M}$; finally
$\g_v = \lg \d_v\rg^{-2}\=\bigl(\o\cdot\sum_{v'\leq v} \a_{v'}\bigr)^{-2}$.

\nin
It is then clear that in order to have compensations it would be 
sufficient to have resonances in the variable 
$\hat\a$ whenever there are resonances in the variable $\a$.
In other words compensations take place if the Fourier modes in \equ{p.1} 
are such that
\beq{p.4}
\sum_{s\in I} n^{(s)} = 0 \quad \implies \sum_{s\in I} m^{(s)} = 0
\eeq
$\forall I\subset \natural$; in fact in such a case one could repeat 
word--by--word the arguments in \cite{CF1}.

\nin
In general, if \equ{p.4} does not hold, compensations (of the type 
described in this paper) do not occur as it is shown by the following 
example.

\giu
Take $N=2$, $M=1$, fix a scalar integer $n\neq 0$ and let
\beq{p.5}
f(x_1,x_2,q) = 2\{\cos x_1 + \cos x_2 + 
\cos (x_1 + x_2 - q) + \cos (n x_1 + x_2 + q)\}\ .
\eeq
Hence, the range of $\hat\a$ is the set
$\{\pm (1,0,0),\ \pm (0,1,0),\ \pm (1,1,-1),\ \pm (n,1,1)\}$; for  
definiteness we also fix $\o = (\sqrt{2},-1)$. 

%
\begin{figure}[hbt]
\begin{center}
\begin{picture}(400,100)
\thicklines
\put(5,60){\circle{8}}
\put(5,60){\circle*{3}}
%\put(10,50){$u_1$}
%\put(50,50){$w_1$}
%\put(100,50){$u_2$}
%\put(140,50){$w_2$}
%\put(190,50){$u_3$}
%\put(230,50){$w_3$}
%\put(280,50){$u_4$}
%\put(320,50){$w_4$}
\multiput(65,60)(90,0){4}{\line(1,0){30}}
\multiput(65,60)(90,0){4}{\circle*{3}}
\multiput(95,60)(90,0){4}{\circle*{3}}
\put(35,60){\oval (75,30)}
\put(125,60){\oval (75,30)}
\put(215,60){\oval (75,30)}
\put(305,60){\oval (75,30)}
\put(350,65){{\scriptsize ($n$,1,1)}}
\put(365,45){$z$}
\put(65,35){$R_1$}
\put(155,35){$R_2$} 
\put(245,35){$R_3$}
\put(335,35){$R_4$}
\parbox{14cm}{\caption{\label{fig:d1} Divergent contribution: the chain 
$C=( R_1,R_2,R_3,R_4 )$ is made of 4 adjacent resonances with $|R_i|=3$ (in 
this symbolic picture only two vertex are drawn).}}
\end{picture}
\end{center}
\end{figure}

\nin
We fix a tree $T\in \widetilde \cT^k$, with $k = 3h+1$, and a function
$\hat \a_v:V(T)\to \integer^3$ so that  $T$ contains 
a chain $C=( R_1,...,R_h )$ made of $h$ identical resonances $R_i\= \{v_1,v_2,
v_3\}$, with $\hat\a_{v_1}=(-1,-1,1)$, $\hat\a_{v_2}=(0,1,0)$,
$\hat\a_{v_3}=(1,0,0)$,  
and the last vertex $z$, following the chain, with 
$\hat\a_z = (n,1,1)$ (see Figure~\ref{fig:d1}).

\noindent
Let $\cF_C$ be the family of all trees $T\in\cT^k$
which contains the chain $C$.
A lengthy but straightforward computation shows that
the $i^{th}$ component of the vector valued function $\bar \pi_C(x)$ 
(see \equ{1.11}) for $h = 1$ is:
\beqa{p.6}
&(\bar\pi_{C}(x))_i \=
(\bar \pi_{R_1}(x))_i & \= \prod_{v\in V} f_{\hat\a_v}
\sum_{u,w\in R_1} 
\hat e_i \cdot \hat \a_u \ \Big(
\prod_{vv'\in E(R_1)} \hat\a_v \cdot \hat\a_{v'}\Big) \ 
\hat \a_w \cdot \hat \a_z \ 
\pi_{R_1}(x)\nonumber \\
&&\ =\ 2\sum_{j=1}^3 (\d_{i3} \d_{j3} + x^2 B_{ij}(x))\hat \a_{zj}
\eeqa
where $\d_{ij}$ is the Kronecker's symbol, $\hat e_i$ is the vector 
of $j^{th}$ component $\d_{ij}$ and
\beq{p.7}
B(x)=\pmatrix{{{-2(6-x^2)}\over {(2-x^2)^2}}&
{{8\sqrt{2}+18 x^2 - 9x^4 + x^6}\over {(2-x^2)^2 (1-x^2)^2}}&
{{6-x^2}\over {(2-x^2)^2}}\cr
&\phantom{a} &\cr
{{8\sqrt{2}+18 x^2 - 9x^4 + x^6}\over {(2-x^2)^2 (1-x^2)^2}}&
{{-2(3-x^2)}\over {(1-x^2)^2}}&
{{3-x^2}\over {(1-x^2)^2}}\cr
&\phantom{a} &\cr
{{6-x^2}\over {(2-x^2)^2}}&
{{3-x^2}\over {(1-x^2)^2}}&0} .
\eeq
Hence $\bar \pi_{R_1}(0)\neq 0$. For $h>1$, one simply has:
\beq{p.8}
\bar \pi_{C}(x) = 2^h (A+x^2 B)^h \hat\a_z\quad,\quad A_{ij}=\d_{i3} \d_{j3}\ .
\eeq
Thus, if $x = \o\cdot \a_z = \sqrt{2}n-1$, taking the third component of 
\equ{p.8} (the component for which condition \equ{p.4} is violated)
and assuming that $n\geq h$, one easily checks that
\beq{p.9}
\bigl( (A+x^2 B)^h \hat\a_z\bigr)_3 \geq 1 + O({1\over n})
\eeq
which implies that $C$ is a non compensable chain.
%%%%%%%%%%%%%%%%%%% bibliography: 

\begin{thebibliography}{99}
%

{\small

\bibitem{A1}
Arnold, V.I. (Ed.): {\em Encyclopaedia of Mathematical Sciences, Dynamical 
Systems}, Vol. 3, Springer--Verlag, 1988 

\bibitem{A2}
Arnold, V.I.: {\em Instability of dynamical systems with several
degrees of freedom} Sov. Math. Dokl. {\bf 5} 581--585 (1964)

\bibitem{Bo}
Bollobas, B.: {\em Graph Theory}, Springer (Graduate text in mathematics: 
63), 1979

\bibitem{Br} 
Brjuno, A. D.: {\em Convergence of transformations of differential 
equations to normal form}, Dokl. Akad. Nauk SSSR {\bf 165}, 987--989
(1965); {\em Analytic form of differential equations},
Trans. Moscow Math. Soc. {\bf 25}, 131--288 (1971) and 
{\bf 26}, 199--239 (1972)

\bibitem{CC}
Celletti, A.; Chierchia, L.: {\em Construction of analytic KAM 
surfaces and effective stability bounds}, Commun. Math.
Phys. {\bf 118}, 119--161 (1988)

\bibitem{CF1}
Chierchia, L.; Falcolini, C.: {\em A direct proof of a theorem of
Kolmogorov in Hamiltonian systems}, 
to appear in Annali Sc. Norm. Super. Pisa, Cl. Sci.

\bibitem{CF2}
Chierchia, L.; Falcolini, C.: {\em Quasi--periodic solutions of some 
elliptic systems}, 
Preprint (1993)

\bibitem{CG}
Chierchia, L.; Gallavotti, G.: {\em Drift and Diffusion in phase space},
Ann. Inst. Henri Poincar\'e (Physique Th\'eorique) {\bf 60}, 
n$^{\rm o}$ 1,
1--144 (1994)

\bibitem{E1}
Eliasson, L. H.: {\em Absolutely convergent series expansions
for quasi periodic motions}, Reports Department of Math., Univ. of
Stockholm, Sweden, No. 2, 1--31 (1988)

\bibitem{E2}
Eliasson, L. H.: {\em Hamiltonian systems with linear normal form near an
invariant torus}, in ``Nonlinear Dynamics", G. Turchetti (Ed.) World
Scientific, Singapore (1989)

\bibitem{E3}
Eliasson, L. H.: {\em Generalization of an estimate of small divisors
by Siegel}, in ``Analysis, et cetera", P.H.~Rabinowitz and E.~Zehnder
(Eds.), Academic Press (1990)

\bibitem{Ga1}
Gallavotti, G.: {\em Twistless KAM tori, quasi flat homoclinic 
intersections, and other cancellations in the perturbation series
of certain completely integrable hamiltonian systems. A review.}
Preprint (1993)

\bibitem{Ga2}
Gallavotti, G.: {\em Twistless KAM tori} to appear in Commun. Math. 
Physics

\bibitem{Ga3}
Gallavotti, G.: {\em Invariant tori: a field theoretic point of view on 
Eliasson's work} Preprint (1993)

\bibitem{GG}
Gallavotti, G.; Gentile, G.: {\em Non recursive proof of the KAM theorem}, 
to appear in Ergodic Theory and Dynamical Systems

\bibitem{Ge1}
Gentile, G.: {\em A proof of existence of whiskered tori with quasi flat 
homoclinic intersections in a class of almost integrable hamiltonian 
systems} Preprint (1994)

\bibitem{Ge2}
Gentile, G.: {\em Whiskered tori with prefixed frequencies and Lyapunov 
spectrum} Preprint (1994) 

\bibitem{Gr}
Graff, S.M.: {\em On the conservation of hyperbolic invariant tori
for Hamiltonian systems}, J. Diff. Equations {\bf 15}, 1--69 (1974)

\bibitem{Po}
Poincar\'e, H.: {\em Les m\'ethodes nouvelles de la m\'ecanique c\'eleste}.
Vols. 1--3. Paris: Gauthier--Villars. (1892/1893/1899)


\bibitem{S}
Siegel, C. L.: {\em Iterations of analytic functions}, Annals of Math. 
{\bf 43}, No. 4, 607--612 (1942)

\bibitem{SZ}
Salamon, D.; Zehnder, E.: {\em KAM theory in configuration space},
Comm. Math. Helv., {\bf 64}, 84 (1988)

\bibitem{X}
Xia, Z.: {\em Arnold diffusion and oscillatory solutions in the three--body
problem}, Preprint (1993)
}
\end{thebibliography}
\end{document}