KAM theory in conguration space and cancellations in the Lindstedt series

Livia Corsi1 , Guido Gentile1 , Michela Procesi2

1 2

Dipartimento di Matematica, Universit` a di Roma Tre, Roma, I-00146, Italy E-mail: lcorsi@mat.uniroma3.it, gentile@mat.uniroma3.it, procesi@unina.it

Dipartimento di Matematica, Universit` a di Napoli Federico II, Napoli, I-80126, Italy

Abstract The KAM theorem for analytic quasi-integrable anisochronous Hamiltonian systems yields that the perturbation expansion (Lindstedt series) for any quasi-periodic solution with Diophantine frequency vector converges. If one studies the Lindstedt series by following a perturbation theory approach, one nds that convergence is ultimately related to the presence of cancellations between contributions of the same perturbation order. In turn, this is due to symmetries in the problem. Such symmetries are easily visualised in action-angle coordinates, where KAM theorem is usually formulated, by exploiting the analogy between Lindstedt series and perturbation expansions in quantum eld theory and, in particular, the possibility of expressing the solutions in terms of tree graphs, which are the analogue of Feynman diagrams. If the unperturbed system is isochronous, Mosers modifying terms theorem ensures that an analytic quasi-periodic solution with the same Diophantine frequency vector as the unperturbed Hamiltonian exists for the system obtained by adding a suitable constant (counterterm) to the vector eld. Also in this case, one can follow the alternative approach of studying the perturbation expansion for both the solution and the counterterm, and again convergence of the two series is obtained as a consequence of deep cancellations between contributions of the same order. In this paper, we revisit Mosers theorem, by studying the perturbation expansion one obtains by working in Cartesian coordinates. We investigate the symmetries giving rise to the cancellations which makes possible the convergence of the series. We nd that the cancellation mechanism works in a completely dierent way in Cartesian coordinates, and the interpretation of the underlying symmetries in terms of tree graphs is much more subtle than in the case of action-angle coordinates.

Introduction

Consider an isochronous Hamiltonian system, described by the Hamiltonian H (, A) = A + f (, A), with f real analytic in Td A and A an open subset of Rd . The corresponding Hamilton equation are = + A f (, A), = f (, A). A (1.1)

Let (0 (t), A0 (t)) = (0 + t, A0 ) be a solution of (1.1) for = 0. For = 0, in general, there is no quasi-periodic solution to (1.1) with frequency vector which reduces to (0 (t), A0 (t)) as 0. However, one can prove that, if is small enough and satises some Diophantine condition, then there is a correction (, A0 ), analytic in both and A0 , such that the modied equations = + A f (, A) + (, A0 ), = f (, A), A (1.2)

admit a quasi-periodic solution with frequency vector which reduces to (0 (t), A0 (t)) as 0. This is a well known result, called the modifying terms theorem, or translated torus theorem, rst proved by Moser [18]. By writing the solution as a power series in (Lindstedt series), the existence of an analytic solution means that the series converges. This is ultimately related to some deep cancellations in the series; see [1] for a review. Equations like (1.1) naturally arise when studying the stability of an elliptic equilibrium point. For instance, one can think of a mechanical system near a minimum point for the potential energy, where the Hamiltonian describing the system looks like H (x1 , . . . , xn , y1 , . . . , yn ) = 1 2
d 2 2 2 yj + j xj + F (x1 , . . . , xn , ), j =1

(1.3)

where F is a real analytic function at least of third order in its arguments, the vector = (1 , . . . , d ) satises some Diophantine condition, and the factor can be assumed to be obtained after a rescaling of the original coordinates. Indeed, the corresponding Hamilton equations, written in action-angle variables, are of the form (1.1). Unfortunately, the action-angle variables are singular near the equilibrium, and hence there are problems in the region where one of the actions is much smaller than the others. Thus, it can be worthwhile to work directly in the original Cartesian coordinates. In fact, there has been a lot of interest for KAM theory in conguration space, that is, without action-angle variables; see for instance [20, 17, 6]. In the light of Mosers theorem of the modifying terms, one expects that, by writing
2 xj = xi F (x1 , . . . , xn , ) + j xj , x j + j

(1.4)

ij t and taking the (arbitrary) unperturbed solution x0,j (t) = Cj cos j t + Sj sin j t = cj eij t + c , je j = 1, . . . , d, there exists a function (, c), analytic both in and c = (c1 , . . . , cd ), such that by xing j = j (, c), there exists a quasi-periodic solution to (1.4) with frequency vector , which reduces to the unperturbed one as 0. One can try to write again the solution as a power series in , and study directly the convergence of the series. In general, when considering the Lindstedt series of some KAM problem, rst of all one identies the terms of the series which are an obstruction to convergence: such terms are usually called resonances (or self-energy clusters, by analogy to what happens in quantum eld theory). Crudely speaking, the series is given by the sum of innitely many terms (nitely many for each perturbation order), and each term looks like a product of small divisors ( i ), i Zd , times some harmless factors: a resonance is a particular structure in the product which allows a dangerous accumulation of small divisors. This phenomenon is very easily visualised when each term of the series is graphically represented as a tree graph (tree tout court in the following), that is, a set of points and lines connecting them in such a way that no loop arises. In terms of trees, each line carries a label Zd (that one call momentum, again inspired by the terminology of quantum eld theory) and with each such line a small divisor ( ) is associated: then a resonance becomes a subgraph which is between two lines with the same and hence the same small divisor ( ). A tree with a chain of resonances represents a term of the series containing a factor ( ) to a very large power, and this produces a factorial k ! to some positive power when bounding some terms contributing to the k -th order in of the Lindstedt series, so preventing a proof of convergence. However, a careful analysis of the resonances shows that there are cancellations to all perturbation orders. This is what can be proved in the case of the standard anisochronous KAM theorem, as rst pointed out by Eliasson [8]; see also [9, 10], for a proof which more deeply exploits the similarity with the techniques of quantum eld theory. More precisely the cancellation mechanism works in the following way. Given two lines 1 and 2 with the same small divisor, consider all possible resonances which can be inserted between 1 and 2 : when

summing together the numerical values corresponding to all such resonances, there are compensations and the sum is in fact much smaller than each summand (for more details we refer to [10, 13]). For the isochronous case, already in action-angle variables [1], there are some kinds of resonances which do not cancel each other. Nevertheless there are other kinds of resonances for which the gain factor due to the cancellation is more than what needed (that is, one has a second order instead of a rst order cancellation). Thus, the hope naturally arises that one can use the extra gain factors to compensate the lack of gain factors for the rst kind of resonances, and in fact this happens. Indeed, the resonances for which there is no cancellation cannot accumulate too much without entailing the presence of as many resonances with the extra gain factors, in such a way that the overall number of gain factors is, in average, one per resonance (this is essentially the meaning of Lemma 5.4 in [1]). When working in Cartesian coordinates, one immediately meets a diculty. If one writes down the lowest order resonances, there is no cancellation at all. This is slightly surprising because a cancellation is expected somewhere: if the resonances do not cancel each other, in principle one can construct chains of arbitrarily many resonances, which produce factorials in the formal power series expansion. However, we shall show that there are cancellations, as soon as one has at least two resonances. So, one has the curious phenomenon that resonances which do not cancel each other are allowed, but they cannot accumulate too much. Moreover, the cancellation mechanism is more involved than in other cases (including the same problem in action-angle variables). First of all, the resonances are no longer diagonal in the momenta, that is, the lines 1 and 2 considered above can have dierent momenta 1 and 2 . Second, the cancellation does not operate simply by collecting together all resonances to a given order and then summing the corresponding numerical values. As we mentioned, in this way no cancellation is produced: to obtain a cancellation one has to consider all possible ways to connect to each other two resonances. Thus, there is a cancellation only if there is a chain of at least two resonances. What emerges eventually is that working in Cartesian coordinates rather complicates the analysis. On the other hand, as remarked above, it can be worthwhile to investigate the problem in Cartesian coordinates. Moreover, the cancellations are due to remarkable symmetries in the problem, which can be of interest by their own; in this regard we mention the problem of the reducibility of the skew-product ows with Bryuno base [11], where the convergence of the corresponding Lindstedt series is also due to some cancellation mechanism and hence, ultimately, to some deep symmetry of the system. In this paper we shall assume the standard Diophantine condition on the frequency vector ; see (2.3) below. Of course one could consider more general Diophantine conditions than the standard one (for instance a Bryuno condition [5]; see also [12] for a discussion using the Lindstedt series expansion). This would make the analysis slightly more complicate, without shedding further light on the problem. An important feature of the Lindstedt series method is that, from a conceptual point of view, the general strategy is exactly the same independently of the kind of coordinates one uses (and independently of the fact that the system is a discrete map or a continuous ow; see [2, 10]). What is really important for the analysis is the form of the unperturbed solution: the simpler is such a solution, the easier is the analysis. Of course, an essential issue is that the system one wants to study is a perturbation of one which is exactly soluble. This is certainly true in the case of quasi-integrable Hamiltonian systems, but of course the range of applicability is much wider, and includes also non-Hamiltonian systems; see for instance [15, 14]. Moreover an assumption of this kind is more or less always implicit in whatsoever method one can envisage to deal with small divisor problems of this kind. In the anisochronous case, the cancellations are due to symmetry properties of the model, as rst pointed out by Eliasson [8]. Ultimately, the cancellation mechanism for the resonances is deeply related to that assuring the formal solubility of the equations of motions, which in turn is due to a symmetry property, as already shown by Poincar e [19]. Subsequently, stressing further the analogy with quantum eld theory, Bricmont et al. showed that the cancellations can be interpreted as a consequence of suitable Ward identities of the corresponding eld theory [4] (see also [7]). In the isochronous case, in terms of Cartesian coordinates the cancellation mechanism works in a completely dierent way with respect to 3

action-angle coordinates. However, as we shall see, the cancellation is still related to underlying symmetry properties: it would be interesting to provide an interpretation of the symmetry properties that we nd in terms of some invariance property of the corresponding quantum eld model.

Statement of the results

2 x j + j xj + fj (x1 , . . . , xd , ) + j xj = 0,

Consider the ordinary dierential equations j = 1, . . . , d, (2.1)

where x = (x1 , . . . , xd ) Rd , is real parameter (perturbation parameter ), the function f (x, ) = (f1 (x, ), . . . , fd (x, )) is real analytic in x and at (x, ) = (0, 0) and at least quadratic in x,

fj (x, ) =
p=1

sd 1 fj,s1 ,...,sd xs 1 . . . xd ,

(2.2)

s1 ,...,sd 0 s1 +...+sd =p+1

= (1 , . . . , d ) is a vector of parameters, and the frequency vector (or rotation vector ) = (1 , . . . , d ) satises the Diophantine condition | | > 0 | | Zd , (2.3)

d with Zd = Z \ {0}, > d 1 and 0 > 0. Here and henceforth denotes the standard scalar product in d R , and | | = |1 | + . . . + |d |. The equations (2.1) naturally arise from equations of the form 2 x j + j xj + gj (x1 , . . . , xd ) + j xj = 0,

j = 1, . . . , d,

with g = (g1 , . . . , gd ) analytic and at least quadratic in x, after rescaling the coordinates: x x. Such rescaling makes sense if one wants to study the behaviour of the system near the origin. We look for quasi-periodic solutions x(t) of (2.1) with frequency vector . Therefore we expand the function x(t) by writing x(t) = ei t x , (2.4)
Zd

and we denote by f (x, ) the -th Fourier coecient of the function that we obtain by Taylor-expanding f (x, ) in powers of x and Fourier-expanding x according to (2.4). Thus, in Fourier space (2.1) becomes
2 xj, = fj, (x, ) + j xj, . ( )2 j

(2.5)

For = 0, = 0, the vector x(0) (t) with components

i j t , xj (t) = cj eij t + c je (0)

j = 1, . . . , d,

(2.6)

is a solution of (2.1) for any choice of the complex constant c = (c1 , . . . , cd ). Here and henceforth denotes complex conjugation. Dene ej as the vector with components ij (Kronecker delta). Then we can split (2.5) into two sets of equations, called respectively the bifurcation equation and the range equation, fj,ej (x, ) + j xj,ej = 0, ( )
2 2 j

j = 1, . . . , d,

= 1, j = 1, . . . , d, = ej .

(2.7a) (2.7b)

xj, = fj, (x, ) + j xj, , 4

We shall study both equations (2.7) simultaneously, by showing that for all choices of the parameters c there exist suitable counterterms , depending analytically on and c, such that (2.7) admits a quasiperiodic solution with frequency vector , which is analytic in , c, and t. Moreover, with the choice xj,ej = cj for all j = 1, . . . , d, the counterterms are uniquely determined. We formulate the following result. Theorem 2.1. Consider the system described by the equations (2.1) and let (2.6) be a solution at = 0, = 0. Set (c) = max{|c1 |, . . . , |cd |, 1}. There exist a positive constant 0 , small enough and independent of , c, and a unique function (, c), holomorphic in the domain ||3 (c) 0 and real for real , such that the system 2 x j + j xj + fj (x1 , . . . , xd , ) + j (, c) xj = 0, j = 1, . . . , d, admits a solution x(t) = x(t, , c) of the form (2.4), holomorphic in the domain ||3 (c)e3|| |Im t| 0 and real for real , t, with Fourier coecients xj,ej = cj and xj, = O() if = ej for j = 1, . . . , d. The proof is organised as follows. After introducing the small divisors and proving some simple preliminary properties in Section 3, we develop in Section 4 a graphical representation for the power series of the counterterms and the solution (tree expansion). In particular we perform a multiscale analysis which allows us to single out the contributions (self-energy clusters) which give problems when trying to bound the coecients of the series. In Section 5 we show that, as far as such contributions are neglected, there is no diculty in obtaining power-like estimates on the coecients: these estimates, which are generalisations of the Siegel-Bryuno bounds holding for anisochronous systems [9, 10], would imply the convergence of the series and hence analyticity. In Section 6 we discuss how to deal with the self-energy clusters: in particular we single out the leading part of their contributions (localised values), which are proved in Section 7 to satisfy some deep symmetry properties. Finally, in Section 8 we show how the symmetry properties can be exploited in order to obtain cancellations involving the localised parts, in such a way that the remaining contributions can still bounded in a summable way. This will yield the convergence of the full series and hence the analyticity of both the solution and the counterterms. Note that the system dealt with in Theorem 2.1 can be non-Hamiltonian. On the other hand the most general case for a Hamiltonian system near a stable equilibrium allows for Hamiltonians of the form H (x1 , . . . , xn , y1 , . . . , yn ) = which lead to the equations x j = yj + yi F (x, y, ), 2 y j = j xj xi F (x, y , ). Also in this case one can consider the modied equations x j = yj + yi F (x, y , ), 2 y j = j xj xi F (x, y , ) + j xj , (2.10) (2.9) 1 2
d 2 2 2 yj + j xj + F (x1 , . . . , xn , y1 , . . . , yn , ), j =1

(2.8)

which are not of the form considered in Theorem 2.1. However, a result in the same spirit as Theorem 2.1 still holds. Theorem 2.2. Consider the system described by the equations (2.10) and let (x(0) (t), y (0) (t)) be a solu (0) (t). Set (c) = max{|c1 |, . . . , |cd |, 1}. tion at = 0, = 0, with x(0) (t) given by (2.6) and y (0) (t) = x Then there exist a positive constant 0 , small enough and independent of , c, and a unique function

 (, c), holomorphic in the domain ||3 (c) 0 and real for real , such that the system x j = yj + yi F (x, y , ), 2 y j = j xj xi F (x, y , ) + j (, c) xj admits a solution (x(t, , c), y (t, , c)), holomorphic in the domain ||3 (c)e3|| |Im t| 0 and real for real , t, with Fourier coecients xj,ej = yj,ej /ij = cj and xj, = yj, = O() if = ej for j = 1, . . . , d. The proof follows the same lines as that of Theorem 2.1, and it is discussed in Appendices A and B. Finally in Appendix C we briey sketch an alternative approach based on the resummation of the perturbation series.

Preliminary results

We shall denote by N the set of (strictly) positive integers, and set Z+ = N {0}. For any j = 1, . . . , d and Zd dene the small divisors j ( ) := min{| j | , | + j |} = | ( ( , j ) ej )|, where ( , j ) is the minimizer. Note that the Diophantine condition (2.3) implies that j ( ) | |

(3.1)

j = 1 . . . , d,

= 0, ( , j ) ej , j, j =1, . . . , d,

(3.2a)

j ( )+ j ( ) | |

= , = ( , j ) ej ( , j ) ej , (3.2b)

for a suitable positive > 0. We can (and shall) assume that is suciently smaller than 0 , and hence than (0) = min{|1 |, . . . , |d |} and := min{||i | |j || : 1 i < j d}. Lemma 3.1. Given , Zd , with = , and j ( ) = j ( ) for some j, j {1, . . . , d}, then either | | | | + | | 2 or | | = 2. Proof. One has j ( ) = | j | and j ( ) = | j |, with = ( , j ) and = ej and = ej . By the Diophantine condition (2.3) one can have = ( , j ). Set | = | |, if and only if = . j ( ) = j ( ), and hence | = then for = one has | | = | | + | |, while for = one obtains | | If = and j = j one has i = i | | + | | 2. If for all i = j and j = j , and hence = and j = j then i = i for all i = j, j , while j = j and j = j |j j | = 2. If , and hence |j j | = |j j | = 1. Lemma 3.2. Let , Zd be such that = and, for some n Z+ , j, j {1, . . . , d}, both j ( ) 2n and j ( ) 2n hold. Then either | | > 2(n2)/ or | | = 2 and j ( ) = j ( ). Proof. Write j ( ) = | j | and j ( ) = | j |, with = ( , j ) and = ( , j ), = ej and = ej as above. and set = , by the Diophantine condition (3.2b), one has If | |

)| | | + | | < 2(n1) , < | (

| > 2(n1)/ , and hence we have | | > 2(n2)/ in such a case. which implies | = then, as in Lemma 3.1, one has | | = 2 and j ( ) = j ( ). If

Remark 3.3. Note that | | 2 and j ( ) = j ( ) if and only if = ( , j )ej ( , j )ej . Lemma 3.4. Let 1 , . . . , p Zd and j1 , . . . , jp {1, . . . , d}, with p 2, be such that | i i1 | 2 and ji ( i ) = j1 ( 1 ) for i = 2, . . . , p. Then | 1 p | 2. i = i i eji for i = 1, . . . , p. For all i = 2, . . . , p, the assumption Proof. Set i = ( i , ji ) and i = i1 , which in turn yields i = i1 since | i i1 | 2. In ji ( i ) = ji1 ( i1 ) implies 1 = p , and hence | 1 p | 2. particular

Multiscale analysis and diagrammatic rules

As we are looking for x(t, , c) and (, c) analytic in , we formally write xj, =

k=0

k xj, ,

(k )

j =
k=1

k j .

(k )

(4.1)

It is not dicult to see that using (4.1) in (2.7) one can recursively compute (at least formally) the (k ) (k ) coecients xj, , j to all orders. Here we introduce a graphical representation for each contribution to xj, , j , which will allow us to study the convergence of the series.
(k ) (k )

4.1

Trees

A graph is a connected set of points and lines. A tree is a graph with no cycle, such that all the lines are oriented toward a unique point (root ) which has only one incident line (root line ). All the points in a tree except the root are called nodes. The orientation of the lines in a tree induces a partial ordering relation ( ) between the nodes and the lines: we can imagine that each line carries an arrow pointing toward the root; see Figure 1. Given two nodes v and w, we shall write w v every time v is along the path (of lines) which connects w to the root.

Figure 1: An unlabelled tree: the arrows on the lines all point toward the root, according to the tree partial ordering.

We call E () the set of end nodes in , that is, the nodes which have no entering line, and V () the set of internal nodes in , that is, the set of nodes which have at least one entering line. Set N () = E () V (). For all v N () denote by sv the number of lines entering the node v . 7

Remark 4.1. One has

v V ( )

sv = |N ()| 1.

We denote by L() the set of lines in . We call internal line a line exiting an internal node and end line a line exiting an end node. Since a line L() is uniquely identied with the node v which it leaves, we may write = v . We write w v if w v ; we say that a node w precedes a line , and write w , if w . Notation 4.2. (1) If and are two comparable lines, i.e., , we denote by P (, ) the (unique) path of lines connecting to . (2) Each internal line L() can be seen as the root line of the tree whose nodes and lines are those of which precede , that is, N ( ) = {v N () : v } and L( ) = { L() : }.

4.2

Tree labels

With each end node v E () we associate a mode label v Zd , a component label jv {1, . . . , d}, and a sign label v {}; see Figure 2. We call Ej () the set of end nodes v E () such that jv = j and v = . With each internal node v V () we associate a component label jv {1, . . . , d}, and an order label kv Z+ . Set V0 () = {v V () : kv = 0} and N0 () = E () V0 (). We also associate a sign label v {} with each v V0 (). The internal nodes v with kv 1 will be drawn as black bullets, while the end nodes and the internal nodes with kv = 0 will be drawn as white bullets and white squares, respectively; see Figure 2.

v jv v (a) v (b)

jv kv (c) v

jv kv v v

Figure 2: Nodes and labels associated with the nodes: (a) end node v with sv = 0, jv {1, . . . , d}, v {}, and v = v ejv (cf. Section 4.3); (b) internal node v with sv 2, jv {1, . . . , d}, and kv = sv 1 (cf. Section 4.3); (c) internal node v with sv = 2, jv {1, . . . , d} kv = 0, v {} (cf. Section 4.3).

With each line we associate a momentum label Zd , a component label j {1, . . . , d}, a sign label {}, and scale label n Z+ {1}; see Figure 3. j n
Figure 3: Labels associated with a line. One has = ( , j ) (cf. Section 4.3) Moreover if = v then j = jv ; if v V0 ( ) one has also = v ; if = ej then n = 1, otherwise n 0 (cf. Section 4.3).

Denote by sv,j the number of lines with component label j = j entering the node v , and with rv,j, the number of end lines with component label j and sign label which enter the node v . Of course sv = sv,1 + . . . + sv,d and sv,j rv,j,+ + rv,j, for all j = 1, . . . , d. Finally call kv k () :=
v V ( )

the order of the tree . In the following we shall call trees tout court the trees with labels, and we shall use the term unlabelled trees for the trees without labels.

4.3

Constraints on the tree labels

Constraint 4.3. We have the following constraints on the labels of the nodes (see Figure 2): (1) if v V () one has sv 2; (2) if v E () one has v = v ejv ; (3) if v V () then kv = sv 1, except for sv = 2, where both kv = 1 and kv = 0 are allowed. Constraint 4.4. The following constraints will be imposed on the labels of the lines: (1) j = jv , = v , and = v if exits v E (); (2) j = jv if exits v V (); (3) if is an internal line then = ( , j ), i.e., j ( ) = | j | (see (3.1) for notations); (4) if v V0 () then (see Figure 4) 1. sv = 2; 2. both lines 1 and 2 entering v are internal and have 1 = 2 = v and j1 = j2 = jv ; 3. either 1 = v ejv and 2 = v ejv or 1 = v ejv and 2 = v ejv ; 4. v = v ; (5) if is an internal line and = ej , then enters a node v V0 (); (6) n 0 if = ej and n = 1 otherwise. v ejv v jv v v jv v jv 0 v v 1 2 v jv 2
Figure 4: If there is an internal node v with kv = 0 then sv = 2 and the following constraints are imposed on the other
labels: v = 1 = 2 = v ; jv = j1 = j2 = jv ; either 1 = v ejv and 2 = v ejv (as in the gure) or 2 = v ejv and 1 = v ejv . (The scale labels are not shown).

Notation 4.5. Given a tree , call 0 its root line and consider the internal lines 1 , . . . , p L() on scale 1 (if any) such that one has n 0 for all P (0 , i ), i = 1, . . . , p; we shall say that 1 , . . . , p are the lines on scale 1 which are closest to the root of . For each such line i , call i = i . Then we the subgraph with set of nodes and set of lines call pruned tree
p p

) = N () \ N (
i=1

N (i ),

) = L() \ L(
i=1

L(i ),

respectively. is a tree, except that, with respect to the constraints listed above, one has sv = 1 By construction, ) except whenever kv = 0; moreover one has = ej (and hence n 0) for all internal lines L( possibly the root line. Constraint 4.6. The modes of the end nodes and the momenta of the lines are related as follows: if = v one has the conservation law =
w E ( ) w v

w
w V0 ( ) w v

w ejw =
) w E ( w v

w.

Note that by Constraint 4.6 one has = v if v E (), and = 1 + . . . + sv if v V (), kv 1, and 1 , . . . , sv are the lines entering v ; see Figure 5. Moreover for any line L() one has )|. | | |E ( 1 1 j1 j j kv v 1 2 2 j2 2 j 3 3 3 3 sv sv sv jsv (b) ej j j j 0 v 2 1 j

(a)

Figure 5: Conservation law: (a) v with kv = sv 1 1, so that = 1 + . . . + sv , (b) v with sv = 2 and kv = 0.

(The scale labels are not shown).

Remark 4.7. In the following we shall repeatedly consider the operation of changing the sign label of the nodes. Of course this changes produces the change of other labels, consistently with the constraints mentioned above: for instance, if we change the label v of an end node v into v , then also v is changed into v ; if we change the sign labels of all the end nodes, then also the momenta of all the lines are changed, according to the conservation law (Constraint 4.6); and so on. Two unlabelled trees are called equivalent if they can be transformed into each other by continuously deforming the lines in such a way that they do not cross each other. We shall call equivalent two trees if the same happens in such a way that all labels match. Notation 4.8. We denote by Tk j, the set of inequivalent trees of order k with tree component j and tree momentum , that is, such that the component label and the momentum of the root line are j and , k respectively. Finally for n 1 dene Tk j, (n) the set of trees Tj, such that n n for all L( ).
+ Remark 4.9. For Tk j, , by writing = (1 , . . . , d ), one has i = |Ei ( )| |Ei ( )| for i = 1, . . . , d. )| + 1 1, and |E ( )| = |E ( )| for all j = j . In particular for = ej , one has |Ej ( )| = |Ej ( j j

Lemma 4.10. The number of unlabelled trees with N nodes is bounded by 4N . If k () = k then |E ()| E0 k and |V ()| V0 k , for suitable positive constants E0 and V0 . Proof. The bound |V ()| |E ()| 1 is easily proved by induction using that sv 2 for all v V (). So it is enough to bound |E ()|. The denition of order and Remark 4.1 yield |E ()| = 1 + k () + |V0 ()|, and the bound |V0 ()| 2k () 1 immediately follows by induction on the order of the tree, simply using that sv 2 for v V (). Thus, the assertions are proved with E0 = V0 = 3.

10

4.4

Tree expansion
(k ) (k )

Now we shall see how to associate with each tree Tk j, a contribution to the coecients xj, and j of the power series in (4.1). For all j = 1, . . . , d set c+ j = cj and cj = cj . We associate with each end node v E ( ) a node factor
v Fv := c jv ,

(4.2)

where the coecients fj,s1 ,...,sd are dened in (2.2). Let be a non-decreasing C function dened in R+ , such that (see Figure 6) (u) = 1, for u 7/8, 0, for u 5/8, (4.4)

and with each internal node v V () a node factor sv,1 ! . . . sv,d ! fjv ,sv,1 ,...,sv,d , sv ! Fv := 1 v , 2cjv

kv 1, (4.3) kv = 0,

and set (u) := 1 (u). For all n Z+ dene n (u) := (2n u) and n (u) := (2n u), and set (see Figure 6) n (u) = n1 (u) n (u), (4.5) where 1 (u) = 1. Note that n1 (u)n (u) = n (u), and hence {n (u)}nZ+ is a partition of unity. (u) n (u)

5 8

7 8

2n1 2n

2n+1

Figure 6: The functions and n .

[n ]

Remark 4.11. The number of scale labels which can be associated with a line in such a way that G = 0 is at most 2. In particular, given a line with momentum = and scale n = n, such that n (j ( )) = 0, then (see Figure 6) 5 n 7 2 j ( ) 2(n1) 2(n1) . 8 8 and if n (j ( ))n+1 (j ( )) = 0, then 2(n+1) 5 n 7 2 j ( ) 2n . 8 8 11 (4.7)

We associate with each line a propagator G := Gj ( ), where n (j (u)) , n 0, [n] 2 u 2 j Gj (u) := 1, n = 1.

(4.6)

(4.8)

We dene V () :=
L ( )

G
v N ( )

Fv ,

(4.9)

and call V () the value of the tree .

k Remark 4.12. The number of trees Tk j, with V ( ) = 0 is bounded proportionally to C , for some positive constant C . This immediately follows from Lemma 4.10 and the observation that the number of trees obtained from a given unlabelled tree by assigning the labels to the nodes and the lines is also bounded by a constant to the power k (use Remark 4.11 to bound the number of allowed scale labels).

Remark 4.13. In any tree there is at least one end node with node factor factor c j for each internal node v with kv = 0, v = and jv = j (this is easily proved by induction on the order of the pruned tree): the node factors 1/2c j do not introduce any singularity at cj = 0. Therefore for any tree the corresponding value V () is well dened because both propagators and node factors are nite quantities. Remark 4.12 implies that also V ()
Tk j,

is well dened for all k N, all j {1, . . . , d}, and all Zd .

k Lemma 4.14. For all k N, all j = 1, . . . , d, and any Tk j,ej , there exists Tj,ej such that c j V ( ) = cj V ( ). The tree is obtained from by changing the sign labels of all the nodes v N0 ( ).

Proof. The proof is by induction on the order of the tree. For any tree Tk j,ej consider the tree obtained from by replacing all the labels of all nodes v N ( ) with v , so that the Tk v 0 j,ej mode labels v are replaced with v and the momenta with (see Remark 4.7). Call 1 , . . . , p the lines on scale 1 (if any) closest to the root of , and for i = 1, . . . , p denote by vi the node i enters and i = i (recall (2) in Notation 4.2). As an eect of the change of the sign labels, each tree i is replaced with a tree i such that cj V (i ) = c jvi V (i ), by the inductive hypothesis. Thus, for each node vi ) nor the vi the quantity Fvi V (i ) is not changed. Moreover, neither the propagators of the lines L( v node factors corresponding to the internal nodes v V ( ) with kv = 0 change, while the node factors c jv + ) are changed into c v . On the other hand one has |E ( )| = |E ( )| for all i = j , of the nodes v E ( jv i i + + whereas |Ej ( )| = |Ej ( )| + 1 and |Ej ( )| + 1 = |Ej ( )|. Therefore one obtains cj V () = c j V ( ), and the assertion follows. For k N, j {1, . . . , d}, and {}, dene j, =
(k )

1 c j

V ().
Tk j,e
j

Lemma 4.15. For all k N and all j = 1, . . . , d one has j,+ = j, . Proof. Lemma 4.14 implies c j
Tk j,ej

(k )

(k )

V () = c+ j
Tk j,e
j

V ()
(k )

for all k N and all j = 1, . . . , d, so that the assertion follows from the denition of j, .

12

Lemma 4.16. Equations (2.7) formally hold, i.e., they hold to all perturbation orders, provided that for all k N and j = 1, . . . , d we set formally

xj,

=
k=1

k xj, , k j ,
k=1 (k )

(k )

xj, =
(k )

(k )

V ()
Tk j,

Zd \ {ej } ,

xj,ej = 0 ,

(k )

(4.10)

1 cj

V ().
Tk j,ej

(4.11)

Proof. The proof is a direct check.

(k ) ) contains at least Remark 4.17. In j , dened as (4.11), there is no singularity in cj = 0 because V ( + one factor cj = cj by Remark 4.9.

In the light of Lemma 4.16 one can wonder why the denition of the propagators for = ej is so involved; as a matter of fact one could dene G = 1 2. ( )2 j

However, since n0 n (u) 1, the two denitions are equivalent. We use the denition (4.6) so that we can immediately identify the factors O(2n ) which could prevent the convergence of the power series (4.1). In what follows we shall make this idea more precise.

4.5

Clusters

A cluster T on scale n is a maximal set of nodes and lines connecting them such that all the lines have scales n n and there is at least one line with scale n; see Figure 7. The lines entering the cluster T and the line coming out from it (unique if existing at all) are called the external lines of the cluster T . We call V (T ), E (T ), and L(T ) the set of internal nodes, of end nodes, and of lines of T , respectively; note that the external lines of T do not belong to L(T ). Dene also Ej (T ) as the set of end nodes v E (T ) such that v = and jv = j . By setting k (T ) :=
v V (T )

kv ,

we say that the cluster T has order k if k (T ) = k .

4.6

Self-energy clusters

We call self-energy cluster any cluster T such that (see Figure 8) (1) T has only one entering line and one exiting line, } 2 for any L(T ), (2) one has n min{nT , n T ). | 2 and jT ( T ) = j ( (3) one has | T T T
T

Notation 4.18. For any self-energy cluster T we denote by T and T the exiting and the entering line of T respectively. We call PT the path of lines L(T ) connecting T to T , i.e., PT = P (T , T ) (recall (1) in Notation 4.2), and set nT = min{nT , nT }.

13

1 1 5 2 5 3 1 (a)

1 1 1 1 1 3 1 1 1 (b)

1 2

1 1 5 1 3

Figure 7: Example of tree and the corresponding clusters: once the scale labels have been assigned to the lines of the tree as in (a), one obtains the cluster structure depicted in (b).

Remark 4.19. Notice that, by Remark 3.3, for any self-energy cluster the label T is uniquely xed by , , . In particular, xed and j such that j ( ) , there are only 2d 1 the labels jT , T , j T T T momenta = such that | | 2 and j ( ) = j ( ) for some j and , depending on . All the other with small divisor equal to j ( ) are far away from , according to Lemma 3.1.

1 2 1 T 2

3 1 4 5

1 6

T
Figure 8: Example of self-energy cluster: consider the cluster T on scale 3 in Figure 7, and suppose that the mode labels of the end nodes are such that |1 + 2 + 3 + 4 + 5 + 6 | 2 and j ( T ) = j ( ). Then T is a self-energy
T

cluster with external lines T (entering line) and T (exiting line). The path PT is such that PT = {}.

We say that a line is a resonant line if it is both the exiting line of a self-energy cluster and the entering line of another self-energy cluster, that is, is resonant if there exist two self-energy clusters T1 and T2 such that = T1 = T2 ; see Figure 9. Remark 4.20. The notion of self-energy cluster was rst introduced by Eliasson, in the context of KAM 14

theorem, in [8], where it was called resonance. We prefer the term self-energy cluster to stress further the analogy with quantum eld theory.

T1 T2

Figure 9: Example of resonant line: is resonant if both T1 and T2 are self-energy clusters.

The notion of equivalence given for trees can be extended in the obvious way to self-energy clusters.
Notation 4.21. We denote by Rk j,,j , ( , n) the set of inequivalent self-energy clusters T on scale = j and = . By denition of cluster for n of order k , such that = , jT = j , T = , j T T T k T Rj,,j , ( , n) one must have n nT 2. For j = j and = dene also Ek j,,j, ( , n) the set of self-energy clusters T Rk j,,j, ( , n) such that (1) T enters the same node v which T exits and k (2) kv = 0. We call vT such a special node and set Rj,,j, ( , n) = Rk j,,j, ( , n) \ Ej,,j, ( , n); see Figure 10. k

ej

j n

vT

j n

Figure 10: A self-energy cluster in Ek j,,j, (n); T contains at least one line on scale n and n such that min{n , n } n +2.

Notation 4.22. For any T Ek j,,j, ( , n) we call T the tree which has as root line the line L(T ) entering vT (one can imagine to obtain T from T by removing the node vT ); see Figure 11. Note that T T k j,ej (n).

Notation 4.23. Consider a self-energy cluster T such that n 0 for all lines PT . Call 1 , . . . , p L(T ) the internal lines on scale 1 (if any) which are closest to the exiting line of T , that is, such that n 0 for all lines P (T , i ), i = 1, . . . , p. For each line i set i = i . Then the pruned self-energy is the subgraph with set of nodes and set of lines cluster T
p p

) = N (T ) \ N (T
i=1

N (i ),

) = L(T ) \ L(T
i=1

L(i ),

respectively. 15

 ej ej ej j n j n
Figure 11: An example of self-energy cluster T Ek j,,j, (n) and the corresponding tree T . (Only the mode labels of the end nodes are shown in T and T )
+ Remark 4.24. For T Rk j,,j , ( , n) such that n 0 for all PT , one has |Ei (T )| = |Ei (T )| )| + 1 and |E (T )| = |E (T )| + 1; if j = j , = and for all i = j, j . If j = j then |E (T )| = |E (T j j j j k (T )| = |Ej (T )|, while if j = j and = then |Ej (T )| = |Ej (T )| +2. T Rj,,j, ( , n) then |Ej k Finally, for any T Ej,,j, ( , n) one has |Ej (T )| = |Ej (T )| + 1 1.

ej ej

ej T =

ej

ej ej

ej

We shall dene V (T, ) := T

L (T )

G
v N (T )

Fv ,

(4.12)

where V (T, ) will be called the value of the self-energy cluster T . T The value V (T, ) depends on through the propagators of the lines PT . T T Remark 4.25. The value of a self-energy cluster T Ek j,,j, (u, n) does not depend on u so that we shall write 1 V (T, u) = V (T ) = V (T ). 2 cj We dene also for future convenience Mj,,j , ( , n) :=
(k ) (k ) (k ) (k )

V (T, ),
(k ) (k )

(4.13)

T Rk ( ,n) j,,j ,

Note that Mj,,j, ( , n) = Mj,,j, (n) + M j,,j, ( , n), where Mj,,j, (n) and M j,,j, ( , n)
are dened as in (4.13) but for the sum restricted to the set Ek j,,j, ( , n) and Rj,,j, ( , n) respectively. k

Remark 4.26. Both the quantities Mj,,j , ( , n) and the coecients xj, and j are well dened to all orders because the number of terms which one sums over is nite (by the same argument in Remark 4.12). At least formally, we can dene

(k )

(k )

(k )

Mj,,j , ( ) =
k=1

k
n1

Mj,,j , ( , n).

(k )

We dene the depth D(T ) of a self-energy cluster T recursively as follows: we set D(T ) = 1 if there is no self-energy cluster containing T , and set D(T ) = D(T ) + 1 if T is contained inside a self-energy 16

cluster T and no other self-energy clusters inside T (if any) contain T . We denote by SD () the set of self-energy clusters of depth D in , and by SD (, T ) the set of self-energy clusters of depth D in contained inside T . Notation 4.27. Call = \ S1 () the subgraph of formed by the set of nodes and lines of which are outside the set S1 () (the external lines of the self-energy clusters T S1 () being included in ), and, = T \ SD+1 (, T ) the subgraph of T formed by the set of nodes and analogously, for T SD () call T ), E (T ), and L(T ) the set of internal lines of T which are outside the set SD+1 (, T ). We denote by V (T nodes, of end nodes, and of lines of T , and by k (T ) the order of T , that is, the sum of the labels kv of all ). the internal nodes v V (T Lemma 4.28. Given a line L(), if T is the self-energy cluster with largest depth containing (if . Then any), PT and there is no line PT preceding with n = 1, one can write = 0 + T 0 one has | 0 | E k ( T ) , for a suitable positive constant E , if k ( T ) 1 , and | | 2 if k ( T ) = 0 . 1 1 Proof. We rst prove that for any tree , if we denote by 0 its root line, one has | 0 | E1 k ( ) 2, if 0 does not exit a self-energy cluster, E1 k ( ), if 0 exits a self-energy cluster, (4.14)

for a suitable constant E1 4. The proof is by induction on the order of the tree . If k () = 1 (and hence = ) then the only internal line of is 0 and | 0 | 2, so that the assertion trivially holds provided E1 4. If k () > 1 let v0 be the node which 0 exits. If v0 is not contained inside a self-energy cluster let 1 , . . . , m , m 0, be the internal lines entering v0 and i = i for all i = 1, . . . , m. Finally 1 ) + . . . + k ( m ). ) = kv0 + k ( let m+1 , . . . , m+m be the end-lines entering v0 . By denition we have k ( If kv0 > 0, we have 0 = 1 + . . . + m+m . This implies in turn 1 ) + . . . + k ( m ) + m E1 (k ( ) m m + 1) + m | 0 | | 1 | + . . . + | m | + m E1 k ( The assertion follows for E1 4 by the inductive hypothesis (the worst possible case is m = 0, m = 2). If kv0 = 0 then sv = 2 and m = 0. moreover one of the lines, say 1 , is on scale n = 1 while for the other line one has 0 = 2 . Once more the bound follows from the inductive hypothesis since 2 ) E1 (k ( ) 1). | 2 | E1 k ( Finally, if v0 is contained inside a self-energy cluster, then 0 exits a self-energy cluster T1 . There will be p self-energy clusters T1 , . . . , Tp , p 1, such that the exiting line of Ti is the entering line of Ti1 , for i = 2, . . . , p, while the entering line of Tp does not exit any self-energy cluster. By ) = k ( ). Then, by the inductive hypothesis, one nds Lemma 3.4, one has | 0 | 2 and k ( ). ) 2 = E1 k ( | 0 | 2 + E1 k ( Now for and T as in the statement we prove, by induction on the order of the self-energy cluster, the bound E1 k (T ) 2, if k (T ) 1, | 0 (4.15) | 2 if k (T ) = 0, where T is the set of nodes and lines of T which precede . The bound is trivially satised when k (T ) = 0. Otherwise let v be the node in V (T ) between and T which is closest to . If kv = 0 the bound follows trivially by using the bound (4.14). If kv 1 call 1 , . . . , m , m 0, the internal lines entering v which are not along the path PT , and m+1 , . . . , m+m the end lines entering v ; one has m + m 1. There is a further line 0 PT entering v such that 0 = 0 ; see Figure 12. Using also Lemma 3.4 one 0 + T 0 0 2 one has k (T has | | 2 + | 0 | + | 1 | + . . . + | m | + m . As n0 n 0 ) 1 and hence, by (4.14) T 17

and the inductive hypothesis, one has

| 0 | 2 + E1 k (T0 ) 2 + E1 k (1 ) + . . . k (m ) + m , where i = i for all i = 1, . . . , m. Thus, since k (T 0 ) + k (1 ) + . . . + k (m ) + (m + m ) = k (T ) and m + m 1, one nds + m E2 k (T | 0 ) 2, | E1 k (T ) m m

provided E1 4. Therefore, the assertion follows with, say, E1 = 4.

1 T v T1 Tp 0

2 2 1 3 2

T
Figure 12:
The self-energy cluster T considered in the proof of Lemma 4.28, with m = 2, m = 3, and a chain of p self-energy clusters between and v (one has p 0, and = v if p = 0).

Notation 4.29. Given a tree and a line L(), call = () the subgraph formed by the set of nodes and lines which do not precede ; see gure. Let us call the set of nodes and lines of which are outside any self-energy cluster contained inside .

Figure 13: The set = () and the subtree determined by the line L(). If is the root line then = .

Lemma 4.30. Given a tree let 0 and be the root line and an arbitrary internal line preceding 0 . If ), for a suitable positive constant E2 . k ( ) 1 one has | 0 | E2 k ( Proof. We prove by induction on the order of the bound | 0 | E2 k ( ) 2, if 0 does not exit a self-energy cluster, E2 k ( ), if 0 exits a self-energy cluster. 18 (4.16)

We mimic the proof of (4.14) in Lemma 4.28. The case k ( ) = 1 is trivial provided E2 3, so let us consider k ( ) > 1 and call v0 the node which 0 exits. If v0 is not contained inside a selfenergy cluster and kv0 1 then 0 = 1 + . . . + m+m , where 1 , . . . , m are the internal lines entering v0 , with (say) m P (0 , ) {}, and m+1 , . . . , m+m are the end lines entering v0 . Hence 1 ) + . . . + k ( m1 ) + k ( m ), where i = i and m = (m ) (m = if m = ). Thus, k ( ) = kv0 + k ( the assertion follows by (4.14) and the inductive hypothesis. If v0 is not contained inside a self-energy cluster and kv0 = 0 then two lines 1 and 2 enter v0 , and one of them, say 1 , is such that | 1 | = 1. If = 2 the result is trivial. If 2 P (0 , ) the bound follows once more from the inductive hypothesis. If = 1 one has 2 ) + 1 E2 k ( ) 2 , | 0 | | 0 | + 1 E1 k ( where 2 = 2 , provided E2 E1 + 3, if E1 is the constant dened in Lemma 4.28. If 1 P (0 , ) denote by 1 the line on scale 1 along the path {1 } P (1 , ) which is closest to . Again call 2 = 2 and J1 the subgraph formed by the set of nodes and lines preceding 1 (with 1 included) but not ; dene also 1 as the tree obtained from J1 by (1) reverting the arrows of all lines along { 1 , } P (1 , ), (2) replacing 1 with an end line carrying the same sign and component labels as 1 , and (3) replacing all the labels v , v N0 (J1 ) with v . One has, by using also (4.14), 1 ) + E1 k ( 2 ) E2 k ( ) 2 , | 0 | | 0 | + | | E1 k ( provided E2 E1 + 2 so that the bound follows once more. Finally, if v0 is contained inside a self-energy cluster, then 0 exits a self-energy cluster T1 . There will be p self-energy clusters T1 , . . . , Tp , p 1, such that the exiting line of Ti is the entering line of Ti1 , for i = 2, . . . , p, while the entering line of Tp ) = k ( ), where does not exit any self-energy cluster. By Lemma 3.4, one has | 0 | 2 and k ( ). = ( ). Then, the inductive hypothesis yields | 0 | 2 + | | 2 + E2 k ( ) 2 = E2 k ( Therefore the assertion follows with, say, E2 = E1 + 3 (and hence E2 = 7 if E1 = 4). Remark 4.31. Lemma 4.28 will be used in Section 6 to control the change of the momenta as an eect of the regularisation procedure (to be dened). Furthermore, both Lemmas 4.28 and 4.30 will be used in Section 8 to show that the resonant lines which are not regularised cannot accumulate too much.

Dimensional bounds

In this section we discuss how to prove that the series (4.10) and (4.11) converge if the resonant lines are excluded. We shall see in the following sections how to take into account the presence of the resonant lines. Call Nn () the number of non-resonant lines L() such that n n, and Nn (T ) the number of non-resonant lines L(T ) such that n n. The analyticity assumption on f yields that one has |Fv | sv +kv for a suitable positive constant . Lemma 5.1. Assume that 2(n +2) j ( ) 2(n 2) for all trees and all lines L(). Then there exists a positive constant c such that for any tree one has Nn () c 2n/ k (). Proof. We prove that Nn () max{0, c 2n/ k () 2} by induction on the order of . 1 (n2)/ 1. First of all note that for a tree to have a line on scale n n one needs k () kn = E0 2 , as it follows from the Diophantine condition (3.2a) and Lemma 4.10. Hence the bound is trivially true 19 v V () \ V0 (), (5.1)

for k < kn . 2. For k () kn , let 0 be the root line of and set = 0 and j = j0 . If n0 < n the assertion follows from the inductive hypothesis. If n0 n, call 1 , . . . , m the lines with scale n 1 which are closest to 0 (that is, such that n n 2 for all p = 1, . . . , m and all lines P (1 , p )). The case m = 0 is trivial. If m 2 the bound follows once more from the inductive hypothesis. 3. If m = 1, then 1 is the only entering line of a cluster T . Set = 1 , j = j1 and n = n1 . By hypothesis one has j ( ) 2(n2) and j ( ) 2(n3) , so that, by Lemma 3.2, either | | > 2(n5)/ or | | 2 and j ( ) = j ( ). In the rst case, since =
w E (T )

w
w V (T ) kw =0

w ejw =
) w E (T

w,

the same argument used to prove Lemma 4.10 yields | | |E (T )| E0 k (T ), and hence k (T ) 1 (n5)/ E0 2 . Thus, if 1 = 1 , one has k () = k (T ) + k (1 ), so that Nn () = 1 + Nn (1 ) c 2n/ k (1 ) 1 c 2n/ k () c 2n/ k (T ) 1 c 2n/ k () 2, provided c E0 25/ . 4. If instead | | 2 and j ( ) = j ( ), then the only way for T not to be a self-energy cluster is that n1 = n0 1 = n 1 and there is at least a line T with n = n 2. But then j ( ) = j ( ) so that | | > 2(n6)/ and we can reason as in the previous case provided c E0 26/ . Otherwise T is a self-energy cluster and 1 can be either resonant or not-resonant. Call 1 , . . . , m the lines with scale n 1 which are closest to 1 . Once more the cases m = 0 and m 2 are trivial. ) , then Nn () = 1 + Nn (1 5. If m = 1, then 1 is the only entering line of a cluster T . If 1 = 1 if 1 is resonant and Nn () 2 + Nn (1 ) if 1 is non-resonant. Consider rst the case of 1 being non-resonant. Set = , j = j and n = n . By reasoning as before we nd that one has 1 1 1 (n5)/ either | | > 2 or | | 2 and j ( ) = j ( ). If | | > 2(n5)/ then 1 (n5)/ k (T ) E0 2 ; thus, by using that k () = k (T ) + k (T ) + k (1 ), we obtain
Nn () 2 + Nn (1 ) c 2n/ k () c 2n/ k (T ) c 2n/ k (T ) c 2n/ k () c 2n/ k (T ) c 2n/ k () 2,

provided c 2E0 25/ . 6. Otherwise one has | | 2, | | 2, and j ( ) = j ( ) = j ( ). Since we are assuming 1 to be non-resonant then, T is not a self-energy cluster. But then there is at least a line T with n = n 2 and we can reason as in item 4. 7. So we are left with the case in which 1 is resonant and hence T is a self-energy cluster. Let 1 be the is either resonant or non-resonant. If it is non-resonant we repeat the entering line of T . Once more 1 same argument as done before for 1 . If it is resonant, we iterate the construction, and so on. Therefore we proceed until either we nd a non-resonant line on scale n, for which we can reason as before, or we reach a tree of order so small that it cannot contain any line on scale n (i.e., k ( ) < kn ). 8. Therefore the assertion follows with, say, c = 2E0 26/ . Remark 5.2. One can wonder why in Lemma 5.1 did we assume 2(n +2) j ( ) 2(n 2) when Remark 4.11 assures the stronger condition 2(n +1) j ( ) 2(n 1) . The reason is that later on we shall need to slightly change the momenta of the lines, in such a way that the scales in general no longer satisfy the condition (4.7) noted in Remark 4.11. However the condition assumed for proving Lemma 5.1 will still be satised.

20

For any tree we call LR () and LNR () the sets of resonant lines and of non-resonant line, respectively, in L(). Then we can write V () =
L R ( )

G V NR (),

V NR () :=
LNR ( )

G
v N ( )

Fv ,

(5.2)

where each propagator G can be bounded as C0 2n , for some constant C0 . Lemma 5.3. For all tree with k () = k one has | V NR ()| C k 3k (c), where (c) := max{|c1 |, . . . , |cd |, 1} and C is a suitable positive constant. Proof. One has
k 3k |V NR ()| C0 (c)k L N R ( ) k 3k C0 (c)k exp c log 2 k n=1 k 3k (c)k 2n C0 n=0

2nNn ()

2n/ n .

The last sum converges: this is enough to prove the lemma. So far the only bound that we have on the propagators of the resonant lines is |G | 1/j j ( ) C0 2n . What we need is to obtain a gain factor proportional to 2n for each resonant line with n 1. Lemma 5.4. Given such that V () = 0, let L() be a resonant line and let T be the self-energy cluster of largest depth containing (if any). Then there is at least one non-resonant line in T on scale n 1 . Proof. Set n = n . There are in general p 2 self-energy clusters T1 , . . . , Tp , contained inside T , connected by resonant lines 1 , . . . , p1 , and is one of such lines, while the entering line p of Tp and the exiting line 0 of T1 are non-resonant. Moreover ( i ) = ( ) for all i = 0, . . . , p, so that all the lines 0 , . . . , p have scales either n, n 1 or n, n + 1, by Remark 4.11. In any case the lines 0 , p must be in T by denition of self-energy cluster.

Renormalisation

Now we shall see how to deal with the resonant lines. In principle, one can have trees containing chains of arbitrarily many self-energy clusters (see Figure 14), and this produces an accumulation of small divisors, and hence a bound proportional to k ! to some positive power for the corresponding values.

T1

T2

T3

Tp

Figure 14: A chain of self-energy clusters.

Let K0 be such that E1 K0 = 28/ . For T Rk j,,j , (u, n), dene the localisation operator L by setting nT / , n 0 PT , V (T, j ), k (T ) K0 2 (6.1) L V (T, u) := 0, otherwise 21

which will be called the localised value of the self-energy cluster T . Dene also R := 1 L , by setting, for T Rk j,,j , (u, n), 1 ) K0 2nT / , n 0 PT , ) d t u V (T, j + t(u j )), k (T ( u j 0 R V (T, u) = (6.2) V (T, u), otherwise so that L Mj,,j , (u, n) = R Mj,,j , (u, n) =
(k ) (k )

L V (T, u), R V (T, u).

(6.3a) (6.3b)

T Rk (u,n) j,,j ,

(u,n) T Rk j,,j ,

We shall call R the regularisation operator and R V (T, u) the regularised value of T . Remark 6.1. If T Ek j,,j, (u, n) the localisation operator acts as V (T ), 0, ) K0 2nT / , k (T ) > K0 2nT / . k (T

L V (T ) =

Remark 6.2. If in a self-energy cluster T there is a line PT such that = ej (and hence n = 1) then L V (T , u) = 0 for all self-energy clusters containing T such that PT . Recall the denition of the sets SD () and SD (, T ) after Remark 4.26. For any tree we can write its value as Fv , (6.4) G ) V (T, V () = T
T S1 ( ) L( \S1 ( )) v N ( \S1 ( ))

and, recursively, for any self-energy cluster T of depth D we have V (T, )= T

T SD+1 (,T )

V (T , ) T
L(T \SD+1 (,T ))

G
v N (T \SD+1 (,T ))

Fv .

(6.5)

Then we modify the diagrammatic rules given in Section 4 by assigning a further label OT {R , L }, which will be called the operator label, to each self-energy cluster T . Then, by writing V () according ) to (6.4) and (6.5), one replaces V (T, ) with L V (T, ) if OT = L and with R V (T, T T T k if OT = R . When considering the regularised value of a self-energy cluster T Rj,,j , (u, n) with ) K0 2nT / and n 0 for all PT , then we have also an interpolation parameter t to consider: k (T we shall denote it by tT to keep trace of the self-energy cluster which it is associated with. We set tT = 1 ) > K0 2nT / or PT containing at least one line for a regularised self-energy cluster T with either k (T with n = 1. We call renormalised trees the trees carrying the further labels OT , associated with the self-energy clusters T of . As an eect of the localisation and regularisation operators the arguments of the propagators of some lines are changed. Remark 6.3. For any self-energy cluster T the localised value L V (T, u) does not depend on the operator labels of the self-energy clusters containing T . 22

Given a self-energy cluster T Rk j,,j , (u, n) such that no line along PT is on scale 1, let be a line such that (1) PT , and (2) T is the self-energy cluster with largest depth containing . If one has OT = R , then the quantity is changed according to the operator labels of all the self-energy clusters T such that (1) T contains T , (2) no line along PT has scale 1, and (3) PT . Call Tp Tp1 . . . T1 such self-energy clusters, with Tp = T . If OTi = R for all i = 1, . . . , p then is replaced with
0 (t ) = 0 + p jp + tp p + p1 jp1 p jp p1

+
i=2

tp . . . ti 0 i + i1 ji1 i ji + tp . . . t1 ( 1 1 j1 ) ,

(6.6)

where we have set t = (t1 , . . . , tp ), Ti = i and tTi = ti for simplicity. Otherwise let Tq be the self-energy cluster of highest depth, among T1 , . . . , Tp1 , with OTq = L (so that OTi = R for i q + 1). In that case, instead of (6.6), one has
0 (t ) = 0 + p jp + tp p + p1 jp1 p jp p1

+
i=q+1

tp . . . ti 0 i + i1 ji1 i ji ,

(6.7)

with the same notations used in (6.6). j j for PT , we can write 0 If OTp = L , since is replaced with 0 + + T T T T as in (6.6) by setting tp = 0. More generally, if we set tT = 0 whenever OT = L , we see that we can always claim that, under the action of the localisation and regularisation operators, the momentum of any line PT is changed to (t ), in such a way that (t ) is given by (6.6). Lemma 6.4. Given such that V () = 0, for all L() one has 4 j ( ) 5 j ( (t )) 6 j ( ). Proof. The proof is by induction on the depth of the self-energy cluster. 1. Consider rst the case that PT , with OT = L . Set n = n , = , = , and j = j . T T T T 0 Then is replaced with j , and, as a consequence, is replaced with (t ) = + j . Dene n such that ( n 1) n +1) , (6.8) 2( j ( 0 + j ) 2
0 0 by the Diophantine condition (3.2b). Therefore where j ( 0 + j ) = | + j j | | | n 1 0 n n 8 )) (E1 K0 ) 2 = 2 2 | | (E1 k (T , and hence n n 7. Since | j | 2n+2 by the inductive hypothesis, one has j ( ) = 0 + j 0 + j j | j |

15 j ( 0 + j ), 16

( n +1) 2n+6 24 | j |. In the same way one can bound because j ( 0 + j ) 2 0 j ( ) | + j j | + | j |, so that we conclude that

17 15 j ( 0 j ( 0 + j ) j ( ) + j ). 16 16

(6.9)

This yields the assertion. 2. Consider now the case that OT = R . In that case (t ) is given by (6.6). Dene n as in (6.8), 23

with = p and j = jp . We want to prove that 7 9 j ( 0 j ( 0 + j ) j ( (t )) + j ). 8 8 (6.10)

for all t = (t1 , . . . , tp ), with ti [0, 1] for i = 1, . . . , p. This immediately implies the assertion because, by using also (6.9), we obtain 9 18 7 14 j ( ) j ( 0 j ( 0 j ( ), + j ) j ( (t )) + j ) 17 8 8 15 and hence 4j ( ) 5j ( (t )) 6j ( ). By the inductive hypothesis and the discussion of the case 1, in (6.8) we have
ni +2 , 0 i + i1 ji1 i ji 2

i = 1, . . . , p,

where ni = ni . Moreover one has ni ni+1 for i = 1, . . . , p 1, so that we obtain

p j ( (t )) j ( 0 + j ) i=1 ( n +1) and n n 7, one nds j ( (t )) (1 23)j ( 0 Since j ( 0 + j ). + j ) 2 3 0 In the same way one has j ( (t )) (1 + 2 )j ( + j ), so that (6.10) follows. n+3 . 2ni +2 j ( 0 + j ) 2

Remark 6.5. Given a renormalised tree , with V () = 0, if a line L() has scale n then n (j ( )(t )) = 0, and hence, by Lemma 6.4, one has 2(n +2) j ( ) 2(n 2) . Therefore, Lemma 5.1 still holds for the renormalised trees without any changes in the proof (see also Remark 5.2). Remark 6.6. Another important consequence of Lemma 6.4 (and of inequality (4.8) in Remark 4.11) is that the number of scale labels which can be associated with each line of a renormalised tree is still at most 2.

Symmetries and identities

Now we shall prove some symmetry properties on the localized value of the self-energy clusters.
Lemma 7.1. If T Ek then j,,j, (u, n) is such that T does not contain any end node v with Fv = cj

there exists T Rj,,j, (u, n) such that 2L V (T ) = L V (T , u).

Proof. If T Ek j,,j, (u, n) one has |Ej (T )| = |Ej (T )| + 1 (see Remark 4.24), so that if |Ej (T )| = 0, ) \ {v0 }, if E (T ) = {v0 }. Consider the then also |Ej (T )| = 1. This means that jv = j for all v E (T j

self-energy cluster T Rj,,j, (u, n) obtained from T by replacing the line exiting v0 with an entering line carrying a momentum such that = u and nT = nT ; see Figure 15. With the exception of v0 , the nodes of T have the same node factors as T ; in particular they have the same combinatorial factors. If we compute the propagators G of L(T ), by setting u = j , then they are the same as the corresponding propagators of T . Finally, as nT = nT , one has L V (T ) = 0 if and only if also L V (T , u) = 0. Thus, by recalling also Remark 4.25, one nds 2L V (T ) = L V (T , u).
For T Ek j,,j, (u, n) let us call F1 (T ) the set of all inequivalent self-energy clusters T Rj,,j, (u, n) obtained from T by replacing a line exiting an end node v Ej (T ) with an entering line carrying a k

24

 ej v0 T ej ej ej v0 j T j

T =

Figure 15: The self-energy cluster T , the tree T , and the self-energy cluster T in the proof of Lemma 7.1.

T1 ej

ej ej ej T2 ej ej ej T3 ej ej

ej ej

ej

Figure 16: The sets F1 (T ) = {T1 , T2 } and F2 (T ) = {T3 } corresponding to the self-energy cluster T in Figure 11.

momentum such that = u and with nT = nT . Call also F2 (T ) the set of all inequivalent self energy clusters T Rk j,,j, (u , n), with u = u 2j , obtained from T by replacing a line exiting an end node v Ej (T ) (if any) with an entering line carrying a momentum such that = u and with nT = nT ; see Figure 16. Lemma 7.2. For all T Ek j,,j, (u, n) one has
2 c j L V (T ) + cj T F
1 (T )

L V (T , u) = c j T F
2 (T )

L V (T , u ),

where u = u 2j and the right hand side is meant as zero if F2 (T ) = . Proof. The case k (T ) > K0 2nT / is trivial so that we consider only the case k (T ) K0 2nT / . By construction any T Ek j,,j, (u, n) is such that T contains at least an end node v such that Fv = cj , k hence |Ej (T )| 1. By Lemma 7.1 either |Ej (T )| 1 or there exists T Rj,,j, (u, n) such that ) = . 2L V (T ) + L V (T , u) = 0. Hence the assertion is proved if E (T
j

So, let us consider the case 1. First of all note that there is a 1-to-1 correspondence between the lines of T and the lines and external lines, respectively, of both T F1 (T ) and T F2 (T ); the same holds for the internal nodes. Moreover the propagators both of any T F1 (T ) and of any T F2 (T ) are equal to the corresponding propagators of T when setting u = j and u = j , respectively. Also the node factors of the internal nodes of all self-energy clusters T F1 (T ) F2 (T ) + are the same as those of T . For T F1 (T ) one has |Ei (T )| = |Ei (T )| for all i = 1, . . . , d, whereas for + T F2 (T ) one has |Ei (T )| = |Ei (T )| for all i = j and |Ej (T )| = |Ej (T )| + 2; thus, one has
v c jv

|Ej (T )|

= c j
) v E (T

v c jv

= c j ) v E (T

v c jv

) v E (T

for all T F1 (T ) and all T F2 (T ). 25

Therefore, if we write 2 c j L V (T ) = V (T ) = A (T )
) v E (T
v c jv ,

(7.1)

where A (T ) depends only on T , then one nds L V (T , u) = A (T )

T F
1 (T )

1 c j

v c jv

rv,j, ,
) v V (T

) v E (T

with the same factor A (T ) as in (7.1). Analogously one has L V (T , u ) = A (T )

T F
2 (T )

1 c j

v c jv ) v E (T ) v V (T

rv,j, ,

again with the same factor A (T ) as in (7.1), so one can write

2 c j V (T ) + cj T F L V (T , u) c j
1 (T )

L V (T , u ) = B (T ) 1 +
2 (T )

(rv,j, rv,j, ) , (7.2)

T F

) v V (T

where B (T ) = A (T )
) v E (T

v c jv .

On the other hand one has

) v V (T

rv,j, = |Ej (T )|,

so that the term in the last parentheses of (7.2) gives 1 + |Ej (T )| |Ej (T )| = 0. Therefore the assertion is proved. For T Rk j,,j , (u, n) with n 0 for all PT , call G1 (T ) the set of self-energy clusters T k Rj,,j , (u, n) obtained from T by exchanging the entering line T with a line exiting an end node v k Ej (T ) (if any). Call also G2 (T ) the set of self-energy clusters T Rj,,j , (u , n), with u = u 2j , obtained from T by (1) replacing the momentum of T with a momentum such that = u , (2) changing the sign label of an end node v Ej (T ) into , and (3) exchanging the lines T and v . ( u, n ), obtained from T by (1) replacing the Finally call G3 (T ) the set of self-energy clusters T Rk j,,j , , (2) replacing all the = e = and entering line with a line exiting a new end node v with j v0 0 v0 T labels v of the nodes v N0 (T ) {v0 } with v and (3) replacing a line exiting an end node v Ej (T ), with the entering line T ; see Figure 17. Again we force nT = nT for all T G1 (T ) G2 (T ) G3 (T ). Lemma 7.3. For all T Rk j,,j , (u, n), with j = j and n 0 for all PT , one has

c j
c j cj

L V (T , u) = c j

L V (T , u )

T G1 (T )

T G2 (T )

L V (T , u) = c j cj

L V (T , u).

T G1 (T )

T G3 (T )

) K0 2nT / . For xed T Rk (u, n), with j = j , let Proof. Again we consider only the case k (T j,,j , k Tj,ej (n) be the tree obtained from T by replacing the entering line T with a line exiting a new 26

1 ej j T T T 2 ej j T T1

1 2 ej ej j T
3

1 ej j T T2 j 3 ej T ej 2

ej j 1 ej

ej

1 ej j T T4 j T 2 3 ej ej v0 j j T T T5

1 ej 2 ej 3
ej j v0 T

j T T3

2 ej 3 ej

Figure 17: A self-energy cluster T and the corresponding sets G1 (T ) = {T, T1 }, G2 (T ) = {T2 , T3 }, and G3 (T ) = {T4 , T5 }.
end node v0 with v0 = and v0 = ej . Note that in particular one has |Ej ( )| = |Ej ( )|. Any T G1 (T ) can be obtained from by replacing a line exiting an end node v Ej ( ) with entering line , with the same labels as , so that T T

c j

L V (T , u) = |Ej ( )| V ( ).

T G1 (T )

On the other hand, any T G2 (T ) can be obtained from by replacing a line exiting an end node v Ej ( ) with entering line T , with labels 2 ej , j , , hence
c j

L V (T , u) = |Ej ()| V (),

T G2 (T )

so that the rst equality is proved. Now, let Tk j,ej (n) be the tree obtained from by replacing all the labels v of the nodes v N0 () with v . Any T G3 (T ) can be obtained from by replacing a line exiting an end node v Ej ( ) with entering line T , carrying the same labels as T . Hence, by Lemma 4.14,
c j cj

L V (T , u) = cj |Ej ( )| V ( ) = c ( )| V ( ) = cj |Ej j cj

L V (T , u),

T G1 (T )

T G3 (T )

which yields the second identity, and hence completes the proof. Lemma 7.4. For all k Z+ , all j, j = 1, . . . , d, and all , {}, one has (i) (k) = (k) (|c1 |2 , . . . , |cd |2 ), i.e., (k) depends on c only through the quantities |c1 |2 , . . . , |cd |2 ; (k ) (k ) (k ) (ii) L Mj,,j , (u, n) = c j cj Mj,j (n), where Mj,j (n) does not depend on the indices , . 27

Proof. One works on the single trees contributing to L Mj,,j , (u, n). Then the proof follows from Lemma 4.14 and the results above. Remark 7.5. Note that Lemma 7.4 could be reformulated as L Mj,,j , (u, n) = c c j L Mj,,j, (n),
j

(k )

(k )

(k )

with Mj,,j, (n) dened after (4.13). We omit the proof of the identity, since it will not be used.

(k )

Cancellations and bounds

We have seen in Section 5 that, as far as resonant lines are not considered, no problems arise in obtaining good bounds, i.e., bounds on the tree values of order k proportional to some constant to the power k (see Lemma 5.3). For the same bound to hold for all tree values we need a gain factor proportional to 2n for each resonant line on scale n 1. Let us consider a tree , and write its value as in (6.4). Let be a resonant line. Then exits a self1 energy cluster T2 and enters a self-energy cluster T1 ; see Figure 9. By construction T1 Rk j1 ,1 ,j , (
2 , n1 ) and T2 Rk , ( , n2 ), for suitable values of the labels, with the constraint j1 = j2 = j2 ,2 ,j2 T1 T2 2 j and 1 = 2 = . If OT1 = OT2 = L , we consider also all trees obtained from by replacing T1 and T2 with other clusters k2 1 T1 Rk , ( , n1 ) and T2 Rj , ,j , ( , n2 ), respectively, with OT = OT = L . In j1 ,1 ,j1 1 2 T1 T2 2 2 2 1 2 this way [n ] L V (T1 , ) Gj ( ) L V (T2 , T2 ) T
1 1 1

is replaced with
2 1 ( , n1 ) Gj ( ) L Mj , L Mj1 , ( , n2 ). 1 ,j , T T ,j , 1 2 2 2

(k )

[n ]

(k )

(8.1)

Then consider also all trees in which the factor (8.1) is replaced with
2 1 , n2 ), , n1 ) Gj ( L Mj1 , ( ) L Mj , ,j ,2 ( 1 ,j , T T 1 2 2

(k )

[n ]

(k )

(8.2)

with such that j = + j ; see Figure 18. Because of Lemmas 7.2 and 7.3 the sum of the two contributions (8.1) and (8.2) gives
1 2 L Mj1 , ( , n1 ) Gj ( ) + Gj ( , n2 ), ) L Mj , ,j , ( 1 ,j , T T 1 2 2 2

(k )

[n ]

[n ]

(k )

where Gj ( ) + Gj ( ) =
[n ] [n ]

1 1 n (j ( )) + ( j ) + j j 2n (j ( )) = , ( + j )( j )

(8.3)

2 n ) we were looking and hence |Gj ( ) + Gj ( )| 2j . This provides the gain factor O(2 n for, with respect to the original bound C0 2 on the propagator G . nT / we can extract a factor C k(T1 ) from V (T1 , ) (C is If OT1 = R then if k (T 1 ) > K0 2 1 T

[n ]

[n ]

the constant appearing in Lemma 5.3), and, after writing C k(T1 ) = C 2k(T1 ) C k(T1 ) , use that C k(T1 ) nT / C K0 2 1 const.2nT1 in order to obtain a gain factor O(2n ). 28

 T1 Rj 1,

k ( ,n1 ), 1 1 ,j , k T2 Rj 2, ,j , ( 2 ,n2 ) 2 2

T1 Rj 1,

k ( ,n1 ), 1 1 ,j , k2 T2 Rj , ,j , ( 2 ,n2 ) 2 2

111111111 000000000 00000000 11111111 000000000 111111111 00000000 j 11111111 2 j2 2 000000000 111111111 00000000 11111111 000000000 111111111 00000000 11111111 1111 0000 11111 00000 11111 00000 000000000 111111111 00000000 11111111 000000000 111111111 00000000 11111111 000000000 111111111 00000000 11111111 T2 T 000000000 111111111 00000000 11111111 1 L L 000000000 111111111 00000000 11111111 000000000 111111111 00000000 11111111 000000000 00000000 11111111 1 j1 1 111111111 2 j2 2 000000000 111111111 00000000 j 11111111 000000000 00000 111111111 00000000 00000 11111111 1111 0000 11111 11111 000000000 111111111 00000000 11111111 000000000 111111111 00000000 11111111 000000000 111111111 00000000 11111111 T2 T 000000000 111111111 00000000 11111111 1
1 j1 1

Figure 18: Graphical representation of the cancellation mechanism discussed in the text: = 2 ej . If we sum the two contributions we obtain a gain factor O (2n ).

nT / and n 0 for all PT , we obtain a gain factor proportional to 2n because If k (T 1 ) K0 2 1 of the rst line of (6.2). Of course whenever one has such a case, then one has a derivative acting on V (T, u) see (6.2). Therefore one needs to control derivatives like

u V (T, u) =
PT

u G
L(T )\{}

G
v N (T )

Fv ,

(8.4)

where u G =

n ( ( )) u n ( ( )) 2 2 2 )2 . 2 ( ) j (( )2 j

(8.5)

The derived propagator (8.5) can be easily bounded by |u G | C1 22n , (8.6)

for some positive constant C1 . In principle, given a line , one could have one derivative of G for each self-energy cluster containing . This should be a problem, because in a tree of order k , a propagator G could be derived up to O(k ) times, and no bound proportional to some constant to the power k can be expected to hold to order k . In fact, it happens that no propagator has to be derived more than once. This can be seen by reasoning as follows. Let T be a self-energy cluster of depth D(T ) = 1. If OT = R then a gain factor O(2nT ) is obtained. When writing u V (T, u) according to (8.4) one obtains |PT | terms, one for each line PT . Then we can bound the derivative of G according to (8.6). By collecting together the gain factor and the bound n (8.6) we obtain 22n 2 T . We can interpret such a bound by saying that, at the cost of replacing the n bound 2n of the propagator G with its square 22n , we have a gain factor 2 T for the self-energy cluster T . Suppose that is contained inside other self-energy cluster besides T , say Tp Tp1 . . . T1 (hence Tp is that with largest depth, and D(Tp ) = p + 1). Then, when taking the contribution to (8.4) with the derivative u acting on the propagator G , we consider together the labels OTi = R and OTi = L for all i = 1, . . . , p (in other words we do not distinguish between localised and regularised values for such self-energy clusters), because we do not want to produce further derivatives on the propagator G .

29

Of course we have obtained no gain factor corresponding to the entering lines of the self-energy clusters T1 , . . . , Tp , and all these lines can be resonant lines. So, eventually we shall have to keep track of this. Then we can iterate the procedure. If the self-energy cluster T does not contain any line whose propagator is derived, we split its value into the sum of the localised value plus the regularised value. On the contrary, if a line along the path PT of T is derived we do not separate the localised value of T from its regularised value. Note that, if T is contained inside a regularised self-energy cluster, then both and in (8.1) and (8.2) must be replaced with (t ) and (t ), respectively, but still (t ) j = (t ) + j , so that the cancellation (8.3) still holds. Let us call ghost line a resonant line such that (1) is along the path PT of a regularised self-energy cluster T and either (2a) enters or exits a self-energy cluster T T containing a line whose propagator is derived or (2b) the propagator of is derived. Then, eventually one obtains a gain 2n for all resonant lines , except for the ghost lines. In other words we can say that there is an overall factor proportional to 2n , (8.7) 2n
L R ( ) L G ( )

where LG () is the set of ghost lines. Indeed, in case (2a) there is no gain corresponding to the line , so that we can insert a good factor 2n provided we allow also a compensating bad factor 2n . In case (2b) one can reason as follows. Call (with some abuse of notation) T1 and T2 the self-energy clusters which enters and exits, respectively. If OT1 = OT2 = L , we consider L V (T1 , ) u Gj ( (t )) L V (T2 , T2 ), T
1

[n ]

and, by summing over all possible self-energy clusters as done in (8.1), we obtain
1 2 ( , n1 ) u Gj ( (t )) L Mj , L Mj1 , ( , n2 ); 1 ,j , T T ,j , 1 2 2 2

(k )

[n ]

(k )

then we sum this contribution with

1 2 L Mj1 , ( , n1 ) u Gj ( , n2 ), (t )) L Mj , ,j ,2 ( 1 ,j , T T 1 2 2

(k )

[n ]

(k )

where

= 2 ej ; again we can use Lemmas 7.2 and 7.3 to obtain

(k ) [n ] [n ] (k )
1 2 2 2

2 1 , n2 ), L Mj1 , ( , n1 ) u Gj ( (t )) + u Gj ( (t )) L Mj , ,j , ( 1 ,j , T T

where u Gj ( (t )) + u Gj ( (t )) =
[n ] [n ]

2u n ( ( (t ))) ( (t ) + j )( (t ) j ) 4( (t ) j )n ( ( (t ))) , 2 ( (t ) + j )2 ( (t ) j )

so that we have not only the gain factor 2n due to the cancellation, but also a factor 2n because of the term u n ( ( )). A trivial but important remark is that all the ghost lines contained inside the same self-energy cluster have dierent scales: in particular there is at most one ghost line on a given scale n. Therefore we can rely upon Lemma 5.4 and Lemma 6.4, to ensure that for each of such lines there is also at least one non-resonant line on scale n 3 (inside the same self-energy cluster). Therefore we can bound the second product in (8.7) as

2n
L G ( ) n=1

2nNn3 () ,

30

which in turn is bounded as a constant to the power k = k (), as argued in the proof of Lemma 5.3. nT / and T1 contains at least one line PT1 with n = 1, in general there Finally if k (T 1 ) K0 2 1 for i = 1, . . . , p, and T are p 1 self-energy clusters Tp Tp 1 . . . T1 = T1 such that PTi p is the one with largest depth containing . For i = 1, . . . , p call i the exiting line of the self-energy cluster Ti and i = i . Denote also, for i = 1, . . . , p 1, by i = i+1 (i ) (recall Notation 4.29). By Lemma 4.30 one 1 )+ . . . + k ( p1 )). i ) for i = 1, . . . , p 1. Moreover one has | 1 ej | E2 (k ( has | i i+1 | E2 k ( On the other hand one has n +2 ji ( i ) + ji+1 ( i+1 ) 2 Ti+1 , | i i+1 | n j1 ( 1 ) 2 T1 , | 1 ej | so that one can write C k(1 )+...+k(p1 )) C 3k(1 )+...+k(p1 )) 2
p nT
1

2
i=2

nT
i

(8.8)

which assures the gain factors for all self-energy clusters T1 , . . . , Tp . To conclude the analysis, if OT1 = L but OT2 = R , one can reason in the same way by noting that |n n | 1. T
2

Lemma 8.1. Set (c) = max{|c1 |, . . . , |cd |, 1}. There exists a positive constant C such that for k N, V ()| C k 3k (c). j {1, . . . , d} and Zd one has |
Tk j,

Proof. Each time one has a resonant line , when summing together the values of all self-energy clusters, a gain B1 2n is obtained (either by the cancellation mechanism described at the beginning of this Section k or as an eect of the regularisation operator R ). The number of trees of order k is bounded by B2 for some constant B2 ; see Remark 4.12. The derived propagators can be bounded by (8.6). By taking into k account also the bound of Lemma 5.3, setting B3 = C0 , and bounding by B4 , with

B4 = exp 3c log 2
n=0

2n/ n ,

the product of the propagators (both derived and non-derived) of the non-resonant lines times the derived propagators of the resonant lines, we obtain the assertion with C = B1 B2 B3 B4 . Lemma 8.2. The function (2.4), with xj, as in (4.10), and the counterterms j dened in (4.11) are analytic in and c, for ||3 (c) 0 with 0 small enough and (c) = max{|c1 |, . . . , |cd |, 1}. Therefore the solution x(t, , c) is analytic in t, , c for ||3 (c)e3|| |Im t| 0 , with 0 small enough. Proof. Just collect together all the results above, in order to obtain the convergence of the series for (k ) 0 small enough and || (c) 0 , for some constant . Moreover xj, = 0 for | | > k , for the same constant . Lemma 4.10 gives = 3.

Momentum-depending perturbation

Here we discuss the Hamiltonian case in which the perturbation depends also on the coordinates y1 , . . . , yd , as in (2.10). As we shall see, dierently from the y -independent case, here the Hamiltonian structure of the system is fundamental. 31

It is more convenient to work in complex variables z , w = z , with zj = (yj + ij xj )/ the Hamilton equations are of the form iz j = j zj + wj F (z , w, ) + j zj , iw j = j wj + zj F (z , w , ) + j wj , with F (z , w, ) =
p=0 s+ s+ s s

2j , where

(A.1)

as+ ,...,s+ ,s ,...,s z11 . . . zdd w11 . . . wdd .

1 d 1 d

(A.2)

+ s+ 1 ,...,sd ,s1 ,...,sd 0 + + s1 +...+sd +s1 +...+s d =p+3

Note that, since the Hamiltonian (2.8) is real, one has as + , s = a s ,s+ , Let us write so that
fj (z , w, ) d s = (s 1 , . . . , sd ) Z+ .

(A.3)

+ fj (z , w, ) = wj F (z , w , ),

fj (z , w, ) = zj F (z , w, )

=
p=1

1 d 1 d fj, s+ ,s z1 . . . zd w1 . . . wd ,

s+

s+

= ,

s+ ,s Zd + + s1 +...+s+ d +s1 +...+sd =p+1

+ + + + with fj, s+ ,s = (sj + 1)as ,s +ej and fj,s+ ,s = (sj + 1)as +ej ,s , and hence + fj, s+ ,s = fj,s ,s+ + (s+ j2 + 1)fj1 ,s+ +ej
2 ,s

j = 1, . . . , d,

s+ , s Zd ,
1

(A.4a) s+ , s Zd , s+ , s Zd , s ,s Z .
+ d

= (s j1 + 1)fj2 ,s+ ,s +ej ,

1

j1 , j2 = 1, . . . , d, j1 , j2 = 1, . . . , d, j1 , j2 = 1, . . . , d,

(A.4b) (A.4c) (A.4d)

+ + (s j2 + 1)fj1 ,s+ ,s +ej = (sj1 + 1)fj2 ,s+ ,s +ej ,

2

(s+ j2

1)fj + 1 ,s +ej2 ,s

(s+ j1

1)fj , + 2 ,s +ej1 ,s

Expanding the solution (z (t), w (t)) in Fourier series with frequency vector , (A.1) gives
+ ( j )zj, = j zj, + fj, (z , w , ), ( j )wj, = j wj, + fj, (z , w , ).

(A.5)

We write the unperturbed solutions as

i j t zj (t) = c+ , j e (0) i j t wj (t) = c , j e (0)

j = 1, . . . , d,

+ with cj = c j C and cj = cj . As in Section 2 we can split (A.5) into + fj, ej (z , w , ) + j zj,ej = 0, fj, ej (z , w , )

j = 1, . . . , d, j = 1, . . . , d, j = 1, . . . , d, = ej , = ej , j = 1, . . . , d,

(A.6a) (A.6b) (A.6c) (A.6d)

+ j wj,ej = 0,

+ [( ) j ] zj, = fj, (z , w , ) + j zj, , [( ) j ] wj, = fj, (z , w , ) + j wj, ,

so that rst of all one has to show that the same choice of j makes both (A.6a) and (A.6b) to hold simultaneously, and that such j is real. 32

We consider a tree expansion very close to the one performed in Section 4: we simply drop (3) in Constraint 4.4. We denote by Tk j, , the set of inequivalent trees of order k , tree component j , tree momentum and tree sign that is, the sign label of the root line is . and We introduce as in Notation 4.5 and 4.27 respectively, and we dene the value of a tree as follows. The node factors are dened as in (4.2) for the end nodes, while for the internal nodes v V () we dene + sv,1 ! . . . s+ v,d !sv,1 ! . . . sv,d ! v fj ,s+ ,s , kv 1, v v v sv ! Fv = (A.7) 1 kv = 0. v , 2cjv The propagators are dened as G = 1 if = ej and G = Gj ( ),
[n ] [n]

Gj (u) =

n (|u j |) , u j

(A.8)

otherwise, and we dene V () as in (4.9). Finally we set zj,ej = wj, ej = cj , and formally dene

zj, =
k=1

k zj, ,
k (k ) wj, ,

(k )

zj, =
Tk j, ,+ (k ) wj,

(k )

V (),

= ej , (A.9)

wj, =
k=1

=
Tk j, ,

V (),

= ej ,

and j, =

k j, ,
k=1

(k )

j, =

(k )

1 c j

V ().
Tk j,ej ,

(A.10)

Note that Remarks 4.9, 4.13 and 4.17 still hold.

Lemma A.1. With the notations introduced above, one has j, + = j, and zj, = wj, . k Proof. By denition of we only have to prove that for any Tk j, ,+ there exists Tj, , such that V () = V ( ). The proof is by induction on the order of the tree. Given Tk j, ,+ , let us consider the tree obtained from by replacing the labels v of all the nodes v N0 () with v and the labels of all the lines L() with . Call 1 , . . . , p the lines on scale 1 (if any) closest to the root of , and denote by vi the node i enters and by i the tree with root line i . Each tree i is then replaced with a tree i such that V (i ) = V (i ) by the inductive hypothesis. Moreover, as for any internal line ) the node in the momentum becomes , the propagators do not change. Finally, for any v V ( factor is changed into + + sv,1 ! . . . s v,d !sv,1 ! . . . sv,d ! v fj ,s ,s+ , kv 1, v v v sv ! (A.11) Fv = 1 kv = 0. v , 2cjv

Hence by (A.4a) one has V () = V ( ).

33

Lemma A.2. With the notations introduced above, one has j,+ R.
k Proof. We only have to prove that for any Tk j,ej ,+ there exists Tj,ej ,+ such that c+ j V ( ) = cj V ( ). + Let v0 Ej ( ) (existing by Remark 4.9) and let us consider the tree obtained from by (1) exchanging the root line 0 with v0 , (2) replacing all the labels v of all the end-nodes v N0 () \ {v0 } with v , and (3) replacing all the labels of all the internal lines with , except for those in P (v0 , 0 ) which remain the same. The propagators do not change; this is trivial for the lines outside P (v0 , 0 ), while ) \ {v0 } into two disjoint sets of for P (v0 , 0 ) one can reason as follows. The line divides E ( ) \ {v0 } : v w} and end nodes E (, p) and E (, s) such that if = w one has E (, p) = {v E ( s) = (E ( ) \ {v0 }) \ E (, p). If E (,

(p) =
) v E (,p

v,

(s) =
) v E (,s

v ,

one has (p) + (s) = 0. When considering as a line in one has = (p) + ej while in one has = (s) + ej . Hence, as we have not changed the sign label , also G does not change. The node factors of the internal nodes are changed into their complex conjugated; this can be obtained as in Lemma A.1 for the internal nodes w such that w / P (v0 , 0 ) while for the other nodes one can reason as follows. P (v0 , 0 ) entering v . We shall denote First of all if v is such that v P (v0 , 0 ), there is a line v = . Moreover we call s the number of lines outside P (v0 , 0 ) = j , and jv = j1 , v = , j 2 i v v with component label i and sign label entering v . Let us consider rst the case = = +. When considering v as node of one has
Fv

+ + s+ 1 ! . . . sd !s1 ! . . . sd !(sj2 + 1) + fj1 ,s+ +ej ,s 2 sv !

+ + s+ 1 ! . . . sd !s1 ! . . . sd !(sj2 + 1) fj1 ,s ,s+ +ej . 2 sv !

+ When considering v as node of one has s+ v = s + ej1 and sv = s , so that Fv = + s+ 1 ! . . . sd !s1 ! . . . sd !(sj1 + 1) + fj2 ,s +ej ,s+ , 1 sv !

and hence by (A.4b) Fv = Fv . Reasoning analogously one obtains Fv = Fv also in the cases = = and = , using again (A.4b) when = = , and (A.4c) and (A.4d) for = , = + and = +, = respectively. Hence the assertion is proved.

We dene the self-energy clusters as in Section 4.6, but replacing the constraint (3) with (3 ) one has and T as in Notation 4.23 j |. We introduce T | 2 and |T T jT | = | | T T T T T and 4.27 respectively, and we can dene V (T ) as in (4.12) and the localisation and the regularisation operators as in Section 6. Note that the main dierence with the y -independent case is in the role of the sign label . In fact, here the sign label of a line does not depend on its momentum and component labels, and the small divisor is given by j, ( ) = | j |. Hence the dimensional bounds of Section 5 and the symmetries discussed in Section 7 and summarised in Lemma 7.1 can be proved word by word as in the y -independent case, except for the second equality in Lemma 7.3 where one has to take into account a change of signs. More precisely for T Rk j,,j , (u, n), with j = j and n 0 for all PT , we dene G1 (T ) as in Section 7 and G3 (T ) as in Section 7 but replacing also the sign labels of the lines L(T ) with . 34

 Lemma A.3. For all T Rk j,,j , (u, n), with j = j and n 0 for all PT , one has c j cj

L V (T , u) = c j cj

L V (T , u).

(A.12)

T G1 (T )

T G3 (T )

) K0 2nT / . For xed T Rk (u, n), with j = j , let Proof. We consider only the case k (T j,,j , Tk j,ej , (n) be the tree obtained from T by replacing the entering line T with a line exiting a new end node v0 with v0 = and v0 = ej . As in the proof of Lemma 7.3 one has c j

L V (T , u) = |Ej ( )| V ( ).

T G1 (T )

Now, let Tk j,ej , (n) be the tree obtained from by replacing all the labels v of the nodes v N0 () with v , and the labels of all the lines L() with . Any T G3 (T ) can be obtained from by replacing a line exiting an end node v Ej ( ) with entering line T , carrying the same labels as T . Hence, by Lemma A.1,
c j cj

L V (T , u) = c j |Ej ( )| V ( ) = cj |Ej ( )| V ( ) = cj cj

(L V (T , u)) .

T G1 (T )

T G3 (T )

On the other hand, exactly as in Lemma A.2 one can prove that for any T G3 (T ) there exists T G3 (T ) such that cj (L V (T , u)) = c j L V (T , u), and hence the assertion follows. The cancellation mechanism and the bounds proved in Section 8 follows by the same reasoning (in fact it is even simpler); see the next appendix for details.

Matrix representation of the cancellations

As we have discussed in Section 6 the only obstacle to convergence of the formal power series of the solution is given by the accumulation of resonant lines; see Figure 14. The cancellation mechanism described in Section 8 can be expressed in matrix notation. This is particularly helpful in the y -dependent case. For this reason, and for the fact that the formalism introduced in Appendix A include the y -independent case, we prefer to work here with the variables (z , w). We rst develop a convenient notation. Given such that ( , 1) = + and 1,+ ( ) < let us group together, in an ordered set S ( ), all the such that = (j, ) := e1 + ej , = 1 and j = 1, . . . , d; see Remark 4.19. By denition one has 1,+ ( ) = j, ( (j, )) for all j = 1, . . . , d and = . Then we construct a 2d 2d localised self-energy matrix L M (k) ( , n) with entries (k ) L Mj,,j , ( (j , ), n). We also dene the 2d 2d diagonal propagator matrix G [n] ( ) with entries Gj,,j, ( ) = j,j , Gj ( (j, )), with Gj (u) dened according to (A.8). As in Section 8 let us consider a chain of two self-energy clusters; see gure 9. By denition its value is [n ] L V (T1 , 1 ) Gj ( ) L V (T2 , 2 ), with 1 = and 2 = T2 . T
1

[n]

[n]

[n]

Notice that, if one sets also for sake of simplicity, 1 = , j1 = jT , 2 = T2 , and j2 = jT2 , by the T1 1 constraint (3 ) in the denition of self-energy clusters given in Appendix A, one has 1 = 1 ej1 ej and 2 = ej 2 ej2 ; moreover 1 , , 2 all belong to a single set S ( ) for some . 35

As done in Section 8 let us sum together the values of all the possible self-energy clusters T1 and T2 with xed labels associated with the external lines, and of xed orders k1 and k2 , respectively. We obtain
2 1 ( (j2 , 2 ), nT2 ). ( ) L Mj , L Mj1 , ( (j , ), nT1 ) Gj , 1 ,j , ,j , ,j2 ,2

(k )

[n ]

(k )

If we also sum over all possible values of the labels j , we get

d
2 1 ( (j2 , 2 ), nT2 ) ( )L Mj , L Mj1 , ( (j , ), nT1 ) Gj , 1 ,j , ,j , ,j2 ,2

(k )

[n ]

(k )

= j =1

= L M (k1 ) ( , nT1 ) G [n ] ( ) L M (k2 ) ( , nT2 )

j1 ,1 ,j2 ,2

and

(i.e. the entry j1 , 1 , j2 , 2 of the matrix in square brackets). By the denition (A.8) of the propagators and by the symmetries of Lemma 7.1, G [n] ( ) and L M (k) ( , n) have the form 1 0 0 0 0 1 . . . . . . . . n (| 1 |) 0 . , (B.1) G [n] ( ) = . .. .. 1 . . . 0 . 1 0 0 0 0 1
(k ) M1,1 (n)

respectively. A direct computation gives L M (k1 ) ( , nT1 ) G [n ] ( ) L M (k2 ) ( , nT2 )

d j1 ,1 ,j2 ,2

(k ) L M ( , n) = (k ) c c Md,1 (n) d 1 cd c1

c 1 c1 c1 c1 . . .

c 1 c1 c1 c1

Mj,j (n)
(k )

(k ) M1,d (n) c j cj cj cj

c j cj cj cj

c 1 cd c1 cd . . . c d cd cd cd

c d c1 cd c 1

Md,d (n)

(k )

c 1 cd c1 c d , c c d d cd c d

n (| 1 |) 1 2 Mj1 ,j (nT1 ) Mj,j2 (nT2 ) |cj |2 (1)1+1 = 0, cj1 cj2 1 = j =1

(B.2)

for all choices of the scales n , nT1 , nT2 and of the orders k1 , k2 . This proves the necessary cancellation. Note that this is an exact cancellation in terms of the variables (z , w ): all chains of localised self-energy clusters of length p 2 can be ignored as their values sum up to zero. In the y -independent case, and in terms of the variables x, the cancellation is only partial, and one only nds L M (k1 ) G[n] L M (k2 ) = O(2n ), as discussed in Section 8.

Resummation of the perturbation series

The fact that the series obtained by systematically eliminating the self-energy clusters converges, as seen in Section 5, suggests that one may follow another approach, alternative to that we have described so far, 36

and leading to the same result. Indeed, one can consider a resummed expansion, where one really gets rid of the self-energy clusters at the price of changing the propagators into new dressed propagators again terminology is borrowed from quantum eld theory. This is a standard procedure, already exploited in the case of KAM tori [10], lower-dimensional tori [10, 12], skew-product systems [11], etc. The convergence of the perturbation series reects the fact that the dressed propagators can be bounded proportionally to (a power of) the original ones for all values of the perturbation parameter . In our case, the latter property can be seen as a consequence of the cancellation mechanism just described. In a few words and oversimplifying the strategy the dressed propagators are obtained starting from a tree expansion where no self-energy clusters are allowed, and then inserting arbitrary chains of self-energy clusters: this means that each propagator G [n] = G [n] ( ) is replaced by a dressed propagator [n] = G [n] + G [n] M G [n] + G [n] M G [n] M G [n] + . . . , (C.1)

where M = M ( ) denotes the insertion of all possible self-energy clusters compatible with the labels of the propagators of the external lines (M is is the matrix with entries Mj,,j ( (j , )) formally dened in Remark 4.26). Then, formally, one can sum together all possible contributions in (C.1), so as to obtain 1 1 = A1 B , A := G [n] , B := M. (C.2) [n] = G [n] 1 M G [n] For sake of simplicity, let us also identify the self-energy values with their localised parts, so as to replace in (C.1), and hence in (C.2), M with L M , if L is the localisation operator. Then, in the notations we are using, the cancellation (B.2) reads BAB = 0, which implies [n] = A + ABA. Therefore one nds [n] A + A 2 B = O(22n ). So the values of the trees appearing in the resummed expansion can be bounded as done in Section 5, with the only dierence that now, instead of the propagators G bounded proportionally to 2n , one has the dressed propagators [n ] bounded proportionally to 22n . Of course, the argument above should be made more precise. First of all one should have to take into account also the regularised values of the self-energy clusters. Moreover, the dressed propagators should be dened recursively, by starting from the lower scales: indeed, the dressed propagator of a line on scale n is dened in terms of the values of the self-energy clusters on scales < n, as in (C.2), and the latter in turn are dened in terms of (dressed) propagators on scales < n, according to (4.13). As a consequence, the cancellation mechanism becomes more involved because the propagators are no longer of the form (B.1); in particular the symmetry properties of the self-energy values should be proved inductively on the scale label. In conclusion, really proceeding by following the strategy outlined above requires some work (essentially the same amount as performed in this paper). We do not push forward the analysis, which in principle could be worked out by reasoning as done in the papers quoted above.

References
[1] M.V. Bartuccelli, G. Gentile, Lindstedt series for perturbations of isochronous systems: a review of the general theory, Rev. Math. Phys. 14 (2002), no. 2, 121171. [2] A. Berretti, G. Gentile, Bryuno function and the standard map, Comm. Math. Phys. 220 (2001), no. 3, 623656. [3] B. Bollob as, Graph theory. An introductory course, Graduate Texts in Mathematics 63, Springer-Verlag, New York-Berlin, 1979.

37

[4] J. Bricmont, K. Gaw edzki, A. Kupiainen, KAM theorem and quantum eld theory, Comm. Math. Phys. 201 (1999), no. 3, 699727. [5] A.D. Bryuno, Analytic form of dierential equations. I, II, Trudy Moskov. Mat. Ob s c. 25 (1971), 119262; ibid. 26 (1972), 199239. English translations: Trans. Moscow Math. Soc. 25 (1971), 131288 (1973); ibid. 26 (1972), 199239 (1974). ` Jorba, J. Villanueva, KAM theory without action-angle variables, Nonlinearity [6] R. de la Llave, A. Gonz alez, A. 18 (2005), no. 2, 855895. [7] E. De Simone, A. Kupiainen, The KAM theorem and renormalization group, Ergodic Theory Dynam. Systems 29 (2009), no. 2, 419431. [8] L.H. Eliasson, Absolutely convergent series expansions for quasi periodic motions, Math. Phys. Electron. J. 2 (1996), Paper 4, 33 pp. (electronic). [9] G. Gallavotti, Twistless KAM tori, Comm. Math. Phys. 164 (1994), no. 1, 145156. [10] G. Gallavotti, F. Bonetto, G. Gentile, Aspects of ergodic, qualitative and statistical theory of motion, Texts and Monographs in Physics, Springer-Verlag, Berlin, 2004. [11] G. Gentile, Resummation of perturbation series and reducibility for Bryuno skew-product ows, J. Stat. Phys. 125 (2006), no. 2, 321361. [12] G. Gentile, Degenerate lower-dimensional tori under the Bryuno condition, Ergodic Theory Dynam. Systems 27 (2007), no. 2, 427457. [13] G. Gentile, Diagrammatic methods in classical perturbation theory, Encyclopedia of Complexity and System Science, Vol. 2, 1932-1948, Ed. R.A. Meyers, Springer, Berlin, 2009. [14] G. Gentile, Quasi-periodic motions in strongly dissipative forced systems, Ergodic Theory Dynam. Systems, to appear. [15] G. Gentile, M. Bartuccelli, J. Deane, Summation of divergent series and Borel summability for strongly dissipative equations with periodic or quasi-periodic forcing terms, J. Math. Phys. 46 (2005), no. 6, 062704, 21 pp. [16] F. Harary, Graph theory, Addison-Wesley Publishing Co., Reading, Mass.-Menlo Park, Calif.-London, 1969. [17] M. Levi, J. Moser, A Lagrangian proof of the invariant curve theorem for twist mappings, Smooth ergodic theory and its applications (Seattle, WA, 1999), 733746, Proc. Sympos. Pure Math., 69, Amer. Math. Soc., Providence, RI, 2001. [18] J. Moser, Convergent series expansions for quasiperiodic motions, Math. Ann. 169 (1967) 136176. [19] H. Poincar e, Les m ethodes nouvelles de la m ecanique c eleste, Vol. IIII, Gauthier-Villars, Paris, 18921899. [20] D. Salamon, E. Zehnder, KAM theory in conguration space, Comment. Math. Helv. 64 (1989), 84132.

38