Journal of Differential Equations 166, 79106 (2000)

doi:10.1006jdeq.2000.3783, available online at http:www.idealibrary.com on

Further Reductions of Normal Forms for

Dynamical Systems
Guoting Chen
Departement de Mathematiques, Universite de Lille 1, 59655 Villeneuve d'Ascq, France
E-mail: Guoting.Chenuniv-lille1.fr

and

Jean Della Dora

LMC-IMAG, 51 Rue des mathematiques, B.P. 53, 38041 Grenoble Cedex 9, France
E-mail: Jean.DellaDoraimag.fr

Received January 28, 1999; revised July 13, 1999

We propose in this paper a method for obtaining a significant refinement of nor-

mal forms for dynamical systems or vector fields, with concrete and interesting
applications. We use lower order nonlinear terms in the normal form for the sim-
plifications of higher order terms. Our approach is applicable for both the non
nilpotent and the nilpotent cases. For dynamical systems of dimensions 2 and 3 we
give an algorithm that leads to interesting finite order normal forms which are
optimal (or unique) with respect to equivalence by formal near identity transforma-
tions. We can compute at the same time a formal diffeormorphism that realizes the
normalization. Comparisons with other methods are given for several examples.
 2000 Academic Press

1. INTRODUCTION

Let K be a commutative field of characteristic zero (which, in practice,

is Q, R). We denote by K[[x]] the algebra of formal power series in n
variables with coefficients belonging to the field K. In this paper we con-
sider the normal form problem for a dynamical system (or its associated
vector field)

dx
x* = =F(x) (1)
dt

where x=(x 1 , ..., x n ) t, and F(x)=( f 1 (x), ..., f n (x)) t with f j (x) # K[[x]]
and F(0)=0. One can write F(x)= k1 F k (x), where F k (x) is a vector
of homogeneous polynomials of degree k and F 1 (x)=Ax, A # gl(n, K), is
79
0022-039600 35.00
Copyright  2000 by Academic Press
All rights of reproduction in any form reserved.
80 CHEN AND DELLA DORA

the linear part of the system, where gl(n, K) denotes the set of n_n
matrices with entries in K.
Let *=(* 1 , ..., * n ) be the eigenvalues of A and suppose that A is in
the Jordan canonical form with * 1 , ..., * n on the diagonal. Let
q=(q 1 , ..., q n ) # N n with

|q| =q 1 + } } } +q n 2.

We define x q =x q11 } } } x qnn . A monomial a q x q in f j (x) is called resonant of

order |q| if * j =( q, *) =q 1 * 1 + } } } +q n * n .
The PoincareDulac normal form theorem states the following

Theorem (PoincareDulac). Consider a dynamical system of the form

(1) where F(x)=Ax+ } } } is a vector of formal power series, and A is in the
Jordan canonical form as above. Then (1) can be reduced to a normal
form y* =Ay+h( y) by a near identity change of variables of the form
x= y+.( y), where .( y) is a vector of formal power series without constant
term or linear part and h( y) contains only resonant monomials.
In practice one uses the PoincareDulac Theorem to get a normal form
up to some finite order, say N, to obtain

y* =Ay+z( y)+O(&y& N+1 )

where z( y) is a polynomial of total degree less than or equal to N without

constant or linear terms and O(&y& N+1 ) represents terms of order greater
than or equal to N+1.
A fundamental improvement of the normal form is given by Takens in
[20] which shall be studied and improved upon in this paper.
Systematic procedures for constructing normal forms have been given in
many ways. A method using Lie brackets is given in [13, 20, 22], a method
by considering an inner product in the space of homogeneous polynomials
is given in [2, 12, 15], a method by direct computation is given in [5], and
a method using Carleman linearization is given in [21]. The nilpotent case
(A is a nilpotent matrix) is treated in [14] using representation theory of
sl 2 (R) and in [9, 10] using Carleman linearization.
All of the above works use the Jordan canonical form of the leading
matrix A. But it is well known that handling eigenvalues and Jordan
canonical forms is a very difficult problem in computer algebra systems.
We propose in this paper to give further reductions of the classical normal
forms for dynamical systems. Our approach is rational in the sense that if
the coefficients of the system are in K (which is Q, R in practice), so is the
normal form and all the computations are done in K. We do not need to
compute the Jordan canonical form of A nor its eigenvalues. Our method
 REDUCTIONS OF NORMAL FORMS 81

is applicable in both the nilpotent and the non nilpotent cases. For
dynamical systems of dimensions 2 and 3 we give an algorithm that leads
to interesting results which improve several normal forms given by Ushiki
in [22] by increasing their orders. The normal forms obtained in these
cases are unique with respect to equivalence by near identity changes of
variables up to the given order. Uniqueness results on some semisimple
examples are given in [3]. Our method can also be used to find degeneracy
conditions and to compute the corresponding degenerate normal forms.
We refer the reader to [8] for a general algorithm in any dimension using
the Carleman linearization procedure.
We begin with a presentation of the classical normal form theorem in
Section 2.1. We recall in Section 2.2 the Takens' normal form of vector
fields in dimension 2 which is used in Section 3. And we then give a refine-
ment of the classical normal form in Section 2.3. In Section 3 we give an
algorithm for the case of dimension 2 by a detailed study of the homologi-
cal equations. In Section 4 we study dynamical systems in dimension 3. We
have a unified treatment for the nilpotent and the non nilpotent cases. We
obtain interesting normal forms for both the nilpotent and the non nilpo-
tent cases. In particular, we get further reductions of normal forms of
Takens.

2. FUNDAMENTAL THEOREMS ON NORMAL FORMS

FOR DYNAMICAL SYSTEMS

2.1. Classical Normal Form Theorem

Consider a dynamical system of dimension n of the form (1), i.e.,

x* =F(x)=Ax+F 2 (x)+F 3 (x)+ } } } , x # K n, (2)

where A is an n_n matrix with coefficients in K and F is a vector of formal

power series with coefficients in K. The notation F k is used to denote the
homogeneous part of degree k of F, F k # H nk , where H k is the space of
homogeneous polynomials of degree k in n variables with coefficients in K.
Consider at first a linear transformation

x=. 1 ( y)=Py (3)

where P is an invertible matrix with coefficients in K. We obtain a new

system in the same form: y* =P &1APy+ } } } . It is clear that one can choose
a matrix P such that A is in some canonical form. In the classical methods
one chooses the Jordan canonical form. In the following we shall use, as in
[8], the Frobenius canonical form whose coefficients remain in the field K.
82 CHEN AND DELLA DORA

Consider next a formal transformation (a near identity change of coor-

dinates)

x= y+.( y) (4)

where .( y) has no constant or linear terms. By substituting this in (2), we

get:

y* =(I+ y .) &1 F( y+.( y)) (5)

where  y . denotes the Jacobian matrix of . with respect to ( y 1 , ..., y n ).

Therefore one can write the new system in the form

y* =Ay+G( y).

We shall say that this new system is equivalent to the original system (2)
by a formal near identity change of variables.
The goal is to determine changes of coordinates (3) and (4) such that the
transformed system will be in the simplest form possible in a sense to be
specified later on. The desired simplification of (2) will be obtained, up to
terms of a specified order, by performing inductively a sequence of near
identity changes of coordinates of the form x= y+. k ( y) where . k ( y) is
a vector of homogeneous polynomials of degree k2. We have

(I+ y . k ( y)) &1 =I& y . k ( y)+O(&y& 2k&2 )

where O(&y& m ) denotes terms of order greater than or equal to m.

Substituting it into (5) we obtain

y* =Ay+ } } } +F k&1 ( y)+[F k ( y)&[ y . k ( y) Ay&A.k ( y)]]+O(&y& k+1 ).

(6)

One then has the equation

F k ( y)&[ y . k ( y) Ay&A. k ( y)]=G k ( y) (7)

which is called the homological equation (of order k). In order to obtain
a G k that contains as few monomials as possible we have to choose a
suitable . k ( y). We introduce for each k2 a linear operator L kA : H nk Ä H nk
defined by

(L kA . k )( y)= y . k ( y) Ay&A. k ( y), . k # H nk .

Then (6) can be expressed as

y* =Ay+ } } } +F k&1 ( y)+[F k ( y)&L kA . k ( y)]+O(&y& k+1 ).

 REDUCTIONS OF NORMAL FORMS 83

Let R k be the range of L kA in H nk and C k be any supplementary subspace

to R k in H nk . We have

H nk =R k ÄC k. (8)

The fundamental idea of the classical normal form theory is in the

following theorem (due to Takens [20]) which is an improvement of the
PoincareDulac theorem (see also [1, 11, 12]).

Theorem 2.1 (Takens). Consider a dynamical system of the form (2).

Let the notations be as above. Suppose that the decomposition (8) is given
for k=2, ..., N. Then there exists a sequence of near identity changes of
variables x= y+. k ( y) where . k ( y) # H nk such that the dynamical system
(2) is transformed into

y* =Ay+G 2 ( y)+ } } } +G N ( y)+O(&y& N+1 )

where G k # C k for k=2, ..., N.

The method in [15] using an inner product in H nk and the method in
[14] using the representation theory of sl 2 (R) lead to the above normal
form of Takens. Their methods construct a basis of a supplementary sub-
space C k. In general this basis consists of homogeneous polynomials (not
always monomials). See also [2, 8, 1012, 21] for other methods.

2.2. An Example of Takens' Normal Form

In this subsection we suppose K=R. For notations and definitions we
refer the reader to [20].
A dynamical system of the form (1) can be seen as a vector field in R n
of the form X= ni=1 f i (x)(x i ) with a singularity at the origin, i.e.,
X(0)=0. We define X 1 to be the vector field in R n which has the same 1-jet
in 0 as X and whose coefficient functions are linear, i.e. the linear part of X.
[X 1 , &]: H nk Ä H nk is the linear map which assigns to each Y # H nk
the Lie product [X 1 , Y] which is again in H nk . For X 1 fixed, we define a
splitting of H nk =R k ÄC k such that R k =Im([X 1 , &]) and C k is some
supplementary subspace of R k in H nk . In a similar way Theorem 2.1 can be
stated in terms of vector fields (see [20]): For any integer k2, there is
a formal diffeomorphism .: (R n, 0) Ä (R n, 0) such that

.*(X)=X 1 + g 2 + } } } + g k +R k ,

where g i # C i, i=2, ..., k, and R k is a vector field, the component functions

of which all have zero k-jet.
84 CHEN AND DELLA DORA

We shall apply this theorem to vector fields in R 2 with a linear part

having only one-block Frobenius canonical form, i.e.,

 
X 1 =x 2 +(c 0 x 1 +c 1 x 2 )
x 1 x 2

with c 0 , c 1 # R. The range R k of [X 1 , &] in H nk is determined by the for-

mulae:

  
_ &
X 1 , x k1
x 1
=kx k&1
1 x2
x 1
&c 0 x k1
x 2
, (9)

  
_ x & =&x
X ,x1
k
1
2
k
1
x 1
+(kx k&1
1 x 2 &c 1 x k1 )
x 2
, (10)

and for 0 jk&1,

 
_X , x x
1
j
1
k& j
2
x 1 &
=&c 0 x 1j x k&
2
j
x 2

+( jx 1j&1 x k&
2
j+1
+(k& j) c 1 x 1j x k&
2
j
+(k& j) c 0 x 1j+1 x k&
2
j&1
) ,
x 1
(11)
 
_X , x x
1
j
1
k& j
2
x 2 &
=&x 1j x 2k& j
x 1

+( jx 1j&1 x k&
2
j+1
+(k& j&1) c 1 x 1j x k&
2
j
+(k& j) c 0 x 1j+1 x k&
2
j&1
) .
x 2
(12)

Now we prove that H nk decomposes into H nk =R k +B k where the sub-

space B k is spanned by x k1 (x 1 ) and x k1 (x 2 ). In fact, we prove, by
downward induction on j, that x 1j x k&2
j
(x 1 ) and x 1j x k&
2
j
(x 2 ) belong
k k k k
to R +B for all j=k, ..., 0. Since x 1 (x i ) # B for i=1, 2, it is clear
from (9) and (10) that


x k&1
1 x2 # R k +B k for i=1, 2.
x i

Suppose now that this is true for l1. Equations (11) and (12) for j=l
imply that


x l&1
1 x k&l+1
2 # R k +B k (i=1, 2).
x i

This proves the following:

 REDUCTIONS OF NORMAL FORMS 85

Proposition 2.1. Let X be a vector field in R 2 whose 1-jet in 0 equals the

1-jet of x 2 (x 1 )+(c 0 x 1 +c 1 x 2 )(x 2 ) with c 0 , c 1 # R. Then for any
integer k2 there is a formal diffeomorphism .: (R 2, 0) Ä (R 2, 0) such that
k
   
.*(X)=x 2
x 1
+(c 0 x 1 +c 1 x 2 )
x 2 l=2 \
+ : a l x l1
x 1
+b l x l1
+
x 2
+R k

where the k-jet of R k is zero.

Of course, some of the a l and b l may turn out to be zero.
In particular with X 1 =x 2 (x 1 ) we obtain the Takens' nilpotent nor-
mal form (see [20])
k
  
.*(X)=x 2
x 1 i=1 \
+ : a i x i1
x 1
+b i x i1
x 2 +
+R k ,

where the k-jet of R k is zero.

Ushiki in [22] has obtained further reductions of normal forms up to
some finite orders in some cases of dimensions 2 and 3. Baider in [3] and
Gaeta in [17] also have given further reduction for systems in dimension
2 with a semisimple linear part. Baider and Sanders in [4] and Gaeta in
[17] also have obtained further reduction of the Takens normal form in
the nilpotent case of dimension 2. But their reductions are not complete in
the latter case.
Our aim in this paper is to give further reductions of the above normal
forms up to some finite order. Our normal forms will be optimal (or
unique) up to the given order with respect to equivalence by formal near
identity transformations.
2.3. A Refinement of the Classical Normal Form
First we note that a normal form is not unique for a fixed A. In fact it
depends on the choices of the supplementary subspaces C k (k=2, ..., N). It
is not unique even with fixed C k. It is also important to note that G l and
. l (2lk) are unique if and only if C l =[0]. The basic idea is that if
. l (l<k) is not unique then it may be used to simplify some terms of G k.
This idea is essential in our method and has been used by other authors
(see [17, 22]). We will show that it is possible to make such a choice.
Let l2 be an integer such that C l {[0]. From the decomposition (8)
one obtains that dim Ker(L lA )=dim C l. Let . l1 # H nl such that

G l ( y)=F l ( y)&L lA(. l1 )( y) # C l

is in a normal form as in Theorem 2.1. Then

L lA(. l1 + l )=L lA(. l1 ) for all  l # Ker(L lA ).

86 CHEN AND DELLA DORA

Consider a transformation of the form

x= y+.( y)= y+. l ( y)+ } } } +. k ( y).

Then the original system is converted to

y* =Ay+F 2 ( y)+ } } } +F l&1 ( y)

+F l ( y) +F l+1 ( y) + } } } +F k&1 ( y) +F k ( y)
&L lA(. l )( y)&L l+1
A (. l+1 )( y)& } } } &L k&1
A (. k )( y)&L kA(. k )( y)
&T l+1
F (.)( y) & } } } &T k&1
F (.)( y) &T kF (.)( y)

where T jF (.) are terms of order j. Then T jF (.) only depends on

F 2, ..., F j&1 and . l, ..., . j&1. Let . l =. l1 + l where  l # Ker(L kA ). Then

G l ( y)=F l ( y)&L lA(. l )( y) # C l.

According to Theorem 2.1, one can choose successively . l+1, ..., . k (which
may depend on  l ) such that the transformed system is in a normal form,
i.e.,

F j &T jF (.)&L Aj (. j ) # C j

for j=l+1, ..., k. One can write

F k ( y)&T kF (.)( y)&L kA(. k )( y)=G k &G k

where G k and G k belong to C k and where G k contains all terms depending

on  l.
Define a nonlinear operator

N l,F k : Ker(L lA ) Ä C k by N l, k l
F ( )=G k.

Let R k2 be a subspace contained in the range of the operator N l,F k in C k

and C k2 be a supplementary subspace to R k2 in C k. Then

C k =R k2 Ä C k2 . (13)

The main theorem of this paper is the following, which is a refinement

of the fundamental Theorem 2.1.

Theorem 2.2. Consider the dynamical system (2) and let the notations be
as above. Assume that Ker(L lA ){[0]. Suppose that there exists a non-trivial
subspace R k2 such that we have decomposition (13). Then there exist
 REDUCTIONS OF NORMAL FORMS 87

. l, ..., . k as above and  l # Ker(L lA ) such that the dynamical system (2) is
transformed into

y* =Ay+G 2 ( y)+ } } } +G k ( y)+O(&y& k+1 )

where G j # C j for 2 jk&1 and G k # C k2 .

Although the above normal form is in general not unique, in some cases
of interest it turns out to be actually unique as we shall show in the next
sections.
Note that Ushiki in [22], and similarly Gaeta in [17], used Ker(L lA ) to
give further reduction of normal forms in a different way. Ushiki has
treated some examples in dimensions 2 and 3 including the nilpotent and
some non nilpotent cases. Gaeta has treated many examples with semisim-
ple linear part and also the nilpotent case of dimension 2. It seems that his
normal form in the latter case is the same as that obtained by Elphick
et al. in [15]. We shall compare our method with that of Ushiki on some
examples in the next sections.
In the following we study applications of the above theorem to dynami-
cal systems of dimensions 2 and 3. We give an algorithm that computes
both the classical and our refined normal forms for dynamical systems of
dimensions 2 and 3.

3. DYNAMICAL SYSTEMS OF DIMENSION 2

3.1. Reduction to Classical Normal Forms
Consider a dynamical system of dimension 2, i.e., x* =F(x)=Ax+ } } }
with n=2. We shall consider a rational normal form using the Frobenius
canonical form of A. Write

0 1
A=
\ c 0 c1 +
where c 0 , c 1 # K. One can determine matrices P such that P &1AP=A.
There may exist arbitrary parameters in the matrix P. The linear transfor-
mation x=Py will not change the linear part of the system but one can use
the arbitrary parameters in P to simplify some higher order terms of the
system.
Let k be an integer 2 and

F(x)=Ax+ } } } +F k (x)+ } } }

where F k is the homogeneous part of degree k in F. Make a change of

variables of the form x= y+.( y) in the system (1), where .( y) is
88 CHEN AND DELLA DORA

homogeneous of degree k, then one can determine .( y) by solving the

homological equation (7), which is

( y .) Ay&A.( y)=F k ( y)&G k ( y). (14)

The problem is then to determine . so that G k ( y) contains the smallest

possible number of monomials.
One writes

: :q xq : :$q y q : :q yq
F k(x)=
\ + \ + \ +
|q| =k

: ;q x
|q| =k
q
, G k(y)=

with a q , b q , :$q , ;$q # K to be determined.

|q| =k

: ;$q y q
|q| =k
, .(y)=
|q| =k

: bq yq
|q|=k

The equations in (14) can be written as

: ((q 1 +1) a q+e1 &e2 +q 2 c 1 a q &b q +(q 2 +1) c 0 a q&e1 +e2 ) y q

|q| =k

= : (: q &:$q ) y q,
|q| =k

: ((q 1 +1) b q+e1 &e2 &c 0 a q +(q 2 &1) c 1 b q +(q 2 +1) c 0 b q&e1 +e2 ) y q
|q| =k

= : ( ; q &;$q ) y q
|q| =k

where e 1 =(1, 0), e 2 =(0, 1) form the standard basis of K 2. We then need
to solve the equations

(q 1 +1) a q+e1 &e2 +q 2 c 1 a q &b q +(q 2 +1) c 0 a q&e1 +e2 =: q &:$q ,

(q 1 +1) b q+e1 &e2 &c 0 a q +(q 2 &1) c 1 b q +(q 2 +1) c 0 b q&e1 +e2 =; q &;$q .

If we can solve these equations with :$q =;$q =0 for all q such that |q| =k
then G k =0. If not, then a normal form obtained in this way will contain
non zero terms in the homogeneous part of degree k.
Let

aq :q :$q
Mq =
\b + ,
q
4q=
\; + ,
q
4$q =
\;$ + ,
q
 REDUCTIONS OF NORMAL FORMS 89

and let I denote the 2_2 identity matrix. Then the above equations can be
written in the form

(q 1 +1) Mq+e1 &e2 +(q 2 c 1 I&A) Mq +(q 2 +1) c 0 Mq&e1 +e2 =4 q &4$q . (15)

It is to be understood that Mq =0 if one of the components of q is strictly

negative. We consider the lexicographical order for the set

[q=(q 1 , q 2 ) # N 2 : |q| =k].

For q 1 =0, q 2 =k one has the equations:

M(1, k&1) +(kc 1 I&A) M(0, k) =4 (0, k) &4$(0, k) .

Let M(0, k) be chosen arbitrarily. The above equations have a unique solu-
tion M(1, k&1) for any value on the right-hand side. Hence one can find a
solution M(1, k&1) for 4$(0, k) =0.
For q=( j, k& j) the above equations become

( j+1) M( j+1, k& j&1) +((k& j) c 1 I&A) M( j, k& j)

+(k& j+1) c 0 M( j&1, k& j+1) =4 ( j, k& j) &4$( j, k& j) .

By induction on j, we can compute .( y) such that 4$( j, k& j) =0 for all

0 jk&1, i.e.,

:$k, 0 y k1
G k ( y)=
\ +
;$k, 0 y k1
.

It is clear that this can be done for any k2. Then a truncated rational
normal form for dynamical systems of dimension 2 is

N
: :$k, 0 y k1
y* =Ay+
\ +k=2

N
: ;$k, 0 y k1
k=2
+O(& y& N+1 ),

which coincides with a normal form of Proposition 2.1.

We now give a detailed study of Eq. (15) to show that further reductions
can be done.
90 CHEN AND DELLA DORA

Let Z j =M( j, k& j) , then

1
Z j+1 = ((A&(k& j) c 1 I ) Z j &(k& j+1) c 0 Z j&1
j+1

+4 ( j, k& j) &4$( j, k& j) ) .

Let Z 0 be given arbitrarily. One can determine Z 1 , ..., Z k by the above

formulae such that 4$( j, k& j) =0 for 0 jk&1. For example, for j=0,
Z 1 =(A&kc 1 I ) Z 0 +4 (0, k) . In particular, all the Z 1 , ..., Z k depend linearly
on Z 0 . In fact, with

Z j =E j&1 Z 0 +B j&1 , 1 jk,

we have E 0 =A&kc 1 I, B 0 =4 (0, k) ,

E 1 = 12 ([A&(k&1) c 1 I] E 0 &kc 0 I) ,

B 1 = 12 (4 (1, k&1) +[A&(k&1) c 1 I] B 0 ) ,

and for j 2,

1
Ej = ([A&(k& j) c 1 I] E j&1 &(k& j+1) c 0 E j&2 ) ,
j+1

1
Bj = (4 ( j, k& j) +[A&(k& j) c 1 I] B j&1 &(k& j+1) c 0 B j&2 ) .
j+1

Then for j=k we have

4$(k, 0) =Z k+1 =4 (k, 0) +AZ k &c 0 Z k&1 =E k Z 0 +B k (16)

where E k is a 2_2 matrix and B k is a vector of dimension 2, which can be

determined by iteration of the above form.
There are three cases to be distinguished:

(a) If det E k {0 then one can find Z 0 such that 4$(k, 0) =0. This
means that one can eliminate all terms of order k in the dynamical system,
i.e., the homogeneous part of degree k in the normal form is reduced to 0.
With notations as in Theorem 2.1, C k =[0] in this case.
(b) If rank(E k )=1 then one can solve Eq. (16) in such a way that
one element of 4$(k, 0) (:$(k, 0) or ;$(k, 0) ) vanishes. This means that one can
 REDUCTIONS OF NORMAL FORMS 91

eliminate all terms of order k except one monomial. Then the

homogeneous part of degree k in the normal form is
:$(k, 0) y k1 0
\ 0 + or
\;$ (k, 0) y k1 +.
In this case we have dim C k =dim Ker(L kA )=1.
(c) E k =0 then in general 4$(k, 0) {0. The homogeneous part of
degree k in the normal form is
:$(k, 0) y k1
\ ;$(k, 0) y k1
.
+
In this case we have dim C k =dim Ker(L kA )=2.
For a given k, we are interested in the values of the determinants of the
matrices E k . We now give a detailed discussion for some values of k. The
following computations have been done in Maple V, Release 4.
If c 0 =0 then E=0. Then we are in case (c).
We now consider the case where c 0 {0. For k=2 one has

det(E 2 )=& 14 c 20 (2 c 21 +9 c 0 ).

Then if c 0 { & 29 c 21 all terms of order 2 can be eliminated in the normal

form of the dynamical system. If c 0 =& 29 c 21 then rank(E 2 )=1. In fact, we
have
2 3 1 2
& 27 c1 c
E2=
_ 2 4
& 81 c 1
9 1
1 3
27 1 c &.
Then we are in case (b). The homogeneous part of degree 2 in the normal
form can be chosen to be in the form

0
\;$ (2, 0) y 21 +.
For k=3, det(E 3 )=& 49 c 20 c 21 (3 c 21 +16 c 0 ), and for k=4,
9 2
det(E 4 )=& 64 c 0(25c 0 +4c 21 )(&2c 21 +c 0 ) 2;

we can give discussions similar to those above and continue with higher
orders.
For the particular case where A=( &10 01 ), we can prove by induction that
E 0 =A and for i1,

k&1 k&2i+1 k&1 k&2i+1

E 2i&1 = }}} I, E 2i = }}} A.
2 2i 2 2i
92 CHEN AND DELLA DORA

Therefore if k=2m is even then E k =(2m&1)2 } } } (12m) A. We have

det(E k ){0. In these cases the homogeneous part of degree k can be
reduced to 0. If k is odd then E k =0, and we will study the further reduc-
tions in the next subsection.
For the case where

0 1
A=
\ 0 0+ ,
we have E k =0 and we are in case (c). The homogeneous part of degree k
in the normal form contains two non-zero parameters. Note that Cushman
and Sanders in [14] and Elphick et al. in [15] by another method
obtained a normal form

x* 1 =x 2 , x* 2 =x 1 x 2 P(x 1 )+x 21 P 2 (x 1 ),

where P 1 (x 1 ) and P 2 (x 1 ) are formal power series in x 1 . In particular, the

homogeneous part of degree k contains also two nonzero parameters.
The above algorithm leads to a classical normal form. In the following,
we apply Theorem 2.2 to study further reductions of the normal forms in
the cases (b) and (c).

3.2. Further Reductions of the Classical Normal Forms

If there remain elements of Z 0 undetermined for the computation of a
normal form of a certain order then one may choose them in looking for
normal forms of higher order. This idea is essential in our method. Now we
explain the method on two typical examples. We study dynamical systems
of dimension 2,

dx x 2 + f 1 (x)
dt
=F(x)=
\
c 0 x 1 +c 1 x 2 + f 2 (x) + (17)

where

f1 (x)= : : q x q; f 2 (x)= : ; q x q.
|q| 2 |q| 2

Example 3.1. Consider at first a dynamical system of the above form

with c 0 =c 1 =0, i.e., the matrix of the linear part is

0 1
A=
\ 0 0+ .
 REDUCTIONS OF NORMAL FORMS 93

Let

u u0
P=
\0 u + (18)

with u, u 0 # K, u{0. Then P is invertible and P &1AP=A.

Suppose that ; 2, 0 {0, then one takes u=1; 2, 0 and ; 2, 0 is reduced to
1. So one can suppose without loss of generality that ; 2, 0 =1 and take
u=1 in P. We take u 0 =0 since it is not used in our normal form.
To obtain a normal form of the dynamical system we apply the algo-
rithm given in the previous section. We use superscripts Z j(k) to show the
t
dependence of the Z j on k. For k=2, let Z (2)
0 =(U 1 , U 2 ) . We then have

1
U 2 +: 0, 2 ; 0, 2 + 12 : 1, 1 : 2, 0 + 12 ; 1, 1
Z (2)
1 =
_ ; 0, 2 &
, Z (2)
2 =
_ 2
1
2 ; 1, 1
& , Z (2)
3 =
_ 1 & .

Hence one can perform the transformation x Ä x+.(x) as given in

Section 3.1. The homogeneous part of degree 2 in the normal form is (for
arbitrary Z (2)
0 ):

(: 2, 0 + 12 ; 1, 1 ) x 21
\ x 21 +
.

With the notations of Section 2.3, the subspace Ker(L 2A ) is spanned by

x 22 0
\ +
0
and
\x + .
2
2

Hence  2 is in the form

U 1 x 22
\ U 2 x 22+.

Theorem 2.2 is applicable with l=2. We shall choose suitable  2 for

further simplifications of higher order normal forms.
t
For k=3, let Z (3)0 =(U 3 , U 4 ) , then one has for example

U 4 +2 : 0, 2 U 2 +: 0, 3 +: 1, 1 U 1 u
Z (3)
1 =
_ U 1 ; 1, 1 +; 0, 3 +2 ; 0, 2 U 2 &
, ..., Z (3)
4 =
v
,
_&
94 CHEN AND DELLA DORA

where

u= 16 ; 0, 2 ; 1, 1 +: 3, 0 &U 2 & 13 : 0, 2 & 13 ; 0, 2 : 2, 0 + 13 ; 2, 1 + 16 : 1, 1 ; 1, 1

and v=&; 1, 1 : 2, 0 +; 3, 0 +: 1, 1 . It is clear now that one can choose

U 2 = 16 ; 0, 2 ; 1, 1 +: 3, 0 + 16 : 1, 1 ; 1, 1 & 13 : 0, 2 & 13 ; 0, 2 : 2, 0 + 13 ; 2, 1

such that the first element of Z (3)

4 is zero. Hence the homogeneous part of
degree 3 in the normal form is

0
\ ( ; 3, 0 +: 1, 1 &; 1, 1 : 2, 0 ) x 31+.

In this case with the notations of Theorem 2.2, R 32 is the subspace

generated by

x 31
\ +
0
.

With C 32 generated by

0
\x + ,
3
1

one has

C 3 =R 32 Ä C 32 .

Theorem 2.1 is applicable with l=2 and k=3.

If ; 2, 0 =0, then there are two parameters in the homogeneous part of
degree 3 in the normal form but only one parameter in the homogeneous
part of degree 2.

Proposition 3.1. Let the notations be as above. Assume that the matrix
of the linear part of the system is A=( 00 10 ).

(a) If ; 2, 0 =1 and # 1 =: 2, 0 + 12 ; 1, 1 {0, then a normal form of order

9 of the dynamical system (17) is

x* 1 =x 2 +# 1 x 21 +O(&x& 10 ),

x* 2 =x 21 +# 3 x 31 +# 4 x 41 +# 5 x 51 +# 6 x 71 +# 7 x 81 +O(&x& 10 ).
 REDUCTIONS OF NORMAL FORMS 95

(b) If ; 2, 0 =1 and 2: 2, 0 +; 1, 1 =0, then a non-degenerate normal

form of order 9 of the dynamical system (17) is

x* 1 =x 2 +#$1 x 41 +#$2 x 51 +#$3 x 71 +#$7 x 81 +O(&x& 10 ),

x* 2 =x 21 +#$5 x 31 +#$6 x 61 +O(&x& 10 ).

(c) If ; 2, 0 =0 and 5; 1, 1 : 2, 0 &4: 22, 0 &; 21, 1 &9; 3, 0 {0, then a non-
degenerate normal form of order 9 of the dynamical system (17) is

x* 1 =x 2 +# 1 x 21 +#"2 x 31 +O(&x& 10 ),
x* 2 =#"3 x 31 +#"4 x 41 +#"5 x 51 +#"6 x 61 +#"7 x 71 +#"8 x 81 +#"9 x 91 +O(&x& 10 ).

# j , #$j , #"j are parameters depending only on the coefficients of F.

Moreover, the above normal forms are optimal or unique in the given order
with respect to equivalence by near identity changes of variables in the sense
that two normal forms are equivalent by a near identity change of variables
if and only if all of the parameters in the normal forms are equal.
The non-degeneracy conditions are algebraic conditions on the coef-
ficients of the system, for instance, #$1 {0, etc. The parameters # j , #$j , #"j can
be given by explicit formulae from the coefficients of F. For example,

# 3 =&; 1, 1 : 2, 0 +; 3, 0 +: 1, 1 and #$5 =: 1, 1 +; 3, 0 +2 : 22, 0 .

In case (c), if # 1 {0 then it can be reduced to 1 by the linear transforma-

tion (18) with u 0 =0, u=1# 1 .
Remark that we have reduced all terms of degrees 6 and 9 to zero.
To simplify the computations we assume that the dynamical system is in
the Takens' normal form as in Section 2.2, i.e.,

f 1 = : : k x k1 , f 2 = : ; k x k1 ,
k2 k2

t t t
with ; 2 =1. Let Z (2) (3) (4)
0 =(U 1 , U 2 ) , Z 0 =(U 3 , U 4 ) , and Z 0 =(U 5 , U 6 ) , we
obtain for k=4,

U 6 +: 2 U 21 & 32 U 21 +: 2 U 3
Z (4)
1 =
_ U 21& ,
_ & Z (4)
2 =
&: 3 U 1 +U 3
,

2 1 2 1 5 2 7
: U + : U & U & U ; + : & U
Z (4)
3 =
_ & U + : +U ; &,
3 3
1
3
1

4
Z =
_
3
1
3
2
; : 2
3
&,
4 3

1
3

3
(4)
4
4 1 3
1
4
6

3
3

3
12 4

2 1 2 3
& : U & : : +: & ; :
Z (4)
5 =
_ & U + : : +; & .
3 2 1
5
3 1
3 2 3
2
3 2 3
4 4

4
3 3
96 CHEN AND DELLA DORA

If : 2 {0 then one can choose

12: 4 &4: 22 : 3 &9; 3 : 3

U1 = ,
8: 2

and the homogeneous part of degree 4 in the new system is

_ 12: 22 : 3 &20: 4 +15; 3 : 3 +8; 4 : 2 4 .

8: 2
x1 &
The algorithm can be used for higher order normal forms. We obtain by
computations in Maple the normal forms of the orders given in the above
proposition.
We give a comparison of the normal forms derived via Takens' method
and Ushiki's method. In [22] Ushiki obtained a normal form of order 4.
The way he used is hardly applicable for computations of higher order nor-
mal forms, so there are some empty spaces in the following table. The nor-
mal form of Cushman and Sanders [14] and that of Elphick et al. [15]
contain two non-zero parameters in each order, the same as the Takens'
normal form. The same is true for the normal form defined in [9] (see also
[10]). Since the goal for obtaining normal forms of dynamical systems is
to eliminate as many monomials from each order as possible, we will list
in the following table the number of monomials of each degree that is still
present in the normal form. For example, 0 means that all terms of a given
degree are eliminated (see also [13, p. 60]).
Degree: 2nd 3rd 4th 5th 6th 7th 8th 9th

Takens 2 2 2 2 2 2 2 2
Ushiki 2 1 1
Case (a) 2 1 1 1 0 1 1 0
Case (b) 1 1 1 1 1 1 1 0
Case (c) 1 2 1 1 1 1 1 1

Example 3.2. We now consider the case where the matrix of the linear
part is A=( &10 10 ). As proved in the preceding section, the homogeneous
part of even degree can be reduced to 0. We suppose that this is done to
simplify the notation.

Proposition 3.2. Consider a dynamical system of the form (17). Assume

that c 0 =&1, c 1 =0. Suppose that the homogeneous parts of degree 2 in f 1
and f 2 are zero.
 REDUCTIONS OF NORMAL FORMS 97

(a) If # 1 = 13 (; 2, 1 + : 1, 2 +3 ; 0, 3 +3 : 3, 0 ){0, then a normal form of

order 9 is

x* 1 =x 2 +# 1 x 31 +# 2 x 51 +O(&x& 10 ),
x* 2 =&x 1 +# 3 x 31 +O(&x& 10 ),

(b) If # 1 =0 and #$3 = 13(&3 : 0, 3 +3 ; 3, 0 + ; 1, 2 & : 2, 1 ){0, then a

non-degenerate normal form of order 9 is

x* 1 =x 2 +#$1 x 51 +#$2 x 71 +#$4 x 91 +O(&x& 10 ),

x* 2 =&x 1 +#$3 x 31 +O(&x& 10 ),

where # j and #$j are parameters depending only on the coefficients of F.

Moreover, the above normal forms are optimal or unique in the given order
in the same sense as in Proposition 3.1.
In case (b) we obtain some non-degeneracy conditions such as #$1 {0.
The parameters # j , #$j can be given by explicit formulae from the coefficients
of F. For example, #$1 = 15 : 1, 4 + 15 : 3, 2 +: 5, 0 + 15 ; 4, 1 + 15 ; 2, 3 +; 0, 5 & 15 : 1, 2 : 2, 1
& 15 : 2, 1 ; 0, 3 & 15 : 1, 2 ; 1, 2 & 35 ; 0, 3 : 0, 3 & 15 : 1, 2 : 0, 3 & 25 : 3, 0 : 2, 1 & 25 : 3, 0 ; 1, 2 &
1 3 1
5 ; 1, 2 ; 0, 3 & 5 ; 0, 3 ; 3, 0 & 5 : 1, 2 ; 3, 0 .
To obtain the normal forms of the proposition, we apply the algorithm
t
given in the above paragraph. For k=3, with Z (3) 0 =(U 1 , U 2 ) , we have

U 2 +: 0, 3 U 1 + 12 ; 0, 3 + 12 : 1, 2
Z (3)
1 =
_ &U 1 +; 0, 3
,
& Z (3)
2 =
_ &
U 2 & 12 : 0, 3 + 12 ; 1, 2
,

U 2 + 13 : 2, 1 + 12 : 0, 3 + 16 ; 1, 2
Z (3)
3 =
_&U 1 + 13 ; 2, 1 + 12 ; 0, 3 & 16 : 1, 2
,
&
: + 13 ; 2, 1 +; 0, 3 + 13 : 1, 2
Z (3)
4 =
_; 3, 0
1 1
3, 0 & 3 : 2, 1 &: 0, 3 + 3 ; 1, 2 & .

One then has in case (a) that (# 1 , # 3 ) t =Z (3) 3

4 . In this case, C is of dimen-
sion 2 spanned by

x 31 0
\ +
0
and
\x + .
3
1

For k=4 we find that C 4 =[0]. And we continue with k=5. If # 1 {0,
Theorem 2.2 is applicable for l=3 and k=5 as stated in the proposition.
98 CHEN AND DELLA DORA

One can continue the algorithm with k=6, 7, ... . The computations are
similar to those above. We obtain by computations in Maple the normal
forms of the orders given in the proposition.
We give in the following table, for comparison, the numbers of non-zero
parameters in the normal forms, as for the preceding example. In [22]
Ushiki obtained a normal form of order 7 in this case. For a normal form
of general order, Baider gave partial results in this semisimple case in [3].
Complete reduction is obtained by Chua and Kokubu, by Ushiki's method
in [13] for case (a) and by Gaeta in [17] for a more general case. Note
also that in this case Gaeta's normal form is the best possible. But these
methods compute only a general normal form of dynamical systems with
the given linear part. They are not concerned with the computations of a
formal diffeomorphism that realizes the normalization. One can compute
by our method at the same time a normal form and a normalization
diffeomorphism.

degree 2nd 3rd 4th 5th 6th 7th 8th 9th

Poincare 0 2 0 2 0 2 0 2
Takens 0 2 0 2 0 2 0
Ushiki 0 2 0 1 0 0
Gaeta 0 2 0 1 0 0 0 0
Case (a) 0 2 0 1 0 0 0 0
Case (b) 0 1 0 1 0 1 0 1

4. DYNAMICAL SYSTEMS OF DIMENSION 3

Consider dynamical systems of dimension 3 of the form

x 2 + f 1 (x)
dx
dt
=F(x)=
\ x 2 + f 2 (x)
' 1 x 1 +' 2 x 2 +' 3 x 3 + f 3 (x) + (19)

where the matrix of the linear part is

0 1 0
A= 0
\ +
'1
0
' 2 '3
1
 REDUCTIONS OF NORMAL FORMS 99

with ' j # K. Write for k2, q=(q 1 , q 2 , q 3 ),

F k(x)=

: :q xq : :$q y q : :q yq

\ + \ + \ +
|q| =k |q| =k |q| =k

: ;q xq , G k ( y)= : ;$q y q , . k( y)= : bq yq ,

|q| =k |q| =k |q| =k

: #q xq : #$q x q : cq yq
|q| =k |q| =k |q| =k

where : q , ; q , # q # K and where a q , b q , c q , :$q , ;$q , #$q # K are to be deter-

mined. Let

aq :q :$q

\+ \+ \+
Mq = b q ,
cq
Fq= ;q ,
#q
G q = ;$q .
#$q

Then the homological equations (7) in dimension 3 can be written as:

(q 1 +1) Mq+e1 &e2 +(q 2 +1) Mq+e2 &e3 +' 1(q 3 +1) Mq&e1 +e3
+' 2(q 3 +1) Mq&e2 +e3 +(' 3 q 3 I&A) Mq =F q &G q

where I denotes the identity matrix of order 3. Let q 1 =k&i& j, q 2 =i,

q 3 = j, then

Z ij =M(k&i& j, i, j) , 4 i, j =F q , and 4$i, j =G q .

We then have

(i+1) Z i+1, j&1 +(k&i& j+1) Z i&1, j +' 1( j+1) Z i, j+1

+' 2( j+1) Z i&1, j+1 +(' 3 jI&A) Z i, j =4 i, j &4$i, j .

Let Z 0, j (0 jk) be given arbitrarily. One can determine Z i+1, j&1 for
0ik&1, 1 jk, and i+ jk such that 4$i, j =0. In fact,

(i+1) Z i+1, j&1 =4 i, j &(k&i& j+1) Z i&1, j &' 1( j+1) Z i, j+1

& ' 2( j+1) Z i&1, j+1 &(' 3 jI&A) Z i, j .

The remaining equations are (with j=0)

4$i, 0 =4 i, 0 &(k&i+1) Z i&1, 0 &' 1 Z i, 1 &' 2 Z i&1, 1 +AZ i, 0 .

100 CHEN AND DELLA DORA

Since we have determined all Z i, j for i1 as functions of Z 0, j , these equa-

tions are linear equations for Z 0, j which may be solved in some cases to
make some of the 4$i, 0 zero.

Example 4.1. Consider now a dynamical system of the form (19) with

0 1 0
A= 0
\ +
0 &1 0
0 1 (20)

as its linear part.

With notations as above, for k=2 one can compute Z 1, 1 , Z 1, 0 , Z 2, 0 to

get

Z 0, 02 +: 2, 0, 0
4$0, 0 =
_ Z 0, 03 +; 2, 0, 0
&Z 0, 02 +# 2, 0, 0 & ,

&2 Z 0, 01 +Z 0, 11 +Z 0, 13 +; 1, 0, 1 +: 1, 1, 0
4$1, 0 =
_ &2 Z 0, 02 +# 1, 0, 1 +; 1, 1, 0
&2 Z 0, 03 &; 1, 0, 1 +# 1, 1, 0 & ,

& 32 Z 0, 12 &: 1, 0, 1 + 32 Z 0, 22 +: 0, 0, 2 + 12 # 0, 0, 2 + 12 ; 0, 1, 1 +: 0, 2, 0
4$2, 0 =
_ & 32 Z 0, 13 &; 1, 0, 1 + 32 Z 0, 23 + 12 ; 0, 0, 2 + 12 # 0, 1, 1 +; 0, 2, 0
3
2 Z 0, 12 &# 1, 0, 1 & 32 Z 0, 22 + 12 # 0, 0, 2 & 12 ; 0, 1, 1 +# 0, 2, 0 &
where we have denoted by Z 0, 1i the i th element of Z 0, 1 . After resolution for
Z 0, 0 by making 4$1, 0 =0 and resolution of the last two elements of Z 0, 1 by
making the first two elements of 4$2, 0 equal zero, we obtain

1
#
2 1, 0, 1 + 12 ; 1, 1, 0 +: 2, 0, 0 0

_ 1
4$0, 0 = & ; 1, 0, 1 + #
2
1
2 1, 1, 0

&
+; 2, 0, 0 ,
& 12 # 1, 0, 1 & 12 ; 1, 1, 0 +# 2, 0, 0 _&
4$1, 0 = 0
0

and

4$2, 0 =(0, 0, # 0, 0, 2 &: 1, 0, 1 +: 0, 0, 2 +: 0, 2, 0 &# 1, 0, 1 +# 0, 2, 0 ) t.

 REDUCTIONS OF NORMAL FORMS 101

We obtain then a new system whose homogeneous part of degree 2

contains four parameters that may be nonzero:

x* 1 =x 2 ++ 1 x 21 +O(&x& 3 ),
x* 2 =x 3 ++ 2 x 21 +O(&x& 3 ), (21)
x* 3 =&x 2 ++ 3 x 21 ++ 4 x 22 +O(&x& 3 ),

where (+ 1 , + 2 , + 3 ) t =4$0, 0 and + 4 is the nonzero element of 4$2, 0 .

Let

u v w

\
P= 0 u&w
0 &v u&w
v
+ (22)

with u, v, w # K such that det P{0, then P &1AP=A. The linear transfor-
mation x=Py makes no change on the linear part of the system. If + 1 {0
then, by taking u=1+ 1 and making the linear transformation (22) on the
system (21), + 1 is reduced to 1. We suppose that this is done to simplify
the notation. The above algorithm applied to the new system leads to the
normal form

x* 1 =x 2 +x 21 +O(&x& 3 ),
x* 2 =x 3 ++ 2 x 21 +O(&x& 3 ),
x* 3 =&x 2 ++ 3 x 21 ++$4 x 22 +O(&x& 3 ),

where +$4 =v 2+ 4 +v 2 +v 2+ 3 &2w++ 4 w 2 +w 2 ++ 4 &2+ 4 w&2w+ 3 +w 2+ 3 .

It is clear that if one looks for a normal form in C or in an algebraic exten-
sion of K then one can generically reduce +$4 to zero by choosing a solution
for v or w. This eliminates one more parameter in degree 2. One can now
give the following example.

Example 4.2. The two dynamical systems

x* 1 =x 2 +x 21 , x* 1 =x 2 +x 21 ,
x* 2 =x 3 +x 21 , x* 2 =x 3 +x 21 ,
x* 3 =&x 2 +x 21 &4x 22 x* 3 =&x 2 +x 21

are not equivalent in R but are equivalent in C up to order 2 (with respect

to equivalence by diffeomorphisms).
102 CHEN AND DELLA DORA

We take v=w=0 in the following, since we are looking for rational nor-
mal forms. We study further reductions of systems with the given linear
part by near identity transformations.
If there are undetermined elements in Z 0, j for the computations of a nor-
mal form of a certain order, then it may be chosen for the reduction of
higher order normal forms. The above algorithm is implemented in Maple
V. We obtain a normal form of the fourth order that we state in the follow-
ing proposition.

Proposition 4.1. Consider a dynamical system of dimension 3 of the

form (19) with the matrix (20) as its linear part. Let notations be as above.
If + 1 {0, then a non-degenerate normal form of order 4 is

x* 1 =x 2 +x 21 ++ 5 x 31 +O(&x& 5 ),

x* 2 =x 3 ++ 2 x 21 ++ 6 x 31 +O(&x& 5 ),

x* 3 =&x 2 ++ 3 x 21 ++ 4 x 22 ++ 7 x 31 ++ 8 x 41 +O(&x& 5 ).

Moreover, the above normal forms are optimal or unique in the given order
in the same sense as in Proposition 3.1.

The following table gives a comparison for the number of non-zero

parameters remaining in the normal forms (see also [13, p. 60]).

Degree 2nd 3rd 4th 5th

Poincare 4 6 7 9
Takens 4 6 7 9
Ushiki 4 3 4 2
Our method 4 3 1

Example 4.3. Consider a dynamical system of the form (19) with the
nilpotent matrix

0 1 0
A= 0
\ + 0
0 1
0 0
(23)

as its linear part.

 REDUCTIONS OF NORMAL FORMS 103

With the above notations we obtain for k=2,

Z 0, 02 +: 2, 0, 0 &2Z 0, 01 +Z 0, 13 +; 1, 0, 1 +: 1, 1, 0

_
4$0, 0 = Z 0, 03 +; 2, 0, 0 ,
# 2, 0, 0 & 4$1, 0 =
_ &2Z 0, 02 +# 1, 0, 1 +; 1, 1, 0
&2Z 0, 03 +# 1, 1, 0
& 32 Z 0, 12 &: 1, 0, 1 + 12 # 0, 0, 2 + 12 ; 0, 1, 1 +: 0, 2, 0
& ,

4$2, 0 =
_ & 32 Z 0, 13 &; 1, 0, 1 + 12 # 0, 1, 1 +; 0, 2, 0
&# 1, 0, 1 +# 0, 2, 0 &
where Z 0, 1i denotes the i th element in Z 0, 1 . After resolution we obtain
finally
1
#
2 1, 0, 1 + 12 ; 1, 1, 0 +: 2, 0, 0
4$0, 0 =
_0
1
#
2 1, 1, 0 +; 2, 0, 0
# 2, 0, 0 &
0
,

4$1, 0 = 0 ,
_&
0
4$2, 0 =
_ 0
&# 1, 0, 1 +# 0, 2, 0 & .

The new system is then

x* 1 =x 2 ++ 1 x 21 +O(&x& 3 ),
x* 2 =x 3 ++ 2 x 21 +O(&x& 3 ), (24)
x* 3 =+ 3 x 21 ++ 4 x 22 +O(&x& 3 ),

where four parameters

+ 1 = 12 # 1, 0, 1 + 12 ; 1, 1, 0 +: 2, 0, 0 , + 2 = 12 # 1, 1, 0 +; 2, 0, 0 ,
+ 3 =# 2, 0, 0 , and + 4 =&# 1, 0, 1 +# 0, 2, 0

may be nonzero. Let

u v w
P= 0
\ +
0
u
0
v
u
(25)
104 CHEN AND DELLA DORA

with u, v, w # K, u{0, then P &1AP=A. The linear transformation x=Py

makes no change on the linear part of the system. If + 3 {0 then we take
u=1+ 3 and make the linear transformation (25) on the system (24). The
above algorithm applied to the new system leads to the normal form

+1 2
x* 1 =x 2 + x +O(&x& 3 ),
+3 1

+2 2
x* 2 =x 3 + x +O(&x& 3 ),
+3 1

&2+ 23 w++ 4 ++ 33 v 2 2
x* 3 =x 21 + x 2 +O(&x& 3 ).
+3

Then one can choose w=(+ 4 ++ 33 v 2 )(2+ 23 ) to eliminate one more term. We
obtain non-degenerate fourth order normal forms which we state in the
following proposition:

Proposition 4.2. Consider a dynamical system of dimension 3 of the

form (19) with the nilpotent matrix (23) as its linear part. Let the notation
be as above.

(a) If # 2, 0, 0 =1, then a non-degenerate normal form of order 4 is

x* 1 =x 2 ++ 1 x 21 ++ 4 x 31 +O(&x& 5 ),

x* 2 =x 3 ++ 2 x 21 ++ 5 x 31 ++ 7 x 41 +O(&x& 5 ),

x* 3 =x 21 ++ 6 x 31 ++ 8 x 41 +O(&x& 5 ).

(b) If #2, 0, 0 =0, +2 {0, then a non-degenerate normal form of order 4 is

x* 1 =x 2 ++ 1 x 21 ++$4 x 31 ++$7 x 41 +O(&x& 5 ),

x* 2 =x 3 +x 21 ++$8 x 41 +O(&x& 5 ),

x* 3 =+$3 x 22 ++$5 x 31 ++$6 x 1 x 22 ++$9 x 41 ++$10 x 21 x 22 +O(&x& 5 )

where + j and + $j are parameters depending on the coefficients of the system.

Moreover, the above normal forms are optimal or unique in the given order
in the same sense as in Proposition 3.1.

The non-degeneracy conditions are non-nullity conditions of some

polynomial expressions in some of the + j .
 REDUCTIONS OF NORMAL FORMS 105

Note that Ushiki (see [22, 13]) obtained non-degenerate normal forms
of order 3 in this nilpotent case. Cushman and Sanders' normal form in this
case is

0 x1 x2
f 1 ( p 1 , p 2 )
\+
0 + f 2 ( p 1 , p 2 )
0 \+ x 2 + f 3 ( p 1 , p 2 )
x3 \+
x3 ,
0

where p 1 =x 3 , p 2 =x 22 &2x 1 x 3 , and f i are formal power series (with leading

matrix A t ). Elphick et al. method leads to the same normal form. This is
a general normal form but it contains as many non-zero parameters as in
the normal form of Takens. One should also note that p 2 is not a
monomial. Note that in [9] we obtained also a normal form by using a
Jordan basis in H k . The normal form up to order 5 defined in [9] contains
the same number of nonzero parameters as that of Takens.
The following table gives a comparison for the number of nonzero
parameters remaining in the normal forms (see also [13, p. 60]), in the
non-degenerate case (a).
Degree 2nd 3rd 4th

Takens 4 6 7
Ushiki 3 3
Our method 3 3 2
Our method can be used to find degeneracy conditions and to compute
the corresponding degenerate normal forms. We compute at the same time
a formal diffeomorphism that realizes the normalization.

ACKNOWLEDGMENTS

The authors express their thanks to the referee for useful suggestions.

REFERENCES

1. V. I. Arnold, ``Geometrical Methods in the Theory of Ordinary Differential Equations,''

Springer-Verlag, New York, 1983.
2. M. Ashkenazi and S. N. Chow, Normal forms near critical points for differential equations
and maps, IEEE Trans. Circuits and Systems 35 (1988), 850862.
3. A. Baider, Unique normal forms for vector fields and hamiltonians, J. Differential
Equations 78 (1989), 3352.
4. A. Baider and J. A. Sanders, Further reductions of the TakensBogdanov normal form,
J. Differential Equations 99 (1992), 205244.
106 CHEN AND DELLA DORA

5. A. D. Bruno, ``Local Method of Nonlinear Analysis of Differential Equations,'' Springer-

Verlag, New YorkBerlin, 1979.
6. T. Carleman, Application de la theorie des equations integrales lineaires aux systemes
d'equations differentielles nonlineaires, Acta Math. 59 (1932), 6368.
7. G. Chen, Computing the normal forms of matrices depending on parameters, in
``Proceedings of ISSAC89,'' pp. 244249, ACM PressAddisonWesley, Portland, OR,
1989.
8. G. Chen and J. Della Dora, Rational normal form for dynamical systems via Carleman
linearization, in ``Proceedings of ISSAC99,'' pp. 165172, ACM PressAddisonWesley,
Vancouver, 1999.
9. G. Chen and J. Della Dora, Nilpotent normal form for systems of nonlinear differential
equations: Algorithm and examples, Rapport de Recherche RR 838 M, Universite de
Grenoble 1, Grenoble, France, Fevrier 1991.
10. G. Chen, J. Della Dora, and L. Stolovitch, Nilpotent normal form via Carleman lineariza-
tion (for systems of nonlinear differential equations), in ``Proceedings of ISSAC91,''
pp. 281288, ACM PressAddisonWesley, Bonn, 1991.
11. S. N. Chow and J. K. Hale, ``Methods of Bifurcation Theory,'' Springer-Verlag, New
York, 1982.
12. S. N. Chow, C. Li, and D. Wang, ``Normal Forms and Bifurcation of Planar Vector
Fields,'' Cambridge Univ. Press, Cambridge, UK, 1994.
13. L. O. Chua and H. Kokubu, Normal forms for nonlinear vector fields. II. Applications,
IEEE Trans. Circuits and Systems 36 (1989), 5170.
14. R. Cushman and J. Sanders, Nilpotent normal forms and representation theory of sl 2 (R),
in ``Multiparameter Bifurcation Theory'' (M. Golubitsky and J. Guckenheimer, Eds.),
Contemporary Mathematics, Vol. 56, pp. 3135, American Math. Society, Providence, RI,
1986.
15. C. Elphick, E. Tirapegui, M. Brachet, P. Coulet, and G. Iooss, A simple characterization
for normal forms of singular vector fields, Physica D 29 (1987), 95127.
16. G. Gaeta, Reduction of Poincare normal forms, Lett. Math. Phys. 42 (1997), 103114.
17. G. Gaeta, Poincare renormalized forms, Ann. Inst. H. Poincare 70 (1999), 461514.
18. J. Martinet and J. P. Ramis, Classification analytique des equations differentielles non
lineaires resonantes du premier ordre, Ann. Sci. Ecole. Norm. Sup. 4 16 (1983), 571621.
19. H. Poincare, Notes sur les proprietes des fonctions definies par des equations differen-
tielles, J. Ecole Polytechn. 45 (1878), 1326.
20. F. Takens, Singularities of vector fields, Publ. Math. Inst. Hautes Etudes Sci. 43 (1974),
47100.
21. C. A. Tsiligiannis and G. Lyberatos, Normal forms, resonance and bifurcation analysis via
the Carleman linearization, J. Math. Anal. Appl. 139 (1989), 123138.
22. S. Ushiki, Normal forms for singularities of vector fields, Jpn. J. Appl. Math. 1 (1984),
134.