Weakly nonlinear oscillations: A perturbative approach

Peter B. Kahn
Department of Physics and Astronomy, State University of New York at Stony Brook,
Stony Brook, New York 11794
Yair Zarmia)
Department of Solar Energy and Environmental Physics, The Jacob Blaustein Institute for Desert Research,
Ben-Gurion University of the Negev, Sede-Boqer Campus, 84990 Israel
共Received 15 September 2003; accepted 23 December 2003兲
The perturbative analysis of a one-dimensional harmonic oscillator subject to a small nonlinear
perturbation is developed within the framework of two popular methods: normal forms and multiple
time scales. The systems analyzed are the Duffing oscillator, an energy conserving oscillatory
system, the cubically damped oscillator, a system that exhibits damped oscillations, and the Van der
Pol oscillator, which represents limit-cycle systems. Special emphasis is given to the exploitation of
the freedom inherent in the calculation of the higher-order terms in the expansion and to the
comparison of the application of the two methods to the three systems. © 2004 American Association
of Physics Teachers.
关DOI: 10.1119/1.1648687兴

I. INTRODUCTION only on time 共so that its governing equation is an ordinary

differential equation兲 and for which the unperturbed equation
An approximate linear model successfully describes many is linear.
physical systems. However, nonlinear phenomena are en- Numerous perturbation schemes have been developed for
countered in all areas of the sciences. Almost always, non- the analysis of systems that are affected by a small nonlinear
linear systems have fundamental aspects that cannot be faith- perturbation. Each method is tailored to address the ques-
fully treated by a linear approximation. Perhaps most tions asked and the desired form of the answers. In this ar-
important is the loss of the principle of superposition which ticle, we follow two fundamentally different approaches. The
plays a critical role in linear systems. first is the method of normal forms,1–10,17 which has a firm
The behavior of nonlinear systems is much richer than that mathematical foundation. Here one introduces a near identity
of linear systems. Unlike linear differential equations, the transformation from the original set of variables to a new set
singularity structure of the solution of a nonlinear differential whose equations are more suitable for analysis. The second
equation is affected by the initial conditions. As a result, a method is an ad hoc perturbation scheme known as the
systematic classification of the solutions based on the struc- method of multiple time scales.11–16 It has a tenuous math-
ture of the differential equations is not generally possible. ematical foundation, but has been applied with much more
In many cases, a simple closed form expression for the success to many types of equations, including perturbed os-
solution cannot be found, as is the case for chaotic systems, cillations, boundary-layer problems, and many partial differ-
especially if their long-time behavior is sought. Even the ential equations.23–26 The idea is to introduce independent
analysis of nonchaotic systems does not lead to simple time scales that characterize different aspects of the solution.
closed-form expressions for the solutions in most cases. Dy- We will show how the freedom in the perturbation expansion
namical systems that can be solved by relatively simple allows us to address the imposition of consistency conditions
that arise in the expansion. In many cases, the methods of
means and to which a small perturbation has been added are
normal forms and multiple time scales are equivalent, and
the most manageable. In such situations, a perturbation ex-
the method of choice depends on the application and on
pansion based on the solution of the unperturbed problem is
one’s inclination. However, open problems exist in both
used to obtain an approximate solution to the full problem.
methods, some of which will be discussed.
Historically, the first perturbed systems that were analyzed
The O symbol is used frequently in the context of pertur-
were those for which the dynamical variables are functions
bation expansions. A quantity, g, that depends on a param-
of time only. Moreover, these are systems for which the un-
eter, ␭, is said to be O(␭) if some positive constant, C, exists
perturbed equations are linear in the dynamical variables.
Simple examples are harmonic oscillators and systems that such that 兩 g(␭) 兩 ⭐C 兩 ␭ 兩 .
decay exponentially.1–17 The methods developed for the We analyze three systems. Despite their simplicity, the
analysis of these systems were later applied to more complex analysis reveals much of what may be encountered in the
dynamical systems of interest in fluid dynamics, plasma application of the two methods to more complex systems.
physics, reaction-diffusion systems, and population Some of the richness as well as the obstacles that perturba-
ecology.18 –26 In many of these latter systems, the dynamical tion theory generates are not evident to first order. Therefore,
variables depend both on time and position, so that they are the analysis is carried at least through second order.
governed by partial differential equations, making the analy- 共1兲 The Duffing oscillator is defined by
sis much more complicated. Because the purpose of this pa-
per is to serve as an introductory tutorial for perturbation
d 2x
methods, we will focus on simpler systems that are described
2
⫹x⫹ ⑀ x 3 ⫽0, 0⬍ ⑀ Ⰶ1, 共1兲
by a single dynamical variable, which is assumed to depend dt

538 Am. J. Phys. 72 共4兲, April 2004 http://aapt.org/ajp © 2004 American Association of Physics Teachers 538
where x(0)⫽A and dx/dt(0)⫽0. Equation 共1兲 provides an II. INTRODUCING THE STANDARD FORM
approximate model for anharmonic effects in crystal vibra-
tions. For ⑀⬍0, it can be viewed as a truncated version of the We first convert our second-order differential equations
time dependence of the angular displacement, ␪, of the pen- into two-dimensional vector equations:
dulum:

d 2␪
d x
dt
冉冊 冉
y
⫽
0
⫺1
1
0
冊冉 冊 冉
x
y
⫹⑀
0
F 共 x,y 兲
, 冊 共5兲

L ⫹g sin ␪ ⫽0, 共2兲 where y⫽dx/dt and

再
d␶2
⫺x 3 共 Duffing oscillator兲
where L is the length of the pendulum and g is the accelera-
F⫽ ⫺ẋ 3
共 Damped oscillator兲 . 共6兲
tion due to gravity. For small oscillation amplitudes, Eq. 共2兲
has the same form as Eq. 共1兲 if only the first two terms of the ẋ 共 1⫺x 兲 2
共 Van der Pol oscillator兲
Taylor series expansion of sin ␪ are kept. Although Eq. 共2兲 is
It is not essential, but it is traditional to diagonalize the ma-
solved by elliptic integrals, we exploit its truncated version
trix associated with the linear problem. The diagonalization
to show how approximate solutions are generated in an ex-
is accomplished by introducing z⫽x⫹iy and its complex
pansion scheme.
For an understanding of perturbative expansion methods, conjugate, z * ⫽x⫺iy. Equation 共5兲 becomes
there is no gain in including higher-order terms in the Taylor dz
series in Eq. 共1兲.11,17 Because Eq. 共1兲 describes a conserva- ⫽⫺iz⫹i ⑀ F 共 z,z * 兲 . 共7兲
dt
tive system, the trajectory in the phase space formed by x and
dx/dt is a closed curve. Note that ⑀ A 2 , which is a measure
of the perturbation to the energy of the system, is the true III. THE METHOD OF NORMAL FORMS
small parameter, not ⑀. We first introduce the near-identity transformation and
共2兲 A linear oscillator with a small cubic damping pertur- write the complex variable z in terms of the zero-order ap-
bation:
proximation, z 0 :
d 2x
dt 2
⫹x⫹ ⑀
dx
dt 再 冎 3
⫽0 共3兲 z⫽z 0 ⫹ 兺
n⭓1
⑀ n z n 共 z 0 ,z *
0 兲, z * ⫽z 0* ⫹ 兺n ⑀ n z *n 共 z 0 ,z 0* 兲 .
共8兲
where 0⬍⑀Ⰶ1. We discuss cubic damping for two reasons. We need only solve the z equation as the equation for z * is
First, linear damping does not generate the richness of phe- its complex conjugate. The dynamical equation obeyed by
nomena generated by nonlinear perturbations. Second, the z 0 , the normal form, is assumed to have the form
interesting effects of quadratic damping appear only in
O( ⑀ 2 ), whereas they emerge in the first-order analysis of the dz 0
cubic damping case.11,17 Hence, the latter is a good example dt
⫽U 0 ⫹
n⭓1
兺
⑀ n U n 共 z 0 ,z 0* 兲 , 共9兲
for illustrating features of perturbation theory.
共3兲 The Van der Pol oscillator:27 where U 0 ⫽⫺iz 0 .
The following points provide some perspective and moti-
vation for the development of the method.
d 2x
⫹x⫽ ⑀ ẋ 共 1⫺x 2 兲 , 共4兲 共1兲 The zero-order approximation, z 0 , need not be re-
dt 2 stricted to obey the unperturbed equation. Rather, it is as-
sumed to be affected by the perturbation parameter, ⑀.
where 0⬍⑀Ⰶ1. Equation 共4兲 was introduced to model an 共2兲 The quantity U 0 is the unperturbed operator; the
electric circuit that contains a nonlinear component called a higher-order terms in the normal form, U n , update the time
triode, and is equivalent to that developed by Lord Rayleigh dependence of z 0 due to the presence of the perturbation.
to model self-sustained sound vibrations.28 The solution of 共3兲 The higher-order terms in the expansion of the solu-
Eq. 共4兲 evolves into a limit cycle, an aspect of nonlinear
tion, z n , depend on z 0 and z *
0 .
oscillations that has no parallel in linear oscillatory system,
and occurs in many dissipative nonlinear systems, such as 共4兲 The key role of the near-identity transformation is to
the onset of convective rolls in a fluid.23 A limit cycle is a exploit z n so that as many nonlinear terms are included as
closed orbit in phase space that is approached asymptotically possible, yet ensuring that the normal form is solvable and
from nearby orbits. For small amplitude, x, the right-hand incorporates the major effect of the perturbation already in
side of Eq. 共4兲 has the same sign as the velocity, and repre- z0 .
sents a force that pumps energy into the system, causing the 共5兲 An obvious choice for the structure of z n is to mimic
amplitude to increase. As the displacement exceeds unity, the the form of the perturbation, leading to a set of linear equa-
dissipative term changes sign, leading to deceleration, and, tions for the desired coefficients. With the normal form
eventually, a halt to amplitude growth. The amplitude then solved for z 0 , the z n are then found and an approximate
decreases and falls below unity, where, once again, the dis- solution is obtained to the desired order of the expansion.
sipative term feeds energy into the system, leading to a re- 共6兲 Central to our approach is the exploitation of the free-
peat of the behavior. For t→⬁, the solution tends to the limit dom in the structure of the z n 共see Refs. 29–38兲 to modify
cycle. The zero-order approximation to the solution is called aspects of the expansion or to satisfy desired constraints.17
the limit circle. The limit cycle is the limit circle modified by 共7兲 The method of normal forms provides an approxima-
higher-order corrections. tion to the exact 共but unknown兲 solution, z(t). The error,

539 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 539
which is the deviation between the exact and approximate
solutions, remains bounded for a limited time span. If we
truncate the near-identity transformation at a finite order,
z 1⫽ ␳ 3 再 1 ⫺3it 3 ⫺it 3 it 1 3it
16
e ⫹ te ⫺ e ⫺ e ⫹ ␣ e ⫺it .
8 16 32 冎
共17兲
thus generating an approximation z a (t),
N The last term in the solution in Eq. 共17兲 is the general
z a ⫽z 0 ⫹ 兺
n⫽1
⑀ z n 共 z 0 ,z *
n
0 兲, 共10兲 solution of the homogeneous part of Eq. 共16兲; its coefficient
is undetermined. The second term on the right-hand side of
Eq. 共16兲 resonates with the differential operator, (ż 1 ⫹iz 1 ),
and, concurrently, the normal form, Eq. 共9兲 is calculated
and hence is called a resonant term. It generates a contribu-
through order N⫹1 共that is, one order beyond that included
tion to z 1 that is linear in t Eq. 共17兲. This contribution, called
in the near-identity transformation兲, then it can be proven
that the error incurred is a secular term, limits the validity of the expansion to short
times of O(1). The reason is that the contribution of the
兩 z 共 t 兲 ⫺z a 共 t 兲 兩 ⫽O 共 ⑀ N⫹1 兲 , 共11兲 first-order correction to the solution, ⑀ z 1 , contains a term
for a time span that is at most of order 共1/⑀兲 关the common that is proportional to ⑀ t. Hence, for ⑀ z 1 to be of magnitude
notation is t⫽O(1/⑀ )]. ⑀ relative to the zero-order approximation, z 0 , t must be
Rather then going through the detailed proof 共see, for ex- O(1). As t grows to O(1/⑀ ), this term spoils the validity of
ample, Ref. 17兲, we provide an intuitive explanation. The the perturbative expansion. Secular terms also appear in the
approximation, z a (t), is computed through O( ⑀ N ). Thus, higher orders.
seemingly, the error ought to be O( ⑀ N⫹1 ). However, the Terms that are linear in t also violate a requirement that is
approximate time dependence, even of the zero-order term, is based on physical intuition. The original system is conserva-
an independent source of error. To be specific, suppose that tive. Hence, its solution must be a periodic function of time.
z 0 (t) is proportional to cos ␻t. If the correct value of ␻ is This behavior does not mean that the perturbative approxi-
replaced by an approximate value, ␻ a , then mation must be periodic in time as well, but such behavior is

冏 冉 冊 冏冏 冉 冊冏
preferable. The avoidance of secular terms and the construc-
␻⫹␻a ␻⫺␻a tion of the approximate solution from only periodic functions
兩 cos ␻ t⫺cos ␻ a t 兩 ⫽2 sin t sin t . work hand in hand and guarantee that the expansion remains
2 2
共12兲 valid for longer times, of O(1/⑀ ). As indicated by the dis-
cussion involving Eqs. 共10兲–共12兲, the error incurred by ap-
If the normal form has been computed through O( ⑀ N⫹1 ), the proximating the full solution by a first-order approximation
method yields ( ␻ ⫺ ␻ a )⫽O( ⑀ N⫹2 ). If t is not too large, the is
second sine function in Eq. 共12兲 can be approximated by its
argument, which is O( ⑀ N⫹2 t), and, hence, the error is 兩 z 共 t 兲 ⫺ 兵 z0 共 t 兲 ⫹ ⑀ z 1 共 z 0 ,z 0* 兲 其 兩 ⫽O 共 ⑀ 兲 共18兲
O( ⑀ N⫹1 ) only for t⫽O(1/⑀ ).
for t⫽O(1/⑀ ). If we substitute Eqs. 共8兲 and 共9兲 in the left-
hand side of Eq. 共13兲 and expand the result through O( ⑀ 2 ),
IV. APPLICATIONS OF THE METHOD OF NORMAL we obtain:

冉 冊
FORMS
dz dz0 ⳵ z 1 dz0 ⳵ z 1 dz*
0
⫽ ⫹⑀ ⫹
A. The Duffing oscillator dt dt ⳵ z0 dt ⳵ z0* dt

In terms of the complex variables z and z * , we can write

Eq. 共1兲 as ⫹⑀2 冉 ⳵ z 2 dz0 ⳵ z 2 dz0*
⳵ z0 dt
⫹
⳵ z*
0
dt 冊
冉 冊
共 z⫹z * 兲 3
ż⫽⫺iz⫺i ⑀ . 共13兲 ⳵z1 ⳵z1
8 ⫽ 兵 U 0⫹ ⑀ U 1⫹ ⑀ 2U 2其 ⫹ ⑀ U 0⫹ U*
⳵ z0 ⳵ z*
0
0
We begin by demonstrating why the method of normal forms
is necessary. Suppose that we ignore the normal form, Eq.
共9兲, and substitute the near-identity transformation, Eq. 共8兲,
in Eq. 共13兲. This procedure is called the naive expansion.
⫹⑀2 冉 ⳵z2
U ⫹
⳵z2
U* ⫹⑀2
⳵ z0 0 ⳵ z0* 0
⳵z1
U ⫹
⳵z1
冊 冉
U* .
⳵ z0 1 ⳵ z0* 1 冊
Because a problem already will arise in first-order, we write 共19兲
the result only through O( ⑀ ):
Next we substitute the expansion for z given by Eq. 共8兲 into
共 z 0 ⫹z *
0兲
3
ż 0 ⫹ ⑀ ż 1 ⫽⫺i 共 z 0 ⫹ ⑀ z 1 兲 ⫺i ⑀ . 共14兲 the right-hand side of Eq. 共13兲:
8
dz 共 z 0 ⫹z 0* 兲 3
We compare terms order by order. In O( ⑀ 0 ) we have ⫽⫺i 兵 z 0 ⫹ ⑀ z 1 ⫹ ⑀ z 2 其 ⫺i ⑀
2
dt 8
ż 0 ⫹iz 0 ⫽0, 共15兲
3 共 z 0 ⫹z 0* 兲 2
which implies that z 0 ⫽ ␳ e ⫺it . In O( ⑀ 1 ) we have ⫺i ⑀ 2 共 z 1 ⫹z *
1 兲. 共20兲
8
⫺3it ⫺it
共e ⫹3e ⫹3e ⫹e
it 3it
兲
ż 1 ⫹iz 1 ⫽⫺i ⑀ ␳ 3 共16兲 We equate Eqs. 共19兲 and 共20兲, collect the ⑀ 0 terms, and find
8
and thus U 0 ⫽⫺iz 0 , 共21兲

540 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 540
which is the eigenvalue equation for the unperturbed prob- ⫹ 163 ␣ 兲 z 40 z *
z 2 ⫽ 1024 z 0 ⫹ 共 ⫺ 256 0 ⫹ 共 512 ⫺ 16 ␣ ⫺ 8 ␣ * 兲 z 0 z 0*
3 15 69 3 3 2 3
lem. We then solve for each order in the small parameter ⑀, a
series of equations for the U i . To O( ⑀ ), we have: ⫺ 323 ␣ * 兲 z 0 z 0* 4 ⫺ 512
⫹ 共 1024 0 ⫹ ␤ z 0z * 共29兲
21 1 5 3 2
z* 0 ,

再
U 1 ⫽ ⫺iz 1 ⫺U 0
⳵z1
⳵z0
⫺U *
0
⳵z1
⳵z*
0
1
8 冎
⫺i 共 z 0 ⫹z *
0兲 .
3
共22兲
and

U 2 ⫽i 共 256
51
⫺ 83 共 ␣ ⫹ ␣ * 兲兲 z 30 z 0* 2 . 共30兲
The structure of the term in curly braces in Eq. 共22兲 is inde- Again, all the resonant terms, which have the capacity to
pendent of the interaction. It is known as the Lie bracket for generate secular behavior, have been assigned to U 2 , the
this problem and in every order has the form: second-order term in the normal form; the remaining terms

再
关 U 0 ,z n 兴 ⬅ ⫺iz n ⫺U 0
⳵zn
⳵z0
⫺U 0*
⳵zn
⳵ z 0*
冎. 共23兲
have been used to construct z 2 . The same pattern recurs in
higher orders. In nth order, the formalism generates polyno-
mials of degree 2n⫹1 in z 0 and z * 0 . The resonant term,
z n⫹1
0 z *
0
n
, which would generate a secular term in z n in the
Because the perturbation is a polynomial in z 0 and z 0* , we
naive expansion, is assigned to U n , the nth-order term in the
expect z n to have a form that mimics this polynomial. The
normal form. Thus, in the normal form expansion, the non-
calculation leads to the result that the nth-order term, z n , is a resonant and the resonant terms that are generated by the
polynomial of degree (2n⫹1) in z 0 and z 0* . Thus, to first perturbation play different roles. We construct z n , the nth-
order we write order correction in the near identity transformation, Eq. 共8兲,
so that it accounts for all the nonresonant terms. The normal
z 1⫽ 兺
k,m⭓0
␣ km z k0 z 0* m . 共24兲 form, Eq. 共9兲, is constructed from all the resonant terms and
determines the time dependence of z 0 , the zero-order ap-
k⫹m⫽3
proximation.
If we substitute Eq. 共24兲 into Eq. 共23兲, we observe that the We note that the numerical coefficients in U n are pure
Lie bracket takes the form: imaginary for all n, leading to real coefficients for all the z n .
关We have shown this explicitly through O( ⑀ 2 ), but one can
关 U 0 ,z 1 兴 ⫽⫺i 兺
k,m⭓0
␣ km z k0 z * m
0 关 1⫺k⫹m 兴 . 共25兲 prove it for all orders by induction.兴 Therefore, the normal
k⫹m⫽3
form has the following simple structure:

We choose the ␣ coefficients to eliminate as many as pos- 0 ;⑀ 兲,

ż 0 ⫽⫺iz 0 g 共 z 0 z * 共31兲
sible of the terms that were generated by the perturbation on where g is a real function. If we write,
the right-hand side of Eq. 共22兲. Only the resonant term,
which is proportional to z 20 z 0* , cannot be eliminated. The z 0 ⫽ ␳ exp共 ⫺i ␪ 兲 , 共32兲
reason is that because ␣ 21 is multiplied by zero in Eq. 共25兲, it is easy to show that the amplitude, ␳, is constant:
it is undetermined. 共In future expressions, we indicate this
free multiplier as ␣.兲 Because the z 20 z 0* term is responsible d␳
⫽0. 共33兲
for the appearance of the secular contribution in Eq. 共17兲, it dt
has to be removed from the equation that determines z 1 by
assigning it to the normal form. This term determines the Hence, the U n affect only the frequency, ␻, which is changed
structure of U 1 . We use Eqs. 共20兲–共25兲 and solve for z 1 to in each order of the expansion. We use the expressions for
find U n , n⫽1,2, to obtain

and
z 1 ⫽ 161 z 30 ⫺ 163 z 0 z 0* 2 ⫺ 321 z 0* 3 ⫹ ␣ z 20 z *
0 , 共26兲
d␪
dt
3
⫽ ␻ ⫽1⫹ ⑀ ␳ 2 ⫺ ⑀ 2
8
51 3
再
⫺ 共 ␣ ⫹ ␣ * 兲 ␳ 4.
256 8 冎 共34兲

As a result, the zero-order approximation, z 0 , is given by

U 1 ⫽⫺i 83 z 20 z 0* . 共27兲
z 0 ⫽ ␳ exp共 ⫺i ␻ t 兲
Note that, were we to include in z 1 a resonant term of the ⫽ ␳ exp共 ⫺i ␻ 0 t 兲 exp共 ⫺i ⑀ ␻ 1 t 兲 exp共 ⫺i ⑀ 2 ␻ 2 t 兲 , 共35兲
form z n⫹1 0 for any n⭓1, Eq. 共25兲 would not determine its
n
0 z*
coefficient, ␣ n⫹1,n . Hence, the most general form of the where
resonant term is ␻ ⫽ ␻ 0⫹ ⑀ ␻ 1⫹ ⑀ 2␻ 2 , ␻ 0 ⫽1, ␻ 1 ⫽ 83 ␳ 2 ,

兺n ␣ n⫹1,n z n⫹1 ␻ 2 ⫽⫺ 兵 256 ⫺ 83 共 ␣ ⫹ ␣ * 兲 其 ␳ 4 . 共36兲

51
0 ⫽ f 共 z 0 •z * 共28兲
n
0 z* 0 兲z0 .
If we use our expressions for z 1 and z 2 , we can now find
For a conservative system with a single degree of freedom, z, z * , and hence x(t) through O( ⑀ 2 ):
this complication turns out to be unnecessary, and we stick to
the resonant term as written in Eq. 共26兲. In more complicated x 共 t 兲 ⫽ ␳ 关 1⫹ ⑀ 共 ⫺ 163 ⫹ ␣ 兲 ␳ 2 ⫹ ⑀ 2 共 512
69
⫺ 169 ␣ ⫹ ␤ 兲 ␳ 4 兴 cos ␪
systems 共in particular, dissipative ones兲 the more general
from of the resonant terms may be needed. ⫹ ␳ 3 共 ⑀ 321 ⫹ ⑀ 2 共 ⫺ 1024
39
⫹ 323 ␣ 兲 ␳ 2 兲 cos 3 ␪
The calculation of the second-order correction follows the
same pattern, and we obtain ⫹ ⑀ 2 ␳ 5 共 1024
1
兲 cos 5 ␪ . 共37兲

541 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 541
 Choice of free terms. The free terms that emerge in the
analysis are an inherent feature of any perturbation theory.
The freedom in their choice is a reflection of the fact that one
does not aim at the exact solution, or at a precisely specified
approximation. Rather, one searches for an approximation
that is required to be close to the correct solution within a
certain error range. Technically, this results in the fact that, in
a given order, n, of the expansion, it is impossible to distin-
guish between the effect of the perturbation on the time de-
pendence of the zero-order term and on the correction term
of order n. The numerical quality of the approximation is not
affected by different choices of the free terms. 关For example,
they cannot improve the error estimate from, say, O( ⑀ ) to
O( ⑀ 2 ) or vice versa.兴 We may choose the free terms in the
expansion 关proportional to ␣ and ␤ in Eqs. 共34兲–共37兲兴 to
satisfy a convenient requirement. Figure 1 provides a nu-
merical example of the phase-plane plot of the full solution
for the Duffing equation, together with a plot of the zero-
order approximation, obtained through a first-order analysis
with ␣⫽0. A relatively large value of ⑀ 共⑀⫽0.3兲 was chosen
to emphasize the effect of the perturbation on distorting the
zero-order circle.
One place where the choice of the free functions can be Fig. 1. Phase-plane plot of the solution for the Duffing oscillator with
⑀⫽0.3, x(0)⫽1, and dx/dt(0)⫽0. Full solution 共line兲 and zero-order circle
useful is in the calculation of the period of the oscillations, approximation 共dashed line兲 with ␣⫽0 in Eq. 共4.12兲.
which is given by an integral over one cycle in phase space:

T⫽ 冖冑 dx
2 共 E⫺V 共 x 兲兲
. 共38兲 the free functions correspond to different ways of arranging
the perturbation series and shifting higher-order terms be-
Here E is the total energy and V(x) is the potential. If we tween the near-identity transformation, Eq. 共8兲, and the nor-
substitute the potential and the total energy for the case of mal form, Eq. 共9兲. Clearly, the physics of the solution cannot
the Duffing oscillator, change, and, indeed, a detailed check17 shows that, through a
given order in ⑀, the error incurred by using the approximate
V 共 x 兲 ⫽ 21 x 2 ⫹ 41 x 4 , E⫽ 12 A 2 ⫹ 14 A 4 , 共39兲 solution rather than the exact one is of the same quality for
where A is the amplitude at the turning point, the integral different choices of the free functions. For instance, if the
becomes near-identity transformation is carried through O( ⑀ ) and the
normal form through O( ⑀ 2 ), then, as Eq. 共11兲 indicates, the
冕 A dx
共40兲
error incurred is O( ⑀ ), namely, the error is bounded by C ⑀ ,

冑
T⫽4 .
0 1 through times of O(1/⑀ ) for some C⬎0. Although this state-
共 A ⫺x 兲 ⫹ ⑀ 共 A 4 ⫺x 4 兲
2 2
ment does not depend on the choice of the free terms, the
2
constant, C, may depend on the choice.
This integral can be expressed in terms of a complete elliptic Of special interest is the minimal normal form
integral of the first kind. If we evaluate the integral through choice.17,34 –38 If we choose ␣⫽17/64, then Eqs. 共30兲 and
O( ⑀ 4 ), we find an expression for the frequency ␻ 共36兲 yield U 2 ⫽ ␻ 2 ⫽0. One can show by induction that U n
⫽(2 ␲ /T) in terms of the A: can be made to vanish for all n⭓2. For instance, ␤⫽131/
3 2 2 21 4 3 81 6 4 6549 8 8192 makes U 3 and ␻ 3 vanish. The numerical approxima-
␻ ⫽1⫹ ⑀ A ⫺⑀ A ⫹⑀ A ⫺⑀ A tions generated by the minimal normal form choice are better
8 256 2048 262 144 than all other choices. This fact is demonstrated in Fig. 2,
⫹O 共 ⑀ 5 兲 . 共41兲 where the error 关the difference between the full numerical
solution of Eq. 共1兲 and the first-order approximation兴 is
In Eq. 共36兲, ␻ is expressed in terms of ␳, the amplitude of the shown for the choice ␣⫽0 and the minimal normal form
zero-order approximation. To properly compare Eqs. 共36兲 choice of ␣⫽17/64.
and 共41兲, we need to find the relation between ␳ and A. We The formal truncation of the normal form is an example of
obtain this relation by imposing the initial conditions, x(0) the concept of renormalization. The origin of the ability to
and dx/dt(0). The choice of the free functions affects this formally truncate the normal form at a finite order is the
relation as well as the manner in which the phase is changed. same as for the multipole expansion of electromagnetic po-
Problem 1. Explain why the period can be calculated as tentials. The functional form of the first nonzero multipole is
four times the integral over a quarter of the cycle and derive physically significant, whereas the functional form of all the
Eq. 共40兲. higher multipoles depends on the choice of the origin of the
The choice of the free functions 共the constants ␣ and ␤ in axes in the frame of reference in which the expansion is
the present problem兲 affects the structure of the higher-order performed. Similarly here, the first nontrivial correction to
correction terms in the near-identity transformation. As the the frequency 共a first-order one in the case of the Duffing
preceding analysis shows, this choice modifies the structure oscillator兲 is physically significant, whereas all the higher-
of the zero-order approximation. Thus, different choices of order corrections depend on the choice of the zero-order ap-

542 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 542
proximation 共the analog of the choice of the origin in the
multipole expansion兲. The latter is intimately related to the
choice of the free terms in each order for the following rea-
son. A resonant term is proportional to z n⫹1 0 z*
0
n

⫽(z 0 z 0* ) n z 0 . Because for a conservative system the ampli-

tude, ␳, is constant 关see, Eq. 共33兲兴, the free resonant terms
that are included in the higher-correction terms z n are pro-
portional to the zero-order term, z 0 . Thus, inclusion of free
resonant terms in z n amounts to shifting parts of the contri-
bution of the zero-order term to higher orders.

B. The oscillator with cubic damping

The main features of the solution for the Duffing oscillator
are typical of conservative systems with one degree of free- Fig. 2. Error in the first-order approximation to the solution for the Duffing
oscillator with the same conditions as in Fig. 1. No free function 共line兲 and
dom: the amplitude, ␳, of the zero-order term, z 0 , is con- minimal normal form 共dashed line兲 choice.
stant, and the frequency is changed by constant terms, so that
the phase is just ␻ t, with ␻ a constant that differs from the
unperturbed frequency. The structure of the solutions in dis-
sipative systems is different. Both the amplitude and the damped motion. In a similar manner, the O( ⑀ 2 ) terms are
phase are complicated functions of time. For the oscillator found to be
with cubic damping, the solution is attracted to the origin due U 2 ⫽⫺ 83 共 f 1 ⫹ f * 0 ⫹ 4f⬘
1 z 0 z 0* ⫹ 256 iz 0 z * 共48兲
2 3 2 3 2
1 兲z 0z *
3 27
0 ,
to damped oscillations. The calculations parallel those of the
Duffing oscillator. If we use Eqs. 共3兲, 共6兲, and 共7兲, we can z 2 ⫽⫺ 1024
9
z 50 ⫹ 256
9 4
z 0 z 0* ⫺ 512 z 0 z 0* 3 ⫺ 1024
27 2 27
z 0 z 0* 4 ⫹ 512
3
z 0* 5
write
⫹ 163 i f 1 z 30 ⫺ 163 i f 1 z 20 z *
0 ⫺ 8i f *
2 3
0 ⫹ 32 i f *
1 z 0z *
2 3
1 z 0*
2
共 z⫺z * 兲 3 共 z⫺z * 兲 3
ż⫽⫺iz⫺i ⫽⫺iz⫹i . 共42兲 ⫹ f 2 共 z 0 •z 0* 兲 z 0 , 共49兲
2i 2
An example of the solution for x(t), which is the real part of where
z(t), is given in Fig. 3. d
We substitute the near-identity transformation, Eq. 共8兲, and f 1⬘ ⫽ f 1 共 z 0 z 0* 兲 , 共50兲
0兲
d共 z 0z *
the normal form, Eq. 共9兲, in Eq. 共42兲 and obtain:

冉 冊
and a free resonant term has been included in z 2 . To O( ⑀ 3 )
dz ⳵z1 ⳵z1 we find
⫽ 兵 U 0⫹ ⑀ U 1⫹ ⑀ 2U 2其 ⫹ ⑀ U ⫹ U*
dt ⳵ z 0 0 ⳵ z 0* 0
U 3 ⫽⫺ 83 共 2 f 21 ⫹3 f 1 f * 2
0 ⫹ 关 128 i 共 f 1 ⫹ f 1* 兲
1 兲z 0z *
27

⫹⑀2 冉 ⳵z2
⳵z0
U 0⫹
⳵z2
⳵ z 0*
冊 冉
U 0* ⫹ ⑀ 2
⳵z1
⳵z0
U 1⫹
⳵z1
⳵z*
0
U*
1 冊 ⫺ 34 f 1 f 1⬘ 兴 z 30 z *
⫺ 38 共 f 2 ⫹ f * 2
0 ⫺ 4共 f ⬘
2 3

0 ⫹ 4f⬘
2 兲z 0z *
3
1 ⫹ f 1⬘ * 兲 f ⬘
4
1 z 0 z 0*

2 z 0 z 0* ⫺ 8192 z 0 z *
3 2 567 4 3
0 .
3

共51兲
1
⫽⫺i 兵 z 0 ⫹ ⑀ z 1 ⫹ ⑀ 2 z 2 其 ⫹ ⑀ 共 z 0 ⫺z *
0兲
3 We now show how the choice of the free resonant terms in
8 the generators, z n , may be exploited. If all the free terms
3 vanish ( f 1 ⫽ f 2 ⫽0 in the approximation presented here兲,
⫹ ⑀ 2 共 z 0 ⫺z 0* 兲 2 共 z 1 ⫹z 1* 兲 . 共43兲 then the normal form is given by
8
dz 0 3 27 3 2 3 567 4 3
We collect terms as before and obtain: ⫽⫺iz 0 ⫺ ⑀ z 20 z 0* ⫹ ⑀ 2 iz z * ⫺ ⑀ z z*
dt 8 256 0 0 8192 0 0
U 0 ⫽⫺iz 0 , 共44兲

再 冎
⫹O 共 ⑀ 4 兲 . 共52兲
⳵z1 ⳵z1 共 z 0 ⫺z *
0兲
3
U 1 ⫽ ⫺iz 1 ⫺U 0 ⫺U * ⫹ . 共45兲 If we use Eq. 共32兲 for z 0 , we find that Eq. 共52兲 leads to the
⳵z0 0
⳵z*
0
8 conclusion that is typical of dissipative systems equations,
namely, that both the frequency, ␻, and the amplitude, ␳,
If we use the expression in Eq. 共24兲 for z 1 , the O( ⑀ ) part of
vary with time in a nontrivial manner:
the solution is
d␳ 3 567 3 7
U 1 ⫽⫺ ⑀ 83 z 20 z *
0 , 共46兲 ⫽⫺ ⑀ ␳ 3 ⫺ ⑀ ␳ ⫹O 共 ⑀ 4 兲 , 共53兲
dt 8 8192
and
d␪ 27 2 4
3 ⫽ ␻ ⫽1⫺ ⑀ ␳ ⫹O 共 ⑀ 4 兲 . 共54兲
z 1 ⫽ 161 iz 30 ⫺ iz 0 z 0* 2 ⫹ 321 iz 0* 3 ⫹ f 1 共 z 0 •z 0* 兲 z 0 . 共47兲 dt 256
16
In higher orders, the amplitude and phase equations quickly
The last term in z 1 is the free resonant term. Observe that z 1 become unwieldy. However, in the present problem, the free
has imaginary coefficients and that U 1 is real, indicating functions can be chosen to eliminate U n for all n⭓2. For

543 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 543
 amplitude of z 0 , is affected by the perturbation. In second
order, for the common choice of f 1 ⫽ f 2 ⫽0, U 2 has an imagi-
nary coefficient 关see Eq. 共48兲兴, so that U 2 affects only the
phase and not the amplitude of z 0 . This pattern persists in all
higher orders. The amplitude equation is changed only in odd
orders in ⑀, and the phase equation only in even orders. 共4兲
For conservative systems, the normal form expansion yields
a phase that is linear in t because the frequency is changed by
successive constant additive corrections. In dissipative sys-
tems, the correction terms are not constant, and the phase
becomes a complicated function of t. This point will be re-
examined in our discussion of the method of multiple time
scales.

Fig. 3. The position x(t) for an harmonic oscillator with cubic damping
with the same conditions as in Fig. 1.
C. The Van der Pol oscillator
The next level of intricacy is attraction to or repulsion
from a closed periodic orbit, a limit cycle. When the trajec-
instance, requiring that U 2 vanish can be achieved for a va- tory is attracted to an asymptotic closed periodic orbit, the
riety of choices for f 1 , the simplest one being latter is called a stable limit cycle. The Van der Pol oscillator
f 1 ⫽⫺ 649 iz 0 z * 共55兲 is the usual system used to illustrate this type of behavior. If
0 .
we use the complex variables, z⫽x⫹iy and z * ⫽x⫺iy, Eq.
In a similar manner, the simplest choice for f 2 that eliminates 共4兲 becomes
U 3 is
ż⫽⫺iz⫹ 81 ⑀ 兵 4⫺ 共 z⫹z * 兲 2 其 共 z⫺z * 兲 , 共60兲
675
f 2⫽ 共 z z * 兲2. 共56兲 where ⑀⬎0. For the discussion that follows, we present the
8192 0 0 z n through O( ⑀ 2 ) and U n through O( ⑀ 3 ),
One can show by induction that higher-order resonant con- z 1 ⫽ 41 iz *
0 ⫺ 16 iz 0 ⫺ 16 iz 0 z 0* ⫺ 32 iz 0* ⫹ f 1 共 z 0 z 0* 兲 z 0 , 共61兲
1 3 1 2 1 3
tributions to the normal form may be eliminated order-by-
order to yield the following expression, correct to all orders, z 2 ⫽⫺ 1024
5
z 50 ⫺ 256
1 4
z 0 z 0* ⫹ 512
5 2
0 ⫹ 1024 z 0 z *
z 0z * 3 1
0 ⫹ 1536 z 0*
4 5 5

for the minimal normal form:

⫺ 128
5 3
z 0 ⫺ 321 z 0 z *
0 ⫹ 128 z *
2 1
0 ⫹ f 2共 z 0z *
3
0 兲z0 , 共62兲
dz 0 3
⫽⫺iz 0 ⫺ ⑀ z 20 z *
0 共57兲 U 1 ⫽ 21 共 1⫺ 41 z 0 z 0* 兲 z 0 , 共63兲
dt 8
which implies that U 2⫽ 11
256 0 ⫺ 16 iz 0 z 0* ⫹ 8 iz 0
iz 30 z * 2 3 2 1

d␳ 3 d␪ ⫹ 共 ⫺ 41 Re f 1 ⫺ f 1⬘ 共 1⫺ 41 z 0 z * 2
0 兲兲 z 0 z *
0 . 共64兲
⫽⫺ ⑀ ␳ 3 , ⫽ ␻ ⫽1. 共58兲
dt 8 dt The minimal normal form option that U 2 vanishes identically
The solution of Eq. 共58兲 is requires that f 1 have a pole at z 0 z 0* ⫽4, and therefore is not
a good choice. Worse than that, a nonzero Re f 1 in Eq. 共64兲
␳0
␪ ⫽t⫹ ␪ 0 , ⇒ ␻ ⫽1, ␳⫽ 共59兲 introduces a spurious time dependence in the amplitude. The
冑
.
3 best choice for the free term in z 1 is f 1 ⫽0, which leads to
1⫹ ⑀ ␳ 20 t
4
U 2 ⫽ 256
11
0 ⫺ 16 iz 0 z 0* ⫹ 8 iz 0 ,
iz 30 z * 2 3 2 1
共65兲
Thus, the minimal normal form captures the behavior of the
zero-order approximation in a very simple manner. Both the and
fundamental frequency of oscillation and the amplitude are U 3 ⫽⫺ 8192
13 4
0 ⫹ 1024 z 0 z *
z 0z * 3 11 3
0 ⫺ 128 z 0 z *
2 3 2
0
determined at this order and are not changed by higher-order
corrections. Higher-order effects show up only in z n in the ⫹ 兵 ⫺ 41 Re f 2 ⫺ f ⬘2 共 1⫺ 41 z 0 z 0* 兲 其 z 20 z 0* . 共66兲
near-identity transformation.
Problem 2. Solve Eqs. 共53兲 and 共54兲 numerically for the We first examine the simple choice of f 2 ⫽0. In polar coor-
initial conditions ␳共0兲⫽1 and ␪共0兲⫽0. What do you learn dinates, Eq. 共32兲, the radial equation is
about the phase ␪ from the solution? In what way does it
differ from the phase for the Duffing oscillator?
Problem 3. Show that the choice, Eq. 共55兲, makes U 2 van-
d␳ 1
dt 2
1
4 冉
⫽ ⑀ ␳ 1⫺ ␳ 2 ⫹ ⑀ 3 ⫺
1
8192 冊 冉
␳ 7⫹
11 5
1024
␳

ish.
We note the following points. 共1兲 The z n have the same
structure as those obtained for the Duffing oscillator. 共2兲 The
⫺
3 3
128 冊
␳ ⫹O 共 ⑀ 5 兲 . 共67兲

minimal normal form choice, so simple in this case, may not Starting at small amplitudes, the system executes a spiral
be a viable option in most dissipative systems. 共The Van der leading asymptotically to a stable limit cycle. In an O( ⑀ )
Pol oscillator is an example.兲 共3兲 We note that, in first order, calculation, the right-hand side of Eq. 共67兲 vanishes at ␳ 2
the coefficient of U 1 is real 关see Eq. 共46兲兴, so that only ␳, the ⫽4, defining the radius of the limit circle. However, in

544 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 544
higher orders, the zero of the right-hand side of Eq. 共67兲 is See, for example, Eq. 共35兲 for the Duffing oscillator, and Eq.
shifted to 共73兲 for the t→⬁ behavior of the van der Pol limit cycle. For
␳ 2 ⫽4⫺ 163 ⑀ 2 ⫹O 共 ⑀ 4 兲 . 共68兲 兩⑀兩Ⰶ1 and short times 关 t⫽O(1) 兴 , only the n⫽0 term is im-
portant, because the terms ⑀ n t(n⭓1) in Eq. 共76兲 are negli-
Thus, the radius of the limit circle is affected by the pertur- gibly small and hardly affect the time dependence. As t
bation. →O(1/⑀ ), the first two terms (n⫽0,1) become significant;
The equation for the 共time-dependent兲 frequency is
when it becomes O(1/⑀ 2 ) the terms with (n⫽0,1,2) are sig-
␻⫽
d␪
dt
⫽1⫺ ⑀ 2
256 冉
11 4 3 2 1
␳ ⫺ ␳ ⫹ ⫹O 共 ⑀ 4 兲 .
16 8 冊 共69兲
nificant, and so on. Thus, as time increases as measured in
powers of 1/⑀, a progressively larger number of terms in the
time dependence become significant. The nth time scale is
To obtain the asymptotic value of the frequency as t→⬁, we defined by
must substitute in Eq. 共69兲 the asymptotic value of ␳, which
is given by an infinite series in ⑀. This complication can be T n ⫽ ⑀ n t. 共75兲
avoided by a radius renormalization choice for f 2 . We look
When Eq. 共74兲 applies, the phase factor, exp(⫺i␪), can be
for an f 2 that will make the asymptotic value of ␳ 2 remain at
written in a factorized form:
4 when the third-order contribution is included in the ampli-
2
tude equation. To this end, we need to make U 3 proportional e ⫺i ␪ ⫽e ⫺i ␻ 0 t e ⫺i ␻ 1 共 ⑀ t 兲 e ⫺i ␻ 2 共 ⑀ t 兲 •••
to U 1 . With U 3 given by Eq. 共66兲, this behavior is achieved
by ⫽e ⫺i ␻ 0 T 0 e ⫺i ␻ 1 T 1 e ⫺i ␻ 2 T 2 •••. 共76兲
Re f 2 ⫽⫺ 8192
13
共 z 0 z 0* 兲 3 ⫹ 1024
11
共 z 0 z 0* 兲 2 ⫺ 128
3
共 z 0 z 0* 兲 For a given duration, hardly anything happens on sufficiently
long time scales 共for example, as the second hand of a clock
共 Im f 2 ⫽0 兲 . 共70兲 moves a few seconds, the minute and hour hands hardly
The expression for U 3 becomes move兲. We are, therefore, tempted to treat the various time
3 2
scales, T n , as independent variables. This assumption is the
U 3 ⫽ 共 4096
13 4
0 ⫺ 1024 z 0 z 0* 兲共 1⫺ 4 z 0 z *
z 0z * 11 3 1
0 兲. 共71兲 starting point of the method of multiple time scales: The
solution is assumed to depend on all time scales, z
The radial equation is modified and Eq. 共67兲 is replaced by
⫽z(T 0 ,T 1 ,T 2 ,...). Hence, in addition to expanding the so-
d␳ 1
dt 2
1
⫽ ⑀ ␳ 1⫺ ␳ 2
4 冉 冊 lution in powers of ⑀ 关the near-identity transformation, Eq.
共8兲兴, the time derivative is replaced by:

再 冉
⫻ 1⫹ ⑀ 2
13 6 11 4
2048
␳ ⫺
512 冊
␳ ⫹O 共 ⑀ 4 兲 . 冎 共72兲
d
→
⳵
dt ⳵ T 0
⫹⑀
⳵
⳵T1
⫹⑀2
⳵
⳵T2
⫹•••. 共77兲

Thus, the value of the limit-circle radius is unaffected by The introduction of independent time scales and the elimina-
higher-order corrections and remains fixed at ␳⫽2. As a re- tion of secular terms lead to conditions on the dependence of
sult, the series for the asymptotic expression for the fre- the solution on the various time scales 共solvability condi-
quency is readily obtained: tions兲. For most dynamical systems, these conditions consti-
tute nontrivial consistency constraints on the structure of the
␻ ⫽1⫺ 161 ⑀ 2 ⫹O 共 ⑀ 4 兲 . 共73兲
approximate solutions. In general, the freedom, enjoyed in
17
A detailed examination shows that, as expected, shifting the method of normal forms, to choose the free resonant
terms in the expansion procedure from the near-identity terms that appear in every order in the near-identity transfor-
transformation to the normal form does not alter the physics. mation, Eq. 共8兲, does not exist in the multiple time scales
The free functions enable us to simplify parts of the calcula- method. The consistency problems encountered in the
tion. However, all choices yield the same numerical accuracy method are an issue of much concern.39– 46
through a given order. For conciseness, it is customary to use the notation
In Fig. 4 a phase-plane plot of two solutions of Eq. 共4兲 is
shown: one spiraling toward the limit cycle from the outside, ⳵
D n⬅ . 共78兲
and one spiraling from the inside. To see the effect of the ⳵Tn
higher-order terms on the solution, a limit circle of ␳⫽2 also
is plotted. If we substitute Eqs. 共8兲 and 共77兲 in Eq. 共7兲, and expand
Problem 4. Solve Eq. 共4兲 for the same initial conditions as through O( ⑀ 2 ), we obtain
in Fig. 4, but with ⑀⫽⫺0.3. What is the behavior of the
共 D 0 ⫹ ⑀ D 1 ⫹ ⑀ 2 D 2 兲共 z 0 ⫹ ⑀ z 1 ⫹ ⑀ 2 z 2 兲
solution?
⫽⫺i 共 z 0 ⫹ ⑀ z 1 ⫹ ⑀ 2 z 2 兲 ⫹i ⑀ F 共 z 0 ,z *
0兲
V. THE METHOD OF MULTIPLE TIME SCALES
To motivate the introduction of the method of multiple
time scales, we observe that in stable oscillatory systems, the
⫹i ⑀ 2 再 ⳵ F 共 z 0 ,z *
⳵z0
0兲
z 1⫹
⳵ F 共 z 0 ,z *
⳵z* 0
0兲
z*
1 . 冎 共79兲

time dependence of the solution is given through a phase, ␪, We collect terms order by order and obtain
which has the form

冉兺 冊
D 0 z 0 ⫹iz 0 ⫽0, 共80兲
␪ ⫽ ␻ t⫽ ⑀ ␻ n t.
n
共74兲 D 0 z 1 ⫹iz 1 ⫽⫺D 1 z 0 ⫹F 共 z 0 ,z * 共81兲
n⭓0 0 兲,

545 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 545
 solution for z 1 , including the solution of the homogeneous
equation:
z 1 ⫽z̃ 1 ⫹B 共 T 1 兲 e ⫺iT 0 , 共87兲
where z̃ 1 is the particular solution that accounts for the non-
resonant contribution of F(z 0 ,z 0* ) in Eq. 共81兲.
With z 1 known, Eq. 共82兲 is analyzed in a similar manner,
and z 2 also will include a solution of the homogeneous part
of Eq. 共82兲. To avoid secular terms, the contribution of all the
resonant terms on the right-hand side of Eq. 共82兲 has to van-
ish, leading to the second-order solvability condition:

冉 ⫺D 1 z 1 ⫺D 2 z 0 ⫹
⳵ F 共 z 0 ,z *
⳵z0
0兲
z 1⫹
⳵ F 共 z 0 ,z 0* 兲
⳵ z 0*
z 1* 冊 R
⫽0.

共88兲
Equation 共88兲 determines the T 1 dependence of the ampli-
Fig. 4. Phase-plane plot of a solution of the Van der Pol oscillator with tude B. Consistency between Eqs. 共86兲 and 共88兲 will turn out
⑀⫽0.3. Solution of Eq. 共4.35兲 共full line兲, limit cycle 共thick lime兲, limit circle
to be a major limitation on the application of the method of
with radius⫽2 共dashed line兲.
multiple time scales.
As in the method of normal forms 关see Eqs. 共11兲 and 共12兲兴,
a comment is in order regarding error estimates. An error of
the magnitude defined by Eq. 共11兲 is obtained if the near-
⳵ F 共 z 0 ,z *
0兲 identity transformation contains higher-order corrections
D 0 z 2 ⫹iz 2 ⫽⫺D 1 z 1 ⫺D 2 z 0 ⫹ z1
⳵z0 through O( ⑀ N ), as in Eq. 共10兲. Concurrently, for each term,
z k , k⭐N, in the approximate solution, we find the depen-
⳵ F 共 z 0 ,z *
0兲
⫹ z 1* . 共82兲 dence of z k on T 0 , T 1 ,...,T N⫹1⫺k . Thus, we have to find the
⳵z*
0 dependence of z 0 on T 0 , T 1 ,...,T N⫹1 , of z 1 on T 0 ,
T 1 ,...,T N , of z 2 on T 0 , T 1 ,...,T N⫺1 , etc. A major point that
Equation 共80兲 is readily solved by
we will discuss in the following is that often the method of
z 0 ⫽A 共 T 1 ,T 2 兲 e ⫺iT 0 . 共83兲 multiple time scales does not yield the dependence on the
higher time scales, and we must augment the method with
The amplitude, A, depends on all higher time scales. How- additional information to find that dependence.
ever, because we will limit our analysis to second order, only
T 1 and T 2 have been written out explicitly.
Although the analysis of Eq. 共81兲 depends on the detailed VI. APPLICATIONS OF THE METHOD OF
structure of the perturbation, F(z 0 ,z 0* ), the following can MULTIPLE TIME SCALES
already be said at this point. First, the homogeneous part of
the equation, A. The Duffing oscillator
D 0 z 1 ⫹iz 1 ⫽0, 共84兲
For F(z 0 ,z 0* ) of the Duffing oscillator 关see Eq. 共6兲兴, and
introduces a contribution z 1 which we write in the form 共83兲 for z 0 , Eq. 共81兲 becomes
B 共 T 1 兲 e ⫺iT 0 . 共85兲 D 0 z 1 ⫹iz 1 ⫽⫺D 1 Ae ⫺iT 0 ⫺ 81 iA 3 e ⫺3iT 0 ⫺ 83 iA 2 A * e ⫺iT 0
The amplitude B may depend on all higher time scales. How- ⫺ 38 iAA * 2 e iT 0 ⫺ 81 iA * 3 e 3iT 0 . 共89兲
ever, in a second-order analysis, only its T 1 dependence
The first and third terms on the right-hand side of Eq. 共89兲
counts. Second, any term on the right-hand side of Eq. 共81兲
are the resonant terms. Their elimination yields the first-
whose T 0 dependence is of the form exp(⫺iT0) resonates
order solvability condition,
with the left-hand side of the equation, and hence will gen-
erate in z 1 a term that is linear in T 0 共a secular term兲. Such D 1 A⫽⫺ 38 iA 2 A * . 共90兲
terms must be eliminated to yield an approximation that is If we write A in the form,
valid for times of O(1/⑀ ) and not just of O(1). There is an
explicit term of this kind, namely, ⫺D 1 z 0 . In addition, any A⫽ 兩 A 兩 e ⫺i ␸ A , 共91兲
monomial of the form z n⫹1 0 z* n
0 in F(z 0 ,z *
0 ) will have this we readily find
T 0 dependence. We denote the resonant part of F(z 0 ,z 0* ) by D 1 兩 A 兩 ⫽0, ␸ A ⫽ 83 兩 A 兩 2 T 1 ⫹ ␺ A 共 T 2 兲 . 共92兲
F(z 0 ,z 0* ) R . To eliminate all the terms that generate secular
behavior, we impose the requirement Thus, the magnitude of the amplitude A is independent of
T 1 , and hence its phase is linear in T 1 .
⫺D 1 z 0 ⫹F 共 z 0 ,z *
0 兲 R ⫽0. 共86兲 Now that we have eliminated the resonant terms in Eq.
共89兲, the solution for z 1 is readily found to be
This first-order solvability condition determines the T 1 de-
pendence of the amplitude A of Eq. 共83兲. Once the resonant z 1 ⫽ 161 A 3 e ⫺3iT 0 ⫺ 163 AA * 2 e iT 0 ⫺ 321 A * 3 e 3iT 0 ⫹Be ⫺iT 0 .
term in Eq. 共81兲 has been removed, we can find the full 共93兲

546 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 546
We turn to the O( ⑀ 2 ) equation, Eq. 共82兲, and substitute Eqs. Over the time span t⫽O(1/⑀ ) for which the approxima-
0 ) of Eq. 共6兲, in Eq. 共82兲,
共90兲 and 共93兲, together with F(z 0 ,z * tion is valid, powers of T 1 in B need not be eliminated: They
to obtain do not spoil the ordering of the expansion in terms of pro-
gressively smaller terms, because ⑀ B contains contributions
D 0 z 2 ⫹iz 2 ⫽⫺D 1 兵 161 A 3 e ⫺3iT 0 ⫺ 163 AA * 2 e iT 0 proportional to
⫺ 321 A * 3 e 3iT 0 ⫹Be ⫺iT 0 其 ⫺D 2 Ae ⫺iT 0 ⑀ T q1 ⫽ ⑀ 共 ⑀ t 兲 q ⬇ O共 ⑀ 兲. 共101兲
⫺iT 0 3 ⫺3iT 0 t⫽O 共 1/⑀ 兲
⫹ 共 Ae
3
8 ⫹A * e 兲 兵 A e
iT 0 2 1
32

⫺ A 2 A * e ⫺iT 0 ⫺ AA * 2 e iT 0 ⫹ 321 A * 3 e 3iT 0

3
16
3
16
Problem 6. What happens to the magnitude of (A * B
⫹AB * ) of Eq. 共99兲 for t⫽O(1/⑀ 2 )?
⫹Be ⫺iT 0 其 . 共94兲 The T 2 dependence of A is undetermined by Eqs. 共90兲 and
共95兲. To specify it, we must either provide additional input,
To avoid secular terms in z 2 , all the terms that are propor-
extraneous to the method, for example, initial data for T 1
tional to exp(⫺iT0) must be eliminated from Eq. 共94兲. This
⫽0. A common way to fix the T 2 dependence of A is to
condition yields the second-order solvability condition:
require that the solution be constructed solely out of periodic
51
D 2 A⫹D 1 B⫽ 256 iA 3 A * 2 ⫺ 43 iAA * B⫺ 38 iA 2 B * . 共95兲 functions. To avoid aperiodic terms, Eq. 共97兲 must be broken
into
Equation 共95兲 and its complex conjugate may be regarded as
linear, first-order differential equations for the T 1 depen- D 2 共 AA * 兲 ⫽0, D 1 共 A * B⫹AB * 兲 ⫽0, 共102兲
dence of B and B * . However, because it contains the deriva-
tive of the amplitude A with respect to T 2 , it also can be which implies that
viewed as a constraint on A. Thus, we must check if the two
solvability constraints, Eqs. 共90兲 and 共95兲, are mutually con- 共 A * B⫹AB * 兲 ⫽H 共 T 2 ,T 3 ,... 兲 . 共103兲
sistent. This check can be done by either solving Eqs. 共90兲
and 共95兲 directly or by requiring that derivatives of A with Hence, 兩 A 兩 is independent also of T 2 . If we use the same
respect to T 2 and T 1 commute, reasoning in Eq. 共98兲, we must require
D 1 D 2 A⫽D 2 D 1 A, 共96兲
3 51
␸ A ⫽ 共 AA * 兲 T 1 ⫺ 共 AA * 兲 2 T 2
when the derivative D 2 is applied to Eq. 共90兲 and D 1 is 8 256
applied to Eq. 共95兲. For the Duffing oscillator, both ways are
simple and easy to implement, while in more complicated
systems, one way may be more advantageous than the other.
⫹ 冕 0
T2 3
8
共 A * B⫹AB * 兲 dT 2 ⫹ ␹ A 共 T 3 ,T 4 ,... 兲 , 共104兲
We now perform the analysis in these two ways.
Problem 5. Derive Eq. 共95兲. and
Direct solution of solvability conditions. If we multiply
Eq. 共97兲 by A * and its complex conjugate by A and use Eq. D 1 共 A * B⫺AB * 兲 ⫽0, 共105兲
共90兲, we obtain

D 1 共 A * B⫹AB * 兲 ⫹D 2 共 AA * 兲 ⫽0, 共97兲 which implies

D 1 共 A * B⫺AB * 兲 ⫺2iAA * D 2 ␸ A 共 A * B⫺AB * 兲 ⫽iJ 共 T 2 ,T 3 ,... 兲 . 共106兲

⫽⫺ 34 iAA * 共 A * B⫹AB * 兲 ⫹ 128

51
i 共 AA * 兲 3 . 共98兲 Consequently, we find that, through O( ⑀ 2 ), the dependence
of A on T 1 and T 2 factors into a product of phase terms, each
Because 兩 A 兩 is independent of T 1 关see Eq. 共92兲兴, the term depending on one time scale only 关although the T 2 depen-
D 2 (AA * ) in Eq. 共97兲 also is independent of T 1 . Hence, the dence of (A * B⫹AB * ) is still unknown兴:
solution for A * B⫹AB * is

共 A * B⫹AB * 兲 ⫽⫺D 2 共 AA * 兲 T 1 ⫹H 共 T 2 ,T 3 ,... 兲 . 共99兲 A⫽ 兩 A 兩 e ⫺iT 0 exp ⫺i 冉 3

8 冊 冉再
共 AA * 兲 T 1 exp i
51
256
共 AA * 兲 2 T 2

冎冊
Consequently, Eq. 共98兲 generates in (A * B⫺AB * ) terms that
are linear and quadratic in T 1 : ⫺ 冕 0
T2 3
8
共 A * B⫹AB * 兲 dT 2 e ⫺i ␹ A 共 T 3 ,T 4 ,... 兲 . 共107兲
共 A * B⫺AB * 兲 ⫽i 兵 共 2AA * D 2 ␸ A ⫹ 128
51
共 AA * 兲 3
⫺ 34 AA * H 共 T 2 ,T 3 ,... 兲兲 T 1 This pattern recurs in higher orders if we require that the
solvability conditions be satisfied solely by periodic func-
⫹ 38 AA * D 2 共 AA * 兲 T 21 ⫹J 共 T 2 ,T 3 ,... 兲 其 . tions. If this requirement is not imposed, then the depen-
共100兲 dence of the amplitudes on higher time scales, T n , n⭓2,
depends on our choice of the initial data at, say, T 1 ⫽0.
Equations 共99兲 and 共100兲 yield the general solution of B with With the choice that the solution does not contain powers
H and J real-valued functions. of T 1 , Eqs. 共103兲, 共106兲, and 共90兲 lead to

547 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 547
D 1 共 A * B 兲 ⫽0 is, there are no powers of T 1 in the solution兲, is that B has the
same phase as A. Also, the choice B⫽0 is the equivalent to
⇒D 1 B
the choice ␣⫽0 in Eq. 共26兲, and the choice of Eq. 共110兲 is the
3 equivalent of the minimal normal form choice of ␣⫽17/64,
⫽⫺ iAA * B mentioned toward the end of Sec. IV A.
8

⇒ 再 B 共 T 1 ,T 2 ,... 兲 ⫽ 兩 B 兩 e ⫺i 共 3/8兲共 AA * 兲 T 1 e i ␺ B 共 T 2 ,T 3 ,... 兲

D 1 共 BB * 兲 ⫽0 B. The oscillator with cubic damping
共108兲 The Duffing oscillator is representative of dynamical sys-
Note that the T 2 dependence of A still cannot be determined tems to which an energy conserving perturbation is added.
unless additional input is provided. Part of the dependence The methods of normal forms and multiple time scales are
has been found due to the requirement that, through O( ⑀ 2 ), completely equivalent, and there is a one-to-one correspon-
the solution does not contain powers of T 1 . This requirement dence of the results. In particular, any choice of the free
led to the result that 兩 A 兩 does not depend on T 2 . However, it terms in the normal form analysis has a corresponding choice
is not obvious that the choices made up to this point are in the multiple time scales analysis and both methods yield
consistent with the solvability conditions in higher orders. the same functional form for the approximate solution. The
Because we have limited our analysis to second order, we two examples of dissipative systems to be discussed in the
only state the answer: For the case of the Duffing oscillator following represent obstacles encountered in the method of
共and all single-oscillator systems with energy conserving per- multiple time scales, which are not encountered in the nor-
turbations兲, the choices are consistent. Finally, we still do not mal form analysis. These difficulties occur in the multiple
time scales analysis of most linear dynamical systems that
know the dependence of ␸ A , the phase of A, on T 2 . Help
are subject to a nonlinear perturbation, be they systems of
comes from observing that, in this order, the solvability con-
one degree of freedom with a nonconservative perturbation,
ditions leave 兩 B 兩 free, allowing, in particular, 兩 B 兩 ⫽0. This or Hamiltonian systems with more than one degree of free-
choice completes the T 2 dependence of A and yields a simple dom.
expression for ␸ A : Because the analysis has been presented in detail for the
␸ A ⫽⫺ 83 共 AA * 兲 T 1 ⫹ 256
51
共 AA * 兲 2 T 2 ⫹ ␹ A 共 T 3 ,T 4 ,... 兲 . Duffing oscillator, we skip some of the steps in the analysis
共109兲 of the cubically damped oscillator, and only discuss the re-
sults and the important points. The zero-order solution, z 0 , is
Note that the T 1 dependence of ␸ B , the phase of B, is the given by Eq. 共83兲. The first-order correction is found to be
same as that of ␸ A , leading to another choice for B that has
aesthetic appeal 共and also is consistent with the higher-order z 1 ⫽ 161 iA 3 e ⫺3iT 0 ⫺ 163 iAA * 2 e iT 0 ⫹ 321 iA * 3 e 3iT 0 ⫹Be ⫺iT 0 .
solvability conditions兲: 共113兲

B⫽ 17 2
共110兲 The first- and second-order solvability conditions are
64 A A * .

With this choice. ␸ A , and, hence, A, become independent of D 1 A⫽⫺ 共 83 兲 A 2 A * , 共114兲

T 2 , and Eq. 共104兲 becomes 27
D 2 A⫹D 1 B⫽ 256 iA 3 A * 2 ⫺ 43 AA * B⫺ 38 A 2 B * . 共115兲
␸ A ⫽ 共 AA * 兲 T 1 ⫹ ␹ A 共 T 3 ,T 4 ,... 兲 .
3
8 共111兲
Direct solution of solvability conditions. If we write A as
In the analysis in higher orders, the requirement that the in Eq. 共91兲, Eq. 共114兲 is solved by
dependence of the amplitudes be expressed in terms of only
periodic functions 共that is, exponential phase factors兲, im- D 1 共 AA * 兲 ⫽⫺ 43 共 AA * 兲 2 , D 1 ␸ A ⫽0, 共116兲
plies that A, B, and the free amplitudes that appear in higher leading to

冒冉 冊
orders will all have the same time dependent phases.
Consistency of solvability conditions. The constraint for
consistency between the two solvability conditions is ob- A⫽ 冑 共 AA * 兲 0
3
1⫹ 共 AA * 兲 0 T 1 e ⫺i ␸ A 共 T 2 ,T 3 兲 ,
4
tained by applying D 2 to Eq. 共90兲, D 1 to Eq. 共95兲, and re- 共117兲
quiring that Eq. 共96兲 be obeyed 关using Eq. 共95兲 repeatedly兴.
where (AA * ) 0 does not depend on T 1 . If we multiply Eq.
The result yields the T 1 dependence of B:
共115兲 by A * , and its complex conjugate equation by A, and
D 21 B⫹ 32 i 共 AA * 兲 D 1 B⫹ 43 iA 2 D 1 B * ⫺ 64
27
共 AA * 兲 2 B sum or subtract the results, we find after repeated use of Eqs.
共114兲 and 共115兲:
⫹ 329 共 AA * 兲 A 2 B * ⫽0. 共112兲
D 1 兵 共 A * B⫹AB * 兲 / 共 AA * 兲 2 其 ⫺D 2 共 1/共 AA * 兲兲 ⫽0, 共118兲
Because Eq. 共112兲 is homogeneous, B⫽0 is an allowed
choice. Similarly, the use of Eq. 共90兲 shows that the choice of D 1 兵 共 A * B⫺AB * 兲 / 共 AA * 兲 ⫹ 329 i 共 AA * 兲 其 ⫹2iD 2 ␸ A ⫽0.
Eq. 共110兲 also is an allowed solution of Eq. 共112兲. 共119兲
We end our analysis of the Duffing oscillator by noting We use Eq. 共117兲 and solve Eq. 共118兲 to find
that the multiple time scales results are completely parallel to
those of the method of normal forms. For instance, in the 共 A * B⫹AB * 兲 ⫽ 兵 D 2 共 1/共 AA * 兲 0 兲 T 1 ⫹H 共 T 2 , . . . 兲 其
latter method, the free term in z 1 is a resonant term, that is,
⫻共 AA * 兲 2 . 共120兲
proportional to z 0 关see Eq. 共26兲兴. The equivalent statement in
the multiple time scales method, if the time dependence of The T 1 linear term in Eq. 共120兲 does not spoil the ordering of
the solution is constructed only from periodic functions 共that the magnitude of the zero- and first-order terms, because

548 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 548
thanks to Eq. 共117兲, this term generates O(T ⫺1
1 ) leading be- D 21 B⫹ 23 共 AA * 兲 D 1 B⫹ 43 A 2 D 1 B * ⫹ 649 共 AA * 兲 2 B
havior in (A * B⫹AB * ).
Equation 共119兲 is solved by ⫹ 329 共 AA * 兲 A 2 B * ⫹ 512
81
iA 4 A * 3 ⫽0. 共126兲
Clearly, B⫽0 is allowed only for the trivial case A⫽0. On
共 A * B⫺AB * 兲 ⫽ 兵 ⫺ 329 i 共 AA * 兲 ⫺2iD 2 ␸ A T 1
the other hand, a detailed check shows that the choice of Eq.
⫹iJ 共 T 2 , . . . 兲 其 共 AA * 兲 . 共121兲 共125兲 is an allowed solution of Eq. 共126兲, and hence is con-
sistent with both solvability conditions, Eqs. 共114兲 and 共115兲.
By using Eqs. 共12兲 and 共121兲, the solution for B is found to In summary, we have to overcome two stumbling blocks.
be First, unlike the Duffing oscillator, the freedom of choice of
the free amplitudes, which the normal form analysis enjoys,
B⫽ 兵 关 21 H 共 T 2 , . . . 兲 ⫺ 649 i 兴 AA * ⫹ 21 关 D 2 共 1/共 AA * 兲 0 AA * T 1 is lost due to the requirement that the solvability conditions
⫹iJ 共 T 2 , . . . 兲兲兴 其 A⫺iD 2 ␸ A AT 1 . 共122兲 for different orders be mutually consistent. Second, as for the
Duffing oscillator, the method of multiple time scales does
Here, H and J are real valued functions of higher time scales. not provide the dependence of the solution on T 2 , or, for that
Again, the T 2 dependence of A is not specified; its determi- matter, any time scales beyond T 1 , without additional infor-
nation requires additional input that is not provided by the mation, consistent with the solvability conditions.
method of multiple time scales and may be fixed as follows. Problem 8. Prove the statement following Eq. 共126兲.
The last term in Eq. 共122兲 generates a term proportional to
1 in B. For the allowed time span, t⫽O(1/⑀ ), for which
T 1/2
the approximation is valid, this term does not destroy the
C. The Van der Pol oscillator
ordering in the near-identity transformation. However, be-
cause a T 1/2
1 term does not agree with the expected structure We follow the steps given in detail in Sec. IV, employ the
of the approximate solution, it makes sense to eliminate it by same notation, and obtain the first-order results:
requiring that
D 1 A⫽ 12 A 共 1⫺ 14 AA * 兲 , 共127兲
D 2 ␸ A ⫽0. 共123兲
z 1 ⫽⫺ 1
16 iA 3 e ⫺3iT 0 ⫹ 兵 ⫺ 116 AA * ⫹ A * 其 ie
2 1
4
iT 0

With this choice, there is no secular behavior in B, because

its asymptotic T 1 dependence becomes that of A, namely, ⫺ 321 iA * 3 e 3iT 0 ⫹Be ⫺iT 0 . 共128兲
O(T ⫺1/2
1 ). Hence, there is no constraint on D 2 (1/(AA * ) 0 ) in The second-order analysis yields
Eq. 共122兲, and, without additional input, we cannot deter- D 1 B⫹D 2 A⫽ 12 B⫺ 14 AA * B⫺ 18 A 2 B * ⫹ 18 iA⫺ 163 iA 2 A *
mine the T 2 dependence of A. An appealing choice is to
assume that A does not depend on T 2 . To this end, we have ⫹ 256
11
iA 3 A * 2 , 共129兲
to add to Eq. 共123兲 the requirement that
z 2 ⫽⫺ 1024
5
A 5 e ⫺5iT 0 ⫹ 兵 256
1
A 4 A * ⫺ 128
5
A 3 ⫺ 163 iA 2 B 其 e ⫺3iT 0
D 2 共 1/共 AA * 兲 0 兲 ⫽0. 共124兲
⫹ 兵 512
5
A 2 A * 3 ⫺ 321 AA * 2 ⫹ 14 iB * ⫺ 161 iA * 2 B
Equation 共122兲 for B then becomes
⫺ 18 iAA * B * 其 e iT 0 ⫹ 兵 128
1
A * 3 ⫹ 1024
1
AA * 4
B⫽ 兵 21 H 共 T 2 , . . . 兲 ⫺ 649 i 其 A 2 A * ⫹ 21 iJ 共 T 2 , . . . 兲 A. 共125兲
⫺ 323 iA * 2 B * 其 e 3iT 0 ⫹ 1536
5
A * 5 e 5iT 0 ⫹Ce ⫺iT 0 . 共130兲
Note that the first-order minimal normal form choice of the
normal form analysis, Eq. 共55兲, is a particular case of Eq. Direct solution of solvability conditions. Equation 共127兲
共125兲, with J⫽H⫽0. This latter choice is consistent with the yields
two solvability conditions. It turns out that the analysis in
higher orders allows for the cancellation of the dependence
of A on the higher time scales (T 3 ,...) as well. This cancel-
1
冉
D 1 共 AA * 兲 ⫽AA * 1⫺ AA * ,
4 冊 共131兲

lation is a reflection of the fact that for the cubically damped D 1 ␸ A ⫽0. 共132兲
oscillator, only two time scales, T 0 and T 1 , are required to
Equation 共131兲 is solved by
describe the physics. However, the equivalence between the
methods of normal forms and multiple time scales ends at 4 共 AA * 兲 0
this point. For example, in the normal form analysis we AA * ⫽ . 共133兲
共共 AA * 兲 0 ⫹ 共 4⫺ 共 AA * 兲 0 兲 exp共 ⫺T 1 兲兲
could chose the free resonant term to vanish, but here B
⫽0 is allowed only in the trivial case, A⫽0, as can be de- ␸ A and (AA * ) 0 do not depend on T 1 , but may be functions
duced directly by a detailed inspection of Eqs. 共114兲 and of T 2 , T 3 ,... .
共115兲, or the solution for B given by Eq. 共122兲. We multiply Eq. 共129兲 by A * and employ Eqs. 共127兲,
Problem 7. What happens to the magnitude of B of Eq. 共129兲, 共132兲, and 共133兲, and obtain for the real and imaginary
共122兲 for t⫽O(1/⑀ 2 )? Use Eq. 共117兲. parts:
Consistency of solvability conditions. The same conclu-
sions regarding the constraints on B may be reached more
easily by the requirement that the two solvability conditions,
D1 再 exp共 T 1 兲共 A * B⫹AB * 兲
共 AA * 兲 2
冎
冉 冊 冉 冊
Eqs. 共114兲 and 共115兲, are mutually consistent. If we apply D 2
exp共 T 1 兲 1
to Eq. 共114兲 and D 1 to Eq. 共115兲 and impose Eq. 共96兲, we ⫽D 2 ⫽D 2 , 共134兲
obtain: 共 AA * 兲 共 AA * 兲 0

549 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 549
D 2␸ A⫺
1
16
⫽D 1 ⫺ 再
共 A * B⫺AB * 兲
2iAA *
AA * ⬅4.
The choice of Eq. 共145兲 makes AA * independent of all
共145兲

⫹
1
16
11
log共 AA * 兲 ⫺ 共 AA * 兲 .
64 冎 共135兲
higher time scales. That is, the zero-order term must be cho-
sen as the asymptotic circle with radius 2:
A⫽2e i 共共 1/16兲 T 2 ⫹ ␺ A 共 T 3 ,T 4 , . . . 兲兲 . 共146兲
Equation 共134兲 is solved by

冉
Equations 共145兲 and 共146兲 imply that z 0 and z 1 are both
共 A * B⫹AB * 兲 ⫽exp共 ⫺T 1 兲 G 共 T 2 ,T 3 , . . . 兲 confined to their asymptotic limit. Hence, this choice does
not cover situations in which the initial condition is removed

⫹D 2 冉 1
共 AA * 兲 0
冊 冊
T 1 共 AA * 兲 2 , 共136兲
from the limit cycle by O( ⑀ ) or O(1), so that exponential
spiraling toward the limit cycle occurs only in orders higher
than first.
where G is a free real-valued function. Although all the terms Thè T 2 dependence of the phase, ␸ A , given by Eq. 共140兲,
in Eq. 共136兲 fall off exponentially, their ordering in the per- is required for consistency with the solvability conditions. In
turbation series is destroyed if the T 1 linear term is not elimi- the normal forms method, this behavior is attained only as
nated in Eq. 共136兲. To eliminate this term, we impose T 2 →⬁. For example, if we use Eqs. 共63兲 and 共64兲 for U 1
and U 2 , respectively, the normal form, Eq. 共9兲 becomes
D 2 共 AA * 兲 0 ⫽0⇒D 2 共 AA * 兲 ⫽0, 共137兲
ż 0 ⫽⫺iz 0 ⫹ 21 ⑀ 共 1⫺ 41 z 0 z *
0 兲 z 0 ⫹ ⑀ 兵 ⫺ 关 4 Re f 1 共 z 0 z 0* 兲
2 1
共 A * B⫹AB * 兲 ⫽exp共 ⫺T 1 兲 G 共 T 2 ,T 3 , . . . 兲共 AA * 兲 2 .
共138兲 ⫹ f ⬘1 共 z 0 z * 0 兲兴 z 0 z 0* ⫹ 256 iz 0 z *
0 ⫺ 16 iz 0 z 0*
2 3 2 2
0 兲共 1⫺ 4 z 0 z *
1 11 3

We note that the left-hand side of Eq. 共135兲 is independent of

T 1 关see Eq. 共132兲兴. The integration of Eq. 共135兲 over T 1 , ⫹ 18 iz 0 其 . 共147兲
therefore, yields 共with H a real valued function兲: The first-order term in Eq. 共147兲 ensures that ␳→2 as t

冉 D 2␸ A⫺
1
冊
T 1 ⫽⫺
共 A * B⫺AB̄ 兲 1
⫹ 16 log共 AA * 兲
→⬁. Through second order, the time dependence of the
phase is

再 冉 冊 冎
16 2iAA *
d␸A 11 4 3 2 1 1
⫺ 11
64 共 AA * 兲 ⫹H 共 T 2 ,T 3 , . . . 兲 . 共139兲 ⫽⫺ ⑀ 2 ␳ ⫺ ␳ ⫹ ⫺Im f 1⬘ 共 ␳ 2 兲 1⫺ ␳ 2 ␳ 2 .
dt 256 16 8 4
The last three terms on the right-hand side of Eq. 共139兲 can- 共148兲
not counterbalance the linear behavior of the left-hand side
Thus, in general, Eq. 共140兲 of the method of multiple time
for T 1 →⬁. Hence, to avoid a T 1 secular term in B, the
scales is reached only for t→⬁. To obtain Eq. 共140兲 as the
left-hand side of Eq. 共139兲 must vanish, yielding
exact ⑀ 2 dependence of ␸ A in the normal forms analysis, we
D 2 ␸ A ⫽ 161 ⇒ ␸ A ⫽ 161 T 2 ⫹ ␺ A 共 T 3 ,T 4 , . . . 兲 , 共140兲 must choose for the free function, f 1 :
共 A * B⫺AB * 兲 ⫽ 兵 161 log共 AA * 兲 ⫺ 64
11
共 AA * 兲 Im f 1 共 ␳ 2 兲 ⫽ 161 ln ␳ 2 ⫺ 64 ␳ ⫹C,
11 2
共149兲

⫹H 共 T 2 ,T 3 , . . . 兲 其 2i 共 AA * 兲 . 共141兲 where C is an arbitrary constant. In the normal forms method

this is a possible choice. Equation 共142兲 shows that in mul-
From Eqs. 共138兲 and 共141兲, we find tiple time scales it is the only possibility.
Consistency of solvability conditions. If we apply the de-
B⫽ 兵 21 G 共 T 2 ,T 3 , . . . 兲 exp共 ⫺T 1 兲共 AA * 兲 rivative D 1 to Eq. 共129兲 and D 2 to Eq. 共127兲 and use both
⫹iH 共 T 2 ,T 3 , . . . 兲 ⫹ 161 i log共 AA * 兲 ⫺ 64
11
i 共 AA * 兲 其 A. equations repeatedly, we find that Eq. 共96兲 leads to:
共142兲 D 21 B⫺D 1 B⫹ 21 AA * D 1 B⫹ 14 A 2 D 1 B * ⫹ 14 B⫹ 641 共 AA * 兲 2 B
The characteristics of the solution are obviously the same as ⫹ 321 A 3 A * B * ⫹ 163 iA 2 A * ⫺ 128
17
iA 3 A * 2 ⫹ 512
11
iA 4 A * 3 ⫽0.
in the method of normal forms. Namely, through second or- 共150兲
der, the zero-order term asymptotes to a limit circle of radius
2. Had we continued the analysis to third order, we would If we multiply Eq. 共150兲 by A * , we obtain for the real and
have found an ⑀ 2 correction to the radius of the limit circle. imaginary parts:
However, here comes the difference. In the method of nor- D 21 兵 exp共 T 1 兲共 A * B⫹AB * 兲 / 共 AA * 兲 2 其 ⫽0, 共151兲

再 冎
mal forms, we saw that it is possible to choose the free term
in first order to vanish 关 f 1 of Eq. 共61兲兴. The equivalent re- 共 A * B⫺AB * 兲 1 11
D 21 ⫹ i log共 AA * 兲 ⫺ i AA * ⫽0.
quirement here is 共 AA * 兲 8 32
B⬅0. 共143兲 共152兲
A detailed analysis shows that Eqs. 共151兲 and 共152兲 are con-
However, unlike the energy conserving Duffing oscillator,
sistent with Eqs. 共134兲 and 共139兲, respectively, from which
the solvability conditions impose restrictions on this choice.
the solution for B, Eq. 共142兲, has been obtained. Hence, the
From Eq. 共142兲 we find that due to the T 1 dependence of A, consistency test is equivalent to the direct analysis of the
Eq. 共143兲 is satisfied only if the solution vanishes identically solution of the solvability conditions. In particular, the
or is already on the limit cycle:
choice B⫽0, is consistent with Eq. 共152兲 only for A⫽0 or
A⫽0, 共144兲 AA * ⫽4.

550 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 550
VII. CONCLUDING COMMENTS bation is added to the Burgers equation 共which describes the
propagation of shock waves in a fluid or traffic flow on a
Our goal has been to give a tutorial on the application of highway兲 or the Korteweg–de Vries equation 共which de-
two commonly used perturbation methods, normal forms and scribes the propagation of shallow water waves兲, then new
multiple time scales. To simplify the discussion, the analysis consistency problems called ‘‘obstacles to integrability’’ are
has been confined to single oscillator systems that are subject encountered.18 –22 The unperturbed equations have interesting
to nonlinear perturbations. We have shown the effects of nonlinear wave solutions 共for example, solitons兲. Often, the
typical nonlinear perturbations: modification of the period perturbative expansion generates terms that spoil the simple
for energy conserving perturbations and damping- or limit- structure of the unperturbed solution, and the zero-order ap-
cycle phenomena for dissipative perturbations. Both pertur- proximation 共which is determined by the normal form兲 may
bation methods focus on the manner in which the perturba- not have the same wave structure. The resolution of this open
tion modifies the zero-order approximation in comparison problem is under investigation.47
with the naive expansion. In both methods, the effect of the
perturbation on the solution for the zero-order approximation a兲
Electronic mail: [email protected]
1
is identified as the resonant terms that are generated by the H. Poincaré, New Methods of Celestial Mechanics, originally published as
perturbation in every order. If, as is done in the naive expan- Les Methodes Nouvelles de la Mechanique Celeste 共Gauthier-Villars,
Paris, 1892兲 and 共AIP Press, New York, 1993兲.
sion, they are not accounted for by their effect on the zero- 2
G. D. Birkhoff, Dynamical Systems 共American Mathematical Society Col-
order term, they lead to the emergence of secular terms. The loquium, New York, 1927兲.
goal is to generate a zero-order approximation that contains 3
G. I. Hori, ‘‘Theory of general perturbations with unspecified canonical
the major physical effects of the perturbation, but is still a variables,’’ Publ. Astron. Soc. Jpn. 18, 287–296 共1966兲.
small modification of the solution of the unperturbed equa- 4
A. Deprit, ‘‘Canonical transformations depending on a parameter,’’ Celest.
tion. Mech. 1, 12–13 共1969兲.
5
Our analysis has stressed the role played by the freedom G. I. Hori, ‘‘Theory of general perturbations for non-canonical systems,’’
inherent in perturbation theory. There is a one-to-one corre- Publ. Astron. Soc. Jpn. 23, 567–587 共1971兲.
6
spondence between the exploitation of this freedom in the V. I. Arnold, Geometrical Methods in the Theory of Ordinary Differential
Equations 共Springer-Verlag, New York, 1988兲, 2nd ed.
application of the two methods to a harmonic oscillator that 7
A. D. Bruno, Local Methods in Nonlinear Differential Equations
is modified by an energy conserving perturbation. This cor- 共Springer-Verlag, New York, 1989兲.
respondence is lost when the methods are applied to dissipa- 8
A. H. Nayfeh, Method of Normal Forms 共Wiley, New York, 1993兲.
tive systems for which only a subset of the choices available 9
J. Guckenheimer and P. J. Holmes, Nonlinear Oscillations, Dynamical
in the method of normal forms is allowed in the method of Systems, and Bifurcations of Vector Fields 共Springer-Verlag, New York,
multiple time scales. 1988兲.
10
Because both methods have been applied to a much wider S. Wiggins, Introduction to Applied Nonlinear Dynamical Systems and
class of systems, we know that the loss of freedom in the Chaos 共Springer-Verlag, New York, 1990兲.
11
A. H. Nayfeh, Perturbation Methods 共Wiley, New York, 1973兲.
multiple time scales method extends beyond the systems 12
C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for
with a single degree of freedom and applies to most dynami- Scientists and Engineers 共McGraw–Hill, New York, 1978兲.
cal systems of physical interest. We provide here an intuitive 13
A. H. Nayfeh, Introduction to Perturbation Techniques 共Wiley, New York,
argument. Consider a system of two harmonic oscillators that 1981兲.
have equal unperturbed frequencies and are coupled by an 14
J. Kevorkian and J. D. Cole, Perturbation Methods in Applied Mathemat-
energy conserving interaction. The combined system is en- ics 共Springer-Verlag, New York, 1985兲.
ergy conserving. However, because of the equality of the
15
J. A. Murdock, Perturbation Theory and Methods 共Wiley, New York,
unperturbed frequencies, energy flows efficiently back and 1991兲.
16
J. Kevorkian and J. D. Cole, Multiple Scale and Singular Perturbation
forth between the two oscillators due to the coupling be-
Methods 共Springer-Verlag, New York, 1996兲.
tween them. ‘‘Efficiently’’ means that, although the coupling 17
P. B. Kahn and Y. Zarmi, Nonlinear-Dynamics: Exploration Through Nor-
is weak, characterized by a small parameter, ⑀, the energy mal Forms 共Wiley, New York, 1998兲.
flow between the two oscillators, while slow 关becoming sig- 18
Y. Kodama and T. Taniuti, ‘‘Higher order approximation in the reductive
nificant only for t⫽O(1/⑀ )], is not small. Thus, each sub- perturbative method,’’ J. Phys. Soc. Jpn. 45, 298 –310 共1978兲.
19
system looks like a dissipative one. The multiple time scales A. S. Fokas and Q. M. Liu, ‘‘Asymptotic integrability of water waves,’’
analysis of this system reveals that the freedom of choice of Phys. Rev. Lett. 77, 2347–2351 共1996兲.
20
the free amplitudes is limited by the requirement that the Y. Kodama, ‘‘Normal form and solitons,’’ in Topics in Soliton Theory and
Exactly Solvable Nonlinear Equation, edited by M. J. Ablowitz et al.
solvability conditions for different orders be mutually
共World Scientific, Singapore, 1987兲, pp. 319–340.
consistent.46 The same applies to the application of the mul- 21
R. A. Kraenkel, J. G. Pereira, and E. C. de Rey Neto, ‘‘Linearizability of
tiple time scales method to systems described by PDEs, be- the perturbed Burgers equation,’’ Phys. Rev. E 58, 2526 –2530 共1998兲.
cause they are systems with an infinite number of degrees of 22
Y. Hiraoka and Y. Kodama, ‘‘Normal form and solitons,’’ Lecture notes,
freedom. 共For example, expanding the solution in a Fourier Euro Summer School 2001, The Isaac Newton Institute, Cambridge,
series in the spatial coordinate, leads to an infinite set of 15–25 August, 2002 and arXiv:nlin.SI/0206021.
23
ODEs for the time dependence of the Fourier coefficients.兲 A. C. Newell, ‘‘Envelope equations,’’ Lect. Appl. Math. 15, 157–163
The lesson from the examples discussed here is that in every 共1974兲.
application of the method that goes beyond a first-order
24
Y. Kuramoto, Chemical Oscillations, Waves and Turbulence 共Springer-
Verlag, New York, 1984兲.
analysis, the choices made for the perturbative solution must 25
P. Manneville, Dissipative Structures and Weak Turbulence 共Academic,
ensure that the solvability conditions for different orders be New York, 1990兲.
obeyed in a mutually consistent manner. 26
C. M. Cross and P. C. Hohenberg, ‘‘Pattern formation outside of equilib-
The application of the normal forms method to systems of rium,’’ Rev. Mod. Phys. 65, 851–112 共1993兲.
physical interest of spatial extent encounters other unex- 27
B. Van der Pol, ‘‘On oscillation hysteresis in a simple triode generator,’’
pected difficulties. For example, if a small nonlinear pertur- Philos. Mag. 43, 700–719 共1926兲.

551 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 551
28
Lord Rayleigh 共J. W. Strutt兲, ‘‘On maintained vibrations,’’ Philos. Mag. harmonic oscillator systems with nonlinear perturbations,’’ Physica D 118,
XV, 229–232 共1883兲. 221–249 共1998兲.
29 39
M. Kummer, ‘‘How to avoid secular terms in classical and quantum me- Z. Rahman and T. Burton, ‘‘On higher order methods of multiple scales in
chanics,’’ Nuovo Cimento Soc. Ital. Fis., B 1B, 123–129 共1971兲. non-linear oscillators—Periodic steady state response,’’ J. Sound Vib. 133,
30
J. C. Liu, ‘‘The uniqueness of normal forms via Lie transformations and its 369–379 共1989兲.
application to Hamiltonian systems,’’ Celest. Mech. 36, 89–95 共1985兲. 40
J. Murdock, ‘‘Some fundamental issues in multiple scale theory,’’ Appl.
31
J. C. Van der Meer, The Hamiltonian Hopf Bifurcation 共Springer-Verlag, Anal. 53, 157–173 共1994兲.
New York, 1985兲. 41
Y. Kodama, ‘‘Normal forms, symmetry and infinite-dimensional Lie alge-
32
A. Baider and R. C. Churchill, ‘‘Uniqueness and non-uniqueness of non- bra for systems of ODE’s,’’ Phys. Lett. A 191, 223–228 共1994兲.
uniqueness of normal form expansions for vector fields,’’ Proc. R. Soc. 42
J. Murdock and L. C. Wang, ‘‘Validity of the multiple scale method for
Edinburgh, Sect. A: Math. 108, 27–33 共1988兲.
33 very long intervals,’’ ZAMP 47, 760–789 共1996兲.
A. Baider, ‘‘Unique normal forms for vector fields and Hamiltonians,’’ J. 43
A. Degasperis, S. V. Manakov, and S. M. Santini, ‘‘Multiple-scale pertur-
Diff. Eqns. 78, 33–39 共1989兲.
34
P. B. Kahn and Y. Zarmi, ‘‘Minimal normal forms in harmonic oscillations bation beyond the nonlinear Schrödinger equation. I,’’ Physica D 100,
with small nonlinear perturbations,’’ Physica D 54, 65–74 共1991兲. 187–211 共1997兲.
44
35
P. B. Kahn and Y. Zarmi, ‘‘Radius renormalization in limit cycles,’’ Proc. C. Rega, W. Lacarbonara, A. H. Nayfeh, and C. M. Chin, ‘‘Multiple reso-
R. Soc. London, Ser. A 440, 189–199 共1993兲. nance in suspended cables: Direct versus reduced-order models,’’ Int. J.
36
P. B. Kahn, D. Murray, and Y. Zarmi, ‘‘Freedom in small parameter ex- Non-Linear Mech. 34, 901–924 共1999兲.
45
pansion for nonlinear perturbations,’’ Proc. R. Soc. London, Ser. A 443, G. Auberson and M. Capdequi Peyranère, ‘‘Quantum anharmonic oscilla-
83–94 共1993兲. tor in the Heisenberg picture and multiple scale techniques,’’ Phys. Rev. A
37
E. Forest and D. Murray, ‘‘Freedom in minimal normal forms,’’ Physica D 65, 032120-1–9 共2002兲.
74, 181–196 共1994兲. 46
P. B. Kahn and Y. Zarmi 共unpublished兲.
38
P. B. Kahn and Y. Zarmi, ‘‘Reduction of secular error in approximations to 47
A. Veksler and Y. Zarmi 共unpublished兲.

552 Am. J. Phys., Vol. 72, No. 4, April 2004 P. B. Kahn and Y. Zarmi 552