Nonlinear Dynamics 38: 322, 2004.

C
2004 Kluwer Academic Publishers. Printed in the Netherlands.
Averaging Oscillations with Small Fractional Damping
and Delayed Terms
PANKAJ WAHI

and ANINDYA CHATTERJEE

Department of Mechanical Engineering, Indian Institute of Science, Bangalore 560012, India;

Author for correspondence (e-mail: pankaj@mecheng.iisc.ernet.in; fax: +91-80-23600648)

(Received: 12 November 2003; accepted: 11 March 2004)
Abstract. We demonstrate the method of averaging for conservative oscillators which may be strongly nonlinear, under small
perturbations including delayed and/or fractional derivative terms. The unperturbed systems studied here include a harmonic
oscillator, a strongly nonlinear oscillator with a cubic nonlinearity, as well as one with a nonanalytic nonlinearity. For the latter
two cases, we use an approximate realization of the asymptotic method of averaging, based on harmonic balance. The averaged
dynamics closely match the full numerical solutions in all cases, verifying the validity of the averaging procedure as well as the
harmonic balance approximations therein. Moreover, interesting dynamics is uncovered in the strongly nonlinear case with small
delayed terms, where arbitrarily many stable and unstable limit cycles can coexist, and innitely many simultaneous saddle-node
bifurcations can occur.
Key words: delay differential equations, fractional derivatives, harmonic balance, averaging, strongly nonlinear oscillations,
saddle-node bifurcations
1. Introduction
In this paper we use the method of averaging to study small perturbations of conservative oscillators.
The unperturbed conservative oscillators need not be linear. The perturbations studied involve delayed
terms as well as fractional derivative terms.
Differential equations with delayed terms (DDEs) have applications in manufacturing processes,
control systems, biology, economics, chemical kinetics and other areas [19]. Fractional derivatives
(which incorporate distributed delay effects with an innite delay) are among other things useful in
describing the frequency dependent damping behavior of many materials [1013].
Let us rst consider problems where the unperturbed system is linear.
Analytical treatment of even linear problems involving delayed terms or fractional derivatives is
nontrivial. However, when the delayed terms or the fractional derivative terms in the equation are small,
the systems can be studied analytically using perturbation methods.
The method of multiple scales, which is similar to the method of averaging [14], has been used to
study linear oscillators perturbed by small delayed terms [1517].
For linear oscillators perturbed by small fractional derivative terms, Li et al. [18] have used averaging.
Their approach indirectly recovers the averaged equations after several modications of the averaging
method for the periodic case. The problem, however, is directly amenable to averaging for the aperiodic
case, for which a short and simple procedure exists [19]. A similar problem has also been studied using
multiple scales by Rossikhin and Shitikova [20].
In this paper, after briey studying the linear/weakly nonlinear cases involving delayed terms and/or
fractional derivatives, we use averaging to study the dynamics of some strongly nonlinear conservative
4 P. Wahi and A. Chatterjee
oscillators with small perturbations in the form of delays and/or fractional derivatives. This is the
main contribution of the present paper. The method of averaging, implemented in an approximate
sense as in [21], gives slow ows governing the evolution of amplitude and phase of the oscilla-
tions under study. The slow ow equations are ODEs, do not contain fractional derivatives or de-
layed terms, and are easy to study further using standard numerical and/or analytical methods for
ODEs.
2. Background and Numerics
2.1. DELAY DIFFERENTIAL EQUATIONS
Consider a DDE of the form
x(t ) = H(x(t ), x(t s
1
), x(t s
2
), . . . , x(t 1)), (1)
where x could be a vector and 0 < s
1
< s
2
< . . . < s
n
= 1 are nite time delays. Assumption of s
n
= 1
is allowed, because time can be scaled to make it so. At instant t , the evolution of x depends explicitly
on values it take in the immediately preceding unit interval [t 1, t ].
For the numerical integration of DDEs, in this paper, we use the following method. We t a cubic
spline over an interval somewhat longer than unity, say [t 1 10h, t ]; and interpolate, using the tted
spline, at all the desired time instants in the past. The portion of stored history longer than unity is
optional. It is used here because splines on intervals t better at points not too close to the endpoints.
This interpolation scheme is convenient for software like Matlab, which has inbuilt commands for
splines. Once the interpolated values are available, we simply use a xed-step fourth-order Runge
Kutta method. Note that all numerics in this paper are solely for checking the accuracy of analytical
approximations.
2.2. DIFFERENTIAL EQUATIONS WITH FRACTIONAL DERIVATIVES
We adopt the Riemann-Liouville denition of the fractional derivative [22],
D
q
a
[x(t )]
d
q
x(t )
[d(t a)]
q

1
(n q)
d
n
dt
n
__
t
a
x(u)
(t u)
qn1
du
_
,
where n is the smallest integer greater than or equal to q and represents the Gamma function. For
systems starting from rest (a case studied by many authors) we have x(t ) = 0 for t < 0, so D
q

is
identical to D
q
0
and we drop the subscript a. In this paper, we assume 0 < q < 1, the case most relevant
to structural damping. Hence,
D
q
x(t ) =
1
(1 q)
d
dt
_
t
0
x(u)
(t u)
q
du =
1
(1 q)
_
x(0)
t
q
+
_
t
0
x(t u)
u
q
du
_
.
The latter form above is useful to us. The rst term in it has a singularity at t = 0, and the integral
involves a singular kernel. The singularity due to the rst term can be treated
1
using a local series
1
Initial conditions for linear problems have also been included in, e.g., Laplace transform based treatments [25].
Averaging Fractional Derivatives and Delayed Terms 5
expansion near t = 0, but for simplicity we assume x(0) = 0 as in [23, 24]. Moreover, we change
variables to u = u
1q
, which removes the singularity in the kernel to give
D
q
x(t ) =
1
(2 q)
_
t
(1q)
0
x
_
t u
1/(1q)
_
d u.
Uniformly spaced points in the original time t are non-uniformly spaced in u. However, tting a
spline to x(t ) and interpolating at uniformly spaced u values, the integral can be evaluated numerically
(Simpsons rule). In this way, knowing the history of x(t ), we evaluate D
q
x(t ). Finally, a modied
fourth-order Runge Kutta routine, with access to the history needed for the integral, is used to integrate
forward in time. Note, again, that our numerics are solely for verifying analytical approximations.
We mention two nal points. First, the cubic spline interpolant has higher-order accuracy than Simp-
sons rule; so, for integration, we should interpolate at a ne mesh. Second, there might be numerical
difculties if q approaches 0 or 1. Here we focus on q away from these limits (e.g. q = 1/2).
3. Weakly Nonlinear Oscillations
3.1. THE METHOD OF AVERAGING
The classical method of averaging (see e.g. [19, 26, 27]) is applicable to systems of the form
x = G(x, t ) , 0 < 1. (2)
In this section, we study
x(t ) + x(t ) =

F, (3)
where

F can depend on delayed values and fractional derivatives of x.
We now cast Equation (3) in the form of Equation (2) (a routine procedure). For = 0, the solution
to Equation (3) is x(t ) = R sin(t +) for R and arbitrary constants. For small = 0 we let
x(t ) = R(t ) sin(t +(t )), (4)
where R(t ) and (t ) both change slowly with time. We also incorporate the usual constraint,
x(t ) = R(t ) cos (t +(t )) . (5)
Proceeding in the usual manner [19, 26, 27], we nally obtain

R = cos(t +) F (6)

=
sin(t +)
R
F, (7)
where F is simply

F with Equations (4) and (5) plugged in. Equations (6) and (7) have the form
of Equation (2). Now, depending on whether the function G is periodic or aperiodic in its explicit
dependence on time t , we have two cases:
6 P. Wahi and A. Chatterjee
1. If G = G(, t ) is periodic in t , i.e., G(, t + T
0
) = G(, t ) for some 0 < T
0
< , then the correct
averaging procedure (see Theorem 3.2.10 in [19]) is to hold x constant in the right hand side of
Equation (2)
2
,while t -averaging over the time period T
0
, as in
1
T
0
_
T
0
0
G dt.
2. If G is aperiodic in t , the correct averaging procedure (see Theorem 3.3.3 in [19]) is to average over
an innitely long time, as in (holding x constant)
lim
T
1
T
_
T
0
G dt.
3.2. WEAKLY NONLINEAR OSCILLATION WITH DELAYED TERMS
For illustration, consider
x(t ) + x(t ) = 2 x(t ) x(t )
3
+2ux
s
(t ) +2v x
s
(t ), (8)
where we introduce the notation x
s
(t ) = x(t s) and x
s
(t ) = x(t s), and , , u and v are small
quantities. This is the same equation as in Section 2.1 of [16], but without forcing. In [16], the method of
multiple scales was used; here, we will briey demonstrate the use of averaging. We substitute =

,
= , u = u and v = v, where 0 < 1. Note that we have replaced the delay (used in [16])
by s. We now have
x(t ) + x(t ) = (2

x(t ) x(t )
3
+2 ux
s
(t ) +2 v x
s
(t )). (9)
Substituting x(t ) = R(t ) sin (t +(t )) and following the routine steps described previously, we obtain

R(t ) = ( R(t )
3
cos(t +(t )) sin
3
(t +(t )) +2 u R
s
(t ) cos(t +(t )) sin(t s +
s
(t ))
+2 v R
s
(t ) cos(t +(t )) cos(t s +
s
(t )) +2

R(t ) sin
2
(t +(t )) 2

R(t ))
and

(t ) =
sin(t +(t ))
R(t )
(2

R(t ) cos(t +(t )) R(t )
3
sin
3
(t +(t ))
+2 u R
s
(t ) sin(t s +
s
(t )) +2 v R
s
(t ) cos(t s +
s
(t ))),
where we again denote delays using R
s
(t ) = R(t s) and
s
(t ) = (t s). For averaging, R and
are to be held constant while t -averaging the right hand sides of the previous equations. Hence, we
substitute R(t ) = R
s
(t ) = R and (t ) =
s
(t ) = , which gives

R = R( R
2
cos(t +) sin
3
(t +) +2 u cos(t +) sin(t s +)
+2 v cos(t +) cos(t s +) +2

(sin(t +))
2
2

)
2
Or to hold R and constant in the right hand sides of Equations (6) and (7).
Averaging Fractional Derivatives and Delayed Terms 7
Figure 1. Solution of Equation (8) and Equations (10) and (11) for s = 1, = = u = v = 0.05. Initial conditions:
x() = 5 sin(), x() = 5 cos(), for [1, 0], R(0) = 4.65, (0) = 0. For a discussion of the choice of initial conditions,
see Appendix A.
and

= sin(t +)(2

cos(t +) R
2
sin
3
(t +) +2 u sin(t s +) +2 v cos(t s +)).
The right hand sides of both equations above are 2-periodic in t . Averaging gives

R = R( u sin(s) v cos(s) +

) (10)
and

8
(3 R
2
+8 v sin(s) +8 u cos(s)). (11)
Going back to the old variables , , u and v, we get the same results as obtained using multiple
scales in [16]. For further verication, we have plotted the numerical solution of Equation (8) along
with that of the averaged equations, i.e., Equations (10) and (11) in Figure 1. The match is good.
We now turn to systems with light fractional derivative damping.
3.3. WEAKLY NONLINEAR OSCILLATIONS WITH FRACTIONAL DAMPING
In what follows, we take q = 1/2 in our fractional derivative, appropriate for a large variety of
viscoelastic damping materials (see Appendix Bfor q = 1/2). To demonstrate the averaging calculation
in this case, we consider
x(t ) +D
1/2
x(t ) + x(t ) = (t ). (12)
8 P. Wahi and A. Chatterjee
In the above (t ) represents an instantaneous impulse
3
at t = 0, and 0 < 1 (light damping).
After the impulse has acted, we have the initial conditions x(t ) = 0 for t 0, and x(0
+
) = 1; and the
governing equation for t > 0 is
x(t ) + x(t ) = D
1/2
x(t ). (13)
We again assume x(t ) = R(t ) sin(t +(t )) as described previously, and obtain

R = cos(t +)D
1/2
x(t ), (14)

=
sin(t +)
R
D
1/2
x(t ). (15)
For averaging, we write D
1/2
x(t ) in terms of R(t ) and (t ) as
D
1/2
x(t ) =
1
(1/2)
_
t
0
x
s
(t )

s
ds =
1

_
t
0
R
s
(t ) cos(t +
s
(t ) s)

s
ds. (16)
We then write (recalling that we will set R(t ) = R
s
(t ) = R and (t ) =
s
(t ) = , to be held constant
within the averaging integral)
D
1/2
x(t ) =
1

_
t
0
R cos(t + s)

s
ds =
R cos(t +)

_
t
0
cos(s)

s
ds +
R sin(t +)

_
t
0
sin(s)

s
ds.
(17)
The integrals above are aperiodic functions of t (Case 2 in Section 3.1), so we use an innite time
average (unlike the case with merely delayed x). Substituting for D
1/2
x(t ) in Equation (14) will require
evaluation of
lim
T
1
T
_
T
0
cos(t +)
_
R cos(t +)

_
t
0
cos(s)

s
ds +
R sin(t +)

_
t
0
sin(s)

s
ds
_
dt. (18)
Integrating the rst term above by parts, we get
lim
T
R
4

_
__
2 t
T
+
sin(2 t )
T
__
t
0
cos(s)

s
ds
_
T
0

1
T
_
T
0
_
(2

t cos(t ) +
sin(2 t ) cos(t )

t
_
dt
_
.
The above can be simplied to
R
2

lim
T
__
T
0
cos(s)

s
ds
1
T
_
T
0

t cos(t )dt
_
. (19)
Integrating the second term above by parts, we have
_
T
0

t cos(t )dt =

t sin(t )

T
0

_
T
0
sin(t )
2

t
dt.
3
The instantaneous impulse is the Dirac delta function, which is a generalized function that is zero everywhere except at t = 0,
undened at t = 0, and whose integral over any nonzero interval containing t = 0 is unity.
Averaging Fractional Derivatives and Delayed Terms 9
Substituting for
_
T
0

t cos(t ) dt in Equation (19), we get (on dropping terms that go to zero)

R
2

lim
T
_
T
0
cos(s)

s
ds =
R
2

2
.
Thus, the rst term of expression (18) is
lim
T
1
T
_
T
0
R cos
2
(t +)

_ _
t
0
cos(s)

s
ds
_
dt =
R
2

2
.
By similar calculations,
lim
T
1
T
_
T
0
R cos(t +) sin(t +)

_ _
t
0
sin(s)

s
ds
_
dt = 0.
Using the above, we get the averaged equations for

R as

R =
1
2

2
R.
In a similar fashion, we also get

=
1
2

2
.
The above results could be obtained using the following alternative and informal argument. The
integrals
_
t
0
sin(s)

s
ds and
_
t
0
cos(s)

s
ds are bounded for small t in spite of the singular integrand in the latter.
In the innite-time average, therefore, only the behaviors for t matter. For large t , we have
_
t
0
cos(s)

s
ds =
_

2
+
sin(t )

t
+O(t
3/2
) (20)
and
_
t
0
sin(s)

s
ds =
_

2

cos(t )

t
+O(t
3/2
). (21)
For our averaging purposes, we need to retain only the leading order, or constant, terms. The con-
tributions from the correction terms go to zero. Therefore, for our purposes, on substituting Equations
(20) and (21) in Equation (17), we have
D
1/2
x(t ) =
R cos(t +)

2
+
R sin(t +)

2
+negligible terms. (22)
Substituting Equation (22) in Equations (14) and (15) and dropping the negligible terms, we get

R =
R

2
(cos
2
(t +) +sin(t +) cos(t +)), (23)
10 P. Wahi and A. Chatterjee
Figure 2. Impulse response of Equation (12) for = 0.1. Initial conditions: x(0) = 0 and x(0) = 1.The solution of Equation
(25) is also shown for R(0) = 1.

=
1

2
(sin(t +) cos(t +) +sin
2
(t +)). (24)
On averaging the above, where we note that the aperiodicity is gone and we are back to the periodic
case, we obtain exactly as mentioned previously

R =
1
2

2
R (25)
and

=
1
2

2
. (26)
These equations govern the evolution of the amplitude R and phase of the unforced oscillations
(though there is an instant of forcing, we examine the subsequent unforced oscillations only). The
solution for R is compared in Figure 2 with the exact solution for the original equation (using the
Laplace transform; see [22]). The match is excellent.
We have now succeeded in averaging harmonic oscillators slightly perturbed with delayed terms and
fractional derivatives. For the delayed terms, our results match those obtained earlier by other authors
using the method of multiple scales. For fractional derivatives, we have shown how aperiodic averaging
can be used more simply than the approach developed in [18].
We now move on to strongly nonlinear oscillators, where the main contributions of this paper lie.
4. Strongly Nonlinear Oscillations
In this section, we adopt a recently developed technique called harmonic balance-based averaging [21]
to tackle small perturbations of some strongly nonlinear oscillators.
Averaging Fractional Derivatives and Delayed Terms 11
4.1. HARMONIC BALANCE BASED AVERAGING
Averaging can not only be used to study small perturbations of the harmonic oscillator, but also (in
principle at least) to perturbations of strongly nonlinear oscillators. However, there are practical dif-
culties in applying averaging to such strongly nonlinear problems. An approximate method for doing
the averaging, called harmonic balance-based averaging (HBBA) [21], is used in this paper. The method
is outlined here for completeness. Consider
x + g(x, x) = F, (27)
where F is possibly a function of x, x, D
1/2
x and/or x(t 1) or x(t 1), and g(x, x) has a two-parameter
family of periodic solutions. Consider Equation (27) for = 0,
x + g(x, x) = 0. (28)
Equation (28) has a periodic solution expressible abstractly as x =

h(R, (R), t ), where R represents
the amplitude of the solution, and the time period is T = 2/(R). The general solution is then
x =

h(R, (R), t +

), (29)
where both R and

are arbitrary constants (dening a two-parameter family of periodic solutions). We
rescale time through
d
dt
= (R()),
where (R()) is the same as in Equation (29). Letting primes denote -derivatives, Equation (27)
becomes

2
x

R
R

+ g = F. (30)
In -time, the unperturbed solution (Equation (29)) may be rewritten as some new function
x = h(R, , ) (31)
that is 2-periodic in , and where R and are constants for = 0. We will take the above to be the
solution for Equation (30), with R and as slowly varying functions of for 0 < 1. We also
choose the constraint equation
x

=
h

. (32)
From Equations (31) and (32), we obtain
h
R
R

+
h

= 0. (33)
12 P. Wahi and A. Chatterjee
Similarly, from Equations (30) and (32), we obtain

2
_

2
h

2
+

2
h
R
R

+

2
h

_
+

R
h

+ g = F. (34)
In Equation (34), we have

2
h

2
+ g = 0
because h is an exact solution. Thus, we obtain
_

2

2
h
R
+

R
h

_
R

+
2

2
h

= F. (35)
Solving Equations (33) and (35), we obtain (using subscripts to denote partial derivatives)
R

=
h

2
(h
R
h

h
R
)
R
h

, (36)

=
h
R
F

2
(h
R
h

h
R
)
R
h

. (37)
Equations (36) and (37) can be -averaged in principle. Subsequently, using the fact that = , we
can recover the slow ow for R and in terms of the original unscaled time t .
The harmonic balance approximations of HBBA enter at this point. Since we do not have

h in closed
form, we use an approximation of the form

h C
0
(R) + R sin((R)t ) +
N

k=2
{B
k
(R) sin(k(R)t ) +C
k
(R) cos(k(R)t )},
where R represents the amplitude of oscillation, the angular frequency (R) determines the fundamental
period, and the B
k
s, C
k
s and are known functions of R obtained using harmonic balance (or, in fact,
any other good approximation method). Replacing (R)t with + , we get an approximation to h
from

h.
Substituting the above approximations for h and in Equations (36) and (37) provides closed form
approximations to the true O() vector eld, in the form of Equation (2), and which can in principle
be averaged. However, averaging typically presents difculties in practice in that the required integrals
cannot be found in closed form. Here again, we use harmonic balance.
We assume the quantity to be averaged can be written as
r
1
(R,,)
r
2
(R,,)
, where r
1
and r
2
are trigonometric
polynomials. We choose an integer N 1 and approximate
r
1
(R, , )
r
2
(R, , )
a
0
+
N

k=1
(a
k
cos k +a
k+N
sin k),
which in turn we rewrite as
r
1
(R, , ) r
2
(R, , )
_
a
0
+
N

k=1
(a
k
cos k +a
k+N
sin k)
_
. (38)
Averaging Fractional Derivatives and Delayed Terms 13
The above yields a system of 2N +1 simultaneous linear equations by harmonic balance [21], which
are solved for a
k
, k = 0, 1, , 2N. The coefcient a
0
then approximates the average. If k harmonics
are used in the initial approximation to the periodic solutions, and if N harmonics are used in the
averaging procedure described previously, then the procedure is referred to as (k, N) averaging.
4.2. DELAYED TERMS
Consider a cubic oscillator with a small delayed term
x(t ) + x(t )
3
= x(t 1). (39)
We take two term harmonic balance and assume a solution of the form
x = R sin t +R sin 3t. (40)
On balancing harmonics (see e.g. [21]), we nd 3/67. Also, 8R/9. The approximate
solution h of Equation (31) is
x = R
_
sin( +)
3
67
sin(3 +3)
_
. (41)
Also, x(t 1) = x( ). Substituting for F = x( ) in Equations (36) and (37) (the expressions
are lengthy and not reproduced here), and carrying out (2, 6) averaging, we get
R

=

R
_
0.41284 sin
_
8R
9
_
+0.00226 sin
_
8R
3
__
,

=

R
2
_
0.41599 cos
_
8R
9
_
+0.00116 cos
_
8R
3
__
.
Multiplying the above with (R), we get the averaged equations in the original time t as

R =
_
0.36697 sin
_
8R
9
_
+0.00201 sin
_
8R
3
__
, (42)
and

=

R
_
0.36977 cos
_
8R
9
_
+0.00103 cos
_
8R
3
__
. (43)
It is interesting to note, in the slow ows above, the trigonometric dependence on R. This is a
consequence of the delay as well as the strong nonlinearity. The slow ow predicts innitely many
limit cycles. For limit cycles with very large R, however, may need to be very small indeed for the
approximation to be good.
Solving the above two equations numerically along with
= (R) =
8
9
R
14 P. Wahi and A. Chatterjee
Figure 3. Numerical solution of Equation (39) and Equations (42) and (43) for = 0.1. Initial conditions: x() = 0.1 cos(),
x() = 0.1 sin(), [1, 0], R(0) = 0.14, (0) =

2
. Averaged amplitude refers to
70
67
R.
gives the response of the system through Equation (41). Results are shown in Figure 3. Note that the
maximum value of the approximate solution is actually x = (70/67)R, and so a (70/67) scaled version
of R has been plotted to represent the averaged amplitude. The match is good, both for amplitude and
phase (the latter shows a small error over many cycles).
4.3. FRACTIONAL DAMPING
We now consider the same cubic oscillator, but with light fractional damping
x + x
3
= D
1/2
x. (44)
We use the same two term harmonic balance as before.
The relation between D
1/2
x(t ) and D
1/2
x() is obtained by introducing u = s in
D
1/2
x(t ) =
1
(1/2)
_
t
0
x(t s)

s
ds =
1
(1/2)
_

0
x

( u)

u/

du

,
where, for our averaging purposes, is constant within the integral. This results in
D
1/2
x(t ) =

D
1/2
x().
Using simplications similar to those of Section 3.3, we nd that we may use
D
1/2
x() =
2
201
R
3/2
(3

3 sin(3 +3 ) +67 cos( +) +67 sin( +)

3

3 cos(3 +3 )). (45)

Averaging Fractional Derivatives and Delayed Terms 15
Substituting for F = D
1/2
x in Equations (36) and (37), we get the slow ow equations governing the
evolution of R and . After (2, 6) averaging and multiplying by (R), we obtain

R = 0.24697

R , (46)

=
0.24770

R
. (47)
For numerical comparisons, initial conditions for the averaged equations are obtained as follows. As
mentioned earlier, x(0) = 0 and x(0) > 0. From Equation (41),
x(0) = R(0)
_
sin((0))
3
67
sin(3(0))
_
= 0,
where R(0) = 0 implies (0) = 0. Also
x(0) =
dx()
d

=0
d
dt

=0
= R(0)
_
cos((0))
9
67
cos(3(0))
_
(R(0))
=
8
9
R
2
(0)
_
cos((0))
9
67
cos(3(0))
_
.
Substituting (0) = 0 and solving, we get
R(0) =
_
9
8
_
67
58
_
x(0) =
_
603
464
_
x(0) = 1.14
_
x(0).
The numerically computed averaged amplitude, (70/67)R as before, compares well with the full
numerical solution in Figure 4.
Figure 4. Numerical solution of Equation (44) and Equations (46) and (47) for = 0.1. Initial conditions : x(0) = 0, x(0) = 3,
R(0) = 1.14

3 = 1.9745, (0) = 0.
16 P. Wahi and A. Chatterjee
4.4. COEXISTING TRADITIONAL, DELAYED AND FRACTIONAL DAMPING
It is interesting to consider the following system, which has three types of damping (traditional, delayed
and fractional order)
x(t ) + x(t )
3
= ( x(t ) + D
1/2
x(t ) + x(t 1)). (48)
Proceeding as above, the averaged equations are

R =
_
0.33155 R 0.32619 R cos
_
8R
9
_
0.00536 R cos
_
8R
3
_
0.24697

R
_
,

=
_
0.24770

R
+0.32868 sin
_
8R
9
_
+0.00274 sin
_
8R
3
__
.
Figures 5 through 8 present numerical results for Equation (48) and the averaged equations, for
different values of , and .
For = 0, i.e., in the absence of delayed damping, the amplitude evolves as

R =

R(0.33155

R 0.24697 ).
There are no limit cycles.
For = 0, the oscillator has traditional as well as delayed damping. The equation for amplitude is
then

R = R
_
0.33155 0.32619 cos
_
8R
9
_
0.00536 cos
_
8R
3
__
.
Figure 5.

R versus R for = 0, (a) = 1, = 0.9, (b) = 1, = 1.
Averaging Fractional Derivatives and Delayed Terms 17
Figure 6. Numerical solution of Equation (48) and the corresponding averaged equations for = 0.1, = 0, = 1, = 0.9.
Initial conditions: x() = 12 cos(), x() = 12 sin(), [1, 0], R(0) = 67/70 12 = 11.4857, (0) =

2
.
Figure 7. Plot of

R versus R for = 0, (a) = 1, = 1, (b) = 1, = 0.25.
Plots of

R versus R for two sets of and are shown in Figure 5. For = 0.9 and = 1, all the
solutions approach R = 0 (there are no limit cycles). Holding = 1 xed and increasing through 1,
there occur innitely many simultaneous saddle-node bifurcations, and innitely many pairs of stable
and unstable limit cycles are born (by rst order averaging). Numerical solutions are shown in Figure
6 for = 0.9 and = 1. The amplitude decay rates correspond well with Figure 5.
Finally, we consider = 0, when traditional linear damping is absent. Now the amplitude evolution
is given by

R =

R
_
0.32619

R cos
_
8R
9
_
0.00536

R cos
_
8R
3
_
0.24697
_
.
There are now innitely many limit cycles whenever is nonzero; for large R the rst order (delayed)
damping term overwhelms the half order derivative.
18 P. Wahi and A. Chatterjee
Figure 8. Phase plot of Equation (48) for = 0, = 1, = 1 and = 0.06.
A plot of

R versus R for two sets of and is shown in Figure 7. A phase plot of the original system
for = 0.06, = 0, = 1 and = 1 is given in Figure 8. It shows the rst few stable limit cycles.
The correspondence with Figure 7(a) is good.
4.5. A NONANALYTIC SYSTEM
The strong cubic nonlinearity considered above could be tackled exactly (without harmonic balance
approximations) using elliptic functions as well (e.g. 28). However, HBBA can handle a much broader
range of nonlinearities. To demonstrate, we now consider the nonanalytic system
x(t ) + x(t ) |x(t )| +
_
x(t ) + D
1/2
x(t ) + x(t 1)
_
= 0. (49)
Using a two term harmonic balance identical to Equation (40), we obtain
63
68

R and approximate
h as
x = R
_
sin( +)
53
2005
sin(3 +3)
_
.
Proceeding as before and using (2, 6) averaging, we obtain

R = R
3/4
_
0.29340 0.40050 R
1/4
0.39785 R
1/4
cos
_
63
68

R
_
0.00265 R
1/4
cos
_
189
68

R
__
and

=

R
1/4
_
0.29214 +0.39729 R
1/4
sin
_
63
68

R
_
+0.00065 R
1/4
sin
_
189
68

R
__
.
Averaging Fractional Derivatives and Delayed Terms 19
Figure 9. Numerical solution of Equation (49) along with the numerical solution of the corresponding averaged equations for
= 0.1, = 0.1, = 1, = 1. Initial conditions : x() = 0, [1, 0], x() = 0, [1, 0), x(0) = 15, R(0) = 6.6,
(0) = 0.
Numerical results, in Figure 9, are good. Thus, nonanalyticity in the unperturbed equation does not lead
to any trouble while using HBBA.
5. Conclusions
The method of averaging has been applied to study small perturbations to conservative oscillators which
may be strongly nonlinear. The perturbations studied include delayed terms and fractional derivative
terms. Averaging for the case of fractional derivatives corresponds to the general or aperiodic case and
involves an innite time average which, however, can be replaced using asymptotic arguments by a
simpler nite time average corresponding to a periodic case.
A slightly perturbed harmonic oscillator is studied rst to validate the averaging procedure in the
presence of delays andfractional derivative terms. For stronglynonlinear systems, the recentlydeveloped
approximate technique of HBBA [21] is used. Strongly nonlinear systems with cubic nonlinearity as
well as a nonanalytic nonlinearity are considered (the latter is not easily tractable using, say, elliptic
functions). The averaged dynamics closely match the full numerical solutions in all cases, verifying the
validity of the averaging procedure as well as the harmonic balance approximations therein. Moreover,
interesting dynamics is uncovered in the strongly nonlinear case with small delayed terms, where
arbitrarily many stable and unstable limit cycles can coexist, and innitely many simultaneous saddle-
node bifurcations can occur.
Appendix A: Initial Conditions in the Presence of Delayed Terms
In applying averaging to conservative oscillators perturbed by terms with delays, we note that they are
innite dimensional for any nonzero . The solutions have innitely many fast decaying components
which rapidly converge to a slow invariant manifold, as shown schematically in Figure 10. The slow
ow equations for amplitude and phase obtained in this paper capture only the dynamics on the slow
manifold. Hence, for purposes of numerical comparison with the full solutions, the choice of the
initial function for the DDEs requires attention.
20 P. Wahi and A. Chatterjee
Figure 10. Solution trajectories for a DDE. Double arrows represent the fast portion of the full-solution trajectory, which
eventually converges to a slow manifold where it coincides with a slow trajectory (single arrow).
The full-solution should not, as an extreme example, approach the slow manifold at the origin: in
such cases there will be no slow dynamics observed at all. Moreover, initial conditions for the averaged
equations should correspond to that trajectory on the slow manifold to which the full solution converges
(adjustments in initial conditions may be required to achieve this). The idea is shown schematically in
Figure 10.
For example, consider a simple second order linear DDE,
x(t ) + x(t ) + x(t 1) = 0. (50)
The corresponding averaged equations are

R = 0.27015 R ,

= 0.42074 .
The rst three characteristic roots of this equation for = 0.05 are 0.01351 i 1.02176, 4.56003
and 5.17593 i 7.21996. The slow manifold in this linear case is simply the subspace spanned by
e
(0.01351i 1.02176)t
. The behavior of any initial function which has no signicant components along
e
4.56003t
or e
(5.17593i 7.21996)t
(neglecting the innitely many roots which decay even faster) will be
captured well by the averaged equations (see Figure 11(a)). However, for the initial function e
4.56003t
,
the solution trajectory has no portion lying on the slowmanifold, has little in common with the averaged
solution (see Figure 11(b)).
Appendix B: Averaging for General q
We have so far used q = 1/2 in our fractional derivative terms, which is the value most encountered
in structural dynamics. However, the procedure is applicable for q = 1/2. A procedural detail is that
we will be required to evaluate
_
t
0
sin

q
d and
_
t
0
cos

q
d
Averaging Fractional Derivatives and Delayed Terms 21
Figure 11. Solution of Equation (50) and the corresponding averaged slow ow equations for = 0.05. Initial conditions:
(a) x() = 1, x() = 0, for [1, 0], R(0) = 1, (b) x() = e
4.56003
, x() = 4.56003 e
4.56003
, for [1, 0],
R(0) = 1.
or equivalently,
_
t
(1q)
0
sin s
1/(1q)
ds and
_
t
(1q)
0
cos s
1/(1q)
ds.
As before, only the behavior of the above as t is of interest to us. Using contour integration, it
can be shown that
_

0
sin s
1/(1q)
ds = (1 q) sin(q/2)
and
_

0
cos s
1/(1q)
ds = (1 q) cos(q/2).
The above integrals are also derivable from related results on page 110 of [22].
Acknowledgements
This work was partially supported by ISRO and DRDO through the Nonlinear Studies Group at IISc
and also by the Department of Science and Technology (DST) through a research grant.
22 P. Wahi and A. Chatterjee
References
1. St ep an, G., Delaydifferential equationmodels for machine tool chatter, inDynamics andChaos inManufacturingProcesses,
F. C. Moon (ed.), Wiley, New York, 1997, pp. 165191.
2. Olgac, N., Elmali, H., Hosek, M., and Renzulli, M., Active vibration control of distributed systems using delayed resonator
with acceleration feedback, ASME Journal of Dynamic Systems, Measurement, and Control 119, 1997, 380389.
3. Santos, O. and Mondi e, S., Control laws involving distributed time delays: Robustness of implementation, in Proceedings
of the American Control Conference, Chicago, Illinois, 2000, pp. 24792480.
4. Insperger, T. and St ep an, G., Remote control of periodic robot motion, in Proceedings of the Thirteenth Symposium on
Theory and Practice of Robots and Manipulators, Zakopane, Poland, 2000, pp. 197203.
5. Batzel, J. J. and Tran, H. T., Stability of the human respiratory control system I. Analysis of a two-dimensional delay
state-space model, Journal of Mathematical Biology 41, 2000, 4579.
6. Szydlowski, M. and Krawiec, A., The Kaldor-Kalecki model of business cycle as a two-dimensional dynamical system,
Journal of Nonlinear Mathematical Physics 8, 2001, 266271.
7. Kalm ar-Nagy, T., St ep an, G., and Moon, F. C., Subcritical Hopf bifurcation in the delay equation model for machine tool
vibrations, Nonlinear Dynamics 26, 2001, 121142.
8. Epstein, I. R., Delay effects and differential delay equations in chemical kinetics, International Reviews in Physical
Chemistry 11(1), 1992, 135160.
9. Roussel, M. R., Approximate state-space manifolds which attract solutions of systems of delay-differential equations,
Journal of Chemical Physics 109(19), 1998, 81548160.
10. Bagley, R. L. and Torvik, P. J., A theoretical basis for the application of fractional calculus to viscoelasticity, Journal of
Rheology 27(3), 1983, 201210.
11. Bagley, R. L. and Torvik, P. J., Fractional calculus a different approach to the analysis of viscoelastically damped structures,
AIAA Journal 21(5), 1983, 741748.
12. Bagley, R. L. and Torvik, P. J., Fractional calculus in the transient analysis of viscoelastically damped structures, AIAA
Journal 23(6), 1985, 918925.
13. Koeller, R. C., Application of fractional calculus to the theory of viscoelasticity, ASME Journal of Applied Mechanics 51,
1984, 299307.
14. Kevorkian, J. and Cole, J. D., Multiple Scales and Singular Perturbation Methods, Springer-Verlag, New York, 1996.
15. Plaut, R. H. and Hseih, J. C., Nonlinear structural vibrations involving a time delay in damping, Journal of Sound and
Vibration 117(3), 1987, 497510.
16. Hu, H., Dowell, E. H., and Virgin, L. N., Resonances of a harmonically forced dufng oscillator with time delay state
feedback, Nonlinear Dynamics 15, 1998, 311327.
17. Maccari, A., The response of a parametrically excited van der Pol oscillator to a time delay state feedback, Nonlinear
Dynamics 26, 2001, 105119.
18. Li, G., Zhu, Z., and Cheng, C., Dynamical stability of viscoelastic column with fractional derivative constitutive relation,
Applied Mathematics and Mechanics 22 (3), 2001, 294303.
19. Sanders, J. A. and Verhulst, F., Averaging Methods in Nonlinear Dynamical Systems, Springer-Verlag, New York, 1985.
20. Rossikhin, Y. A. and Shitikova, M. V., Application of fractional calculus for analysis of nonlinear damped vibrations of
suspension bridges, Journal of Engineering Mechanics 124 (9), 1998, 10291036.
21. Chatterjee, A., Harmonic balance based averaging: Approximate realizations of an asymptotic technique, Nonlinear
Dynamics 32, 2003, pp. 323-343.
22. Oldham, K. B. and Spanier, J., The Fractional Calculus Theory and Applications of Differentiation and Integration to
Arbitrary Order, Academic Press, New York, 1974.
23. Koh, C. G. and Kelly, J. M., Application of fractional derivatives to seismic analysis of base-isolated models, Earthquake
Engineering and Structural Dynamics 19, 1990, 229241.
24. Yuan, L. and Agrawal, O. P., A numerical scheme for dynamic systems containing fractional derivatives, in Proceedings of
1998 ASME Design Engineering Technical Conferences, Atlanta, Georgia, September 1316, 1998.
25. Suarez, L. E. and Shokooh, A., An eigenvector expansion method for the solution of motion containing fractional derivatives,
ASME Journal of Applied Mechanics 64(3), 1997, 629635.
26. Rand, R.H., Lecture Notes on Nonlinear Vibrations, version 36. Available online at http://tam.cornell.edu/randdocs/
nlvibe36.pdf, 2001.
27. Verhulst, F., Nonlinear Differential Equations and Dynamical Systems, Springer-Verlag, Berlin, 1990.
28. Coppola, V. T. and Rand, R. H., Averaging using elliptic functions : Approximation of limit cycles, Acta Mechanica 81,
1990, pp. 125142.