Asymmetric Mathieu Equations

Author(s): Amol Marathe and Anindya Chatterjee

Source: Proceedings: Mathematical, Physical and Engineering Sciences, Vol. 462, No. 2070
(Jun. 8, 2006), pp. 1643-1659
Published by: Royal Society
Stable URL: http://www.jstor.org/stable/20208965
Accessed: 10-08-2016 12:20 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].

Royal Society is collaborating with JSTOR to digitize, preserve and extend access to Proceedings:
Mathematical, Physical and Engineering Sciences

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 PROCEEDINGS
-OF- A Proc. R. Soc. A (2006) 462, 1643-1659
THE ROYAL i^\ doi:10.1098/rspa.2005.1632
SOCIETY Jl?? Published online 14 February 2006

Asymmetric Mathieu equations

By Amol Marathe* and Anindya Chatterjee
Mechanical Engineering, Indian Institute of Science, Bangalore 560012, India

An inverted pendulum with asymmetric elastic restraints (e.g. a one-sided spring), when
subjected to harmonic vertical base excitation, on linearizing trigonometric terms, is
governed by an asymmetric Mathieu equation. This system is parametrically forced and
strongly nonlinear (linearization for small motions is not possible). However, solutions
are scaleable: if x(t) is a solution, then so is ax(t) for any real a>0. We numerically study
the stability regions in the parameter plane of this system for a fixed degree of
asymmetry in the elastic restraints. A Lyapunov-like exponent is defined and
numerically evaluated to find these regions of stable and unstable behaviour. These
numerics indicate that there are infinitely many possibilities of instabilities in this system
that are missing in the usual or symmetric Mathieu equation. We find numerically that
there are periodic solutions at the boundaries of stable regions in the parameter plane,
analogous to the symmetric Mathieu equation. We compute and plot several of these
solution branches, which provide a relatively simpler means of computing the stability
transition curves of this system. We prove theoretically that such periodic solutions must
exist on all stability boundaries. Our theoretical results apply to the asymmetric Hill's
equation, of which the pendulum system is a special case. We demonstrate this with
numerical studies of a more general asymmetric Mathieu equation.
Keywords: asymmetric; Mathieu; Lyapunov exponent; stability; transition curves

1. Introduction

Consider an inverted pendulum with asymmetric elastic restraints in which the

pivot is given a vertical periodic oscillation, as shown schematically in figure 1.
Spring stiffnesses on the left and right are taken as <5(1 ? a) and <5(1 + a),
respectively, with 0< a< 1. The springs may be taken as straight, as shown,
with a long free length so that they remain effectively horizontal even when
their endpoints move up and down; they may alternatively, for simplicity, be
thought of as torsion springs at the pivot point. Gravity may be incorporated in
the spring stiffnesses and is not explicitly accounted for. There is no damping.
Let the angular displacement of the pendulum be 0(t). Taking the length of
the pendulum to be / and the origin of the coordinate system at some point on the
0 = 0 line, the coordinates of the centre of mass of the pendulum are

x = I sin d, y = u(t) + I cos 6,

* Author for correspondence ([email protected]).

Received 20 July 2005

Accepted 29 November 2005 1643 ? 2006 The Royal Society

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 1644 A. Marathe and A. Chatterjee

spring is unloaded t springis compressed

5(1-5)

ecos(?)t)

Figure 1. An asymmetrically supported inverted pendulum with base excitation.

where u(t) is the imposed displacement of the pivot. Differentiating,

x = I cos 66, y = ii ?I sin 66.
The system kinetic energy is

T = ?m(x2 + y2),
where m is the mass. We take the potential energy to be of the form (recall that
gravity is not explicitly included)
1
V=-O(l+Sgn(d)a)l?6?,
which is correct for torsion springs, and approximately correct for straight
springs with long free lengths as mentioned earlier.
The equation of motion can now be routinely obtained using Lagrange's method.
On linearizing trigonometric terms by substituting cos 6=1 and sin 0 = 0, and
letting 1= 1, m= 1, we obtain

6-?6 + o(l +sgn(6)?)6 = 0.

Assuming u=e cos(?), writing x in place of 6, and noting that
sgn(x)x = |x|,
we obtain
x + (? + e cos(t))x + ??\x\ = 0. (1.1)
We call equation (1.1) an asymmetric Mathieu equation; setting ?= 0 gives the
usual Mathieu equation (Stoker 1950; Winkler 1966; Rand 2003). We mention in
passing that the dynamics of a gear-pair system involving backlash can be

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1645

modelled by a similar equation (Theodossiades & Natsiavas 2000). In this paper,

we investigate some of the similarities and differences between the stability
diagrams (in the ?-e parameter plane) for the usual and asymmetric Mathieu
equations. Our theoretical treatment allows parametric forcing of the \x\ term as
well, which introduces two more free parameters as discussed later.
As is well known, the usual Mathieu equation has alternating stable and
unstable regions in the parameter plane. On the transition curves (boundaries of
the stability regions), there are periodic solutions with period either 2n or Arc (2:1
subharmonic motion) (Stoker 1950; Winkler 1966; Rand 2003). How are these
related to the stability regions, transition curves, and periodic solutions of the
asymmetric Mathieu equation?
Equation (1.1) is a homogeneous second-order differential equation with real
periodic coefficients. It is also essentially nonlinear: linearization near x=0 is not
possible. If Xi(t) and a^(i) are two solutions, then axi(t) + ?x2(t) is not a solution
in general. However, solutions are scaleable, i.e. if x(t) is a solution, then ax(t) is
also a solution for any real a>0. This scaleability will be important in the
subsequent analysis.

2. Unforced system

The unforced system (e = 0) is

x + ?x + ??\x\ = 0. (2.1)
The asymmetric potential energy of this unforced system and the correspo
asymmetric spring force are shown in figure 2, left and right, respectively
All solutions of the unforced system are periodic, and have the same pe
Each such periodic solution consists of two half-sinusoids, one for x>0 and
for x<0, with different amplitudes and time durations. The time period T
corresponding fundamental frequency ?) are
r^ 7T 7T 27T . .
T= / =+ /_= and co =?. (2.2)

3. A Lyapunov-like exponent for scaleable systems

Since the fully nonlinear forced system seems analytically intractab

to numerics. A few numerically obtained solutions for equation (1.1)
figure 3. It is seen that solutions grow without bound (unstab
parameter values, but remain bounded for others. The unstable solu
exponentially, and we begin by characterizing their exponential grow
A Lyapunov-like exponent is defined as
1 N
a(?,6)=^im-^ln(||Xr||), (3.1)
where the state vector X={x, x} ; ||-|| denotes the Euclidean
subscript k in X2n denotes the time in number of forcing per
superscript, either 0 or 27T, denotes the start or end of a forc
respectively.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 1646 A. Marathe and A. Chatterjee

(a)\

A 8

V(x)[

-10 -8 -6 -4 -2

Figure 2. 0 = 0.5, a =0.7. (a) Potential energy. (6) Restoring spring force.

The above Lyapunov-like exponent is calculated as follows. We numer

integrate equation (1.1) over an interval of 2tt with random initial condition
that satisfy ||Xo|| = 1. At the end of this integration, we have the stat
Initial conditions for the next cycle are then obtained by rescaling to unit n
Y"2tt
x,u I Y"27T|

Integration over another period of forcing then gives X2n. The above steps are
repeated N times, for some large N. We discard the first several states in the
calculation to get rid of transients, and take the first retained state as X??. N is
taken large enough to obtain convergence, which in this case is relatively rapid
presumably because solutions are not chaotic.
An exponent a = 0 implies the corresponding ?-e point is stable, while a>0
implies it is unstable. In numerics, cr = 0 is not exactly observed, and a non-zero
numerical tolerance or threshold is set by the analyst.

4. Numerical stability diagram

We first present the results of a numerical stability analysis of the asymmetric

Mathieu equation on the ?-e parameter plane for the arbitrarily chosen value of
a= 0.7 (we will show some more results for a= 0.8 later in this same section). In
these calculations, numerical integration of ODEs was done using Matlab's
built-in ODE solver 'ODE45' with 'event detection' so as to ensure that every
zero-crossing of x coincides with a mesh point (or sampling instant) in the time
discretized solution.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1647

xlO8
(a) 6 (b) 5r

(e) 5000

-5000

-lOOOOl
50 100 150
t

Figure 3. Time histories of solutions for e = 0.2 fixed and ? values as indicated. Cases (a, c, e) are
unstable and (?, d, /) are stable.

Figure 4 shows the Lyapunov-like exponent on a 500X500 grid covering

<5e[ ? 0.1,1.3] and ee[0,l], for 7V=600. The calculation took several days,
distributed among a few PCs. Figure 5 shows a closer and refined view of the
boxed, bottom left, region of figure 4. Now ?e [0,0.45] and eG [0,0.4], the grid is
450X400, and iV=2400.
The figures show several instability regions. Limitations in numerics and
graphics limit the number of visible regions; however, it seems that there are
infinitely many such regions, successively narrower and more weakly unstable,
nested between a few large and strongly unstable regions. These large and
strongly unstable regions have a one-to-one correspondence with regions of
instability for the usual Mathieu equation, as may be seen from the following.
Narrower and finer resonance regions are associated with smaller numerical
values of Lyapunov-like exponents: decreasing N or increasing the numerical
value of the cutoff tolerance in the stability diagram will prevent us from
detecting these regions.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
1648 A. Marathe and A. Chatterjee

Figure 4. Stability diagram for a =0.7, N=600. Grey, stable; black, unstable.

0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45

S

Figure 5. Zoomed portion of figure 4: ?=0.7, N=2400. Grey, stable; black, unstable.

For the usual or symmetric Mathieu equation (a=0), there are regions
instability emanating from the ?-axis at <5sym = 0, 0.25, 1, . These correspon
simple rational relationships between the unforced time period, 2n/^/?sym,
the forcing period, 2tt.
For the asymmetric Mathieu equation with a=0.7, the time period of t
unforced system is, from equation (2.2),
2tt
1.2964
V^ asym

For the same rational relationships to hold so that the same resonances may
occur, we require this time period to bear the same ratio to 27T as in the
symmetric case.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1649

This gives
2tt 2tt
1.2964
y ^asym Y ^s;^sym
or

<5asym = l-29642<5sym = 1.6805<5sym.

Thus, we expect corresponding instability regions for the a= 0.7 case to emanate
from the ?-axis at <5 = 0, 0.4201, 1.6805, . This is supported by the numerical
results.
The asymmetric and symmetric Mathieu equations are identical on the e-axis
(with ? = 0),

x + cos(t)x = 0.
Therefore, stability intervals on the e-axis for both equations (for any a) are
the same. So, all the new instability regions of the asymmetric equation must
have zero widths on the e-axis. This is observed in the numerics. Similarly, the
widths of the corresponding strong instability regions that occur for both the
asymmetric as well as symmetric equations coincide on the e-axis; this is outside
the region covered by the numerical results, and cannot be seen in the figure.
We now consider other possible resonances. Let r be a simple rational number,
such as (say) 2.5, 3 or 4. Consider a situation where the ratio between the
unforced time period and the forcing period is r, i.e.
27T
1.2964 X -= = 2irr (4.1)
v<5
or
t 1.29642 1.6805
0 =-2- =-2~~*
The ? values corresponding to r=2.5, 3 and 4 are then 0.2689, 0.1867 and 0.1050,
respectively. There are, in fact, instability regions corresponding to these values
on the ?-axis, as may be seen with a little faith from figure 5, or more
convincingly from figure 7 in ?5.
The qualitative results obtained above are not special for a=0.7. Figure 6
shows the Lyapunov-like exponent for ? = 0.8 on a 500X500 grid covering
<5e[ ? 0.1,1.5] and ee[0,l], for 7V=600. There is agreement with the results for
5=0.7.
From the numerical results of this section, it seems likely that all the narrow
instability regions seen in figure 5 do in fact continue all the way to the ?-axis.
This is verified numerically in ?5, and supported further with theory later in
the paper.

5. Periodic solutions on stability boundaries

As is well known, for the usual or symmetric Mathieu equation, there are periodic
solutions of period either 2it or 47T on each stability boundary on the parameter
plane. Are there periodic solutions (of possibly other periods) for each stability

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
1650 A. Marathe and A. Chatterjee

0.06 0.22 0.38 0.54 0.70 0.86 1.02 1.18 1.34 1.50
?
Figure 6. Stability diagram for a=0.8, iV=600. Grey, stable; black, unstable.

boundary for the asymmetric Mathieu equation? In this section, we in

this question numerically. In particular, we seek curves on the ?-e
plane where periodic solutions exist; and anticipate that these curves w
from the ?-axis at the points from which instability regions emanate (a
earlier).
We present here the results obtained for 5 = 0.4201, 0.2689, 0.1867 and 0.1050.
The corresponding time periods expected, and obtained, are Att, 1?7T, 67T and 87T,
respectively (note: 107T corresponds to r=2.5 and not 5 in equation (4.1)). From
each of these points on the ?-axis, two curves are found to emanate. These curves
were obtained using a numerical arc-length based continuation method which is
described in appendix A. Results, shown using white lines on the stability
diagram, are given in figure 7. The picture, presented here for e>0, is symmetric
about the ?-axis.
It is clear that, at least for the instability regions considered, stability
boundaries correspond to the existence of periodic solutions, even for the
unstable regions that are absent for the symmetric Mathieu equation. It seems
likely that there are also periodic solutions at all other stability boundaries. In
?7, we show theoretically that this is in fact the case.
A few points regarding the numerical search for periodic solutions are now
presented.
Since solutions are scaleable, we may assume that the initial conditions at t=0
satisfy either Xq= +1, or i0 = ?1, or even 0$ + ijj = 1. It turns out that for the
periodic solutions plotted in figure 7, the solution branch to the left of each
instability region has x0 = 0 and Xq=1; while for the solution branch on the right
side of each region, Xq = 0 and x0 = 1. The numerical strategy, however, assumes
one of these (xq or x0) to be non-zero (equal to +1), and lets the numerical
routine discover that the other is zero (if in fact it is).
Numerically, we proceed as follows. Starting at t=0 with some values of ? and
e (along with, say, ?= 0.7 fixed and Xq = 1), we numerically integrate forward in
time to time T (for suitable T, which is an even multiple of n that we know in

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1651

O 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45

?
Figure 7. 47T, 67r, Stt and 1?7T period solutions on the stability boundaries in the ?-e
plane for the asymmetric Mathieu equation with ? =0.7.

advance as indicated earlier), and check that the final state is ident
initial state. We then iteratively look for a nearby point on the <5-e pl
these conditions are satisfied again. This well-known procedure, called a
based continuation, is described for completeness in appendix A.
Note that these periodic solution branches can (at least poten
computed in a small fraction of the time needed to generate, say, figure
as for the symmetric Mathieu equation, finding periodic solutions is an
way of computing the stability transition curves for the asymmetric
equation.

6. Theoretical considerations

The asymmetric Mathieu equation (1.1) can be rewritten as

x=f(x,t), (6.1)
with

and

[ ? (? + e cos(t))x ? ?a\x\ J
Formally, we observe that equation (6.1) is divergence free, i.e. V/=0. This
means areas are preserved in the (x, x) plane.
Let x=Rcos((/)) and y=R sin(0). Consider a Poincar? map that takes x from
the start of a forcing period to the start of the next (i.e. through t=2ir). Let

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 1652 A. Marathe and A. Chatterjee

</?o,</>o+A0o)[ (*o/?b)>s?b>)
(?o/(0o+A0o), g(</>o+A0o))

Figure 8. Area preservation of the flow.

(R0, 0o) be an initial point. The Poincar? map sends (RQ,(/)0)^ (^i?0i) with
?l=/(?b,0o),
0i = g(Ro,<l>o),
for some as yet unknown continuous functions / and g. Since solutions to
equation (1.1) are scaleable, we can write

7(?o,0o)=?o7(Mo)=-Ro/(0o)
Similarly, scaleability requires

g(Ro^o) = g(4>o)
Thus, the point (R0, 0O) gets mapped to (Aq/(0o)j s(0o))
The system is reversible in time. Zero initial conditions lead to zero solutions
for all time. These two together imply that non-zero solutions do not become zero
within finite time. In turn, this means

/(?o)*0.
We may, without loss of generality, assume />0.
Consider two nearby initial conditions (Rq, </>0) and (R0, (/>0 +A0O), as sketched
in figure 8. The Poincar? map sends these initial conditions to (?2q/(0o)> ?(0o))
and (R0f'(</>o + A0O), #(</>o + A0O)) as sketched in figure 8.
If we think of the entire triangle as composed of initial conditions, then the
triangle gets mapped to a triangle because solutions are scaleable (the edges
remain straight lines).
Since / and g are continuous, we write
/(</>o+A4>o)=/(0o)+A/, and g(<j>0 + A0O) #(0o) + A<7,
where the A symbol is now taken to denote 'small'.
Area preservation now gives
|A0(0o)l _ 1
Wo f(4>o. ,2 '
where we have ignored small quantities of second-order.
We now come to an important point. No matter what finite values we assign to
the parameters ? and e, the functions / and g depend continuously on them as

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1653

well as on 0O. So, if we now change any combination of ?, e and/or 0O in any way
that we like, /(0o) always remains finite and non-zero. This means Ag(0o) is
always non-zero as well, because A0O is non-zero and positive by choice; it never
changes sign. It is possible to conclude from a consideration of the case <5 = e = 0
and 0o = 7r/2, i.e.
x =0,
that Ag>0. From here, by continuous changes in parameters and initial
conditions, we can arrive at the point of interest to conclude that the absolute
value sign may be removed, and so
Ag(4>o) _ 1
Wo /(</>0)2 '
which in the limit shows that g is differentiable and satisfies

It follows that </(0o)>O for all 0O.

^)=74)' (6-2)
We observe from equation (6.2) that if any solution settles down to some
stable point 0*, then #(0*) = 0* and </(0*)<l (the condition for stability of fixed
points of iterated scalar maps). This in turn implies that /(0*)>1. Thus, any
solution that settles exponentially to some 0* must grow exponentially in
magnitude.
What happens if, instead of a fixed point, g has a fc-cycle, i.e. the kth iterate of
some 0* equals itself, or <7A;(0*) = 0*? Let 5(0*) = 0i, ?,(0i) = 02> and so on. Then,
0&=0* for a fc-cycle.
If, in addition, the &-cycle is exponentially stable, i.e.

g'(<?>*)-g'((t>i)-9'(4>2)-9'(^i)<?, (6.3)
then it follows that
/(0*)-/(^i)-/(02)-/(^*-i)> i,
and the solution grows exponentially in magnitude (the Lyapunov-like exponent
is positive).
Consider, now, a gradual change in parameters that causes this unstable point
in parameter space to approach a stability boundary. The fc-cycle (corresponding
to 0*) depends on parameters, and changes gradually as well; it is structurally
stable as long as inequality (6.3) holds. Thus, loss of stability can only occur
when
9 (</>*) g'((t>i)'9(^2)'"9(^k-i) = 1,
at which point we also have
/(0*)-/(^i)7(^)-/(^i) = i.
In other words, an unstable solution (i.e. a growing solution with a positive value
of the Lyapunov-like exponent used in this paper; but also a solution where 0*
corresponds to a stable fc-cycle of the iterated function g) can only lose instability
(or gain stability; or reach a stability margin) by deforming continuously into
a periodic solution as parameters are slowly changed so as to reach a point on a
stability boundary.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 1654 A. Marathe and A. Chatterjee

Conclusion 6.1. A stable &-cycle in the iterated function g implies instability in

the system solution. From the corresponding (unstable) point on the parameter
plane, moving towards a stable point requires appearance of a periodic solution
on the stability boundary.
What we wish to prove, however, is more general. We wish to prove that every
stability boundary corresponds to the existence of a periodic solution. We will do
this by showing that every unstable point in the parameter plane corresponds to
the existence of a stable fc-cycle in the iterated function g. Conclusion 6.1 will
then be applicable, and the desired result will be established.
Accordingly, we now assume that the solution is unstable, i.e. for some
parameter values ? and e, and referring to equation (3.1),

ff=iim^Eln(n^2,rii)>0- (6-4)
This is equivalent to
?(ln(/(/(0o))))>O, (6.5)
where E represents expected value, and n is sufficiently large that initial
transients are not important and the final steady-state behaviour of the iterates
of g, whether periodic or not, is obtained. From equation (6.2), we have

?(ln(i/))=-2?(ln(/)),

E(ln(9>(gn(<l>)))) = -2E(\n(f(gn (<!>))))

Therefore

?(ln(5V(?))))<0.
Let
E(\n(gl(gn(<l>))))=-a, (6-6)
for some strictly positive number a. Moreover, since we assume the system has
reached steady-state behaviour, we can also define the variance of In g' as

var(ln(ffV(*)))) = b\ (6.7)
for some b>0.
Let (?)p=gp(^Q) for p> 1. Since 0 is an angle, we can look at its values modulo 27T.
The dynamics of the system generates a sequence 0O, </>i, 02? 03? * ' ' 5 which is now
bounded (because we look at the values modulo 2?r). Every bounded sequence has a
convergent subsequence (Rudin 1978). Let the subsequence be 0Ci, 0^, 0C3, ,
where c?> Cj if i>j. Let the subsequence converge to 0*. Then, for any given e>0,
there exists a finite M such that for all n> M, <f>Cn ? 0* < e. We choose two points
from the subsequence, not necessarily consecutive, say 0Cn , 0C with n^ > n? > M.
Now consider the function

h((j)) =gc^~c^(<p)-(?). (6.8)

Then, applying the chain rule of differentiation,

h'{4>) - gXg^-^mg'ig^-'m </(<?(</>))</(<*>)-l. (6.9)

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1655

Considering /?'(0m) for some ra>Min the subsequence Ci, c2, , we have

^(0m) = 9 (0cn2-S+m-l)^(0cn2-S+m-2)'-*5/(0m+l)5/(0m) "1

Considering the logarithm of the first-term on the right-hand side, we have
(calling it, say, Z)

Z= JZt=0 W(0ro+i)
For c^ sufficiently larger than cni (and we are free to choose it so), the central
limit theorem applies; in particular, the expected value of Zis ? (c^ ? cnJa (see
equation (6.6)), and its standard deviation is b^/c^ ? cni (see equation (6.7)),
which is much smaller. Thus, Z is strictly negative, and can be as large as we
wish to make it: its exponential is a positive number which can be as small as we
like, independent of e. It follows that

ti((t>m) = 5/(0Cn2_s+m_i)^(0Cn2_s+m_2)---^(0m+l)5f/(0m)-l,
lies between ?1 and 0. In particular, it can be bounded away from 0 by a non
zero amount, such as 1/2, independent of e.
Now, consider equation (6.8)

\K4>m)\ = |0m+S-cni -0m| < 2e.

Thus, /i(0m) is small; and h'((/)m) is non-zero. By the implicit function theorem, h
has a zero close to 0m, say at 0. This zero corresponds to a stable fc-cycle of the
iterated function g.
Conclusion 6.2. An unstable solution of the system implies the existence of a
stable fc-cycle of the iterated function g.

By Conclusion 6.1, every point on a stability boundary in the parameter plane

(or parameter space, if we introduce more parameters) corresponds to the existence
of a periodic solution. However, unlike the usual or linear Mathieu (or Hill)
equation, the period need not be solely 27r or 47r, but could be a higher multiple of 27T.

7. General asymmetric Mathieu equations

More generally, our foregoing results apply to

x + (? + e0 cos(i))x + ?(? + ex cos(t + \p))\x\ = 0, (7.1)
where ex and \p are additional free parameters (compare with equation (1.1)).
The theoretical results presented above actually hold for the general
asymmetric Hill's equation
x + Pl(t)x + p2(t)\x\ =0, (7.2)
with pi(i+ T)=px(t) and p2(t+ T)=p2(t)Vi and for some T>0 (Tcan be taken
as 27T upon scaling time suitably). We are not presently aware of actual physical
systems governed by such equations, except for the restricted case we began this
paper with.
A few numerically generated stability diagrams for equation (7.1) are now
presented. We restrict ourselves to a= 0.7 as before along with ^ = 7r/4.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
1656 A. Marathe and A. Chatterjee

tt

?,= 0 plane
jL

,= 0.2
plane

VT

6q= 0 plane ~Pr

Figure 9. Parameter planes in the (?, e(), \) space where numerical stability results are presented.

*i 0.5

Figure 10. Stability diagram for equation (7.1) with e0 = 0, ^ = 7r/4, a = 0.7 and A/=600. Grey,
stable; black, unstable; white, periodic solution.

A sketch of a subset of the full parameter space is shown in figure 9. Figure 10

shows the stability diagram in the parameter plane 0 = 0, and figure 11 shows the
stability diagrams on portions of the two orthogonal planes i = 0 and 60 = 0,
respectively.
Figure 12 shows the stability diagram on ei = 0.2 plane and figure 13 shows the
stability diagrams on planes 6x = 0.2 and e0 = 0, respectively. Numerical

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1657

Figure 11. Stability diagram for equation (7.1) with ^ = 7r/4, a= 0.7 and 7V=600, on two different
parameter planes. Figure 10 is included but now is horizontal.

Figure 12. Stability diagram for equation (7.1) with e^O.2, \?/ = tt/A, a=0.7 and A/=600. Grey,
stable; black, unstable; white, periodic solution.

Figure 13. Stability diagram on the planes ! = 0.2 and e() = 0; again, a =0.7 and \p = ir/A and
AT=600.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 1658 A. Marathe and A. Chatterjee

resolution limits the degree of detail that can be trusted in the figures. With more
computation time, any of these figures could be regenerated with higher precision
and resolution. However, our key point in producing these figures is to emphasize
that the white lines, representing numerically obtained periodic solutions, were
in fact computed accurate to nine decimal places. Moreover, we did find periodic
solutions on every stability boundary that we examined, verifying the theoretical
results obtained above.

8. Conclusions and further work

We have numerically and theoretically studied the asymmetric Mathieu equations,

which are strongly nonlinear but conservative, and have scaleable solutions (if x(t)
is a solution, then so is ax(t) with a> 0). We have found that there are infinitely
many more instabilities for this system than for the usual Mathieu equation. There
are periodic solutions on every stability boundary in the parameter space. The
periods of these solutions are not confined to either 27T or 47r; higher multiples of 27T
occur. Our theoretical results are also applicable to asymmetric Hill's equations.
Several questions seem interesting and relevant which we have been unable to
answer here. For equation (1.1), are there in fact infinitely many instability regions
of strictly non-zero width emanating from any finite interval on the ?-axis? How do
the widths of these regions, say for small e and for an m:n resonance, depend on m, n
and e? Given an arbitrary point (?, e) with e>0, does every open set in the
parameter plane that contains this point also contain an unstable point? Given
some small damping, how much of which instability regions will survive? We hope
that future work may shed some light on these issues.

Appendix A. An arc-length based continuation method

We numerically seek periodic solutions to equation (1.1). Consider a solution

starting at t=0. If Xq and i0 are the initial conditions and we numerically
integrate equation (1.1) to t= Tfor suitable T, then x(T) and x(T) should satisfy
x(T)-xo=0, (Al)
x(T)-x0=0. (A 2)
The routine continuation method described below was also used in
Nandakumar & Chatterjee (2005). Assume that we are on a branch where we
can take Xq = 1 (recall that solutions are scaleable; the corresponding procedure
for Xq= ? 1 or i:0 =+1 will be obvious below). We define

y=i e \. (A3)
UJ
Assume that for some known y, i.e. for some known choice of <5, e as well as x0, we have
a periodic solution of period T. (Such a y can be found by an initial numerical search
that is simpler than the continuation method and is not described here. Note, in
particular, that resonance conditions provide the ? values corresponding to e = 0,

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms
 Asymmetrie Mathieu equations 1659

so only ?g need to be found for these points. ) Let us refer to this value of y by the name
2/0id. The continuation method is based on defining a function g(ynew) as follows.
Given any vector ynew, we have ?new, enew and i0,new Inserting <5new and enew
into equation (1.1), and using initial conditions Xq = 1 and ?o = ^o,new5 we
numerically integrate forward to time Tto obtain x(T) and x(T). We also choose
a small positive number s, which represents arc-length. Then, we define
( <T)-1 \
#(2/new)V =H?/old
| i(T)-X0,new
??/newll ? ? J
> (A4)
We now numerically evaluate the function g as accurately as we like; and use any
standard iterative technique (like Newton-Raphson) to find a value ynew that
satisfies p(2/new) = 0- Finally, we redefine j/old to be the freshly obtained ynew, and
repeat the above procedure. This completes the arc-length based continuation
method.
In numerical implementation, in the Newton-Raphson iterative stage, it is
helpful to have a good initial guess. Using the two previously obtained y vectors,
say y i and y2, to extrapolate linearly to
2/guess = 2y2-2/i,

gives a good guess for small s in most cases.

References
Nandakumar, K. & Chatterjee, A. 2005 Resonance, parameter estimation, and modal interactions
in a strongly nonlinear benchtop oscillator. Nonlin. Dyn. 40, 149-167. (doi: 10.1007/sl 1071-005
4228-3)
Rand, R. H. 2003 Lecture notes on nonlinear vibrations?version 45. Available at http://www.
tam.cornell.edu/randdocs/.
Rudin, W. 1976 Principles of mathematical analysis. Singapore: McGraw-Hill.
Stoker, J. J. 1950 Nonlinear vibrations in mechanical and electrical systems. New York: Wiley.
Theodossiades, S. & Natsiavas, S. 2000 Nonlinear dynamics of gear pair systems with periodic
stiffness and backlash. J. Sound Vib. 229, 287-310. (doi: 10.1006/jsvi. 1999.2490)
Winkler, S. 8? Magnus, W. 1966 HilVs equation. New York: Dover.

Proc. R. Soc. A (2006)

This content downloaded from 202.78.175.199 on Wed, 10 Aug 2016 12:20:27 UTC
All use subject to http://about.jstor.org/terms