Nonlinear Dynamical Systems,

Their Stability, and Chaos

Amol Marathe

Lecture notes from the FLOW-NORDITA Summer School

on Advanced Instability Methods for Complex Flows,
Stockholm, Sweden, 2013

Birla Institute of Technology and Science,

Pilani 333031, India
e-mail: [email protected]

This introduction to nonlinear systems is written for students of fluid mechanics, so connections are made throughout the text to familiar fluid flow systems. The aim is to present how
nonlinear systems are qualitatively different from linear and to outline some simple procedures by which an understanding of nonlinear systems may be attempted. Considerable
attention is paid to linear systems in the vicinity of fixed points, and it is discussed why this
is relevant for nonlinear systems. A detailed explanation of chaos is not given, but a flavor
of chaotic systems is presented. The focus is on physical understanding and not on mathematical rigor. [DOI: 10.1115/1.4026864]

TIFR Centre for Interdisciplinary Sciences,

Tata Institute of Fundamental Research,
Narsingi, Hyderabad 500075, India
e-mail: [email protected]

Introduction

What distinguishes a nonlinear system from a linear one? For a

stability theorist, a gut reaction on being confronted with a nonlinear system is to first linearize it about some equilibrium point if
he/she can find one, and then study the linear system in detail [1].
This is an excellent approach for many purposes. When a system
goes from laminar to turbulent or from periodic to chaotic as a
control parameter is increased, the first step in this process is often
linear. However, to understand the entire process, we need to
understand how nonlinearities change the answers. Most systems
are nonlinear, and even if a given system is not chaotic, nonlinearity can, and often will, play a big role in the dynamics. We begin
by discussing how nonlinear systems are completely and qualitatively different from linear systems. First, nonlinear systems do
not need to obey the same rules about existence and uniqueness of
solutions that linear systems do. Let us discuss a counter intuitive
example, in a boundary value problem. The FalknerSkan
equation,
f 000 ff 00 b1  f 02 0

(1)

describing the nondimensional streamfunction f g in the incompressible boundary layer over a solid wedge placed at angle bp to
a flow, is a third-order differential equation in a nondimensional
normal distance g from the solid wall. There are three boundary
conditions, f 0 f 0 0 0, and f 0 1 1. In a linear thirdorder system with three boundary conditions, we would get either
a unique solution or no solution. The FalknerSkan equation,
however, displays at least two solutions for each b < 0; an example is shown in Fig. 1. Such multiple solutions are not too unusual
in nonlinear boundary value problems such as this one. In contrast, dynamical systems are most often posed as initial value
problems. In an initial value problem, we have only now way to
march forward and will obtain a unique solution. However, the
fact that multiple solutions exist in boundary value problems is
relevant to us. To understand which of these solutions is manifested in a real flow, we must linearize the NavierStokes equations about each of these solutions and solve the resulting
dynamical system. Usually separated flows (where the flow next
to the wall is in the direction opposite to the average flow) such as
the profile shown by the red line (dashed) are very unstable

Manuscript received July 8, 2013; final manuscript received December 6, 2013;

published online March 24, 2014. Assoc. Editor: Ardeshir Hanifi.

Applied Mechanics Reviews

Rama Govindarajan

compared to attached flows such as that shown by the black line

(continuous), so the latter will be manifested.
Second, nonlinear systems can display finite-time singularities,
which change the qualitative behavior of the solutions. The worst (or
best) thing about nonlinear systems is that the solutions do not addup. The dependent variable in a linear partial differential equation
can be Fourier transformed in one of the variables, and you can then
solve for each Fourier mode in isolation. The terms in a nonlinear
equation can be Fourier transformed, but every mode will depend on
every other. This makes nonlinear systems harder to solve but also
far richer than linear systems. For example, a cascade or inverse cascade of turbulent kinetic energy, creating small or large scales, is
only possible due to different modes exchanging energy.
The most interesting difference between linear and nonlinear
systems is that the latter can display chaotic dynamics. We shall
distinguish in the section on chaotic dynamics the fundamental
difference between a chaotic system that is deterministic and simply a noisy system.
We shall discuss several ways to treat a nonlinear system. It is
instructive to first look for equilibrium points, also known as fixed
points, or steady solutions. These are solutions of the system that
do not change with time.

2 Fixed Points and the Behavior of the System

in Their Vicinity
We have already mentioned linearizing the system about a fixed
point and studying the linear stability of the fixed point solution.
A first-order nonlinear system may be written in the form
x_ f x

(2)

where x is an n-component vector, and the overdot indicates a

time derivative. It is sufficient to study a vector first-order differential equation of this kind since, without loss of generality, an
nth order differential equation can be turned into n first-order
equations and written in vector form, as above. A dynamical system is usually presented in the form of an initial value problem,
with the initial conditions
x0 x

(3)

Second, a system such as that given by Eq. (2), where time t does
not appear explicitly, is called an autonomous system. An
n  1th order nonautonomous system, where time appears

C 2014 by ASME
Copyright V

MARCH 2014, Vol. 66 / 024802-1

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

Fig. 1 A sample solution of the velocity profiles in a boundary

layer, as described by the FalknerSkan equation. Both solutions satisfy all three boundary conditions. Here b 5 20:1. The
solid line corresponds to f 00 0 5 0:1644, whereas the dashed
line corresponds to a separated velocity profile, with
f 00 0 5 20:0545.

Fig. 2 Standard fixed points and their phase portraits, along

with their canonical linear equations. Made using1. The tiny
arrows indicate the direction in which the solutions move as
time progresses.

explicitly, can be turned into an nth order autonomous system by

adding time as another component of x; xn t, giving x_n 1.
The solution of an n-dimensional system of the form of Eq. (2)
may be plotted in n-dimensional space known as phase space. We
remark here that trajectories cannot cross in phase space. In other
words, a given equation can have more than one solution as we
have seen, but a given point in phase space can belong to at most
one solution. Each point in phase space can be thought of as an
initial value for the rest of time, so there in a unique way forward,
as we saw earlier. Solutions of a two-dimensional system may be
plotted in a phase-plane. Representative trajectories beginning
from different initial conditions, plotted in phase space, make up a
phase portrait.
Let us suppose that x0 is a fixed point of the above system, i.e.,
f x0 0

(4)

so the system, if initially placed at the fixed point, stays there for
all time. In its immediate neighborhood, we may write
x x0 dxt, where the magnitude of the perturbation is small,
neglect terms containing powers of dx greater than 1 and obtain a
linear system in dx. We drop the d for convenience in future discussions of linear systems. From the eigenvalues of the linear system, the fixed point may be classified, for example, as a stable or
unstable spiral point, a stable or unstable node, a saddle point or a
center. There are also some other unusual fixed points that we
shall not discuss here. Simple equations that support fixed points
of each type, and phase portraits in the vicinity of the fixed-points,
are shown in Fig. 2. Such fixed points may be obtained for simple
systems on a plane, for example, by the MATLAB software available
on the Rice University website,1 as done here. Arrows may be
drawn on the trajectories shown in the phase portraits to show the
direction of time. Trajectories approach a stable (attracting) fixed
point, while they emerge out of an unstable or repelling point. The
reader may wish to attempt the following exercise.
Exercise (i) Work out the connection between the eigenvalues
of a system and the nature of the fixed point.
A simple example of a linear system is that of Fig. 3. The density q of the fluid shown is stratified so as to be heavier at the bottom. A small blob of fluid of density q0 is displaced along the
vertical direction z by a small distance dz to a location where the
density of the blob is different from that of the surrounding fluid
by dq dq=dzdz. Putting dz x1 , the dynamics may be given
by
1

http://math.rice.edu/dfield/:dfield8.m and pplane8.m.

024802-2 / Vol. 66, MARCH 2014

Fig. 3 A cylindrical blob of fluid displaced slightly from its

original location in a stratified fluid will display simple harmonic motion in the absence of diffusion. The z-axis is upwards
and density increases as z decreases.

x_1 x2
x_2 N 2 x1 ;

s
g dq
where N 
q0 dz

(5)
(6)

is the BruntVaisala frequency, g the magnitude of acceleration

due to gravity and the z-coordinate increases upwards. If the fluid
is bottom heavy, the density stratification is negative, and we have
a real N. This will result in oscillatory motion. On the other hand,
a top-heavy system, with dq=dz > 0, will result in imaginary values of N, one positive and one negative. It is easy to see that in
this case, the perturbation will grow unboundedly.
Exercise (ii) It is left to the reader to solve this simple system,
and to then include a viscous damping term proportional to x_1 in
Eq. (6), to get a center and a stable spiral, respectively. The direction of density stratification can then be reversed to get an unstable system.
Fixed points whose associated linear systems display eigenvalues with nonzero real part are known as hyperbolic and those with
Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

a zero real part of the eigenvalues are nonhyperbolic. Nonhyperbolic fixed points are sometimes referred to as elliptic. The
HartmanGrobman theorem assures us that the behavior of a nonlinear system close to a hyperbolic fixed point is very similar to
that of the linear system that approximates it. A saddle-point, for
example, will remain a saddle in the nonlinear system as well, and
moreover, the invariant manifolds emanating from it will look
conformally the same as their linear approximates. Thus, a hyperbolic fixed point is robust when faced with the addition of nonlinear terms. However, an elliptic point, or center, of a linear system
is a fragile fixed point. The nonlinear system that it approximates
could display a stable or unstable spiral instead or change the local
flow in other ways. Once we know all the fixed points in a given
domain, if none of them is a center, we may then obtain the qualitative features of the nonlinear phase space by knowing the linear
behavior in each neighborhood.
Figure 4 is an example of a nonlinear system with several fixed
points in the domain shown. The equations solved are
x_ y  y3  0:25x x3 y
y_ x  y  xy xy

(7)
(8)

An examination of this figure reveals that the stable node has a

basin of attraction, i.e., a region in phase space inside which
every trajectory leads to that fixed point. This is characteristic of
every attracting fixed point. There are basin boundaries separating
these basins of attraction. Repelling fixed points sit on the basin
boundaries. A saddle point on the basin boundary contains the
boundary itself as its attracting direction and has repelling directions
within the basins of attraction of (typically) two different attracting
fixed points. A trajectory beginning exactly on the basin boundary
will travel towards the saddle point, and reach it in infinite time. In
shear flows we often have a situation when the laminar state is linearly stable but turbulent states exist as well. Here the laminar and turbulent states may each be thought of as attracting fixed points, with
their own basin boundaries. Recent studies, e.g., Refs. [24] find that
the basin boundary between the laminar and turbulent states usually
contains one such saddle, which is termed the edge state.
Exercise (iii) Design a pair of simultaneous nonlinear first-order
differential equations that contain several fixed points. Plot the
phase space and mark the basin boundaries. What happens if one
of the fixed points is a center in the linear limit and not in the nonlinear limit?

Bifurcations

Fig. 5 The flow past a cylinder of square cross section. At a

Reynolds number Re 5 30, a steady bubble is formed behind
the cylinder. A Hopf bifurcation occurs at a Reynolds number of
about 46, and oscillatory states with Karman vortex streets are
evident at Re 5 50 and 100. Note that the length scale in different at Re 5 30. Figure courtesy: Srikanth Toppaladoddi.

Bifurcations of various types frequently occur in fluid mechanical

situations. The transition from a laminar (ordered) state to a turbulent (chaotic) state is an extremely complicated one, but in some
flows, some parts of this process occur via bifurcations. As one
example, consider a cylinder placed with its axis perpendicular to
the direction of flow. At low Reynolds numbers, the flow is
steady. At slightly higher Reynolds numbers, an oscillatory flow
is displayed behind the cylinder, with an alternate row of vortices
moving downstream, as shown in Fig. 5. This transition from a
steady flow to an oscillatory flow is the result of a Hopf bifurcation [5]. A Hopf bifurcation is characterized by a change from a
steady state to an oscillatory state as a particular parameter, in this
case the Reynolds number, crosses a critical value. Moreover, as
is seen in the figure, the amplitude of oscillation gets larger as the
control parameter moves further away from critical. At higher
Reynolds numbers than those shown, other instabilities occur,
until the wake of the cylinder becomes turbulent. We shall discuss
more about the transition to turbulence in the context of a boundary layer below, but let us first discuss some standard bifurcations.
The canonical equations for some standard bifurcations in simple nonlinear systems and the type of bifurcation that results from
the change of a parameter r are shown in Figs. 6 to 9. A transcritical bifurcation occurs when there are two fixed points, one stable
and one unstable. These exchange their stabilities across the bifurcation. Phase portraits for different values of the control parameter
r are shown in Fig. 6. In fluid flows, r is the Reynolds number, the

A bifurcation point is one where the dynamics changes in

character. Why are we interested in studying bifurcations?

Fig. 4 A phase plane of a nonlinear system with multiple fixed

points. There are two saddles, at (0,0) and (1.1, 21.59), a stable
node, at 20:59; 20:77, and an unstable node at (0.82,1.17).

Applied Mechanics Reviews

Fig. 6 Phase portrait in a transcritical bifurcation. The

exchange of stabilities between the two fixed points as r
crosses zero is evident in this figure. In this and subsequent
figures, filled circles correspond to stable fixed points, and
open circles indicate unstable fixed points.

MARCH 2014, Vol. 66 / 024802-3

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

Rayleigh number etc. At the value of r where the bifurcation

occurs (here r 0) the stable fixed point loses its stability while
the unstable one gains stability. There is thus an exchange of
stabilities, as shown in Fig. 7. A saddle-node bifurcation, shown
in Fig. 8, goes by many other names as well. Here a stable and an
unstable fixed point are neutralized at the bifurcation, so for values of r beyond the bifurcation point, there are no fixed points. It
is not evident in a one-dimensional system why a saddle-node
bifurcation goes by this name (this becomes evident in a twodimensional system), but it is very clear from Fig. 9 why a pitch-

fork bifurcation is called by that name! In the supercritical pitchfork, a stable fixed point becomes unstable at the bifurcation
point, but two other stable fixed points are generated, as seen. The
subcritical pitchfork is the same except for a sign change in all the
growth rates. The two unstable fixed points can be stabilized (and
often are, in real systems) by a fifth-order term of the opposite
sign, and such systems display hysteresis. This is an interesting
feature of nonlinear systems that is not discussed further here, but
occurs in thermoacoustic [6] and many other systems.
In a continuation of exercises from the previous section:
Exercise (iv): Plot trajectories in phase space on either side of
(a) a transcritical bifurcation, (b) a saddle-node bifurcation, (c) a
supercritical pitchfork bifurcation, (d) a subcritical pitchfork
bifurcation with a higher-order stabilization.
Exercise (v): Plot trajectories in phase space on either side
(negative and positive r) of a Hopf bifurcation, given by the
equations
a_ ra  a3 ;

Fig. 7 The nature of fixed points in a transcritical bifurcation.

As r crosses zero, the two fixed points exchange stabilities. In
such diagrams solid lines correspond to stable fixed points,
whereas dashed lines show how unstable fixed points move as
r varies.

Fig. 8 Phase portrait for a prototypical saddle-node bifurcation. The two fixed points annihilate each other.

Fig. 9 Subcritical and supercritical bifurcations; this figure is

taken from Wikipedia

024802-4 / Vol. 66, MARCH 2014

h_ 1 a2

(9)

Here a and h are the radial and angular coordinates, respectively.

At r < 0, the steady state a 0 is attained. As r crosses zero, an
oscillatory state, characterized by constant angular velocity
motion on a circle, is attained at large times. Notice that the amplitude of oscillation, i.e., the radius of the circle, increases as the
square root of the parameter r. The Hopf bifurcation can be
thought of as an oscillatory version of the supercritical pitchfork.
In the Hopf bifurcation seen in Fig. 5, the Reynolds number Re
plays the role of r in the above exercise.
Can we study a simple equation that displays a Hopf bifurcation, and therefore some of the features of the flow past a cylinder? A canonical nonlinear system that a fluid mechanicist can
identify with is the van der Pol oscillator, which is described in
Sec. 4. Before that we return to the boundary layer, to contrast it
with the flow past a cylinder, and to reiterate that transition to turbulence is far more complicated than just a series of bifurcations
that can be modeled by simple equations. While the steady state
in the flow past a cylinder gave way to self-sustained oscillations
of a certain frequency, the onset of the Tollmien-Schlichting instability in a boundary layer (at low free-stream turbulence levels)
marks a transition from a steady state to an oscillatory state. This
seems at first sight similar to a Hopf bifurcation, but there is a basic difference. These oscillatory solutions are not limit cycles. To
distinguish this basic difference, flows such as this one are
referred to as amplifiers, where a perturbation of a given frequency can be made to amplify by the system. The oscillations
grow in time with a very small growth rate, so the amplitude is
not a fixed function of Reynolds number. At every Reynolds number, or equivalently, every downstream location, the growth rate is
different. These experimentally observed oscillations are well
predicted by linear theory, whereas a Hopf bifurcation is the result
of a nonlinear saturation to a new steady state. Figure 10 shows
Roddam Narasimhas sketch of the transition process at low
free-stream turbulence levels. The Reynolds number increases
downstream, so the entire route to turbulence can be seen in one
snapshot. The straight solid lines (green in color) represent peaks
in amplitude of the TollmienSchlichting waves that appear periodically and move downstream. On the other hand, globally unstable systems such as that of Fig. 5 are called oscillators, which
display their inherent frequency and amplitude.
The growth of TollmienSchlichting waves is very slow, so at a
given time and Reynolds number, we may approximate the flow
to one with constant amplitude oscillations. Under certain conditions of free-stream turbulence, this oscillatory state can be followed by another instability, to give a new state termed as the
Klebanoff mode, shown by the wavy green lines. This instability
is called the secondary instability and can be predicted by the Floquet approach described later in this article. The new state is still
periodic but has a wavy structure in the spanwise direction. Note
that the term instability is more appropriate here than
Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

wires in the wind and several other situations. Wherever one finds
a self-sustained oscillation, one immediately thinks of the van der
Pol oscillator. Other reviews in this issue discuss flows that are
globally unstable, and it may be seen that they often display selfsustained oscillations of a characteristic frequency. The frequency
and the shape of the disturbance structure are independent of the
details of the disturbance environment. This is a completely nonlinear phenomenon, more commonly known as a stable limit cycle
in phase space. A van der Pol oscillator displays a limit cycle. The
frequency of this limit cycle may be estimated from the linear
approximation, as we shall see. This will give us a basic idea of
how to obtain the periodicity of a nonlinear system by expanding
about the linear answer.
The differential equation
Fig. 10 Top view of the boundary layer on a flat plate, sketch
of Roddam Narasimha. The thick vertical line marks the transition from periodic to chaotic flow.

bifurcation, although the character of the flow has changed from

one type of oscillations to another. One reason is that the new
oscillations do not constitute limit cycles either. Moreover, the
change can occur over a range of Reynolds numbers rather than a
sharply identified Reynolds number, depending on the nature and
levels of disturbance in the external flow. As we go further downstream a host of not completely understood changes follow, until
turbulent spots appear. The turbulent spots grow and merge as
they move further downstream until the entire flow is turbulent.
At a slightly higher free stream turbulence than that shown in
Fig. 10, we get, instead of the Klebanoff mode, another spanwise
structure called the Craik mode, where the wavy patterns are
arranged as shown in Fig. 11. It is seen that the pattern is repeated
once every two amplitude-maxima, so the period has doubled as
compared to the previous stage of two-dimensional Tollmien
Schlichting waves. We shall talk about period doubling in the
context of the logistic map.

The van der Pol Oscillator

In Sec. 3, we saw how vortices of a certain amplitude are shed

behind a cylinder at a particular frequency, when the cylinder is
placed in an oncoming flow at a moderate Reynolds number. The
Hopf bifurcation that occurs there bears similarity to a far simpler
counterpart: the van der Pol equation. The equation and its variants have been used to model lift and drag in such Karman
streets, the flow past a slender cone [7], flow past airfoils with
drag-reducing hair like structures [8], the whistling of electric

Fig. 11 A schematic of the Craik mode of secondary instability: an example of spatial period doubling. The lines show maxima in the disturbance amplitude, which move downstream as
well as in the spanwise direction in a sinusoidal manner.

Applied Mechanics Reviews

x x2  1x_ x 0 with 0 < 

(10)

describes the van der Pol oscillator (see, e.g., Ref. [9]). This
belongs to a class of planar systems, i.e., systems modeled as
second-order ordinary differential equations, where a special periodic solution as described above, namely a limit cycle, exists in the
x-x_ plane. (Note that in the phase plane x r cos h.) In Fig. 12, the
origin is seen to be an unstable equilibrium point of Eq. (10). With
 > 0, a trajectory starting sufficiently close to the origin (let us say
_  0:5; 0, shown by the blue dashed-dotted line) will spiat x; x
ral outward away from the origin, since for this initial condition,
the damping term has a negative coefficient, and so acts to
_ On
amplify. Note that the damping is given by the coefficient of x.
the other hand a trajectory that starts with an initial condition suffi_  5; 0, shown by the
ciently far from the origin, (e.g., at x; x
pink solid line in the figure), spirals inwards since the damping is
positive. The two trajectories cannot intersect each other, as discussed above. Both asymptotically approach the same isolated,
simple, closed curve in the phase-plane, in other words, the limit
cycle. This curve divides the phase plane into two regions. If
instead  in Eq. (10) were to be less than 0, we would have an
unstable limit cycle. In this case, the origin will be stable and trajectories from the neighborhood of the limit cycle will tend to go
away from it as t ! 1.
Limit cycles are to be found in systems that are much more general than the van der Pol oscillator. How do we characterize a
limit cycle quantitatively and qualitatively? We need its shape
and size as well as its stability.

Fig. 12 Phase plane of van der Pol oscillator with  5 0.1. The
dash-dot trajectory emerges out of the origin and approaches
the limit cycle, while the pink line begins from infinity and
moves inwards towards the limit cycle. The limit cycle is where
the pink and blue trajectories meet.

MARCH 2014, Vol. 66 / 024802-5

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

4.1 Period of the Limit Cycle by LindstedtPoincare

Method. We wish to obtain the time period of one traversal of the
limit cycle for the van der Pol oscillator. We approach this problem via a singular perturbation method called the
LindstedtPoincare. This is a routinely applied method to construct an analytical but approximate periodic solution of a nonlinear ordinary differential equation.
For applying the perturbation method, we further assume
0 <   1. As usual with perturbation methods [9, 10], we
expand the solution in a power series in , i.e.,
xt x0 t x1 t 2 x2 t   

(11)

We expect the above expansion to be uniformly valid, i.e., each

term added to the series is one order smaller than the previous
term [11]. Given the assumption imposed on , this is equivalent
to saying xk t  O1 for every k and for all time t > 0.
Let T be the time period of the periodic solution we are looking
for, and x 2p=T be the corresponding frequency. The key idea
is to expand x as a power series in  in terms of unknown coefficients x1 ; x2 , etc. Such an approach is also called the method of
strained parameters [10, 12]. We write
x 1 x1 2 x2   

(12)

where x0 1 is the frequency of the unperturbed oscillator, i.e.,

with  0. This is the most important step in the LindstedtPoincare method. The unknown xi s are to be found using
conditions that make the expansion uniformly valid. One of the
significant differences between linear and nonlinear oscillators is
the independence and dependence, respectively, of their time periods on the amplitude of oscillation. The next step in the process is
to stretch the time by a change of the independent co-ordinate
from t to s using
s xt

(13)

Substituting this into the equation at O; we get

1
1
x001 x1 A2  1  2x1 A coss A2 cos2s
2
2
Solving the above for x1(s) gives us the term that grows secularly,
i.e., unboundedly in time. This is because we now have a forcing
term, 2x1 A coss, whose frequency [1=(2p)] is the same as
the natural frequency of the oscillator. Such a forcing term is
termed the secular term. As x1(s) grows, after a time O1= (in s
units), the expansion will break down because x1 s will be as
big (or as small) as x0(s). In order to make the expansion uniformly valid, we make the coefficient of coss equal to zero. This
key step, peculiar to singular perturbation methods, is called the
removal of secular terms [9]. This gives
x1 0
After the removal of the secular term, we solve O equation for
x1 s with x1 0 0 and x_ 1 0 0 to get


1 2
1 2
1
A  1 coss A2 cos2s
(14)
x1 s 1  A
2
3
6
Substituting for x0 s from Eq. (13) and x1 s from Eq. (14) into
equation at O2 ; we get


5 3
00
x2 x2 2A 2x2 A A coss
6


1 3
1
A  A 1 cos2s A3 cos3s
3
6
Removal of another secular term now gives

We rewrite Eq. (10) in stretched time as

x2 x00 xx2  1x0 x 0
where prime denotes differentiation with respect to s. Substituting
x from Eq. (12) and using Eq. (11) in the above equation, collecting terms of various powers of , we obtain
O0 : x000 x0 0
O : x001 x1 x0 x20  1 2x1 x000
O2 : x002 x2 x20  1x1 x00 x01  2x0 x1 x00  2x1 x001
 2x2 x21 x000
We solve Eq. (10) for the initial condition (IC)
x0 A and

x0 t A cost

_
x0
0

_
We choose x0
0, and we can do this without loss of generality,
as the van der Pol oscillator is autonomous. In other words, we can
choose the origin for time to correspond to a location anywhere on
the limit cycle, since the oscillator is autonomous. Assuming A to
be of O1; we write down ICs for x0 t; x1 t;    as
x0 0 A;

x1 0 0;

x2 0 0;   

x_0 0 0;

x_ 1 0 0;

x_ 2 0 0;   

and

With above mentioned ICs, solving the O0 equation above, we

get
024802-6 / Vol. 66, MARCH 2014

x2 1 

5 2
A
12

Thus, we obtain an amplitude dependent frequency and the time

period of the van der Pol oscillator, correct up to O3 , respectively, as




5
5
and T 1  2 1  A2
x 1 2 1  A 2
12
12
We recall that the van der Pol equation is a good model for
describing the dynamics on a cylinder in a range of Reynolds
numbers, and we can thus find an application for estimating its
frequency of oscillation. There are other applications.
Exercise (vi): Using the above method, obtain the period of the
following nonlinear oscillator up to O2 :
x x x2 2 x3 0
4.2 Stability of the Limit Cycle via Method of Multiple
Scales. Now, we apply another singular perturbation method
called the method of the multiple scales to the van der Pol oscillator. The objective is to get an estimate of the size and shape of the
limit cycle in the phase plane and to know whether the limit cycle
is stable or not, meaning, do the nearby trajectories get attracted
toward the limit cycle or do they go away from the limit cycle?
Note that we need not assume here that we are looking for a periodic solution, as we needed to in the LindstedtPoincare method.
We begin with the assumption that the solution x(t) we are
looking for depends on various time scales [13]
T0 t; T1 t; T2 2 t;   
Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

2
Rt v
!
u
u
4
t1
 1 et
R02

where T0 is the fast time scale and T1 ; T2 ;    are slower ones, so

we have
xt XT0 ; T1 ;   
Further the solution X is expanded as a power series in ,
2

XT0 ; T1 ;    X0 T0 ; T1 ;    X1 T0 ; T1 ;    O (15)

The derivatives with respect to t become
d @
@


O2
dt
@T0
@T1

(16)

d2  @ 2 
@ 2 

2
O2
2
dt2
@T0 @T1
@T0

(17)

Substituting Eqs. (15) through (17) in Eq. (10) and expanding and
collecting terms, we obtain at O1 and at O
@ 2 X0
X0 0
@T02

(18)

@ 2 X1
@ 2 X0
@X0
X1 2
 X02  1
@T1 @T0
@T0
@T02

(19)

respectively. Remembering that X0 is a function of more than one

variable, Eq. (18) can be solved for X0 as
X0 AT1 ; cos T0 BT1 ; sin T0
Substituting for X0 from the above into Eq. (19), we get


@ 2 X1
@A
A 2
2
A

2

A

sin T0
1
@T1
4
@T02


@B
B 2
2
2
B  A B cos T0
@T1
4
A 2
B
A  3B2 sin 3T0 B2  3A2 cos 3T0
4
4
Removing the secular terms, i.e., equating coefficients of the resonant terms to zero, we get the slow flow
@A A A 2
 A B2
@T1 2 8
@B B B 2
 A B2
@T1 2 8
Transforming to polar coordinates A R cos /; B R sin / and
using
@A
A_ 
@T1

and

@B
B_ 
@T1

we get, up to O2 ,


R R3

and
R_ 
2
8

/_ 0

Slow flow equilibria correspond to periodic solutions of the system. For the van der Pol oscillator, these are
R 0 and

R2

R 2 is nothing but the limit cycle amplitude as can be seen from

Fig. 12. We solve equation for R_ with IC R(0) to get
Applied Mechanics Reviews

As t ! 1; Rt ! 2; hence, the limit cycle is stable.

It is suggested to the reader that they may consult Ref. [13] to
study the method of multiple scales further. To know more about
the van der Pol oscillator and limit cycles, refer to Ref. [14].
Let us now consider another canonical nonlinear system, the
Duffing oscillator, and apply the methods we have discussed
above to understand it. The Duffing oscillator is a simplified version of a simple pendulum, and we will see that under certain conditions it can display chaotic behavior.

Duffing Oscillator, Unforced, and Forced

The motion of a simple pendulum is governed by
g
h sin h 0
l

where g is the acceleration due to gravity and l is the length of the

3
5
pendulum. Expanding sin h h  hp
=6
Oh , changing the
dependent variable from h to x h= , stretching the independent variable from t to s g=l1=2 t, and ignoring O2 terms, we
get
x00 x ax3 0
where a 1=6 and prime denotes differentiation with respect to
s. This is the equation for the Duffing oscillator, one of the simplest nonlinear oscillators. The LindstedtPoincare method gives
the time period T (correct up to O2 ) as


3
T 2p 1   aA2
8
indicating again an amplitude dependence of the period. This amplitude dependence is an essential feature of nonlinear oscillators
and differentiates nonlinear systems from linear ones.
We now consider the damped, externally forced Duffing
oscillator
x dx_ bx ax3 c cosxt

(20)

For a 0, the oscillator is linear and has a unique steady state solution. For each set of initial conditions, the solution settles down
to a unique amplitude. Its frequency is the same as the forcing frequency and its amplitude depends upon the forcing amplitude c,
damping coefficient d and linear stiffness b. Nonlinear systems
(e.g., the Duffing oscillator in the a 6 0 case) differ from linear
systems as they may settle down to multiple periodic solutions,
stable or unstable, depending upon initial conditions. Figure 13
shows two such stable solutions found numerically. The frequency
of each is the same as the forcing frequency, but the solutions
differ in amplitudes, which are decided by their initial conditions,
_
_
here IC1, i.e., x0; x0
1; 0 and IC2, i.e., x0; x0
7; 0. Every stable periodic solution has its own basin of
attraction, i.e., a set of initial conditions from which the system
settles down to that particular periodic solution.
We now use the heuristic analytical but approximate method
of harmonic balance (MHB) in conjuction with a numerical
arclength-based continuation method to obtain the nonlinear harmonic response of the Duffing oscillator. MHB, a special case of
Galerkin projections, approximates periodic solutions of nonlinear
oscillators using a truncated Fourier series representation. Here,
we assume the periodic solution to be a single-term harmonic balance solution of the frequency of forcing [15],
MARCH 2014, Vol. 66 / 024802-7

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

Fig. 13 Multiple periodic solutions for the forced Duffing oscillator with a 5 1; d 5 0:1; b 5 1; c 5 1; x 5 2:5 with initial conditions
IC1 5 (1,0) (small amplitude) and IC2 5 (7,0) (large amplitude).

Fig. 14 Harmonic response of the Duffing equation for different strengths of nonlinearity a

x  A sinxt B cosxt
Substituting the above into Eq. (20), multiplying by sinxt and
cosxt, then integrating with respect to t from 0 to 2p=x (which
is nothing but collecting coefficients of sinxt and cosxt), we
get two simultaneous nonlinear algebraic equations
3
3
 Ax2  dxB bA aA3 aAB2  c 0
4
4

(21)

3
3
 Bx2 dxA bB aB3 aA2 B 0
4
4

(22)

and

We can solve Eqs. (21) and (22) simultaneously for A and B for
some fixed values of other parameters. Using a numerical technique called arclength-based continuation (also called pseudo
arclength continuation
p method) [16, 17], we find the approximate
amplitude A2 B2 as a function of x, i.e., the nonlinear harmonic response of the system. Putting d 0 and c 0 in Eqs. 21
and 22, we obtain B 0 from Eq. (22) and then substituting so in
Eq. (21), we get
r
2 x2  1
(23)
A6
3
a
The above result though approximate in nature, tells us that for
undamped, unforced periodic oscillations, the frequency of oscillations (and, hence, the period) depends on the amplitude. As we
have discussed, this is typical of nonlinear systems. Equation (23)
is the equation of a backbone curve (the continuous curve in
Fig. 16). The case of a > 0 is called a stiffening nonlinearity while
a < 0 case is a softening nonlinearity [18].
Figures 1416 show the effect of varying nonlinearity strength
a, damping d and forcing amplitude c, respectively, on the frequency response of Eq. (20). As can be seen from Fig. 14, increasing a while holding other parameters constant, the resonance
curve along with the backbone curve leans over to the right for a
stiffening nonlinearity and to the left (not shown) for a softening
nonlinearity. Figure 15 shows the effect of varying the damping d
while keeping other parameters constant. The hump in the amplitude versus frequency curve follows the backbone curve in each
case, hence, the importance of the backbone curve. Decreasing d
raises the hump, i.e., raises the maximum possible response
024802-8 / Vol. 66, MARCH 2014

Fig. 15 Nonlinear harmonic response of the Duffing equation

for different damping coefficients d

Fig. 16 Harmonic response of Duffing equation for different

forcing amplitudes c

Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

amplitude for the given amplitude of harmonic forcing. For smaller

and smaller values of d and c, the response curve draws closer and
closer to the backbone curve. Figure 16 shows the effect of varying
forcing amplitude c while keeping other parameters fixed. As we
increase the magnitude of forcing amplitude c, the system response
grows. The response follows the backbone as well.
Equation (20) is nonautonomous in nature since the independent variable t appears explicitly on the right hand side. Converting
it to an autonomous equation will increase its dimension by 1,
making it three-dimensional. Therefore, the phase plane is an
inadequate tool to analyze this system. The vector field at a given
point changes in time, allowing a trajectory in the phase plane to
appear to intersect itself. The fact that the system is threedimensional allows, as mentioned above, for the possibility of
chaotic behavior. In order to study the behavior, we now use a
Poincare section. Poincare sections [9] are routinely used to
reduce the dynamics of such three-dimensional systems to two by
introducing a plane transverse to the flow, and marking all points
where a trajectory intersects this plane from a given direction of
approach. The map that takes the current intersection to the next
is called the Poincare return map. Studying the map dynamics
helps us to gain insight into the original systems dynamics. For
example, a fixed point of the Poincare map, where the trajectory
cuts the plane of interest at the same location every time,
corresponds to a periodic motion of the system. A series of neverrepeating points corresponds usually to a chaotic case. Figure 17
shows such a Poincare section for Eq. (20) generated using script
file Poincaresection_script.m and function file duffing.m. The
function file writes the Duffing oscillator equation in a standard
format so that built-in routine ode45 can numerically integrate
the oscillator equation. The script file plots the Poincare section.
For the specified parameter values, the system is chaotic and the
Poincare section is a fractal set. It is clear that while the points are
not repeating, they are not completely randomly distributed in
space. The points seem to all lie on a particular pattern occupying
a subset of phase space. Trajectories intersecting the plane somewhere on this pattern clearly form an attractor for this dynamics.
This attractor is not a fixed point or a limit cycle, but the threedimensional extension of the fractal set we see here. Such an
attractor is called a strange attractor, as we shall see. To study the
Duffing oscillator further, the reader is suggested to go through
Refs. [14,18,19].
The reader, after successful completion of these exercises, will
be able to generate Poincare sections for different nonautonomous
oscillators.
Exercise (vii): As shown in Fig. 17, plot the graph for the following parameter values d 0:5; b 1:3; a 0:1125; c 2;
x 3:14 with the help of the following MATLAB files.

Exercise (viii): On similar lines to the code given, write a code

for the forced van der Pol oscillator
x x2  1x_ x F sinxt
Generate
Poincare
 0:1; F 2; x 5.

plot

for

parameter

values

%Function le dufng.m
function xdotdufng(t,x)
% dene variables globally,
%i.e., across the les
global gamma omega epsilon GAMMA OMEGA
%write 2nd order forced Dufng
%oscillator as two rst order ODEs
xdot(1)-gamma*x(1)omega 2*x(2)epsilon*x(2) 3GAMMA*cos(OMEGA*t);
xdot(2)x(1);
xdotxdot;
end
%Script le Poincaresection_script.m
% dene variables globally, i.e.,
%across the les
global gamma omega epsilon GAMMA OMEGA
%numerical values of the parameters involved
gamma0.1; omega1; epsilon0.25;
OMEGA2; GAMMA1.5;
%number of cycles
N50000;
%using ode45 numerical integration routine,
%integrate forced Dufng oscillator
[t x]ode45(dufng,
linspace(0,N*2*pi/OMEGA,100*N),[0 1]);
%plot Poincare section, x(nT)
%against xdot(nT)
Psectionplot(x(1:100:end,1),x(1:100:end,2),
r.);
%label the gure &
%set x & y label font and fontsize
set(gca,fontsize,21);
set(Psection,MarkerSize,2)
xlabel(x(nT)), ylabel(x(nT));

Floquet Theory and Hills Equation

When we talked about the Klebanoff and Craik modes, the observant reader would have wondered how to study the instability
of an oscillatory flow. We do this by Floquet theory, and a canonical example is discussed below.
6.1 Floquet Theory. Consider an n-dimensional, linear nonautonomous system
x_ Atx
(24)
where the dot denotes differentiation with respect to time t, and
A(t) is periodic in time with a fixed time period T, i.e.,
At T At for all t. Let X(t) be a fundamental solution matrix (FSM) to this system, i.e.,
Xt x1 t; x2 t; ; xn t
where x1 t; x2 t; ; xn t are n linearly independent solutions
of Eq. (24).
It is straightforward to see that if x(t) is a solution to Eq. (24),
then so is x(t T). Since x(t) is a solution, we have
 section for the forced Duffing oscillator for
Fig. 17 Poincare
d 0:1; b 1; a 0:25; c 1:38; x 2

Applied Mechanics Reviews

d
xt Atxt
dt
MARCH 2014, Vol. 66 / 024802-9

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

Replacing t by t T in the above equation, we get

d
xt T At Txt T
dt T
Since dt T dt and A(t) is periodic with period T, we rewrite
the above as
d
xt T Atxt T
dt

6.2 Hills Equation. Consider an inverted pendulum of

length l and mass m as shown in Fig. 18, and that its pivot point is
given a displacement u(t) in the y-direction. The angular displacement h(t) is a generalized coordinate describing the motion of the
pendulum. Acceleration due to gravity g acts in the negative
y-direction. Taking the origin of the coordinate system at some
point on the h 0 line, the coordinates of the center of mass of the
pendulum are
x l sin h;

It follows that if X(t) is a FSM to Eq. (24), then so is X(t T).

Therefore, these two must be related to each other via some constant nonsingular matrix C. We then have

y ut l cos h

where u(t) is the imposed displacement of the pivot. Differentiating with respect to time, we have
_
x_ l cos hh;

y_ u_  l sin hh_

Xt T XtC
~ we denote a special
Setting t 0, we get XT X0C. By Xt
~ I , where I denotes the n n
FSM to Eq. (24) such that X0
~
identity matrix. Then we have C XT
[9] and, therefore,
~
Xt T XtXT

(25)

We now consider another FSM Z(t) to Eq. (24) such that

Zt XtR
where R is arbitrary. Substituting for X(t) from the above equation
in Eq. (25), we get
~
Zt T Zt R1 XTR
~
is diagonal.2 Then we
We now choose R such that R1 XTR
write n decoupled equations as
zi t T ki zi t;

i 1; ; n

(26)

~
ki is ith eigenvalue of XT.
Each zi t is a linear combination of
x1 t; x2 t; ; xn t. Equation (26) is a functional equation. We
assume its solution to be of the form
zi t kkt
i pi t

The Lagrangian [20] of the system is given by


1 
L T  V m x_2 y_ 2 mgl cos h ut
2
and the EulerLagrange equation governing the angular displacement h(t) is given by [20]
 
d @L
@L

0
(27)
dt @ h_
@h
which in the present case gives


u g
sin h 0
h 
l l
Linearizing the trigonometric term for small h, we get


u g
h  h 0
l l
Assuming the pivot displacement to be harmonic, i.e.,
u  cost, the above equation becomes


g  cost

h0
h
l
l

where pi t is some function periodic with period T, the same as

that of A(t). k is an unknown to be determined. Substituting the
assumed form of zi(t) into Eq. (26), we get k 1=T. Therefore,
t=T

zi t ki pi t
At the end of each time period, zi attains the values shown below:
t 0;
t T;
..
.
t mT;

zi 0 pi 0
zi T ki pi 0 ki zi 0

m
zi mT km
i pi 0 ki zi 0

~
Eigenvalues of XT,
i.e., ki s, determine the long-term behavior
of solutions to Eq. (24). If jki j > 1 for some i, then the corresponding solution grows unboundedly and the system is unstable.
If jki j < 1 for all i 1;    ; n, then all solutions remain bounded
as t ! 1 and the system is stable. The matrix T is called the
Floquet matrix and its eigenvalues ki s are called Floquet
multipliers.

~ is diagonalizable.
We assume that Xt

024802-10 / Vol. 66, MARCH 2014

Fig. 18 Inverted pendulum, harmonically excited in y-direction

Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

The above may be considered as an oscillator whose frequency

varies in a periodic fashion with respect to time. In a more compact setting, we can write such an oscillator in a form
x f tx 0

(28)

with f t T f t for all t. Equation (28) is called Hills equation, which appeared in a memoir on the motion of the lunar perigee by Hill in 1886 [21].
We study the asymptotic stability of solutions to Eq. (28) with
the help of Floquet theory. Writing Eq. (28) as a system of two
first-order ordinary differential equations, we get
( ) "
#( )
0
1
x1
d x1

dt x2
x2
f t 0
_ By
where x1 x and x2 x.
"
Xt x1 t x2 t 

x11 t

x12 t

x21 t

x22 t

we denote a fundamental solution matrix (FSM) to Eq. (28). By

definition, we have
"
# "
#"
#
1
1
0
1 x11 t x12 t
d x1 t x2 t

dt x21 t x22 t
f t 0 x21 t x22 t
i.e., x_11 t x21 t;
f tx12 t
We have

x_ 12 t x22 t;

x_21 t f tx11 t and x_22 t

and the other decays exponentially. The solution in general

grows and, hence, the system is unstable.
~
(2) jtrXTj
2: k1;2 1 or k1;2 1. For k1;2 1, out of
two linearly independent solutions x1 t and x2 t, one
grows linearly and the other is periodic with period T i.e.
the period of f(t). For k1;2 1, one grows linearly and the
other is periodic with period 2T. In both cases, the solution
in general grows and, hence, the system is unstable.
~
(3) jtrXTj
< 2: k1 and k2 are a complex conjugate pair. The
two lie on a unit circle since k1 k2 1. Both linearly independent solutions are periodic but with incommensurate
frequencies; so the solution in general is quasi-periodic.
The system is stable.
Note that solutions x1 t and x2 t cannot be expressed in terms
of elementary functions such as sine and cosine. However, since
Eq. (28) is linear, the principle of superposition holds and the
most general solution is given by a linear combination of x1 t
and x2 t. Floquet multipliers move on the real line or on the unit
circle in the complex plane with respect to parameters of function
f(t) as indicated in Fig. 19.
The aim of the following example is to obtain the Floquet multipliers of the Hills equation for a given set of parameter values.
Example: if we set g l  1 for the inverted pendulum, we
have
h 1 costh 0

(29)

We integrate Eq. (29) in MATLAB with ode45 for a period T 2p

with ICs f1; 0gT and f0; 1gT using function file mathieu.m and
script file FloquetMultipliers_script.m given below.
%Function le mathieu.m
function xdotmathieu(t,x)

det Xt x11 tx22 t  x21 tx12 t

Therefore,
d
det Xt x_11 tx22 t x11 tx_ 22 t  x_ 21 tx12 t  x21 tx_ 12 t
dt
Substituting for time derivatives from the above four equations,
we get
d
det Xt 0 and therefore
dt

det Xt constant

~ be the FSM which evolves from the

As denoted earlier let Xt
identity matrix at t 0. Therefore,

%parameter values
delta 1;
epsilon 1;
%Mathieu equation written as
%two rst order ODEs
xdot[x(2);-(deltaepsilon*cos(t))*x(1)];
%Script le FloquetMultipliers_script.m
%set absolute and relative tolerances
optionsodeset(AbsTol,1e-7,RelTol,1e-8);
% Numerical integration of Mathieu equation
%for IC (1,0)
[t,x]ode45(mathieu,[0 2*pi],[1 0],options);
x1x(end,:);

~ det X0
~ 1 det XT
~
det Xt
According to Floquet theory, the asymptotic stability of solutions
to the Hills equation is governed by the Floquet multipliers, i.e.,
~
the eigenvalues of XT.
These satisfy the quadratic equation
~
10
k2  trXTk
~
where 1 det XT.
So the Floquet multipliers are

k1;2

q
~
~ 24
trXT6
trXT
2

~
Based on numerical value of trXT,
we have the following cases:
~
(1) jtrXTj
> 2: Both k1 and k2 are real. If k1 > 1, then
0 < k2 < 1 since k1 k2 1. The second possibility is
k1 < 1, so 0 > k2 > 1. Out of the two linearly independent solutions x1 t and x2 t, one grows exponentially
Applied Mechanics Reviews

Fig. 19 Movement of Floquet multipliers with respect to

parameters of f(t)

MARCH 2014, Vol. 66 / 024802-11

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

% Numerical integration of Mathieu equation

%for IC (0,1)
[t,x]ode45(mathieu,[0 2*pi],[0 1],options);
x2x(end,:);
% Combine x1 and x2 to create Floquet Matrix
X[x1;x2];
det(X), trace(X)
% Floquet multipliers are
%eigenvalues of Floquet Matrix
Fmeig(X);

The Floquet multipliers for the given choice of parameters are

k1 1:7716 and k1 0:5645. So the system is unstable.
Exercise (ix): Obtain Floquet multipliers for g l 10 and
 0.1 for the inverted pendulum. Determine whether the system
is stable for these parameter values.
To study Hills equation further, it is suggested to the reader to
go through Ref. [22].

Chaotic Systems

When we obtained a Poincare section with nonrepeating intersection points for the forced Duffing oscillator, we termed it a chaotic system but did not say why. When do we call the dynamics of
a system chaotic? The most often quoted property of a chaotic
system is that it is sensitive to initial conditions. By this term,
we mean that trajectories beginning from two initial conditions
that are arbitrarily close to each other will diverge exponentially,
so that, prescribing an initial condition to any degree of accuracy
is insufficient to predict the complete dynamics of the system. For
example, let us say Eq. (2) describes chaotic dynamics that we
wish to compute the value of x at all times and that we know the
initial condition x(0) correct to 10 decimal places. Even if we
have a completely accurate solver, our predictions will go very
wrong after some time t. If we know x(0) correct to 20 decimal
places, our predictions would be good for longer, but again at
some long time we would go very wrong. Chaotic dynamics thus
cannot be predicted completely. Such systems by definition will
display trajectories that are not periodic, because periodic dynamics, however long the period, is completely predictable once we
obtain the dynamics within one period. All chaotic systems, however, do contain many orbits that are periodic and have various
periods. In fact, one feature of a chaotic system is that it is dense
in periodic orbits. We note that these periodic orbits are unstable,
unlike limit cycles.
It is pertinent to mention here that not all dynamical systems
that are not periodic are chaotic. An Hamiltonian system of n
degrees of freedom and n integrals or constant of motion, will in
general display quasi-periodic motion on the surface of an n-torus.
Such a system is termed as integrable. Quasi-periodicity is created
by the existence of two or more time scales that are incommensurate (do not have a rational ratio). Trajectories winding around the
torus will cover its surface densely but will not return to the point
they started from. If on the other hand such a system has fewer
than n integrals, it is likely to display chaotic or nonintegrable
behavior. A convincing explanation of this is beyond the scope of
this article, but the interested reader is directed towards any good
text book on classical mechanics, such as Ref. [23].
The discussion above brings to light a fundamental difference
between a system that is deterministic but chaotic, and one that is
noisy. In the former, we know the equations governing the system
exactly but since the smallest changes in initial conditions take us
ultimately on completely different trajectories, we cannot predict
the dynamics completely. In a noisy system, we do not know the
equations of motion exactly, i.e., there are small external forces,
usually random, which produce changes in the dynamics. Of biggest interest to fluid dynamicists is the feature of topological mixing that chaotic systems possess. The effect is the same as when
024802-12 / Vol. 66, MARCH 2014

we mix sugar into tea. By this, every region of phase space will
overlap at some time or other with every other region.
When we talk ofunpredictability, we are usually comfortable
with the idea that a system of many variables, such as our lives,
shows unpredictable turns. On the other hand, when a system that
contains only a few independent unknowns (degrees of freedom)
displays chaotic behavior (three is the minimum number required
for chaos), it is very surprising. In turbulence, although there are
an infinite number of variables (the velocity, temperature, etc., at
each point in space is a variable), it seems that some aspects of
the physics can be reduced to that dominated by a much smaller
set of variables. We should, however, warn the reader of the following: at high Reynolds numbers, it is not clear what the dimensionality of turbulence is. Even at moderate Reynolds numbers,
we often need quite a high dimensionality in our models to get
reasonable answers. In other words, we have not clearly understood even what we need in order to understand turbulence as the
Reynolds number goes higher and higher. In spite of this, lowdimensional models are instructive to study and make a good
starting point toward the understanding of more complicated problems such as turbulence, and life. The section on the Lorenz system will describe the most famous example of low-dimensional
chaos, written for a toy weather problem.
Systems often go through a series of bifurcations to go from ordered to chaotic. The routes to chaos can most often be classified
under the period-doubling route, the RuelleTakensNewhouse
quasi-periodic route, the crisis route, and the intermittency route.
We will talk only about the first here. We had seen that period
doubling occurs in a Craik mode of disturbance growth in the
boundary layer, although it was not via a standard bifurcation.
The following section gives the most famous example of this
route to chaos. The similarity in detail of this simple system to
many far more complicated real systems is simply amazing. The
reader is encouraged to try out some of these.
7.1 The Logistic Map: A Prototype for Period-Doubling
Bifurcations. Human population growth seems unlimited even
today. It appears to be following what Malthus described as an exponential growth. Malthusian law does not take into consideration
the finiteness of natural resources. Verhulst suggested a different
law where growth rate is proportional to the existing population
and at the same time is limited by the carrying capacity of the
environment. Populations of almost all species appear to follow
this law. The discrete version of this law is the famous logistic
map
xn1 rxn 1  xn

(30)

This is a quadratic first-order difference equation. Population evolution as predicted by Eq. (30) strongly depends on the value of
the parameter r. For r < 2.992 trajectories evolve toward a fixed
point x*. The approach to the fixed point is monotonic when
r < 2.35 (refer to Fig. 20(a)). For r lying in the range
2.35 < r < 2.992, the trajectory still approaches its stable fixed
point, but in an oscillatory manner (Fig. 20(b)). As we increase r
further till it reaches the value 2.992, oscillation amplitudes
increase and the decay rate decreases. The fixed point x*
approaches a value of 0.643 as r approaches the numerical value
2.992 from below. As r increases steadily past the value of 2.992,
trajectories of increasing period, doubling at every bifurcation
point, are displayed. A period-k orbit is a trajectory that eventually
settles down to dynamics given by xnk xn for every n > n0 for
some finite n0. It may be given by, for example,
xn0 1 ; xn0 2 ; ; xn0 k . The period is 2 for r up to 3.448. Beyond
r 3.448, the dynamics changes from that on a period-2 orbit
(Fig. 20(c)) to that on a period-4 orbit (Fig. 20(d)). Again as r
crosses 3.544, the steady state changes from period-4 orbit to
period-8 orbit (Fig. 20(e)). Similarly as r crosses 3.568, the steady
state changes from period-8 orbit to period-16 orbit. At r
Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

rar(i);
a0.1;
for j1:900
afeval(@logis,a);
end;
a(1)a;
for j2:length(r)
a(j)feval(@logis,a(j-1));
end
rpra*ones(1,length(r));
bdplotplot(rp,a,.);
end;
%font, fontsize of title and labels
set(gca,fontsize,18);
set(bdplot,MarkerSize,1)
xlabel(r), ylabel(steady state);
title(Bifurcation diagram of a Logistic map)
axis([2.6 4 0 1.1]);
Fig. 20 Trajectories of the logistic map for different values of r

approximately 3.56995 is the onset of chaos, and the end of the

period-doubling cascade. This cascade of period-doubling bifurcations is one of the major routes from nonchaotic to chaotic behavior. If we denote by rk the value of r for which the trajectory
settles down to period-2k cycle for the first time, then we have the
Feigenbaum constant defined as
rn  rn1
4:66920   
d lim
n!1 rn1  rn
It is seen that period doubling bifurcations occur in increasingly
quick succession for larger periods. This happens in all period
doubling bifurcations, and the Feigenbaum constant is the same
for a wide variety of systems! To generate Fig. 20(a), we run the
script file bdlogis.m, which in turn calls function file logis.m. The
function file just describes the logistic map in MATLAB syntax,
while the script file runs the function file for parameter r values
from 2.6 to 4 in step of 0.004 and plots the solution after transients
die out.
The aim of the following exercise is to generate a bifurcation
diagram of a given map.
Exercise (x): Modifying the routines bdlogis.m and logis.m
given below appropriately, generate a bifurcation diagram for the
following cubic map:
xn1 rxn 1  x2n

The Lorenz System

Lorenz in his remarkable 1963 paper described a simple threedimensional approximation of an infinite-dimensional model of a
weather system, which is now famous as the Lorenz equations. It
describes the convective motion of a 2D fluid cell that is warmed
from below and cooled from above. The resulting partial differential equation is approximated as
x_ ry  x
y_ rx  y  xz
z_ xy  bz
where x physically describes the rate of convective overturning, y
is the horizontal temperature variation, and z the vertical temperature variation. Real positive parameters r and r are proportional to
the Prandtl number, Rayleigh number, while b is proportional to
physical dimensions of the system under consideration (Fig. 21).
Since the system is dissipative all trajectories eventually settle
down to a bounded set of zero volume lying in phase space. This
set may be an equilibrium point, a periodic orbit, or some complicated set of noninteger dimension. The last of these is called as a
strange attractor. Figure 22 shows a trajectory of the system starting with point (0.01,0,0) in the phase portrait. The trajectory is not
periodic. Continuing the numerical integration, the trajectory continues to wind around on one side and then switches from that
side to the other, and continues to wind around that side until the
next switch, without settling down to either periodic or stationary
behavior. This aspect is not a function of a particular choice of

for r varying from 0 to 3. Set the x-axis to [0.6 3] and y-axis to

[1.2 1.2] while plotting.
%Function le logis.m
function flogis(x)
global ra
fra*x*(1-x);
%Script le bdlogis.m
%set ra globally so that
it can be used across the les
global ra;
%range of the parameter r
r2.6:0.004:4;
hold off;
gure(1)
hold on
% for loop to vary parameter r
%in steps of 0.004
for i1:length(r);

Applied Mechanics Reviews

Fig. 21 Bifurcation diagram of a logistic map

MARCH 2014, Vol. 66 / 024802-13

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

Fig. 22 Phase portrait for r 5 10:0; b 5 8=3; r 5 28:0 with IC

(0.01,0,0)

initial conditions or the choice of the numerical scheme employed

but only of the choice of parameter values. Note that a choice of
r 28 makes the system operate in a parameter regime far from
the original fluid dynamical problem, for which r  1. Lorenz
equations are chaotic for the particular choice of parameters
shown in the phase portrait of Fig. 22. The figure is generated by
running the script file Butterfly_script.m, which calls the function
file Lorenz.m. The function file writes down Lorenz system in
MATLABs built-in ode45 syntax and the script file generates the
phase space for Lorenz system using plot3 command.
The origin is an equilibrium point for all parameter choices.
Keeping r 10 and b 8=3; we see interesting dynamics by
changing r. For 0 < r < 1, the origin is globally stable and all trajectories are attracted to it. Stability analysis is based on the linearized flow around an equilibrium point and is determined by the
positive or negative real parts of the eigenvalues of the Jacobian
evaluated at this point. At r 1, a supercritical pitchfork bifurcationpoccurs
two
new equilibrium points appear, given by
and
p

6 br  1; 6 br  1; r  1 [24]. The origin loses its stability. At r 13:926    ; unstable limit cycles are created (due to
a homoclinic bifurcation, which we have not discussed). At this
point, a strange invariant set is born. For 1 < r < 24:74    ; the
other two equilibrium points are stable and the origin is unstable.
At r 24:74    ; a subcritical Hopf bifurcation occurs, the two
equilibrium points other than the origin lose their stability by
absorbing unstable limit cycles. As we increase r further, the Lorenz attractor is displayed and the system continues to be chaotic.
Finally as r is increased beyond 24:74    ; all three equilibrium
points are unstable, the system undergoes period doubling bifurcations and intermittent chaos is observed.
The dynamics on the strange invariant set shows sensitive dependence on initial conditions, usually quantified by Lyapunov
exponents. Figure 23 shows the divergence of two trajectories that
start quite close to each other. The figure is generated by running
script file sensitivity_IC_script.m, which calls function file Lorenz.m. The script file numerically integrates the Lorenz system
with two nearby initial conditions. Lyapunov exponents are a
measure of the exponential rate of divergence of two trajectories
that begin arbitrarily close to each other. A Lyapunov exponent
greater than zero signifies the system to be chaotic.
The aim of this exercise is to generate phase portraits of the
Lorenz system for various parameter values.
Exercise (xi): Generate phase portraits of the Lorenz equations
for the parameter values r 10:0; b 8=3 and for different r
values such as r 0:5; 2; 15 using Lorenz.m for initial conditions
such as (0.01,0,0).
024802-14 / Vol. 66, MARCH 2014

Fig. 23 Two trajectories for the Lorenz system with ICs

(1.5,0,0) and (1.5 1 0.001,0,0) diverge appreciably after t 5 25

%Function le Lorenz.m
function ydotLorenz(t,y)
global z
% parameter values
sigma10;
r28;
b8/3;
% Lorenz system
ydot[sigma*(y(2)-y(1));y(1)*(r-y(3))-y(2);
y(1)*y(2)-b*y(3)];
%Script le sensitivity_IC_script.m
epsil0.001;
% numerical integration of Lorenz system
%for initial condition (1.5,0,0)
[t,x]ode45(Lorenz1,linspace(0,50,1000)
,[1.5 0 0]);
%plotting the solution
plot(t,x(:,1));
hold on;
% numerical integration of Lorenz system
%for initial condition (1.501,0,0)
[t,x]ode45(Lorenz1,linspace(0,50,1000)
,[1.5epsil 0 0]);
%plotting the solution over
the earlier solution
plot(t,x(:,1),r);
%Script le Buttery_script.m
%numerically integrate Lorenz
equations using ode45
[t,x]ode45(Lorenz1,linspace(0,50,30000)
,[0.01 0 0]);
%3-D plot of the phase space
Butteryplot3(x(:,1),x(:,2),x(:,3));
%setting font and fontsize of label and title
set(gca,fontsize,18); set(Buttery,
MarkerSize,5);
xlabel(x(t)), ylabel(y(t)),
zlabel(z(t));
title(Phase portrait for Lorenz equations)
axis([10,20,20,20,0,50]);
grid on;

Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

s 0ek1 s ;
s k2 s
e ; ;
2s
jjsjj
n  1s kn s
e
ns
jjn  1sjj
We define k as
k lim

Fig. 24 Numerical calculation of the largest Lyapunov

exponent

Consider two trajectories that start from nearby initial conditions. Let t be the distance between them. We write it in the linearized form
_ Jt

(31)

with initial condition 0 assumed to be a very small quantity.

The Jacobian for the Lorenz system is given by
2

r

6
Jt 4 r  zt
yt

r
1
xt

n
1X

n!1 ns

k1

lnjjksjj lim

n
1X

n!1 n

kk

k1

Figure 25 shows all the three Lyapunov exponents for the Lorenz
system. The largest is greater than zero and, therefore, the system
is chaotic. The system displays a strange attractor for the chosen
values of the parameters, in accordance with the largest exponent
being positive, the second one zero and the smallest negative. The
magnitude of the largest negative exponent is greater than that of
the largest positive one, given that the Lorenz system is dissipative. To learn more about the Lorenz system and related chaos
theory, it is suggested to the reader to study Ref. [24].

7
xt 5
b

We solve Eq. (31) to get

t eJt 0
To get the Lyapunov exponent numerically, we integrate Eq. (31)
for a large number of cycles (n), each for small but fixed time
interval s. We begin with the IC 0 as shown in Fig. 24. At the
end of the first cycle, we store the natural logarithm of the Euclidean norm (jj  jj) of the end condition s (Euclidean distance
between two given points a and b in terms of Euclidean norm is
given by jja  bjj). We normalize s to unit norm (refer to
Fig. 24) to use this normalized end condition as the IC for the next
cycle to integrate Eq. (31). We repeat this process n times. An
arithmetic average of these stored logarithms of end conditions in
the limit of number of cycles n tending to infinity is nothing but
the largest Lyapunov exponent k [25].

Conclusions

Nonlinear dynamics is too wide a subject to be treated fairly in

a brief review. Our objective has only been to ignite the readers
enthusiasm for this field. Apart from the references we recommend above at the end of some of the subsections, there is an
enormous amount of literature on this topic available in Wikipedia, Scholarpedia, and in many other cites, journals, and books.
The difference in the present treatment is that we have tried to
make connections with flow stability and transition to turbulence.
For further reading on this vast subject, we recommend for general reading excellent books [2628] and for advanced treatment
on a few topics touched upon in this document, we recommend
Ref. [30].

Acknowledgment
Sharath Jose and Mamta Jotkar are thanked for their contributions. The anonymous referees have contributed significantly to
improving the article.

References

Fig. 25 Three Lyapunov exponents of the Lorenz system for

r 5 10:0; b 5 8=3; r 5 28:0

Applied Mechanics Reviews

[1] Strogatz, S., 2001, Nonlinear Dynamics and Chaos: with Applications to Physics,
Biology, Chemistry and Engineering, 1st ed., Westview Press Boulder, CO.
[2] Schneider, T., Marinc, D., and Eckhardt, B., 2010, Localized Edge States Nucleate Turbulence in Extended Plane Couette Cells, J. Fluid Mech., 646, pp.
441451.
[3] Willis, A., and Kerswell, R., 2009, Turbulent Dynamics of Pipe Flow Captured
in a Reduced Model: pu? Relaminarization and Localized Edge States,J. Fluid
Mech., 619, pp. 213233.
[4] Duguet, Y., Schlatter, P., and Henningson, D. S., 2009, Localised Edge States
in Plane Couette Flow, Phys. Fluids, 21, p. 111701.
[5] Manneville, P., and Pomeau, Y., 2009, Transition to Turbulence, Scholarpedia, 4(3), p. 2072.
[6] Subramanian, P., Mariappan, S., Sujith, R. I., and Wahi, P., 2010, Bifurcation
Analysis of Thermoacoustic Instability in a Horizontal Rijke Tube, Int. J.
Spray Combus. Dyn., 2, pp. 325355.
[7] Gaster, M., 1969, Vortex Shedding From Slender Cones at Low Reynolds
Numbers, J. Fluid Mech., 38(3), pp. 565576.
[8] Venkatraman, D., 2013, Computational Models of Dorsal Coverts on Birds
Wings, Ph.D. thesis, University of Genova, Italy.
[9] Rand, R., Lecture notes on Nonlinear Vibrations. Version 52. Available at:
http://www.tam.cornell.edu/randdocs/ed
[10] Hinch, E. J., 1991, Perturbation Methods, Cambridge University Press, Cambridge, UK.
[11] Johnson, R. S., 2005, Singular Perturbation Theory, Springer ebooks.
[12] Holmes, M. H., 1991, In Introduction to Perturbation Methods, Springer-Verlag,
New York.
[13] Kevorkian, J., and Cole, J. D., 1996, Multiple Scale and Singular Perturbation
Methods, Springer-Verlag, New York.

MARCH 2014, Vol. 66 / 024802-15

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms

[14] Mickens, R. E., 1981, An Introduction to Nonlinear Oscillations, 1st ed., Cambridge University Press Cambridge, UK.
[15] Mickens, R. E., 1996, Oscillations in Planar Dynamic Systems, World Scientific Publishing Co., River Edge, New Jersey.
[16] Chatterjee, A., 2002, An Elementary Continuation Technique, http://home.iitk.
ac.in/anindya/continuation.pdf
[17] Morgan, A., 2009, Solving Polynomial Systems using Continuation for Engineering and Scientific Problems, Classics in Applied Mathematics, SIAM.
[18] Stoker, J. J., 1950, Nonlinear Vibrations in Mechanical and Electrical Systems,
John Wiley & Sons, New York.
[19] Verhulst, F., 1996, Nonlinear Differential Equations and Dynamical Systems,
2nd ed., Springer, New York.
[20] Lanczos, C., 1986, The Variational Principles of Mechanics, Dover Publications, New York.
[21] Hill, G. W., 1886, On the Part of the Motion of the Lunar Perigee is a Function
of the Mean Motions of the Sun and the Moon, Acta Math., 8, pp. 136.
[22] Magnus, W., and Winkler, S., 2004, Hills equation, Dover Publications, Mineola, NY.

024802-16 / Vol. 66, MARCH 2014

[23] Rana, N., and Joag, P., 2011, Classical Mechanics, 30th ed., Tata-McGraw
Hill.
[24] Sparrow, C., 1982, The Lorenz Equations: Bifurcations, Chaos and
Strange Attractors, Applied Mathematical Sciences 41, Springer-Verlag,
New York.
[25] Rugonyi, S., and Bathe, K., 2003, An Evaluation of the Lyapunov Characteristic Exponent of Chaotic Continuous Systems, Int. J. Numer. Methods Eng., 56,
pp. 145163.
[26] Wiggins, S., 2003, Introduction to Applied Nonlinear Dynamics and Chaos,
2nd ed., Springer, New York.
[27] Hale, J. K., 1963, Oscillations in Nonlinear Systems, 1st ed., Dover Publications
Mineola, NY.
[28] Drazin, P. G., 1992, Nonlinear Systems, 1st ed., Cambridge University Press,
Cambridge, UK.
[29] Grimshaw, R., 1991, Nonlinear Ordinary Differential Equations, Taylor &
Francis, London.
[30] Glendinning, P., 1994, Stability, Instability and Chaos, Cambridge University
Press, Cambridge, UK.

Transactions of the ASME

Downloaded From: http://appliedmechanicsreviews.asmedigitalcollection.asme.org/ on 06/30/2015 Terms of Use: http://asme.org/terms