Dynamical systems 2019 kristian.gustafsson@physics.gu.

se

8 Example 1: The van der Pol oscillator

(Strogatz Chapter 7)
So far we have seen some different possibilities of what can happen in
two-dimensional systems (local and global attractors and bifurcations)
using mainly constructed examples. In this and the next lectures we
consider two examples with real-world applications.

8.1 Self-sustained oscillations

Self-sustained oscillations are frequent in nature and in technology,
some examples being stick-slip oscillations, unwanted mechanical vi-
brations, and oscillators in biology. Usually a system with self-sustained
oscillations has a feedback mechanism such that small-amplitude os-
cillations grow in size. The small oscillations are often modeled using
an oscillator with negative damping, γ = −µ < 0

ẍ − µẋ + ω02x = 0 .

The system for x and y = ẋ has an unstable spiral at the origin, blow-
ing up small-scale oscillations. As oscillations grow, non-linear terms
may form a stable limit cycle and the system shows self-sustained os-
cillations. One simple and common model in mathematics, physics,
engineering, biology and economic theory, is the van der Pol oscillator:

ẍ + µ(x2 − 1) ẋ + x = 0 (1)
| {z }
f (x)

Nonlinear damping coefficient f (x) damps large oscillations (friction

when |x| > 1) and amplifies small oscillations (forcing when |x| < 1)
⇒ We expect self-sustained oscillations to be possible.
Indeed the corresponding dynamical system

ẋ = y
ẏ = −x − µ(x2 − 1)y

1
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

√
shows a√stable limit cycle if µ > 0 (after a rescaling x = x0/ µ and
y = y 0/ µ, the system has a supercritical Hopf bifurcation):

Period time and shape of the cycle depends on µ.

The van der Pol oscillator is an example of a Liénard system

ẍ + f (x) ẋ + g(x) = 0
| {z } |{z}
even odd
For such systems a unique stable limit cycle surrounds the origin if
certain criteria (Strogatz 7.4) are fulfilled that ensures that:
• the nonlinear damping f is negative for small |x| and positive
for large x
• displacements are reduced by the nonlinear restoring force g
The van der Pol oscillator, Eq. (1), can not be solved analytically for
general values of µ. In certain limits however, we can find approximate
solutions, as seen by the following sections.

2
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

8.2 Relaxation oscillations: Case of large µ

Now consider the van der Pol oscillator (1)

ẍ + µ(x2 − 1)ẋ + x = 0

with µ  1. Let  = 1/µ  1 be small. Let y = ẋ + F (x) with

F (x) = x3/3 − x:

ẋ = y − F (x)
ẏ = ẍ + F 0(x)ẋ = −x .

Two kinds of dynamics emerge:

Fast: If |y − F (x)| ∼ 1: |ẋ|  1 and |ẏ| ∼   1
Slow: If |y − F (x)| ∼ 2: |ẋ| ∼  and |ẏ| ∼ 

Dynamics can be understood by plotting the nullclines ẋ = 0

(y = F (x)) and ẏ = 0

Starting from any point (except fixed point in origin) the trajectory
moves quickly horizontally onto the cubic nullcline y = F (x), then it
moves slowly along nullcline until the ’jump-off points’ (max and min
of F (x)) where the direction of the flow and the nullcline starts to
deviate. After the jump-off point the trajectory quickly moves over to

3
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

the branch on the opposite side, and so it continues (the red underly-
ing limit cycle is approached).

The period time can be approximated by the travel time along the
two slow branches. On the slow branches y ≈ F (x) ⇒ ẏ ≈ F 0(x)ẋ.
But we also have ẏ = −x
dx 1
⇒ F 0(x) ≈ −x , ⇒ dt ≈ − F 0(x)dx
dt Z x  x2
x2
1 0 1 x2
⇒ Tslow ≈− F (x)dx = − − ln x
x1 x  2 x1
1
= [Take slow branch from x1 = 2 to x2 = 1] = [3 − 2 ln 2]
2
Slow compared to Tf ast ∼  ⇒ period time is ≈ 2Tslow .

This is an example of a relaxation oscillator:

Relaxation oscillations: very slow build-up and sudden discharge

(for example periodic firing of nerve cells, geysers, stick-slip oscilla-
tions, e.g. squeaking of door hinges, of chalk on a blackboard, of
violin bow setting strings in vibration),. . .
Relaxation oscillators have two widely separated time scales act-
ing sequentially. In the opposite limit µ  1 the situation is more
complicated: two time scales act at the same time.

4
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

8.3 van der Pol oscillator with small µ

Consider the van der Pol oscillator Eq. (1) with 0 < µ  1 and some
initial condition x(0) = 0, ẋ(0) = 1 (arbitrary condition that starts
off the limit cycle).

8.3.1 Regular perturbation theory

Search for a solution for small values of µ by a series expansion

x(t) = x0(t) + µx1(t) + µ2x2(t) + . . .

and insert this expansion into Eq. (1) and collect terms to order µ

0 = ẍ + µ(x2 − 1)ẋ + x
= [ẍ0 + µẍ1] + µ([x0 + µx1]2 − 1)[ẋ0 + µẋ1] + [x0 + µx1] + O(µ2)
= ẍ0 + x0 + µ[ẍ1 + (x20 − 1)ẋ0 + x1] + O(µ2) .

With initial conditions

0 = x(0) = x0(0) +µ x1(0)

| {z } | {z }
=0 =0
1 = ẋ(0) = ẋ0(0) +µ ẋ1(0) .
| {z } | {z }
=1 =0

To order µ0 we have

ẍ0 + x0 = 0 ⇒ x0(t) = sin t .

To order µ1 we have

ẍ1 + (sin2 t − 1) cos t + x1 = 0 ⇒ x1(t) = (6t + sin(2t)) sin t/16 .

Problem: x1 contains secular terms (terms that approaches infinity

as t → ∞):

5
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

Exact
Pert. theory
t = 1/µ
t
Thus the perturbation theory fails to describe the formation of the
limit cycle (the amplitude of oscillations grow to infinity). Even
though the perturbation theory is identical to the series expansion
of the actual solution, it fails for times of order t ∼ 1/µ.
In order to obtain a perturbation theory valid for large values of t,
we need to make a high-order expansion. Alternatively, we can notice
that there are (at least) two time scales in the problem: one for the
oscillations (fast, O(1)) and one for the peak amplitude (slow,O(1/µ)).
Separating these time scales in the perturbation expansion, so called
two-timing, gives more accurate results for large t even though we only
consider the lowest order in µ.

8.3.2 Two-timing
Let τ = t denote the fast time scale and T = µt denote the slow
time scale (T is of order unity when t ∼ 1/µ  1) and treat these
as independent variables: x = x(τ, T ). The reason this works is that
when we have a large time separation T  τ , x is roughly constant
w.r.t. T during the time scale τ , and x fluctuates so rapidly that the
variable τ is effectively averaged during the time scale T . Evaluate

6
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

time derivatives:
dx ∂x dτ ∂x dT
ẋ = = + = ∂τ x + µ∂T x
dt |{z}
∂τ |{z}
dt |{z}
∂T |{z}
dt
∂τ x 1 ∂T x µ
2
ẍ = ∂τ ẋ + µ∂T ẋ = ∂τ x + 2µ∂τ ∂T x + O(µ2)
Make expansion in terms of small µ, x(τ, T ) = x0(τ, T ) + µx1(τ, T ) +
. . . , and insert this into the van der Pol equation (1)
0 = ẍ + µ(x2 − 1)ẋ + x
= ∂τ2x + 2µ∂τ ∂T x + µ(x2 − 1)[∂τ x + µ∂T x] + x + O(µ2)
= ∂τ2x + x + µ[2∂τ ∂T x + (x2 − 1)∂τ x] + O(µ2)
= ∂τ2[x0 + µx1] + [x0 + µx1] + µ[2∂τ ∂T x0 + ([x0]2 − 1)∂τ [x0]] + O(µ2)
= ∂τ2x0 + x0 + µ[∂τ2x1 + x1 + 2∂τ ∂T x0 + ([x0]2 − 1)∂τ [x0]] + O(µ2)

To order µ0 we have
ẍ0 + x0 = 0 ⇒ x0 = A(T ) sin τ + B(T ) cos τ .
Where A(T ) and B(T ) are T -dependent coefficients. The initial con-
dition for x0
0 = x(0) = x0(0, 0) +µ x1(0, 0) + . . .
| {z } | {z }
=0 =0
1 = ẋ(0) = ∂τ x0(0, 0) +µ[∂T x0(0, 0) + ∂τ x1(0, 0)] + . . . .
| {z } | {z }
=1 =0

gives A(T ) and B(T ) are any functions satisfying A(0) = 1, B(0) = 0.
To order µ1 we have
0 = ∂τ2x1 + x1 + 2∂τ ∂T [A(T ) sin τ + B(T ) cos τ ]
+ ([A(T ) sin τ + B(T ) cos τ ]2 − 1)∂τ [A(T ) sin τ + B(T ) cos τ ]]
= ∂τ2x1 + x1 + 2(A0(T ) cos τ − B 0(T ) sin τ )
+ ([A(T ) sin τ + B(T ) cos τ ] − 1)[A(T ) cos τ − B(T ) sin τ ] .

7
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

This equation can be solved for x1 (preferably Mathematica). The

solution contains a secular term on the form f1(T )τ sin τ +f2(T )τ cos τ
with
f1(T ) = −A0(T ) + A(T )(4 − A(T )2 − B(T )2)/8
f2(T ) = −B 0(T ) + B(T )(4 − A(T )2 − B(T )2)/8 .
We choose the coefficients A and B in a self-consistent manner as to
remove the secular divergence from the terms proportional to τ sin τ
and τ cos τ , i.e. we solve f1 = f2 = 0 for A(T ) and B(T ) with
A(0) = 1 and B(0) = 0 [Mathematica]:
2
A(T ) = √ , B(T ) = 0 . (2)
1 + 3e−T
In conclusion, the two-timing gives the solution
2
x(t) = A(T ) sin τ = √ sin t
1 + 3e−µt
To lowest order in µ the van der Pol oscillator approaches a circular
limit cycle with amplitude limT →∞ A(T ) = 2. The two-timing result
for x(t) agrees very well with the numerical solution, also for large
times:

Exact
Pert. theory
t = 1/µ
t

8
Dynamical systems 2019 kristian.gustafsson@physics.gu.se

8.3.3 Average over fast variable

If we are not interested in the fast dynamics, an easier method than
two-timing, frequently applied in mechanics, is to average over the fast
variable. Consider van der Pol’s equation (1) as a dynamical system
ẋ = y
ẏ = −x − µ(x2 − 1)y .
Change to polar coordinates
p
r= x2 + y 2 , φ = arctan(y/x)

to obtain (these expressions for ṙ and θ̇ were derived in Lecture 4)

xẋ + y ẏ
ṙ = = −µr sin2 φ(r2 cos2 φ − 1)
r
xẏ − y ẋ
φ̇ = 2
= −1 − µ(r2 cos2 φ − 1) cos φ sin φ .
r
When |µ|  1, r changes slowly (time scale ∼ 1/µ ) compared to φ
(time scale ∼ 1 ). Introduce a slow, smoothed variable R obtained
by filtering out the fast oscillations in φ. Its dynamics is given by
averaging the ṙ equation over the fast variable φ
1 2π 1 2π
Z Z
Ṙ = dφ ṙ|r→R = dφ[−µR sin2 φ(R2 cos2 φ − 1)]
2π 0 2π 0
µR
= − (R2 − 4) . (3)
8
When µ > 0 we have unstable fixed point at R = 0 and stable fixed
point at R = 2. Thus, to lowest order in µ the system has a stable
limit cycle of radius 2. An exact solution of Eq. (3) gives
2
R=q .
1+ (4R0−2 − 1)e−µt

This is the same result that was obtained for the amplitude Eq. (2)
with R0 = A(0) = 1.