Ruprecht-Karls University of Heidelberg

Faculty of Physics and Astronomy

Non-linear Dynamics
Prof. Ulrich Schwarz
Winter term 2013/2014
Last update: January 29, 2014
Contents
1 The central equation 1
2 Flow on a line 3
3 Bifurcations in 1d 8
3.1 Saddle-node bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Transcritical bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Pitchfork bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 Supercritical pitchfork bifurcation . . . . . . . . . . . . . . . . . . 14
3.3.2 Subcritical pitchfork bifurcation . . . . . . . . . . . . . . . . . . . 15
3.4 Inuence of high order terms . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.5 Summary of 1d bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 Flow on a circle 19
5 Flow in linear 2d systems 25
5.1 General remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Phase plane ow for linear systems . . . . . . . . . . . . . . . . . . . . . 27
6 Flow in non-linear 2d systems 30
7 Oscillations in 2d 39
8 Bifurcations in 2d 49
9 Excitable systems 55
10 Pattern formation in reaction-diusion systems 64
I
1 The central equation
1 The central equation
We consider dynamical systems of dimension d which are described by ODEs. This
implies that we use continuous time. (One alternative would be dierent equations with
discrete time.) Calling x the state vector of the system we consider the equation
dx
dt
=

f(x)
with a vector-valued function

f which can be non-linear. In case of a linear function

f
the equation simplies to

x = A x
with a matrix A and the system shows exponential behavior.
Examples
1. Harmonic oscillator

x +
2
0
x = 0
This looks like an one-dimensional system. But the following trick eliminates
the second derivative and shows the linear but two-dimensional character of the
harmonic oscillator:
Choose x
1
= x and x
2
= v = x with the velocity v. Then, the equation written in
the general form is

x =
_
x
1
x
2
_
=
_
0 1

2
0
0
_
x.
2. Overdamped particle x +k x = 0
is the viscosity of the surrounding medium.
Solving for x shows that the equation is al-
ready in the general form:
x =
k

x.
The system is one-dimensional and linear.
Because of this, no oscillations occur.
1
1 The central equation
3. Pendulum x +
g
l
sin(x) = 0
Using the same trick as in example 1, we
get x
1
= x
2
and x
2
=
g
l
sin(x
1
), hence a
non-linear, two-dimensional system. Only
for small angles x (sin(x) x) we end up
with a harmonic oscillator.
4. Driven harmonic oscillator m x +k x = F cos(t)
The equation of the driven harmonic oscillator explicitly depends on time t. But
rewriting the equation using x
1
= x, x
2
= x and introducing a third variable
x
3
= t leads to the relations x
1
= x
2
, x
2
=
1
m
(kx
1
+F cos(x
3
)) , x
3
= 1.
The system is non-linear with d = 3.
5. Electric curcuit R I +
Q
C
= V
0
Using Kirchhos law and the relation I =

Q we get the ODE of an overdamped
particle:

Q =
V
0
R

1
RC
Q.
Remember this is a linear, one-dimensional system. If we put in a solenoid the
dependence of

Q will lead to d = 2 and oscillations can occur.
2
2 Flow on a line
2 Flow on a line
In this chapter, we are looking at one-dimensional systems. Therefore, the central equa-
tion becomes x = f(x) with an arbitrary function f.
The rst example we want to discuss is non-linear: x = sin(x). The separation of
variables leads to
dx
sin(x)
= csc(x)dx = dt
which can be integrated with the result
t = ln
_
csc(x
0
) + cot(x
0
)
csc(x) + cot(x)
_
.
Even in this simple non-linear example, the behavior of the system is not easy to under-
stand from this solution. But graphical analysis shows the most important properties.
Plotting a phase portrait (left gure), stable and unstable xed points can be deter-
mined. In 1d, the systems dynamics corresponds to ow on the line. The corresponding
trajectories are shown in the right gure.
For a stable xed point a little change in x drives the system back, whereas for an
unstable xed point it causes a ow away from the xed point.
Choosing dierent starting points x

the time-dependence of the acceleration computes

as follows: for starting points [x

[

2
the acceleration directly decreases. But if
x

= x with

2
< x <= the acceleration rst increases and decreases after the
deection point.
3
2 Flow on a line
The graphical analysis can be performed for the earlier examples as well:
Overdamped particle
x = x
x

= 0 , stable xed point

Electrical curcuit
x

,= 0
The applied method works for any graph.
4
2 Flow on a line
In an one-dimensional system, there are three possibilities in total the system can behave:
1. staying at a xed point
2. owing to a stable xed point
3. owing to innity
There is also a mathematical method to analyze xed points. It is called linear stability
analysis. Firstly, one determines the xed points by solving x = f(x) = 0 for x. Take x

to be a xed point. Then, the deviation from this xed point is given by = x x

.
The derivative can be written in dependence of the sum x

+.
= x = f(x) = f(x

+)
The rst order Taylor expansion = f(x

)
=0
+f

(x

) +O(
2
) leads to a rst order ODE
= f

(x

)
which can be integrated to a time-dependent deviation
(t) =
0
exp(f

(x

) t)
with the starting deviation
0
at t = 0. Introducing the relaxation time t
0
= [
1
f

(x

)
[ this
yields
(t) =
0
exp (sign(f

(x

)) t/t
0
) .
Thus, f

(x

) hints towards the characteristics of a xed point x

.
Conclusion
If f

(x

) < 0: stable xed point, exponential decay

f

(x

) > 0: unstable xed point, blow-up

f

(x

) = 0: further investigations are needed

5
2 Flow on a line
Examples
In both cases x = x
3
the character of the xed point is not clear from f

.
For
x = x
2
the system has a half-stable point at x = 0. If x is constant, the result is a line
of xed points.
Uniqueness theorem
If f(x) and f

(x) are continuous on an open interval around x

0
, then a solution exists
and is unique.
This leads to the impossibility of oscillations in an one-dimensional system. Instead,
everything is overdamped. Consider a potential V . In 1d, we can write
x = f(x) =
dV (x)
dx
.
6
2 Flow on a line
Hence, the time derivative of V yields
dV
dt
=
dV
dx
dx
dt
=
_
dV
dx
_
2
0.
So, the energy of the system can never increase. It always decreases during ow.
Examples
1. Overdamped particle
2. Mexican hat
The Mexican hat is an example of a bistable system.
The gures show the following relation for d = 1:
minimum in V : stable xed point
maximum in V : unstable xed point
7
3 Bifurcations in 1d
3 Bifurcations in 1d
As we have seen in chapter 2, ow can easily be understood in d = 1. However, up to now
we did not consider any parameter which in principal could change the ow structure.
A sudden change in the character of the solution is called bifurcation. Physical exam-
ples for this are phase transitions, mechanical instabilities, laser thresholds, population
thresholds etc.
There are exactly three types of bifurcations in d = 1. Mathematically it can be shown
that each type can be described by one general form using the bifurcation parameter r.
The general properties are summarized in the following table.
Saddle-node bifurcation x = r +x
2
xed points can appear or disappear
depending on r
Transcritical bifurcation x = rx x
2
xed points always exist for all r
but they can exchange stability
Pitchfork bifurcation x = rx x
3
xed points appear or disappear
as a symmetrical pair
3.1 Saddle-node bifurcation
Consider x = r+x
2
with the bifurcation parameter r. The roots are given by x

r.
For various r the system behaves dierently.
8
3 Bifurcations in 1d
This allows to analyze the inuence of r on the ow behavior in a bifurcation diagram.
Since the branches appear suddenly for x

0 the saddle-node bifurcation is also called

out of the blue sky bifurcation.
This corresponds to a phase transition of second order like the magnetization of an Ising
magnet.
9
3 Bifurcations in 1d
But why does x = r + x
2
describe all saddle-node bifurcations? We assume x to be a
function of x and the parameter r.
Expanding x = f(x, r) around x = x

and r = r
c
leads to
x f(x

, r
c
) +
f
x
[
(x

,r
c
)
(x x

) +
f
r
[
(x

,r
c
)
(r r
c
) +
1
2

2
f
x
2
[
(x

,r
c
)
(x x

)
2
.
Considering f(x

, r
c
) = 0 and
f
x
[
(x

,r

)
= 0 the general form computes as
x = a(r r
c
) +b(x x

)
2
.
Example: Stability of adhesion cluster under constant force
We consider an adhesion cluster with bonds (receptor-ligand pairs) that can be either
open or closed. If N(t) represents the time-dependent number of closed bonds and N
t
10
3 Bifurcations in 1d
the total number of bonds, the number of open bonds is given as N
t
N. Two rates
k
o
, k
on
are used to describe the systems dynamics. In contrast to k
on
, k
o
depends on the
acting force F. It can be written as k
o
= k
0
exp(
F
F
0
N
) = k
0
exp(
f
N
). Furthermore
using =
k
on
k
0
and the dimensionless force f =
F
F
0
, the time-dependent number of bonds
is given by
dN
dt
= k
on
(N
t
N) k
o
N
= k
on
(N
t
N) k
0
exp(
f
N
) N.

N =
dN
d
= N exp(
f
N
) + (N
t
N)
in dimensionless time = k
0
t. The graphical analysis shows the dierent cases for
f < f
c
and f > f
c
. The xed points are given by N exp(
f
N
) = (N
t
N).
Below, the bifurcation point f
c
is calculated using two equations.
N
c
exp(
f
c
N
c
) = (N
t
N
c
) (3.1)
N
c
exp(
f
c
N
c
)
_
1
f
c
N
c
_
= N
c
(3.2)
=
f
c
N
c
exp(
f
c
N
c
) (3.3)
11
3 Bifurcations in 1d
Equation (1) represents the xed points. Dividing (2) by (1) results in
_
1
f
c
N
c
_
=
N
c
N
t
N
c
f
c
=
N
t
N
c
N
t
N
c

f
c
N
c
=
f
c
N
t
+ 1
Using this, can be written as
=
f
c
N
t
exp(
f
c
N
t
+ 1)
f
c
N
t
exp(
f
c
N
t
) =

e
f
c
= N
t
plog
_

e
_
. In the last step, the function plog is used, dened by x exp(x) = Q x = plog(Q).
3.2 Transcritical bifurcation
The general form of a transcritical bifurcation x = r x x
2
leads to the xed point
x

= 0 which exists for an arbitrary bifurcation parameter r and also represents the
bifurcation point r
c
= 0. The second xed point x

= r is stable for r > 0 and unstable

for r < 0. Depending on the application, not every xed point is reasonable, e.g. when
the population size is always positive.
A good example for a transcritical bifurcation is a laser. The rate of the photons in
the laser n(t) is determined by the dierence between gain and loss. Since the photons
stimulate the atoms, the gain is proportional to both the number of photons in the laser
n(t) and the number of excited atoms N(t). The gain coecient is positive: G > 0. The
loss of photons is determined by the rate constant k > 0.
n(t) = gain loss
= G n N k n
Introducing the maximal possible number of excited atoms N
0
, the number of excited
atoms can be written as N(t) = N
0
n(t). Hence, the rate n computes as
n(t) = (GN
0
k) n G n
2
.
Thus, the xed points are n

1
= 0 and n

2
=
GN
0
k
G
. Since n describes a particle number,
12
3 Bifurcations in 1d
demanding n

2
> 0 is reasonable and leads to N
0
>
k
G
.
The xed points stability analysis is done by evaluating
g

N
(n

) =
d n
dn
(n

) = (GN
0
k) 2Gn

.
g

n
(n

1
) = GN
0
k =
_
_
_
< 0, for N
0
<
k
G
, stable xed point
> 0, for N
0
>
k
G
, unstable xed point
g

n
(n

2
) = k GN
0
< 0, since N
0
>
k
G
, stable xed point
Obviously, N
0
=
k
G
is a bifurcation point.
13
3 Bifurcations in 1d
3.3 Pitchfork bifurcation
As the general form reads x = rx x
3
two dierent types exist: the supercritical and
the subcritical pitchfork bifurcation.
3.3.1 Supercritical pitchfork bifurcation
Considering rstly f(x, r) = x = rx x
3
, the xed points compute as x

1
= 0 and
x

2/3
=

r if r > 0. The stability analysis yields

f

x
(x, r) = r 3x
2
.
f

x
(x

1
) = r =
_
_
_
< 0, for r < 0, stable xed point
> 0, for r > 0, unstable xed point
f

x
(x

2/3
) = 2r, stable xed point since x

2/3
only exist for r > 0
14
3 Bifurcations in 1d
3.3.2 Subcritical pitchfork bifurcation
Now, f(x, r) = x = rx + x
3
. The xed points are similar: x

1
= 0 and x

2/3
=

r if
r < 0. In this case, the properties are:
f

x
(x, r) = r + 3x
2
.
f

x
(x

1
) = r =
_
_
_
< 0, for r < 0, stable xed point
> 0, for r > 0, unstable xed point
f

x
(x

2/3
) = 2r, unstable xed point since x

2/3
only exist for r < 0
15
3 Bifurcations in 1d
So, the two types of pitchfork bifurcation dier in the second and third xed point which
are symmetrical in both cases. By now, the system is unstable. But it can be stabilized
by using high order terms.
3.4 Inuence of high order terms
In order to stabilize pitchfork bifurcations, a fth order term is used. For instance,
consider the following equation:
x = rx +x
3
x
5
.
The xed points are
x

1
= 0
x

2/3
=

1 +

1 + 4r
2
, exists for 1 + 4r > 0
x

4/5
=

1 + 4r
2
, exists for 1 < 4r < 0
The existing xed points in dependence of the bifurcation parameter r are shown in the
next gure.
The bifurcation diagram is particularly interesting because there are dierent bifurcation
types visible. The surrounding of r = 0 is characterized by a subcritical pitchfork
bifurcation, whereas the transition of the unstable to the stable branch at r

c
describes
a saddle-node bifurcation.
16
3 Bifurcations in 1d
As well, a hysteresis eect is possible in this conguration. Starting at r > r

c
, x = 0
and increasing r up to r > 0 leads to an unstable condition. A small perturbation results
in a transition to a stable branch at same r. Decreasing r again, the system remains on
the stable branch. This shows that the system does not come back to the original xed
point.
3.5 Summary of 1d bifurcations
The general form is f(x, r) = x and behaves as shown in the following table:
normal form f(x

) f

x
(x

, r
c
) f

r
(x

, r
c
) f

xx
(x

, r
c
) f

xr
(x

, r
c
) f

xxx
(x

, r
c
)
saddle-node x = r +x
2
0 0 ,= 0 ,= 0
bifurcation
transcritical x = rx x
2
0 0 0 ,= 0 ,= 0
bifurcation
pitchfork x = rx x
3
0 0 0 0 0 ,= 0
bifurcation
17
3 Bifurcations in 1d
Example: Overdamped bead on a rotating hoop
The hoop is rotating around the axis with an angular velocity . The acting forces
are the gravitational force F
G
, the centrifugal force F
C
and a friction force F
R
which
describes the system in a uid. They are projected on the -plane.
F
G
= m g mg sin()
F
C
= m
2
m
2
cos()
F
R
= b

Using = r sin(), the total acting force yields

F = m r

= b

mg sin() +mr
2
sin() cos().
In order to receive a rst order equation, the term m r

shall be neglected. The time

=
t
T
with the timescale T is introduced. In a second step, the equation is reformulated
dimensionless by dividing by the gravitational force F
G
.
_
m
T
2
_
2
d
2

d
2
=
b
T
d
d
mg sin() +mr
2
sin() cos()
_

gT
2
_
2
d
2

d
2
=
b
Tmg
d
d
sin() +
r
2
g
sin() cos()
How to dene T so that
2
:=
_

gT
2
_
2
is negligible?
18
4 Flow on a circle
Choosing the prefactor of the friction force
b
Tmg
to be of order 1, T is dened to be
T =
b
mg
. With this, is negligible if =

gT
2
=
m
2
g
g
2
1, so if the inertia is much
smaller than the friction.
Set = 0 from now on and introduce :=

2
g
. The equation computes as
d
d
= sin() ( cos() 1) .
The applied procedure has reduced the number of parameters from ve to two: and
. But setting = 0 transforms the equation to dimension one. So, only one initial
condition can be considered. Thus, the behavior of the system at the very beginning
is neglected. After that, the system behaves as if it was of the order of 1.
The xed points result from sin() = 0 and cos()
1

= 0. Therefore, the number of

xed points depends on :
For [[ > 1 there are only the xed points due to sin() = 0. For the bifurcation point
[[ = 1, one additional xed point exists for each period of and for [[ < 1, there are
actually two.
4 Flow on a circle
So far, we considered linear systems x = f(x) and visualized their dynamics as ow on
a line. Now, we are taking into account periodic behavior using the dierential equation

= f(). Then, f( + 2) = f(). This corresponds to a vector eld on the circle.

The simplest case is a constant velocity

= f() = const = leading to an oscillation
19
4 Flow on a circle
with period T =
2

but without amplitude, (t) = t +

0
.
Ths simplest non-trivial case is

= a sin(). It has various applications in dif-
ferent branches of science: e.g. Josephson junctions, electronics, biological oscillations,
mechanics, etc.
a is a bifurcation parameter (a > 0). Depending on its relation to , the following phase
portraits (lhs) and corresponding ow diagrams (rhs) exist.
20
4 Flow on a circle
What is the oscillation period T?
T =
_
dt =
_
2
0
dt
d
d =
_
2
0
1
d
dt
d =
_
2
0
1
a sin()
d =
2

2
a
2
Obviously, the oscillation period depends on a.
a = 0 T =
2

a T =
2
_
( a)( +a)
=
2

a
The scaling is generic for a saddle-node bifurcation.
21
4 Flow on a circle
A Taylor expansion around the critical value

=

2
by introducing =

= a sin
_
+

2
_
= a cos()
a +
1
2
a
2
results in the normal form of saddle-node bifurcations:
x :=
_
a
2
_
1/2

r := a

_
2
a
_
1/2
x = r +x
2
.
Most of the time will be spent close to

.
T =
_
dt =
_

dt
dx
dx =
_
2
a
_
1/2
_
dr
1
r +x
2
=
_
2
a
_
1/2

r
=
_
2

_
1/2

a
This is the same result as above. Therefore, extending the integration boundaries to
innity is indeed not a problem.
22
4 Flow on a circle
Examples
1. Driven overdamped pendulum b

+mgLsin() =
Dividing by mgL leads to the dimensionless
equation
b
mgL
. .
=
0

+ sin() =

mgL
=: .
The bifurcation parameter is the quotient of the applied torque to the maxi-
mum gravitational torque. Having also introduced the dimensionless time
0
=
t

the system is described by the equation

=
d
d
= sin().
At = 1 the pendulum stops its motion. For < 1 the external torque is too
weak to drive the pendulum around.
2. Firey synchronization
Identify = 0 with the emission of the ash.
Without external stimulus, each rey has

= . Now, consider a periodic stimulus

with phase which satises

= . (4.1)
The basic model to simulate the reys re-
action to the stimulus is given by

= +A sin( ) (4.2)
with the resetting or coupling strength A.
23
4 Flow on a circle
The last equation can be expressed using the phase dierence = . Sub-
stracting (4.2) from (4.1) leads to = A sin(). Dening furthermore
:=

A
and = A t the nal equation reads

=
d
d
= sin()
.
This leads to the following phase space diagrams:
For = 0, there is a perfect synchrony at = 0. If > 1 (or < 1) there
arent any stable xed points. So, there is no synchrony but a phase drift. The
oscillation occurs with T =
2

()
2
A
2
.
The interval 1 < < 1 is called the
range of entrainment. There is syn-
chrony at the stable xed point but
with a phaselag. The stimulus entrains
the oscillation with a frequency if
A < < + A. This is called
phase locking.
In our example, now all reies ash in
synchrony, but with a possible lag to the
external stimulus (e.g. a ash light).
24
5 Flow in linear 2d systems
5 Flow in linear 2d systems
5.1 General remarks
In 2d, the varity of dynamical behavior is much larger than in 1d.
As a rst step, we look at linear systems in two dimensions. Then, a complete classi-
cation is possible and starts from
x = A x =
_
a b
c d
__
x
1
x
2
_
.
In general, if x
1
(t) and x
2
(t) are solutions of the equation, so is c
1
x
1
+c
2
x
2
.
In addition, x = 0 is always a solution.
Graphical analysis can be done by drawing and analyzing the phase plane (x
1
, x
2
).
Examples
1. Harmonic oscillator: m x +kx = 0
Dening the frequency
0
=
_
k
m
and choosing x
1
= x, x
2
= x = v as done in
section 1 the matrix is A =
_
0 1

2
0
0
_
.
Why are the trajectories ellipses?
x
v
=
v

2
0
x

2
0
x dx = v dv

2
0
x
2
v
2
= const
This corresponds to energy conservation:
1
2
kx
2
+
1
2
mv
2
= const.
25
5 Flow in linear 2d systems
2. A linear 2d system without oscillation:
_
x
y
_
=
_
a 0
0 1
__
x
y
_
For the given matrix A =
_
a 0
0 1
_
, the two equations are uncoupled. They can
be separately solved using an exponential ansatz.
x = a x x(t) = x
0
exp(a t)
y = y y(t) = y
0
exp(t)
The phase portraits dier depending on a.
26
5 Flow in linear 2d systems
5.2 Phase plane ow for linear systems
Consider the linear case

x = A x =
_
a b
c d
_
x.
Our analysis is based on the two eigenvalues
1
,
2
of A =
_
a b
c d
_
. They are calculated
using the characteristic equation A v = v.
0
!
= det
_
a b
c d
_
=
2

= (
1
)(
2
)

1/2
=

2
4
2
with = a +d = tr A =
1
+
2
and = ad cb = det A =
1

2
.
In general, the eigenvalues are complex numbers depending on the trace and the
determinant of A. If the eigenvalues are dierent
1
,=
2
, the eigenvectors v
1/2
are
linearly independent. Hence, the characteristics of A determines the phase plane ow
as shown below. The time-dependent eigenvectors can be calculated using x
1/2
(t) =
exp(
1/2
t) v
1/2
.
We now completely enumerate all possible cases.
1. Real eigenvalues
27
5 Flow in linear 2d systems
2. Complex eigenvalues
If the discriminant is smaller than zero,
2
4 < 0, the eigenvalues are complex.
Dening =
1
2
_
(
2
4 ), the eigenvalues read
1/2
=

2
i. The general
solution can be decomposed in the directions of the eigenvalues.
x(t) = (c
1
v
1
exp(it) +c
2
v
2
exp(it)) exp(t)
Now, there are oscillations in the system. If < 0 the amplitude is decaying. In
contrast, if > 0 it is exploding. Only if = 0 the amplitude is constant. In this
case, the eigenvalues are purely imaginary. It is the boundary between stability
and instability.
28
5 Flow in linear 2d systems
3. Equal eigenvalues
The eigenvalues are equal if the discriminant is zero
2
4 = 0. Then, the
eigenvectors exhibit the same velocity. They can either be dierent or the same.
In both cases, stable and unstable phase portraits exist. As an example, the stable
ones are plotted.

1
=
2
, v
1
= v
2
, degenerate node
4. At least one eigenvalue is zero
In this case, the phase portrait is a line or plane of xed points.
29
6 Flow in non-linear 2d systems
Summary in one scheme
Saddles, nodes and spirals are the major types of xed points.
6 Flow in non-linear 2d systems
Non-linear systems show a much larger variety of ow behavior.

x =
_
f
1
(x)
f
2
(x)
_
Example
30
6 Flow in non-linear 2d systems
Recipe for phase space analysis:
1. Identify nullclines: lines with x = 0 or y = 0
2. Identify xed points: intersections of nullclines
3. Linear stability analysis around xed points
Example x = x + exp(y), y = y
The phase portrait shows four dierent regions varying in the sign of x and y. At (-1,0),
there is a saddle. Having drawn the nullclines it is easy to compute the remaining ow
behavior.
Linear stability analysis:
At a xed point (x

, y

), we look at small deviations u = x x

and v = y y

using
Cartesian coordinates. The derivatives are approximated in a Taylor expansion up to
rst order.
u = x = f
1
(x

+u, y

+v)
= f
1
(x

, y

) +
f
1
x
u +
f
1
y
v +O(u
2
, v
2
, . . . )
v = y = f
2
(x

+u, y

+v)
= f
2
(x

, y

) +
f
2
x
u +
f
2
y
v +O(u
2
, v
2
, . . . )
31
6 Flow in non-linear 2d systems
Approximated up to rst order, this can also be written as matrix equation.
_
u
v
_
=
_

x
f
1

y
f
1

x
f
2

y
f
2
_
. .
=A
_
u
v
_
A is the Jacobian at the xed point. Calculating the eigenvalues of A, linear stability
analysis can be performed. This works well for saddles, nodes and spirals. But it does
not always work for borderline cases such as centers, stars, lines and planes of xed
points or degenerate nodes as shown in the following example.
Example: Pathological example
x = y +ax(x
2
+y
2
)
y = x +ay(x
2
+y
2
)
Obviously, (x

, y

) = (0, 0) is a xed point. In order to calculate the Jacobian, only the

linear terms have to be considered.
A =
_
0 1
1 0
_
= 0; = 1;
1/2
= i
In linear approximation, the xed point is a center.
But this is not true for the non-linear case. To analyze this in more detail, we switch to
polar coordinates x = r cos , y = r sin and derive the equation of motion for r
r
2
= x
2
+y
2
r r = x x +y y = x(y +axr
2
) +y(x +ayr
2
) = a r
4
r = a r
3
and :
= arctan
_
y
x
_

=
1
1 +
_
y
x
_
2

x y xy
x
2
= = 1.
Hence, the angular velocity

is constant. a is the important parameter. The situation
is similar to ow on a circle, yet the radius as dynamic variable can either explode or
decay.
32
6 Flow in non-linear 2d systems
a > 0 a < 0 a = 0
A center occurs only for a = 0. But linear stability analysis predicted this for all values
of a. Instead, the typical case is a spiral.
The following biological model shows the power of phase space analysis.
Example: Phase space analysis of a biophysical model
Consider the Lotka-Volterra model for two competing species x, y. The variables x and
y name the population size of rabbits and sheep, respectively. The rabbit population
exhibits a faster logistic growth than the sheep population. As the sheep compete for
grass with the rabbits, the growth rate of the rabbits x decreases if more sheep exist
while the sheep suer only little under more rabbits.
x = x(3 x 2y)
y = y(2 y x)
33
6 Flow in non-linear 2d systems
The Jacobian computes as A =
_
3 2x 2y 2x
y 2 2y x
_
. The character of the four
xed points is dierent.
xed point (0, 0) (0, 2) (3, 0) (1, 1)
A
_
3 0
0 2
_ _
1 0
2 2
_ _
3 6
0 1
_ _
1 2
1 1
_
5 -3 -4 -2
6 2 4 -1

1
3 -1 -2

2 1

2
2 -2 -2 (

2 1)
classication unstable node stable node stable node saddle
Furthermore, we have a brief look at two dierent types of special situations: con-
servative (e.g. earth orbiting around the sun) and reversible systems (systems with
time-reversal).
1. Conservative systems
In a conservative system, the acting force F can be derived for a potential V . For
34
6 Flow in non-linear 2d systems
example, in 1d we have:
m x = F(x) =
dV
dx
.
Multiplying with the velocity x leads to
m
2
d
dt
( x
2
) =
dV
dt

d
dt
(
1
2
m x
2
+V
. .
=E
) = 0.
There exists a quantity E which is constant along trajectories (but not in an open
set in x). This corresponds to energy conservation.
Theorem: In a conservative system, an attractive xed point cannot exist.
Proof: In such a case, there would be a bassin of attraction and thus E could not
be constant in a nontrivial way.
Example: Mexican hat V (x) =
1
2
x
2
+
1
4
x
4
The second derivative is x = x x
3
. Three xed points exist: (0,0), (1,0) and
(-1,0). Applying a simple trick x = y and y = x x
3
the phase portrait can be
drawn.
35
6 Flow in non-linear 2d systems
Example: Hamiltonian system H(q, p)
It is q =
H
p
and p =
H
q
. From this, energy conservation simply follows:

H =
p
H p +
q
H q = 0.
We know from the theorem that there are no attractive xed points in the system.
Instead, a typical xed point is a center and thus often oscillations occur in the
system.
2. Reversible systems
Time reversal symmetry is more general than energy conservation. Reversible,
non-conservative systems occur e.g. uid ow, laser, superconductors, etc.
Mechanical systems are invariant under t t. Consider in 1d, m x = F. We
introduce the velocity
v = x v =
1
m
F(x).
Both (x(t), v(t)) and (x(t), v(t)) are
solutions of the system. In general,
there is a twin for each trajectory. Note
the similarity to centers, which have
trajectories that have merged at the
ends.
Examples:
a) x = y y
3
y = x y
2
The system is invariant under t t
and y y. There are three xed
points: two saddles and a center.
There is mirror symmetry around the
x-axis in regard to the ow lines (but
not the ow distribution).
36
6 Flow in non-linear 2d systems
b) x = 2 cos(x) cos(y) y = 2 cos(y) cos(x)
The system is invariant under t t,
x x and y y.
Four xed points exist (x

, y

) =
_

2
,

2
_
: two saddles, one stable node
and one unstable node. Since the stable
node is an attractive xed point, this is
not a conservative system.
Now, there is mirror symmetry around
the bisector.
Numerical integration of ODEs
Several numerical integration methods for ODEs exist, diering in their accuracy. They
are based on a Taylor expansion up to a certain order. In the following, we have a closer
look at three dierent methods.
1. Euler method:
The time t is discretized. Starting at t
n
, the next step is computed multiplying
the velocity at t
n
with the time step t. This corresponds to a rst order Taylor
expansion
x
n+1
= x
n
+f(x
n
)t +O(t
2
).
Since this accuracy is not very good, higher order methods are often applied.
2. Runge-Kutta methods:
The Runge-Kutta methods combine several Euler-style steps. A second order
accuracy is achieved by using the mid-point velocity of the integration step.
k
1
= f(x
n
)t, k
2
= f
_
x
n
+
k
1
2
_
t
x
n+1
= x
n
+k
2
+O(t
3
)
We see that two function evaluations are needed. In an analogous manner, we
37
6 Flow in non-linear 2d systems
maintain fourth order accuracy:
k
1
= f(x
n
)t, k
2
= f
_
x
n
+
k
1
2
_
t,
k
3
= f
_
x
n
+
k
2
2
_
t, k
4
= f
_
x
n
+
k
3
2
_
t,
x
n+1
= x
n
+
1
6
(k
1
+ 2k
2
+ 2k
3
+k
4
) +O(t
5
)
Obviously, this requires four function evaluations. It is the a very good choice if
the number of evaluations is not essential. To realize this, one standard choice is
Matlab ODE 45 (see also on the web page).
3. Stormer-Verlet methods: (leaping frog)
St ormer-Verlet methods are especially suited for Hamiltonian systems, e.g. molec-
ular dynamics. The simplest version is:
x = f(x) f(x
n
) =
x
n+1
+x
n1
2x
n
t
2
x
n+1
= 2x
n
x
n1
+f(x
n
)t
2
We now rewrite this as 2d system. We dene the velocity v = x and discretize the
function f(x) = v. For each integration step, the position is updated for a free
time step and the velocity half a time step. The resulting equations are:
v
n+1/2
= v
1
+
t
2
f(x
n
)
x
n+1
= x
n
+v
n+1/2
t
v
n+1
= v
n+1/2
+
t
2
f(x
n+1
)
This procedure is called leaping frog because we have staggered jumps.
38
7 Oscil lations in 2d
7 Oscillations in 2d
In contrast to linear systems, non-linear ones allow for limit cycles. These are isolated
closed trajectories in the phase plane.
Poincare-Bendixson-Theorem:
If R is a closed, bounded subset of the plane without any xed point and if there is a
trajectory that is conned in R, then R contains a closed orbit.
The second condition is satised if a trapping region R exists. To prove that a stable
limit cycle exists, we have to show that a trapping region exists without a xed point
inside.
Examples:
1. r = r(1 r
2
),

= 1
In Cartesian coordinates:
x = (1 x
2
y
2
)x y
y = (1 x
2
y
2
)y x.
2. r = r(1 r
2
) +r cos(),

= 1
In this more complicated example, for which values of 0 does a stable limit
cycle exist?
The trapping region can be determined for the given example by constructing mini-
mum and maximum radius r
min
, r
max
and demanding an increasing and decreasing
ow, respectively.
r
min
: r r(1 r
2
) r > 0 r
min
<

1
r
max
: r r(1 r
2
) +r < 0 r
max
>

1 +
39
7 Oscil lations in 2d
Since

1 + is always real for 0, the only restriction for the trapping region
comes from r
min
: 0 < 1. Due to the Poincare-Bendixson theorem, we have a
stable limit cycle for these values of .
3. Biological example:
The substrate-depletion oscillator is one of the rst discovered biochemical oscil-
lators (1964). E.g. it describes the sugar metabolism. In 1968, Selkov was rst to
model the basic understanding of glycolysis.
Sugar metabolism is extremly ecient.
C
6
H
12
O
6
. .
glycose
+ 6 O
2
6 CO
2
+ 6 H
2
O + E
The produced energy is in the form of 36 ATP molecules. Having a closer look at
the glycolysis, we see that one part of the system can oscillate:
glycose
ATPADP
glycose 6-P fructose 6-P
. .
=F6P
ATPADP

()
fructose 1,6-P
. .
=FBP
Therefore, though somehow counterintuitive, the rst steps of the glycolysis path-
way use up ATP rather than producing it. For (*), the enzyme PFK functions as a
catalyst. It helps to produce ADP, but it is also activated by ADP. We call this an
autocatalysis or a positive feedback loop. The biological function of the rst steps of
the glycolysis pathway is that ADP switches on the pathway of ATP production.
This ATP is produced on demand. Schematically, we can group
y =
_
_
_
ATP
F6P
and x =
_
_
_
ADP
FBP
since those substances play the same role. Introducing the reaction rates a, b, we
get:
glucose
b
y

a
xproducts.
x is produced with a constant rate a from y and it reacts to further products. The
production rate of x increases with its amount. In this model, the equations are:
x = x +ay +x
2
y
y = b ay x
2
y.
40
7 Oscil lations in 2d
The nullclines are y =
x
a+x
2
and
y =
b
a+x
2
. So, there is a xed point
(x

, y

) = (b,
b
a+b
2
). In order to show
that a trapping region exists, we con-
sider the region bounded by the green
lines. We know for the left part x = 0
and 0 y
b
a
. So we get 0 x b
and 0 y b. The ow goes in-
side. For the right part we calculate
x ( y) = x + y = b x < 0 since
x > b. This yields y > x. Therefore,
the ow is more negative than 1 and
goes inside.
Now, we have found a trapping region. The Poincare-Bendixson theorem demands
that no xed points exist. Hence, we make a hole around the xed point and show
that no trajectory goes into the hole. This is equivialent to a repulsive (unstable)
xed point. The Jacobian is
A =
_
1 + 2xy a +x
2
2xy (a +x
2
)
_
.
We need a positive trace and determinant of the matrix A.
= a +b
2
> 0
=
b
4
+ (2a 1)b
2
+ (a +a
2
)
a +b
2
!
0
Thus, the boundary between stable and unstable xed points = 0 is given by
b
2
=
1
2
(1 2a)

1 8a. This can be represented via a state-diagram.

The arrow indicates an increasing b. If
a is small enough, the system performs
oscillations in a certain interval of b. We
call this a re-entrance process.
41
7 Oscil lations in 2d
Lienard systems / van der Pol oscillator
The following structure often occurs in mechanics and electronics:
x
..
inertia
+f(x) x
. .
damping
+ g(x)
. .
restoring force
= 0.
It is called Lienard-system. Note, that this is the equation of the harmonic oscillator for
f = 0 and g = x.
We consider the most famous example of Lienard systems, the van der Pol oscil lator:
f = (x
2
1)
g = x.
For x 1 we have negative damping. The system can be driven by putting energy into
the system. For large x, the damping is positive. Energy is dissipated.
Example: Tetrode circuit (electronics)
The system is described as follows:
L

I +V +F(I) = 0
L

I +

V +F

(I)

I = 0
LC

I +I +CF

(I)

I = 0.
This is a van der Pol oscillator with f(I) = CF

(I) and g = 1.
Lienard systems are very widespread: e.g.
1. neural activity, action potential
2. biological oscillators (ear, circadian rhythms)
3. stick-slip oscillations in sliding friction
42
7 Oscil lations in 2d
Lienard theorem:
A Lienard system has a stable limit cycle around the origin at the phase plane if
1. g(x) = g(x), g(x) > 0 for x > 0
2. f(x) = f(x), F(x) =
_
x
0
f(x

)dx

has to have a zero at a > 0

and F(x) < 0 for 0 < x < a, F(x) > 0 for x > a, F() = .
Obviously, the rst condition is fulllled for the van der Pol oscillator. Consider the
second condition:
F(x) =
_
x
3
3
x
_
=
x
3
_
x
2
3
_
a =

3.
As for the harmonic oscillator the deection behaves sine-shaped, the deection of the
van der Pol oscillator follows a sawtooth. The phase portrait shows a deformed circle.
Now, we analyze the van der Pol oscillator in two limits: 1 and 1.
1. 1: Lienard phase plane analysis
The dynamic equation is
x = x + x(x
2
1) =
d
dt
( x +F(x)) =:
d
dt
w(x).
x = w F(x)
In the last step, we are using
dF(x)
dt
= f(x) x. Dene
y :=
w

y =
x

, x = (y F(x)).
Hence, the nullclines are x = 0 and y = F(x).
We assume the initial condition y F(x) O(1). Then, the velocity is very fast
in x-direction but very slow in y-direction:
x O()
y O(1/)
43
7 Oscil lations in 2d
The rst part shows a fast movement to
the right which stops at the x-nullcline.
In the second part, y F(x) O(1/
2
)
can be arbitrarily small. This leads to
x O(1/) and y O(1/). Thus, we
have a slow movement along the null-
cline. The third and fourth part are like
the rst and second, respectively, only
with changing signs in the velocities.
What is the oscillation period T? We neglect the fast paths. Hence, we have to
calculate
T = 2(T(x = 2 x = 1)) = 2
_
t
1
t
2
dt.
We know
dy
dt
= F

(x)
dx
dt
= (x
2
1)
dx
dt
=
x

and from this

dt = dx
(x
2
1)
x
.
Therefore, the oscillation period is T O():
T = 2
_
1
2
dx
(x
2
1)
x
= 2
_
x
2
2
ln(x)
_
2
1
= (3 2 ln(2))
T O().
2. 1:
This case is a small perturbation to the harmonic oscillator
x +x + h(x, x) = 0.
The systems dynamics depend on h(x, x).
a) For h = (x
2
1) x, the system is a van der Pol oscillator.
b) For h = 2 x, we have a weakly damped harmonic oscillator. The system is
linear.
44
7 Oscil lations in 2d
c) For h = x
3
, the system corresponds to an unharmonic spring with a spring
constant k = 1 + x
2
that increases by extending the spring. This is called
strain sting. The system is a Dung oscil lator.
For simplicity, we consider the second case h = 2 x.
For the initial conditions
x(0) = 0, x = 1,
the exact solution holds
x(t) = (1
2
)
1/2
exp(t) sin((1
2
)
1/2
t).
If we expand for 1, we nd
x(t) = (1 t) sin(t) +O(
2
).
But, this is only valid for t <
1

and blows up for large times. Thus, the small

epsilon limit appears to be problematic.
We now try to solve the problem using regular perturbation theory. Plugging
x(t) = x
0
(t) +x
1
(t) +. . .
in the dynamic equation yields
d
2
dt
2
(x
0
+x
1
+. . . ) + 2
d
dt
(x
0
+x
1
+. . . ) + (x
0
+x
1
+. . . ) = 0.
Compare the parameters for dierent orders of .
O(1) : x
0
+x
0
= 0
x
0
= sin(t)
O() : x
1
+ 2 x
0
+x
1
= 0
x
1
+x
1
= 2 cos(t)
x
1
(t) = t sin(t)
45
7 Oscil lations in 2d
Now, we see
x = x
0
+x
1
+O(
2
) (1 t) sin(t).
Again, we end up with a term that is linear in t. The solution to this problem
comes from singular perturbation theory. We separate time scale into a fast time
= t and a slow time T = t. This procedure is called two timing.
We use the new ansatz:
x(t) = x
0
(, T) +x
1
(, T) +O(
2
)
and therefore calculate the derivatives
x =

x
t
+
T
x
t
T
=

x +
T
x
=

x
0
+(

x
1
+
T
x
0
) +O(
2
)
x =

x
0
+(

x
1
+ 2
T
x
0
) +O(
2
)
and plug them into the dynamic equation
0 =

x
0
+(

x
1
+ 2
T
x
0
) + 2

x
0
+ (x
0
+x
1
) +O(
2
).
O(1) :

x
0
+x
0
= 0
O() :

x
1
+x
1
= 2(
T
x
0
+

x
0
)
O(1) : x
0
= A(T) sin() +B(T) cos()
O() :

x
1
+x
1
= 2 (A

(T) +A(T)) cos() + 2 (B

(T) +B(T)) sin()

In order to end up with a well-behaved solution, demand the prefactors (A

+ A)
and (B

+B) to be zero.
A = A
0
exp(T), B = B
0
exp(T)
For the initial conditions x(0) = 0, x(0) = 1 B
0
= 0, A
0
= 1, the general
solution is sine-shaped with an envelope decaying in time
x
0
= exp(T) sin() = exp(t) sin().
This is identical to the exact solution in order O(
2
).
46
7 Oscil lations in 2d
Application to van der Pol oscillator
The dynamic equation
x +(x
2
1) x +x = 0
holds
[

x
0
+(

x
1
+ 2
T
x
0
)] +(x
2
0
1)

x
0
+ (x
0
+x
1
) = 0.
O(1) :

x
0
+x
0
= 0
O() :

x
1
+x
1
= 2
T
x
0
(x
2
0
1)

x
0
In polar coordinates:
x
0
= r(T) cos( + (T))

x
1
+x
1
= 2[r

sin( + ) +r

cos( + )] +r sin( + )[r

2
cos
2
( + ) 1]
[2r

r +
1
4
r
3
] sin( + ) + 2r

cos( + ) +
1
4
r
3
sin(3( + )).
In the last step, we used the relation sin() cos
2
() =
1
4
[sin() + sin(3)]. In order to
avoid a resonance catastrophe, demand the prefactors [2r

r +
1
4
r
3
] and 2r

to be zero.
r

=
1
8
r(4 r
2
) and

= 0
The result is a logistic growth in r.
There is a xed point for
r

= 2 and = const =
0
.
Note, that we get a limit cycle irrespec-
tive of the value of . Therefore, is a
singular perturbation.
Averaging method:
We now discuss a more general method to solve these kinds of problems. Consider again
x +x +h(x, x) = 0, which represents a large class of non-linear oscillators.

x
0
+x
0
= 0

x
1
+x
1
= 2
T
x
0
h
x
0
= r(T) cos( + (T))

x
1
+x
1
= 2[r

sin( + ) +r

cos( + )] h
47
7 Oscil lations in 2d
Since we have h(x, x) = h(sin( + ), cos( + )), a 2-periodic function in = + ,
we know the Fourier expansion h():
h() =

k=0
a
k
cos(k) +

k=1
b
k
sin(k).
Up to O(), the only resonant terms for x
1
are (2r

b
1
) sin() and (2r

a
1
) cos().
From this, we get the conditions
r

=
b
1
2
, r

=
a
1
2
to avoid the resonance catastrophe. We write the prefactors in terms of averages of :
a
1
=
1

_
2
0
d h() cos()
= 2 hcos())

b
1
= 2 hsin())

= hsin())

= hcos())

.
_
dynamical equations for (r, )
Example: van der Pol oscillator
Consider h = (x
2
1) x = (r
2
cos
2
() 1)(r sin()).
r

= hsin())

= r
3
_
cos
2
() sin
2
()
_

+r
_
sin
2
()
_
=
1
2
r
1
8
r
3
=
1
8
r(4 r
2
) r

= 2
r

= hcos())

= rsin() cos())

. .
=0
r
3
_
cos
3
() sin()
_

. .
=0
= 0 = const
This is the same result as before.
48
8 Bifurcations in 2d
8 Bifurcations in 2d
Like in 1d, in 2d existence and stability of xed points depend on the parameters of the
system. In contrast to 1d, however, now also oscillations can be switched on and o. As
an example, look at the substrate-depletion-oscillator.
There are three types of bifurcations in 2d:
1. 1d-like bifurcations (4 types)
2. Hopf bifurcation (local switch on/o of oscillations)
3. global bifurcations
In the following, we have a closer look at them.
1. 1d-like bifurcations
All four types of 1d bifurcations exist in 2d (cf. center manifold theorem).
Example: Grith model for genetic switch
Consider a gene that codes for a certain protein. The activity of the gene shall
be induced by the protein and its copies which are translated from the messenger
RNA. The system is described as
x = y ax
y =
x
2
1 +x
2
by,
where, x and y are the concentrations of the protein and the mRNA, respectively.
49
8 Bifurcations in 2d
The following gure shows a protein acting as transcription factor.
The nullclines are y = a x and y =
x
2
b(1+x
2
)
. We see that the system depends on
the parameters a, b. For increasing a, the two upper xed points approach each
other until they fall together when the nullclines intersect tangentially. For even
larger a, only the xed point in the origin remains.
The two upper xed points are given by
x

= ab(1 +x

2
)
x

=
1

1 4a
2
b
2
2ab
.
For 2ab = 1, the xed points collide. The critical values are
a
c
=
1
2b
x

c
= 1.
If a < a
c
, the system is bistable and acts like a genetic switch: depending on the
initial conditions, the gene is on or o.
We note that in this case, the 2d analysis gives essentially the same results like
a 1d analysis along the x-nullcline. In fact, this is an example of a saddle-node
bifurcation.
50
8 Bifurcations in 2d
The prototype of a saddle-node bifurcation
in 2d is
x = x
2
y = y.
The phase portrait varies with .
One xed point is a saddle, x

= (

, 0), the other one a stable node, x

=
(+

, 0). The behavior of the saddle depends on the eigenvalues of the Jacobian
A =
_
2x

0
0 1
_
with x

= (

, 0). The bifurcation is given for = 0. This is

called a zero eigenvalue-bifurcation. It exists for all bifurcations from type 1).
Recall that the eigenvalues can be calculated using

1/2
=

2
4
2
.
Either both eigenvalues are real (shown on the left), which corresponds to (
2

4) > 0, or they are complex (shown on the right), (

2
4) < 0
1/2
=

2
i.
The saddle-node bifurcation is of the rst type. We now turn to the second type.
If the complex eigenvalues
1/2
=

2
i cross the yaxis from left to right,
oscillations are switched on. This is called Hopf bifurcation.
51
8 Bifurcations in 2d
2. Hopf bifurcation
a) Supercritical Hopf bifurcation
r = r r
3

= +br
2
The systems dynamics strongly depends on .
In order to analyze the behavior of the eigenvalues during the bifurcation,
rewrite the system in Cartesian coordinates:
x = r cos()
y = r sin()
x = r cos() r sin()

= x y +O(xr
2
, yr
2
)
y = x +y
A =
_


_

1/2
= i
52
8 Bifurcations in 2d
Thus, if we increase , the eigenvalues cross the imaginary axis from left to
right as expected.
The structure is similar to the supercritical pitchfork bifurcation. But now,
we have a supercritical Hopf bifurcation.
b) Subcritical Hopf bifurcation
r = r +r
3
r
5

= +br
2
The term (r
5
) causes certain trajectories to drive away from the origin. For
< 0, both exist a stable limit cycle and an attractive xed point in the
origin. But for increasing , the attractive area around the origin decreases
and nally disappears for = 0. That is when the subcritical Hopf bifurcation
occurs. The origin is now unstable and there is an abrupt transition to large-
amplitude oscillations.
53
8 Bifurcations in 2d
Note that a Hopf bifurcation theorem exists, demanding rigorous conditions on

1/2
for a Hopf bifurcation to occur.
3. Global bifurcations
Apart from Hopf bifurcations, global bifurcations are another way in which limit
cycles are created or destroyed. They are a combination of 1) and 2). Cycle
interactions with other xed points exist. As well, there is a global change in ow
structure.
a) Saddle-node bifurcation of circles
This is a prototypical example of global bifurcations.
r = yr +r
3
r
5

= +br
2

c
=
1
4
b) Innite period bifurcation
54
9 Excitable systems
r = r(1 r
2
)

= sin()

c
= 1
c) Homoclinic bifurcation
x = y
y = y +x x
2
+xy

c
8.6
9 Excitable systems
In this chapter, we discuss so-called excitable systems. One example is grass, that is
burned down and then regrows. Obviously, this process can be repeated over and over
again. We also see that such a process can occur in the spatial domain as a wave. Our
biological example is neuronal activity in the brain. Here, the wave is called an action
potential and it travels along the axon of a neuron.
55
9 Excitable systems
A human neuron has a diameter of r
50 m for the cell body and an axon
that is up to 1 m long. The axon trans-
ports the signals called action potentials
or spikes. We think about them as trav-
eling waves V (x, t).
It is approximately cylindrically shaped
containing mainly potassium (K
+
) and
organic anions (A

) and being sur-

rounded predominantly by sodium
(Na
+
), calcium (Ca
2+
) and chlorine
(Cl

).
Typical values of the concentrations of the two dominant ions potassium and sodium
are given in the following table:
inside outside V
[mM] [mM] [mV]
K
+
155 4 -98
Na
+
12 154 67
The axon membrane is typically 5 nm thick and made of dielectric material ( = 2) which
is an insulator. To allow ion ow, there are many channels in the surface. Channels
can dynamically open and close. The resting potential is approximately V = 60 mV.
Here is an equivalent electrical circuit:
A biological membrane acts as a capaci-
tor for which the equation C
A
=
C
A
=

0

d
holds. Typically,
C
A
is around
F
cm
2
. The
timescale is =
C
A
g
A
=
C
g
2 ms with
the conductivity per area g
A
= 5
1
m
2
.
56
9 Excitable systems
Assuming equilibrium, the concentration c is proportional to a Boltzmann distribution
c exp(eV/k
B
T). The ratio of the concentrations inside and outside of the axon
therefore depends on the dierence of the potential
c
i
c
o
exp(e(V
i
V
o
)/k
B
T). Solving
for the dierence of the potential, this yields the Nernst potential :
V =
k
B
T
e
ln
_
c
i
c
o
_
.
This equation has been used to calculate V in the previous table (
k
B
T
e
= 25 mV). Since
V
1
,= V
2
, the system is not in equilibrium. Therefore, we need dynamical models to
understand.
The exact form of an action potential can be measured with a space clamp. Taking
a giant axon of a squid and leading a silver wire into the axon, the action potential
V (t) = V
i
(t) V
o
(t) can be determined independent from spatial components.
Lets have a short look at the historical background.
1937/38 Hodgkin worked at Woods Hole and invented with others the space clamp.
1947 He performed experiments at Plymouth together with the student Huxley.
1952 They published ve papers in the Journal of Physiology.
1963 They got the Nobel prize for their work.
1991 Neher and Sakmann got the Nobel prize for nding the rst evidence
for ion channels using the patch clamp method.
Hodgkin-Huxley model (1952)
Starting with the electrical circuit,
Hodgkin and Huxley invented a model
for action potentials. They assumed a
linear relation between voltage and cur-
rent
I
Na
= g
Na
(V V
Na
).
57
9 Excitable systems
Faradays law states Q = C V I = C

V .

V =
1
C
_

_g
Na
(V V
Na
) +g
K
(V V
K
) +g
L
(V V
L
)
. .
leakage current
_

_
The leakage current contains all contributions except those from potassium and sodium.
This equations describes a linear system
with a stable xed point. If a little per-
turbation occurs the system will relax-
ing back. But this behavior does not t
to the experiments.
Hence, to get an action potential we consider a conductivity that depends on voltage,
g = g(V ). The basic idea is a two-state process for opening and closing of the gate.
Such a process can be described in the following way using the opening and closing
probabilities
n
and
n
, respectively.
n =
n
(1 n)
n
n.
A closed gate corresponds to n = 0, an open one to n = 1. Therefore, the time-dependent
processes are
opening: n(t) = 1 exp(t)
closing: n(t) = exp(t).
A voltage clamp experiment showed a
good agreement for the closing process.
But the t result of the opening pro-
cess corresponds to power four. In gen-
eral, higher powers and three gates are
needed to t the data.
We use a four-dimensional model (V, n, m, h) dened by the additional equation
n =
n
(1 n)
n
n
m =
m
(1 m)
m
m

h =
h
(1 h)
h
h.
58
9 Excitable systems
n, m and h label the potassium and sodium activation and the sodium deactivation
gates, respectively. Phenomenologically, the coupling to the conductivities is
g
Na
= 120 m
3
h
g
K
= 36 n
4
g
L
= 0.3.
The justication for these microscopic processes (the ion channels) came only 30 years
later. We can interpret the powers as
n
4
: potassium 4 gates open
m
3
: sodium 3 gates open
h: sodium 1 gate closed
The parameters (
i
,
i
) are non-trivial functions of V . For example for n

n
=
10 V
100(exp((10 V )/10) 1)
,
n
=
125 exp(V/80)
1000
.
Today, again, the six parameters can be understood from the physics of ion channels.
Now, weve got a complete 4d model and we have to integrate the ODEs. We end up
with the following time-dependence of the gates and the conductivities.
59
9 Excitable systems
A more general NLD-analysis shows that the system can oscillate.
Up to now, we did not consider the eects of space.
Describing the signal as traveling wave, Ohms law holds in lateral direction
V (x + x) V (x) = I(x) R.
Looking at a node of the circuit, current conservation demands
I(x x) I(x) = g
A
V (x).
Taking x 0, this yields
V

(x) =

r
2
I(x)
I

(x) = g
A
2r V (x)
V

(x) =
1

2
V
with the radius r of the axon (r 0.25 mm for a squid), the conductivity of the
medium and the resistance per length

r
2
= 0.3 m. In the last step, the decay length
=
_
r
2g
A
= 9 mm has been introduced.
60
9 Excitable systems
Injecting a voltage V
0
at left, the space dependence is
V (x) = V
0
exp(
x

).
Because of this, the signal would just decay if the conductivities were independent from
the voltage.
We next combine the longitudinal equation with the full transverse Hodkin-Huxley cur-
rent for the current conservation
I

= [g
Na
(V V
Na
N
) +g
K
(V V
K
N
) +g
L
(V V
L
N
)] C
U
t
.

2
V

V =
g
Na
(V V
Na
N
) +g
K
(V V
K
N
) +g
L
(V V
L
N
)
g
.
The result is a non-linear PDE, the cable equation with time-dependent conductivity.
Considering one type of channel (e.g. Na) and injecting a voltage V
0
at the left, a front
propagates to the right. To get a wave propagation, one has to add a counteracting
process (e.g. opening of K channels). Hodgkin and Huxley showed the waves in 1952.
To understand this better, we consider the bistable cable equation

V = V

+f(V )
with f(V ) = V (V )(V 1). We look for solutions of the form V (x, t) = U(x+c t)
which describe waves propagating from right to left with velocity c. Dening y := x+ct,
we can convert the system into an ODE.

y
U c =
2
y
U +f(U)
Now, we do a phase plane analysis and therefore dene
W =
y
U
y
W = c W f(U), f(U) = U(U )(U 1).
A traveling front solution must connect the xed points (U = 0, W = 0) and (U =
1, W = 0) in the (U, W)-plane as we vary y from to +. There is a unique c

which results in such a trajectory. It can be calculated analytically. For this, we guess
that the connection between the two resting states is given by
W(U) =
dU
dy
= BU(U 1).
61
9 Excitable systems
So, we calculate
0 =
y
U c +
2
y
U +f(U)
= W c +
y
W +f(U)
= cBU(U 1) +B
2
(2U 1) U(U 1) U(U )(U 1)
= cB +B
2
(2U 1) (U )
and nd
B =
1

2
, c =
1

2
(1 2).
The speed is a decreasing function of and the direction of propagation changes at
=
1
2
. In this case, there is no propagation. The prole of the traveling front is found
by integrating the assumption
W(U) =
dU
dy
= BU(U 1).
U(y) =
1
2
_
1 + tanh
_
1
2

2
y
__
.
A propagating wave is a trajectory which comes back to the original state. This occurs in
the Hodgkin-Huxley model as well as in the Fitzhugh-Nagumo model (which is discussed
below) with diusive coupling

V = V

+f(V, W)

W = g(V, W).
Until now, we studied 1d wave propagation. The cable equation can also be extended to
2d or 3d. We then obtain propagation fronts (planar or circular), but also spirals. In the
context of heart and muscle biology, spirals are a sign of a pathophysiological situation.
Fitzhugh-Nagumo model
In depth analysis of the Hodgkin-Huxley model, Fitzhugh and Nagumo tried to analyze
the phase plane for a cut through phase space. On the rst attempt, they used two
variables V and m and kept n and h constant. The two equations then read
e

V = g
K
n
4
0
(V V
K
) g
Na
m
3
h
0
(V V
Na
) g
L
(V V
L
)
m =
(V )
m
(1 m)
(V )
m
m
with the parameters g denoting constants.
62
9 Excitable systems
The phase plane dependency of m and V is shown in the gure below.
This analysis explains the threshold but not the relaxation. Because of this, on the
second attempt Fitzhugh and Nagumo introduced one fast variable x =(V, m) responsible
for the excitation and a slow variable y =(h, n) responsible for the relaxation. z = z(t)
denotes the stimulus and c is an inverse time constant. This yields the Fitzhugh-Nagumo
model which is a simplication of the Hodgkin-Huxley model:
x = c (y +x
x
3
3
+z(t))
y =
x a +by
c
.
For a = b = z = 0, this is a van der Pol oscillator which leads to relaxation oscillations.
The phase portraits are shown in comparison: left the van der Pol oscillator and in the
middle and right the Fitzhugh-Naguma model with dierent parameters. The middle
gure shows a large excursion ending in a stable xed point (the action potential) and
the right one that oscillations are possible for certain parameters.
63
10 Pattern formation in reaction-diusion systems
10 Pattern formation in reaction-diusion systems
One of the most fascinating subjects in mathematical biology is called morphogenesis:
a part of embryology leading to various patterns and forms. The pioneering paper was
published in 1952 by Alan Turing. He wondered under which condition a reaction-
diusion system produces a heterogeneous spatial pattern. To answer this question, he
considered a two-dimensional system of the type:

A = F(A, B) +D
A
A

B = G(A, B) +D
B
B.
The simplest case was studied by J. Schnackenberg with the functions F and G
F = k
1
k
2
A +k
3
A
2
B
G = k
4
k
3
A
2
B.
Thus, A is autocatalytic and inhibits B, while B activates A. We rst non-dimensionalize
the system:
u = A
_
k
3
k
2
_
2
, v = B
_
k
3
k
2
_
2
,
t =
D
A
t
L
2
, x =
x
L
, d =
D
B
D
A
,
a =
k
1
k
2
_
k
3
k
2
_
1/2
, b =
k
4
k
2
_
k
3
k
2
_
1/2
,
=
L
2
k
2
D
A
Note that the variables u and v are positive since they are concentrations of reactants.
By introducing the variables above, the system is described as follows
u = (a u +u
2
v)
. .
=:f(u,v)
+ u
v = (b u
2
v)
. .
=:f(u,v)
+dv
with the ratio of the diusion d and the relative strength of the reaction versus the
diusion terms which scales as L
2
. We start with a linear stability analysis of this
general form with a zero-ux boundary condition and a given initial congurations for
64
10 Pattern formation in reaction-diusion systems
u and v. We rst consider the case without diusion D = 0 and ask for a homogeneous
state which is stable; we then ask under which conditions diusion leads to an instability,
the Turing instability.
Stabilizing diusion. 1D reaction-diusion system
For single reactant u(x, t) there is no Turing instability:
u
t
= Du
xx
+f(u), x [0, l] + BC
An uniform (x - independent) ODE
u = f(u)
has either: stable m = f

(u

) < 0, or unstable m > 0 equilibrium at stationary point

u

: f(u

) = 0.
The diusion will maintain a stable one, and it can stabilize the unstable one.
To check it we take the linearized problem about u

, for v(x, t) = u(x, t) u

:
_
_
_
v
t
= Dv
xx
+mv, x [0, l]
m = f

(u

)
We use the Neumann BC (no ux):

x
v

0,l
= 0
and eigenfunction expansion.
Separation of variables: look for the solution in the form:
v(x, t) = (t) (x)
Leading to:

= (D

+m )

=
D

+m

=:
Hence, the solution stability depends on the sign of :
(t) exp(t)
65
10 Pattern formation in reaction-diusion systems
The second equation:
D

+ (m) = 0, (0, l) = 0
or

+
2
= 0,
2
=
m
D
The eigenfunctions are:

k
(x) = cos
k
x = cos
kx
l
, k = 0, 1, ...
Using the expression for we get:
= mD
_
kx
l
_
2
So:
stable case (m < 0) remains stable,
For unstable case the system get bifurcation values at D
_
kx
l
_
2
= m. The diusion
stabilize the unstable uniform equillibrium.
2d reaction-diusion system
We now turn to two dimensions, where the Turing instability occurs. So lets start with
a linear stability analysis of the reaction part using

W =
_
u u

v v

_
. We denote the
steady state with

W

=
_
u

_
and a partial derivative with f
u
=
f
u
etc. This yields

W = A

W
with the matrix
A =
_
f
u
f
v
g
u
g
v
_
[

W

.
Linear stability is guaranteed if the real part of the eigenvalues is smaller than zero,
Re < 0. Thus, the trace of A is smaller than zero
tr A = f
u
+g
v
< 0 (10.1)
and the determinant larger than zero
det A = f
u
g
v
f
v
g
u
> 0. (10.2)
66
10 Pattern formation in reaction-diusion systems
The u- and v- nullcline is given by setting f = 0 and g = 0, respectively.
u-nullcline: v =
u a
u
2
v-nullcline: v =
b
u
2
For the steady state

W

=
_
u

_
, we demand u

and v

to be positive for physical

reasons.
(u

, v

) =
_
a +b,
b
(a +b)
2
_
Thus, it is a +b > 0 and b > 0.
A =
_
1 + 2uv u
2
2uv u
2
_
[

W

=
_
ba
b+a
(a +b)
2
2ab
a+b
(a +b)
2
_
det A = (a +b)
2
> 0
We get a stable spiral for b = 2 and
a = 0.2.
We now turn to the full reaction-diusion system and linearize it about the steady state

W = A

W +D

W
with D =
_
1 0
0 d
_
.
In order to obtain an ODE from this PDE, we use the Fourier solutions of the spatial
eigenvalue problem

W +k
2

W = 0
with no-ux boundary of size p in 1d, we have

W
k
(x) cos(k x)
67
10 Pattern formation in reaction-diusion systems
with wavenumber k =
n
p
and wavelength =
2
k
=
2p
n
(n integer).

W(r, t) =

k
c
k
exp(t)

W
k
(r)

W
k
= A

W
k
Dk
2

W
k
We now have to solve this eigenvalue problem. A Turing instability occurs if Re (k) > 0.
Our side constraint is that the eigenvalue problem for D = 0 (only reactions) is assumed
to be stable, that is Re (k = 0) < 0.
0 =
2
+[k
2
(1 +d) tr A] + [dk
4
(df
u
+g
v
)k
2
+
2
det A].
We rst note that the coecient of is always positive because k
2
(1 + d) > 0 and tr
A < 0 (for reasons of the stability of the reaction system). In order to get an instability,
Re > 0, the constant part has to be negative. Since the rst and last terms are
positive, this implies
df
u
+g
v
> 0 d ,= 1. (10.3)
This is the main result by Turing: An instability can occur if one component diuses
faster than the other. (10.3) is only a necessary, but not a sucient condition. We
require that the constant term as a function of k
2
has a negative minimum.
(df
u
+g
v
)
2
4d
> det A = f
u
g
v
f
v
g
u
(10.4)
The critical wavenumber can be calculated to be
k
c
=
_
det A
d
c
_
1/2
with the critical diusion constant from
d
2
c
f
2
u
+ 2(2f
v
g
u
f
u
g
v
)d
c
+g
2
v
= 0.
For d > d
c
, we have a band of instable wavenumbers. The relation = (k
2
) is called
the dispersion relation. The maximum singles out the fastest growing mode. This one
dominates the solution

W(r, t) =

k
c
k
exp((k
2
)t)
for large t. Note however, that in this case also non-linear eects will become important
and thus will determine the nal pattern.
In summary, we have found four conditions (10.1) - (10.4) for the Turing instability to
occur. We now analyze the Schnackenberg-model in one spatial dimension. We already
noted that a+b > 0 and b > 0 for the steady state to make sense. From the phase plane
we see that f > 0 for large u and f < 0 for small u. Hence, f
u
> 0 around the steady
68
10 Pattern formation in reaction-diusion systems
state. Thus, b > a.
All in all, there are four relations:
f
u
, f
v
> 0 and g
u
, g
v
< 0.
A
_
+ +

_
From condition (10.1) and (10.3), we now calculate that d > 1 in this case (the activation
B diuses faster in this model). In general, the conditions (10.1)-(10.4) dene a domain
in (a, b, d)space, the Turing space, in which the instabilities occurs. The structure
of the matrix A tells us how this will happen: as u or v increases, u increases and v
decreases. So, the two species will grow out of phase.
The domain size p has to be large enough for a wavenumber k =
n
p
to be within the
range of the unstable wavenumbers ( L
2
):
L(a, b, d) <
_
n
p
_
2
< M(a, b, d)
where L and M can be calculated exactly. Typically, the mode which grows has n = 1.
Whether the left or right solution occurs depends on the initial conditions. If the domain
grows to double size, than changes by four ( L
2
). p stays the same because it is
measured in units of L. Now, the mode n = 2 behaves as shown in the following gures.
On this way, a growing system will develop a stripe pattern.
69
10 Pattern formation in reaction-diusion systems
In d=2 spatial dimensions, many more possible scenarios exist in a model-dependent
manner: stripes (left), checkerboard (middle), hexagonal or triangular (right) patterns,
etc.
In the last 20 years, several steady state nite amplitude spatial patterns (Turing pat-
terns) have been found experimentally. Today, the corresponding models are usually
explored numerically, thus making it easy to also include the non-linear parts.
However, if this is the way the zebra gets its stripes, is still an open question.
70